diff --git a/.gitattributes b/.gitattributes index 649a15e2971283cc7fe168d2478d7d9b2053555e..aa6b718948fd6dbd809f4a0347fb5622dc4f88a6 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5849,3 +5849,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/Efficient[[:space:]]Continue[[:space:]]Training[[:space:]]of[[:space:]]Temporal[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Structural[[:space:]]Information/4e01312d-f9ea-48a2-ac0d-45b87a14409b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Efficient[[:space:]]Cross-Task[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Few-Shot[[:space:]]Conversational[[:space:]]Emotion[[:space:]]Recognition/d55c3273-2161-46ae-adc0-b3cf30e5ba4b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Efficient[[:space:]]Data[[:space:]]Learning[[:space:]]for[[:space:]]Open[[:space:]]Information[[:space:]]Extraction[[:space:]]with[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models/2930ad8d-b924-4379-920e-36da32fb8d1e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Efficient[[:space:]]Latent[[:space:]]Variable[[:space:]]Modeling[[:space:]]for[[:space:]]Knowledge-Grounded[[:space:]]Dialogue[[:space:]]Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Efficient[[:space:]]Long-Range[[:space:]]Transformers_[[:space:]]You[[:space:]]Need[[:space:]]to[[:space:]]Attend[[:space:]]More,[[:space:]]but[[:space:]]Not[[:space:]]Necessarily[[:space:]]at[[:space:]]Every[[:space:]]Layer/b2cbac77-9108-42c4-9fd3-211395757830_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Efficient[[:space:]]Multilingual[[:space:]]Language[[:space:]]Model[[:space:]]Compression[[:space:]]through[[:space:]]Vocabulary[[:space:]]Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Efficient[[:space:]]k-NN[[:space:]]Search[[:space:]]with[[:space:]]Cross-Encoders[[:space:]]using[[:space:]]Adaptive[[:space:]]Multi-Round[[:space:]]CUR[[:space:]]Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Efficiently[[:space:]]Enhancing[[:space:]]Zero-Shot[[:space:]]Performance[[:space:]]of[[:space:]]Instruction[[:space:]]Following[[:space:]]Model[[:space:]]via[[:space:]]Retrieval[[:space:]]of[[:space:]]Soft[[:space:]]Prompt/91bfba92-6122-4a1c-8806-4f5015582836_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Emergent[[:space:]]Inabilities_[[:space:]]Inverse[[:space:]]Scaling[[:space:]]Over[[:space:]]the[[:space:]]Course[[:space:]]of[[:space:]]Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Empowering[[:space:]]Psychotherapy[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]Cognitive[[:space:]]Distortion[[:space:]]Detection[[:space:]]through[[:space:]]Diagnosis[[:space:]]of[[:space:]]Thought[[:space:]]Prompting/04980a83-f896-4166-a550-6164ef62b39f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Emptying[[:space:]]the[[:space:]]Ocean[[:space:]]with[[:space:]]a[[:space:]]Spoon_[[:space:]]Should[[:space:]]We[[:space:]]Edit[[:space:]]Models_/04529405-4e73-4a21-827a-c746e1975cce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enabling[[:space:]]Unsupervised[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Word-level[[:space:]]Visual[[:space:]]Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/End-to-End[[:space:]]Autoregressive[[:space:]]Retrieval[[:space:]]via[[:space:]]Bootstrapping[[:space:]]for[[:space:]]Smart[[:space:]]Reply[[:space:]]Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/End-to-end[[:space:]]Adversarial[[:space:]]Sample[[:space:]]Generation[[:space:]]for[[:space:]]Data[[:space:]]Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Energy[[:space:]]and[[:space:]]Carbon[[:space:]]Considerations[[:space:]]of[[:space:]]Fine-Tuning[[:space:]]BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhanced[[:space:]]Simultaneous[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Word-level[[:space:]]Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Abstractiveness[[:space:]]of[[:space:]]Summarization[[:space:]]Models[[:space:]]through[[:space:]]Calibrated[[:space:]]Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Accessible[[:space:]]Communication_[[:space:]]from[[:space:]]European[[:space:]]Portuguese[[:space:]]to[[:space:]]Portuguese[[:space:]]Sign[[:space:]]Language/d1d9cef1-f716-4931-80c7-979f571de27b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Argument[[:space:]]Structure[[:space:]]Extraction[[:space:]]with[[:space:]]Efficient[[:space:]]Leverage[[:space:]]of[[:space:]]Contextual[[:space:]]Information/5994108c-c5db-4e48-8a82-84394aa38bd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Conversational[[:space:]]Search_[[:space:]]Large[[:space:]]Language[[:space:]]Model-Aided[[:space:]]Informative[[:space:]]Query[[:space:]]Rewriting/28ee10b5-c87b-4d0e-bf9d-1bdf62700be6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Emotion[[:space:]]Recognition[[:space:]]in[[:space:]]Conversation[[:space:]]via[[:space:]]Multi-view[[:space:]]Feature[[:space:]]Alignment[[:space:]]and[[:space:]]Memorization/17140297-58bc-40ff-8c98-7cf6826f913f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Semantic[[:space:]]Units/0d4671ac-5112-4116-b8a5-14c52dd9b5ee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Reasoning[[:space:]]Capabilities[[:space:]]by[[:space:]]Instruction[[:space:]]Learning[[:space:]]and[[:space:]]Chain-of-Thoughts[[:space:]]for[[:space:]]Implicit[[:space:]]Discourse[[:space:]]Relation[[:space:]]Recognition/7420c6ad-7f33-41fb-93fd-25202ce16a30_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Retrieval-Augmented[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Iterative[[:space:]]Retrieval-Generation[[:space:]]Synergy/f3064320-be4a-4110-a373-933f61879a44_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Scalability[[:space:]]of[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Efficient[[:space:]]Parameter[[:space:]]Sharing/3932f23d-3769-4987-b111-0647844ee92c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Enhancing[[:space:]]Text-to-SQL[[:space:]]Capabilities[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Study[[:space:]]on[[:space:]]Prompt[[:space:]]Design[[:space:]]Strategies/e8887642-37cb-41f2-9b01-f9b63f1db872_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Ensemble-Instruct_[[:space:]]Instruction[[:space:]]Tuning[[:space:]]Data[[:space:]]Generation[[:space:]]with[[:space:]]a[[:space:]]Heterogeneous[[:space:]]Mixture[[:space:]]of[[:space:]]LMs/d2596f0c-003a-4753-9668-60e19104ae23_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Entity[[:space:]]Disambiguation[[:space:]]on[[:space:]]a[[:space:]]Tight[[:space:]]Labeling[[:space:]]Budget/2c04b8d1-9e8c-4ffc-bffc-ac1bf6c8dc26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Entity-Based[[:space:]]Evaluation[[:space:]]of[[:space:]]Political[[:space:]]Bias[[:space:]]in[[:space:]]Automatic[[:space:]]Summarization/08ae56a4-e6e9-4b36-a19c-00576429cf26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Epsilon[[:space:]]Sampling[[:space:]]Rocks_[[:space:]]Investigating[[:space:]]Sampling[[:space:]]Strategies[[:space:]]for[[:space:]]Minimum[[:space:]]Bayes[[:space:]]Risk[[:space:]]Decoding[[:space:]]for[[:space:]]Machine[[:space:]]Translation/48fcc534-7bee-4863-bdb2-b6a2234b21dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Error[[:space:]]Detection[[:space:]]for[[:space:]]Text-to-SQL[[:space:]]Semantic[[:space:]]Parsing/d355af85-b1e6-44f4-a71f-a816fe8d3ccc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Estimating[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Capabilities[[:space:]]without[[:space:]]Labeled[[:space:]]Test[[:space:]]Data/564e32e7-5d1d-48a4-9a3a-317db9c128cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Ethical[[:space:]]Reasoning[[:space:]]over[[:space:]]Moral[[:space:]]Alignment_[[:space:]]A[[:space:]]Case[[:space:]]and[[:space:]]Framework[[:space:]]for[[:space:]]In-Context[[:space:]]Ethical[[:space:]]Policies[[:space:]]in[[:space:]]LLMs/34213480-0ccb-4047-90fc-d32c1764ae27_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]Dependencies[[:space:]]in[[:space:]]Fact[[:space:]]Editing[[:space:]]for[[:space:]]Language[[:space:]]Models_[[:space:]]Specificity[[:space:]]and[[:space:]]Implication[[:space:]]Awareness/3640cf72-25d4-441f-993a-55e025a4950c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]Emotion[[:space:]]Arcs[[:space:]]Across[[:space:]]Languages_[[:space:]]Bridging[[:space:]]the[[:space:]]Global[[:space:]]Divide[[:space:]]in[[:space:]]Sentiment[[:space:]]Analysis/a57e16b5-2fc7-4515-b210-640caeab99f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]Parameter-Efficient[[:space:]]Finetuning[[:space:]]Approaches[[:space:]]for[[:space:]]Pre-trained[[:space:]]Models[[:space:]]on[[:space:]]the[[:space:]]Financial[[:space:]]Domain/d0ac477d-9b37-4d97-b2db-d6bcdaf34946_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]Subjective[[:space:]]Cognitive[[:space:]]Appraisals[[:space:]]of[[:space:]]Emotions[[:space:]]from[[:space:]]Large[[:space:]]Language[[:space:]]Models/aa4b53ff-a1c5-4a7a-9481-af63ece728ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]Verifiability[[:space:]]in[[:space:]]Generative[[:space:]]Search[[:space:]]Engines/ba1d3d57-179d-4030-8083-e23e2548fdca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]and[[:space:]]Enhancing[[:space:]]the[[:space:]]Robustness[[:space:]]of[[:space:]]Code[[:space:]]Pre-trained[[:space:]]Models[[:space:]]through[[:space:]]Structure-Aware[[:space:]]Adversarial[[:space:]]Samples[[:space:]]Generation/921f7e52-3ed3-4834-831a-fc47235920be_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Evaluating[[:space:]]the[[:space:]]Knowledge[[:space:]]Base[[:space:]]Completion[[:space:]]Potential[[:space:]]of[[:space:]]GPT/3bc7a74b-d239-4b8b-b2cb-3deafa244b6b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Examining[[:space:]]Inter-Consistency[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Collaboration_[[:space:]]An[[:space:]]In-depth[[:space:]]Analysis[[:space:]]via[[:space:]]Debate/b32bdb31-1d53-4040-95b7-04c960da1c7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Example-based[[:space:]]Hypernetworks[[:space:]]for[[:space:]]Multi-source[[:space:]]Adaptation[[:space:]]to[[:space:]]Unseen[[:space:]]Domains/6a25d3ed-bb96-4cf8-8aff-a9b986445a61_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Execution-Based[[:space:]]Evaluation[[:space:]]for[[:space:]]Open-Domain[[:space:]]Code[[:space:]]Generation/50b60d58-a06a-4443-b30a-7189530b6241_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ExpNote_[[:space:]]Black-box[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]better[[:space:]]Task[[:space:]]Solvers[[:space:]]with[[:space:]]Experience[[:space:]]Notebook/2318b8d2-88c6-4b2c-b206-d536a2a8c7ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Explain-then-translate_[[:space:]]an[[:space:]]analysis[[:space:]]on[[:space:]]improving[[:space:]]program[[:space:]]translation[[:space:]]with[[:space:]]self-generated[[:space:]]explanations/6961a868-bbf6-4496-9b2a-85639430657c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ExplainCPE_[[:space:]]A[[:space:]]Free-text[[:space:]]Explanation[[:space:]]Benchmark[[:space:]]of[[:space:]]Chinese[[:space:]]Pharmacist[[:space:]]Examination/a72e8b93-6ab1-47c4-a077-cb4fdc5162f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Explainable[[:space:]]Claim[[:space:]]Verification[[:space:]]via[[:space:]]Knowledge-Grounded[[:space:]]Reasoning[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/18a84354-cfce-4db6-b179-cd851ed39019_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Explicit[[:space:]]Alignment[[:space:]]and[[:space:]]Many-to-many[[:space:]]Entailment[[:space:]]Based[[:space:]]Reasoning[[:space:]]for[[:space:]]Conversational[[:space:]]Machine[[:space:]]Reading/59c18b14-085d-4a8c-b402-a89170c5790b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploiting[[:space:]]Contrastive[[:space:]]Learning[[:space:]]and[[:space:]]Numerical[[:space:]]Evidence[[:space:]]for[[:space:]]Confusing[[:space:]]Legal[[:space:]]Judgment[[:space:]]Prediction/d05877f8-fd38-4c60-9186-57237bcbef9d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploiting[[:space:]]Emotion-Semantic[[:space:]]Correlations[[:space:]]for[[:space:]]Empathetic[[:space:]]Response[[:space:]]Generation/422244f0-35c6-4ef5-a4ee-44a79edcd966_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Explore[[:space:]]the[[:space:]]Way_[[:space:]]Exploring[[:space:]]Reasoning[[:space:]]Path[[:space:]]by[[:space:]]Bridging[[:space:]]Entities[[:space:]]for[[:space:]]Effective[[:space:]]Cross-Document[[:space:]]Relation[[:space:]]Extraction/811743fd-9262-48c8-bc24-8536da413fcb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]Context-Aware[[:space:]]Evaluation[[:space:]]Metrics[[:space:]]for[[:space:]]Machine[[:space:]]Translation/99043eb5-44fb-4cb0-86ae-9780d6a9dfa7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]Graph[[:space:]]Pre-training[[:space:]]for[[:space:]]Aspect-based[[:space:]]Sentiment[[:space:]]Analysis/9039fba0-55c9-4e2a-8562-04b973233fbe_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]In-Context[[:space:]]Learning[[:space:]]for[[:space:]]Knowledge[[:space:]]Grounded[[:space:]]Dialog[[:space:]]Generation/9a741ca0-d517-414e-a790-3e4240b3f8c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Multi-Modal[[:space:]]Out-of-Distribution[[:space:]]Detection/370fe86a-89bc-481c-9b45-b7d6a721173f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]Linguistic[[:space:]]Properties[[:space:]]of[[:space:]]Monolingual[[:space:]]BERTs[[:space:]]with[[:space:]]Typological[[:space:]]Classification[[:space:]]among[[:space:]]Languages/a31ff030-1182-47b8-b7c6-a074a3b82434_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Cognitive[[:space:]]Knowledge[[:space:]]Structure[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models_[[:space:]]An[[:space:]]Educational[[:space:]]Diagnostic[[:space:]]Assessment[[:space:]]Approach/0da50962-a33f-4e66-9155-6d20ab4a56e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Effectiveness[[:space:]]of[[:space:]]Multi-Lingual[[:space:]]Commonsense[[:space:]]Knowledge-Aware[[:space:]]Open-Domain[[:space:]]Dialogue[[:space:]]Response[[:space:]]Generation/25a4e6b3-4fd2-49f7-995a-4282bd56d09e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Impact[[:space:]]of[[:space:]]Corpus[[:space:]]Diversity[[:space:]]on[[:space:]]Financial[[:space:]]Pretrained[[:space:]]Language[[:space:]]Models/1e2f7c88-6efb-49cc-8d76-2bb8d6dcd423_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Numerical[[:space:]]Reasoning[[:space:]]Capabilities[[:space:]]of[[:space:]]Language[[:space:]]Models_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Analysis[[:space:]]on[[:space:]]Tabular[[:space:]]Data/e99267a4-bd81-4d11-b21c-b5fa84defd3d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]in[[:space:]]Generating[[:space:]]Code-Tracing[[:space:]]Questions[[:space:]]for[[:space:]]Introductory[[:space:]]Programming[[:space:]]Courses/b9f84753-3729-4b71-93d3-f729145be984_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Exploring[[:space:]]the[[:space:]]Sensitivity[[:space:]]of[[:space:]]LLMs’[[:space:]]Decision-Making[[:space:]]Capabilities_[[:space:]]Insights[[:space:]]from[[:space:]]Prompt[[:space:]]Variations[[:space:]]and[[:space:]]Hyperparameters/1598ac8a-8b03-4771-9f24-0b52c5f3a22d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Extractive[[:space:]]Summarization[[:space:]]via[[:space:]]ChatGPT[[:space:]]for[[:space:]]Faithful[[:space:]]Summary[[:space:]]Generation/662ef5d1-84fe-4bce-b1ac-1bad079a094e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Extrapolating[[:space:]]Multilingual[[:space:]]Understanding[[:space:]]Models[[:space:]]as[[:space:]]Multilingual[[:space:]]Generators/4bb07834-e1b1-484f-8bfb-13359a1ca71d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Eyes[[:space:]]Show[[:space:]]the[[:space:]]Way_[[:space:]]Modelling[[:space:]]Gaze[[:space:]]Behaviour[[:space:]]for[[:space:]]Hallucination[[:space:]]Detection/b8253aff-e44c-453c-b907-0e33981569fa_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/FFAEval_[[:space:]]Evaluating[[:space:]]Dialogue[[:space:]]System[[:space:]]via[[:space:]]Free-For-All[[:space:]]Ranking/99952f8c-1043-48f7-9a3f-b36b121ed08f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/FREDSum_[[:space:]]A[[:space:]]Dialogue[[:space:]]Summarization[[:space:]]Corpus[[:space:]]for[[:space:]]French[[:space:]]Political[[:space:]]Debates/bbdc2244-8d38-4b21-a1bb-467cee78bada_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_content_list.json b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e54eb14b60487ab56c8a3db8161fc7c4bc45dee7 --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_content_list.json @@ -0,0 +1,2790 @@ +[ + { + "type": "text", + "text": "Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation", + "text_level": 1, + "bbox": [ + 272, + 84, + 724, + 123 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gunsoo Han $^{1}$ Daejin Jo $^{1}$ Daniel Wontae Nam $^{1}$ Eunseop Yoon $^{2}$ Taehwan Kwon $^{3*}$ Seungeun Rho $^{4*}$ Kyoung-Woon On $^{1}$ Chang D. Yoo $^{2}$ Sungwoong Kim $^{5\\dagger}$", + "bbox": [ + 132, + 134, + 880, + 168 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Kakao Brain $^{2}$ KAIST $^{3}$ KRAFTON $^{4}$ Georgia Institute of Technology $^{5}$ Korea University", + "bbox": [ + 114, + 171, + 882, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gunsoo.han@kakaobrain.com swkim01@korea.ac.kr", + "bbox": [ + 253, + 200, + 741, + 216 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge-grounded dialogue generation requires to first retrieve appropriate external knowledge based on a conversational context and then generate a response grounded on the retrieved knowledge. In general, these two sequential modules, a knowledge retriever and a response generator, have been separately trained by supervised data for each module. However, obtaining intermediate labels of the ground-truth knowledge is expensive and difficult especially in open-domain conversation. Latent variable modeling can circumvent it and enables a joint training without the knowledge supervision. In this paper, we propose an efficient algorithm for this latent variable modeling that is able to leverage a large amount of dialogue data. In specific, rather than directly training the complex retriever, we adapt a query generator with an off-the-shelf retriever, and the query generator and response generator are simultaneously trained over the latent variable of query. Moreover, we employ the evidence lower bound as a training objective and modify it to efficiently and robustly perform the joint training. Experimental results on diverse knowledge-grounded dialogue datasets show that the proposed algorithm achieves state-of-the-art performances even without the use of the annotated knowledge while maintaining the efficiency and scalability.", + "bbox": [ + 142, + 281, + 460, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 714, + 258, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, knowledge-grounded dialogue generation has drawn increasing attention especially as a key ingredient for open-domain conversational agents (Dinan et al., 2018; Zhou et al., 2018b; Zhan et al., 2021; Liu et al., 2018; Zhou et al., 2018a; Lian et al., 2019; Zhao et al., 2020; Kim et al., 2020; Chen et al., 2022; kom; Shuster et al., 2022a,b; Cai et al., 2023; Wu et al., 2022; Lewis et al., 2020; Huang et al., 2021; Thulke et al., 2021; Anantha", + "bbox": [ + 112, + 740, + 489, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "et al., 2021). It usually obtains knowledge from external resources such as Wikipedia-based databases and web search engines since internal knowledge even in large-scale parametric language models (Brown et al., 2020; Thoppilan et al., 2022; Chowdhery et al., 2022) is incomplete or outdated, and moreover it can provide hallucinated information.", + "bbox": [ + 507, + 252, + 884, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order for a dialogue response to be grounded on such external knowledge, the conversational agent generally consists of a knowledge retriever, which retrieves knowledge corresponding to a given dialogue context, followed by a response generator that produces an informative response based on the dialogue context and the retrieved knowledge. In many previous methods, supervised learning has often been applied to independently optimize each module using the ground-truth or gold knowledge (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). However, human annotation of knowledge information is cumbersome, expensive, and often incomplete due to the existence of multiple possible knowledge candidates and the overlooking of the target response. In addition, existing automatic annotations are generally limited to a simple extractive question answering (Zhao et al., 2020; Liu et al., 2021). This difficulty in obtaining annotations of knowledge information could be severe under the open-domain conversation and hinders the use of large-scale dialogue data.", + "bbox": [ + 507, + 365, + 884, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, there have been a number of recent approaches to learn the knowledge retriever and the response generator without the knowledge supervision. In specific, they have treated the retrieved knowledge, document, or passage as an unobserved latent variable and adapt latent variable modeling based on approximated marginalization (e.g. top-k) (Lewis et al., 2020; Huang et al., 2021; Cai et al., 2023; Guu et al., 2020), reinforcement learning (Zhao et al., 2020; Zhang et al., 2022; Chen et al., 2022; Wu et al., 2022) or variational methods (Zhan", + "bbox": [ + 507, + 745, + 885, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done at Kakao Brain", + "bbox": [ + 139, + 891, + 319, + 904 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 139, + 904, + 294, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2683", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2683-2702 \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 216, + 945, + 779, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020; Chen et al., 2020; Xu et al., 2023). However, joint training of the retriever along with the generator under this latent variable modeling has some restrictions in utilizing the retriever. For example, a retriever needs to produce a differentiable prior probability for the gradient propagation through the marginalization or to be intermittently updated to rebuild the whole passage index during training (Karpukhin et al., 2020; Lewis et al., 2020; Zhang et al., 2022). This leads to limitation in facilitating complicated or large-scale retrievers and further in leveraging large-scale data. In addition, the posterior probability used in the previous variational methods has been modeled by a separated network, which grows the training complexity and also incurs the discrepancy between the training-time and test-time knowledge samplings.", + "bbox": [ + 112, + 84, + 489, + 368 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome such restrictions, in this paper, we propose an efficient latent variable modeling, named $ELVM$ , for knowledge-grounded dialogue generation. In particular, we reduce the burden for training a whole retriever by employing a query generator followed by an off-the-shelf retriever that is fixed during training. More precisely, in tandem with the response generator, the query generator rather than the retriever is jointly optimized by latent variable modeling, in which a query is taken as a latent variable, on paired observations of dialogue contexts and responses.", + "bbox": [ + 112, + 373, + 489, + 562 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we exploit the variational method using the evidence lower bound (ELBO) (Kingma and Welling, 2013; Jordan et al., 1999) and approximate the expected conditional likelihood in the ELBO by subset sampling from the prior distribution, which acts as the training objective for the response and the query generators. This approximation gets rid of an extra modeling of a surrogate posterior distribution or online posterior inference such as Markov Chain Monte Carlo (MCMC) and also reduces the training-inference discrepancy in knowledge retrieval. Moreover, we further modify the Kullback-Leibler (KL) regularization of the ELBO by constructing the approximated posterior distribution from the conditional likelihood and prior distribution and set this posterior distribution as a teacher for a distillation objective to further learn the query generator.", + "bbox": [ + 112, + 567, + 489, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results show that the proposed ELVM allows to efficiently and robustly perform training without (1) the use of the annotated knowledge, (2) an explicit training of the knowledge", + "bbox": [ + 112, + 854, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "retrieval, and (3) a complex posterior sampling. Especially, it significantly outperforms previous state-of-the-art methods for knowledge-grounded dialogue generation. In addition, the proposed posterior distillation improves the performances over the baseline that solely maximizes the expectation of the conditional likelihood.", + "bbox": [ + 507, + 84, + 884, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as:", + "bbox": [ + 526, + 196, + 875, + 211 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An efficient latent variable modeling is proposed in joint training of query generator and dialogue response generator without the knowledge supervision for knowledge-intensive dialogue generation.", + "- For realizing efficient yet robust joint training, the ELBO for the marginal likelihood is modified as the combination of the conditional likelihood objective and the posterior distillation objective, based on multiple prior samples.", + "- The proposed ELVM demonstrates its effectiveness in performing unsupervised joint training and even significant performance improvements over previous methods, being a new state-of-the-art on benchmark datasets." + ], + "bbox": [ + 509, + 215, + 885, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 509, + 469, + 665, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Knowledge-Grounded Dialogue Generation", + "text_level": 1, + "bbox": [ + 507, + 500, + 805, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge-grounded conversation has been extensively studied through recently released public datasets (Dinan et al., 2018; kom; Xu et al., 2022b; Kwiatkowski et al., 2019; Anantha et al., 2021; Moghe et al., 2018). Existing approaches for this task have mostly exploited two successive modules, the knowledge retriever and the knowledge-grounded response generator. Many previous works have used manual annotations of gold knowledge provided by some of public datasets to optimize the modules (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). Specifically, Seeker (Shuster et al., 2022a) and BlenderBot3 (Shuster et al., 2022b) have recently proposed to build a series of modules by a single transformer with different prompts for each module, and trained its transformer on a large number of modular tasks using annotated datasets. However, this manual annotation is expensive, time-consuming, and often inaccurate, which impedes the utilization of a large-scale dialogue data that is especially necessary for open-domain conversation.", + "bbox": [ + 507, + 539, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2684", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Unsupervised Joint Training", + "text_level": 1, + "bbox": [ + 112, + 84, + 386, + 99 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, unsupervised training methods have been widely applied, and many of them have tries to jointly learn the modules without knowledge labels (Lewis et al., 2020; Huang et al., 2021; Zhao et al., 2020; Zhang et al., 2022; Zhan et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020). For example, PLATO-KAG (Huang et al., 2021) has approximated the marginal likelihood by top-k knowledge samples while RetGen (Zhang et al., 2022) has trained to reward knowledge retrieval with the highest utilization in response generation by reinforcement learning and mixture-of-experts ensembling.", + "bbox": [ + 112, + 99, + 489, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, latent variable modeling based on variational methods has also been developed by several recent works. CoLV (Zhan et al., 2021) has introduced collaborative latent spaces to reflect the inherent correlation between the knowledge selection and the response generation. SKT (Kim et al., 2020) has developed a sequential latent variable model for the multi-turn knowledge-grounded dialogue generation. In order to perform the posterior sampling of knowledge selection during joint training, some works have proposed to separately train the posterior distribution model (Paranjape et al., 2022; Lian et al., 2019) or the posterior information prediction model (Chen et al., 2020). Very recently, SPI (Xu et al., 2023) has applied short-run MCMC (Erik et al., 2019) for posterior sampling on the collaborative latent spaces.", + "bbox": [ + 115, + 305, + 489, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While these latent variable modeling algorithms can effectively perform unsupervised joint training, the entire training of the retriever is still difficult and imposes restrictions on the selection of the retriever in terms of both a search algorithm and a knowledge resource. Furthermore, the additional posterior sampling through separate networks or multiple iterations in the previous variational methods also leads to increased training complexity as well as the training-inference discrepancy in knowledge generation. In contrast, the proposed ELVM can employ any kind of retriever in latent variable modeling since it controls a retrieval output by only changing of a generated query. Moreover, ELVM removes an extra posterior modeling or sampling by prior subset sampling and approximated posterior distillation, which leads to efficient training without the discrepancy.", + "bbox": [ + 112, + 573, + 489, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Query Generation", + "text_level": 1, + "bbox": [ + 112, + 866, + 305, + 882 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When an off-the-shelf retriever is used for knowledge retrieval, how to make an input text query", + "bbox": [ + 112, + 887, + 489, + 919 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is important for obtaining appropriate documents. Especially, in knowledge-grounded dialogue, a self-contained query should be generated from the multi-turn dialogue context. A number of prior works have tried to train a supervised query rewriting model by human rewrites (Yu et al., 2020; Lin et al., 2020; Vakulenko et al., 2021; Voskarides et al., 2020). Similar to the knowledge annotation, the limitations of manual query annotation have come up with unsupervised learning of the query generator. A number of works have proposed a novel rewards for reinforcement learning of a query rewritten (Wu et al., 2022; Chen et al., 2022).", + "bbox": [ + 507, + 84, + 884, + 288 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "QKConv (Cai et al., 2023) has proposed an unsupervised query enhanced method for knowledge-grounded conversation. This work is similar to ours in that it consists of a query generator, an off-the-shelf knowledge retriever, and a response generator, and unsupervised joint training is applied to the query generator and the response generator. However, their training objective is based on the approximated marginal likelihood over candidate queries, which is different from our objective that includes the posterior distillation based on the ELBO loss.", + "bbox": [ + 507, + 290, + 885, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Efficient Latent Variable Model for Knowledge-Grounded Dialogue Generation", + "text_level": 1, + "bbox": [ + 507, + 494, + 848, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Setup", + "text_level": 1, + "bbox": [ + 509, + 546, + 601, + 562 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we explain an overall structure of our ELVM, as described in Figure 1a, and define notations. Unlike previous works of document (passage) retrieval (Zhang et al., 2022; Huang et al., 2021; Lewis et al., 2020) using a query of an embedded vector of a dialogue context $x$ , we use a text query of natural language, which enables to utilize any kind of retriever for dialogue response generation. We define the probability of generating a natural language query $u$ for a given dialogue context $x$ , $p_{\\phi}(u|x)$ , as below:", + "bbox": [ + 507, + 565, + 884, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\phi} (u | x) = \\prod_ {t} p _ {\\phi} \\left(u _ {t} \\mid u _ {0: t - 1}, x\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 574, + 753, + 882, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u_{t}$ is the $t$ -th token in $u$ and $p_{\\phi}$ is the probability obtained from the model parameterized by $\\phi$ . Similarly, the probability of generating a response $y$ for a given query $u$ and dialogue context $x$ is defined as", + "bbox": [ + 507, + 799, + 884, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (y | Z (u), x) = \\prod_ {t} p _ {\\theta} \\left(y _ {t} \\mid y _ {0: t - 1}, Z (u), x\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 889, + 882, + 921 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2685", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cfe6d0673e496a8a96875adf2b2e3dfa927b4c687d9710d9bafbea2183e7fe0.jpg", + "image_caption": [ + "(a) Overall structure of ELVM." + ], + "image_footnote": [], + "bbox": [ + 127, + 84, + 670, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c140f2c8b4e80b181f7d1a434c3dbfa3a15da0576b0dd08f773430c18f14263.jpg", + "image_caption": [ + "(b) Example inference of ELVM.", + "Figure 1: Overall structure and example inference case of ELVM. During training, for a given dialogue context $x$ , the query generator generates a set of unique queries $Q^s$ and for each query $u$ a set of top- $k$ documents $Z(u)$ is retrieved using the off-the-shelf retriever. Then, response generator utilizes both $x$ and $Z(u)$ : (a) The approximated marginal likelihood $\\tilde{p}_{\\theta, \\phi}(y|x)$ used during training is defined by $Q^s$ and the re-normalized prior $\\bar{p}_{\\phi}(u|x)$ ; (b) At inference time, a single sampled query $u$ is used for retrieving multiple documents and producing a final response." + ], + "image_footnote": [], + "bbox": [ + 697, + 82, + 875, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y_{t}$ is the $t$ -th token in $y$ , $Z(u) = \\{z_1, z_2, \\dots, z_k\\}$ is the set of retrieved documents (passages) by $u$ , and $\\theta$ is the model parameters to produce the response probability. Here, we use top- $k$ result obtained by an off-the-shelf retriever such as BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) and freeze the retriever. We assume that there is a deterministic mapping from $u$ to $Z$ . We now set a query $u$ as a latent variable. Then, using the prior $p_{\\phi}(u|x)$ associated with the query generator and the conditional likelihood $p_{\\theta}(y|Z(u), x)$ corresponding to the response generator, we can define the marginal likelihood of the knowledge-grounded response $y$ given the dialogue context $x$ as below:", + "bbox": [ + 112, + 445, + 487, + 682 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q} p _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 139, + 692, + 487, + 725 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ is the set of all possible queries. Since it is intractable to marginalize over all queries, we sample a small subset composed of $m$ unique queries, $Q^{s}(x) = \\{u_{1}, u_{2}, \\dots, u_{m}\\}$ using $p_{\\phi}(\\cdot | x)$ and then marginalize over this set as below:", + "bbox": [ + 112, + 734, + 489, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q ^ {s}} \\bar {p} _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 136, + 822, + 487, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 112, + 865, + 164, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {p} _ {\\phi} (u | x) = \\frac {p _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 885, + 487, + 922 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We construct $Q^s$ by obtaining $m$ -unique queries sampled from $p_{\\phi}(\\cdot | x)$ . It is noted that our response generator takes multiple retrieved documents efficiently using FiD (Izacard and Grave, 2020), which is different from the use of a single document for response generation in previous methods. In addition, while we can utilize relative matching scores from the retriever to assign varying likelihoods to each document given a query for $p(Z(u)|u)$ , we have noticed that this approach can occasionally hinder the effective and comprehensive learning of the query generator. During the training process, that will be described in the next subsection, our objective is to amplify the impact of the generated query. Consequently, we opt for a deterministic mapping from a query to a set of retrieved documents as a unified whole, aiming to enhance the gradient flow.", + "bbox": [ + 507, + 445, + 884, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Joint Training", + "text_level": 1, + "bbox": [ + 507, + 741, + 670, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our training objective is the ELBO (Kingma and Welling, 2013; Jordan et al., 1999) of $\\log \\tilde{p}_{\\theta,\\phi}$ which can be written as below:", + "bbox": [ + 507, + 760, + 882, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log \\tilde {p} _ {\\theta , \\phi} (y | x) \\geq \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ - D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 820, + 880, + 860 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $q(u|x,y)$ is the variational posterior. Note that the equality holds when $q(u|x,y)$ is identical to the true posterior distribution of it. In this work,", + "bbox": [ + 507, + 871, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2686", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we do not parameterize $q(u|x,y)$ and approximate the expected conditional likelihood in the ELBO by sampling $u$ from not $q(u|x,y)$ but $\\bar{p}_{\\phi}(u|x)$ such as", + "bbox": [ + 112, + 84, + 487, + 145 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ \\approx \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ]. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 160, + 485, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approximation is due to that it minimizes the discrepancy between training and test time inference and it does not separately model $q(u|x,y)$ or complicatedly sample from it during training. Then, we obtain the following two losses corresponding to the expected conditional likelihood and the KL regularization such as", + "bbox": [ + 112, + 210, + 487, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {y, u} (\\theta , \\phi) = - \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 334, + 487, + 353 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {u} (\\phi) = D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 152, + 355, + 487, + 374 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, we define $q(u|x,y)$ in $\\mathcal{L}^u (\\phi)$ by the approximated posterior derived from the prior and the likelihood such as", + "bbox": [ + 112, + 386, + 485, + 432 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq (u | x, y) = \\frac {p _ {\\theta} (y | Z (u) , x) \\bar {p} _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\theta} (y | Z (u ^ {\\prime}) , x) \\bar {p} _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 443, + 485, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set this posterior distribution $q(u|x,y)$ as a teacher for a distillation to update the prior associated with the query generation at test time. Namely, $\\mathcal{L}^u (\\phi)$ becomes the posterior distillation loss, and our prior is explicitly updated to be similar to the posterior. Here, the gradient is not propagated from $q(u|x,y)$ , and $q(u|x,y)$ can be computed easily without additional model or approximation since we define it over the subset $Q^s$ .", + "bbox": [ + 112, + 492, + 487, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To sum up, we obtain the conditional likelihood loss and the posterior distillation loss by multiple samples from the prior distribution that is aligned with query sampling at the inference time. We would mitigate the bias of prior sampling by multiple query samples for each input during training. In addition, our modification of ELBO is different from the expectation-maximization (EM) algorithm in that we sample a latent variable from the prior distribution rather than the posterior distribution that is used for sampling of E-step in EM. Moreover, in contrast to the M-step loss, we use the prior-weighted conditional likelihood and the gradient is also propagated to the sampled latents and the corresponding prior model. For experiments in this paper, the query generator $\\phi$ and the response generator $\\theta$ share parameters, with different input prompts to distinguish each module, for simplicity.", + "bbox": [ + 112, + 634, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the loss for our joint training is", + "bbox": [ + 527, + 84, + 823, + 99 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta , \\phi) = \\mathcal {L} ^ {y, u} (\\theta , \\phi) + \\beta \\mathcal {L} ^ {u} (\\phi), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 112, + 882, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is the relative weight for the posterior distillation loss.", + "bbox": [ + 507, + 143, + 880, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experiment", + "text_level": 1, + "bbox": [ + 507, + 187, + 648, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 214, + 712, + 230 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Our experiments are conducted on two widely used knowledge-grounded dialogue datasets, Wizard of Wikipedia (WoW) (Dinan et al., 2018) and QReCC (Anantha et al., 2021). WoW is structured around dialogues that are grounded in knowledge sentences sourced from Wikipedia, consisting of roughly 18K, 2K, and 2K dialogues, for training, validation, and test subsets, respectively. QReCC is a comprehensive compilation of conversational data, consisting of 14K open-domain conversations comprising 80K question-answer pairs.", + "bbox": [ + 507, + 235, + 882, + 409 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87698417f33a1cffae8002e1577bc3c05079eaa86f1a3dc976d87cf553d64661.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDatasetTrainValid
QueryWizInt (Komeili et al., 2022)351,3752,467
Fits (xu2)3,587392
KG ResponseWizInt (Komeili et al., 2022)22,4881,687
Fits (xu2)6,279656
Ms Marco (Nguyen et al., 2017)281,63636,859
NQ Open (Adolphs et al., 2021b)79,1688.757
Dialogue ResponseWizInt (Komeili et al., 2022)8,335587
MSC (Xu et al., 2022a)105,54917,691
SaferDialogues (Ung et al., 2022)6,3067,88
PersonaChat (Zhang et al., 2018)131,4387.801
EmpatheicDialogues (Rashkin et al., 2019)64,6365,738
", + "bbox": [ + 512, + 438, + 878, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: List of dataset and number of train and validation examples used for fine-tuning R2C2 model to obtain R2C2-PT. KG is an abbreviation for knowledge-grounded.", + "bbox": [ + 507, + 577, + 882, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We begin with publicly available R2C2 pre-trained transformer model (Shuster et al., 2022a) with 400M parameter size and fine-tune it on multiple datasets to acquire the essential skill sets for knowledge-grounded dialogue such as query generation and knowledge-grounded response generation. Detailed tasks and statistics are shown in Table 1. In line with Shuster et al. (2022b), we adopt a modular system with a single model that incorporates control tokens into the dialogue context to perform different modules. For example, appending the control token __generate-query__ to the dialogue context promotes the model to generate relevant search query. We name the initial model from this pretraining procedure as R2C2-PT. Subsequently, we apply our proposed algorithm to R2C2-PT model,", + "bbox": [ + 507, + 650, + 882, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2687", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6b4309a7b555fa2602c1f82b893988590d36ebed56160cae604ea961c3ccfa7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
PostKS (Lian et al., 2019)79.1--13.01.0193.8--13.11.0
ITDD (Li et al., 2019)17.84.02.516.2-44.82.11.111.4-
SKT (Kim et al., 2020)52.0--19.36.881.4--16.14.2
PIPM (Chen et al., 2020)42.7-3.319.97.365.7-2.517.65.4
ZRKGK (Li et al., 2021)40.4----41.5----
CoLV (Zhan et al., 2021)39.6-2.920.67.954.3-2.119.76.3
KnowledGPT (Zhao et al., 2020)19.29.57.222.07.922.38.36.020.56.7
SPI (Xu et al., 2023)17.110.27.722.78.819.19.67.322.08.5
R2C240.71.70.912.31.946.61.60.911.95.4
R2C2-PT24.76.04.417.35.230.46.24.617.75.4
ELVM-OK19.812.09.525.310.725.812.19.725.210.5
ELVM-EM-Like16.211.69.125.710.719.412.09.525.710.8
ELVM-from-R2C215.814.511.828.813.318.914.311.528.813.3
ELVM14.614.711.929.313.918.014.511.829.213.7
", + "bbox": [ + 132, + 80, + 867, + 272 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/809f2367c8bf6a80ab0f94ed3f6dd83f2121e8d4396294a55838875cf7c6e1d2.jpg", + "table_caption": [ + "Table 2: Performance comparison in WoW. PPL represents perplexity, B3 and B4 denote BLEU-3 and BLEU-4 scores, respectively, and R1 and R2 indicate Rouge-1 and Rouge-2 scores. Our proposed model, ELVM, achieves new SOTA performance for both seen and unseen tasks with a substantial margin." + ], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
R@1R@5R@10R@1R@5R@10
R2C2-PT25.763.874.329.264.876.7
ELVM-EM-Like26.463.974.728.964.575.8
ELVM-β=026.164.675.728.965.376.8
ELVM29.168.480.330.468.380.3
", + "bbox": [ + 117, + 344, + 485, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "resulting in ELVM. Furthermore, to dispel concerns that our training gains might predominantly stem from R2C2-PT's comprehensive fine-tuning across multiple datasets rather than ELVM's inherent efficacy, we opted to utilize the base R2C2 model. Upon integrating it with ELVM, the resulting configuration is denoted as ELVM-from-R2C2.", + "bbox": [ + 112, + 533, + 489, + 644 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training of ELVM, we employ a sampling-based generation method, specifically nucleus sampling (Holtzman et al., 2020), to promote diverse query generation in $p_{\\phi}$ . When retrieving documents based on the generated queries, we utilize an off-the-shelf sparse retriever such as BM25 (Robertson et al., 2009), prioritizing low latency over the dense retriever (DR). We train for 3 epochs with $m = 4$ and $k = 5$ for WoW and 1 epoch with $m = 8$ and $k = 10$ for QReCC. $\\beta$ is set to 1 for both tasks, unless stated otherwise. To ensure a thorough evaluation of ELVM's performance and its ability to generalize to unseen dialogues, we purposely avoid dataset overlap between training of R2C2-PT and ELVM. More detailed information on the training process can be found in Appendix A.", + "bbox": [ + 112, + 650, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/57baa9223cfd301befa41f54992799d1e3786e6a7aef496d4d902602cae4ab5f.jpg", + "table_caption": [ + "Table 3: Comparison of document recall in WoW between ELVM variants. Note that we omit the performance of ELVM-OK whose performance is identical to R2C2-PT since it is trained with keeping its query generator frozen." + ], + "table_footnote": [], + "table_body": "
ModelF1 ↑EM ↑R1 ↑R2 ↑PPL ↓
Q.Rewriting (Raposo et al., 2022)18.91.0---
DPR-IHN (Kim and Kim, 2022)30.44.7---
QKConv (Cai et al., 2023)33.55.9---
R2C2-PT22.81.325.711.59.4
ELVM-OK33.04.535.621.95.6
ELVM36.56.239.025.84.8
", + "bbox": [ + 512, + 344, + 878, + 425 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75e19563e84b5f032a0f9c724bc724f9d27d6e1ae8acc4600df9f00740892194.jpg", + "table_caption": [ + "Table 4: Performance comparison in QReCC with QKConv (Cai et al., 2023) which is the previous SOTA on this task. We report F1 scores, exact match (EM), Rouge score (R) and perplexity (PPL)." + ], + "table_footnote": [], + "table_body": "
ModelFluency (%)Relevance (%)
SeenUnseenSeenUnseen
KnowledGPT69.862.941.541.1
ELVM73.577.5*49.563.6**
", + "bbox": [ + 526, + 505, + 865, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5: Human evaluation results on WoW test. A pairwise t-test is conducted to verify statistical significance of the improvements, and the corresponding results in bold are significantly better than those from the baseline model $(^{**}: p < 0.01, ^*: p < 0.025)$ .", + "bbox": [ + 507, + 571, + 882, + 644 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During inference, unlike the training phase, only a single query is generated by the query generator $p_{\\phi}$ . The default number of retrieved documents $k$ from the off-the-shelf retriever is set to 5 for WoW and 10 for QReCC. The knowledge-grounded response is generated by the response generator, $p_{\\theta}$ . An illustration of the inference is depicted in Figure 1b. During the inference, both query and response generation are conducted using beam search with a beam size of 5.", + "bbox": [ + 507, + 670, + 882, + 827 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Variants of ELVM. To examine the impact of the proposed ELVM training, we explore another variant, ELVM-OK, which is trained on the response generation task with annotated oracle knowledge documents while keeping the query generator $p_{\\phi}$", + "bbox": [ + 507, + 840, + 884, + 920 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2688", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/adbeeedf36f7c9f7917c84bd396388749de2e908a905c19f6fa193281f866f78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mSeenUnseen
B4↑R2↑B4↑R2↑
110.812.410.512.2
211.113.011.112.8
4*11.913.911.813.7
811.813.711.713.6
", + "bbox": [ + 181, + 80, + 420, + 165 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c41d2ee83258d09039bbef510dab3e91e85e10c6e4ac1a3c5b11057a33cf3f12.jpg", + "table_caption": [ + "Table 6: Effect of varying the number of queries, $m$ , sampled during training on WoW. We see the highest performance on unseen task when $m = 4$ . The superscripted value by * is the default setting for ELVM." + ], + "table_footnote": [], + "table_body": "
βWow UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
09.310.665.334.55.155.8
1*11.813.768.336.56.965.6
510.812.466.731.33.531.9
", + "bbox": [ + 117, + 244, + 485, + 317 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frozen. During inference, given a dialogue context the frozen query generator $p_{\\phi}$ generates query and use the off-the-shelf retriever to bring relevant documents. Finally the response generator of ELVM-OK generates the final response.", + "bbox": [ + 112, + 424, + 487, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Section 3.2, we describe the difference between ELVM training and EM algorithm. In order to quantitatively compare these two methods, we perform EM-like training in ELVM where we apply the posterior-weighted conditional likelihood in $\\mathcal{L}^{y,u}(\\theta ,\\phi)$ .. $\\mathcal{L}^{y,u}(\\theta ,\\phi) =$ $-\\sum_{u}SG(q(u|x,y))\\log p_{\\theta}(y|Z(u),x).$ Here, the query latents $u$ are still sampled from the prior distribution, and $SG$ means the stop-gradient that prevents the gradient propagation to the query generator. We name this variant as ELVM-EM-Like.", + "bbox": [ + 112, + 504, + 489, + 677 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Automatic Evaluation", + "text_level": 1, + "bbox": [ + 112, + 690, + 332, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The evaluation results of WoW test seen and unseen tasks are shown in Table 2. ELVM surpasses all its variants and previous state-of-the-art (SOTA) model, achieving a new SOTA on both WoW test seen unseen tasks, across all metrics. In addition, ELVM-OK demonstrates notable improvements in BLEU and Rouge scores compared to previous models with a slight drawback in PPL. Similarly, ELVM-EM-Like also exhibits competitive performance across multiple metrics. Moreover, when comparing ELVM-from-R2C2 and ELVM, the experiment demonstrates that while R2C2-PT offers marginal benefits, the marked performance", + "bbox": [ + 112, + 713, + 489, + 919 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b56a112fc882e9e7d7ba4399359a38ffe4339de8db84dcf0d14a73447d821b8e.jpg", + "table_caption": [ + "Table 7: Ablation study of the posterior distillation weight $\\beta$ on WoW. From the results, we can see that the trivial value of $\\beta = 1$ shows the best performance. The super-scripted value by $*$ is the default setting for ELVM." + ], + "table_footnote": [], + "table_body": "
ModelWoW UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
ELVM11.813.768.336.56.259.9
ELVMDR12.815.581.237.96.965.6
", + "bbox": [ + 517, + 80, + 873, + 131 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/913b147fb9c8a7c91b40b06555ff4f908470f0f10169fe95854159b93d6ea559.jpg", + "table_caption": [ + "Table 8: Performance results achieved by employing the dense retriever instead of the default BM25 retriever for WoW unseen and QReCC evaluation." + ], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑B4 ↑R2 ↑R@5 ↑
R2C2-PT4.45.263.84.65.464.8
ELVM-OK9.510.7-9.710.5-
ELVM11.913.968.411.813.768.3
R2C2-PTLarge8.710.262.98.910.364.4
ELVM-OKLarge9.610.6-10.211.2-
ELVMLarge12.614.372.612.414.174.3
", + "bbox": [ + 514, + 196, + 878, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9: Effect of model scaling tested on WoW. ELVM scales out to larger parameter size with increase in performance. The subscript Large indicates the model size of 2.7B parameters.", + "bbox": [ + 507, + 298, + 882, + 355 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "increase is predominantly attributable to ELVM.", + "bbox": [ + 507, + 380, + 868, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further examine the performance of query generator by assessing the recall of the ground-truth documents. The results summarized in Table 3 reveal that training with our proposed algorithm leads to an increase in recall for both ELVM- $\\beta = 0$ and ELVM compared to R2C2-PT while the latter yields a larger performance gain. This highlights the effectiveness of distilling the posterior distribution of the response generator into the prior distribution associated with the query generator.", + "bbox": [ + 505, + 397, + 882, + 552 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For QReCC, as shown in Table 4, ELVM surpasses the previous SOTA, QKConv (Cai et al., 2023), the recent previous works (Raposo et al., 2022; Kim and Kim, 2022), and the variants of ELVM including R2C2-PT and ELVM-OK in terms of all metrics. The recall performance on QReCC is included in Table C.", + "bbox": [ + 507, + 554, + 882, + 663 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Human Evaluation", + "text_level": 1, + "bbox": [ + 507, + 674, + 705, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gauge performance across multiple aspects of the quality of generated responses, human evaluation is performed on the generated responses of ELVM and KnoawedGPT (Zhao et al., 2020) $^{1}$ . Similar to (Rashkin et al., 2021) and (Xu et al., 2023), we assess the response quality in two aspects: Fluency and Relevance. Fluency measures whether the response is understandable, self-consistent without repetition, and proficient. Relevance assesses the extent to which the response aligns with the dialogue context, incorporates pertinent knowledge, and maintains appropriateness. In total, 50 data", + "bbox": [ + 505, + 695, + 882, + 884 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1We choose KnowledgeGPT because it is the best performing model among publicly available models.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2689", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c7673e4e670b5d874a006d650c1eb5ff85030cf4abfe011801d9ffea4b3dba07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
CoLV39.62.97.954.32.16.3
KnowledGPT19.27.27.922.36.06.7
SPI17.17.78.819.17.38.5
BART9741.92.69731.32.6
ELVM-from-BART15.38.411.116.48.210.9
ELVM14.611.913.918.011.813.7
", + "bbox": [ + 122, + 80, + 480, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f8b30158651c9cdd761a9fc64d06a6b5a2e9a72e85665efe6001e1031fefb539.jpg", + "table_caption": [ + "Table 10: Performance comparison with BART as our initial model." + ], + "table_footnote": [], + "table_body": "
ModelTrain RetTest RetWoW Unseen
PPL ↓B4 ↑R2 ↑F1 ↑
CoLV--54.32.16.318.5
KnowledgeGPT--22.36.06.720.5
SPI--19.17.38.5-
ELVMBM25BM2518.011.813.727.0
ELVMBM25Dense14.812.615.229.3
ELVMDenseDense12.912.815.529.0
ELVMDenseBM2521.611.813.128.2
", + "bbox": [ + 117, + 221, + 485, + 322 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11: Impact of alternating the retriever between training and testing phases on WoW unseen test. Dense indicates utilizing all-MiniLM-L6-v2 retriever as described in Section 4 and Ret is an abbreviation for retriever.", + "bbox": [ + 112, + 331, + 489, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "samples are randomly selected from WoW tasks where 25 samples are randomly selected from each seen and unseen task. The qualities of the responses are measured by A/B testing on the two aspects. The feedback is collected from 11 human experts. Further details and annotator instructions can be found in Table G. As shown in Table 5, ELVM significantly outperforms KnowledGPT in all aspects especially on the unseen task.", + "bbox": [ + 112, + 428, + 489, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Ablation Studies", + "text_level": 1, + "bbox": [ + 112, + 583, + 292, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Queries. The size of sampled query set $Q^{s}$ , $m$ , plays a huge role in training of ELVM, as it directly influences the bias of the marginal likelihood of the knowledge-grounded response $y$ . To investigate the impact of $m$ , we gradually increase $m$ from 1 to 8 and measure the performance on WoW while keeping other hyperparameters fixed. As shown in Table 6, we observe a positive correlation where increasing $m$ leads to improved performance. Notably, setting $m = 4$ yields the optimal performance for both seen and unseen tasks.", + "bbox": [ + 112, + 606, + 489, + 778 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Posterior Distillation Weight. We examine the influence of the posterior distillation loss on ELVM training by adjusting the value of $\\beta$ . The experimental results presented in Table 7 reveal that the optimal performance is attained when $\\beta = 1$ for both the WoW unseen and QReCC tasks, indicating the effectiveness of incorporating the posterior distillation loss. However, a substantial decline in", + "bbox": [ + 112, + 791, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance is observed when $\\beta$ is increased to 5 or set to 0, particularly in the case of QReCC.", + "bbox": [ + 507, + 84, + 880, + 116 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying the Retriever. We extend our experiments by incorporating the DR in two distinct ways. For WoW, we train and evaluate using all MiniLM-L6-v2 $^2$ DR. In contrast, for QReCC, due to its extensive knowledge pool, we leverage DR solely during evaluation by employing both the BM25 retriever and all-mpnet-base-v2 $^3$ DR in a hybrid manner to retrieve the top- $k$ documents. Concretely, we score each document by summing up the normalized score of each retriever and rerank to obtain top- $k$ documents. Table 8 presents a summary of the performance results of employing DR, demonstrating a successful integration of DR in different settings of ELVM.", + "bbox": [ + 507, + 127, + 884, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling Up Model. We investigate the impact of increasing the model size on performance. Starting with the publicly available 2.7B R2C2 pretrained transformer, we follow exactly the same procedure outlined in Section 3.2 to obtain R2C2-PTLarge, ELVM-OKLarge and ELVMLarge. Detailed hyperparameters for training 2.7B model is given in Appendix A. The results presented in Table 9 demonstrate that as the model size increases, there is an improvement in performance across all metrics, supporting the scalability of ELVM.", + "bbox": [ + 507, + 360, + 884, + 533 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BART as Base. To ensure a fair comparison and further demonstrate the robustness of our algorithm on diverse model scales, we conduct an experiment employing BART (Lewis et al., 2019) base as an initial model and then apply our proposed algorithm on WoW dataset. We choose BART base model since the previous SOTA model (Xu et al., 2023) utilized BART base as their base model. The outcome of this training is termed ELVM-from-BART. As shown in Table 10, the initial BART model shows very poor performances. However, upon training the BART-base with our ELVM algorithm, there is a great improvement in performance across both WoW Seen and Unseen tasks. Importantly, the results surpass the performance benchmarks set by the previous SOTA model (SPI).", + "bbox": [ + 507, + 545, + 882, + 797 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retriever Switching To assess the robustness of our algorithm to the retriever mismatch between training and inference phases, we undertake an", + "bbox": [ + 507, + 809, + 882, + 857 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 \n $^{3}$ https://huggingface.co/sentence-transformers/all-mpnet-base-v2", + "bbox": [ + 507, + 866, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2690", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34ed99973cf77f8408438675e7626ab246405613d575ca400a8bbe740c9be767.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dialogue ContextI want to teach my family about investing by buying them a few shares of stock, but I really don't know much about it. The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses. Oh, so I could tell my kids they own part of a company. That's cool. Maybe I should get them a share in a company they know then, like Apple or XBox or Snapchat? [...] Examples of the latter include shares of private companies which are sold to investors through equity crowdfunding platforms. [...]
GT knowledge
GT ResponseYes, it's like the shares of a company are sold to investors through equity crowdfunding platforms.
KnownGPTQuery: n/aGT knowledge selected: False
Response: yes, you can get shares in a company like apple or xbox. you can also get shares in company like stock market.
R2C2-PTQuery: How to teach kids about investingGT knowledge selected: False
Response: The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses.
ELVM-β = 0Query: Investing tips for beginnersGT knowledge selected: False
Response: That's a good idea. The stock market is a place where you can buy and sell stocks.
ELVM w/ BM25Query: Apple xbox snapshot stock share buy apple stock instagram stock investingGT knowledge selected: True
Response: That is a good idea. You can also buy shares of private companies which are sold to investors through equity crowdfunding platforms.
ELVM w/ DRQuery: Best stocks for teaching kids about investingGT knowledge selected: True
Response: That is a good idea. The latter include shares of private companies which are sold to investors through equity crowdfunding platforms.
", + "bbox": [ + 114, + 80, + 884, + 397 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 12: Generated samples from WoW unseen task. The first row represents the dialogue context, ground truth (GT) knowledge (highlighted in olive), and GT response. Following rows contain the model generated query, response and whether the GT knowledge is included in top-5 retrieved documents. More samples can be found in Appendix H. We truncate the GT knowledge and display only relevant spans of the GT knowledge to the ground-truth response.", + "bbox": [ + 112, + 404, + 882, + 475 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "additional experiment, of which the results are depicted in Table 11. In this experiment, we train the WoW task utilizing either BM25 or dense retriever (utilizing all-MiniLM-L6-v2 retriever, as detailed Section 4), and subsequently evaluate using either BM25 or dense retriever. The results on the WoW unseen test indicate that utilizing a different retriever during testing doesn't necessitate model re-training, particularly given the superior performance of the dense retriever compared to BM25.", + "bbox": [ + 112, + 500, + 489, + 674 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.5 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 112, + 700, + 319, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For qualitative analysis, we examine the generated samples from KnownGPT, R2C2-PT and other variants of ELVM, whose samples are presented in Table 12. The example reveals that the queries generated by KnownGPT, R2C2-PT and ELVM- $\\beta = 0$ fail to retrieve the relevant knowledge effectively while the queries generated by ELVM with BM25 and DR (all-MiniLM-L6-v2) reflects the ability to access the relevant knowledge, leading to improved responses. The results also demonstrate that the query generator trained with our algorithm effectively adapts to the chosen retriever.", + "bbox": [ + 112, + 728, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 499, + 640, + 514 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we propose ELVM as an efficient latent variable modeling for knowledge-intensive dialogue generation. In ELVM, the knowledge retrieval is realized by the query generation followed by document retrieval using the off-the-shelf retriever for ease of joint training of both the retriever and the response generator, without the knowledge supervision. Furthermore, the training scheme of ELVM modifies the ELBO for the marginal likelihood to effectively perform the joint training without the use of complex posterior sampling, which also eliminates the discrepancy between the training-time and test-time knowledge samplings. ELVM empirically demonstrates that it significant outperforms over previous latent variable methods with in-depth analysis on diverse dialogue datasets.", + "bbox": [ + 507, + 527, + 884, + 781 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 796, + 613, + 810 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There are a number of limitations and possible directions to future works of ELVM proposed in this paper. First, as more large language models beyond the scale of hundreds of billion parameters are emerging, there are rooms for emergent behaviors in ELVM as the sizes of data and model grow that", + "bbox": [ + 507, + 824, + 882, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2691", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "has not been tested in this paper. Second, instead of the prior sampling that we use, fast posterior sampling such as short-run MCMC (Erik et al., 2019; Xu et al., 2023) is also an alternative candidate. Third, the experiments in this paper retrieve documents from confined set (e.g., Wikipedia) using light retriever such as BM25 and dense retriever. However, real-world applications often rely on the web which has much bigger pool of documents from broader open-domain. In turn, we should consider future extension of this work to utilizing larger public search engines. Last, we may consider alternative yet more effective means of query sampling during training other than a small set of random samples, as the current design. Other possible methods include imposing additional guidance for query sampling or multiple refinements of query sampling and knowledge retrieval for more diverse set.", + "bbox": [ + 115, + 84, + 490, + 382 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 117, + 398, + 263, + 411 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This work presents no direct ethical issues. However, there is a possibility that the generated responses may inadvertently contain inadequate, biased, or discriminatory content. We do not introduce new datasets, and all experiments are conducted using publicly available datasets.", + "bbox": [ + 115, + 424, + 487, + 516 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 117, + 530, + 260, + 542 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University). In addition, this work was also supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea) Gwangju Metropolitan City. Finally, we would also like to thank Brain Cloud Team at Kakao Brain for their support.", + "bbox": [ + 115, + 548, + 487, + 736 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 117, + 766, + 211, + 778 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Leonard Adolphs, Benjamin Boerschinger, Christian Buck, Michelle Chen Huebscher, Massimiliano Ciaramita, Lasse Espeholt, Thomas Hofmann, Yannic Kilcher, Sascha Rothe, Pier Giuseppe Sessa, et al. 2021a. Boosting search engines with interactive agents. arXiv preprint arXiv:2109.00527.", + "bbox": [ + 117, + 839, + 485, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021b. Reason first, then respond: Modular generation for knowledge-infused dialogue. arXiv preprint arXiv:2111.05204.", + "bbox": [ + 512, + 86, + 880, + 137 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Online. Association for Computational Linguistics.", + "bbox": [ + 512, + 148, + 882, + 253 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165.", + "bbox": [ + 510, + 263, + 882, + 419 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mingzhu Cai, Siqi Bao, Xin Tian, Huang He, Fan Wang, and Hua Wu. 2023. Query enhanced knowledge-intensive conversation via unsupervised joint modeling. In ACL.", + "bbox": [ + 512, + 430, + 882, + 482 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Xiuyi Chen, Fandong Meng, Peng Li, Feilong Chen, Shuang Xu, Bo Xu, and Jie Zhou. 2020. Bridging the gap between prior and posterior knowledge selection for knowledge-grounded dialogue generation. In EMNLP, pages 3426-3437.", + "bbox": [ + 512, + 493, + 882, + 558 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Oleg Rokhlenko, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering. arXiv preprint arXiv:2210.15777.", + "bbox": [ + 512, + 569, + 882, + 620 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling", + "bbox": [ + 510, + 631, + 882, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2692", + "bbox": [ + 482, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language modeling with pathways. arXiv preprint arXiv:2204.02311.", + "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241.", + "Nijkamp Erik, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. 2019. Learning non-convergent non-persistent short-run mcmc toward energy-based model. In Advances in Neural Information Processing Systems (NeurIPS).", + "Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In NAACL.", + "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909.", + "Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration.", + "Xinxian Huang, Huang He, Siqi Bao, Fan Wang, Hua Wu, and Haifeng Wang. 2021. PLATO-KAG: Unsupervised knowledge-grounded conversation via joint modeling. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, pages 143-154, Online. Association for Computational Linguistics.", + "Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282.", + "Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. 1999. An introduction to variational methods for graphical models. Machine learning, 37:183-233.", + "Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906.", + "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR.", + "Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search.", + "Diederik P Kingma and Max Welling. 2013. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8460-8478, Dublin, Ireland. Association for Computational Linguistics.", + "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension.", + "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474.", + "Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2021. Zero-resource knowledge-grounded dialogue generation.", + "Zekang Li, Cheng Niu, Fandong Meng, Yang Feng, Qian Li, and Jie Zhou. 2019. Incremental transformer with deliberation decoder for document grounded conversations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 12-21, Florence, Italy. Association for Computational Linguistics.", + "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. arXiv preprint arXiv:1902.04911.", + "Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909.", + "Shilei Liu, Xiaofeng Zhao, Bochao Li, Feiliang Ren, Longhui Zhang, and Shujuan Yin. 2021. A Three-Stage Learning Framework for Low-Resource Knowledge-Grounded Dialogue Generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.", + "Shuman Liu, Hongshen Chen, Zhaochun Ren, Yang Feng, Qun Liu, and Dawei Yin. 2018. Knowledge diffusion for neural dialogue generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)," + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2693", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pages 1489-1498, Melbourne, Australia. Association for Computational Linguistics.", + "Longxuan Ma, Weinan Zhang, Runxin Sun, and Ting Liu. 2020. A compare aggregate transformer for understanding document-grounded dialogue. arXiv preprint arXiv:2010.00190.", + "Nikita Moghe, Siddhartha Arora, Suman Banerjee, and Mitesh M Khapra. 2018. Towards exploiting background knowledge for building conversation systems. arXiv preprint arXiv:1809.08205.", + "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2017. MS MARCO: A human-generated MAchine reading COprehension dataset.", + "Rodrigo Nogueira and Kyunghyun Cho. 2017. Task-oriented query reformulation with reinforcement learning. arXiv preprint arXiv:1704.04572.", + "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In ICLR.", + "Gonçalo Raposo, Rui Ribeiro, Bruno Martins, and Luísca Coheur. 2022. Question rewriting? assessing its importance for conversational question answering. In Lecture Notes in Computer Science, pages 199-206. Springer International Publishing.", + "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics.", + "Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics.", + "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389.", + "Kurt Shuster, Mojtaba Komeili, Leonard Adolphs, Stephen Roller, Arthur Szlam, and Jason Weston. 2022a. Language models that seek for knowledge: Modular search & generation for dialogue and prompt completion. arXiv preprint arXiv:2203.13224.", + "Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya Chen, Kushal Arora, Joshua Lane, et al. 2022b." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage. arXiv preprint arXiv:2208.03188.", + "Weiwei Sun, Pengjie Ren, and Zhaochun Ren. 2023. Generative knowledge selection for knowledge-grounded dialogues. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2077-2088, Dubrovnik, Croatia. Association for Computational Linguistics.", + "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, abs/2201.08239.", + "David Thulke, Nico Daheim, Christian Dugast, and Hermann Ney. 2021. Efficient retrieval augmented generation from unstructured knowledge for task-oriented dialog. arXiv preprint arXiv:2102.04643.", + "Megan Ung, Jing Xu, and Y-Lan Boureau. 2022. SaFeR-Dialogues: Taking feedback gracefully after conversational safety failures. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6462-6481, Dublin, Ireland. Association for Computational Linguistics.", + "Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021. Question rewriting for conversational question answering. In Proceedings of the 14th ACM international conference on web search and data mining, pages 355-363.", + "Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 921-930.", + "Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, Hannaneh Hajishirzi, Mari Ostendorf, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning. In EMNLP." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2694", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966.", + "bbox": [ + 115, + 85, + 487, + 165 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jing Xu, Arthur Szlam, and Jason Weston. 2022a. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics.", + "bbox": [ + 114, + 173, + 487, + 254 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Jing Xu, Megan Ung, Mojtaba Komeili, Kushal Arora, Y-Lan Boureau, and Jason Weston. 2022b. Learning new skills after deployment: Improving open-domain internet-driven dialogue with human feedback. arXiv preprint arXiv:2208.03270.", + "bbox": [ + 114, + 262, + 487, + 329 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yan Xu, Deqian Kong, Dehong Xu, Ziwei Ji, Bo Pang, Pascale Fung, and Ying Nian Wu. 2023. Diverse and faithful knowledge-grounded dialogue generation via sequential posterior inference.", + "bbox": [ + 115, + 338, + 487, + 391 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 1933-1936.", + "bbox": [ + 114, + 400, + 487, + 479 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Haolan Zhan, Lei Shen, Hongshen Chen, and Hainan Zhang. 2021. Colv: A collaborative latent variable model for knowledge-grounded dialogue generation. In EMNLP.", + "bbox": [ + 115, + 489, + 487, + 541 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics.", + "bbox": [ + 114, + 551, + 487, + 643 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Yizhe Zhang, Siqi Sun, Xiang Gao, Yuwei Fang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2022. Retgen: A joint framework for retrieval and grounded text generation modeling. In AAAI.", + "bbox": [ + 115, + 653, + 485, + 706 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP.", + "bbox": [ + 114, + 715, + 487, + 768 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018a. Common-sense knowledge aware conversation generation with graph attention. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 4623-4629. AAAI Press.", + "bbox": [ + 114, + 778, + 487, + 857 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W Black. 2018b. A dataset for document grounded conversations. In EMNLP.", + "bbox": [ + 115, + 866, + 487, + 906 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/d0a7a1157173c9833922c948fcaeebe5c9161c7c66a63f40f21c981412d73a50.jpg", + "table_caption": [ + "A Training Details" + ], + "table_footnote": [], + "table_body": "
ParameterValue
Training Steps10,000
Batch Size128
OptimizerAdamW
Learning Rate (LR)5e-5
LR SchedulerCosine
Adam β10.9
Adam β20.999
Adam ε1e-8
Warmup Steps1000
", + "bbox": [ + 539, + 118, + 853, + 258 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/1c730c27136aa5a9da7f3fd4151d394265af8f771186b43f3b76c2f44fb989e7.jpg", + "table_caption": [ + "Table 13: Hyperparameters used for stage-1 training. $k$ indicates number of retrieved documents." + ], + "table_footnote": [], + "table_body": "
WoWQReCC
Train Epochs31
Batch Size6464
OptimizerAdamWAdamW
Learning Rate (LR)1e-45e-4
LR SchedulerCosineCosine
Adam β10.90.9
Adam β20.9990.999
Adam ε1e-81e-8
Warmup Steps1000
Min Query Length11
Max Query Length3232
Top-k100100
Top-p0.950.5
Temperature1.02.0
", + "bbox": [ + 544, + 338, + 847, + 544 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 14: Hyperparameters used for training ELVM and its.", + "bbox": [ + 507, + 552, + 880, + 581 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 13 we provide hyperparameters utilized during this training process. We also report hyperparameters used for training ELVM and its variants model (ELVM-OK and ELVM-EM-Like) in Table 14. In addition, during training of QReCC, we adopt a strategy to decrease the knowledge pool size from 54M to 1M. This reduction aims to mitigate latency issues during document retrieval. Concretely, the 1M pool is constructed by initially gathering the ground-truth relevant passage for each instance in the training dataset then adding randomly sampled documents. During the evaluation stage, we utilize the original full-size knowledge pool of 54M passages.", + "bbox": [ + 507, + 598, + 882, + 819 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Upper Bound Performance", + "text_level": 1, + "bbox": [ + 507, + 829, + 781, + 847 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We believe in the value of understanding the utmost capability in an ideal scenario. To this end, we pursue an upper-bound performance analysis for both the WoW and QReCC tasks. This approach", + "bbox": [ + 507, + 854, + 880, + 917 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "2695", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/0c14ace2b5871033696403c1686574f0b5bf91cdc0f68dd252e001d4c7c2717c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
ELVM14.614.711.929.313.918.014.511.829.213.7
ELVM-w-GT-Doc9.722.118.540.923.910.521.517.940.723.3
", + "bbox": [ + 176, + 80, + 820, + 131 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1e8a502aa77db1d1cbc69f99251611957c32cb3b72860cbd7ccb66aca877894b.jpg", + "table_caption": [ + "Table 15: Performance comparison in WoW in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + ], + "table_footnote": [], + "table_body": "
ModelF1 ↑EM ↑
Q Rewriting (Raposo et al., 2022)18.91.0
DPR-IHN (Kim and Kim, 2022)30.44.7
QKConv (Cai et al., 2023)33.55.9
ELVM36.56.2
ELVM-w-GT-Doc55.416.2
", + "bbox": [ + 119, + 192, + 480, + 286 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "involves training our response generator under the assumption that the model has access to the ground-truth document during both training and evaluation. The performance metrics of this specific setup are designated as \"ELVM-w-GT-Doc\" in Table 15 and Table 16. From both tables, it's evident that there still remains a margin for improvement for both tasks.", + "bbox": [ + 112, + 368, + 485, + 492 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C Document Recall Performance", + "text_level": 1, + "bbox": [ + 112, + 513, + 415, + 527 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/1babfe3a5ac5cf8e69512ebd39e8c77924ff6297e1c66560f6d1cdefc5dcc9ad.jpg", + "table_caption": [ + "Table 16: Performance comparison in QReCC in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + ], + "table_footnote": [], + "table_body": "
Query fromR@1R@5R@10
Dialogue History33.458.668.7
Last Utterance25.550.862.6
R2C2-PT17.433.039.5
ELVM-β=030.655.865.7
ELVM35.259.968.6
ELVMDR37.565.675.1
", + "bbox": [ + 154, + 554, + 445, + 652 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We report the document recall performance on QReCC, including different types of queries, as show in Table 17. Our proposed model, ELVM, surpasses both its variants and non-model generated queries such as using dialogue context as a query. The document recall performance on QReCC, encompassing various query types, is presented in Table 17. Our proposed ELVM model outperforms its variants as well as non-model generated queries, including the use of dialogue context as a query. Moreover, incorporating the dense retriever for reranking documents $(\\mathrm{ELVM}_{\\mathrm{DR}})$ leads to further improvements in recall.", + "bbox": [ + 112, + 713, + 487, + 917 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/4a4f0f21a12b452a41062e3c44606dda4c413cd8526270faf1eff818377f1ae6.jpg", + "table_caption": [ + "Table 17: Query performance comparison in QReCC for different types of queries." + ], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑R@10 ↑B4 ↑R2 ↑R@5 ↑R@10 ↑
k=511.913.968.480.311.813.768.380.3
k=5Bulk9.410.962.773.98.29.547.261.9
k=1010.312.264.876.410.612.764.876.9
k=10Bulk10.212.262.174.19.010.547.261.6
", + "bbox": [ + 512, + 200, + 887, + 277 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 18: Effect of increasing the document pool size to 1000 on the performance of ELVM. As before, $k$ indicates the number of retrieved documents used during train and inference.", + "bbox": [ + 507, + 286, + 880, + 341 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Knowledge Pool Scaling", + "text_level": 1, + "bbox": [ + 507, + 367, + 756, + 384 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In order to create a more realistic and challenging retrieval scenario, we expand the original setting of the WoW dataset by increasing the number of relevant documents per instance to 1000. This modification better reflects real-world information retrieval scenarios where a vast array of both relevant and irrelevant documents are typically encountered. We construct this extended dataset by randomly augmenting $1000 - K$ documents from the WoW dataset to the document pool for each instance, where $K$ represents the number of annotated relevant documents for each instance. On average, $K \\approx 20$ for each instance in WoW dataset.", + "bbox": [ + 505, + 392, + 882, + 596 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Results in Table 18 show performances of ELVM models with different number of retrieved documents, $k$ , and knowledge pool size, where models with increased knowledge pool has tailing keyword Bulk at then end of their names. Despite the increase in complexity of the retrieval task, ELVM demonstrates robust performance, with only a marginal drop in metrics compared to the original setting. Furthermore, it is important to highlight that even under these more difficult conditions with 1000 relevant documents, ELVM still outperforms SPI by a significant margin.", + "bbox": [ + 507, + 598, + 882, + 785 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "E Effect of Parameter Sharing", + "text_level": 1, + "bbox": [ + 507, + 799, + 789, + 815 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "In our ELVM training, parameters are shared between the query and response generators. To probe this design choice, we decoupled these parameters, resulting in the ELVM-Decouple configuration, as detailed in Table 19. Upon evaluation on WoW's test seen and unseen tasks, ELVM-", + "bbox": [ + 507, + 824, + 882, + 917 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "2696", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/dc402caeb57da018a3932ab96f09dac3b7b55c399894fae2ca181de5a1aff157.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
ELVM14.611.913.918.011.813.7
ELVM-Decouple16.212.614.521.612.214.0
", + "bbox": [ + 122, + 80, + 478, + 128 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/7d93c1e91159d46e3fd98d3e1f26bb3dd58396e43d6ceddee28f334e05e4b8b6.jpg", + "table_caption": [ + "Table 19: Performance comparison on the WoW dataset using the ELVM algorithm. ELVM denotes training with shared parameters between query and response generators, while ELVM-Decouple denotes a non-shared parameter approach." + ], + "table_footnote": [], + "table_body": "
ModelSeenUnseen
F1 ↑KF1 ↑F1 ↑KF1 ↑
CoLV20.318.218.517.5
KnownGPT22.023.820.522.1
SPI----
R2C2-PT15.420.515.722.3
ELVM-OK23.033.022.933.0
ELVM-EM-Like24.533.223.533.5
ELVM27.240.827.040.3
", + "bbox": [ + 122, + 219, + 478, + 350 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 20: F1 and KF1 scores for ELVM and baseline models.", + "bbox": [ + 112, + 361, + 485, + 388 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Decouple showed a slight edge over the standard ELVM. Nevertheless, the benefits of shared parameters—particularly regarding computational efficiency and knowledge transfer during training and inference—cannot be understated.", + "bbox": [ + 112, + 414, + 489, + 492 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "F Additional Metrics", + "text_level": 1, + "bbox": [ + 114, + 505, + 315, + 520 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We report the evaluation of our models on the WoW task using both F1 and knowledge F1 (KF1) metrics in Table 20. For reference, values pertaining to previous models are derived from Sun et al. (2023).", + "bbox": [ + 112, + 530, + 489, + 593 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "G Human Evaluation", + "text_level": 1, + "bbox": [ + 114, + 605, + 319, + 619 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "For a comprehensive evaluation, we conduct the human evaluation between ELVM and KnownGPT on seen and unseen tasks in WoW (Dinan et al., 2018). We randomly select 50 samples (25 samples per task), and each sample is evaluated by 11 different human experts. In specific, the two generated responses from each model in the same context are assigned to the annotators. For A/B testing, we give one score to the model if its response is received an equally good or better than the other one. Figure 2 shows the annotator instructions for the two aspects.", + "bbox": [ + 112, + 630, + 489, + 819 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "H Generated Samples", + "text_level": 1, + "bbox": [ + 114, + 831, + 324, + 847 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Generated samples for the WoW test unseen are presented in Table 21 and Table 22. Similarly, Table 23 and Table 24 showcase the generated samples for the QReCC test.", + "bbox": [ + 112, + 854, + 489, + 917 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "2697", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/0e6037791cb85ab75bb71f9d8bd0a9a250b15dd9ccfbd0bc8f135e7a88833997.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dialogue ContextHey, do you like cardigans? I like the modern version without no buttons and hangs open by design. Me too. My fav is the waterfall style but it's going out of fashion now I think [...] Cardigan (sweater) By contrast, a pullover does not open in front but must be "pulled over" the head to be worn. [...] But the pullover does not open in front but must be "pulled over" the head to be worn.
GT knowledge
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: i like the hand-knitted version. it's more comfortable. i like mine made of a cardigan.
R2C2-PTQuery: fallen fashion GT knowledge selected: False Response: I'm sorry, I didn't mean to offend you.
ELVM-β = 0Query: fallen fashion cardigans GT knowledge selected: True Response: I like the one that is tied is instead considered a robe.
ELVM w/ BM25Query: fallen waterfall cardigan style fallout cardigans GT knowledge selected: True Response: I like the pullover style that does not open in front but must be "pulled over" the head to be worn.
ELVM w/ DRQuery: fallen waterfall cardigan GT knowledge selected: True Response: I like the pullover that does not open in front but must be "pulled over" the head to be worn.
Dialogue Context[...] Was she diagnosed as a child? It is much easier to diagnose children, since the symptoms begin occurring early in life
GT knowledgeYes it started when she was 6 years old. She was 23 when I met her she was already recieving treatment [...] These fellowships include pediatric neurosurgery, trauma/neurocritical care, functional and stereotactic surgery, surgical neuro-oncology, radiosurgery, neurovascular surgery, skull-base surgery, peripheral nerve and spine surgery. [...] My first fellowship was in pediatric neurosurgery. It is great that she is doing better.
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: did she cry during the periods of depression? she was crying, a negative outlook on life, and poor eye contact with others.
R2C2-PTQuery: types of epilepsy GT knowledge selected: False Response: I am glad to hear she is receiving treatment. Dementia is such a devastating disease.
ELVM-β = 0Query: types of epilepsy GT knowledge selected: False Response: I am glad she is receiving treatment. Dementia has a significant effect on a person's caregivers.
ELVM w/ BM25Query: neurosurgery epilepsy neurocognitive neurosciences GT knowledge selected: True Response: I am glad to hear that she is receiving treatment. Neurosurgery is a branch of medicine that specializes in the treatment of epilepsy.
ELVM w/ DRQuery: neurosurgery treatment for epilepsy GT knowledge selected: True Response: Did she undergo a fellowship? Many neurosurgeons pursue additional training in the form of a fellowship, after residency or in some cases, as a senior resident.
", + "bbox": [ + 115, + 148, + 884, + 809 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Table 21: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response.", + "bbox": [ + 112, + 818, + 882, + 847 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "2698", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/fc63143279353e0e504e68c20a06ce19be1f988a10d973c6632f3ba79785021a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dialogue ContextDo you know anything about the band called Insane Clown Posse? \nInsane Clown Posse (ICP) is an American hip hop duo composed of Violent J, and Shaggy 2 Dope. [...] \nThat's quite a lot of albums, what else do you know about ICP? \n[...] The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas" and "Big Money Rustlas". [...] \nThe duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas
GT knowledge
GT Response
KnownedGPTResponse: Query: n/a \nGT knowledge selected: False
Response: the duo has earned two platinum and five gold albums. they have also earned two platinum and five gold albums
R2C2-PTQuery: ICP wiki \nGT knowledge selected: False
Response: The duo has earned two platinum ,with 5 golden album
ELVM-β = 0Query: ICP facts \nGT knowledge selected: False
Response: The group has been around for a long time .
ELVM w/ BM25Query: Insane clown posse ICP facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
ELVM w/ DRQuery: Insane clown posse facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
Dialogue Context[...] Athletes normally eat meat in large quantities. \nExactly. I have also added much meat eating as part of my special diet. They are rich in protein. Especially, red meat
GT knowledge[...] In nutritional science, on the other hand, red meat is defined as any meat that has more myoglobin than white meat, white meat being defined as non-dark meat from chicken (excluding leg or thigh), or fish. [...]
GT ResponseRed meat is divined as any meat that has myoglobin than white
KnownedGPTQuery: n/a \nGT knowledge selected: False
Response: red meat is classified as white meat. i think it's interesting that only flesh from mammals or fowl is classified as red.
R2C2-PTQuery: protein powder \nGT knowledge selected: False
Response: Athletes normally eat meat in large quantities.
ELVM-β = 0Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ BM25Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ DRQuery: red meat protein \nGT knowledge selected: True
Response: Red meat is defined as any meat that has more myoglobin than white meat.
", + "bbox": [ + 115, + 128, + 884, + 828 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table 22: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response.", + "bbox": [ + 112, + 835, + 882, + 866 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "2699", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/a900e3cf666337bc10dbbf58a3b055dc955c9f88d06eeda9e3538b678ca509f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dialogue ContextWhen did Mark Spitz' swimming career end? Following the Munich Olympics [...] What else did Mark Spitz do besides Speedo swimsuit ads after his swimming career? [...] However, as the memory of his feats receded, so did his endorsement and promotional deals. As his endorsements faded, he started a successful real-estate company in Beverly Hills. He bought a Ferrari and says he made [...]
GT knowledge
GT ResponseMark Spitz started a successful real-estate company in Beverly Hills
R2C2-PTQuery: What did mark spitz do after his olympic career? \nGT knowledge selected: False \nResponse: In 1999 Spitz ranked No. 33 on ESPN SportsCentury 50 Greatest Athletes, the only aquatic athlete to make the list.
ELVM-β = 0Query: Mark Spitz Speedo Swimsuit ads after speedo swimming career ends mark spitz m Munich olympics 2012 \nGT knowledge selected: False \nResponse: Spitz did Schick razor commercials and Speedo swimsuit ads
ELVM w/ BM25Query: Spitz endorsements after swimming career ends \nGT knowledge selected: True \nResponse: Mark Spitz started a successful real-estate company in Beverly Hills.
Dialogue ContextWhat was the New Scheme? \nThe 'new scheme' and the 'new dedication' respectively, initially seen as a [...] \n[...] \nDid Ahmadi followers of the New Scheme act as missionaries? \nAs part of this Mahmood Ahmad appointed 5 men to survey the Punjab [...] \nDid Mirza Basheer-ud-Din Mahmood Ahmad use missionaries to try to spread the New Scheme to foreign countries?
GT knowledge[...] In time the scheme produced a vast amount of literature in defence of Islam in general and the Ahmadiyya beliefs in particular. The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance. As part of this Mahmood Ahmad appointed [...]
GT ResponseThe funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
R2C2-PTQuery: Mirza Basher-ud Din Mahmood Ahmad missionaries \nGT knowledge selected: False \nResponse: Mirza Basheer-ud-Din Mahmood Ahmad used missionaries to try to spread the New Scheme to foreign countries.
ELVM-β = 0Query: New Scheme Mahmud Ahmad missionaries to spread New Scheme in foreign countries \nGT knowledge selected: True \nResponse: Yes, I have. I have experienced heartache, also known as loss of a loved one.
ELVM w/ BM25Query: New scheme missionaries used to spread Ahmadiyyah teachings in foreign countries \nGT knowledge selected: True \nResponse: The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
", + "bbox": [ + 115, + 187, + 882, + 770 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table 23: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response.", + "bbox": [ + 112, + 778, + 882, + 809 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "2700", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/6d4efa4c5edc635a4576d6ee14969c14be87d094b8218bb9d915a2a8cf1b96c5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dialogue ContextHow did Stephen Sondheim collaborate with James Lapine?\nStephen Sondheim found a new collaborator in [...]What did Stephen Sondheim and James Lapine do together in the 1980s?\nTheir first collaboration was Sunday in the Park with George (1984).Was Stephen Sondheim and James Lapine's first collaboration a success?\nStephen Sondheim and James Lapine won the 1985 Pulitzer Prize [...]Did Stephen Sondheim and James Lapine collaborate in other projects, besides Sunday in the Park with George?[...] last work together was the rhapsodic Passion (1994).Was Stephen Sondheim and James Lapine's last work a success?\nSondheim and Lapine's last work together was the rhapsodic Passion (1994), adapted from Ettore Scola's Italian film Passione D'Amore. With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.[46] Later work Assassins opened [...]With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
GT knowledge
GT Response
R2C2-PTQuery: Was the play "rhapsodic passion" a success or a failure?\nGT knowledge selected: TrueResponse: Sondheim and Lapine's last work together was the rhapsodic Passion (1994).
ELVM-β = 0Query: Stephen Sonds and Lapine Rhapsodic passion1994 Stephen sondheim, James lapine plays 12 dreams off-broadway p Pulitzer prize for play Sunday in Park.GT knowledge selected: FalseResponse: Stephen Sondheim and James Lapine's last work together was the rhapsodic Passion (1994).
ELVM w/ BM25Query: Sondheim last work was the Rhapsodic passion (1994) Pulitzer prize won for Drama. Lapine wrote the play Twelve Dreams.GT knowledge selected: TrueResponse: With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
Dialogue ContextDid Frank Hague commit any questionable acts?\nYes, he took kickbacks in cash, and relied on two ordinances of dubious [...]Did Frank Hague do anything illegal?\nHe passed an unconstitutional ordinance that gave himself as commissioner the power to [...]Did anyone attempt to prosecute Frank Hague?[...] Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. [38] He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash. [...]Yes, but he was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
GT knowledge
GT Response
R2C2-PTQuery: Frank Hague prosecutionGT knowledge selected: FalseResponse: No, no one tried to prosecute Frank Hague.
ELVM-β = 0Query: Frank Hague ordinance ordinance frank Hague charges Frank Hague Hague for flouting constitutionality ordinancesGT knowledge selected: FalseResponse: Hague was found guilty of contempt of court for ignoring the subpoena. He was fined $100 and stripped of his duties as Deputy Sheriff.
ELVM w/ BM25Query: Hague prosecution attempt failed after Hague passed unconstitutional ordinance. no Hague was prosecuted.GT knowledge selected: TrueResponse: Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
", + "bbox": [ + 115, + 135, + 884, + 822 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Table 24: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response.", + "bbox": [ + 112, + 829, + 882, + 860 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "2701", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Human Evaluation Annotation Tool", + "text_level": 1, + "bbox": [ + 250, + 155, + 593, + 171 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Components of each example", + "text_level": 1, + "bbox": [ + 250, + 183, + 426, + 195 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Each example may include one of more of the following", + "bbox": [ + 250, + 205, + 463, + 214 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- History: topic related to the conversation and the dialogue history. Examples may not contain 'history'", + "- Input text: Utterance of the opponent", + "- Ground truth response of the example, given as a reference. The annotator may use the reference in decision.", + "- Two responses from different models." + ], + "bbox": [ + 250, + 223, + 668, + 262 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Annotation Standards", + "text_level": 1, + "bbox": [ + 250, + 274, + 381, + 285 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "The annotator marks which model of the given two is better on three different categories: Fluency and Relevance. The deifintion of each category is", + "bbox": [ + 248, + 294, + 746, + 316 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Fluency measures the correctness, completeness, and semantic consistency of the response.", + "- Relevance measures the degree of relevance of the knowledge and response to the history." + ], + "bbox": [ + 250, + 324, + 606, + 343 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "History:", + "text_level": 1, + "bbox": [ + 250, + 386, + 289, + 395 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Singing", + "text_level": 1, + "bbox": [ + 250, + 398, + 289, + 409 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "I love to sing. Some of my favorite genres to sing are blues, pop, and rock. Do you like to sing?", + "bbox": [ + 250, + 411, + 685, + 420 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Yes very much so. I love older artists like Ray Charles and Billie Holiday.", + "bbox": [ + 250, + 423, + 583, + 434 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Oh yeah I love Billie Holiday. She had no formal music training but her vocal delivery made up for it.", + "bbox": [ + 250, + 436, + 714, + 447 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Input Text:", + "text_level": 1, + "bbox": [ + 250, + 457, + 302, + 467 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "I've been getting into the strokes as of late. Do you like them?", + "bbox": [ + 250, + 469, + 537, + 480 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Reference:", + "text_level": 1, + "bbox": [ + 250, + 489, + 292, + 497 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "I used to listen to The Strokes a lot back in high school. Their debut album was added to Rolling Stones 500 greatest albums of all time at number 199.", + "bbox": [ + 250, + 499, + 737, + 518 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Model A Response:", + "text_level": 1, + "bbox": [ + 250, + 561, + 337, + 570 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Yes I do like them. I like all kinds of music, but I'm not a big fan of the Strokes.", + "bbox": [ + 250, + 573, + 610, + 583 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Model B Response:", + "text_level": 1, + "bbox": [ + 250, + 594, + 339, + 604 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "I do like them. I also like to wear my Strokes shirt during May because it is national stroke awareness month.", + "bbox": [ + 250, + 606, + 717, + 629 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Annotation", + "text_level": 1, + "bbox": [ + 250, + 649, + 361, + 664 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Fluency", + "text_level": 1, + "bbox": [ + 250, + 675, + 287, + 684 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Both responses are fluent.", + "$\\bigcirc$ Model A response is more fluent.", + "Model B response is more fluent.", + "$\\bigcirc$ Neither response is fluent." + ], + "bbox": [ + 252, + 690, + 394, + 765 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Previous", + "bbox": [ + 260, + 790, + 304, + 797 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Relevance", + "text_level": 1, + "bbox": [ + 418, + 675, + 463, + 684 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Both knowledges and responses are relevant", + "Model A response is more relevant.", + "Model B response is more relevant.", + "$\\bigcirc$ Neither is relevant." + ], + "bbox": [ + 418, + 690, + 559, + 777 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Next", + "bbox": [ + 510, + 790, + 536, + 797 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Made with Streamlit", + "bbox": [ + 250, + 887, + 336, + 896 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Figure 2: Screenshot of the annotation tool.", + "bbox": [ + 349, + 912, + 645, + 926 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": ">", + "bbox": [ + 186, + 91, + 196, + 99 + ], + "page_idx": 19 + }, + { + "type": "header", + "text": "三", + "bbox": [ + 798, + 91, + 813, + 99 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "2702", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 19 + } +] \ No newline at end of file diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_model.json b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fb6fbff23145bd8e6dd7abb071465f4be805155b --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_model.json @@ -0,0 +1,3441 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.273, + 0.085, + 0.725, + 0.124 + ], + "angle": 0, + "content": "Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.135, + 0.882, + 0.17 + ], + "angle": 0, + "content": "Gunsoo Han\\(^{1}\\) Daejin Jo\\(^{1}\\) Daniel Wontae Nam\\(^{1}\\) Eunseop Yoon\\(^{2}\\) Taehwan Kwon\\(^{3*}\\) Seungeun Rho\\(^{4*}\\) Kyoung-Woon On\\(^{1}\\) Chang D. Yoo\\(^{2}\\) Sungwoong Kim\\(^{5\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.172, + 0.884, + 0.19 + ], + "angle": 0, + "content": "\\(^{1}\\)Kakao Brain \\(^{2}\\)KAIST \\(^{3}\\)KRAFTON \\(^{4}\\)Georgia Institute of Technology \\(^{5}\\)Korea University" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.201, + 0.742, + 0.217 + ], + "angle": 0, + "content": "gunsoo.han@kakaobrain.com swkim01@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.282, + 0.461, + 0.701 + ], + "angle": 0, + "content": "Knowledge-grounded dialogue generation requires to first retrieve appropriate external knowledge based on a conversational context and then generate a response grounded on the retrieved knowledge. In general, these two sequential modules, a knowledge retriever and a response generator, have been separately trained by supervised data for each module. However, obtaining intermediate labels of the ground-truth knowledge is expensive and difficult especially in open-domain conversation. Latent variable modeling can circumvent it and enables a joint training without the knowledge supervision. In this paper, we propose an efficient algorithm for this latent variable modeling that is able to leverage a large amount of dialogue data. In specific, rather than directly training the complex retriever, we adapt a query generator with an off-the-shelf retriever, and the query generator and response generator are simultaneously trained over the latent variable of query. Moreover, we employ the evidence lower bound as a training objective and modify it to efficiently and robustly perform the joint training. Experimental results on diverse knowledge-grounded dialogue datasets show that the proposed algorithm achieves state-of-the-art performances even without the use of the annotated knowledge while maintaining the efficiency and scalability." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.715, + 0.26, + 0.73 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.741, + 0.49, + 0.884 + ], + "angle": 0, + "content": "Recently, knowledge-grounded dialogue generation has drawn increasing attention especially as a key ingredient for open-domain conversational agents (Dinan et al., 2018; Zhou et al., 2018b; Zhan et al., 2021; Liu et al., 2018; Zhou et al., 2018a; Lian et al., 2019; Zhao et al., 2020; Kim et al., 2020; Chen et al., 2022; kom; Shuster et al., 2022a,b; Cai et al., 2023; Wu et al., 2022; Lewis et al., 2020; Huang et al., 2021; Thulke et al., 2021; Anantha" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.885, + 0.364 + ], + "angle": 0, + "content": "et al., 2021). It usually obtains knowledge from external resources such as Wikipedia-based databases and web search engines since internal knowledge even in large-scale parametric language models (Brown et al., 2020; Thoppilan et al., 2022; Chowdhery et al., 2022) is incomplete or outdated, and moreover it can provide hallucinated information." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.366, + 0.885, + 0.744 + ], + "angle": 0, + "content": "In order for a dialogue response to be grounded on such external knowledge, the conversational agent generally consists of a knowledge retriever, which retrieves knowledge corresponding to a given dialogue context, followed by a response generator that produces an informative response based on the dialogue context and the retrieved knowledge. In many previous methods, supervised learning has often been applied to independently optimize each module using the ground-truth or gold knowledge (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). However, human annotation of knowledge information is cumbersome, expensive, and often incomplete due to the existence of multiple possible knowledge candidates and the overlooking of the target response. In addition, existing automatic annotations are generally limited to a simple extractive question answering (Zhao et al., 2020; Liu et al., 2021). This difficulty in obtaining annotations of knowledge information could be severe under the open-domain conversation and hinders the use of large-scale dialogue data." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.746, + 0.887, + 0.919 + ], + "angle": 0, + "content": "Therefore, there have been a number of recent approaches to learn the knowledge retriever and the response generator without the knowledge supervision. In specific, they have treated the retrieved knowledge, document, or passage as an unobserved latent variable and adapt latent variable modeling based on approximated marginalization (e.g. top-k) (Lewis et al., 2020; Huang et al., 2021; Cai et al., 2023; Guu et al., 2020), reinforcement learning (Zhao et al., 2020; Zhang et al., 2022; Chen et al., 2022; Wu et al., 2022) or variational methods (Zhan" + }, + { + "type": "page_footnote", + "bbox": [ + 0.141, + 0.892, + 0.321, + 0.905 + ], + "angle": 0, + "content": "* Work done at Kakao Brain" + }, + { + "type": "page_footnote", + "bbox": [ + 0.141, + 0.905, + 0.295, + 0.919 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.141, + 0.892, + 0.321, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2683" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.78, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2683-2702 \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.369 + ], + "angle": 0, + "content": "et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020; Chen et al., 2020; Xu et al., 2023). However, joint training of the retriever along with the generator under this latent variable modeling has some restrictions in utilizing the retriever. For example, a retriever needs to produce a differentiable prior probability for the gradient propagation through the marginalization or to be intermittently updated to rebuild the whole passage index during training (Karpukhin et al., 2020; Lewis et al., 2020; Zhang et al., 2022). This leads to limitation in facilitating complicated or large-scale retrievers and further in leveraging large-scale data. In addition, the posterior probability used in the previous variational methods has been modeled by a separated network, which grows the training complexity and also incurs the discrepancy between the training-time and test-time knowledge samplings." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.374, + 0.49, + 0.563 + ], + "angle": 0, + "content": "To overcome such restrictions, in this paper, we propose an efficient latent variable modeling, named \\(ELVM\\), for knowledge-grounded dialogue generation. In particular, we reduce the burden for training a whole retriever by employing a query generator followed by an off-the-shelf retriever that is fixed during training. More precisely, in tandem with the response generator, the query generator rather than the retriever is jointly optimized by latent variable modeling, in which a query is taken as a latent variable, on paired observations of dialogue contexts and responses." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.568, + 0.49, + 0.852 + ], + "angle": 0, + "content": "To this end, we exploit the variational method using the evidence lower bound (ELBO) (Kingma and Welling, 2013; Jordan et al., 1999) and approximate the expected conditional likelihood in the ELBO by subset sampling from the prior distribution, which acts as the training objective for the response and the query generators. This approximation gets rid of an extra modeling of a surrogate posterior distribution or online posterior inference such as Markov Chain Monte Carlo (MCMC) and also reduces the training-inference discrepancy in knowledge retrieval. Moreover, we further modify the Kullback-Leibler (KL) regularization of the ELBO by constructing the approximated posterior distribution from the conditional likelihood and prior distribution and set this posterior distribution as a teacher for a distillation objective to further learn the query generator." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.856, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Experimental results show that the proposed ELVM allows to efficiently and robustly perform training without (1) the use of the annotated knowledge, (2) an explicit training of the knowledge" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.195 + ], + "angle": 0, + "content": "retrieval, and (3) a complex posterior sampling. Especially, it significantly outperforms previous state-of-the-art methods for knowledge-grounded dialogue generation. In addition, the proposed posterior distillation improves the performances over the baseline that solely maximizes the expectation of the conditional likelihood." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.197, + 0.877, + 0.212 + ], + "angle": 0, + "content": "Our main contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.216, + 0.887, + 0.293 + ], + "angle": 0, + "content": "- An efficient latent variable modeling is proposed in joint training of query generator and dialogue response generator without the knowledge supervision for knowledge-intensive dialogue generation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.295, + 0.887, + 0.373 + ], + "angle": 0, + "content": "- For realizing efficient yet robust joint training, the ELBO for the marginal likelihood is modified as the combination of the conditional likelihood objective and the posterior distillation objective, based on multiple prior samples." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.373, + 0.887, + 0.451 + ], + "angle": 0, + "content": "- The proposed ELVM demonstrates its effectiveness in performing unsupervised joint training and even significant performance improvements over previous methods, being a new state-of-the-art on benchmark datasets." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.216, + 0.887, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.47, + 0.666, + 0.485 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.501, + 0.806, + 0.531 + ], + "angle": 0, + "content": "2.1 Knowledge-Grounded Dialogue Generation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.541, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Knowledge-grounded conversation has been extensively studied through recently released public datasets (Dinan et al., 2018; kom; Xu et al., 2022b; Kwiatkowski et al., 2019; Anantha et al., 2021; Moghe et al., 2018). Existing approaches for this task have mostly exploited two successive modules, the knowledge retriever and the knowledge-grounded response generator. Many previous works have used manual annotations of gold knowledge provided by some of public datasets to optimize the modules (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). Specifically, Seeker (Shuster et al., 2022a) and BlenderBot3 (Shuster et al., 2022b) have recently proposed to build a series of modules by a single transformer with different prompts for each module, and trained its transformer on a large number of modular tasks using annotated datasets. However, this manual annotation is expensive, time-consuming, and often inaccurate, which impedes the utilization of a large-scale dialogue data that is especially necessary for open-domain conversation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "2684" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.388, + 0.1 + ], + "angle": 0, + "content": "2.2 Unsupervised Joint Training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.101, + 0.49, + 0.305 + ], + "angle": 0, + "content": "Recently, unsupervised training methods have been widely applied, and many of them have tries to jointly learn the modules without knowledge labels (Lewis et al., 2020; Huang et al., 2021; Zhao et al., 2020; Zhang et al., 2022; Zhan et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020). For example, PLATO-KAG (Huang et al., 2021) has approximated the marginal likelihood by top-k knowledge samples while RetGen (Zhang et al., 2022) has trained to reward knowledge retrieval with the highest utilization in response generation by reinforcement learning and mixture-of-experts ensembling." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.306, + 0.49, + 0.573 + ], + "angle": 0, + "content": "Meanwhile, latent variable modeling based on variational methods has also been developed by several recent works. CoLV (Zhan et al., 2021) has introduced collaborative latent spaces to reflect the inherent correlation between the knowledge selection and the response generation. SKT (Kim et al., 2020) has developed a sequential latent variable model for the multi-turn knowledge-grounded dialogue generation. In order to perform the posterior sampling of knowledge selection during joint training, some works have proposed to separately train the posterior distribution model (Paranjape et al., 2022; Lian et al., 2019) or the posterior information prediction model (Chen et al., 2020). Very recently, SPI (Xu et al., 2023) has applied short-run MCMC (Erik et al., 2019) for posterior sampling on the collaborative latent spaces." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.574, + 0.49, + 0.858 + ], + "angle": 0, + "content": "While these latent variable modeling algorithms can effectively perform unsupervised joint training, the entire training of the retriever is still difficult and imposes restrictions on the selection of the retriever in terms of both a search algorithm and a knowledge resource. Furthermore, the additional posterior sampling through separate networks or multiple iterations in the previous variational methods also leads to increased training complexity as well as the training-inference discrepancy in knowledge generation. In contrast, the proposed ELVM can employ any kind of retriever in latent variable modeling since it controls a retrieval output by only changing of a generated query. Moreover, ELVM removes an extra posterior modeling or sampling by prior subset sampling and approximated posterior distillation, which leads to efficient training without the discrepancy." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.868, + 0.307, + 0.883 + ], + "angle": 0, + "content": "2.3 Query Generation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.49, + 0.92 + ], + "angle": 0, + "content": "When an off-the-shelf retriever is used for knowledge retrieval, how to make an input text query" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.289 + ], + "angle": 0, + "content": "is important for obtaining appropriate documents. Especially, in knowledge-grounded dialogue, a self-contained query should be generated from the multi-turn dialogue context. A number of prior works have tried to train a supervised query rewriting model by human rewrites (Yu et al., 2020; Lin et al., 2020; Vakulenko et al., 2021; Voskarides et al., 2020). Similar to the knowledge annotation, the limitations of manual query annotation have come up with unsupervised learning of the query generator. A number of works have proposed a novel rewards for reinforcement learning of a query rewritten (Wu et al., 2022; Chen et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.291, + 0.887, + 0.48 + ], + "angle": 0, + "content": "QKConv (Cai et al., 2023) has proposed an unsupervised query enhanced method for knowledge-grounded conversation. This work is similar to ours in that it consists of a query generator, an off-the-shelf knowledge retriever, and a response generator, and unsupervised joint training is applied to the query generator and the response generator. However, their training objective is based on the approximated marginal likelihood over candidate queries, which is different from our objective that includes the posterior distillation based on the ELBO loss." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.495, + 0.849, + 0.543 + ], + "angle": 0, + "content": "3 Efficient Latent Variable Model for Knowledge-Grounded Dialogue Generation" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.548, + 0.602, + 0.563 + ], + "angle": 0, + "content": "3.1 Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.566, + 0.885, + 0.74 + ], + "angle": 0, + "content": "In this section, we explain an overall structure of our ELVM, as described in Figure 1a, and define notations. Unlike previous works of document (passage) retrieval (Zhang et al., 2022; Huang et al., 2021; Lewis et al., 2020) using a query of an embedded vector of a dialogue context \\( x \\), we use a text query of natural language, which enables to utilize any kind of retriever for dialogue response generation. We define the probability of generating a natural language query \\( u \\) for a given dialogue context \\( x \\), \\( p_{\\phi}(u|x) \\), as below:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.754, + 0.884, + 0.786 + ], + "angle": 0, + "content": "\\[\np _ {\\phi} (u | x) = \\prod_ {t} p _ {\\phi} \\left(u _ {t} \\mid u _ {0: t - 1}, x\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.8, + 0.885, + 0.877 + ], + "angle": 0, + "content": "where \\( u_{t} \\) is the \\( t \\)-th token in \\( u \\) and \\( p_{\\phi} \\) is the probability obtained from the model parameterized by \\( \\phi \\). Similarly, the probability of generating a response \\( y \\) for a given query \\( u \\) and dialogue context \\( x \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.89, + 0.884, + 0.922 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (y | Z (u), x) = \\prod_ {t} p _ {\\theta} \\left(y _ {t} \\mid y _ {0: t - 1}, Z (u), x\\right), \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2685" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.085, + 0.672, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.327, + 0.493, + 0.34 + ], + "angle": 0, + "content": "(a) Overall structure of ELVM." + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.083, + 0.877, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.315, + 0.878, + 0.34 + ], + "angle": 0, + "content": "(b) Example inference of ELVM." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.351, + 0.885, + 0.423 + ], + "angle": 0, + "content": "Figure 1: Overall structure and example inference case of ELVM. During training, for a given dialogue context \\( x \\), the query generator generates a set of unique queries \\( Q^s \\) and for each query \\( u \\) a set of top- \\( k \\) documents \\( Z(u) \\) is retrieved using the off-the-shelf retriever. Then, response generator utilizes both \\( x \\) and \\( Z(u) \\): (a) The approximated marginal likelihood \\( \\tilde{p}_{\\theta, \\phi}(y|x) \\) used during training is defined by \\( Q^s \\) and the re-normalized prior \\( \\bar{p}_{\\phi}(u|x) \\); (b) At inference time, a single sampled query \\( u \\) is used for retrieving multiple documents and producing a final response." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.447, + 0.489, + 0.683 + ], + "angle": 0, + "content": "where \\( y_{t} \\) is the \\( t \\)-th token in \\( y \\), \\( Z(u) = \\{z_1, z_2, \\dots, z_k\\} \\) is the set of retrieved documents (passages) by \\( u \\), and \\( \\theta \\) is the model parameters to produce the response probability. Here, we use top- \\( k \\) result obtained by an off-the-shelf retriever such as BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) and freeze the retriever. We assume that there is a deterministic mapping from \\( u \\) to \\( Z \\). We now set a query \\( u \\) as a latent variable. Then, using the prior \\( p_{\\phi}(u|x) \\) associated with the query generator and the conditional likelihood \\( p_{\\theta}(y|Z(u), x) \\) corresponding to the response generator, we can define the marginal likelihood of the knowledge-grounded response \\( y \\) given the dialogue context \\( x \\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.693, + 0.488, + 0.726 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q} p _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.735, + 0.49, + 0.814 + ], + "angle": 0, + "content": "where \\(Q\\) is the set of all possible queries. Since it is intractable to marginalize over all queries, we sample a small subset composed of \\(m\\) unique queries, \\(Q^{s}(x) = \\{u_{1}, u_{2}, \\dots, u_{m}\\}\\) using \\(p_{\\phi}(\\cdot | x)\\) and then marginalize over this set as below:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.823, + 0.488, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q ^ {s}} \\bar {p} _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.866, + 0.165, + 0.879 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.887, + 0.488, + 0.923 + ], + "angle": 0, + "content": "\\[\n\\bar {p} _ {\\phi} (u | x) = \\frac {p _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.447, + 0.885, + 0.732 + ], + "angle": 0, + "content": "We construct \\( Q^s \\) by obtaining \\( m \\)-unique queries sampled from \\( p_{\\phi}(\\cdot | x) \\). It is noted that our response generator takes multiple retrieved documents efficiently using FiD (Izacard and Grave, 2020), which is different from the use of a single document for response generation in previous methods. In addition, while we can utilize relative matching scores from the retriever to assign varying likelihoods to each document given a query for \\( p(Z(u)|u) \\), we have noticed that this approach can occasionally hinder the effective and comprehensive learning of the query generator. During the training process, that will be described in the next subsection, our objective is to amplify the impact of the generated query. Consequently, we opt for a deterministic mapping from a query to a set of retrieved documents as a unified whole, aiming to enhance the gradient flow." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.742, + 0.671, + 0.758 + ], + "angle": 0, + "content": "3.2 Joint Training" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.762, + 0.883, + 0.809 + ], + "angle": 0, + "content": "Our training objective is the ELBO (Kingma and Welling, 2013; Jordan et al., 1999) of \\(\\log \\tilde{p}_{\\theta,\\phi}\\) which can be written as below:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.821, + 0.882, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log \\tilde {p} _ {\\theta , \\phi} (y | x) \\geq \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ - D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.872, + 0.884, + 0.919 + ], + "angle": 0, + "content": "where \\(q(u|x,y)\\) is the variational posterior. Note that the equality holds when \\(q(u|x,y)\\) is identical to the true posterior distribution of it. In this work," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.146 + ], + "angle": 0, + "content": "we do not parameterize \\( q(u|x,y) \\) and approximate the expected conditional likelihood in the ELBO by sampling \\( u \\) from not \\( q(u|x,y) \\) but \\( \\bar{p}_{\\phi}(u|x) \\) such as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.161, + 0.487, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ \\approx \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ]. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.211, + 0.489, + 0.321 + ], + "angle": 0, + "content": "This approximation is due to that it minimizes the discrepancy between training and test time inference and it does not separately model \\( q(u|x,y) \\) or complicatedly sample from it during training. Then, we obtain the following two losses corresponding to the expected conditional likelihood and the KL regularization such as" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.335, + 0.488, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {y, u} (\\theta , \\phi) = - \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.356, + 0.488, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {u} (\\phi) = D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.387, + 0.487, + 0.434 + ], + "angle": 0, + "content": "Moreover, we define \\(q(u|x,y)\\) in \\(\\mathcal{L}^u (\\phi)\\) by the approximated posterior derived from the prior and the likelihood such as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.444, + 0.487, + 0.492 + ], + "angle": 0, + "content": "\\[\nq (u | x, y) = \\frac {p _ {\\theta} (y | Z (u) , x) \\bar {p} _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\theta} (y | Z (u ^ {\\prime}) , x) \\bar {p} _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.493, + 0.489, + 0.634 + ], + "angle": 0, + "content": "We set this posterior distribution \\( q(u|x,y) \\) as a teacher for a distillation to update the prior associated with the query generation at test time. Namely, \\( \\mathcal{L}^u (\\phi) \\) becomes the posterior distillation loss, and our prior is explicitly updated to be similar to the posterior. Here, the gradient is not propagated from \\( q(u|x,y) \\), and \\( q(u|x,y) \\) can be computed easily without additional model or approximation since we define it over the subset \\( Q^s \\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.636, + 0.49, + 0.92 + ], + "angle": 0, + "content": "To sum up, we obtain the conditional likelihood loss and the posterior distillation loss by multiple samples from the prior distribution that is aligned with query sampling at the inference time. We would mitigate the bias of prior sampling by multiple query samples for each input during training. In addition, our modification of ELBO is different from the expectation-maximization (EM) algorithm in that we sample a latent variable from the prior distribution rather than the posterior distribution that is used for sampling of E-step in EM. Moreover, in contrast to the M-step loss, we use the prior-weighted conditional likelihood and the gradient is also propagated to the sampled latents and the corresponding prior model. For experiments in this paper, the query generator \\(\\phi\\) and the response generator \\(\\theta\\) share parameters, with different input prompts to distinguish each module, for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.085, + 0.824, + 0.101 + ], + "angle": 0, + "content": "Overall, the loss for our joint training is" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.114, + 0.883, + 0.131 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta , \\phi) = \\mathcal {L} ^ {y, u} (\\theta , \\phi) + \\beta \\mathcal {L} ^ {u} (\\phi), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.145, + 0.882, + 0.175 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the relative weight for the posterior distillation loss." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.189, + 0.649, + 0.206 + ], + "angle": 0, + "content": "4 Experiment" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.215, + 0.714, + 0.231 + ], + "angle": 0, + "content": "4.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.236, + 0.884, + 0.41 + ], + "angle": 0, + "content": "Dataset. Our experiments are conducted on two widely used knowledge-grounded dialogue datasets, Wizard of Wikipedia (WoW) (Dinan et al., 2018) and QReCC (Anantha et al., 2021). WoW is structured around dialogues that are grounded in knowledge sentences sourced from Wikipedia, consisting of roughly 18K, 2K, and 2K dialogues, for training, validation, and test subsets, respectively. QReCC is a comprehensive compilation of conversational data, consisting of 14K open-domain conversations comprising 80K question-answer pairs." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.439, + 0.88, + 0.57 + ], + "angle": 0, + "content": "
TaskDatasetTrainValid
QueryWizInt (Komeili et al., 2022)351,3752,467
Fits (xu2)3,587392
KG ResponseWizInt (Komeili et al., 2022)22,4881,687
Fits (xu2)6,279656
Ms Marco (Nguyen et al., 2017)281,63636,859
NQ Open (Adolphs et al., 2021b)79,1688.757
Dialogue ResponseWizInt (Komeili et al., 2022)8,335587
MSC (Xu et al., 2022a)105,54917,691
SaferDialogues (Ung et al., 2022)6,3067,88
PersonaChat (Zhang et al., 2018)131,4387.801
EmpatheicDialogues (Rashkin et al., 2019)64,6365,738
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.578, + 0.884, + 0.634 + ], + "angle": 0, + "content": "Table 1: List of dataset and number of train and validation examples used for fine-tuning R2C2 model to obtain R2C2-PT. KG is an abbreviation for knowledge-grounded." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.651, + 0.884, + 0.92 + ], + "angle": 0, + "content": "Training. We begin with publicly available R2C2 pre-trained transformer model (Shuster et al., 2022a) with 400M parameter size and fine-tune it on multiple datasets to acquire the essential skill sets for knowledge-grounded dialogue such as query generation and knowledge-grounded response generation. Detailed tasks and statistics are shown in Table 1. In line with Shuster et al. (2022b), we adopt a modular system with a single model that incorporates control tokens into the dialogue context to perform different modules. For example, appending the control token __generate-query__ to the dialogue context promotes the model to generate relevant search query. We name the initial model from this pretraining procedure as R2C2-PT. Subsequently, we apply our proposed algorithm to R2C2-PT model," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2687" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.082, + 0.868, + 0.273 + ], + "angle": 0, + "content": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
PostKS (Lian et al., 2019)79.1--13.01.0193.8--13.11.0
ITDD (Li et al., 2019)17.84.02.516.2-44.82.11.111.4-
SKT (Kim et al., 2020)52.0--19.36.881.4--16.14.2
PIPM (Chen et al., 2020)42.7-3.319.97.365.7-2.517.65.4
ZRKGK (Li et al., 2021)40.4----41.5----
CoLV (Zhan et al., 2021)39.6-2.920.67.954.3-2.119.76.3
KnowledGPT (Zhao et al., 2020)19.29.57.222.07.922.38.36.020.56.7
SPI (Xu et al., 2023)17.110.27.722.78.819.19.67.322.08.5
R2C240.71.70.912.31.946.61.60.911.95.4
R2C2-PT24.76.04.417.35.230.46.24.617.75.4
ELVM-OK19.812.09.525.310.725.812.19.725.210.5
ELVM-EM-Like16.211.69.125.710.719.412.09.525.710.8
ELVM-from-R2C215.814.511.828.813.318.914.311.528.813.3
ELVM14.614.711.929.313.918.014.511.829.213.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.282, + 0.883, + 0.325 + ], + "angle": 0, + "content": "Table 2: Performance comparison in WoW. PPL represents perplexity, B3 and B4 denote BLEU-3 and BLEU-4 scores, respectively, and R1 and R2 indicate Rouge-1 and Rouge-2 scores. Our proposed model, ELVM, achieves new SOTA performance for both seen and unseen tasks with a substantial margin." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.346, + 0.486, + 0.416 + ], + "angle": 0, + "content": "
ModelSeenUnseen
R@1R@5R@10R@1R@5R@10
R2C2-PT25.763.874.329.264.876.7
ELVM-EM-Like26.463.974.728.964.575.8
ELVM-β=026.164.675.728.965.376.8
ELVM29.168.480.330.468.380.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.424, + 0.49, + 0.496 + ], + "angle": 0, + "content": "Table 3: Comparison of document recall in WoW between ELVM variants. Note that we omit the performance of ELVM-OK whose performance is identical to R2C2-PT since it is trained with keeping its query generator frozen." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.534, + 0.49, + 0.645 + ], + "angle": 0, + "content": "resulting in ELVM. Furthermore, to dispel concerns that our training gains might predominantly stem from R2C2-PT's comprehensive fine-tuning across multiple datasets rather than ELVM's inherent efficacy, we opted to utilize the base R2C2 model. Upon integrating it with ELVM, the resulting configuration is denoted as ELVM-from-R2C2." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.651, + 0.49, + 0.919 + ], + "angle": 0, + "content": "During the training of ELVM, we employ a sampling-based generation method, specifically nucleus sampling (Holtzman et al., 2020), to promote diverse query generation in \\( p_{\\phi} \\). When retrieving documents based on the generated queries, we utilize an off-the-shelf sparse retriever such as BM25 (Robertson et al., 2009), prioritizing low latency over the dense retriever (DR). We train for 3 epochs with \\( m = 4 \\) and \\( k = 5 \\) for WoW and 1 epoch with \\( m = 8 \\) and \\( k = 10 \\) for QReCC. \\( \\beta \\) is set to 1 for both tasks, unless stated otherwise. To ensure a thorough evaluation of ELVM's performance and its ability to generalize to unseen dialogues, we purposely avoid dataset overlap between training of R2C2-PT and ELVM. More detailed information on the training process can be found in Appendix A." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.346, + 0.88, + 0.426 + ], + "angle": 0, + "content": "
ModelF1 ↑EM ↑R1 ↑R2 ↑PPL ↓
Q.Rewriting (Raposo et al., 2022)18.91.0---
DPR-IHN (Kim and Kim, 2022)30.44.7---
QKConv (Cai et al., 2023)33.55.9---
R2C2-PT22.81.325.711.59.4
ELVM-OK33.04.535.621.95.6
ELVM36.56.239.025.84.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.435, + 0.885, + 0.493 + ], + "angle": 0, + "content": "Table 4: Performance comparison in QReCC with QKConv (Cai et al., 2023) which is the previous SOTA on this task. We report F1 scores, exact match (EM), Rouge score (R) and perplexity (PPL)." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.506, + 0.866, + 0.564 + ], + "angle": 0, + "content": "
ModelFluency (%)Relevance (%)
SeenUnseenSeenUnseen
KnowledGPT69.862.941.541.1
ELVM73.577.5*49.563.6**
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.573, + 0.884, + 0.645 + ], + "angle": 0, + "content": "Table 5: Human evaluation results on WoW test. A pairwise t-test is conducted to verify statistical significance of the improvements, and the corresponding results in bold are significantly better than those from the baseline model \\((^{**}: p < 0.01, ^*: p < 0.025)\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.671, + 0.884, + 0.828 + ], + "angle": 0, + "content": "Inference. During inference, unlike the training phase, only a single query is generated by the query generator \\( p_{\\phi} \\). The default number of retrieved documents \\( k \\) from the off-the-shelf retriever is set to 5 for WoW and 10 for QReCC. The knowledge-grounded response is generated by the response generator, \\( p_{\\theta} \\). An illustration of the inference is depicted in Figure 1b. During the inference, both query and response generation are conducted using beam search with a beam size of 5." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.841, + 0.885, + 0.921 + ], + "angle": 0, + "content": "Variants of ELVM. To examine the impact of the proposed ELVM training, we explore another variant, ELVM-OK, which is trained on the response generation task with annotated oracle knowledge documents while keeping the query generator \\( p_{\\phi} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2688" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.183, + 0.082, + 0.421, + 0.166 + ], + "angle": 0, + "content": "
mSeenUnseen
B4↑R2↑B4↑R2↑
110.812.410.512.2
211.113.011.112.8
4*11.913.911.813.7
811.813.711.713.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.176, + 0.49, + 0.232 + ], + "angle": 0, + "content": "Table 6: Effect of varying the number of queries, \\( m \\) , sampled during training on WoW. We see the highest performance on unseen task when \\( m = 4 \\) . The superscripted value by * is the default setting for ELVM." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.246, + 0.486, + 0.318 + ], + "angle": 0, + "content": "
βWow UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
09.310.665.334.55.155.8
1*11.813.768.336.56.965.6
510.812.466.731.33.531.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.327, + 0.489, + 0.397 + ], + "angle": 0, + "content": "Table 7: Ablation study of the posterior distillation weight \\(\\beta\\) on WoW. From the results, we can see that the trivial value of \\(\\beta = 1\\) shows the best performance. The super-scripted value by \\(*\\) is the default setting for ELVM." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.425, + 0.489, + 0.503 + ], + "angle": 0, + "content": "frozen. During inference, given a dialogue context the frozen query generator \\( p_{\\phi} \\) generates query and use the off-the-shelf retriever to bring relevant documents. Finally the response generator of ELVM-OK generates the final response." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.505, + 0.49, + 0.678 + ], + "angle": 0, + "content": "In Section 3.2, we describe the difference between ELVM training and EM algorithm. In order to quantitatively compare these two methods, we perform EM-like training in ELVM where we apply the posterior-weighted conditional likelihood in \\(\\mathcal{L}^{y,u}(\\theta ,\\phi)\\) .. \\(\\mathcal{L}^{y,u}(\\theta ,\\phi) =\\) \\(-\\sum_{u}SG(q(u|x,y))\\log p_{\\theta}(y|Z(u),x).\\) Here, the query latents \\(u\\) are still sampled from the prior distribution, and \\(SG\\) means the stop-gradient that prevents the gradient propagation to the query generator. We name this variant as ELVM-EM-Like." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.692, + 0.334, + 0.706 + ], + "angle": 0, + "content": "4.2 Automatic Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.714, + 0.49, + 0.92 + ], + "angle": 0, + "content": "The evaluation results of WoW test seen and unseen tasks are shown in Table 2. ELVM surpasses all its variants and previous state-of-the-art (SOTA) model, achieving a new SOTA on both WoW test seen unseen tasks, across all metrics. In addition, ELVM-OK demonstrates notable improvements in BLEU and Rouge scores compared to previous models with a slight drawback in PPL. Similarly, ELVM-EM-Like also exhibits competitive performance across multiple metrics. Moreover, when comparing ELVM-from-R2C2 and ELVM, the experiment demonstrates that while R2C2-PT offers marginal benefits, the marked performance" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.082, + 0.875, + 0.133 + ], + "angle": 0, + "content": "
ModelWoW UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
ELVM11.813.768.336.56.259.9
ELVMDR12.815.581.237.96.965.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.142, + 0.883, + 0.185 + ], + "angle": 0, + "content": "Table 8: Performance results achieved by employing the dense retriever instead of the default BM25 retriever for WoW unseen and QReCC evaluation." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.197, + 0.88, + 0.291 + ], + "angle": 0, + "content": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑B4 ↑R2 ↑R@5 ↑
R2C2-PT4.45.263.84.65.464.8
ELVM-OK9.510.7-9.710.5-
ELVM11.913.968.411.813.768.3
R2C2-PTLarge8.710.262.98.910.364.4
ELVM-OKLarge9.610.6-10.211.2-
ELVMLarge12.614.372.612.414.174.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.299, + 0.884, + 0.356 + ], + "angle": 0, + "content": "Table 9: Effect of model scaling tested on WoW. ELVM scales out to larger parameter size with increase in performance. The subscript Large indicates the model size of 2.7B parameters." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.381, + 0.869, + 0.396 + ], + "angle": 0, + "content": "increase is predominantly attributable to ELVM." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.398, + 0.884, + 0.554 + ], + "angle": 0, + "content": "We further examine the performance of query generator by assessing the recall of the ground-truth documents. The results summarized in Table 3 reveal that training with our proposed algorithm leads to an increase in recall for both ELVM- \\(\\beta = 0\\) and ELVM compared to R2C2-PT while the latter yields a larger performance gain. This highlights the effectiveness of distilling the posterior distribution of the response generator into the prior distribution associated with the query generator." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.555, + 0.884, + 0.664 + ], + "angle": 0, + "content": "For QReCC, as shown in Table 4, ELVM surpasses the previous SOTA, QKConv (Cai et al., 2023), the recent previous works (Raposo et al., 2022; Kim and Kim, 2022), and the variants of ELVM including R2C2-PT and ELVM-OK in terms of all metrics. The recall performance on QReCC is included in Table C." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.675, + 0.707, + 0.689 + ], + "angle": 0, + "content": "4.3 Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.696, + 0.884, + 0.885 + ], + "angle": 0, + "content": "To gauge performance across multiple aspects of the quality of generated responses, human evaluation is performed on the generated responses of ELVM and KnoawedGPT (Zhao et al., 2020)\\(^{1}\\). Similar to (Rashkin et al., 2021) and (Xu et al., 2023), we assess the response quality in two aspects: Fluency and Relevance. Fluency measures whether the response is understandable, self-consistent without repetition, and proficient. Relevance assesses the extent to which the response aligns with the dialogue context, incorporates pertinent knowledge, and maintains appropriateness. In total, 50 data" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.883, + 0.919 + ], + "angle": 0, + "content": "1We choose KnowledgeGPT because it is the best performing model among publicly available models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2689" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.082, + 0.481, + 0.171 + ], + "angle": 0, + "content": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
CoLV39.62.97.954.32.16.3
KnowledGPT19.27.27.922.36.06.7
SPI17.17.78.819.17.38.5
BART9741.92.69731.32.6
ELVM-from-BART15.38.411.116.48.210.9
ELVM14.611.913.918.011.813.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.179, + 0.489, + 0.207 + ], + "angle": 0, + "content": "Table 10: Performance comparison with BART as our initial model." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.222, + 0.486, + 0.323 + ], + "angle": 0, + "content": "
ModelTrain RetTest RetWoW Unseen
PPL ↓B4 ↑R2 ↑F1 ↑
CoLV--54.32.16.318.5
KnowledgeGPT--22.36.06.720.5
SPI--19.17.38.5-
ELVMBM25BM2518.011.813.727.0
ELVMBM25Dense14.812.615.229.3
ELVMDenseDense12.912.815.529.0
ELVMDenseBM2521.611.813.128.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.332, + 0.49, + 0.402 + ], + "angle": 0, + "content": "Table 11: Impact of alternating the retriever between training and testing phases on WoW unseen test. Dense indicates utilizing all-MiniLM-L6-v2 retriever as described in Section 4 and Ret is an abbreviation for retriever." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.429, + 0.49, + 0.571 + ], + "angle": 0, + "content": "samples are randomly selected from WoW tasks where 25 samples are randomly selected from each seen and unseen task. The qualities of the responses are measured by A/B testing on the two aspects. The feedback is collected from 11 human experts. Further details and annotator instructions can be found in Table G. As shown in Table 5, ELVM significantly outperforms KnowledGPT in all aspects especially on the unseen task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.585, + 0.293, + 0.6 + ], + "angle": 0, + "content": "4.4 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.607, + 0.49, + 0.78 + ], + "angle": 0, + "content": "Number of Queries. The size of sampled query set \\( Q^{s} \\), \\( m \\), plays a huge role in training of ELVM, as it directly influences the bias of the marginal likelihood of the knowledge-grounded response \\( y \\). To investigate the impact of \\( m \\), we gradually increase \\( m \\) from 1 to 8 and measure the performance on WoW while keeping other hyperparameters fixed. As shown in Table 6, we observe a positive correlation where increasing \\( m \\) leads to improved performance. Notably, setting \\( m = 4 \\) yields the optimal performance for both seen and unseen tasks." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.793, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Posterior Distillation Weight. We examine the influence of the posterior distillation loss on ELVM training by adjusting the value of \\(\\beta\\). The experimental results presented in Table 7 reveal that the optimal performance is attained when \\(\\beta = 1\\) for both the WoW unseen and QReCC tasks, indicating the effectiveness of incorporating the posterior distillation loss. However, a substantial decline in" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.117 + ], + "angle": 0, + "content": "performance is observed when \\(\\beta\\) is increased to 5 or set to 0, particularly in the case of QReCC." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.128, + 0.885, + 0.349 + ], + "angle": 0, + "content": "Varying the Retriever. We extend our experiments by incorporating the DR in two distinct ways. For WoW, we train and evaluate using all MiniLM-L6-v2\\(^2\\) DR. In contrast, for QReCC, due to its extensive knowledge pool, we leverage DR solely during evaluation by employing both the BM25 retriever and all-mpnet-base-v2\\(^3\\) DR in a hybrid manner to retrieve the top-\\(k\\) documents. Concretely, we score each document by summing up the normalized score of each retriever and rerank to obtain top-\\(k\\) documents. Table 8 presents a summary of the performance results of employing DR, demonstrating a successful integration of DR in different settings of ELVM." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.361, + 0.885, + 0.534 + ], + "angle": 0, + "content": "Scaling Up Model. We investigate the impact of increasing the model size on performance. Starting with the publicly available 2.7B R2C2 pretrained transformer, we follow exactly the same procedure outlined in Section 3.2 to obtain R2C2-PTLarge, ELVM-OKLarge and ELVMLarge. Detailed hyperparameters for training 2.7B model is given in Appendix A. The results presented in Table 9 demonstrate that as the model size increases, there is an improvement in performance across all metrics, supporting the scalability of ELVM." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.546, + 0.884, + 0.799 + ], + "angle": 0, + "content": "BART as Base. To ensure a fair comparison and further demonstrate the robustness of our algorithm on diverse model scales, we conduct an experiment employing BART (Lewis et al., 2019) base as an initial model and then apply our proposed algorithm on WoW dataset. We choose BART base model since the previous SOTA model (Xu et al., 2023) utilized BART base as their base model. The outcome of this training is termed ELVM-from-BART. As shown in Table 10, the initial BART model shows very poor performances. However, upon training the BART-base with our ELVM algorithm, there is a great improvement in performance across both WoW Seen and Unseen tasks. Importantly, the results surpass the performance benchmarks set by the previous SOTA model (SPI)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.81, + 0.884, + 0.858 + ], + "angle": 0, + "content": "Retriever Switching To assess the robustness of our algorithm to the retriever mismatch between training and inference phases, we undertake an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.868, + 0.884, + 0.918 + ], + "angle": 0, + "content": "\\(^{2}\\)https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 \n\\(^{3}\\)https://huggingface.co/sentence-transformers/all-mpnet-base-v2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2690" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.081, + 0.885, + 0.398 + ], + "angle": 0, + "content": "
Dialogue ContextI want to teach my family about investing by buying them a few shares of stock, but I really don't know much about it. The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses. Oh, so I could tell my kids they own part of a company. That's cool. Maybe I should get them a share in a company they know then, like Apple or XBox or Snapchat? [...] Examples of the latter include shares of private companies which are sold to investors through equity crowdfunding platforms. [...]
GT knowledge
GT ResponseYes, it's like the shares of a company are sold to investors through equity crowdfunding platforms.
KnownGPTQuery: n/aGT knowledge selected: False
Response: yes, you can get shares in a company like apple or xbox. you can also get shares in company like stock market.
R2C2-PTQuery: How to teach kids about investingGT knowledge selected: False
Response: The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses.
ELVM-β = 0Query: Investing tips for beginnersGT knowledge selected: False
Response: That's a good idea. The stock market is a place where you can buy and sell stocks.
ELVM w/ BM25Query: Apple xbox snapshot stock share buy apple stock instagram stock investingGT knowledge selected: True
Response: That is a good idea. You can also buy shares of private companies which are sold to investors through equity crowdfunding platforms.
ELVM w/ DRQuery: Best stocks for teaching kids about investingGT knowledge selected: True
Response: That is a good idea. The latter include shares of private companies which are sold to investors through equity crowdfunding platforms.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.405, + 0.884, + 0.476 + ], + "angle": 0, + "content": "Table 12: Generated samples from WoW unseen task. The first row represents the dialogue context, ground truth (GT) knowledge (highlighted in olive), and GT response. Following rows contain the model generated query, response and whether the GT knowledge is included in top-5 retrieved documents. More samples can be found in Appendix H. We truncate the GT knowledge and display only relevant spans of the GT knowledge to the ground-truth response." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.675 + ], + "angle": 0, + "content": "additional experiment, of which the results are depicted in Table 11. In this experiment, we train the WoW task utilizing either BM25 or dense retriever (utilizing all-MiniLM-L6-v2 retriever, as detailed Section 4), and subsequently evaluate using either BM25 or dense retriever. The results on the WoW unseen test indicate that utilizing a different retriever during testing doesn't necessitate model re-training, particularly given the superior performance of the dense retriever compared to BM25." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.701, + 0.32, + 0.717 + ], + "angle": 0, + "content": "4.5 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.73, + 0.49, + 0.919 + ], + "angle": 0, + "content": "For qualitative analysis, we examine the generated samples from KnownGPT, R2C2-PT and other variants of ELVM, whose samples are presented in Table 12. The example reveals that the queries generated by KnownGPT, R2C2-PT and ELVM-\\(\\beta = 0\\) fail to retrieve the relevant knowledge effectively while the queries generated by ELVM with BM25 and DR (all-MiniLM-L6-v2) reflects the ability to access the relevant knowledge, leading to improved responses. The results also demonstrate that the query generator trained with our algorithm effectively adapts to the chosen retriever." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.5, + 0.642, + 0.516 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.529, + 0.885, + 0.782 + ], + "angle": 0, + "content": "In this paper, we propose ELVM as an efficient latent variable modeling for knowledge-intensive dialogue generation. In ELVM, the knowledge retrieval is realized by the query generation followed by document retrieval using the off-the-shelf retriever for ease of joint training of both the retriever and the response generator, without the knowledge supervision. Furthermore, the training scheme of ELVM modifies the ELBO for the marginal likelihood to effectively perform the joint training without the use of complex posterior sampling, which also eliminates the discrepancy between the training-time and test-time knowledge samplings. ELVM empirically demonstrates that it significant outperforms over previous latent variable methods with in-depth analysis on diverse dialogue datasets." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.797, + 0.615, + 0.812 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.825, + 0.884, + 0.919 + ], + "angle": 0, + "content": "There are a number of limitations and possible directions to future works of ELVM proposed in this paper. First, as more large language models beyond the scale of hundreds of billion parameters are emerging, there are rooms for emergent behaviors in ELVM as the sizes of data and model grow that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "2691" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.491, + 0.384 + ], + "angle": 0, + "content": "has not been tested in this paper. Second, instead of the prior sampling that we use, fast posterior sampling such as short-run MCMC (Erik et al., 2019; Xu et al., 2023) is also an alternative candidate. Third, the experiments in this paper retrieve documents from confined set (e.g., Wikipedia) using light retriever such as BM25 and dense retriever. However, real-world applications often rely on the web which has much bigger pool of documents from broader open-domain. In turn, we should consider future extension of this work to utilizing larger public search engines. Last, we may consider alternative yet more effective means of query sampling during training other than a small set of random samples, as the current design. Other possible methods include imposing additional guidance for query sampling or multiple refinements of query sampling and knowledge retrieval for more diverse set." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.399, + 0.264, + 0.412 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.425, + 0.488, + 0.517 + ], + "angle": 0, + "content": "This work presents no direct ethical issues. However, there is a possibility that the generated responses may inadvertently contain inadequate, biased, or discriminatory content. We do not introduce new datasets, and all experiments are conducted using publicly available datasets." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.531, + 0.262, + 0.543 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.549, + 0.488, + 0.737 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University). In addition, this work was also supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea) Gwangju Metropolitan City. Finally, we would also like to thank Brain Cloud Team at Kakao Brain for their support." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.767, + 0.212, + 0.78 + ], + "angle": 0, + "content": "References" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.84, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Leonard Adolphs, Benjamin Boerschinger, Christian Buck, Michelle Chen Huebscher, Massimiliano Ciaramita, Lasse Espeholt, Thomas Hofmann, Yannic Kilcher, Sascha Rothe, Pier Giuseppe Sessa, et al. 2021a. Boosting search engines with interactive agents. arXiv preprint arXiv:2109.00527." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.087, + 0.881, + 0.139 + ], + "angle": 0, + "content": "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021b. Reason first, then respond: Modular generation for knowledge-infused dialogue. arXiv preprint arXiv:2111.05204." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.149, + 0.883, + 0.254 + ], + "angle": 0, + "content": "Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.883, + 0.42 + ], + "angle": 0, + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.431, + 0.883, + 0.483 + ], + "angle": 0, + "content": "Mingzhu Cai, Siqi Bao, Xin Tian, Huang He, Fan Wang, and Hua Wu. 2023. Query enhanced knowledge-intensive conversation via unsupervised joint modeling. In ACL." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.494, + 0.883, + 0.559 + ], + "angle": 0, + "content": "Xiuyi Chen, Fandong Meng, Peng Li, Feilong Chen, Shuang Xu, Bo Xu, and Jie Zhou. 2020. Bridging the gap between prior and posterior knowledge selection for knowledge-grounded dialogue generation. In EMNLP, pages 3426-3437." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.883, + 0.621 + ], + "angle": 0, + "content": "Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Oleg Rokhlenko, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering. arXiv preprint arXiv:2210.15777." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "2692" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.113 + ], + "angle": 0, + "content": "language modeling with pathways. arXiv preprint arXiv:2204.02311." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.126, + 0.487, + 0.179 + ], + "angle": 0, + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.192, + 0.487, + 0.258 + ], + "angle": 0, + "content": "Nijkamp Erik, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. 2019. Learning non-convergent non-persistent short-run mcmc toward energy-based model. In Advances in Neural Information Processing Systems (NeurIPS)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.271, + 0.487, + 0.324 + ], + "angle": 0, + "content": "Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.337, + 0.487, + 0.39 + ], + "angle": 0, + "content": "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.403, + 0.487, + 0.443 + ], + "angle": 0, + "content": "Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.456, + 0.487, + 0.548 + ], + "angle": 0, + "content": "Xinxian Huang, Huang He, Siqi Bao, Fan Wang, Hua Wu, and Haifeng Wang. 2021. PLATO-KAG: Unsupervised knowledge-grounded conversation via joint modeling. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, pages 143-154, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.562, + 0.487, + 0.613 + ], + "angle": 0, + "content": "Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.627, + 0.487, + 0.68 + ], + "angle": 0, + "content": "Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. 1999. An introduction to variational methods for graphical models. Machine learning, 37:183-233." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.693, + 0.487, + 0.758 + ], + "angle": 0, + "content": "Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.772, + 0.487, + 0.813 + ], + "angle": 0, + "content": "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.825, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.878, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Diederik P Kingma and Max Welling. 2013. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.166 + ], + "angle": 0, + "content": "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8460-8478, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.175, + 0.883, + 0.293 + ], + "angle": 0, + "content": "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.304, + 0.883, + 0.37 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.38, + 0.883, + 0.46 + ], + "angle": 0, + "content": "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.47, + 0.883, + 0.511 + ], + "angle": 0, + "content": "Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2021. Zero-resource knowledge-grounded dialogue generation." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.521, + 0.883, + 0.613 + ], + "angle": 0, + "content": "Zekang Li, Cheng Niu, Fandong Meng, Yang Feng, Qian Li, and Jie Zhou. 2019. Incremental transformer with deliberation decoder for document grounded conversations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 12-21, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.623, + 0.883, + 0.676 + ], + "angle": 0, + "content": "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. arXiv preprint arXiv:1902.04911." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.686, + 0.883, + 0.753 + ], + "angle": 0, + "content": "Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.763, + 0.883, + 0.843 + ], + "angle": 0, + "content": "Shilei Liu, Xiaofeng Zhao, Bochao Li, Feiliang Ren, Longhui Zhang, and Shujuan Yin. 2021. A Three-Stage Learning Framework for Low-Resource Knowledge-Grounded Dialogue Generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Shuman Liu, Hongshen Chen, Zhaochun Ren, Yang Feng, Qun Liu, and Dawei Yin. 2018. Knowledge diffusion for neural dialogue generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2693" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.113 + ], + "angle": 0, + "content": "pages 1489-1498, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.123, + 0.487, + 0.175 + ], + "angle": 0, + "content": "Longxuan Ma, Weinan Zhang, Runxin Sun, and Ting Liu. 2020. A compare aggregate transformer for understanding document-grounded dialogue. arXiv preprint arXiv:2010.00190." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.185, + 0.487, + 0.238 + ], + "angle": 0, + "content": "Nikita Moghe, Siddhartha Arora, Suman Banerjee, and Mitesh M Khapra. 2018. Towards exploiting background knowledge for building conversation systems. arXiv preprint arXiv:1809.08205." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.248, + 0.487, + 0.301 + ], + "angle": 0, + "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2017. MS MARCO: A human-generated MAchine reading COprehension dataset." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.311, + 0.487, + 0.35 + ], + "angle": 0, + "content": "Rodrigo Nogueira and Kyunghyun Cho. 2017. Task-oriented query reformulation with reinforcement learning. arXiv preprint arXiv:1704.04572." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.36, + 0.487, + 0.412 + ], + "angle": 0, + "content": "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.422, + 0.487, + 0.488 + ], + "angle": 0, + "content": "Gonçalo Raposo, Rui Ribeiro, Bruno Martins, and Luísca Coheur. 2022. Question rewriting? assessing its importance for conversational question answering. In Lecture Notes in Computer Science, pages 199-206. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.498, + 0.487, + 0.616 + ], + "angle": 0, + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.626, + 0.487, + 0.718 + ], + "angle": 0, + "content": "Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.728, + 0.487, + 0.78 + ], + "angle": 0, + "content": "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.79, + 0.487, + 0.868 + ], + "angle": 0, + "content": "Kurt Shuster, Mojtaba Komeili, Leonard Adolphs, Stephen Roller, Arthur Szlam, and Jason Weston. 2022a. Language models that seek for knowledge: Modular search & generation for dialogue and prompt completion. arXiv preprint arXiv:2203.13224." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.878, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya Chen, Kushal Arora, Joshua Lane, et al. 2022b." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.881, + 0.126 + ], + "angle": 0, + "content": "Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage. arXiv preprint arXiv:2208.03188." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.138, + 0.883, + 0.217 + ], + "angle": 0, + "content": "Weiwei Sun, Pengjie Ren, and Zhaochun Ren. 2023. Generative knowledge selection for knowledge-grounded dialogues. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2077-2088, Dubrovnik, Croatia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.228, + 0.883, + 0.502 + ], + "angle": 0, + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, abs/2201.08239." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.515, + 0.883, + 0.568 + ], + "angle": 0, + "content": "David Thulke, Nico Daheim, Christian Dugast, and Hermann Ney. 2021. Efficient retrieval augmented generation from unstructured knowledge for task-oriented dialog. arXiv preprint arXiv:2102.04643." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.58, + 0.883, + 0.672 + ], + "angle": 0, + "content": "Megan Ung, Jing Xu, and Y-Lan Boureau. 2022. SaFeR-Dialogues: Taking feedback gracefully after conversational safety failures. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6462-6481, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.684, + 0.883, + 0.75 + ], + "angle": 0, + "content": "Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021. Question rewriting for conversational question answering. In Proceedings of the 14th ACM international conference on web search and data mining, pages 355-363." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.762, + 0.883, + 0.841 + ], + "angle": 0, + "content": "Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 921-930." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, Hannaneh Hajishirzi, Mari Ostendorf, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning. In EMNLP." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2694" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.166 + ], + "angle": 0, + "content": "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.174, + 0.489, + 0.255 + ], + "angle": 0, + "content": "Jing Xu, Arthur Szlam, and Jason Weston. 2022a. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.263, + 0.488, + 0.33 + ], + "angle": 0, + "content": "Jing Xu, Megan Ung, Mojtaba Komeili, Kushal Arora, Y-Lan Boureau, and Jason Weston. 2022b. Learning new skills after deployment: Improving open-domain internet-driven dialogue with human feedback. arXiv preprint arXiv:2208.03270." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.339, + 0.488, + 0.392 + ], + "angle": 0, + "content": "Yan Xu, Deqian Kong, Dehong Xu, Ziwei Ji, Bo Pang, Pascale Fung, and Ying Nian Wu. 2023. Diverse and faithful knowledge-grounded dialogue generation via sequential posterior inference." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.401, + 0.489, + 0.48 + ], + "angle": 0, + "content": "Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 1933-1936." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.49, + 0.488, + 0.542 + ], + "angle": 0, + "content": "Haolan Zhan, Lei Shen, Hongshen Chen, and Hainan Zhang. 2021. Colv: A collaborative latent variable model for knowledge-grounded dialogue generation. In EMNLP." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.552, + 0.488, + 0.644 + ], + "angle": 0, + "content": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.654, + 0.487, + 0.707 + ], + "angle": 0, + "content": "Yizhe Zhang, Siqi Sun, Xiang Gao, Yuwei Fang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2022. Retgen: A joint framework for retrieval and grounded text generation modeling. In AAAI." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.716, + 0.488, + 0.769 + ], + "angle": 0, + "content": "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.779, + 0.488, + 0.858 + ], + "angle": 0, + "content": "Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018a. Common-sense knowledge aware conversation generation with graph attention. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 4623-4629. AAAI Press." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.867, + 0.488, + 0.907 + ], + "angle": 0, + "content": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W Black. 2018b. A dataset for document grounded conversations. In EMNLP." + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.084, + 0.689, + 0.101 + ], + "angle": 0, + "content": "A Training Details" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.12, + 0.855, + 0.259 + ], + "angle": 0, + "content": "
ParameterValue
Training Steps10,000
Batch Size128
OptimizerAdamW
Learning Rate (LR)5e-5
LR SchedulerCosine
Adam β10.9
Adam β20.999
Adam ε1e-8
Warmup Steps1000
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.268, + 0.882, + 0.296 + ], + "angle": 0, + "content": "Table 13: Hyperparameters used for stage-1 training. \\( k \\) indicates number of retrieved documents." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.339, + 0.848, + 0.545 + ], + "angle": 0, + "content": "
WoWQReCC
Train Epochs31
Batch Size6464
OptimizerAdamWAdamW
Learning Rate (LR)1e-45e-4
LR SchedulerCosineCosine
Adam β10.90.9
Adam β20.9990.999
Adam ε1e-81e-8
Warmup Steps1000
Min Query Length11
Max Query Length3232
Top-k100100
Top-p0.950.5
Temperature1.02.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.554, + 0.882, + 0.582 + ], + "angle": 0, + "content": "Table 14: Hyperparameters used for training ELVM and its." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.599, + 0.884, + 0.82 + ], + "angle": 0, + "content": "Table 13 we provide hyperparameters utilized during this training process. We also report hyperparameters used for training ELVM and its variants model (ELVM-OK and ELVM-EM-Like) in Table 14. In addition, during training of QReCC, we adopt a strategy to decrease the knowledge pool size from 54M to 1M. This reduction aims to mitigate latency issues during document retrieval. Concretely, the 1M pool is constructed by initially gathering the ground-truth relevant passage for each instance in the training dataset then adding randomly sampled documents. During the evaluation stage, we utilize the original full-size knowledge pool of 54M passages." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.831, + 0.782, + 0.848 + ], + "angle": 0, + "content": "B Upper Bound Performance" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.882, + 0.919 + ], + "angle": 0, + "content": "We believe in the value of understanding the utmost capability in an ideal scenario. To this end, we pursue an upper-bound performance analysis for both the WoW and QReCC tasks. This approach" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2695" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.177, + 0.082, + 0.821, + 0.132 + ], + "angle": 0, + "content": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
ELVM14.614.711.929.313.918.014.511.829.213.7
ELVM-w-GT-Doc9.722.118.540.923.910.521.517.940.723.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.142, + 0.884, + 0.171 + ], + "angle": 0, + "content": "Table 15: Performance comparison in WoW in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.193, + 0.481, + 0.287 + ], + "angle": 0, + "content": "
ModelF1 ↑EM ↑
Q Rewriting (Raposo et al., 2022)18.91.0
DPR-IHN (Kim and Kim, 2022)30.44.7
QKConv (Cai et al., 2023)33.55.9
ELVM36.56.2
ELVM-w-GT-Doc55.416.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.297, + 0.487, + 0.339 + ], + "angle": 0, + "content": "Table 16: Performance comparison in QReCC in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.369, + 0.487, + 0.493 + ], + "angle": 0, + "content": "involves training our response generator under the assumption that the model has access to the ground-truth document during both training and evaluation. The performance metrics of this specific setup are designated as \"ELVM-w-GT-Doc\" in Table 15 and Table 16. From both tables, it's evident that there still remains a margin for improvement for both tasks." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.514, + 0.416, + 0.529 + ], + "angle": 0, + "content": "C Document Recall Performance" + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.555, + 0.446, + 0.653 + ], + "angle": 0, + "content": "
Query fromR@1R@5R@10
Dialogue History33.458.668.7
Last Utterance25.550.862.6
R2C2-PT17.433.039.5
ELVM-β=030.655.865.7
ELVM35.259.968.6
ELVMDR37.565.675.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.663, + 0.486, + 0.692 + ], + "angle": 0, + "content": "Table 17: Query performance comparison in QReCC for different types of queries." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.714, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We report the document recall performance on QReCC, including different types of queries, as show in Table 17. Our proposed model, ELVM, surpasses both its variants and non-model generated queries such as using dialogue context as a query. The document recall performance on QReCC, encompassing various query types, is presented in Table 17. Our proposed ELVM model outperforms its variants as well as non-model generated queries, including the use of dialogue context as a query. Moreover, incorporating the dense retriever for reranking documents \\((\\mathrm{ELVM}_{\\mathrm{DR}})\\) leads to further improvements in recall." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.202, + 0.889, + 0.278 + ], + "angle": 0, + "content": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑R@10 ↑B4 ↑R2 ↑R@5 ↑R@10 ↑
k=511.913.968.480.311.813.768.380.3
k=5Bulk9.410.962.773.98.29.547.261.9
k=1010.312.264.876.410.612.764.876.9
k=10Bulk10.212.262.174.19.010.547.261.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.287, + 0.882, + 0.342 + ], + "angle": 0, + "content": "Table 18: Effect of increasing the document pool size to 1000 on the performance of ELVM. As before, \\( k \\) indicates the number of retrieved documents used during train and inference." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.368, + 0.757, + 0.385 + ], + "angle": 0, + "content": "D Knowledge Pool Scaling" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.393, + 0.884, + 0.597 + ], + "angle": 0, + "content": "In order to create a more realistic and challenging retrieval scenario, we expand the original setting of the WoW dataset by increasing the number of relevant documents per instance to 1000. This modification better reflects real-world information retrieval scenarios where a vast array of both relevant and irrelevant documents are typically encountered. We construct this extended dataset by randomly augmenting \\(1000 - K\\) documents from the WoW dataset to the document pool for each instance, where \\(K\\) represents the number of annotated relevant documents for each instance. On average, \\(K \\approx 20\\) for each instance in WoW dataset." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.599, + 0.884, + 0.787 + ], + "angle": 0, + "content": "Results in Table 18 show performances of ELVM models with different number of retrieved documents, \\( k \\), and knowledge pool size, where models with increased knowledge pool has tailing keyword Bulk at then end of their names. Despite the increase in complexity of the retrieval task, ELVM demonstrates robust performance, with only a marginal drop in metrics compared to the original setting. Furthermore, it is important to highlight that even under these more difficult conditions with 1000 relevant documents, ELVM still outperforms SPI by a significant margin." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.8, + 0.791, + 0.816 + ], + "angle": 0, + "content": "E Effect of Parameter Sharing" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.825, + 0.884, + 0.918 + ], + "angle": 0, + "content": "In our ELVM training, parameters are shared between the query and response generators. To probe this design choice, we decoupled these parameters, resulting in the ELVM-Decouple configuration, as detailed in Table 19. Upon evaluation on WoW's test seen and unseen tasks, ELVM-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2696" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.082, + 0.48, + 0.129 + ], + "angle": 0, + "content": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
ELVM14.611.913.918.011.813.7
ELVM-Decouple16.212.614.521.612.214.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.138, + 0.489, + 0.21 + ], + "angle": 0, + "content": "Table 19: Performance comparison on the WoW dataset using the ELVM algorithm. ELVM denotes training with shared parameters between query and response generators, while ELVM-Decouple denotes a non-shared parameter approach." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.221, + 0.48, + 0.351 + ], + "angle": 0, + "content": "
ModelSeenUnseen
F1 ↑KF1 ↑F1 ↑KF1 ↑
CoLV20.318.218.517.5
KnownGPT22.023.820.522.1
SPI----
R2C2-PT15.420.515.722.3
ELVM-OK23.033.022.933.0
ELVM-EM-Like24.533.223.533.5
ELVM27.240.827.040.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.362, + 0.486, + 0.389 + ], + "angle": 0, + "content": "Table 20: F1 and KF1 scores for ELVM and baseline models." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.416, + 0.49, + 0.493 + ], + "angle": 0, + "content": "Decouple showed a slight edge over the standard ELVM. Nevertheless, the benefits of shared parameters—particularly regarding computational efficiency and knowledge transfer during training and inference—cannot be understated." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.506, + 0.317, + 0.521 + ], + "angle": 0, + "content": "F Additional Metrics" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.531, + 0.49, + 0.594 + ], + "angle": 0, + "content": "We report the evaluation of our models on the WoW task using both F1 and knowledge F1 (KF1) metrics in Table 20. For reference, values pertaining to previous models are derived from Sun et al. (2023)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.606, + 0.32, + 0.62 + ], + "angle": 0, + "content": "G Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.631, + 0.49, + 0.82 + ], + "angle": 0, + "content": "For a comprehensive evaluation, we conduct the human evaluation between ELVM and KnownGPT on seen and unseen tasks in WoW (Dinan et al., 2018). We randomly select 50 samples (25 samples per task), and each sample is evaluated by 11 different human experts. In specific, the two generated responses from each model in the same context are assigned to the annotators. For A/B testing, we give one score to the model if its response is received an equally good or better than the other one. Figure 2 shows the annotator instructions for the two aspects." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.832, + 0.325, + 0.848 + ], + "angle": 0, + "content": "H Generated Samples" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.856, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Generated samples for the WoW test unseen are presented in Table 21 and Table 22. Similarly, Table 23 and Table 24 showcase the generated samples for the QReCC test." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2697" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.149, + 0.885, + 0.81 + ], + "angle": 0, + "content": "
Dialogue ContextHey, do you like cardigans? I like the modern version without no buttons and hangs open by design. Me too. My fav is the waterfall style but it's going out of fashion now I think [...] Cardigan (sweater) By contrast, a pullover does not open in front but must be "pulled over" the head to be worn. [...] But the pullover does not open in front but must be "pulled over" the head to be worn.
GT knowledge
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: i like the hand-knitted version. it's more comfortable. i like mine made of a cardigan.
R2C2-PTQuery: fallen fashion GT knowledge selected: False Response: I'm sorry, I didn't mean to offend you.
ELVM-β = 0Query: fallen fashion cardigans GT knowledge selected: True Response: I like the one that is tied is instead considered a robe.
ELVM w/ BM25Query: fallen waterfall cardigan style fallout cardigans GT knowledge selected: True Response: I like the pullover style that does not open in front but must be "pulled over" the head to be worn.
ELVM w/ DRQuery: fallen waterfall cardigan GT knowledge selected: True Response: I like the pullover that does not open in front but must be "pulled over" the head to be worn.
Dialogue Context[...] Was she diagnosed as a child? It is much easier to diagnose children, since the symptoms begin occurring early in life
GT knowledgeYes it started when she was 6 years old. She was 23 when I met her she was already recieving treatment [...] These fellowships include pediatric neurosurgery, trauma/neurocritical care, functional and stereotactic surgery, surgical neuro-oncology, radiosurgery, neurovascular surgery, skull-base surgery, peripheral nerve and spine surgery. [...] My first fellowship was in pediatric neurosurgery. It is great that she is doing better.
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: did she cry during the periods of depression? she was crying, a negative outlook on life, and poor eye contact with others.
R2C2-PTQuery: types of epilepsy GT knowledge selected: False Response: I am glad to hear she is receiving treatment. Dementia is such a devastating disease.
ELVM-β = 0Query: types of epilepsy GT knowledge selected: False Response: I am glad she is receiving treatment. Dementia has a significant effect on a person's caregivers.
ELVM w/ BM25Query: neurosurgery epilepsy neurocognitive neurosciences GT knowledge selected: True Response: I am glad to hear that she is receiving treatment. Neurosurgery is a branch of medicine that specializes in the treatment of epilepsy.
ELVM w/ DRQuery: neurosurgery treatment for epilepsy GT knowledge selected: True Response: Did she undergo a fellowship? Many neurosurgeons pursue additional training in the form of a fellowship, after residency or in some cases, as a senior resident.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.819, + 0.884, + 0.848 + ], + "angle": 0, + "content": "Table 21: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2698" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.129, + 0.885, + 0.829 + ], + "angle": 0, + "content": "
Dialogue ContextDo you know anything about the band called Insane Clown Posse? \nInsane Clown Posse (ICP) is an American hip hop duo composed of Violent J, and Shaggy 2 Dope. [...] \nThat's quite a lot of albums, what else do you know about ICP? \n[...] The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas" and "Big Money Rustlas". [...] \nThe duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas
GT knowledge
GT Response
KnownedGPTResponse: Query: n/a \nGT knowledge selected: False
Response: the duo has earned two platinum and five gold albums. they have also earned two platinum and five gold albums
R2C2-PTQuery: ICP wiki \nGT knowledge selected: False
Response: The duo has earned two platinum ,with 5 golden album
ELVM-β = 0Query: ICP facts \nGT knowledge selected: False
Response: The group has been around for a long time .
ELVM w/ BM25Query: Insane clown posse ICP facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
ELVM w/ DRQuery: Insane clown posse facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
Dialogue Context[...] Athletes normally eat meat in large quantities. \nExactly. I have also added much meat eating as part of my special diet. They are rich in protein. Especially, red meat
GT knowledge[...] In nutritional science, on the other hand, red meat is defined as any meat that has more myoglobin than white meat, white meat being defined as non-dark meat from chicken (excluding leg or thigh), or fish. [...]
GT ResponseRed meat is divined as any meat that has myoglobin than white
KnownedGPTQuery: n/a \nGT knowledge selected: False
Response: red meat is classified as white meat. i think it's interesting that only flesh from mammals or fowl is classified as red.
R2C2-PTQuery: protein powder \nGT knowledge selected: False
Response: Athletes normally eat meat in large quantities.
ELVM-β = 0Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ BM25Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ DRQuery: red meat protein \nGT knowledge selected: True
Response: Red meat is defined as any meat that has more myoglobin than white meat.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.837, + 0.883, + 0.867 + ], + "angle": 0, + "content": "Table 22: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2699" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.188, + 0.884, + 0.771 + ], + "angle": 0, + "content": "
Dialogue ContextWhen did Mark Spitz' swimming career end? Following the Munich Olympics [...] What else did Mark Spitz do besides Speedo swimsuit ads after his swimming career? [...] However, as the memory of his feats receded, so did his endorsement and promotional deals. As his endorsements faded, he started a successful real-estate company in Beverly Hills. He bought a Ferrari and says he made [...]
GT knowledge
GT ResponseMark Spitz started a successful real-estate company in Beverly Hills
R2C2-PTQuery: What did mark spitz do after his olympic career? \nGT knowledge selected: False \nResponse: In 1999 Spitz ranked No. 33 on ESPN SportsCentury 50 Greatest Athletes, the only aquatic athlete to make the list.
ELVM-β = 0Query: Mark Spitz Speedo Swimsuit ads after speedo swimming career ends mark spitz m Munich olympics 2012 \nGT knowledge selected: False \nResponse: Spitz did Schick razor commercials and Speedo swimsuit ads
ELVM w/ BM25Query: Spitz endorsements after swimming career ends \nGT knowledge selected: True \nResponse: Mark Spitz started a successful real-estate company in Beverly Hills.
Dialogue ContextWhat was the New Scheme? \nThe 'new scheme' and the 'new dedication' respectively, initially seen as a [...] \n[...] \nDid Ahmadi followers of the New Scheme act as missionaries? \nAs part of this Mahmood Ahmad appointed 5 men to survey the Punjab [...] \nDid Mirza Basheer-ud-Din Mahmood Ahmad use missionaries to try to spread the New Scheme to foreign countries?
GT knowledge[...] In time the scheme produced a vast amount of literature in defence of Islam in general and the Ahmadiyya beliefs in particular. The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance. As part of this Mahmood Ahmad appointed [...]
GT ResponseThe funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
R2C2-PTQuery: Mirza Basher-ud Din Mahmood Ahmad missionaries \nGT knowledge selected: False \nResponse: Mirza Basheer-ud-Din Mahmood Ahmad used missionaries to try to spread the New Scheme to foreign countries.
ELVM-β = 0Query: New Scheme Mahmud Ahmad missionaries to spread New Scheme in foreign countries \nGT knowledge selected: True \nResponse: Yes, I have. I have experienced heartache, also known as loss of a loved one.
ELVM w/ BM25Query: New scheme missionaries used to spread Ahmadiyyah teachings in foreign countries \nGT knowledge selected: True \nResponse: The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.78, + 0.884, + 0.81 + ], + "angle": 0, + "content": "Table 23: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "2700" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.136, + 0.885, + 0.824 + ], + "angle": 0, + "content": "
Dialogue ContextHow did Stephen Sondheim collaborate with James Lapine?\nStephen Sondheim found a new collaborator in [...]What did Stephen Sondheim and James Lapine do together in the 1980s?\nTheir first collaboration was Sunday in the Park with George (1984).Was Stephen Sondheim and James Lapine's first collaboration a success?\nStephen Sondheim and James Lapine won the 1985 Pulitzer Prize [...]Did Stephen Sondheim and James Lapine collaborate in other projects, besides Sunday in the Park with George?[...] last work together was the rhapsodic Passion (1994).Was Stephen Sondheim and James Lapine's last work a success?\nSondheim and Lapine's last work together was the rhapsodic Passion (1994), adapted from Ettore Scola's Italian film Passione D'Amore. With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.[46] Later work Assassins opened [...]With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
GT knowledge
GT Response
R2C2-PTQuery: Was the play "rhapsodic passion" a success or a failure?\nGT knowledge selected: TrueResponse: Sondheim and Lapine's last work together was the rhapsodic Passion (1994).
ELVM-β = 0Query: Stephen Sonds and Lapine Rhapsodic passion1994 Stephen sondheim, James lapine plays 12 dreams off-broadway p Pulitzer prize for play Sunday in Park.GT knowledge selected: FalseResponse: Stephen Sondheim and James Lapine's last work together was the rhapsodic Passion (1994).
ELVM w/ BM25Query: Sondheim last work was the Rhapsodic passion (1994) Pulitzer prize won for Drama. Lapine wrote the play Twelve Dreams.GT knowledge selected: TrueResponse: With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
Dialogue ContextDid Frank Hague commit any questionable acts?\nYes, he took kickbacks in cash, and relied on two ordinances of dubious [...]Did Frank Hague do anything illegal?\nHe passed an unconstitutional ordinance that gave himself as commissioner the power to [...]Did anyone attempt to prosecute Frank Hague?[...] Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. [38] He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash. [...]Yes, but he was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
GT knowledge
GT Response
R2C2-PTQuery: Frank Hague prosecutionGT knowledge selected: FalseResponse: No, no one tried to prosecute Frank Hague.
ELVM-β = 0Query: Frank Hague ordinance ordinance frank Hague charges Frank Hague Hague for flouting constitutionality ordinancesGT knowledge selected: FalseResponse: Hague was found guilty of contempt of court for ignoring the subpoena. He was fined $100 and stripped of his duties as Deputy Sheriff.
ELVM w/ BM25Query: Hague prosecution attempt failed after Hague passed unconstitutional ordinance. no Hague was prosecuted.GT knowledge selected: TrueResponse: Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.831, + 0.884, + 0.862 + ], + "angle": 0, + "content": "Table 24: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "2701" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.188, + 0.092, + 0.197, + 0.1 + ], + "angle": 0, + "content": ">" + }, + { + "type": "header", + "bbox": [ + 0.799, + 0.092, + 0.814, + 0.1 + ], + "angle": 0, + "content": "三" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.156, + 0.594, + 0.172 + ], + "angle": 0, + "content": "Human Evaluation Annotation Tool" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.184, + 0.428, + 0.196 + ], + "angle": 0, + "content": "Components of each example" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.206, + 0.465, + 0.215 + ], + "angle": 0, + "content": "Each example may include one of more of the following" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.224, + 0.642, + 0.233 + ], + "angle": 0, + "content": "- History: topic related to the conversation and the dialogue history. Examples may not contain 'history'" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.234, + 0.401, + 0.243 + ], + "angle": 0, + "content": "- Input text: Utterance of the opponent" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.245, + 0.669, + 0.254 + ], + "angle": 0, + "content": "- Ground truth response of the example, given as a reference. The annotator may use the reference in decision." + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.255, + 0.399, + 0.263 + ], + "angle": 0, + "content": "- Two responses from different models." + }, + { + "type": "list", + "bbox": [ + 0.251, + 0.224, + 0.669, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.275, + 0.382, + 0.286 + ], + "angle": 0, + "content": "Annotation Standards" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.296, + 0.747, + 0.317 + ], + "angle": 0, + "content": "The annotator marks which model of the given two is better on three different categories: Fluency and Relevance. The deifintion of each category is" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.325, + 0.608, + 0.334 + ], + "angle": 0, + "content": "- Fluency measures the correctness, completeness, and semantic consistency of the response." + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.335, + 0.601, + 0.344 + ], + "angle": 0, + "content": "- Relevance measures the degree of relevance of the knowledge and response to the history." + }, + { + "type": "list", + "bbox": [ + 0.251, + 0.325, + 0.608, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.387, + 0.29, + 0.397 + ], + "angle": 0, + "content": "History:" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.399, + 0.29, + 0.41 + ], + "angle": 0, + "content": "Singing" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.412, + 0.686, + 0.422 + ], + "angle": 0, + "content": "I love to sing. Some of my favorite genres to sing are blues, pop, and rock. Do you like to sing?" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.424, + 0.584, + 0.435 + ], + "angle": 0, + "content": "Yes very much so. I love older artists like Ray Charles and Billie Holiday." + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.437, + 0.715, + 0.448 + ], + "angle": 0, + "content": "Oh yeah I love Billie Holiday. She had no formal music training but her vocal delivery made up for it." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.458, + 0.303, + 0.468 + ], + "angle": 0, + "content": "Input Text:" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.47, + 0.538, + 0.481 + ], + "angle": 0, + "content": "I've been getting into the strokes as of late. Do you like them?" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.49, + 0.294, + 0.498 + ], + "angle": 0, + "content": "Reference:" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.5, + 0.739, + 0.519 + ], + "angle": 0, + "content": "I used to listen to The Strokes a lot back in high school. Their debut album was added to Rolling Stones 500 greatest albums of all time at number 199." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.562, + 0.338, + 0.571 + ], + "angle": 0, + "content": "Model A Response:" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.574, + 0.611, + 0.585 + ], + "angle": 0, + "content": "Yes I do like them. I like all kinds of music, but I'm not a big fan of the Strokes." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.595, + 0.34, + 0.605 + ], + "angle": 0, + "content": "Model B Response:" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.607, + 0.718, + 0.63 + ], + "angle": 0, + "content": "I do like them. I also like to wear my Strokes shirt during May because it is national stroke awareness month." + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.65, + 0.362, + 0.665 + ], + "angle": 0, + "content": "Annotation" + }, + { + "type": "title", + "bbox": [ + 0.251, + 0.676, + 0.288, + 0.686 + ], + "angle": 0, + "content": "Fluency" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.692, + 0.395, + 0.703 + ], + "angle": 0, + "content": "Both responses are fluent." + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.705, + 0.393, + 0.727 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Model A response is more fluent." + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.73, + 0.393, + 0.752 + ], + "angle": 0, + "content": "Model B response is more fluent." + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.755, + 0.395, + 0.766 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Neither response is fluent." + }, + { + "type": "list", + "bbox": [ + 0.253, + 0.692, + 0.395, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.791, + 0.305, + 0.799 + ], + "angle": 0, + "content": "Previous" + }, + { + "type": "title", + "bbox": [ + 0.42, + 0.676, + 0.464, + 0.685 + ], + "angle": 0, + "content": "Relevance" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.692, + 0.545, + 0.715 + ], + "angle": 0, + "content": "Both knowledges and responses are relevant" + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.717, + 0.559, + 0.739 + ], + "angle": 0, + "content": "Model A response is more relevant." + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.742, + 0.56, + 0.764 + ], + "angle": 0, + "content": "Model B response is more relevant." + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.767, + 0.528, + 0.778 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Neither is relevant." + }, + { + "type": "list", + "bbox": [ + 0.42, + 0.692, + 0.56, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.537, + 0.799 + ], + "angle": 0, + "content": "Next" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.888, + 0.337, + 0.897 + ], + "angle": 0, + "content": "Made with Streamlit" + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.913, + 0.646, + 0.927 + ], + "angle": 0, + "content": "Figure 2: Screenshot of the annotation tool." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2702" + } + ] +] \ No newline at end of file diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_origin.pdf b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5bcda92612829878ad7f439a3ae3d1884c0cc737 --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/ef187d64-77d7-4019-9ab7-2e41bbdafa31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f35c6ceb6a51604da5093de565f54ad017fea0f9757e270e4fac4152c16ebea +size 2420284 diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/full.md b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41de8933417d36b0574047608869eacef2f573ac --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/full.md @@ -0,0 +1,550 @@ +# Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation + +Gunsoo Han $^{1}$ Daejin Jo $^{1}$ Daniel Wontae Nam $^{1}$ Eunseop Yoon $^{2}$ Taehwan Kwon $^{3*}$ Seungeun Rho $^{4*}$ Kyoung-Woon On $^{1}$ Chang D. Yoo $^{2}$ Sungwoong Kim $^{5\dagger}$ + +$^{1}$ Kakao Brain $^{2}$ KAIST $^{3}$ KRAFTON $^{4}$ Georgia Institute of Technology $^{5}$ Korea University + +gunsoo.han@kakaobrain.com swkim01@korea.ac.kr + +# Abstract + +Knowledge-grounded dialogue generation requires to first retrieve appropriate external knowledge based on a conversational context and then generate a response grounded on the retrieved knowledge. In general, these two sequential modules, a knowledge retriever and a response generator, have been separately trained by supervised data for each module. However, obtaining intermediate labels of the ground-truth knowledge is expensive and difficult especially in open-domain conversation. Latent variable modeling can circumvent it and enables a joint training without the knowledge supervision. In this paper, we propose an efficient algorithm for this latent variable modeling that is able to leverage a large amount of dialogue data. In specific, rather than directly training the complex retriever, we adapt a query generator with an off-the-shelf retriever, and the query generator and response generator are simultaneously trained over the latent variable of query. Moreover, we employ the evidence lower bound as a training objective and modify it to efficiently and robustly perform the joint training. Experimental results on diverse knowledge-grounded dialogue datasets show that the proposed algorithm achieves state-of-the-art performances even without the use of the annotated knowledge while maintaining the efficiency and scalability. + +# 1 Introduction + +Recently, knowledge-grounded dialogue generation has drawn increasing attention especially as a key ingredient for open-domain conversational agents (Dinan et al., 2018; Zhou et al., 2018b; Zhan et al., 2021; Liu et al., 2018; Zhou et al., 2018a; Lian et al., 2019; Zhao et al., 2020; Kim et al., 2020; Chen et al., 2022; kom; Shuster et al., 2022a,b; Cai et al., 2023; Wu et al., 2022; Lewis et al., 2020; Huang et al., 2021; Thulke et al., 2021; Anantha + +et al., 2021). It usually obtains knowledge from external resources such as Wikipedia-based databases and web search engines since internal knowledge even in large-scale parametric language models (Brown et al., 2020; Thoppilan et al., 2022; Chowdhery et al., 2022) is incomplete or outdated, and moreover it can provide hallucinated information. + +In order for a dialogue response to be grounded on such external knowledge, the conversational agent generally consists of a knowledge retriever, which retrieves knowledge corresponding to a given dialogue context, followed by a response generator that produces an informative response based on the dialogue context and the retrieved knowledge. In many previous methods, supervised learning has often been applied to independently optimize each module using the ground-truth or gold knowledge (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). However, human annotation of knowledge information is cumbersome, expensive, and often incomplete due to the existence of multiple possible knowledge candidates and the overlooking of the target response. In addition, existing automatic annotations are generally limited to a simple extractive question answering (Zhao et al., 2020; Liu et al., 2021). This difficulty in obtaining annotations of knowledge information could be severe under the open-domain conversation and hinders the use of large-scale dialogue data. + +Therefore, there have been a number of recent approaches to learn the knowledge retriever and the response generator without the knowledge supervision. In specific, they have treated the retrieved knowledge, document, or passage as an unobserved latent variable and adapt latent variable modeling based on approximated marginalization (e.g. top-k) (Lewis et al., 2020; Huang et al., 2021; Cai et al., 2023; Guu et al., 2020), reinforcement learning (Zhao et al., 2020; Zhang et al., 2022; Chen et al., 2022; Wu et al., 2022) or variational methods (Zhan + +et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020; Chen et al., 2020; Xu et al., 2023). However, joint training of the retriever along with the generator under this latent variable modeling has some restrictions in utilizing the retriever. For example, a retriever needs to produce a differentiable prior probability for the gradient propagation through the marginalization or to be intermittently updated to rebuild the whole passage index during training (Karpukhin et al., 2020; Lewis et al., 2020; Zhang et al., 2022). This leads to limitation in facilitating complicated or large-scale retrievers and further in leveraging large-scale data. In addition, the posterior probability used in the previous variational methods has been modeled by a separated network, which grows the training complexity and also incurs the discrepancy between the training-time and test-time knowledge samplings. + +To overcome such restrictions, in this paper, we propose an efficient latent variable modeling, named $ELVM$ , for knowledge-grounded dialogue generation. In particular, we reduce the burden for training a whole retriever by employing a query generator followed by an off-the-shelf retriever that is fixed during training. More precisely, in tandem with the response generator, the query generator rather than the retriever is jointly optimized by latent variable modeling, in which a query is taken as a latent variable, on paired observations of dialogue contexts and responses. + +To this end, we exploit the variational method using the evidence lower bound (ELBO) (Kingma and Welling, 2013; Jordan et al., 1999) and approximate the expected conditional likelihood in the ELBO by subset sampling from the prior distribution, which acts as the training objective for the response and the query generators. This approximation gets rid of an extra modeling of a surrogate posterior distribution or online posterior inference such as Markov Chain Monte Carlo (MCMC) and also reduces the training-inference discrepancy in knowledge retrieval. Moreover, we further modify the Kullback-Leibler (KL) regularization of the ELBO by constructing the approximated posterior distribution from the conditional likelihood and prior distribution and set this posterior distribution as a teacher for a distillation objective to further learn the query generator. + +Experimental results show that the proposed ELVM allows to efficiently and robustly perform training without (1) the use of the annotated knowledge, (2) an explicit training of the knowledge + +retrieval, and (3) a complex posterior sampling. Especially, it significantly outperforms previous state-of-the-art methods for knowledge-grounded dialogue generation. In addition, the proposed posterior distillation improves the performances over the baseline that solely maximizes the expectation of the conditional likelihood. + +Our main contributions can be summarized as: + +- An efficient latent variable modeling is proposed in joint training of query generator and dialogue response generator without the knowledge supervision for knowledge-intensive dialogue generation. +- For realizing efficient yet robust joint training, the ELBO for the marginal likelihood is modified as the combination of the conditional likelihood objective and the posterior distillation objective, based on multiple prior samples. +- The proposed ELVM demonstrates its effectiveness in performing unsupervised joint training and even significant performance improvements over previous methods, being a new state-of-the-art on benchmark datasets. + +# 2 Related Work + +# 2.1 Knowledge-Grounded Dialogue Generation + +Knowledge-grounded conversation has been extensively studied through recently released public datasets (Dinan et al., 2018; kom; Xu et al., 2022b; Kwiatkowski et al., 2019; Anantha et al., 2021; Moghe et al., 2018). Existing approaches for this task have mostly exploited two successive modules, the knowledge retriever and the knowledge-grounded response generator. Many previous works have used manual annotations of gold knowledge provided by some of public datasets to optimize the modules (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). Specifically, Seeker (Shuster et al., 2022a) and BlenderBot3 (Shuster et al., 2022b) have recently proposed to build a series of modules by a single transformer with different prompts for each module, and trained its transformer on a large number of modular tasks using annotated datasets. However, this manual annotation is expensive, time-consuming, and often inaccurate, which impedes the utilization of a large-scale dialogue data that is especially necessary for open-domain conversation. + +# 2.2 Unsupervised Joint Training + +Recently, unsupervised training methods have been widely applied, and many of them have tries to jointly learn the modules without knowledge labels (Lewis et al., 2020; Huang et al., 2021; Zhao et al., 2020; Zhang et al., 2022; Zhan et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020). For example, PLATO-KAG (Huang et al., 2021) has approximated the marginal likelihood by top-k knowledge samples while RetGen (Zhang et al., 2022) has trained to reward knowledge retrieval with the highest utilization in response generation by reinforcement learning and mixture-of-experts ensembling. + +Meanwhile, latent variable modeling based on variational methods has also been developed by several recent works. CoLV (Zhan et al., 2021) has introduced collaborative latent spaces to reflect the inherent correlation between the knowledge selection and the response generation. SKT (Kim et al., 2020) has developed a sequential latent variable model for the multi-turn knowledge-grounded dialogue generation. In order to perform the posterior sampling of knowledge selection during joint training, some works have proposed to separately train the posterior distribution model (Paranjape et al., 2022; Lian et al., 2019) or the posterior information prediction model (Chen et al., 2020). Very recently, SPI (Xu et al., 2023) has applied short-run MCMC (Erik et al., 2019) for posterior sampling on the collaborative latent spaces. + +While these latent variable modeling algorithms can effectively perform unsupervised joint training, the entire training of the retriever is still difficult and imposes restrictions on the selection of the retriever in terms of both a search algorithm and a knowledge resource. Furthermore, the additional posterior sampling through separate networks or multiple iterations in the previous variational methods also leads to increased training complexity as well as the training-inference discrepancy in knowledge generation. In contrast, the proposed ELVM can employ any kind of retriever in latent variable modeling since it controls a retrieval output by only changing of a generated query. Moreover, ELVM removes an extra posterior modeling or sampling by prior subset sampling and approximated posterior distillation, which leads to efficient training without the discrepancy. + +# 2.3 Query Generation + +When an off-the-shelf retriever is used for knowledge retrieval, how to make an input text query + +is important for obtaining appropriate documents. Especially, in knowledge-grounded dialogue, a self-contained query should be generated from the multi-turn dialogue context. A number of prior works have tried to train a supervised query rewriting model by human rewrites (Yu et al., 2020; Lin et al., 2020; Vakulenko et al., 2021; Voskarides et al., 2020). Similar to the knowledge annotation, the limitations of manual query annotation have come up with unsupervised learning of the query generator. A number of works have proposed a novel rewards for reinforcement learning of a query rewritten (Wu et al., 2022; Chen et al., 2022). + +QKConv (Cai et al., 2023) has proposed an unsupervised query enhanced method for knowledge-grounded conversation. This work is similar to ours in that it consists of a query generator, an off-the-shelf knowledge retriever, and a response generator, and unsupervised joint training is applied to the query generator and the response generator. However, their training objective is based on the approximated marginal likelihood over candidate queries, which is different from our objective that includes the posterior distillation based on the ELBO loss. + +# 3 Efficient Latent Variable Model for Knowledge-Grounded Dialogue Generation + +# 3.1 Setup + +In this section, we explain an overall structure of our ELVM, as described in Figure 1a, and define notations. Unlike previous works of document (passage) retrieval (Zhang et al., 2022; Huang et al., 2021; Lewis et al., 2020) using a query of an embedded vector of a dialogue context $x$ , we use a text query of natural language, which enables to utilize any kind of retriever for dialogue response generation. We define the probability of generating a natural language query $u$ for a given dialogue context $x$ , $p_{\phi}(u|x)$ , as below: + +$$ +p _ {\phi} (u | x) = \prod_ {t} p _ {\phi} \left(u _ {t} \mid u _ {0: t - 1}, x\right), \tag {1} +$$ + +where $u_{t}$ is the $t$ -th token in $u$ and $p_{\phi}$ is the probability obtained from the model parameterized by $\phi$ . Similarly, the probability of generating a response $y$ for a given query $u$ and dialogue context $x$ is defined as + +$$ +p _ {\theta} (y | Z (u), x) = \prod_ {t} p _ {\theta} \left(y _ {t} \mid y _ {0: t - 1}, Z (u), x\right), \tag {2} +$$ + +![](images/4cfe6d0673e496a8a96875adf2b2e3dfa927b4c687d9710d9bafbea2183e7fe0.jpg) +(a) Overall structure of ELVM. + +![](images/3c140f2c8b4e80b181f7d1a434c3dbfa3a15da0576b0dd08f773430c18f14263.jpg) +(b) Example inference of ELVM. +Figure 1: Overall structure and example inference case of ELVM. During training, for a given dialogue context $x$ , the query generator generates a set of unique queries $Q^s$ and for each query $u$ a set of top- $k$ documents $Z(u)$ is retrieved using the off-the-shelf retriever. Then, response generator utilizes both $x$ and $Z(u)$ : (a) The approximated marginal likelihood $\tilde{p}_{\theta, \phi}(y|x)$ used during training is defined by $Q^s$ and the re-normalized prior $\bar{p}_{\phi}(u|x)$ ; (b) At inference time, a single sampled query $u$ is used for retrieving multiple documents and producing a final response. + +where $y_{t}$ is the $t$ -th token in $y$ , $Z(u) = \{z_1, z_2, \dots, z_k\}$ is the set of retrieved documents (passages) by $u$ , and $\theta$ is the model parameters to produce the response probability. Here, we use top- $k$ result obtained by an off-the-shelf retriever such as BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) and freeze the retriever. We assume that there is a deterministic mapping from $u$ to $Z$ . We now set a query $u$ as a latent variable. Then, using the prior $p_{\phi}(u|x)$ associated with the query generator and the conditional likelihood $p_{\theta}(y|Z(u), x)$ corresponding to the response generator, we can define the marginal likelihood of the knowledge-grounded response $y$ given the dialogue context $x$ as below: + +$$ +p _ {\theta , \phi} (y | x) = \sum_ {u \in Q} p _ {\phi} (u | x) p _ {\theta} (y | Z (u), x), \tag {3} +$$ + +where $Q$ is the set of all possible queries. Since it is intractable to marginalize over all queries, we sample a small subset composed of $m$ unique queries, $Q^{s}(x) = \{u_{1}, u_{2}, \dots, u_{m}\}$ using $p_{\phi}(\cdot | x)$ and then marginalize over this set as below: + +$$ +\tilde {p} _ {\theta , \phi} (y | x) = \sum_ {u \in Q ^ {s}} \bar {p} _ {\phi} (u | x) p _ {\theta} (y | Z (u), x), \tag {4} +$$ + +where + +$$ +\bar {p} _ {\phi} (u | x) = \frac {p _ {\phi} (u | x)}{\sum_ {u ^ {\prime} \in Q ^ {s}} p _ {\phi} \left(u ^ {\prime} \mid x\right)}. \tag {5} +$$ + +We construct $Q^s$ by obtaining $m$ -unique queries sampled from $p_{\phi}(\cdot | x)$ . It is noted that our response generator takes multiple retrieved documents efficiently using FiD (Izacard and Grave, 2020), which is different from the use of a single document for response generation in previous methods. In addition, while we can utilize relative matching scores from the retriever to assign varying likelihoods to each document given a query for $p(Z(u)|u)$ , we have noticed that this approach can occasionally hinder the effective and comprehensive learning of the query generator. During the training process, that will be described in the next subsection, our objective is to amplify the impact of the generated query. Consequently, we opt for a deterministic mapping from a query to a set of retrieved documents as a unified whole, aiming to enhance the gradient flow. + +# 3.2 Joint Training + +Our training objective is the ELBO (Kingma and Welling, 2013; Jordan et al., 1999) of $\log \tilde{p}_{\theta,\phi}$ which can be written as below: + +$$ +\begin{array}{l} \log \tilde {p} _ {\theta , \phi} (y | x) \geq \mathbb {E} _ {u \sim q (u | x, y)} [ \log p _ {\theta} (y | Z (u), x) ] \\ - D _ {K L} \left(q (u | x, y) | | \bar {p} _ {\phi} (u | x)\right), \tag {6} \\ \end{array} +$$ + +where $q(u|x,y)$ is the variational posterior. Note that the equality holds when $q(u|x,y)$ is identical to the true posterior distribution of it. In this work, + +we do not parameterize $q(u|x,y)$ and approximate the expected conditional likelihood in the ELBO by sampling $u$ from not $q(u|x,y)$ but $\bar{p}_{\phi}(u|x)$ such as + +$$ +\begin{array}{l} \mathbb {E} _ {u \sim q (u | x, y)} [ \log p _ {\theta} (y | Z (u), x) ] \\ \approx \mathbb {E} _ {u \sim \bar {p} _ {\phi} (u | x)} [ \log p _ {\theta} (y | Z (u), x) ]. \tag {7} \\ \end{array} +$$ + +This approximation is due to that it minimizes the discrepancy between training and test time inference and it does not separately model $q(u|x,y)$ or complicatedly sample from it during training. Then, we obtain the following two losses corresponding to the expected conditional likelihood and the KL regularization such as + +$$ +\mathcal {L} ^ {y, u} (\theta , \phi) = - \mathbb {E} _ {u \sim \bar {p} _ {\phi} (u | x)} [ \log p _ {\theta} (y | Z (u), x) ], \tag {8} +$$ + +$$ +\mathcal {L} ^ {u} (\phi) = D _ {K L} \left(q (u | x, y) | | \bar {p} _ {\phi} (u | x)\right). \tag {9} +$$ + +Moreover, we define $q(u|x,y)$ in $\mathcal{L}^u (\phi)$ by the approximated posterior derived from the prior and the likelihood such as + +$$ +q (u | x, y) = \frac {p _ {\theta} (y | Z (u) , x) \bar {p} _ {\phi} (u | x)}{\sum_ {u ^ {\prime} \in Q ^ {s}} p _ {\theta} (y | Z (u ^ {\prime}) , x) \bar {p} _ {\phi} \left(u ^ {\prime} \mid x\right)}. \tag {10} +$$ + +We set this posterior distribution $q(u|x,y)$ as a teacher for a distillation to update the prior associated with the query generation at test time. Namely, $\mathcal{L}^u (\phi)$ becomes the posterior distillation loss, and our prior is explicitly updated to be similar to the posterior. Here, the gradient is not propagated from $q(u|x,y)$ , and $q(u|x,y)$ can be computed easily without additional model or approximation since we define it over the subset $Q^s$ . + +To sum up, we obtain the conditional likelihood loss and the posterior distillation loss by multiple samples from the prior distribution that is aligned with query sampling at the inference time. We would mitigate the bias of prior sampling by multiple query samples for each input during training. In addition, our modification of ELBO is different from the expectation-maximization (EM) algorithm in that we sample a latent variable from the prior distribution rather than the posterior distribution that is used for sampling of E-step in EM. Moreover, in contrast to the M-step loss, we use the prior-weighted conditional likelihood and the gradient is also propagated to the sampled latents and the corresponding prior model. For experiments in this paper, the query generator $\phi$ and the response generator $\theta$ share parameters, with different input prompts to distinguish each module, for simplicity. + +Overall, the loss for our joint training is + +$$ +\mathcal {L} (\theta , \phi) = \mathcal {L} ^ {y, u} (\theta , \phi) + \beta \mathcal {L} ^ {u} (\phi), \tag {11} +$$ + +where $\beta$ is the relative weight for the posterior distillation loss. + +# 4 Experiment + +# 4.1 Experimental Setup + +Dataset. Our experiments are conducted on two widely used knowledge-grounded dialogue datasets, Wizard of Wikipedia (WoW) (Dinan et al., 2018) and QReCC (Anantha et al., 2021). WoW is structured around dialogues that are grounded in knowledge sentences sourced from Wikipedia, consisting of roughly 18K, 2K, and 2K dialogues, for training, validation, and test subsets, respectively. QReCC is a comprehensive compilation of conversational data, consisting of 14K open-domain conversations comprising 80K question-answer pairs. + +
TaskDatasetTrainValid
QueryWizInt (Komeili et al., 2022)351,3752,467
Fits (xu2)3,587392
KG ResponseWizInt (Komeili et al., 2022)22,4881,687
Fits (xu2)6,279656
Ms Marco (Nguyen et al., 2017)281,63636,859
NQ Open (Adolphs et al., 2021b)79,1688.757
Dialogue ResponseWizInt (Komeili et al., 2022)8,335587
MSC (Xu et al., 2022a)105,54917,691
SaferDialogues (Ung et al., 2022)6,3067,88
PersonaChat (Zhang et al., 2018)131,4387.801
EmpatheicDialogues (Rashkin et al., 2019)64,6365,738
+ +Table 1: List of dataset and number of train and validation examples used for fine-tuning R2C2 model to obtain R2C2-PT. KG is an abbreviation for knowledge-grounded. + +Training. We begin with publicly available R2C2 pre-trained transformer model (Shuster et al., 2022a) with 400M parameter size and fine-tune it on multiple datasets to acquire the essential skill sets for knowledge-grounded dialogue such as query generation and knowledge-grounded response generation. Detailed tasks and statistics are shown in Table 1. In line with Shuster et al. (2022b), we adopt a modular system with a single model that incorporates control tokens into the dialogue context to perform different modules. For example, appending the control token __generate-query__ to the dialogue context promotes the model to generate relevant search query. We name the initial model from this pretraining procedure as R2C2-PT. Subsequently, we apply our proposed algorithm to R2C2-PT model, + +
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
PostKS (Lian et al., 2019)79.1--13.01.0193.8--13.11.0
ITDD (Li et al., 2019)17.84.02.516.2-44.82.11.111.4-
SKT (Kim et al., 2020)52.0--19.36.881.4--16.14.2
PIPM (Chen et al., 2020)42.7-3.319.97.365.7-2.517.65.4
ZRKGK (Li et al., 2021)40.4----41.5----
CoLV (Zhan et al., 2021)39.6-2.920.67.954.3-2.119.76.3
KnowledGPT (Zhao et al., 2020)19.29.57.222.07.922.38.36.020.56.7
SPI (Xu et al., 2023)17.110.27.722.78.819.19.67.322.08.5
R2C240.71.70.912.31.946.61.60.911.95.4
R2C2-PT24.76.04.417.35.230.46.24.617.75.4
ELVM-OK19.812.09.525.310.725.812.19.725.210.5
ELVM-EM-Like16.211.69.125.710.719.412.09.525.710.8
ELVM-from-R2C215.814.511.828.813.318.914.311.528.813.3
ELVM14.614.711.929.313.918.014.511.829.213.7
+ +Table 2: Performance comparison in WoW. PPL represents perplexity, B3 and B4 denote BLEU-3 and BLEU-4 scores, respectively, and R1 and R2 indicate Rouge-1 and Rouge-2 scores. Our proposed model, ELVM, achieves new SOTA performance for both seen and unseen tasks with a substantial margin. + +
ModelSeenUnseen
R@1R@5R@10R@1R@5R@10
R2C2-PT25.763.874.329.264.876.7
ELVM-EM-Like26.463.974.728.964.575.8
ELVM-β=026.164.675.728.965.376.8
ELVM29.168.480.330.468.380.3
+ +resulting in ELVM. Furthermore, to dispel concerns that our training gains might predominantly stem from R2C2-PT's comprehensive fine-tuning across multiple datasets rather than ELVM's inherent efficacy, we opted to utilize the base R2C2 model. Upon integrating it with ELVM, the resulting configuration is denoted as ELVM-from-R2C2. + +During the training of ELVM, we employ a sampling-based generation method, specifically nucleus sampling (Holtzman et al., 2020), to promote diverse query generation in $p_{\phi}$ . When retrieving documents based on the generated queries, we utilize an off-the-shelf sparse retriever such as BM25 (Robertson et al., 2009), prioritizing low latency over the dense retriever (DR). We train for 3 epochs with $m = 4$ and $k = 5$ for WoW and 1 epoch with $m = 8$ and $k = 10$ for QReCC. $\beta$ is set to 1 for both tasks, unless stated otherwise. To ensure a thorough evaluation of ELVM's performance and its ability to generalize to unseen dialogues, we purposely avoid dataset overlap between training of R2C2-PT and ELVM. More detailed information on the training process can be found in Appendix A. + +Table 3: Comparison of document recall in WoW between ELVM variants. Note that we omit the performance of ELVM-OK whose performance is identical to R2C2-PT since it is trained with keeping its query generator frozen. + +
ModelF1 ↑EM ↑R1 ↑R2 ↑PPL ↓
Q.Rewriting (Raposo et al., 2022)18.91.0---
DPR-IHN (Kim and Kim, 2022)30.44.7---
QKConv (Cai et al., 2023)33.55.9---
R2C2-PT22.81.325.711.59.4
ELVM-OK33.04.535.621.95.6
ELVM36.56.239.025.84.8
+ +Table 4: Performance comparison in QReCC with QKConv (Cai et al., 2023) which is the previous SOTA on this task. We report F1 scores, exact match (EM), Rouge score (R) and perplexity (PPL). + +
ModelFluency (%)Relevance (%)
SeenUnseenSeenUnseen
KnowledGPT69.862.941.541.1
ELVM73.577.5*49.563.6**
+ +Table 5: Human evaluation results on WoW test. A pairwise t-test is conducted to verify statistical significance of the improvements, and the corresponding results in bold are significantly better than those from the baseline model $(^{**}: p < 0.01, ^*: p < 0.025)$ . + +Inference. During inference, unlike the training phase, only a single query is generated by the query generator $p_{\phi}$ . The default number of retrieved documents $k$ from the off-the-shelf retriever is set to 5 for WoW and 10 for QReCC. The knowledge-grounded response is generated by the response generator, $p_{\theta}$ . An illustration of the inference is depicted in Figure 1b. During the inference, both query and response generation are conducted using beam search with a beam size of 5. + +Variants of ELVM. To examine the impact of the proposed ELVM training, we explore another variant, ELVM-OK, which is trained on the response generation task with annotated oracle knowledge documents while keeping the query generator $p_{\phi}$ + +
mSeenUnseen
B4↑R2↑B4↑R2↑
110.812.410.512.2
211.113.011.112.8
4*11.913.911.813.7
811.813.711.713.6
+ +Table 6: Effect of varying the number of queries, $m$ , sampled during training on WoW. We see the highest performance on unseen task when $m = 4$ . The superscripted value by * is the default setting for ELVM. + +
βWow UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
09.310.665.334.55.155.8
1*11.813.768.336.56.965.6
510.812.466.731.33.531.9
+ +frozen. During inference, given a dialogue context the frozen query generator $p_{\phi}$ generates query and use the off-the-shelf retriever to bring relevant documents. Finally the response generator of ELVM-OK generates the final response. + +In Section 3.2, we describe the difference between ELVM training and EM algorithm. In order to quantitatively compare these two methods, we perform EM-like training in ELVM where we apply the posterior-weighted conditional likelihood in $\mathcal{L}^{y,u}(\theta ,\phi)$ .. $\mathcal{L}^{y,u}(\theta ,\phi) =$ $-\sum_{u}SG(q(u|x,y))\log p_{\theta}(y|Z(u),x).$ Here, the query latents $u$ are still sampled from the prior distribution, and $SG$ means the stop-gradient that prevents the gradient propagation to the query generator. We name this variant as ELVM-EM-Like. + +# 4.2 Automatic Evaluation + +The evaluation results of WoW test seen and unseen tasks are shown in Table 2. ELVM surpasses all its variants and previous state-of-the-art (SOTA) model, achieving a new SOTA on both WoW test seen unseen tasks, across all metrics. In addition, ELVM-OK demonstrates notable improvements in BLEU and Rouge scores compared to previous models with a slight drawback in PPL. Similarly, ELVM-EM-Like also exhibits competitive performance across multiple metrics. Moreover, when comparing ELVM-from-R2C2 and ELVM, the experiment demonstrates that while R2C2-PT offers marginal benefits, the marked performance + +Table 7: Ablation study of the posterior distillation weight $\beta$ on WoW. From the results, we can see that the trivial value of $\beta = 1$ shows the best performance. The super-scripted value by $*$ is the default setting for ELVM. + +
ModelWoW UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
ELVM11.813.768.336.56.259.9
ELVMDR12.815.581.237.96.965.6
+ +Table 8: Performance results achieved by employing the dense retriever instead of the default BM25 retriever for WoW unseen and QReCC evaluation. + +
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑B4 ↑R2 ↑R@5 ↑
R2C2-PT4.45.263.84.65.464.8
ELVM-OK9.510.7-9.710.5-
ELVM11.913.968.411.813.768.3
R2C2-PTLarge8.710.262.98.910.364.4
ELVM-OKLarge9.610.6-10.211.2-
ELVMLarge12.614.372.612.414.174.3
+ +Table 9: Effect of model scaling tested on WoW. ELVM scales out to larger parameter size with increase in performance. The subscript Large indicates the model size of 2.7B parameters. + +increase is predominantly attributable to ELVM. + +We further examine the performance of query generator by assessing the recall of the ground-truth documents. The results summarized in Table 3 reveal that training with our proposed algorithm leads to an increase in recall for both ELVM- $\beta = 0$ and ELVM compared to R2C2-PT while the latter yields a larger performance gain. This highlights the effectiveness of distilling the posterior distribution of the response generator into the prior distribution associated with the query generator. + +For QReCC, as shown in Table 4, ELVM surpasses the previous SOTA, QKConv (Cai et al., 2023), the recent previous works (Raposo et al., 2022; Kim and Kim, 2022), and the variants of ELVM including R2C2-PT and ELVM-OK in terms of all metrics. The recall performance on QReCC is included in Table C. + +# 4.3 Human Evaluation + +To gauge performance across multiple aspects of the quality of generated responses, human evaluation is performed on the generated responses of ELVM and KnoawedGPT (Zhao et al., 2020) $^{1}$ . Similar to (Rashkin et al., 2021) and (Xu et al., 2023), we assess the response quality in two aspects: Fluency and Relevance. Fluency measures whether the response is understandable, self-consistent without repetition, and proficient. Relevance assesses the extent to which the response aligns with the dialogue context, incorporates pertinent knowledge, and maintains appropriateness. In total, 50 data + +
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
CoLV39.62.97.954.32.16.3
KnowledGPT19.27.27.922.36.06.7
SPI17.17.78.819.17.38.5
BART9741.92.69731.32.6
ELVM-from-BART15.38.411.116.48.210.9
ELVM14.611.913.918.011.813.7
+ +Table 10: Performance comparison with BART as our initial model. + +
ModelTrain RetTest RetWoW Unseen
PPL ↓B4 ↑R2 ↑F1 ↑
CoLV--54.32.16.318.5
KnowledgeGPT--22.36.06.720.5
SPI--19.17.38.5-
ELVMBM25BM2518.011.813.727.0
ELVMBM25Dense14.812.615.229.3
ELVMDenseDense12.912.815.529.0
ELVMDenseBM2521.611.813.128.2
+ +Table 11: Impact of alternating the retriever between training and testing phases on WoW unseen test. Dense indicates utilizing all-MiniLM-L6-v2 retriever as described in Section 4 and Ret is an abbreviation for retriever. + +samples are randomly selected from WoW tasks where 25 samples are randomly selected from each seen and unseen task. The qualities of the responses are measured by A/B testing on the two aspects. The feedback is collected from 11 human experts. Further details and annotator instructions can be found in Table G. As shown in Table 5, ELVM significantly outperforms KnowledGPT in all aspects especially on the unseen task. + +# 4.4 Ablation Studies + +Number of Queries. The size of sampled query set $Q^{s}$ , $m$ , plays a huge role in training of ELVM, as it directly influences the bias of the marginal likelihood of the knowledge-grounded response $y$ . To investigate the impact of $m$ , we gradually increase $m$ from 1 to 8 and measure the performance on WoW while keeping other hyperparameters fixed. As shown in Table 6, we observe a positive correlation where increasing $m$ leads to improved performance. Notably, setting $m = 4$ yields the optimal performance for both seen and unseen tasks. + +Posterior Distillation Weight. We examine the influence of the posterior distillation loss on ELVM training by adjusting the value of $\beta$ . The experimental results presented in Table 7 reveal that the optimal performance is attained when $\beta = 1$ for both the WoW unseen and QReCC tasks, indicating the effectiveness of incorporating the posterior distillation loss. However, a substantial decline in + +performance is observed when $\beta$ is increased to 5 or set to 0, particularly in the case of QReCC. + +Varying the Retriever. We extend our experiments by incorporating the DR in two distinct ways. For WoW, we train and evaluate using all MiniLM-L6-v2 $^2$ DR. In contrast, for QReCC, due to its extensive knowledge pool, we leverage DR solely during evaluation by employing both the BM25 retriever and all-mpnet-base-v2 $^3$ DR in a hybrid manner to retrieve the top- $k$ documents. Concretely, we score each document by summing up the normalized score of each retriever and rerank to obtain top- $k$ documents. Table 8 presents a summary of the performance results of employing DR, demonstrating a successful integration of DR in different settings of ELVM. + +Scaling Up Model. We investigate the impact of increasing the model size on performance. Starting with the publicly available 2.7B R2C2 pretrained transformer, we follow exactly the same procedure outlined in Section 3.2 to obtain R2C2-PTLarge, ELVM-OKLarge and ELVMLarge. Detailed hyperparameters for training 2.7B model is given in Appendix A. The results presented in Table 9 demonstrate that as the model size increases, there is an improvement in performance across all metrics, supporting the scalability of ELVM. + +BART as Base. To ensure a fair comparison and further demonstrate the robustness of our algorithm on diverse model scales, we conduct an experiment employing BART (Lewis et al., 2019) base as an initial model and then apply our proposed algorithm on WoW dataset. We choose BART base model since the previous SOTA model (Xu et al., 2023) utilized BART base as their base model. The outcome of this training is termed ELVM-from-BART. As shown in Table 10, the initial BART model shows very poor performances. However, upon training the BART-base with our ELVM algorithm, there is a great improvement in performance across both WoW Seen and Unseen tasks. Importantly, the results surpass the performance benchmarks set by the previous SOTA model (SPI). + +Retriever Switching To assess the robustness of our algorithm to the retriever mismatch between training and inference phases, we undertake an + +
Dialogue ContextI want to teach my family about investing by buying them a few shares of stock, but I really don't know much about it. The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses. Oh, so I could tell my kids they own part of a company. That's cool. Maybe I should get them a share in a company they know then, like Apple or XBox or Snapchat? [...] Examples of the latter include shares of private companies which are sold to investors through equity crowdfunding platforms. [...]
GT knowledge
GT ResponseYes, it's like the shares of a company are sold to investors through equity crowdfunding platforms.
KnownGPTQuery: n/aGT knowledge selected: False
Response: yes, you can get shares in a company like apple or xbox. you can also get shares in company like stock market.
R2C2-PTQuery: How to teach kids about investingGT knowledge selected: False
Response: The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses.
ELVM-β = 0Query: Investing tips for beginnersGT knowledge selected: False
Response: That's a good idea. The stock market is a place where you can buy and sell stocks.
ELVM w/ BM25Query: Apple xbox snapshot stock share buy apple stock instagram stock investingGT knowledge selected: True
Response: That is a good idea. You can also buy shares of private companies which are sold to investors through equity crowdfunding platforms.
ELVM w/ DRQuery: Best stocks for teaching kids about investingGT knowledge selected: True
Response: That is a good idea. The latter include shares of private companies which are sold to investors through equity crowdfunding platforms.
+ +Table 12: Generated samples from WoW unseen task. The first row represents the dialogue context, ground truth (GT) knowledge (highlighted in olive), and GT response. Following rows contain the model generated query, response and whether the GT knowledge is included in top-5 retrieved documents. More samples can be found in Appendix H. We truncate the GT knowledge and display only relevant spans of the GT knowledge to the ground-truth response. + +additional experiment, of which the results are depicted in Table 11. In this experiment, we train the WoW task utilizing either BM25 or dense retriever (utilizing all-MiniLM-L6-v2 retriever, as detailed Section 4), and subsequently evaluate using either BM25 or dense retriever. The results on the WoW unseen test indicate that utilizing a different retriever during testing doesn't necessitate model re-training, particularly given the superior performance of the dense retriever compared to BM25. + +# 4.5 Qualitative Analysis + +For qualitative analysis, we examine the generated samples from KnownGPT, R2C2-PT and other variants of ELVM, whose samples are presented in Table 12. The example reveals that the queries generated by KnownGPT, R2C2-PT and ELVM- $\beta = 0$ fail to retrieve the relevant knowledge effectively while the queries generated by ELVM with BM25 and DR (all-MiniLM-L6-v2) reflects the ability to access the relevant knowledge, leading to improved responses. The results also demonstrate that the query generator trained with our algorithm effectively adapts to the chosen retriever. + +# 5 Conclusion + +In this paper, we propose ELVM as an efficient latent variable modeling for knowledge-intensive dialogue generation. In ELVM, the knowledge retrieval is realized by the query generation followed by document retrieval using the off-the-shelf retriever for ease of joint training of both the retriever and the response generator, without the knowledge supervision. Furthermore, the training scheme of ELVM modifies the ELBO for the marginal likelihood to effectively perform the joint training without the use of complex posterior sampling, which also eliminates the discrepancy between the training-time and test-time knowledge samplings. ELVM empirically demonstrates that it significant outperforms over previous latent variable methods with in-depth analysis on diverse dialogue datasets. + +# Limitations + +There are a number of limitations and possible directions to future works of ELVM proposed in this paper. First, as more large language models beyond the scale of hundreds of billion parameters are emerging, there are rooms for emergent behaviors in ELVM as the sizes of data and model grow that + +has not been tested in this paper. Second, instead of the prior sampling that we use, fast posterior sampling such as short-run MCMC (Erik et al., 2019; Xu et al., 2023) is also an alternative candidate. Third, the experiments in this paper retrieve documents from confined set (e.g., Wikipedia) using light retriever such as BM25 and dense retriever. However, real-world applications often rely on the web which has much bigger pool of documents from broader open-domain. In turn, we should consider future extension of this work to utilizing larger public search engines. Last, we may consider alternative yet more effective means of query sampling during training other than a small set of random samples, as the current design. Other possible methods include imposing additional guidance for query sampling or multiple refinements of query sampling and knowledge retrieval for more diverse set. + +# Ethics Statement + +This work presents no direct ethical issues. However, there is a possibility that the generated responses may inadvertently contain inadequate, biased, or discriminatory content. We do not introduce new datasets, and all experiments are conducted using publicly available datasets. + +# Acknowledgments + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University). In addition, this work was also supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea) Gwangju Metropolitan City. Finally, we would also like to thank Brain Cloud Team at Kakao Brain for their support. + +# References + +Leonard Adolphs, Benjamin Boerschinger, Christian Buck, Michelle Chen Huebscher, Massimiliano Ciaramita, Lasse Espeholt, Thomas Hofmann, Yannic Kilcher, Sascha Rothe, Pier Giuseppe Sessa, et al. 2021a. Boosting search engines with interactive agents. arXiv preprint arXiv:2109.00527. + +Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021b. Reason first, then respond: Modular generation for knowledge-infused dialogue. arXiv preprint arXiv:2111.05204. + +Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Online. Association for Computational Linguistics. + +Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165. + +Mingzhu Cai, Siqi Bao, Xin Tian, Huang He, Fan Wang, and Hua Wu. 2023. Query enhanced knowledge-intensive conversation via unsupervised joint modeling. In ACL. + +Xiuyi Chen, Fandong Meng, Peng Li, Feilong Chen, Shuang Xu, Bo Xu, and Jie Zhou. 2020. Bridging the gap between prior and posterior knowledge selection for knowledge-grounded dialogue generation. In EMNLP, pages 3426-3437. + +Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Oleg Rokhlenko, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering. arXiv preprint arXiv:2210.15777. + +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling + +language modeling with pathways. arXiv preprint arXiv:2204.02311. +Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241. +Nijkamp Erik, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. 2019. Learning non-convergent non-persistent short-run mcmc toward energy-based model. In Advances in Neural Information Processing Systems (NeurIPS). +Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In NAACL. +Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909. +Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration. +Xinxian Huang, Huang He, Siqi Bao, Fan Wang, Hua Wu, and Haifeng Wang. 2021. PLATO-KAG: Unsupervised knowledge-grounded conversation via joint modeling. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, pages 143-154, Online. Association for Computational Linguistics. +Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282. +Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. 1999. An introduction to variational methods for graphical models. Machine learning, 37:183-233. +Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906. +Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR. +Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search. +Diederik P Kingma and Max Welling. 2013. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114. + +Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8460-8478, Dublin, Ireland. Association for Computational Linguistics. +Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. +Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474. +Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2021. Zero-resource knowledge-grounded dialogue generation. +Zekang Li, Cheng Niu, Fandong Meng, Yang Feng, Qian Li, and Jie Zhou. 2019. Incremental transformer with deliberation decoder for document grounded conversations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 12-21, Florence, Italy. Association for Computational Linguistics. +Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. arXiv preprint arXiv:1902.04911. +Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909. +Shilei Liu, Xiaofeng Zhao, Bochao Li, Feiliang Ren, Longhui Zhang, and Shujuan Yin. 2021. A Three-Stage Learning Framework for Low-Resource Knowledge-Grounded Dialogue Generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. +Shuman Liu, Hongshen Chen, Zhaochun Ren, Yang Feng, Qun Liu, and Dawei Yin. 2018. Knowledge diffusion for neural dialogue generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), + +pages 1489-1498, Melbourne, Australia. Association for Computational Linguistics. +Longxuan Ma, Weinan Zhang, Runxin Sun, and Ting Liu. 2020. A compare aggregate transformer for understanding document-grounded dialogue. arXiv preprint arXiv:2010.00190. +Nikita Moghe, Siddhartha Arora, Suman Banerjee, and Mitesh M Khapra. 2018. Towards exploiting background knowledge for building conversation systems. arXiv preprint arXiv:1809.08205. +Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2017. MS MARCO: A human-generated MAchine reading COprehension dataset. +Rodrigo Nogueira and Kyunghyun Cho. 2017. Task-oriented query reformulation with reinforcement learning. arXiv preprint arXiv:1704.04572. +Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In ICLR. +Gonçalo Raposo, Rui Ribeiro, Bruno Martins, and Luísca Coheur. 2022. Question rewriting? assessing its importance for conversational question answering. In Lecture Notes in Computer Science, pages 199-206. Springer International Publishing. +Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics. +Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics. +Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389. +Kurt Shuster, Mojtaba Komeili, Leonard Adolphs, Stephen Roller, Arthur Szlam, and Jason Weston. 2022a. Language models that seek for knowledge: Modular search & generation for dialogue and prompt completion. arXiv preprint arXiv:2203.13224. +Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya Chen, Kushal Arora, Joshua Lane, et al. 2022b. + +Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage. arXiv preprint arXiv:2208.03188. +Weiwei Sun, Pengjie Ren, and Zhaochun Ren. 2023. Generative knowledge selection for knowledge-grounded dialogues. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2077-2088, Dubrovnik, Croatia. Association for Computational Linguistics. +Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, abs/2201.08239. +David Thulke, Nico Daheim, Christian Dugast, and Hermann Ney. 2021. Efficient retrieval augmented generation from unstructured knowledge for task-oriented dialog. arXiv preprint arXiv:2102.04643. +Megan Ung, Jing Xu, and Y-Lan Boureau. 2022. SaFeR-Dialogues: Taking feedback gracefully after conversational safety failures. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6462-6481, Dublin, Ireland. Association for Computational Linguistics. +Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021. Question rewriting for conversational question answering. In Proceedings of the 14th ACM international conference on web search and data mining, pages 355-363. +Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 921-930. +Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, Hannaneh Hajishirzi, Mari Ostendorf, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning. In EMNLP. + +Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966. + +Jing Xu, Arthur Szlam, and Jason Weston. 2022a. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics. + +Jing Xu, Megan Ung, Mojtaba Komeili, Kushal Arora, Y-Lan Boureau, and Jason Weston. 2022b. Learning new skills after deployment: Improving open-domain internet-driven dialogue with human feedback. arXiv preprint arXiv:2208.03270. + +Yan Xu, Deqian Kong, Dehong Xu, Ziwei Ji, Bo Pang, Pascale Fung, and Ying Nian Wu. 2023. Diverse and faithful knowledge-grounded dialogue generation via sequential posterior inference. + +Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 1933-1936. + +Haolan Zhan, Lei Shen, Hongshen Chen, and Hainan Zhang. 2021. Colv: A collaborative latent variable model for knowledge-grounded dialogue generation. In EMNLP. + +Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics. + +Yizhe Zhang, Siqi Sun, Xiang Gao, Yuwei Fang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2022. Retgen: A joint framework for retrieval and grounded text generation modeling. In AAAI. + +Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP. + +Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018a. Common-sense knowledge aware conversation generation with graph attention. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 4623-4629. AAAI Press. + +Kangyan Zhou, Shrimai Prabhumoye, and Alan W Black. 2018b. A dataset for document grounded conversations. In EMNLP. + +A Training Details + +
ParameterValue
Training Steps10,000
Batch Size128
OptimizerAdamW
Learning Rate (LR)5e-5
LR SchedulerCosine
Adam β10.9
Adam β20.999
Adam ε1e-8
Warmup Steps1000
+ +Table 13: Hyperparameters used for stage-1 training. $k$ indicates number of retrieved documents. + +
WoWQReCC
Train Epochs31
Batch Size6464
OptimizerAdamWAdamW
Learning Rate (LR)1e-45e-4
LR SchedulerCosineCosine
Adam β10.90.9
Adam β20.9990.999
Adam ε1e-81e-8
Warmup Steps1000
Min Query Length11
Max Query Length3232
Top-k100100
Top-p0.950.5
Temperature1.02.0
+ +Table 14: Hyperparameters used for training ELVM and its. + +Table 13 we provide hyperparameters utilized during this training process. We also report hyperparameters used for training ELVM and its variants model (ELVM-OK and ELVM-EM-Like) in Table 14. In addition, during training of QReCC, we adopt a strategy to decrease the knowledge pool size from 54M to 1M. This reduction aims to mitigate latency issues during document retrieval. Concretely, the 1M pool is constructed by initially gathering the ground-truth relevant passage for each instance in the training dataset then adding randomly sampled documents. During the evaluation stage, we utilize the original full-size knowledge pool of 54M passages. + +# B Upper Bound Performance + +We believe in the value of understanding the utmost capability in an ideal scenario. To this end, we pursue an upper-bound performance analysis for both the WoW and QReCC tasks. This approach + +
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
ELVM14.614.711.929.313.918.014.511.829.213.7
ELVM-w-GT-Doc9.722.118.540.923.910.521.517.940.723.3
+ +Table 15: Performance comparison in WoW in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc). + +
ModelF1 ↑EM ↑
Q Rewriting (Raposo et al., 2022)18.91.0
DPR-IHN (Kim and Kim, 2022)30.44.7
QKConv (Cai et al., 2023)33.55.9
ELVM36.56.2
ELVM-w-GT-Doc55.416.2
+ +involves training our response generator under the assumption that the model has access to the ground-truth document during both training and evaluation. The performance metrics of this specific setup are designated as "ELVM-w-GT-Doc" in Table 15 and Table 16. From both tables, it's evident that there still remains a margin for improvement for both tasks. + +# C Document Recall Performance + +Table 16: Performance comparison in QReCC in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc). + +
Query fromR@1R@5R@10
Dialogue History33.458.668.7
Last Utterance25.550.862.6
R2C2-PT17.433.039.5
ELVM-β=030.655.865.7
ELVM35.259.968.6
ELVMDR37.565.675.1
+ +We report the document recall performance on QReCC, including different types of queries, as show in Table 17. Our proposed model, ELVM, surpasses both its variants and non-model generated queries such as using dialogue context as a query. The document recall performance on QReCC, encompassing various query types, is presented in Table 17. Our proposed ELVM model outperforms its variants as well as non-model generated queries, including the use of dialogue context as a query. Moreover, incorporating the dense retriever for reranking documents $(\mathrm{ELVM}_{\mathrm{DR}})$ leads to further improvements in recall. + +Table 17: Query performance comparison in QReCC for different types of queries. + +
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑R@10 ↑B4 ↑R2 ↑R@5 ↑R@10 ↑
k=511.913.968.480.311.813.768.380.3
k=5Bulk9.410.962.773.98.29.547.261.9
k=1010.312.264.876.410.612.764.876.9
k=10Bulk10.212.262.174.19.010.547.261.6
+ +Table 18: Effect of increasing the document pool size to 1000 on the performance of ELVM. As before, $k$ indicates the number of retrieved documents used during train and inference. + +# D Knowledge Pool Scaling + +In order to create a more realistic and challenging retrieval scenario, we expand the original setting of the WoW dataset by increasing the number of relevant documents per instance to 1000. This modification better reflects real-world information retrieval scenarios where a vast array of both relevant and irrelevant documents are typically encountered. We construct this extended dataset by randomly augmenting $1000 - K$ documents from the WoW dataset to the document pool for each instance, where $K$ represents the number of annotated relevant documents for each instance. On average, $K \approx 20$ for each instance in WoW dataset. + +Results in Table 18 show performances of ELVM models with different number of retrieved documents, $k$ , and knowledge pool size, where models with increased knowledge pool has tailing keyword Bulk at then end of their names. Despite the increase in complexity of the retrieval task, ELVM demonstrates robust performance, with only a marginal drop in metrics compared to the original setting. Furthermore, it is important to highlight that even under these more difficult conditions with 1000 relevant documents, ELVM still outperforms SPI by a significant margin. + +# E Effect of Parameter Sharing + +In our ELVM training, parameters are shared between the query and response generators. To probe this design choice, we decoupled these parameters, resulting in the ELVM-Decouple configuration, as detailed in Table 19. Upon evaluation on WoW's test seen and unseen tasks, ELVM- + +
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
ELVM14.611.913.918.011.813.7
ELVM-Decouple16.212.614.521.612.214.0
+ +Table 19: Performance comparison on the WoW dataset using the ELVM algorithm. ELVM denotes training with shared parameters between query and response generators, while ELVM-Decouple denotes a non-shared parameter approach. + +
ModelSeenUnseen
F1 ↑KF1 ↑F1 ↑KF1 ↑
CoLV20.318.218.517.5
KnownGPT22.023.820.522.1
SPI----
R2C2-PT15.420.515.722.3
ELVM-OK23.033.022.933.0
ELVM-EM-Like24.533.223.533.5
ELVM27.240.827.040.3
+ +Table 20: F1 and KF1 scores for ELVM and baseline models. + +Decouple showed a slight edge over the standard ELVM. Nevertheless, the benefits of shared parameters—particularly regarding computational efficiency and knowledge transfer during training and inference—cannot be understated. + +# F Additional Metrics + +We report the evaluation of our models on the WoW task using both F1 and knowledge F1 (KF1) metrics in Table 20. For reference, values pertaining to previous models are derived from Sun et al. (2023). + +# G Human Evaluation + +For a comprehensive evaluation, we conduct the human evaluation between ELVM and KnownGPT on seen and unseen tasks in WoW (Dinan et al., 2018). We randomly select 50 samples (25 samples per task), and each sample is evaluated by 11 different human experts. In specific, the two generated responses from each model in the same context are assigned to the annotators. For A/B testing, we give one score to the model if its response is received an equally good or better than the other one. Figure 2 shows the annotator instructions for the two aspects. + +# H Generated Samples + +Generated samples for the WoW test unseen are presented in Table 21 and Table 22. Similarly, Table 23 and Table 24 showcase the generated samples for the QReCC test. + +
Dialogue ContextHey, do you like cardigans? I like the modern version without no buttons and hangs open by design. Me too. My fav is the waterfall style but it's going out of fashion now I think [...] Cardigan (sweater) By contrast, a pullover does not open in front but must be "pulled over" the head to be worn. [...] But the pullover does not open in front but must be "pulled over" the head to be worn.
GT knowledge
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: i like the hand-knitted version. it's more comfortable. i like mine made of a cardigan.
R2C2-PTQuery: fallen fashion GT knowledge selected: False Response: I'm sorry, I didn't mean to offend you.
ELVM-β = 0Query: fallen fashion cardigans GT knowledge selected: True Response: I like the one that is tied is instead considered a robe.
ELVM w/ BM25Query: fallen waterfall cardigan style fallout cardigans GT knowledge selected: True Response: I like the pullover style that does not open in front but must be "pulled over" the head to be worn.
ELVM w/ DRQuery: fallen waterfall cardigan GT knowledge selected: True Response: I like the pullover that does not open in front but must be "pulled over" the head to be worn.
Dialogue Context[...] Was she diagnosed as a child? It is much easier to diagnose children, since the symptoms begin occurring early in life
GT knowledgeYes it started when she was 6 years old. She was 23 when I met her she was already recieving treatment [...] These fellowships include pediatric neurosurgery, trauma/neurocritical care, functional and stereotactic surgery, surgical neuro-oncology, radiosurgery, neurovascular surgery, skull-base surgery, peripheral nerve and spine surgery. [...] My first fellowship was in pediatric neurosurgery. It is great that she is doing better.
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: did she cry during the periods of depression? she was crying, a negative outlook on life, and poor eye contact with others.
R2C2-PTQuery: types of epilepsy GT knowledge selected: False Response: I am glad to hear she is receiving treatment. Dementia is such a devastating disease.
ELVM-β = 0Query: types of epilepsy GT knowledge selected: False Response: I am glad she is receiving treatment. Dementia has a significant effect on a person's caregivers.
ELVM w/ BM25Query: neurosurgery epilepsy neurocognitive neurosciences GT knowledge selected: True Response: I am glad to hear that she is receiving treatment. Neurosurgery is a branch of medicine that specializes in the treatment of epilepsy.
ELVM w/ DRQuery: neurosurgery treatment for epilepsy GT knowledge selected: True Response: Did she undergo a fellowship? Many neurosurgeons pursue additional training in the form of a fellowship, after residency or in some cases, as a senior resident.
+ +Table 21: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response. + +
Dialogue ContextDo you know anything about the band called Insane Clown Posse? +Insane Clown Posse (ICP) is an American hip hop duo composed of Violent J, and Shaggy 2 Dope. [...] +That's quite a lot of albums, what else do you know about ICP? +[...] The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas" and "Big Money Rustlas". [...] +The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas
GT knowledge
GT Response
KnownedGPTResponse: Query: n/a +GT knowledge selected: False
Response: the duo has earned two platinum and five gold albums. they have also earned two platinum and five gold albums
R2C2-PTQuery: ICP wiki +GT knowledge selected: False
Response: The duo has earned two platinum ,with 5 golden album
ELVM-β = 0Query: ICP facts +GT knowledge selected: False
Response: The group has been around for a long time .
ELVM w/ BM25Query: Insane clown posse ICP facts +GT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
ELVM w/ DRQuery: Insane clown posse facts +GT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
Dialogue Context[...] Athletes normally eat meat in large quantities. +Exactly. I have also added much meat eating as part of my special diet. They are rich in protein. Especially, red meat
GT knowledge[...] In nutritional science, on the other hand, red meat is defined as any meat that has more myoglobin than white meat, white meat being defined as non-dark meat from chicken (excluding leg or thigh), or fish. [...]
GT ResponseRed meat is divined as any meat that has myoglobin than white
KnownedGPTQuery: n/a +GT knowledge selected: False
Response: red meat is classified as white meat. i think it's interesting that only flesh from mammals or fowl is classified as red.
R2C2-PTQuery: protein powder +GT knowledge selected: False
Response: Athletes normally eat meat in large quantities.
ELVM-β = 0Query: red meat protein +GT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ BM25Query: red meat protein +GT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ DRQuery: red meat protein +GT knowledge selected: True
Response: Red meat is defined as any meat that has more myoglobin than white meat.
+ +Table 22: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response. + +
Dialogue ContextWhen did Mark Spitz' swimming career end? Following the Munich Olympics [...] What else did Mark Spitz do besides Speedo swimsuit ads after his swimming career? [...] However, as the memory of his feats receded, so did his endorsement and promotional deals. As his endorsements faded, he started a successful real-estate company in Beverly Hills. He bought a Ferrari and says he made [...]
GT knowledge
GT ResponseMark Spitz started a successful real-estate company in Beverly Hills
R2C2-PTQuery: What did mark spitz do after his olympic career? +GT knowledge selected: False +Response: In 1999 Spitz ranked No. 33 on ESPN SportsCentury 50 Greatest Athletes, the only aquatic athlete to make the list.
ELVM-β = 0Query: Mark Spitz Speedo Swimsuit ads after speedo swimming career ends mark spitz m Munich olympics 2012 +GT knowledge selected: False +Response: Spitz did Schick razor commercials and Speedo swimsuit ads
ELVM w/ BM25Query: Spitz endorsements after swimming career ends +GT knowledge selected: True +Response: Mark Spitz started a successful real-estate company in Beverly Hills.
Dialogue ContextWhat was the New Scheme? +The 'new scheme' and the 'new dedication' respectively, initially seen as a [...] +[...] +Did Ahmadi followers of the New Scheme act as missionaries? +As part of this Mahmood Ahmad appointed 5 men to survey the Punjab [...] +Did Mirza Basheer-ud-Din Mahmood Ahmad use missionaries to try to spread the New Scheme to foreign countries?
GT knowledge[...] In time the scheme produced a vast amount of literature in defence of Islam in general and the Ahmadiyya beliefs in particular. The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance. As part of this Mahmood Ahmad appointed [...]
GT ResponseThe funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
R2C2-PTQuery: Mirza Basher-ud Din Mahmood Ahmad missionaries +GT knowledge selected: False +Response: Mirza Basheer-ud-Din Mahmood Ahmad used missionaries to try to spread the New Scheme to foreign countries.
ELVM-β = 0Query: New Scheme Mahmud Ahmad missionaries to spread New Scheme in foreign countries +GT knowledge selected: True +Response: Yes, I have. I have experienced heartache, also known as loss of a loved one.
ELVM w/ BM25Query: New scheme missionaries used to spread Ahmadiyyah teachings in foreign countries +GT knowledge selected: True +Response: The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
+ +Table 23: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response. + +
Dialogue ContextHow did Stephen Sondheim collaborate with James Lapine? +Stephen Sondheim found a new collaborator in [...]What did Stephen Sondheim and James Lapine do together in the 1980s? +Their first collaboration was Sunday in the Park with George (1984).Was Stephen Sondheim and James Lapine's first collaboration a success? +Stephen Sondheim and James Lapine won the 1985 Pulitzer Prize [...]Did Stephen Sondheim and James Lapine collaborate in other projects, besides Sunday in the Park with George?[...] last work together was the rhapsodic Passion (1994).Was Stephen Sondheim and James Lapine's last work a success? +Sondheim and Lapine's last work together was the rhapsodic Passion (1994), adapted from Ettore Scola's Italian film Passione D'Amore. With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.[46] Later work Assassins opened [...]With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
GT knowledge
GT Response
R2C2-PTQuery: Was the play "rhapsodic passion" a success or a failure? +GT knowledge selected: TrueResponse: Sondheim and Lapine's last work together was the rhapsodic Passion (1994).
ELVM-β = 0Query: Stephen Sonds and Lapine Rhapsodic passion1994 Stephen sondheim, James lapine plays 12 dreams off-broadway p Pulitzer prize for play Sunday in Park.GT knowledge selected: FalseResponse: Stephen Sondheim and James Lapine's last work together was the rhapsodic Passion (1994).
ELVM w/ BM25Query: Sondheim last work was the Rhapsodic passion (1994) Pulitzer prize won for Drama. Lapine wrote the play Twelve Dreams.GT knowledge selected: TrueResponse: With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
Dialogue ContextDid Frank Hague commit any questionable acts? +Yes, he took kickbacks in cash, and relied on two ordinances of dubious [...]Did Frank Hague do anything illegal? +He passed an unconstitutional ordinance that gave himself as commissioner the power to [...]Did anyone attempt to prosecute Frank Hague?[...] Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. [38] He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash. [...]Yes, but he was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
GT knowledge
GT Response
R2C2-PTQuery: Frank Hague prosecutionGT knowledge selected: FalseResponse: No, no one tried to prosecute Frank Hague.
ELVM-β = 0Query: Frank Hague ordinance ordinance frank Hague charges Frank Hague Hague for flouting constitutionality ordinancesGT knowledge selected: FalseResponse: Hague was found guilty of contempt of court for ignoring the subpoena. He was fined $100 and stripped of his duties as Deputy Sheriff.
ELVM w/ BM25Query: Hague prosecution attempt failed after Hague passed unconstitutional ordinance. no Hague was prosecuted.GT knowledge selected: TrueResponse: Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
+ +Table 24: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response. + +# Human Evaluation Annotation Tool + +# Components of each example + +Each example may include one of more of the following + +- History: topic related to the conversation and the dialogue history. Examples may not contain 'history' +- Input text: Utterance of the opponent +- Ground truth response of the example, given as a reference. The annotator may use the reference in decision. +- Two responses from different models. + +# Annotation Standards + +The annotator marks which model of the given two is better on three different categories: Fluency and Relevance. The deifintion of each category is + +- Fluency measures the correctness, completeness, and semantic consistency of the response. +- Relevance measures the degree of relevance of the knowledge and response to the history. + +# History: + +# Singing + +I love to sing. Some of my favorite genres to sing are blues, pop, and rock. Do you like to sing? + +Yes very much so. I love older artists like Ray Charles and Billie Holiday. + +Oh yeah I love Billie Holiday. She had no formal music training but her vocal delivery made up for it. + +# Input Text: + +I've been getting into the strokes as of late. Do you like them? + +# Reference: + +I used to listen to The Strokes a lot back in high school. Their debut album was added to Rolling Stones 500 greatest albums of all time at number 199. + +# Model A Response: + +Yes I do like them. I like all kinds of music, but I'm not a big fan of the Strokes. + +# Model B Response: + +I do like them. I also like to wear my Strokes shirt during May because it is national stroke awareness month. + +# Annotation + +# Fluency + +Both responses are fluent. +$\bigcirc$ Model A response is more fluent. +Model B response is more fluent. +$\bigcirc$ Neither response is fluent. + +Previous + +# Relevance + +Both knowledges and responses are relevant +Model A response is more relevant. +Model B response is more relevant. +$\bigcirc$ Neither is relevant. + +Next + +Made with Streamlit + +Figure 2: Screenshot of the annotation tool. \ No newline at end of file diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/images.zip b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e465383d861b6919372365ca120227abfa8c53bf --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d44808006c2cf66432f286aa1421d32ec14a5bb2fd73ec6c10161b29c88e17f +size 2038494 diff --git a/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/layout.json b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2fd9eb2145e381bb51f9bbf38879c34fac496b5d --- /dev/null +++ b/2023/Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation/layout.json @@ -0,0 +1,13066 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 162, + 71, + 431, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 71, + 431, + 104 + ], + "spans": [ + { + "bbox": [ + 162, + 71, + 431, + 104 + ], + "type": "text", + "content": "Efficient Latent Variable Modeling for Knowledge-Grounded Dialogue Generation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "spans": [ + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": "Gunsoo Han" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Daejin Jo" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Daniel Wontae Nam" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Eunseop Yoon" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Taehwan Kwon" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Seungeun Rho" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Kyoung-Woon On" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Chang D. Yoo" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "text", + "content": " Sungwoong Kim" + }, + { + "bbox": [ + 79, + 113, + 524, + 142 + ], + "type": "inline_equation", + "content": "^{5\\dagger}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "spans": [ + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "content": "Kakao Brain " + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "content": "KAIST " + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "content": "KRAFTON " + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "content": "Georgia Institute of Technology " + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 68, + 144, + 525, + 159 + ], + "type": "text", + "content": "Korea University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 169, + 441, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 169, + 441, + 182 + ], + "spans": [ + { + "bbox": [ + 151, + 169, + 441, + 182 + ], + "type": "text", + "content": "gunsoo.han@kakaobrain.com swkim01@korea.ac.kr" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 237, + 274, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 237, + 274, + 589 + ], + "spans": [ + { + "bbox": [ + 85, + 237, + 274, + 589 + ], + "type": "text", + "content": "Knowledge-grounded dialogue generation requires to first retrieve appropriate external knowledge based on a conversational context and then generate a response grounded on the retrieved knowledge. In general, these two sequential modules, a knowledge retriever and a response generator, have been separately trained by supervised data for each module. However, obtaining intermediate labels of the ground-truth knowledge is expensive and difficult especially in open-domain conversation. Latent variable modeling can circumvent it and enables a joint training without the knowledge supervision. In this paper, we propose an efficient algorithm for this latent variable modeling that is able to leverage a large amount of dialogue data. In specific, rather than directly training the complex retriever, we adapt a query generator with an off-the-shelf retriever, and the query generator and response generator are simultaneously trained over the latent variable of query. Moreover, we employ the evidence lower bound as a training objective and modify it to efficiently and robustly perform the joint training. Experimental results on diverse knowledge-grounded dialogue datasets show that the proposed algorithm achieves state-of-the-art performances even without the use of the annotated knowledge while maintaining the efficiency and scalability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 601, + 154, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 601, + 154, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 601, + 154, + 613 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 623, + 291, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 623, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 623, + 291, + 743 + ], + "type": "text", + "content": "Recently, knowledge-grounded dialogue generation has drawn increasing attention especially as a key ingredient for open-domain conversational agents (Dinan et al., 2018; Zhou et al., 2018b; Zhan et al., 2021; Liu et al., 2018; Zhou et al., 2018a; Lian et al., 2019; Zhao et al., 2020; Kim et al., 2020; Chen et al., 2022; kom; Shuster et al., 2022a,b; Cai et al., 2023; Wu et al., 2022; Lewis et al., 2020; Huang et al., 2021; Thulke et al., 2021; Anantha" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 212, + 526, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 526, + 306 + ], + "type": "text", + "content": "et al., 2021). It usually obtains knowledge from external resources such as Wikipedia-based databases and web search engines since internal knowledge even in large-scale parametric language models (Brown et al., 2020; Thoppilan et al., 2022; Chowdhery et al., 2022) is incomplete or outdated, and moreover it can provide hallucinated information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 307, + 526, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 307, + 526, + 625 + ], + "spans": [ + { + "bbox": [ + 302, + 307, + 526, + 625 + ], + "type": "text", + "content": "In order for a dialogue response to be grounded on such external knowledge, the conversational agent generally consists of a knowledge retriever, which retrieves knowledge corresponding to a given dialogue context, followed by a response generator that produces an informative response based on the dialogue context and the retrieved knowledge. In many previous methods, supervised learning has often been applied to independently optimize each module using the ground-truth or gold knowledge (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). However, human annotation of knowledge information is cumbersome, expensive, and often incomplete due to the existence of multiple possible knowledge candidates and the overlooking of the target response. In addition, existing automatic annotations are generally limited to a simple extractive question answering (Zhao et al., 2020; Liu et al., 2021). This difficulty in obtaining annotations of knowledge information could be severe under the open-domain conversation and hinders the use of large-scale dialogue data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 627, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 627, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 627, + 527, + 772 + ], + "type": "text", + "content": "Therefore, there have been a number of recent approaches to learn the knowledge retriever and the response generator without the knowledge supervision. In specific, they have treated the retrieved knowledge, document, or passage as an unobserved latent variable and adapt latent variable modeling based on approximated marginalization (e.g. top-k) (Lewis et al., 2020; Huang et al., 2021; Cai et al., 2023; Guu et al., 2020), reinforcement learning (Zhao et al., 2020; Zhang et al., 2022; Chen et al., 2022; Wu et al., 2022) or variational methods (Zhan" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 83, + 750, + 190, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 750, + 190, + 761 + ], + "spans": [ + { + "bbox": [ + 83, + 750, + 190, + 761 + ], + "type": "text", + "content": "* Work done at Kakao Brain" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "spans": [ + { + "bbox": [ + 83, + 761, + 175, + 772 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2683" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2683-2702 \nDecember 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 310 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 310 + ], + "type": "text", + "content": "et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020; Chen et al., 2020; Xu et al., 2023). However, joint training of the retriever along with the generator under this latent variable modeling has some restrictions in utilizing the retriever. For example, a retriever needs to produce a differentiable prior probability for the gradient propagation through the marginalization or to be intermittently updated to rebuild the whole passage index during training (Karpukhin et al., 2020; Lewis et al., 2020; Zhang et al., 2022). This leads to limitation in facilitating complicated or large-scale retrievers and further in leveraging large-scale data. In addition, the posterior probability used in the previous variational methods has been modeled by a separated network, which grows the training complexity and also incurs the discrepancy between the training-time and test-time knowledge samplings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 314, + 291, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 314, + 291, + 473 + ], + "spans": [ + { + "bbox": [ + 67, + 314, + 291, + 473 + ], + "type": "text", + "content": "To overcome such restrictions, in this paper, we propose an efficient latent variable modeling, named " + }, + { + "bbox": [ + 67, + 314, + 291, + 473 + ], + "type": "inline_equation", + "content": "ELVM" + }, + { + "bbox": [ + 67, + 314, + 291, + 473 + ], + "type": "text", + "content": ", for knowledge-grounded dialogue generation. In particular, we reduce the burden for training a whole retriever by employing a query generator followed by an off-the-shelf retriever that is fixed during training. More precisely, in tandem with the response generator, the query generator rather than the retriever is jointly optimized by latent variable modeling, in which a query is taken as a latent variable, on paired observations of dialogue contexts and responses." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 477, + 291, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 477, + 291, + 716 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 291, + 716 + ], + "type": "text", + "content": "To this end, we exploit the variational method using the evidence lower bound (ELBO) (Kingma and Welling, 2013; Jordan et al., 1999) and approximate the expected conditional likelihood in the ELBO by subset sampling from the prior distribution, which acts as the training objective for the response and the query generators. This approximation gets rid of an extra modeling of a surrogate posterior distribution or online posterior inference such as Markov Chain Monte Carlo (MCMC) and also reduces the training-inference discrepancy in knowledge retrieval. Moreover, we further modify the Kullback-Leibler (KL) regularization of the ELBO by constructing the approximated posterior distribution from the conditional likelihood and prior distribution and set this posterior distribution as a teacher for a distillation objective to further learn the query generator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 773 + ], + "type": "text", + "content": "Experimental results show that the proposed ELVM allows to efficiently and robustly perform training without (1) the use of the annotated knowledge, (2) an explicit training of the knowledge" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 163 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 163 + ], + "type": "text", + "content": "retrieval, and (3) a complex posterior sampling. Especially, it significantly outperforms previous state-of-the-art methods for knowledge-grounded dialogue generation. In addition, the proposed posterior distillation improves the performances over the baseline that solely maximizes the expectation of the conditional likelihood." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 165, + 521, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 521, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 521, + 178 + ], + "type": "text", + "content": "Our main contributions can be summarized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 181, + 527, + 379 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 304, + 181, + 527, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 527, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 527, + 246 + ], + "type": "text", + "content": "- An efficient latent variable modeling is proposed in joint training of query generator and dialogue response generator without the knowledge supervision for knowledge-intensive dialogue generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 248, + 527, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 527, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 527, + 313 + ], + "type": "text", + "content": "- For realizing efficient yet robust joint training, the ELBO for the marginal likelihood is modified as the combination of the conditional likelihood objective and the posterior distillation objective, based on multiple prior samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 313, + 527, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 313, + 527, + 379 + ], + "spans": [ + { + "bbox": [ + 303, + 313, + 527, + 379 + ], + "type": "text", + "content": "- The proposed ELVM demonstrates its effectiveness in performing unsupervised joint training and even significant performance improvements over previous methods, being a new state-of-the-art on benchmark datasets." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 303, + 395, + 396, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 395, + 396, + 407 + ], + "spans": [ + { + "bbox": [ + 303, + 395, + 396, + 407 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 421, + 479, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 421, + 479, + 446 + ], + "spans": [ + { + "bbox": [ + 302, + 421, + 479, + 446 + ], + "type": "text", + "content": "2.1 Knowledge-Grounded Dialogue Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 454, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 454, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 454, + 526, + 772 + ], + "type": "text", + "content": "Knowledge-grounded conversation has been extensively studied through recently released public datasets (Dinan et al., 2018; kom; Xu et al., 2022b; Kwiatkowski et al., 2019; Anantha et al., 2021; Moghe et al., 2018). Existing approaches for this task have mostly exploited two successive modules, the knowledge retriever and the knowledge-grounded response generator. Many previous works have used manual annotations of gold knowledge provided by some of public datasets to optimize the modules (Dinan et al., 2018; Shuster et al., 2022a,b; Glass et al., 2022; Adolphs et al., 2021a; Nogueira and Cho, 2017; Ma et al., 2020; Xie et al., 2022). Specifically, Seeker (Shuster et al., 2022a) and BlenderBot3 (Shuster et al., 2022b) have recently proposed to build a series of modules by a single transformer with different prompts for each module, and trained its transformer on a large number of modular tasks using annotated datasets. However, this manual annotation is expensive, time-consuming, and often inaccurate, which impedes the utilization of a large-scale dialogue data that is especially necessary for open-domain conversation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "2684" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 230, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 230, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 230, + 84 + ], + "type": "text", + "content": "2.2 Unsupervised Joint Training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 84, + 291, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 84, + 291, + 256 + ], + "spans": [ + { + "bbox": [ + 67, + 84, + 291, + 256 + ], + "type": "text", + "content": "Recently, unsupervised training methods have been widely applied, and many of them have tries to jointly learn the modules without knowledge labels (Lewis et al., 2020; Huang et al., 2021; Zhao et al., 2020; Zhang et al., 2022; Zhan et al., 2021; Paranjape et al., 2022; Lian et al., 2019; Kim et al., 2020). For example, PLATO-KAG (Huang et al., 2021) has approximated the marginal likelihood by top-k knowledge samples while RetGen (Zhang et al., 2022) has trained to reward knowledge retrieval with the highest utilization in response generation by reinforcement learning and mixture-of-experts ensembling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 257, + 291, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 257, + 291, + 481 + ], + "spans": [ + { + "bbox": [ + 69, + 257, + 291, + 481 + ], + "type": "text", + "content": "Meanwhile, latent variable modeling based on variational methods has also been developed by several recent works. CoLV (Zhan et al., 2021) has introduced collaborative latent spaces to reflect the inherent correlation between the knowledge selection and the response generation. SKT (Kim et al., 2020) has developed a sequential latent variable model for the multi-turn knowledge-grounded dialogue generation. In order to perform the posterior sampling of knowledge selection during joint training, some works have proposed to separately train the posterior distribution model (Paranjape et al., 2022; Lian et al., 2019) or the posterior information prediction model (Chen et al., 2020). Very recently, SPI (Xu et al., 2023) has applied short-run MCMC (Erik et al., 2019) for posterior sampling on the collaborative latent spaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 482, + 291, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 482, + 291, + 721 + ], + "spans": [ + { + "bbox": [ + 67, + 482, + 291, + 721 + ], + "type": "text", + "content": "While these latent variable modeling algorithms can effectively perform unsupervised joint training, the entire training of the retriever is still difficult and imposes restrictions on the selection of the retriever in terms of both a search algorithm and a knowledge resource. Furthermore, the additional posterior sampling through separate networks or multiple iterations in the previous variational methods also leads to increased training complexity as well as the training-inference discrepancy in knowledge generation. In contrast, the proposed ELVM can employ any kind of retriever in latent variable modeling since it controls a retrieval output by only changing of a generated query. Moreover, ELVM removes an extra posterior modeling or sampling by prior subset sampling and approximated posterior distillation, which leads to efficient training without the discrepancy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 729, + 182, + 742 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 182, + 742 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 182, + 742 + ], + "type": "text", + "content": "2.3 Query Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "content": "When an off-the-shelf retriever is used for knowledge retrieval, how to make an input text query" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 243 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 243 + ], + "type": "text", + "content": "is important for obtaining appropriate documents. Especially, in knowledge-grounded dialogue, a self-contained query should be generated from the multi-turn dialogue context. A number of prior works have tried to train a supervised query rewriting model by human rewrites (Yu et al., 2020; Lin et al., 2020; Vakulenko et al., 2021; Voskarides et al., 2020). Similar to the knowledge annotation, the limitations of manual query annotation have come up with unsupervised learning of the query generator. A number of works have proposed a novel rewards for reinforcement learning of a query rewritten (Wu et al., 2022; Chen et al., 2022)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 244, + 527, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 527, + 403 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 527, + 403 + ], + "type": "text", + "content": "QKConv (Cai et al., 2023) has proposed an unsupervised query enhanced method for knowledge-grounded conversation. This work is similar to ours in that it consists of a query generator, an off-the-shelf knowledge retriever, and a response generator, and unsupervised joint training is applied to the query generator and the response generator. However, their training objective is based on the approximated marginal likelihood over candidate queries, which is different from our objective that includes the posterior distillation based on the ELBO loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 416, + 505, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 416, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 302, + 416, + 505, + 456 + ], + "type": "text", + "content": "3 Efficient Latent Variable Model for Knowledge-Grounded Dialogue Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 460, + 358, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 460, + 358, + 473 + ], + "spans": [ + { + "bbox": [ + 303, + 460, + 358, + 473 + ], + "type": "text", + "content": "3.1 Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "content": "In this section, we explain an overall structure of our ELVM, as described in Figure 1a, and define notations. Unlike previous works of document (passage) retrieval (Zhang et al., 2022; Huang et al., 2021; Lewis et al., 2020) using a query of an embedded vector of a dialogue context " + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "content": ", we use a text query of natural language, which enables to utilize any kind of retriever for dialogue response generation. We define the probability of generating a natural language query " + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "content": " for a given dialogue context " + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "inline_equation", + "content": "p_{\\phi}(u|x)" + }, + { + "bbox": [ + 302, + 476, + 526, + 622 + ], + "type": "text", + "content": ", as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 634, + 525, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 634, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 342, + 634, + 525, + 661 + ], + "type": "interline_equation", + "content": "p _ {\\phi} (u | x) = \\prod_ {t} p _ {\\phi} \\left(u _ {t} \\mid u _ {0: t - 1}, x\\right), \\tag {1}", + "image_path": "a8621a7003d48000c88df7328f9398fcb16e7678244e1232e961f1507ea5045e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "spans": [ + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "u_{t}" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": "-th token in " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " is the probability obtained from the model parameterized by " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": ". Similarly, the probability of generating a response " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " for a given query " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " and dialogue context " + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 302, + 672, + 526, + 737 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 748, + 525, + 775 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 748, + 525, + 775 + ], + "spans": [ + { + "bbox": [ + 310, + 748, + 525, + 775 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (y | Z (u), x) = \\prod_ {t} p _ {\\theta} \\left(y _ {t} \\mid y _ {0: t - 1}, Z (u), x\\right), \\tag {2}", + "image_path": "1835d16a803fd4d410545ef839a8a4c4ef71cdf411fcc023193312bc10f4e933.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2685" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 71, + 399, + 270 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 399, + 270 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 399, + 270 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 399, + 270 + ], + "type": "image", + "image_path": "4cfe6d0673e496a8a96875adf2b2e3dfa927b4c687d9710d9bafbea2183e7fe0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 275, + 293, + 285 + ], + "lines": [ + { + "bbox": [ + 179, + 275, + 293, + 285 + ], + "spans": [ + { + "bbox": [ + 179, + 275, + 293, + 285 + ], + "type": "text", + "content": "(a) Overall structure of ELVM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 415, + 69, + 521, + 262 + ], + "blocks": [ + { + "bbox": [ + 415, + 69, + 521, + 262 + ], + "lines": [ + { + "bbox": [ + 415, + 69, + 521, + 262 + ], + "spans": [ + { + "bbox": [ + 415, + 69, + 521, + 262 + ], + "type": "image", + "image_path": "3c140f2c8b4e80b181f7d1a434c3dbfa3a15da0576b0dd08f773430c18f14263.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 264, + 522, + 285 + ], + "lines": [ + { + "bbox": [ + 414, + 264, + 522, + 285 + ], + "spans": [ + { + "bbox": [ + 414, + 264, + 522, + 285 + ], + "type": "text", + "content": "(b) Example inference of ELVM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "lines": [ + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "spans": [ + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": "Figure 1: Overall structure and example inference case of ELVM. During training, for a given dialogue context " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": ", the query generator generates a set of unique queries " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "Q^s" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " and for each query " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " a set of top- " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " documents " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "Z(u)" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " is retrieved using the off-the-shelf retriever. Then, response generator utilizes both " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "Z(u)" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": ": (a) The approximated marginal likelihood " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta, \\phi}(y|x)" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " used during training is defined by " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "Q^s" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " and the re-normalized prior " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "\\bar{p}_{\\phi}(u|x)" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": "; (b) At inference time, a single sampled query " + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 295, + 526, + 355 + ], + "type": "text", + "content": " is used for retrieving multiple documents and producing a final response." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": "-th token in " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "Z(u) = \\{z_1, z_2, \\dots, z_k\\}" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " is the set of retrieved documents (passages) by " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " is the model parameters to produce the response probability. Here, we use top- " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " result obtained by an off-the-shelf retriever such as BM25 (Robertson et al., 2009) and DPR (Karpukhin et al., 2020) and freeze the retriever. We assume that there is a deterministic mapping from " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": ". We now set a query " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " as a latent variable. Then, using the prior " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "p_{\\phi}(u|x)" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " associated with the query generator and the conditional likelihood " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|Z(u), x)" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " corresponding to the response generator, we can define the marginal likelihood of the knowledge-grounded response " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " given the dialogue context " + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 375, + 290, + 574 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 582, + 290, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 582, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 83, + 582, + 290, + 610 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q} p _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {3}", + "image_path": "92666c2330020963e9d9bafc243788e98d6a0adbc0ab16d348a9a579f16d2648.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "content": " is the set of all possible queries. Since it is intractable to marginalize over all queries, we sample a small subset composed of " + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "content": " unique queries, " + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "inline_equation", + "content": "Q^{s}(x) = \\{u_{1}, u_{2}, \\dots, u_{m}\\}" + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\cdot | x)" + }, + { + "bbox": [ + 67, + 618, + 291, + 684 + ], + "type": "text", + "content": " and then marginalize over this set as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 81, + 692, + 290, + 719 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 692, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 81, + 692, + 290, + 719 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta , \\phi} (y | x) = \\sum_ {u \\in Q ^ {s}} \\bar {p} _ {\\phi} (u | x) p _ {\\theta} (y | Z (u), x), \\tag {4}", + "image_path": "adf5489b13c1b07e6966b48f89b24268b42e3e87d7900a5444aaa33fcf621a57.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 728, + 98, + 739 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 728, + 98, + 739 + ], + "spans": [ + { + "bbox": [ + 67, + 728, + 98, + 739 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 745, + 290, + 776 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 745, + 290, + 776 + ], + "spans": [ + { + "bbox": [ + 111, + 745, + 290, + 776 + ], + "type": "interline_equation", + "content": "\\bar {p} _ {\\phi} (u | x) = \\frac {p _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {5}", + "image_path": "1934406bf7364732502a083e7595cb5467154fbd0056474473ed77df47c7350c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "spans": [ + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "content": "We construct " + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "inline_equation", + "content": "Q^s" + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "content": " by obtaining " + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "content": "-unique queries sampled from " + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\cdot | x)" + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "content": ". It is noted that our response generator takes multiple retrieved documents efficiently using FiD (Izacard and Grave, 2020), which is different from the use of a single document for response generation in previous methods. In addition, while we can utilize relative matching scores from the retriever to assign varying likelihoods to each document given a query for " + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "inline_equation", + "content": "p(Z(u)|u)" + }, + { + "bbox": [ + 302, + 375, + 526, + 615 + ], + "type": "text", + "content": ", we have noticed that this approach can occasionally hinder the effective and comprehensive learning of the query generator. During the training process, that will be described in the next subsection, our objective is to amplify the impact of the generated query. Consequently, we opt for a deterministic mapping from a query to a set of retrieved documents as a unified whole, aiming to enhance the gradient flow." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 624, + 399, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 399, + 637 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 399, + 637 + ], + "type": "text", + "content": "3.2 Joint Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 640, + 525, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 640, + 525, + 680 + ], + "spans": [ + { + "bbox": [ + 302, + 640, + 525, + 680 + ], + "type": "text", + "content": "Our training objective is the ELBO (Kingma and Welling, 2013; Jordan et al., 1999) of " + }, + { + "bbox": [ + 302, + 640, + 525, + 680 + ], + "type": "inline_equation", + "content": "\\log \\tilde{p}_{\\theta,\\phi}" + }, + { + "bbox": [ + 302, + 640, + 525, + 680 + ], + "type": "text", + "content": " which can be written as below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 309, + 690, + 524, + 724 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 690, + 524, + 724 + ], + "spans": [ + { + "bbox": [ + 309, + 690, + 524, + 724 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log \\tilde {p} _ {\\theta , \\phi} (y | x) \\geq \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ - D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right), \\tag {6} \\\\ \\end{array}", + "image_path": "68b2afd46a244b17843900446fb6ba0d81ebe9d5b9c97c97e4b5eac04bee2823.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": " is the variational posterior. Note that the equality holds when " + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": " is identical to the true posterior distribution of it. In this work," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2686" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "content": "we do not parameterize " + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "content": " and approximate the expected conditional likelihood in the ELBO by sampling " + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "content": " from not " + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "inline_equation", + "content": "\\bar{p}_{\\phi}(u|x)" + }, + { + "bbox": [ + 67, + 71, + 290, + 122 + ], + "type": "text", + "content": " such as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 135, + 289, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 135, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 86, + 135, + 289, + 168 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {u \\sim q (u | x, y)} [ \\log p _ {\\theta} (y | Z (u), x) ] \\\\ \\approx \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ]. \\tag {7} \\\\ \\end{array}", + "image_path": "f8987a8f6273e94f99a031f9caa592a7f51e948fa64d582acac4714230817db5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 177, + 290, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 177, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 290, + 269 + ], + "type": "text", + "content": "This approximation is due to that it minimizes the discrepancy between training and test time inference and it does not separately model " + }, + { + "bbox": [ + 67, + 177, + 290, + 269 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 177, + 290, + 269 + ], + "type": "text", + "content": " or complicatedly sample from it during training. Then, we obtain the following two losses corresponding to the expected conditional likelihood and the KL regularization such as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 281, + 290, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 281, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 73, + 281, + 290, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {y, u} (\\theta , \\phi) = - \\mathbb {E} _ {u \\sim \\bar {p} _ {\\phi} (u | x)} [ \\log p _ {\\theta} (y | Z (u), x) ], \\tag {8}", + "image_path": "067fe603a6276bee7b81a5389d96ffc9a49e97ccd1c3a06587fcb5a63acbaa6b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 299, + 290, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 299, + 290, + 315 + ], + "spans": [ + { + "bbox": [ + 91, + 299, + 290, + 315 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {u} (\\phi) = D _ {K L} \\left(q (u | x, y) | | \\bar {p} _ {\\phi} (u | x)\\right). \\tag {9}", + "image_path": "9dd35aff8463b60157a4b6fc4de4201b7b854acadb5aa0f513303f7f622906d3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "text", + "content": "Moreover, we define " + }, + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^u (\\phi)" + }, + { + "bbox": [ + 67, + 325, + 289, + 364 + ], + "type": "text", + "content": " by the approximated posterior derived from the prior and the likelihood such as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 373, + 289, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 373, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 77, + 373, + 289, + 413 + ], + "type": "interline_equation", + "content": "q (u | x, y) = \\frac {p _ {\\theta} (y | Z (u) , x) \\bar {p} _ {\\phi} (u | x)}{\\sum_ {u ^ {\\prime} \\in Q ^ {s}} p _ {\\theta} (y | Z (u ^ {\\prime}) , x) \\bar {p} _ {\\phi} \\left(u ^ {\\prime} \\mid x\\right)}. \\tag {10}", + "image_path": "2acbc41b83b02f4c1df485f77151cf818d17121308f13a781e1bf2e71f15520c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": "We set this posterior distribution " + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": " as a teacher for a distillation to update the prior associated with the query generation at test time. Namely, " + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^u (\\phi)" + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": " becomes the posterior distillation loss, and our prior is explicitly updated to be similar to the posterior. Here, the gradient is not propagated from " + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "inline_equation", + "content": "q(u|x,y)" + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": " can be computed easily without additional model or approximation since we define it over the subset " + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "inline_equation", + "content": "Q^s" + }, + { + "bbox": [ + 67, + 414, + 290, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "text", + "content": "To sum up, we obtain the conditional likelihood loss and the posterior distillation loss by multiple samples from the prior distribution that is aligned with query sampling at the inference time. We would mitigate the bias of prior sampling by multiple query samples for each input during training. In addition, our modification of ELBO is different from the expectation-maximization (EM) algorithm in that we sample a latent variable from the prior distribution rather than the posterior distribution that is used for sampling of E-step in EM. Moreover, in contrast to the M-step loss, we use the prior-weighted conditional likelihood and the gradient is also propagated to the sampled latents and the corresponding prior model. For experiments in this paper, the query generator " + }, + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "text", + "content": " and the response generator " + }, + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 534, + 291, + 773 + ], + "type": "text", + "content": " share parameters, with different input prompts to distinguish each module, for simplicity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 71, + 490, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 490, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 490, + 84 + ], + "type": "text", + "content": "Overall, the loss for our joint training is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 95, + 525, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 95, + 525, + 110 + ], + "spans": [ + { + "bbox": [ + 340, + 95, + 525, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta , \\phi) = \\mathcal {L} ^ {y, u} (\\theta , \\phi) + \\beta \\mathcal {L} ^ {u} (\\phi), \\tag {11}", + "image_path": "176bd3ff994e45782f82c4f28696c1b87b9747d858091b89bc5bd25996f0e075.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 121, + 524, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 121, + 524, + 147 + ], + "spans": [ + { + "bbox": [ + 302, + 121, + 524, + 147 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 121, + 524, + 147 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 302, + 121, + 524, + 147 + ], + "type": "text", + "content": " is the relative weight for the posterior distillation loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 158, + 386, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 158, + 386, + 173 + ], + "spans": [ + { + "bbox": [ + 302, + 158, + 386, + 173 + ], + "type": "text", + "content": "4 Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 180, + 424, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 180, + 424, + 194 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 424, + 194 + ], + "type": "text", + "content": "4.1 Experimental Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 198, + 525, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 198, + 525, + 344 + ], + "spans": [ + { + "bbox": [ + 302, + 198, + 525, + 344 + ], + "type": "text", + "content": "Dataset. Our experiments are conducted on two widely used knowledge-grounded dialogue datasets, Wizard of Wikipedia (WoW) (Dinan et al., 2018) and QReCC (Anantha et al., 2021). WoW is structured around dialogues that are grounded in knowledge sentences sourced from Wikipedia, consisting of roughly 18K, 2K, and 2K dialogues, for training, validation, and test subsets, respectively. QReCC is a comprehensive compilation of conversational data, consisting of 14K open-domain conversations comprising 80K question-answer pairs." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 305, + 369, + 523, + 479 + ], + "blocks": [ + { + "bbox": [ + 305, + 369, + 523, + 479 + ], + "lines": [ + { + "bbox": [ + 305, + 369, + 523, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 523, + 479 + ], + "type": "table", + "html": "
TaskDatasetTrainValid
QueryWizInt (Komeili et al., 2022)351,3752,467
Fits (xu2)3,587392
KG ResponseWizInt (Komeili et al., 2022)22,4881,687
Fits (xu2)6,279656
Ms Marco (Nguyen et al., 2017)281,63636,859
NQ Open (Adolphs et al., 2021b)79,1688.757
Dialogue ResponseWizInt (Komeili et al., 2022)8,335587
MSC (Xu et al., 2022a)105,54917,691
SaferDialogues (Ung et al., 2022)6,3067,88
PersonaChat (Zhang et al., 2018)131,4387.801
EmpatheicDialogues (Rashkin et al., 2019)64,6365,738
", + "image_path": "87698417f33a1cffae8002e1577bc3c05079eaa86f1a3dc976d87cf553d64661.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 486, + 525, + 533 + ], + "lines": [ + { + "bbox": [ + 302, + 486, + 525, + 533 + ], + "spans": [ + { + "bbox": [ + 302, + 486, + 525, + 533 + ], + "type": "text", + "content": "Table 1: List of dataset and number of train and validation examples used for fine-tuning R2C2 model to obtain R2C2-PT. KG is an abbreviation for knowledge-grounded." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 547, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 547, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 547, + 525, + 773 + ], + "type": "text", + "content": "Training. We begin with publicly available R2C2 pre-trained transformer model (Shuster et al., 2022a) with 400M parameter size and fine-tune it on multiple datasets to acquire the essential skill sets for knowledge-grounded dialogue such as query generation and knowledge-grounded response generation. Detailed tasks and statistics are shown in Table 1. In line with Shuster et al. (2022b), we adopt a modular system with a single model that incorporates control tokens into the dialogue context to perform different modules. For example, appending the control token __generate-query__ to the dialogue context promotes the model to generate relevant search query. We name the initial model from this pretraining procedure as R2C2-PT. Subsequently, we apply our proposed algorithm to R2C2-PT model," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2687" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 68, + 516, + 229 + ], + "blocks": [ + { + "bbox": [ + 79, + 68, + 516, + 229 + ], + "lines": [ + { + "bbox": [ + 79, + 68, + 516, + 229 + ], + "spans": [ + { + "bbox": [ + 79, + 68, + 516, + 229 + ], + "type": "table", + "html": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
PostKS (Lian et al., 2019)79.1--13.01.0193.8--13.11.0
ITDD (Li et al., 2019)17.84.02.516.2-44.82.11.111.4-
SKT (Kim et al., 2020)52.0--19.36.881.4--16.14.2
PIPM (Chen et al., 2020)42.7-3.319.97.365.7-2.517.65.4
ZRKGK (Li et al., 2021)40.4----41.5----
CoLV (Zhan et al., 2021)39.6-2.920.67.954.3-2.119.76.3
KnowledGPT (Zhao et al., 2020)19.29.57.222.07.922.38.36.020.56.7
SPI (Xu et al., 2023)17.110.27.722.78.819.19.67.322.08.5
R2C240.71.70.912.31.946.61.60.911.95.4
R2C2-PT24.76.04.417.35.230.46.24.617.75.4
ELVM-OK19.812.09.525.310.725.812.19.725.210.5
ELVM-EM-Like16.211.69.125.710.719.412.09.525.710.8
ELVM-from-R2C215.814.511.828.813.318.914.311.528.813.3
ELVM14.614.711.929.313.918.014.511.829.213.7
", + "image_path": "6b4309a7b555fa2602c1f82b893988590d36ebed56160cae604ea961c3ccfa7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 290, + 289, + 349 + ], + "blocks": [ + { + "bbox": [ + 67, + 237, + 525, + 273 + ], + "lines": [ + { + "bbox": [ + 67, + 237, + 525, + 273 + ], + "spans": [ + { + "bbox": [ + 67, + 237, + 525, + 273 + ], + "type": "text", + "content": "Table 2: Performance comparison in WoW. PPL represents perplexity, B3 and B4 denote BLEU-3 and BLEU-4 scores, respectively, and R1 and R2 indicate Rouge-1 and Rouge-2 scores. Our proposed model, ELVM, achieves new SOTA performance for both seen and unseen tasks with a substantial margin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 290, + 289, + 349 + ], + "lines": [ + { + "bbox": [ + 70, + 290, + 289, + 349 + ], + "spans": [ + { + "bbox": [ + 70, + 290, + 289, + 349 + ], + "type": "table", + "html": "
ModelSeenUnseen
R@1R@5R@10R@1R@5R@10
R2C2-PT25.763.874.329.264.876.7
ELVM-EM-Like26.463.974.728.964.575.8
ELVM-β=026.164.675.728.965.376.8
ELVM29.168.480.330.468.380.3
", + "image_path": "809f2367c8bf6a80ab0f94ed3f6dd83f2121e8d4396294a55838875cf7c6e1d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 449, + 291, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 291, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 291, + 542 + ], + "type": "text", + "content": "resulting in ELVM. Furthermore, to dispel concerns that our training gains might predominantly stem from R2C2-PT's comprehensive fine-tuning across multiple datasets rather than ELVM's inherent efficacy, we opted to utilize the base R2C2 model. Upon integrating it with ELVM, the resulting configuration is denoted as ELVM-from-R2C2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": "During the training of ELVM, we employ a sampling-based generation method, specifically nucleus sampling (Holtzman et al., 2020), to promote diverse query generation in " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": ". When retrieving documents based on the generated queries, we utilize an off-the-shelf sparse retriever such as BM25 (Robertson et al., 2009), prioritizing low latency over the dense retriever (DR). We train for 3 epochs with " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": " for WoW and 1 epoch with " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "m = 8" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": " for QReCC. " + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 547, + 291, + 772 + ], + "type": "text", + "content": " is set to 1 for both tasks, unless stated otherwise. To ensure a thorough evaluation of ELVM's performance and its ability to generalize to unseen dialogues, we purposely avoid dataset overlap between training of R2C2-PT and ELVM. More detailed information on the training process can be found in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 305, + 290, + 523, + 358 + ], + "blocks": [ + { + "bbox": [ + 67, + 356, + 291, + 417 + ], + "lines": [ + { + "bbox": [ + 67, + 356, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 291, + 417 + ], + "type": "text", + "content": "Table 3: Comparison of document recall in WoW between ELVM variants. Note that we omit the performance of ELVM-OK whose performance is identical to R2C2-PT since it is trained with keeping its query generator frozen." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 290, + 523, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 290, + 523, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 523, + 358 + ], + "type": "table", + "html": "
ModelF1 ↑EM ↑R1 ↑R2 ↑PPL ↓
Q.Rewriting (Raposo et al., 2022)18.91.0---
DPR-IHN (Kim and Kim, 2022)30.44.7---
QKConv (Cai et al., 2023)33.55.9---
R2C2-PT22.81.325.711.59.4
ELVM-OK33.04.535.621.95.6
ELVM36.56.239.025.84.8
", + "image_path": "57baa9223cfd301befa41f54992799d1e3786e6a7aef496d4d902602cae4ab5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 425, + 515, + 474 + ], + "blocks": [ + { + "bbox": [ + 302, + 365, + 526, + 414 + ], + "lines": [ + { + "bbox": [ + 302, + 365, + 526, + 414 + ], + "spans": [ + { + "bbox": [ + 302, + 365, + 526, + 414 + ], + "type": "text", + "content": "Table 4: Performance comparison in QReCC with QKConv (Cai et al., 2023) which is the previous SOTA on this task. We report F1 scores, exact match (EM), Rouge score (R) and perplexity (PPL)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 425, + 515, + 474 + ], + "lines": [ + { + "bbox": [ + 313, + 425, + 515, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 515, + 474 + ], + "type": "table", + "html": "
ModelFluency (%)Relevance (%)
SeenUnseenSeenUnseen
KnowledGPT69.862.941.541.1
ELVM73.577.5*49.563.6**
", + "image_path": "75e19563e84b5f032a0f9c724bc724f9d27d6e1ae8acc4600df9f00740892194.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 481, + 525, + 542 + ], + "lines": [ + { + "bbox": [ + 302, + 481, + 525, + 542 + ], + "spans": [ + { + "bbox": [ + 302, + 481, + 525, + 542 + ], + "type": "text", + "content": "Table 5: Human evaluation results on WoW test. A pairwise t-test is conducted to verify statistical significance of the improvements, and the corresponding results in bold are significantly better than those from the baseline model " + }, + { + "bbox": [ + 302, + 481, + 525, + 542 + ], + "type": "inline_equation", + "content": "(^{**}: p < 0.01, ^*: p < 0.025)" + }, + { + "bbox": [ + 302, + 481, + 525, + 542 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "spans": [ + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "text", + "content": "Inference. During inference, unlike the training phase, only a single query is generated by the query generator " + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "text", + "content": ". The default number of retrieved documents " + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "text", + "content": " from the off-the-shelf retriever is set to 5 for WoW and 10 for QReCC. The knowledge-grounded response is generated by the response generator, " + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 302, + 564, + 525, + 696 + ], + "type": "text", + "content": ". An illustration of the inference is depicted in Figure 1b. During the inference, both query and response generation are conducted using beam search with a beam size of 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 707, + 526, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 707, + 526, + 774 + ], + "spans": [ + { + "bbox": [ + 302, + 707, + 526, + 774 + ], + "type": "text", + "content": "Variants of ELVM. To examine the impact of the proposed ELVM training, we explore another variant, ELVM-OK, which is trained on the response generation task with annotated oracle knowledge documents while keeping the query generator " + }, + { + "bbox": [ + 302, + 707, + 526, + 774 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2688" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 68, + 250, + 139 + ], + "blocks": [ + { + "bbox": [ + 108, + 68, + 250, + 139 + ], + "lines": [ + { + "bbox": [ + 108, + 68, + 250, + 139 + ], + "spans": [ + { + "bbox": [ + 108, + 68, + 250, + 139 + ], + "type": "table", + "html": "
mSeenUnseen
B4↑R2↑B4↑R2↑
110.812.410.512.2
211.113.011.112.8
4*11.913.911.813.7
811.813.711.713.6
", + "image_path": "adbeeedf36f7c9f7917c84bd396388749de2e908a905c19f6fa193281f866f78.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 206, + 289, + 267 + ], + "blocks": [ + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "lines": [ + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "type": "text", + "content": "Table 6: Effect of varying the number of queries, " + }, + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "type": "text", + "content": " , sampled during training on WoW. We see the highest performance on unseen task when " + }, + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 67, + 148, + 291, + 195 + ], + "type": "text", + "content": " . The superscripted value by * is the default setting for ELVM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 206, + 289, + 267 + ], + "lines": [ + { + "bbox": [ + 70, + 206, + 289, + 267 + ], + "spans": [ + { + "bbox": [ + 70, + 206, + 289, + 267 + ], + "type": "table", + "html": "
βWow UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
09.310.665.334.55.155.8
1*11.813.768.336.56.965.6
510.812.466.731.33.531.9
", + "image_path": "c41d2ee83258d09039bbef510dab3e91e85e10c6e4ac1a3c5b11057a33cf3f12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 357, + 290, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 357, + 290, + 423 + ], + "spans": [ + { + "bbox": [ + 67, + 357, + 290, + 423 + ], + "type": "text", + "content": "frozen. During inference, given a dialogue context the frozen query generator " + }, + { + "bbox": [ + 67, + 357, + 290, + 423 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 67, + 357, + 290, + 423 + ], + "type": "text", + "content": " generates query and use the off-the-shelf retriever to bring relevant documents. Finally the response generator of ELVM-OK generates the final response." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "spans": [ + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "content": "In Section 3.2, we describe the difference between ELVM training and EM algorithm. In order to quantitatively compare these two methods, we perform EM-like training in ELVM where we apply the posterior-weighted conditional likelihood in " + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{y,u}(\\theta ,\\phi)" + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{y,u}(\\theta ,\\phi) =" + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "inline_equation", + "content": "-\\sum_{u}SG(q(u|x,y))\\log p_{\\theta}(y|Z(u),x)." + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "content": " Here, the query latents " + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "content": " are still sampled from the prior distribution, and " + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "inline_equation", + "content": "SG" + }, + { + "bbox": [ + 67, + 424, + 291, + 570 + ], + "type": "text", + "content": " means the stop-gradient that prevents the gradient propagation to the query generator. We name this variant as ELVM-EM-Like." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 581, + 198, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 198, + 593 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 198, + 593 + ], + "type": "text", + "content": "4.2 Automatic Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 600, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 600, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 600, + 291, + 773 + ], + "type": "text", + "content": "The evaluation results of WoW test seen and unseen tasks are shown in Table 2. ELVM surpasses all its variants and previous state-of-the-art (SOTA) model, achieving a new SOTA on both WoW test seen unseen tasks, across all metrics. In addition, ELVM-OK demonstrates notable improvements in BLEU and Rouge scores compared to previous models with a slight drawback in PPL. Similarly, ELVM-EM-Like also exhibits competitive performance across multiple metrics. Moreover, when comparing ELVM-from-R2C2 and ELVM, the experiment demonstrates that while R2C2-PT offers marginal benefits, the marked performance" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 68, + 520, + 111 + ], + "blocks": [ + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "lines": [ + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "spans": [ + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "text", + "content": "Table 7: Ablation study of the posterior distillation weight " + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "text", + "content": " on WoW. From the results, we can see that the trivial value of " + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "text", + "content": " shows the best performance. The super-scripted value by " + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 67, + 275, + 290, + 333 + ], + "type": "text", + "content": " is the default setting for ELVM." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 68, + 520, + 111 + ], + "lines": [ + { + "bbox": [ + 308, + 68, + 520, + 111 + ], + "spans": [ + { + "bbox": [ + 308, + 68, + 520, + 111 + ], + "type": "table", + "html": "
ModelWoW UnseenQReCC
B4 ↑R2 ↑R@5 ↑F1 ↑EM ↑R@5 ↑
ELVM11.813.768.336.56.259.9
ELVMDR12.815.581.237.96.965.6
", + "image_path": "b56a112fc882e9e7d7ba4399359a38ffe4339de8db84dcf0d14a73447d821b8e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 165, + 523, + 244 + ], + "blocks": [ + { + "bbox": [ + 302, + 119, + 525, + 155 + ], + "lines": [ + { + "bbox": [ + 302, + 119, + 525, + 155 + ], + "spans": [ + { + "bbox": [ + 302, + 119, + 525, + 155 + ], + "type": "text", + "content": "Table 8: Performance results achieved by employing the dense retriever instead of the default BM25 retriever for WoW unseen and QReCC evaluation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 165, + 523, + 244 + ], + "lines": [ + { + "bbox": [ + 306, + 165, + 523, + 244 + ], + "spans": [ + { + "bbox": [ + 306, + 165, + 523, + 244 + ], + "type": "table", + "html": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑B4 ↑R2 ↑R@5 ↑
R2C2-PT4.45.263.84.65.464.8
ELVM-OK9.510.7-9.710.5-
ELVM11.913.968.411.813.768.3
R2C2-PTLarge8.710.262.98.910.364.4
ELVM-OKLarge9.610.6-10.211.2-
ELVMLarge12.614.372.612.414.174.3
", + "image_path": "913b147fb9c8a7c91b40b06555ff4f908470f0f10169fe95854159b93d6ea559.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 251, + 525, + 299 + ], + "lines": [ + { + "bbox": [ + 302, + 251, + 525, + 299 + ], + "spans": [ + { + "bbox": [ + 302, + 251, + 525, + 299 + ], + "type": "text", + "content": "Table 9: Effect of model scaling tested on WoW. ELVM scales out to larger parameter size with increase in performance. The subscript Large indicates the model size of 2.7B parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 320, + 517, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 320, + 517, + 333 + ], + "spans": [ + { + "bbox": [ + 302, + 320, + 517, + 333 + ], + "type": "text", + "content": "increase is predominantly attributable to ELVM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 301, + 334, + 525, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 334, + 525, + 465 + ], + "spans": [ + { + "bbox": [ + 301, + 334, + 525, + 465 + ], + "type": "text", + "content": "We further examine the performance of query generator by assessing the recall of the ground-truth documents. The results summarized in Table 3 reveal that training with our proposed algorithm leads to an increase in recall for both ELVM- " + }, + { + "bbox": [ + 301, + 334, + 525, + 465 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 301, + 334, + 525, + 465 + ], + "type": "text", + "content": " and ELVM compared to R2C2-PT while the latter yields a larger performance gain. This highlights the effectiveness of distilling the posterior distribution of the response generator into the prior distribution associated with the query generator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 466, + 525, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 466, + 525, + 558 + ], + "spans": [ + { + "bbox": [ + 302, + 466, + 525, + 558 + ], + "type": "text", + "content": "For QReCC, as shown in Table 4, ELVM surpasses the previous SOTA, QKConv (Cai et al., 2023), the recent previous works (Raposo et al., 2022; Kim and Kim, 2022), and the variants of ELVM including R2C2-PT and ELVM-OK in terms of all metrics. The recall performance on QReCC is included in Table C." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 567, + 420, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 567, + 420, + 579 + ], + "spans": [ + { + "bbox": [ + 302, + 567, + 420, + 579 + ], + "type": "text", + "content": "4.3 Human Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 301, + 585, + 525, + 744 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 585, + 525, + 744 + ], + "spans": [ + { + "bbox": [ + 301, + 585, + 525, + 744 + ], + "type": "text", + "content": "To gauge performance across multiple aspects of the quality of generated responses, human evaluation is performed on the generated responses of ELVM and KnoawedGPT (Zhao et al., 2020)" + }, + { + "bbox": [ + 301, + 585, + 525, + 744 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 301, + 585, + 525, + 744 + ], + "type": "text", + "content": ". Similar to (Rashkin et al., 2021) and (Xu et al., 2023), we assess the response quality in two aspects: Fluency and Relevance. Fluency measures whether the response is understandable, self-consistent without repetition, and proficient. Relevance assesses the extent to which the response aligns with the dialogue context, incorporates pertinent knowledge, and maintains appropriateness. In total, 50 data" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "1We choose KnowledgeGPT because it is the best performing model among publicly available models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2689" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 68, + 286, + 143 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 286, + 143 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 286, + 143 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 286, + 143 + ], + "type": "table", + "html": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
CoLV39.62.97.954.32.16.3
KnowledGPT19.27.27.922.36.06.7
SPI17.17.78.819.17.38.5
BART9741.92.69731.32.6
ELVM-from-BART15.38.411.116.48.210.9
ELVM14.611.913.918.011.813.7
", + "image_path": "c7673e4e670b5d874a006d650c1eb5ff85030cf4abfe011801d9ffea4b3dba07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 186, + 289, + 271 + ], + "blocks": [ + { + "bbox": [ + 67, + 150, + 290, + 174 + ], + "lines": [ + { + "bbox": [ + 67, + 150, + 290, + 174 + ], + "spans": [ + { + "bbox": [ + 67, + 150, + 290, + 174 + ], + "type": "text", + "content": "Table 10: Performance comparison with BART as our initial model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 186, + 289, + 271 + ], + "lines": [ + { + "bbox": [ + 70, + 186, + 289, + 271 + ], + "spans": [ + { + "bbox": [ + 70, + 186, + 289, + 271 + ], + "type": "table", + "html": "
ModelTrain RetTest RetWoW Unseen
PPL ↓B4 ↑R2 ↑F1 ↑
CoLV--54.32.16.318.5
KnowledgeGPT--22.36.06.720.5
SPI--19.17.38.5-
ELVMBM25BM2518.011.813.727.0
ELVMBM25Dense14.812.615.229.3
ELVMDenseDense12.912.815.529.0
ELVMDenseBM2521.611.813.128.2
", + "image_path": "f8b30158651c9cdd761a9fc64d06a6b5a2e9a72e85665efe6001e1031fefb539.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 279, + 291, + 338 + ], + "lines": [ + { + "bbox": [ + 67, + 279, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 291, + 338 + ], + "type": "text", + "content": "Table 11: Impact of alternating the retriever between training and testing phases on WoW unseen test. Dense indicates utilizing all-MiniLM-L6-v2 retriever as described in Section 4 and Ret is an abbreviation for retriever." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 360, + 291, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 360, + 291, + 480 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 291, + 480 + ], + "type": "text", + "content": "samples are randomly selected from WoW tasks where 25 samples are randomly selected from each seen and unseen task. The qualities of the responses are measured by A/B testing on the two aspects. The feedback is collected from 11 human experts. Further details and annotator instructions can be found in Table G. As shown in Table 5, ELVM significantly outperforms KnowledGPT in all aspects especially on the unseen task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 491, + 174, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 491, + 174, + 504 + ], + "spans": [ + { + "bbox": [ + 67, + 491, + 174, + 504 + ], + "type": "text", + "content": "4.4 Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "spans": [ + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": "Number of Queries. The size of sampled query set " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "Q^{s}" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": ", plays a huge role in training of ELVM, as it directly influences the bias of the marginal likelihood of the knowledge-grounded response " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": ". To investigate the impact of " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": ", we gradually increase " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": " from 1 to 8 and measure the performance on WoW while keeping other hyperparameters fixed. As shown in Table 6, we observe a positive correlation where increasing " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": " leads to improved performance. Notably, setting " + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 67, + 510, + 291, + 655 + ], + "type": "text", + "content": " yields the optimal performance for both seen and unseen tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "text", + "content": "Posterior Distillation Weight. We examine the influence of the posterior distillation loss on ELVM training by adjusting the value of " + }, + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "text", + "content": ". The experimental results presented in Table 7 reveal that the optimal performance is attained when " + }, + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 67, + 666, + 291, + 772 + ], + "type": "text", + "content": " for both the WoW unseen and QReCC tasks, indicating the effectiveness of incorporating the posterior distillation loss. However, a substantial decline in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 524, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 98 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 98 + ], + "type": "text", + "content": "performance is observed when " + }, + { + "bbox": [ + 302, + 71, + 524, + 98 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 302, + 71, + 524, + 98 + ], + "type": "text", + "content": " is increased to 5 or set to 0, particularly in the case of QReCC." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "spans": [ + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "content": "Varying the Retriever. We extend our experiments by incorporating the DR in two distinct ways. For WoW, we train and evaluate using all MiniLM-L6-v2" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "content": " DR. In contrast, for QReCC, due to its extensive knowledge pool, we leverage DR solely during evaluation by employing both the BM25 retriever and all-mpnet-base-v2" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "content": " DR in a hybrid manner to retrieve the top-" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "content": " documents. Concretely, we score each document by summing up the normalized score of each retriever and rerank to obtain top-" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 107, + 526, + 293 + ], + "type": "text", + "content": " documents. Table 8 presents a summary of the performance results of employing DR, demonstrating a successful integration of DR in different settings of ELVM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 303, + 526, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 303, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 302, + 303, + 526, + 449 + ], + "type": "text", + "content": "Scaling Up Model. We investigate the impact of increasing the model size on performance. Starting with the publicly available 2.7B R2C2 pretrained transformer, we follow exactly the same procedure outlined in Section 3.2 to obtain R2C2-PTLarge, ELVM-OKLarge and ELVMLarge. Detailed hyperparameters for training 2.7B model is given in Appendix A. The results presented in Table 9 demonstrate that as the model size increases, there is an improvement in performance across all metrics, supporting the scalability of ELVM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 459, + 525, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 459, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 302, + 459, + 525, + 671 + ], + "type": "text", + "content": "BART as Base. To ensure a fair comparison and further demonstrate the robustness of our algorithm on diverse model scales, we conduct an experiment employing BART (Lewis et al., 2019) base as an initial model and then apply our proposed algorithm on WoW dataset. We choose BART base model since the previous SOTA model (Xu et al., 2023) utilized BART base as their base model. The outcome of this training is termed ELVM-from-BART. As shown in Table 10, the initial BART model shows very poor performances. However, upon training the BART-base with our ELVM algorithm, there is a great improvement in performance across both WoW Seen and Unseen tasks. Importantly, the results surpass the performance benchmarks set by the previous SOTA model (SPI)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 681, + 525, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 681, + 525, + 721 + ], + "spans": [ + { + "bbox": [ + 302, + 681, + 525, + 721 + ], + "type": "text", + "content": "Retriever Switching To assess the robustness of our algorithm to the retriever mismatch between training and inference phases, we undertake an" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "type": "text", + "content": "https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2 \n" + }, + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 302, + 729, + 525, + 772 + ], + "type": "text", + "content": "https://huggingface.co/sentence-transformers/all-mpnet-base-v2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2690" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 526, + 334 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 526, + 334 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 526, + 334 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 526, + 334 + ], + "type": "table", + "html": "
Dialogue ContextI want to teach my family about investing by buying them a few shares of stock, but I really don't know much about it. The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses. Oh, so I could tell my kids they own part of a company. That's cool. Maybe I should get them a share in a company they know then, like Apple or XBox or Snapchat? [...] Examples of the latter include shares of private companies which are sold to investors through equity crowdfunding platforms. [...]
GT knowledge
GT ResponseYes, it's like the shares of a company are sold to investors through equity crowdfunding platforms.
KnownGPTQuery: n/aGT knowledge selected: False
Response: yes, you can get shares in a company like apple or xbox. you can also get shares in company like stock market.
R2C2-PTQuery: How to teach kids about investingGT knowledge selected: False
Response: The stock market can also be called the equity market or share market. That is where you buy and sell stocks. The stock represent ownership claims on businesses.
ELVM-β = 0Query: Investing tips for beginnersGT knowledge selected: False
Response: That's a good idea. The stock market is a place where you can buy and sell stocks.
ELVM w/ BM25Query: Apple xbox snapshot stock share buy apple stock instagram stock investingGT knowledge selected: True
Response: That is a good idea. You can also buy shares of private companies which are sold to investors through equity crowdfunding platforms.
ELVM w/ DRQuery: Best stocks for teaching kids about investingGT knowledge selected: True
Response: That is a good idea. The latter include shares of private companies which are sold to investors through equity crowdfunding platforms.
", + "image_path": "34ed99973cf77f8408438675e7626ab246405613d575ca400a8bbe740c9be767.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 340, + 525, + 400 + ], + "lines": [ + { + "bbox": [ + 67, + 340, + 525, + 400 + ], + "spans": [ + { + "bbox": [ + 67, + 340, + 525, + 400 + ], + "type": "text", + "content": "Table 12: Generated samples from WoW unseen task. The first row represents the dialogue context, ground truth (GT) knowledge (highlighted in olive), and GT response. Following rows contain the model generated query, response and whether the GT knowledge is included in top-5 retrieved documents. More samples can be found in Appendix H. We truncate the GT knowledge and display only relevant spans of the GT knowledge to the ground-truth response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 421, + 291, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 567 + ], + "type": "text", + "content": "additional experiment, of which the results are depicted in Table 11. In this experiment, we train the WoW task utilizing either BM25 or dense retriever (utilizing all-MiniLM-L6-v2 retriever, as detailed Section 4), and subsequently evaluate using either BM25 or dense retriever. The results on the WoW unseen test indicate that utilizing a different retriever during testing doesn't necessitate model re-training, particularly given the superior performance of the dense retriever compared to BM25." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 589, + 190, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 589, + 190, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 589, + 190, + 602 + ], + "type": "text", + "content": "4.5 Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 613, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 613, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 613, + 291, + 772 + ], + "type": "text", + "content": "For qualitative analysis, we examine the generated samples from KnownGPT, R2C2-PT and other variants of ELVM, whose samples are presented in Table 12. The example reveals that the queries generated by KnownGPT, R2C2-PT and ELVM-" + }, + { + "bbox": [ + 67, + 613, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 67, + 613, + 291, + 772 + ], + "type": "text", + "content": " fail to retrieve the relevant knowledge effectively while the queries generated by ELVM with BM25 and DR (all-MiniLM-L6-v2) reflects the ability to access the relevant knowledge, leading to improved responses. The results also demonstrate that the query generator trained with our algorithm effectively adapts to the chosen retriever." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 420, + 381, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 420, + 381, + 433 + ], + "spans": [ + { + "bbox": [ + 303, + 420, + 381, + 433 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 444, + 526, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 444, + 526, + 657 + ], + "spans": [ + { + "bbox": [ + 302, + 444, + 526, + 657 + ], + "type": "text", + "content": "In this paper, we propose ELVM as an efficient latent variable modeling for knowledge-intensive dialogue generation. In ELVM, the knowledge retrieval is realized by the query generation followed by document retrieval using the off-the-shelf retriever for ease of joint training of both the retriever and the response generator, without the knowledge supervision. Furthermore, the training scheme of ELVM modifies the ELBO for the marginal likelihood to effectively perform the joint training without the use of complex posterior sampling, which also eliminates the discrepancy between the training-time and test-time knowledge samplings. ELVM empirically demonstrates that it significant outperforms over previous latent variable methods with in-depth analysis on diverse dialogue datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 670, + 365, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 670, + 365, + 682 + ], + "spans": [ + { + "bbox": [ + 303, + 670, + 365, + 682 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "type": "text", + "content": "There are a number of limitations and possible directions to future works of ELVM proposed in this paper. First, as more large language models beyond the scale of hundreds of billion parameters are emerging, there are rooms for emergent behaviors in ELVM as the sizes of data and model grow that" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "2691" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 292, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 292, + 322 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 292, + 322 + ], + "type": "text", + "content": "has not been tested in this paper. Second, instead of the prior sampling that we use, fast posterior sampling such as short-run MCMC (Erik et al., 2019; Xu et al., 2023) is also an alternative candidate. Third, the experiments in this paper retrieve documents from confined set (e.g., Wikipedia) using light retriever such as BM25 and dense retriever. However, real-world applications often rely on the web which has much bigger pool of documents from broader open-domain. In turn, we should consider future extension of this work to utilizing larger public search engines. Last, we may consider alternative yet more effective means of query sampling during training other than a small set of random samples, as the current design. Other possible methods include imposing additional guidance for query sampling or multiple refinements of query sampling and knowledge retrieval for more diverse set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 335, + 157, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 335, + 157, + 346 + ], + "spans": [ + { + "bbox": [ + 70, + 335, + 157, + 346 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 357, + 290, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 357, + 290, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 357, + 290, + 434 + ], + "type": "text", + "content": "This work presents no direct ethical issues. However, there is a possibility that the generated responses may inadvertently contain inadequate, biased, or discriminatory content. We do not introduce new datasets, and all experiments are conducted using publicly available datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 446, + 155, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 446, + 155, + 456 + ], + "spans": [ + { + "bbox": [ + 70, + 446, + 155, + 456 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 461, + 290, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 461, + 290, + 619 + ], + "spans": [ + { + "bbox": [ + 69, + 461, + 290, + 619 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT)(No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University). In addition, this work was also supported by Artificial intelligence industrial convergence cluster development project funded by the Ministry of Science and ICT(MSIT, Korea) Gwangju Metropolitan City. Finally, we would also like to thank Brain Cloud Team at Kakao Brain for their support." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 645, + 126, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 645, + 126, + 655 + ], + "spans": [ + { + "bbox": [ + 70, + 645, + 126, + 655 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 706, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 706, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 70, + 706, + 289, + 772 + ], + "type": "text", + "content": "Leonard Adolphs, Benjamin Boerschinger, Christian Buck, Michelle Chen Huebscher, Massimiliano Ciaramita, Lasse Espeholt, Thomas Hofmann, Yannic Kilcher, Sascha Rothe, Pier Giuseppe Sessa, et al. 2021a. Boosting search engines with interactive agents. arXiv preprint arXiv:2109.00527." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 73, + 524, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 524, + 116 + ], + "type": "text", + "content": "Leonard Adolphs, Kurt Shuster, Jack Urbanek, Arthur Szlam, and Jason Weston. 2021b. Reason first, then respond: Modular generation for knowledge-infused dialogue. arXiv preprint arXiv:2111.05204." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 125, + 525, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 125, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 525, + 213 + ], + "type": "text", + "content": "Raviteja Anantha, Svitlana Vakulenko, Zhucheng Tu, Shayne Longpre, Stephen Pulman, and Srinivas Chappidi. 2021. Open-domain question answering goes conversational via question rewriting. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 222, + 525, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 525, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 525, + 353 + ], + "type": "text", + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 362, + 525, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 362, + 525, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 362, + 525, + 406 + ], + "type": "text", + "content": "Mingzhu Cai, Siqi Bao, Xin Tian, Huang He, Fan Wang, and Hua Wu. 2023. Query enhanced knowledge-intensive conversation via unsupervised joint modeling. In ACL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 415, + 525, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 525, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 525, + 470 + ], + "type": "text", + "content": "Xiuyi Chen, Fandong Meng, Peng Li, Feilong Chen, Shuang Xu, Bo Xu, and Jie Zhou. 2020. Bridging the gap between prior and posterior knowledge selection for knowledge-grounded dialogue generation. In EMNLP, pages 3426-3437." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 479, + 525, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 525, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 525, + 522 + ], + "type": "text", + "content": "Zhiyu Chen, Jie Zhao, Anjie Fang, Besnik Fetahu, Oleg Rokhlenko, and Shervin Malmasi. 2022. Reinforced question rewriting for conversational question answering. arXiv preprint arXiv:2210.15777." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 531, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 525, + 772 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 287, + 781, + 309, + 791 + ], + "type": "text", + "content": "2692" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "text", + "content": "language modeling with pathways. arXiv preprint arXiv:2204.02311." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 105, + 289, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 105, + 289, + 150 + ], + "spans": [ + { + "bbox": [ + 69, + 105, + 289, + 150 + ], + "type": "text", + "content": "Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 161, + 289, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 161, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 161, + 289, + 216 + ], + "type": "text", + "content": "Nijkamp Erik, Mitch Hill, Song-Chun Zhu, and Ying Nian Wu. 2019. Learning non-convergent non-persistent short-run mcmc toward energy-based model. In Advances in Neural Information Processing Systems (NeurIPS)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 227, + 289, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 227, + 289, + 272 + ], + "spans": [ + { + "bbox": [ + 69, + 227, + 289, + 272 + ], + "type": "text", + "content": "Michael Glass, Gaetano Rossiello, Md Faisal Mahbub Chowdhury, Ankita Rajaram Naik, Pengshan Cai, and Alfio Gliozzo. 2022. Re2g: Retrieve, rerank, generate. In NAACL." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 283, + 289, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 283, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 69, + 283, + 289, + 327 + ], + "type": "text", + "content": "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Realm: Retrievalaugmented language model pre-training. arXiv preprint arXiv:2002.08909." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 338, + 289, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 338, + 289, + 372 + ], + "spans": [ + { + "bbox": [ + 69, + 338, + 289, + 372 + ], + "type": "text", + "content": "Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2020. The curious case of neural text degeneration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 383, + 289, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 383, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 69, + 383, + 289, + 460 + ], + "type": "text", + "content": "Xinxian Huang, Huang He, Siqi Bao, Fan Wang, Hua Wu, and Haifeng Wang. 2021. PLATO-KAG: Unsupervised knowledge-grounded conversation via joint modeling. In Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI, pages 143-154, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 472, + 289, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 472, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 289, + 515 + ], + "type": "text", + "content": "Gautier Izacard and Edouard Grave. 2020. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 527, + 289, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 527, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 527, + 289, + 571 + ], + "type": "text", + "content": "Michael I Jordan, Zoubin Ghahramani, Tommi S Jaakkola, and Lawrence K Saul. 1999. An introduction to variational methods for graphical models. Machine learning, 37:183-233." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 582, + 289, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 582, + 289, + 637 + ], + "spans": [ + { + "bbox": [ + 69, + 582, + 289, + 637 + ], + "type": "text", + "content": "Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 649, + 289, + 683 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 649, + 289, + 683 + ], + "spans": [ + { + "bbox": [ + 69, + 649, + 289, + 683 + ], + "type": "text", + "content": "Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. In ICLR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 693, + 289, + 726 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 289, + 726 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 289, + 726 + ], + "type": "text", + "content": "Sungdong Kim and Gangwoo Kim. 2022. Saving dense retriever from shortcut dependency in conversational search." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "type": "text", + "content": "Diederik P Kingma and Max Welling. 2013. Autoencoding variational bayes. arXiv preprint arXiv:1312.6114." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "text", + "content": "Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8460-8478, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 147, + 525, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 525, + 246 + ], + "type": "text", + "content": "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 255, + 525, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 525, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 525, + 311 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 319, + 525, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 525, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 525, + 386 + ], + "type": "text", + "content": "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459-9474." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 395, + 525, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 525, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 525, + 429 + ], + "type": "text", + "content": "Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2021. Zero-resource knowledge-grounded dialogue generation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 438, + 525, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 525, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 525, + 515 + ], + "type": "text", + "content": "Zekang Li, Cheng Niu, Fandong Meng, Yang Feng, Qian Li, and Jie Zhou. 2019. Incremental transformer with deliberation decoder for document grounded conversations. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 12-21, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 523, + 525, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 525, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 525, + 568 + ], + "type": "text", + "content": "Rongzhong Lian, Min Xie, Fan Wang, Jinhua Peng, and Hua Wu. 2019. Learning to select knowledge for response generation in dialog systems. arXiv preprint arXiv:1902.04911." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 576, + 525, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 525, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 525, + 633 + ], + "type": "text", + "content": "Sheng-Chieh Lin, Jheng-Hong Yang, Rodrigo Nogueira, Ming-Feng Tsai, Chuan-Ju Wang, and Jimmy Lin. 2020. Conversational question reformulation via sequence-to-sequence architectures and pretrained language models. arXiv preprint arXiv:2004.01909." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 525, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 525, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 525, + 708 + ], + "type": "text", + "content": "Shilei Liu, Xiaofeng Zhao, Bochao Li, Feiliang Ren, Longhui Zhang, and Shujuan Yin. 2021. A Three-Stage Learning Framework for Low-Resource Knowledge-Grounded Dialogue Generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "text", + "content": "Shuman Liu, Hongshen Chen, Zhaochun Ren, Yang Feng, Qun Liu, and Dawei Yin. 2018. Knowledge diffusion for neural dialogue generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)," + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2693" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "text", + "content": "pages 1489-1498, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 103, + 289, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 289, + 147 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 289, + 147 + ], + "type": "text", + "content": "Longxuan Ma, Weinan Zhang, Runxin Sun, and Ting Liu. 2020. A compare aggregate transformer for understanding document-grounded dialogue. arXiv preprint arXiv:2010.00190." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 155, + 289, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 155, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 155, + 289, + 200 + ], + "type": "text", + "content": "Nikita Moghe, Siddhartha Arora, Suman Banerjee, and Mitesh M Khapra. 2018. Towards exploiting background knowledge for building conversation systems. arXiv preprint arXiv:1809.08205." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "type": "text", + "content": "Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2017. MS MARCO: A human-generated MAchine reading COprehension dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 261, + 289, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 261, + 289, + 294 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 289, + 294 + ], + "type": "text", + "content": "Rodrigo Nogueira and Kyunghyun Cho. 2017. Task-oriented query reformulation with reinforcement learning. arXiv preprint arXiv:1704.04572." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 302, + 289, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 302, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 69, + 302, + 289, + 346 + ], + "type": "text", + "content": "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In ICLR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 354, + 289, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 354, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 69, + 354, + 289, + 410 + ], + "type": "text", + "content": "Gonçalo Raposo, Rui Ribeiro, Bruno Martins, and Luísca Coheur. 2022. Question rewriting? assessing its importance for conversational question answering. In Lecture Notes in Computer Science, pages 199-206. Springer International Publishing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 418, + 289, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 418, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 69, + 418, + 289, + 518 + ], + "type": "text", + "content": "Hannah Rashkin, David Reitter, Gaurav Singh Tomar, and Dipanjan Das. 2021. Increasing faithfulness in knowledge-grounded dialogue with controllable features. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 704-718, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 526, + 289, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 526, + 289, + 603 + ], + "spans": [ + { + "bbox": [ + 69, + 526, + 289, + 603 + ], + "type": "text", + "content": "Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5370-5381, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 612, + 289, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 612, + 289, + 655 + ], + "spans": [ + { + "bbox": [ + 69, + 612, + 289, + 655 + ], + "type": "text", + "content": "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends® in Information Retrieval, 3(4):333-389." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 664, + 289, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 664, + 289, + 729 + ], + "spans": [ + { + "bbox": [ + 69, + 664, + 289, + 729 + ], + "type": "text", + "content": "Kurt Shuster, Mojtaba Komeili, Leonard Adolphs, Stephen Roller, Arthur Szlam, and Jason Weston. 2022a. Language models that seek for knowledge: Modular search & generation for dialogue and prompt completion. arXiv preprint arXiv:2203.13224." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 738, + 289, + 772 + ], + "type": "text", + "content": "Kurt Shuster, Jing Xu, Mojtaba Komeili, Da Ju, Eric Michael Smith, Stephen Roller, Megan Ung, Moya Chen, Kushal Arora, Joshua Lane, et al. 2022b." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 105 + ], + "type": "text", + "content": "Blenderbot 3: a deployed conversational agent that continually learns to responsibly engage. arXiv preprint arXiv:2208.03188." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 116, + 525, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 525, + 182 + ], + "type": "text", + "content": "Weiwei Sun, Pengjie Ren, and Zhaochun Ren. 2023. Generative knowledge selection for knowledge-grounded dialogues. In Findings of the Association for Computational Linguistics: EACL 2023, pages 2077-2088, Dubrovnik, Croatia. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 191, + 525, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 191, + 525, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 191, + 525, + 422 + ], + "type": "text", + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen S. Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed H. Chi, and Quoc Le. 2022. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, abs/2201.08239." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 433, + 525, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 525, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 525, + 477 + ], + "type": "text", + "content": "David Thulke, Nico Daheim, Christian Dugast, and Hermann Ney. 2021. Efficient retrieval augmented generation from unstructured knowledge for task-oriented dialog. arXiv preprint arXiv:2102.04643." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 487, + 525, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 525, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 525, + 565 + ], + "type": "text", + "content": "Megan Ung, Jing Xu, and Y-Lan Boureau. 2022. SaFeR-Dialogues: Taking feedback gracefully after conversational safety failures. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6462-6481, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 575, + 525, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 525, + 630 + ], + "type": "text", + "content": "Svitlana Vakulenko, Shayne Longpre, Zhucheng Tu, and Raviteja Anantha. 2021. Question rewriting for conversational question answering. In Proceedings of the 14th ACM international conference on web search and data mining, pages 355-363." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 640, + 525, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 525, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 525, + 707 + ], + "type": "text", + "content": "Nikos Voskarides, Dan Li, Pengjie Ren, Evangelos Kanoulas, and Maarten de Rijke. 2020. Query resolution for conversational search with limited supervision. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 921-930." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 525, + 772 + ], + "type": "text", + "content": "Zeqiu Wu, Yi Luan, Hannah Rashkin, David Reitter, Hannaneh Hajishirzi, Mari Ostendorf, and Gaurav Singh Tomar. 2022. Conqrr: Conversational query rewriting for retrieval with reinforcement learning. In EMNLP." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2694" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "text", + "content": "Tianbao Xie, Chen Henry Wu, Peng Shi, Ruiqi Zhong, Torsten Scholak, Michihiro Yasunaga, Chien-Sheng Wu, Ming Zhong, Pengcheng Yin, Sida I Wang, et al. 2022. Unifiedsgk: Unifying and multi-tasking structured knowledge grounding with text-to-text language models. arXiv preprint arXiv:2201.05966." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 146, + 290, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 146, + 290, + 214 + ], + "spans": [ + { + "bbox": [ + 68, + 146, + 290, + 214 + ], + "type": "text", + "content": "Jing Xu, Arthur Szlam, and Jason Weston. 2022a. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 221, + 290, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 221, + 290, + 277 + ], + "spans": [ + { + "bbox": [ + 68, + 221, + 290, + 277 + ], + "type": "text", + "content": "Jing Xu, Megan Ung, Mojtaba Komeili, Kushal Arora, Y-Lan Boureau, and Jason Weston. 2022b. Learning new skills after deployment: Improving open-domain internet-driven dialogue with human feedback. arXiv preprint arXiv:2208.03270." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 285, + 290, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 285, + 290, + 329 + ], + "spans": [ + { + "bbox": [ + 69, + 285, + 290, + 329 + ], + "type": "text", + "content": "Yan Xu, Deqian Kong, Dehong Xu, Ziwei Ji, Bo Pang, Pascale Fung, and Ying Nian Wu. 2023. Diverse and faithful knowledge-grounded dialogue generation via sequential posterior inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 337, + 290, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 337, + 290, + 403 + ], + "spans": [ + { + "bbox": [ + 68, + 337, + 290, + 403 + ], + "type": "text", + "content": "Shi Yu, Jiahua Liu, Jingqin Yang, Chenyan Xiong, Paul Bennett, Jianfeng Gao, and Zhiyuan Liu. 2020. Few-shot generative conversational query rewriting. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, pages 1933-1936." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 412, + 290, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 412, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 412, + 290, + 455 + ], + "type": "text", + "content": "Haolan Zhan, Lei Shen, Hongshen Chen, and Hainan Zhang. 2021. Colv: A collaborative latent variable model for knowledge-grounded dialogue generation. In EMNLP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 464, + 290, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 464, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 68, + 464, + 290, + 541 + ], + "type": "text", + "content": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 550, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 550, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 69, + 550, + 289, + 594 + ], + "type": "text", + "content": "Yizhe Zhang, Siqi Sun, Xiang Gao, Yuwei Fang, Chris Brockett, Michel Galley, Jianfeng Gao, and Bill Dolan. 2022. Retgen: A joint framework for retrieval and grounded text generation modeling. In AAAI." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 602, + 290, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 602, + 290, + 646 + ], + "spans": [ + { + "bbox": [ + 68, + 602, + 290, + 646 + ], + "type": "text", + "content": "Xueliang Zhao, Wei Wu, Can Xu, Chongyang Tao, Dongyan Zhao, and Rui Yan. 2020. Knowledge-grounded dialogue generation with pre-trained language models. In EMNLP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 655, + 290, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 655, + 290, + 721 + ], + "spans": [ + { + "bbox": [ + 68, + 655, + 290, + 721 + ], + "type": "text", + "content": "Hao Zhou, Tom Young, Minlie Huang, Haizhou Zhao, Jingfang Xu, and Xiaoyan Zhu. 2018a. Common-sense knowledge aware conversation generation with graph attention. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, IJCAI'18, page 4623-4629. AAAI Press." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 729, + 290, + 762 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 729, + 290, + 762 + ], + "spans": [ + { + "bbox": [ + 69, + 729, + 290, + 762 + ], + "type": "text", + "content": "Kangyan Zhou, Shrimai Prabhumoye, and Alan W Black. 2018b. A dataset for document grounded conversations. In EMNLP." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 100, + 508, + 217 + ], + "blocks": [ + { + "bbox": [ + 303, + 70, + 409, + 84 + ], + "lines": [ + { + "bbox": [ + 303, + 70, + 409, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 409, + 84 + ], + "type": "text", + "content": "A Training Details" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 100, + 508, + 217 + ], + "lines": [ + { + "bbox": [ + 321, + 100, + 508, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 100, + 508, + 217 + ], + "type": "table", + "html": "
ParameterValue
Training Steps10,000
Batch Size128
OptimizerAdamW
Learning Rate (LR)5e-5
LR SchedulerCosine
Adam β10.9
Adam β20.999
Adam ε1e-8
Warmup Steps1000
", + "image_path": "d0a7a1157173c9833922c948fcaeebe5c9161c7c66a63f40f21c981412d73a50.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 324, + 285, + 504, + 458 + ], + "blocks": [ + { + "bbox": [ + 302, + 225, + 524, + 248 + ], + "lines": [ + { + "bbox": [ + 302, + 225, + 524, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 225, + 524, + 248 + ], + "type": "text", + "content": "Table 13: Hyperparameters used for stage-1 training. " + }, + { + "bbox": [ + 302, + 225, + 524, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 225, + 524, + 248 + ], + "type": "text", + "content": " indicates number of retrieved documents." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 285, + 504, + 458 + ], + "lines": [ + { + "bbox": [ + 324, + 285, + 504, + 458 + ], + "spans": [ + { + "bbox": [ + 324, + 285, + 504, + 458 + ], + "type": "table", + "html": "
WoWQReCC
Train Epochs31
Batch Size6464
OptimizerAdamWAdamW
Learning Rate (LR)1e-45e-4
LR SchedulerCosineCosine
Adam β10.90.9
Adam β20.9990.999
Adam ε1e-81e-8
Warmup Steps1000
Min Query Length11
Max Query Length3232
Top-k100100
Top-p0.950.5
Temperature1.02.0
", + "image_path": "1c730c27136aa5a9da7f3fd4151d394265af8f771186b43f3b76c2f44fb989e7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 465, + 524, + 489 + ], + "lines": [ + { + "bbox": [ + 302, + 465, + 524, + 489 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 524, + 489 + ], + "type": "text", + "content": "Table 14: Hyperparameters used for training ELVM and its." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 503, + 525, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 503, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 302, + 503, + 525, + 689 + ], + "type": "text", + "content": "Table 13 we provide hyperparameters utilized during this training process. We also report hyperparameters used for training ELVM and its variants model (ELVM-OK and ELVM-EM-Like) in Table 14. In addition, during training of QReCC, we adopt a strategy to decrease the knowledge pool size from 54M to 1M. This reduction aims to mitigate latency issues during document retrieval. Concretely, the 1M pool is constructed by initially gathering the ground-truth relevant passage for each instance in the training dataset then adding randomly sampled documents. During the evaluation stage, we utilize the original full-size knowledge pool of 54M passages." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 698, + 465, + 713 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 698, + 465, + 713 + ], + "spans": [ + { + "bbox": [ + 302, + 698, + 465, + 713 + ], + "type": "text", + "content": "B Upper Bound Performance" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 719, + 524, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 524, + 772 + ], + "type": "text", + "content": "We believe in the value of understanding the utmost capability in an ideal scenario. To this end, we pursue an upper-bound performance analysis for both the WoW and QReCC tasks. This approach" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2695" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 68, + 488, + 111 + ], + "blocks": [ + { + "bbox": [ + 105, + 68, + 488, + 111 + ], + "lines": [ + { + "bbox": [ + 105, + 68, + 488, + 111 + ], + "spans": [ + { + "bbox": [ + 105, + 68, + 488, + 111 + ], + "type": "table", + "html": "
ModelSeenUnseen
PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑PPL ↓B3 ↑B4 ↑R1 ↑R2 ↑
ELVM14.614.711.929.313.918.014.511.829.213.7
ELVM-w-GT-Doc9.722.118.540.923.910.521.517.940.723.3
", + "image_path": "0c14ace2b5871033696403c1686574f0b5bf91cdc0f68dd252e001d4c7c2717c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 71, + 162, + 286, + 241 + ], + "blocks": [ + { + "bbox": [ + 67, + 119, + 525, + 143 + ], + "lines": [ + { + "bbox": [ + 67, + 119, + 525, + 143 + ], + "spans": [ + { + "bbox": [ + 67, + 119, + 525, + 143 + ], + "type": "text", + "content": "Table 15: Performance comparison in WoW in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 162, + 286, + 241 + ], + "lines": [ + { + "bbox": [ + 71, + 162, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 71, + 162, + 286, + 241 + ], + "type": "table", + "html": "
ModelF1 ↑EM ↑
Q Rewriting (Raposo et al., 2022)18.91.0
DPR-IHN (Kim and Kim, 2022)30.44.7
QKConv (Cai et al., 2023)33.55.9
ELVM36.56.2
ELVM-w-GT-Doc55.416.2
", + "image_path": "1e8a502aa77db1d1cbc69f99251611957c32cb3b72860cbd7ccb66aca877894b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 310, + 289, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 310, + 289, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 310, + 289, + 414 + ], + "type": "text", + "content": "involves training our response generator under the assumption that the model has access to the ground-truth document during both training and evaluation. The performance metrics of this specific setup are designated as \"ELVM-w-GT-Doc\" in Table 15 and Table 16. From both tables, it's evident that there still remains a margin for improvement for both tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 432, + 247, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 432, + 247, + 444 + ], + "spans": [ + { + "bbox": [ + 67, + 432, + 247, + 444 + ], + "type": "text", + "content": "C Document Recall Performance" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 92, + 466, + 265, + 549 + ], + "blocks": [ + { + "bbox": [ + 67, + 249, + 289, + 285 + ], + "lines": [ + { + "bbox": [ + 67, + 249, + 289, + 285 + ], + "spans": [ + { + "bbox": [ + 67, + 249, + 289, + 285 + ], + "type": "text", + "content": "Table 16: Performance comparison in QReCC in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 466, + 265, + 549 + ], + "lines": [ + { + "bbox": [ + 92, + 466, + 265, + 549 + ], + "spans": [ + { + "bbox": [ + 92, + 466, + 265, + 549 + ], + "type": "table", + "html": "
Query fromR@1R@5R@10
Dialogue History33.458.668.7
Last Utterance25.550.862.6
R2C2-PT17.433.039.5
ELVM-β=030.655.865.7
ELVM35.259.968.6
ELVMDR37.565.675.1
", + "image_path": "1babfe3a5ac5cf8e69512ebd39e8c77924ff6297e1c66560f6d1cdefc5dcc9ad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 600, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 600, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 600, + 290, + 772 + ], + "type": "text", + "content": "We report the document recall performance on QReCC, including different types of queries, as show in Table 17. Our proposed model, ELVM, surpasses both its variants and non-model generated queries such as using dialogue context as a query. The document recall performance on QReCC, encompassing various query types, is presented in Table 17. Our proposed ELVM model outperforms its variants as well as non-model generated queries, including the use of dialogue context as a query. Moreover, incorporating the dense retriever for reranking documents " + }, + { + "bbox": [ + 67, + 600, + 290, + 772 + ], + "type": "inline_equation", + "content": "(\\mathrm{ELVM}_{\\mathrm{DR}})" + }, + { + "bbox": [ + 67, + 600, + 290, + 772 + ], + "type": "text", + "content": " leads to further improvements in recall." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 305, + 169, + 528, + 233 + ], + "blocks": [ + { + "bbox": [ + 67, + 557, + 289, + 581 + ], + "lines": [ + { + "bbox": [ + 67, + 557, + 289, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 557, + 289, + 581 + ], + "type": "text", + "content": "Table 17: Query performance comparison in QReCC for different types of queries." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 169, + 528, + 233 + ], + "lines": [ + { + "bbox": [ + 305, + 169, + 528, + 233 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 528, + 233 + ], + "type": "table", + "html": "
ModelSeenUnseen
B4 ↑R2 ↑R@5 ↑R@10 ↑B4 ↑R2 ↑R@5 ↑R@10 ↑
k=511.913.968.480.311.813.768.380.3
k=5Bulk9.410.962.773.98.29.547.261.9
k=1010.312.264.876.410.612.764.876.9
k=10Bulk10.212.262.174.19.010.547.261.6
", + "image_path": "4a4f0f21a12b452a41062e3c44606dda4c413cd8526270faf1eff818377f1ae6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 241, + 524, + 287 + ], + "lines": [ + { + "bbox": [ + 302, + 241, + 524, + 287 + ], + "spans": [ + { + "bbox": [ + 302, + 241, + 524, + 287 + ], + "type": "text", + "content": "Table 18: Effect of increasing the document pool size to 1000 on the performance of ELVM. As before, " + }, + { + "bbox": [ + 302, + 241, + 524, + 287 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 241, + 524, + 287 + ], + "type": "text", + "content": " indicates the number of retrieved documents used during train and inference." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 309, + 450, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 309, + 450, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 309, + 450, + 323 + ], + "type": "text", + "content": "D Knowledge Pool Scaling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "text", + "content": "In order to create a more realistic and challenging retrieval scenario, we expand the original setting of the WoW dataset by increasing the number of relevant documents per instance to 1000. This modification better reflects real-world information retrieval scenarios where a vast array of both relevant and irrelevant documents are typically encountered. We construct this extended dataset by randomly augmenting " + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "inline_equation", + "content": "1000 - K" + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "text", + "content": " documents from the WoW dataset to the document pool for each instance, where " + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "text", + "content": " represents the number of annotated relevant documents for each instance. On average, " + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "inline_equation", + "content": "K \\approx 20" + }, + { + "bbox": [ + 301, + 330, + 525, + 502 + ], + "type": "text", + "content": " for each instance in WoW dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 503, + 525, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 503, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 503, + 525, + 661 + ], + "type": "text", + "content": "Results in Table 18 show performances of ELVM models with different number of retrieved documents, " + }, + { + "bbox": [ + 302, + 503, + 525, + 661 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 503, + 525, + 661 + ], + "type": "text", + "content": ", and knowledge pool size, where models with increased knowledge pool has tailing keyword Bulk at then end of their names. Despite the increase in complexity of the retrieval task, ELVM demonstrates robust performance, with only a marginal drop in metrics compared to the original setting. Furthermore, it is important to highlight that even under these more difficult conditions with 1000 relevant documents, ELVM still outperforms SPI by a significant margin." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 672, + 470, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 672, + 470, + 686 + ], + "spans": [ + { + "bbox": [ + 302, + 672, + 470, + 686 + ], + "type": "text", + "content": "E Effect of Parameter Sharing" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 693, + 525, + 772 + ], + "type": "text", + "content": "In our ELVM training, parameters are shared between the query and response generators. To probe this design choice, we decoupled these parameters, resulting in the ELVM-Decouple configuration, as detailed in Table 19. Upon evaluation on WoW's test seen and unseen tasks, ELVM-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2696" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 68, + 285, + 108 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 285, + 108 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 285, + 108 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 285, + 108 + ], + "type": "table", + "html": "
ModelSeenUnseen
PPL ↓B4 ↑R2 ↑PPL ↓B4 ↑R2 ↑
ELVM14.611.913.918.011.813.7
ELVM-Decouple16.212.614.521.612.214.0
", + "image_path": "dc402caeb57da018a3932ab96f09dac3b7b55c399894fae2ca181de5a1aff157.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 185, + 285, + 295 + ], + "blocks": [ + { + "bbox": [ + 67, + 116, + 290, + 176 + ], + "lines": [ + { + "bbox": [ + 67, + 116, + 290, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 116, + 290, + 176 + ], + "type": "text", + "content": "Table 19: Performance comparison on the WoW dataset using the ELVM algorithm. ELVM denotes training with shared parameters between query and response generators, while ELVM-Decouple denotes a non-shared parameter approach." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 185, + 285, + 295 + ], + "lines": [ + { + "bbox": [ + 73, + 185, + 285, + 295 + ], + "spans": [ + { + "bbox": [ + 73, + 185, + 285, + 295 + ], + "type": "table", + "html": "
ModelSeenUnseen
F1 ↑KF1 ↑F1 ↑KF1 ↑
CoLV20.318.218.517.5
KnownGPT22.023.820.522.1
SPI----
R2C2-PT15.420.515.722.3
ELVM-OK23.033.022.933.0
ELVM-EM-Like24.533.223.533.5
ELVM27.240.827.040.3
", + "image_path": "7d93c1e91159d46e3fd98d3e1f26bb3dd58396e43d6ceddee28f334e05e4b8b6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 304, + 289, + 327 + ], + "lines": [ + { + "bbox": [ + 67, + 304, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 67, + 304, + 289, + 327 + ], + "type": "text", + "content": "Table 20: F1 and KF1 scores for ELVM and baseline models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 349, + 291, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 349, + 291, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 349, + 291, + 414 + ], + "type": "text", + "content": "Decouple showed a slight edge over the standard ELVM. Nevertheless, the benefits of shared parameters—particularly regarding computational efficiency and knowledge transfer during training and inference—cannot be understated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 425, + 188, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 425, + 188, + 438 + ], + "spans": [ + { + "bbox": [ + 68, + 425, + 188, + 438 + ], + "type": "text", + "content": "F Additional Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 446, + 291, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 446, + 291, + 499 + ], + "spans": [ + { + "bbox": [ + 67, + 446, + 291, + 499 + ], + "type": "text", + "content": "We report the evaluation of our models on the WoW task using both F1 and knowledge F1 (KF1) metrics in Table 20. For reference, values pertaining to previous models are derived from Sun et al. (2023)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 509, + 190, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 509, + 190, + 521 + ], + "spans": [ + { + "bbox": [ + 68, + 509, + 190, + 521 + ], + "type": "text", + "content": "G Human Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 530, + 291, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 530, + 291, + 689 + ], + "spans": [ + { + "bbox": [ + 67, + 530, + 291, + 689 + ], + "type": "text", + "content": "For a comprehensive evaluation, we conduct the human evaluation between ELVM and KnownGPT on seen and unseen tasks in WoW (Dinan et al., 2018). We randomly select 50 samples (25 samples per task), and each sample is evaluated by 11 different human experts. In specific, the two generated responses from each model in the same context are assigned to the annotators. For A/B testing, we give one score to the model if its response is received an equally good or better than the other one. Figure 2 shows the annotator instructions for the two aspects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 699, + 193, + 713 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 699, + 193, + 713 + ], + "spans": [ + { + "bbox": [ + 68, + 699, + 193, + 713 + ], + "type": "text", + "content": "H Generated Samples" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "content": "Generated samples for the WoW test unseen are presented in Table 21 and Table 22. Similarly, Table 23 and Table 24 showcase the generated samples for the QReCC test." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2697" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 125, + 526, + 681 + ], + "blocks": [ + { + "bbox": [ + 69, + 125, + 526, + 681 + ], + "lines": [ + { + "bbox": [ + 69, + 125, + 526, + 681 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 526, + 681 + ], + "type": "table", + "html": "
Dialogue ContextHey, do you like cardigans? I like the modern version without no buttons and hangs open by design. Me too. My fav is the waterfall style but it's going out of fashion now I think [...] Cardigan (sweater) By contrast, a pullover does not open in front but must be "pulled over" the head to be worn. [...] But the pullover does not open in front but must be "pulled over" the head to be worn.
GT knowledge
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: i like the hand-knitted version. it's more comfortable. i like mine made of a cardigan.
R2C2-PTQuery: fallen fashion GT knowledge selected: False Response: I'm sorry, I didn't mean to offend you.
ELVM-β = 0Query: fallen fashion cardigans GT knowledge selected: True Response: I like the one that is tied is instead considered a robe.
ELVM w/ BM25Query: fallen waterfall cardigan style fallout cardigans GT knowledge selected: True Response: I like the pullover style that does not open in front but must be "pulled over" the head to be worn.
ELVM w/ DRQuery: fallen waterfall cardigan GT knowledge selected: True Response: I like the pullover that does not open in front but must be "pulled over" the head to be worn.
Dialogue Context[...] Was she diagnosed as a child? It is much easier to diagnose children, since the symptoms begin occurring early in life
GT knowledgeYes it started when she was 6 years old. She was 23 when I met her she was already recieving treatment [...] These fellowships include pediatric neurosurgery, trauma/neurocritical care, functional and stereotactic surgery, surgical neuro-oncology, radiosurgery, neurovascular surgery, skull-base surgery, peripheral nerve and spine surgery. [...] My first fellowship was in pediatric neurosurgery. It is great that she is doing better.
GT Response
KnownGPTQuery: n/a GT knowledge selected: False Response: did she cry during the periods of depression? she was crying, a negative outlook on life, and poor eye contact with others.
R2C2-PTQuery: types of epilepsy GT knowledge selected: False Response: I am glad to hear she is receiving treatment. Dementia is such a devastating disease.
ELVM-β = 0Query: types of epilepsy GT knowledge selected: False Response: I am glad she is receiving treatment. Dementia has a significant effect on a person's caregivers.
ELVM w/ BM25Query: neurosurgery epilepsy neurocognitive neurosciences GT knowledge selected: True Response: I am glad to hear that she is receiving treatment. Neurosurgery is a branch of medicine that specializes in the treatment of epilepsy.
ELVM w/ DRQuery: neurosurgery treatment for epilepsy GT knowledge selected: True Response: Did she undergo a fellowship? Many neurosurgeons pursue additional training in the form of a fellowship, after residency or in some cases, as a senior resident.
", + "image_path": "0e6037791cb85ab75bb71f9d8bd0a9a250b15dd9ccfbd0bc8f135e7a88833997.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 688, + 525, + 713 + ], + "lines": [ + { + "bbox": [ + 67, + 688, + 525, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 525, + 713 + ], + "type": "text", + "content": "Table 21: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2698" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 108, + 526, + 697 + ], + "blocks": [ + { + "bbox": [ + 69, + 108, + 526, + 697 + ], + "lines": [ + { + "bbox": [ + 69, + 108, + 526, + 697 + ], + "spans": [ + { + "bbox": [ + 69, + 108, + 526, + 697 + ], + "type": "table", + "html": "
Dialogue ContextDo you know anything about the band called Insane Clown Posse? \nInsane Clown Posse (ICP) is an American hip hop duo composed of Violent J, and Shaggy 2 Dope. [...] \nThat's quite a lot of albums, what else do you know about ICP? \n[...] The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas" and "Big Money Rustlas". [...] \nThe duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas
GT knowledge
GT Response
KnownedGPTResponse: Query: n/a \nGT knowledge selected: False
Response: the duo has earned two platinum and five gold albums. they have also earned two platinum and five gold albums
R2C2-PTQuery: ICP wiki \nGT knowledge selected: False
Response: The duo has earned two platinum ,with 5 golden album
ELVM-β = 0Query: ICP facts \nGT knowledge selected: False
Response: The group has been around for a long time .
ELVM w/ BM25Query: Insane clown posse ICP facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
ELVM w/ DRQuery: Insane clown posse facts \nGT knowledge selected: True
Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films "Big Money Hustlas"
Dialogue Context[...] Athletes normally eat meat in large quantities. \nExactly. I have also added much meat eating as part of my special diet. They are rich in protein. Especially, red meat
GT knowledge[...] In nutritional science, on the other hand, red meat is defined as any meat that has more myoglobin than white meat, white meat being defined as non-dark meat from chicken (excluding leg or thigh), or fish. [...]
GT ResponseRed meat is divined as any meat that has myoglobin than white
KnownedGPTQuery: n/a \nGT knowledge selected: False
Response: red meat is classified as white meat. i think it's interesting that only flesh from mammals or fowl is classified as red.
R2C2-PTQuery: protein powder \nGT knowledge selected: False
Response: Athletes normally eat meat in large quantities.
ELVM-β = 0Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ BM25Query: red meat protein \nGT knowledge selected: True
Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.
ELVM w/ DRQuery: red meat protein \nGT knowledge selected: True
Response: Red meat is defined as any meat that has more myoglobin than white meat.
", + "image_path": "fc63143279353e0e504e68c20a06ce19be1f988a10d973c6632f3ba79785021a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 703, + 525, + 729 + ], + "lines": [ + { + "bbox": [ + 67, + 703, + 525, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 525, + 729 + ], + "type": "text", + "content": "Table 22: Generated samples from WoW test unseen. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2699" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 158, + 525, + 648 + ], + "blocks": [ + { + "bbox": [ + 69, + 158, + 525, + 648 + ], + "lines": [ + { + "bbox": [ + 69, + 158, + 525, + 648 + ], + "spans": [ + { + "bbox": [ + 69, + 158, + 525, + 648 + ], + "type": "table", + "html": "
Dialogue ContextWhen did Mark Spitz' swimming career end? Following the Munich Olympics [...] What else did Mark Spitz do besides Speedo swimsuit ads after his swimming career? [...] However, as the memory of his feats receded, so did his endorsement and promotional deals. As his endorsements faded, he started a successful real-estate company in Beverly Hills. He bought a Ferrari and says he made [...]
GT knowledge
GT ResponseMark Spitz started a successful real-estate company in Beverly Hills
R2C2-PTQuery: What did mark spitz do after his olympic career? \nGT knowledge selected: False \nResponse: In 1999 Spitz ranked No. 33 on ESPN SportsCentury 50 Greatest Athletes, the only aquatic athlete to make the list.
ELVM-β = 0Query: Mark Spitz Speedo Swimsuit ads after speedo swimming career ends mark spitz m Munich olympics 2012 \nGT knowledge selected: False \nResponse: Spitz did Schick razor commercials and Speedo swimsuit ads
ELVM w/ BM25Query: Spitz endorsements after swimming career ends \nGT knowledge selected: True \nResponse: Mark Spitz started a successful real-estate company in Beverly Hills.
Dialogue ContextWhat was the New Scheme? \nThe 'new scheme' and the 'new dedication' respectively, initially seen as a [...] \n[...] \nDid Ahmadi followers of the New Scheme act as missionaries? \nAs part of this Mahmood Ahmad appointed 5 men to survey the Punjab [...] \nDid Mirza Basheer-ud-Din Mahmood Ahmad use missionaries to try to spread the New Scheme to foreign countries?
GT knowledge[...] In time the scheme produced a vast amount of literature in defence of Islam in general and the Ahmadiyya beliefs in particular. The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance. As part of this Mahmood Ahmad appointed [...]
GT ResponseThe funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
R2C2-PTQuery: Mirza Basher-ud Din Mahmood Ahmad missionaries \nGT knowledge selected: False \nResponse: Mirza Basheer-ud-Din Mahmood Ahmad used missionaries to try to spread the New Scheme to foreign countries.
ELVM-β = 0Query: New Scheme Mahmud Ahmad missionaries to spread New Scheme in foreign countries \nGT knowledge selected: True \nResponse: Yes, I have. I have experienced heartache, also known as loss of a loved one.
ELVM w/ BM25Query: New scheme missionaries used to spread Ahmadiyyah teachings in foreign countries \nGT knowledge selected: True \nResponse: The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.
", + "image_path": "a900e3cf666337bc10dbbf58a3b055dc955c9f88d06eeda9e3538b678ca509f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 655, + 525, + 681 + ], + "lines": [ + { + "bbox": [ + 67, + 655, + 525, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 525, + 681 + ], + "type": "text", + "content": "Table 23: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "2700" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 114, + 526, + 692 + ], + "blocks": [ + { + "bbox": [ + 69, + 114, + 526, + 692 + ], + "lines": [ + { + "bbox": [ + 69, + 114, + 526, + 692 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 526, + 692 + ], + "type": "table", + "html": "
Dialogue ContextHow did Stephen Sondheim collaborate with James Lapine?\nStephen Sondheim found a new collaborator in [...]What did Stephen Sondheim and James Lapine do together in the 1980s?\nTheir first collaboration was Sunday in the Park with George (1984).Was Stephen Sondheim and James Lapine's first collaboration a success?\nStephen Sondheim and James Lapine won the 1985 Pulitzer Prize [...]Did Stephen Sondheim and James Lapine collaborate in other projects, besides Sunday in the Park with George?[...] last work together was the rhapsodic Passion (1994).Was Stephen Sondheim and James Lapine's last work a success?\nSondheim and Lapine's last work together was the rhapsodic Passion (1994), adapted from Ettore Scola's Italian film Passione D'Amore. With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.[46] Later work Assassins opened [...]With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
GT knowledge
GT Response
R2C2-PTQuery: Was the play "rhapsodic passion" a success or a failure?\nGT knowledge selected: TrueResponse: Sondheim and Lapine's last work together was the rhapsodic Passion (1994).
ELVM-β = 0Query: Stephen Sonds and Lapine Rhapsodic passion1994 Stephen sondheim, James lapine plays 12 dreams off-broadway p Pulitzer prize for play Sunday in Park.GT knowledge selected: FalseResponse: Stephen Sondheim and James Lapine's last work together was the rhapsodic Passion (1994).
ELVM w/ BM25Query: Sondheim last work was the Rhapsodic passion (1994) Pulitzer prize won for Drama. Lapine wrote the play Twelve Dreams.GT knowledge selected: TrueResponse: With a run of 280 performances, Passion was the shortest-running show to win a Tony Award for Best Musical.
Dialogue ContextDid Frank Hague commit any questionable acts?\nYes, he took kickbacks in cash, and relied on two ordinances of dubious [...]Did Frank Hague do anything illegal?\nHe passed an unconstitutional ordinance that gave himself as commissioner the power to [...]Did anyone attempt to prosecute Frank Hague?[...] Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. [38] He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash. [...]Yes, but he was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
GT knowledge
GT Response
R2C2-PTQuery: Frank Hague prosecutionGT knowledge selected: FalseResponse: No, no one tried to prosecute Frank Hague.
ELVM-β = 0Query: Frank Hague ordinance ordinance frank Hague charges Frank Hague Hague for flouting constitutionality ordinancesGT knowledge selected: FalseResponse: Hague was found guilty of contempt of court for ignoring the subpoena. He was fined $100 and stripped of his duties as Deputy Sheriff.
ELVM w/ BM25Query: Hague prosecution attempt failed after Hague passed unconstitutional ordinance. no Hague was prosecuted.GT knowledge selected: TrueResponse: Hague was able to stay in power despite a nearly constant effort to turn him out of office from 1921 onward. He was also able to avoid prosecution despite numerous federal and state investigations in part due to the fact he took most of his kickbacks in cash.
", + "image_path": "6d4efa4c5edc635a4576d6ee14969c14be87d094b8218bb9d915a2a8cf1b96c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 698, + 525, + 724 + ], + "lines": [ + { + "bbox": [ + 67, + 698, + 525, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 698, + 525, + 724 + ], + "type": "text", + "content": "Table 24: Generated samples from QReCC test. We truncate the ground-truth knowledge and display only relevant spans of the ground-truth knowledge to the ground-truth response." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2701" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 149, + 131, + 353, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 131, + 353, + 144 + ], + "spans": [ + { + "bbox": [ + 149, + 131, + 353, + 144 + ], + "type": "text", + "content": "Human Evaluation Annotation Tool" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 154, + 254, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 154, + 254, + 164 + ], + "spans": [ + { + "bbox": [ + 149, + 154, + 254, + 164 + ], + "type": "text", + "content": "Components of each example" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 173, + 276, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 173, + 276, + 180 + ], + "spans": [ + { + "bbox": [ + 149, + 173, + 276, + 180 + ], + "type": "text", + "content": "Each example may include one of more of the following" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 188, + 398, + 221 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 149, + 188, + 381, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 188, + 381, + 195 + ], + "spans": [ + { + "bbox": [ + 149, + 188, + 381, + 195 + ], + "type": "text", + "content": "- History: topic related to the conversation and the dialogue history. Examples may not contain 'history'" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 149, + 196, + 238, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 196, + 238, + 204 + ], + "spans": [ + { + "bbox": [ + 149, + 196, + 238, + 204 + ], + "type": "text", + "content": "- Input text: Utterance of the opponent" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 149, + 206, + 398, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 206, + 398, + 213 + ], + "spans": [ + { + "bbox": [ + 149, + 206, + 398, + 213 + ], + "type": "text", + "content": "- Ground truth response of the example, given as a reference. The annotator may use the reference in decision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 214, + 237, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 214, + 237, + 221 + ], + "spans": [ + { + "bbox": [ + 149, + 214, + 237, + 221 + ], + "type": "text", + "content": "- Two responses from different models." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 149, + 231, + 227, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 231, + 227, + 240 + ], + "spans": [ + { + "bbox": [ + 149, + 231, + 227, + 240 + ], + "type": "text", + "content": "Annotation Standards" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 148, + 248, + 444, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 248, + 444, + 266 + ], + "spans": [ + { + "bbox": [ + 148, + 248, + 444, + 266 + ], + "type": "text", + "content": "The annotator marks which model of the given two is better on three different categories: Fluency and Relevance. The deifintion of each category is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 149, + 273, + 361, + 289 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 149, + 273, + 361, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 273, + 361, + 280 + ], + "spans": [ + { + "bbox": [ + 149, + 273, + 361, + 280 + ], + "type": "text", + "content": "- Fluency measures the correctness, completeness, and semantic consistency of the response." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 149, + 281, + 357, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 281, + 357, + 289 + ], + "spans": [ + { + "bbox": [ + 149, + 281, + 357, + 289 + ], + "type": "text", + "content": "- Relevance measures the degree of relevance of the knowledge and response to the history." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 149, + 325, + 172, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 325, + 172, + 333 + ], + "spans": [ + { + "bbox": [ + 149, + 325, + 172, + 333 + ], + "type": "text", + "content": "History:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 149, + 335, + 172, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 335, + 172, + 344 + ], + "spans": [ + { + "bbox": [ + 149, + 335, + 172, + 344 + ], + "type": "text", + "content": "Singing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 149, + 346, + 408, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 346, + 408, + 354 + ], + "spans": [ + { + "bbox": [ + 149, + 346, + 408, + 354 + ], + "type": "text", + "content": "I love to sing. Some of my favorite genres to sing are blues, pop, and rock. Do you like to sing?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 149, + 356, + 347, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 356, + 347, + 365 + ], + "spans": [ + { + "bbox": [ + 149, + 356, + 347, + 365 + ], + "type": "text", + "content": "Yes very much so. I love older artists like Ray Charles and Billie Holiday." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 149, + 367, + 425, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 367, + 425, + 376 + ], + "spans": [ + { + "bbox": [ + 149, + 367, + 425, + 376 + ], + "type": "text", + "content": "Oh yeah I love Billie Holiday. She had no formal music training but her vocal delivery made up for it." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 149, + 385, + 180, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 385, + 180, + 393 + ], + "spans": [ + { + "bbox": [ + 149, + 385, + 180, + 393 + ], + "type": "text", + "content": "Input Text:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 149, + 395, + 320, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 395, + 320, + 404 + ], + "spans": [ + { + "bbox": [ + 149, + 395, + 320, + 404 + ], + "type": "text", + "content": "I've been getting into the strokes as of late. Do you like them?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 149, + 412, + 174, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 412, + 174, + 418 + ], + "spans": [ + { + "bbox": [ + 149, + 412, + 174, + 418 + ], + "type": "text", + "content": "Reference:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 149, + 420, + 439, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 420, + 439, + 436 + ], + "spans": [ + { + "bbox": [ + 149, + 420, + 439, + 436 + ], + "type": "text", + "content": "I used to listen to The Strokes a lot back in high school. Their debut album was added to Rolling Stones 500 greatest albums of all time at number 199." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 149, + 472, + 201, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 472, + 201, + 480 + ], + "spans": [ + { + "bbox": [ + 149, + 472, + 201, + 480 + ], + "type": "text", + "content": "Model A Response:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 149, + 482, + 363, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 482, + 363, + 491 + ], + "spans": [ + { + "bbox": [ + 149, + 482, + 363, + 491 + ], + "type": "text", + "content": "Yes I do like them. I like all kinds of music, but I'm not a big fan of the Strokes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 149, + 500, + 202, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 500, + 202, + 508 + ], + "spans": [ + { + "bbox": [ + 149, + 500, + 202, + 508 + ], + "type": "text", + "content": "Model B Response:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 149, + 510, + 427, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 510, + 427, + 529 + ], + "spans": [ + { + "bbox": [ + 149, + 510, + 427, + 529 + ], + "type": "text", + "content": "I do like them. I also like to wear my Strokes shirt during May because it is national stroke awareness month." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 149, + 546, + 215, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 546, + 215, + 559 + ], + "spans": [ + { + "bbox": [ + 149, + 546, + 215, + 559 + ], + "type": "text", + "content": "Annotation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 149, + 568, + 171, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 568, + 171, + 576 + ], + "spans": [ + { + "bbox": [ + 149, + 568, + 171, + 576 + ], + "type": "text", + "content": "Fluency" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 150, + 581, + 235, + 644 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 150, + 581, + 235, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 581, + 235, + 591 + ], + "spans": [ + { + "bbox": [ + 150, + 581, + 235, + 591 + ], + "type": "text", + "content": "Both responses are fluent." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 150, + 592, + 233, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 592, + 233, + 611 + ], + "spans": [ + { + "bbox": [ + 150, + 592, + 233, + 611 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 150, + 592, + 233, + 611 + ], + "type": "text", + "content": " Model A response is more fluent." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 150, + 613, + 233, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 613, + 233, + 632 + ], + "spans": [ + { + "bbox": [ + 150, + 613, + 233, + 632 + ], + "type": "text", + "content": "Model B response is more fluent." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 150, + 634, + 235, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 634, + 235, + 644 + ], + "spans": [ + { + "bbox": [ + 150, + 634, + 235, + 644 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 150, + 634, + 235, + 644 + ], + "type": "text", + "content": " Neither response is fluent." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 155, + 665, + 181, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 665, + 181, + 671 + ], + "spans": [ + { + "bbox": [ + 155, + 665, + 181, + 671 + ], + "type": "text", + "content": "Previous" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 249, + 568, + 276, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 568, + 276, + 576 + ], + "spans": [ + { + "bbox": [ + 249, + 568, + 276, + 576 + ], + "type": "text", + "content": "Relevance" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 249, + 581, + 333, + 654 + ], + "type": "list", + "angle": 0, + "index": 41, + "blocks": [ + { + "bbox": [ + 249, + 581, + 324, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 581, + 324, + 601 + ], + "spans": [ + { + "bbox": [ + 249, + 581, + 324, + 601 + ], + "type": "text", + "content": "Both knowledges and responses are relevant" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 249, + 602, + 332, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 602, + 332, + 621 + ], + "spans": [ + { + "bbox": [ + 249, + 602, + 332, + 621 + ], + "type": "text", + "content": "Model A response is more relevant." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 249, + 624, + 333, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 624, + 333, + 642 + ], + "spans": [ + { + "bbox": [ + 249, + 624, + 333, + 642 + ], + "type": "text", + "content": "Model B response is more relevant." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 249, + 645, + 314, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 249, + 645, + 314, + 654 + ], + "spans": [ + { + "bbox": [ + 249, + 645, + 314, + 654 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 249, + 645, + 314, + 654 + ], + "type": "text", + "content": " Neither is relevant." + } + ] + } + ], + "index": 40 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 665, + 319, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 319, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 319, + 671 + ], + "type": "text", + "content": "Next" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 149, + 746, + 200, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 746, + 200, + 754 + ], + "spans": [ + { + "bbox": [ + 149, + 746, + 200, + 754 + ], + "type": "text", + "content": "Made with Streamlit" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 208, + 767, + 384, + 779 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 767, + 384, + 779 + ], + "spans": [ + { + "bbox": [ + 208, + 767, + 384, + 779 + ], + "type": "text", + "content": "Figure 2: Screenshot of the annotation tool." + } + ] + } + ], + "index": 44, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 111, + 77, + 117, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 77, + 117, + 84 + ], + "spans": [ + { + "bbox": [ + 111, + 77, + 117, + 84 + ], + "type": "text", + "content": ">" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 475, + 77, + 484, + 84 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 77, + 484, + 84 + ], + "spans": [ + { + "bbox": [ + 475, + 77, + 484, + 84 + ], + "type": "text", + "content": "三" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2702" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 19 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_content_list.json b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a6c8a0426a7e354412bc82d901c0f13fa47884e8 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_content_list.json @@ -0,0 +1,1718 @@ +[ + { + "type": "text", + "text": "Efficient Long-Range Transformers: You Need to Attend More, but Not Necessarily at Every Layer", + "text_level": 1, + "bbox": [ + 171, + 89, + 831, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingru Zhang†, Dhananjay Ram, Cole Hawkins, Sheng Zha, Tuo Zhao", + "bbox": [ + 171, + 148, + 835, + 165 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{\\dagger}$ Georgia Institute of Technology $\\diamond$ Amazon Web Service", + "bbox": [ + 268, + 166, + 734, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{qingru.zhang,tourzhao}@gatech.edu", + "bbox": [ + 327, + 184, + 673, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{radhna,colehawk,zhasheng}@amazon.com", + "bbox": [ + 314, + 200, + 689, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 342, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pretrained transformer models have demonstrated remarkable performance across various natural language processing tasks. These models leverage the attention mechanism to capture long- and short-range dependencies in the sequence. However, the (full) attention mechanism incurs high computational cost – quadratic in the sequence length, which is not affordable in tasks with long sequences, e.g., inputs with 8k tokens. Although sparse attention can be used to improve computational efficiency, as suggested in existing work, it has limited modeling capacity and often fails to capture complicated dependencies in long sequences. To tackle this challenge, we propose MASFormer, an easy-to-implement transformer variant with Mixed Attention Spans. Specifically, MASFormer is equipped with full attention to capture long-range dependencies, but only at a small number of layers. For the remaining layers, MASformer only employs sparse attention to capture short-range dependencies. Our experiments on natural language modeling and generation tasks show that a decoder-only MASFormer model of 1.3B parameters can achieve competitive performance to vanilla transformers with full attention while significantly reducing computational cost (up to $75\\%$ ). Additionally, we investigate the effectiveness of continual training with long sequence data and how sequence length impacts downstream generation performance, which may be of independent interest.", + "bbox": [ + 144, + 274, + 460, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 751, + 260, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained transformer models have manifested superior performance in various natural language processing tasks such as natural language modeling (NLM) (Dai et al., 2019; Radford et al., 2019), natural language generation (NLG) (Brown et al., 2020) and natural language understanding (NU) (Devlin et al., 2019; Liu et al., 2019; He et al., 2021b).", + "bbox": [ + 112, + 775, + 490, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These models leverage the attention mechanism (Vaswani et al., 2017) to compute the dependency score for each pair of tokens in an input sequence.", + "bbox": [ + 507, + 253, + 884, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some practical tasks require these transformer models to handle long-sequence inputs like 8k tokens. For example, chatbot systems gather long-term contexts of user interactions to generate informative texts (Roller et al., 2021). Summarization for news, government reports, and academic papers request models to take inputs of long sequences to generate comprehensive summaries (Shaham et al., 2022), otherwise models often miss important information. Note that typical transformer models apply full attention to capture token dependencies pair-wise. It leads to a quadratic time and space complexity w.r.t. input length. However, such a complexity is prohibitive for long sequences. In particular, it incurs massive memory consumption during the back propagation. For example, a transformer model with 250M parameters consumes over 80G GPU memory when sequence length is 8k (Zuo et al., 2022).", + "bbox": [ + 507, + 304, + 885, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this scalability issue, various approaches have been proposed to reduce the complexity. One approach is sparse attention, which restricts each token to attend a subset of tokens based on predefined sparsity patterns (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020). For instance, block sparse attention (Kitaev et al., 2020; Ma et al., 2023) divides the input sequence into several blocks, and only intra-block attention is performed. Besides, sliding-window attention (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020) allows each token to attend to its neighboring tokens within a sliding window. These methods, though reducing the complexity of full attention, cannot sufficiently capture long-range dependencies. Other variants, such as kernel approximation (Peng et al., 2021) and low-rank approximation (Wang et al., 2020; Chen et al., 2021) methods, share the similar spirit and drawbacks. To com", + "bbox": [ + 507, + 613, + 885, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work was done during Qingru Zhang's internship at Amazon Web Service.", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2775", + "bbox": [ + 478, + 927, + 521, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2775-2786 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 216, + 945, + 779, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pensate for the lack of long-range dependencies, LongT5 (Guo et al., 2021) introduces global tokens that are obtained by average pooling on every block of tokens (Ainslie et al., 2020). However, the block pooling operations can weaken the signal of crucial tokens and prevent the long-range dependencies from being detected.", + "bbox": [ + 112, + 84, + 487, + 197 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to these methods, state space models (SSMs) prespecify global dependency patterns to capture the long-range dependencies only (Gu et al., 2020, 2021; Li et al., 2022; Zuo et al., 2022; Ma et al., 2023; Smith et al., 2023). These models can be regarded as linear recurrent neural networks with specifically designed fixed weights. As tailored for global dependencies, SSMs fail to effectively capture local dependencies. In order to combine both local and global dependencies, SPADE (Zuo et al., 2022) and MEGA (Ma et al., 2023) augment SSM layers into transformer layers equipped with local attention. However, state space methods require sophisticated implementation, and often encounter computational instability during the back propagation, especially when scaling up to large model size (Gupta et al., 2022). SPADE and MEGA hence inherit these drawbacks.", + "bbox": [ + 115, + 204, + 489, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that the aforementioned methods apply same attention mechanism for every layer. We challenge this conventional wisdom and propose a transformer variant - MASFormer (Mixed Attention Span Transformer). MASFormer utilizes full attention only at a subset of layers whereas employs sparse attention at the remaining layers. Our design is motivated by the phenomenon - that most contexts in NLP data display a great deal of locality of reference (Zaheer et al., 2020; Beltagy et al., 2020). That is, most of information about a token can be derived from its neighboring tokens. In contrast, long-range dependencies among tokens are sparse and infrequent. Consider an academic paper as an example. Within a paragraph, there exist numerous short-term dependencies. Neighboring tokens are closely connected to convey meaningful semantics. Across paragraphs, there can be a small number of long-range dependencies. For example, tokens associated to the primary theme of the paper exhibit rare and weak dependencies across a long span. Since long-range dependencies occur much less frequently, a few layers of full attention are adequate to capture them. In stark contrast, short-term dependencies are more frequent, necessitating local attention in the majority of layers to fully extract", + "bbox": [ + 115, + 501, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these signals.", + "bbox": [ + 509, + 84, + 611, + 99 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of MASFormer, We conduct experiments on natural language modeling (ArXiv and PubMed Cohen et al. (2018)) and natural language generation (ArXiv, Cohen et al. (2018) and SCROLLS, Shaham et al. (2022)) tasks. Specifically, we compare the performance of MASFormer to other attention methods using a pretrained GPT-2 model (Radford et al., 2019) of 1.3 billion parameters. Our empirical results demonstrate that MASFormer consistently outperforms baseline methods across different attention cost (i.e. the total number of computed attention scores). In particular, MASFormer can achieve comparable performance to full attention while significantly reducing the computational cost. For example, with $27\\%$ of its attention cost, MASFormer achieves a close R2 score as full attention on QMSUM dataset.", + "bbox": [ + 507, + 107, + 884, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also make additional discoveries with MASFormer, which are of independent interest. Firstly, we investigate the effectiveness of continual training for long sequence modeling. Many publicly available models are pre-trained with sequences shorter than 2048, and often fail to perform well on longer sequences (e.g. 8k/16k tokens). To bridge the gap, we explore the option of continual training to adapt these models to long sequences, thereby avoiding pre-training from the scratch. We discuss its effectiveness with MASFormer in Section 4.3. Secondly, we showcase that increasing sequence length can yield more performance gains on downstream tasks than NLM tasks evaluated by perplexity. We are aware of the recent findings by Sun et al. (2021) that increasing context length exhibits limited impact on NLM perplexity. Nevertheless, when applying MASFormer to downstream tasks like long-context summarization, we find that model performance benefits significantly from extending context length. Such a difference arises from the fact that predicting the next tokens in NLM primarily relies on locality of reference. Capturing infrequent long-range tokens can improve perplexity but not significantly. Therefore, we emphasize the necessity to evaluate model performance on downstream tasks that require long-range dependencies. Furthermore, our empirical evidence suggests that increasing the length can improve the performance only if models possess sufficient capability to handle additional long-range information. Local attention, as a counterexample, often fails to capture long-range signals and hence", + "bbox": [ + 507, + 385, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2776", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "benefits much less from long sequences.", + "bbox": [ + 112, + 84, + 413, + 99 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 Background", + "text_level": 1, + "bbox": [ + 112, + 112, + 255, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1 Pretrained Language Models", + "text_level": 1, + "bbox": [ + 112, + 137, + 389, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained transformer models (Devlin et al., 2019; Liu et al., 2019; Brown et al., 2020; Dosovitskiy et al., 2020; He et al., 2021b,a) have manifested superior performance in various NLP tasks. These models are often pre-trained on enormous amounts of unlabeled data in a unsupervised/self-supervised manner such that they can learn rich semantic knowledge. By further fine-tuning these pre-trained models, we can effectively transfer such knowledge to benefit downstream tasks (Zhang et al., 2023).", + "bbox": [ + 112, + 159, + 487, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing research on long-range transformers commonly requires pre-training the proposed models from scratch to accommodate new architectures and long inputs (Guo et al., 2021; Zuo et al., 2022). However, the significant training overheads raise a barrier for the widespread utilization of these methods across different language models. Motivated by this, we explore the possibility of leveraging existing pre-trained models and adapting them to long sequences though continual training.", + "bbox": [ + 112, + 321, + 487, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Attention Mechanism", + "text_level": 1, + "bbox": [ + 112, + 493, + 331, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose the input to the layer is $X\\in \\mathbb{R}^{n\\times d}$ , where $n$ is the input sequence length and $d$ is embedding dimension, then self-attention mechanism outputs", + "bbox": [ + 112, + 513, + 485, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} (X) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 569, + 487, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q = XW_{q}, K = XW_{k}, V = VW_{v}$ and $W_{q}, W_{k}, W_{v}$ are learnable projection weights. Such full attention can simultaneously evaluate the alignment between any pair of tokens in the sequence. Specifically, denote the attention score $A = \\text{softmax}(QK^{\\top} / \\sqrt{d})$ , then $A_{ij}$ captures the alignment between tokens $i$ and $j$ . A typical transformer model applies the full attention at every layer. Denote the number of layers as $L$ . Then its attention cost is $Ln^{2}$ .", + "bbox": [ + 112, + 612, + 487, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse attention variants are introduced to mitigate the computational cost of full attention. Figures 1a and 1b illustrates the attention patterns of block sparse attention and sliding-window attention. For instance, block sparse attention divides tokens into blocks of size $b$ and performs intrablock attention only, resulting in an attention cost of $bn$ . Sliding-window attention allows each token to attend its left/right neighboring tokens within a", + "bbox": [ + 112, + 774, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "local window of size $w$ . In most of cases, block sparse attention exhibits similar performance as sliding-window attention (Zuo et al., 2022).", + "bbox": [ + 507, + 84, + 880, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Our Approach", + "text_level": 1, + "bbox": [ + 507, + 140, + 672, + 156 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our method - MASFormer, a long-range transformer variant that mixes different attention spans across layers.", + "bbox": [ + 507, + 162, + 882, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 MASFormer: Mixed Attention Span", + "text_level": 1, + "bbox": [ + 507, + 216, + 840, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MASFormer leverages full attention exclusively at a subset of transformer layers, whereas it employs block sparse attention at the remaining layers. The structure of MASFormer is illustrated in Figure 1c. We choose full attention to encode long-range information due to the following reasons: (i) full attention exhibits superior capability to capture long-range dependencies compared to sparse attention; (ii) full attention does not require sophisticated implementation and hence is computationally stable compared to SSMs (Zuo et al., 2022; Gupta et al., 2022); (iii) full attention is compatible with existing pre-trained transformer models, enabling us to conduct continual training which we elaborate in Section 3.2. To mitigate the computational cost, we restrict the number of layers using full attention.", + "bbox": [ + 505, + 237, + 882, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MASFormer is motivated by empirical investigations on performance comparison between models that apply the same attention span at every layer. Figure 2 presents the performance of block sparse attention and full attention on language modeling and summarization tasks. We find that, given long-sequence inputs, sparse attention is often insufficient to capture long-range dependencies beyond its attention span. As a result, it shows unsatisfactory performance. To remedy it, one can either increase attention span or switch to full attention to improve model capability of capturing sophisticated dependencies. Though improving model performance, it incurs high computational cost.", + "bbox": [ + 507, + 495, + 882, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Confronting such a trade-off between computational cost and model performance, we challenge the common practice - that applies the same attention span at every layer. MASFormer provides an alternative solution. Instead of increasing attention span evenly, MASFormer allocates a large portion of attention computations to a subset of $l$ layers by equipping them with full attention. Specifically, equipping bottom layers with full attention can yield the best performance as suggested by our empirical analysis in Section 4.31. At the remain-", + "bbox": [ + 507, + 720, + 882, + 896 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Please see Section 4.3 for detailed explanations", + "bbox": [ + 529, + 904, + 826, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2777", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b14e862e159937c0a8d0aba62f3dcf18dc7295a58d45f99d7a2c8abb6f7d2625.jpg", + "image_caption": [ + "(a) Block Attention" + ], + "image_footnote": [], + "bbox": [ + 117, + 112, + 255, + 211 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e385a1ca1c74eb15b1f237f6586b185c1f36c971d83bf972e19739324ca47bb6.jpg", + "image_caption": [ + "(b) Window Attention" + ], + "image_footnote": [], + "bbox": [ + 265, + 112, + 406, + 211 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e07b6b4719286115e818c89a0512ae2593a10b031ab89993a43a5e73348c59ee.jpg", + "image_caption": [ + "(c) MASFormer" + ], + "image_footnote": [], + "bbox": [ + 423, + 85, + 880, + 233 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cdc86d8d963cb9cd155602b3ad9c20143a6c51bbf63bd87fc5c7884176cbac8c.jpg", + "image_caption": [ + "Figure 1: Illustration of attention patterns of (a) block sparse attention with block size $b = 3$ ; (b) sliding-window attention with window size $w = 1$ (on each side); (c) MASFormer that integrates full and sparse attention.", + "(a) NLM for ArXiv", + "Figure 2: (a,b): We evaluate the perplexity of a pre-trained GPT-2 model with block attention of differnet block size after continual training. (c,d): We fine-tune a GPT-2 model with block attention and compare the summarization performance on ArXiv and GovReport under different block size. Here the input length $n$ is 8192." + ], + "image_footnote": [], + "bbox": [ + 124, + 300, + 297, + 394 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9ac5c06d068f9a7c4a4b26666d22d047f491027c96e6ad0643d82ebc258f6e7.jpg", + "image_caption": [ + "(b) NLM for PubMed" + ], + "image_footnote": [], + "bbox": [ + 317, + 300, + 489, + 394 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3d353730f2b9c2e78c69ce82fb754ba1bc1c9b583bdf4740544689399a05475b.jpg", + "image_caption": [ + "(c) Sum. for ArXiv" + ], + "image_footnote": [], + "bbox": [ + 507, + 300, + 680, + 394 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/76170219267e1227827b845500c7f6484232a8edfc99eafc43cbac1dbf902c3d.jpg", + "image_caption": [ + "(d) Sum. for GovReport" + ], + "image_footnote": [], + "bbox": [ + 700, + 300, + 870, + 394 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing layers, MASFormer utilizes block attention of small size $m$ , resulting in a controlled attention cost of $(L - l)mn + ln^2$ . As mentioned in Section 1, such a design is inspired by the phenomenon that most of contexts in NLP data exhibit a great deal of locality of reference. Long-range dependencies, in contrast, are less frequent. Therefore, it is not necessary to enhance attention span at every layer. Instead, a few layers of full attention are sufficient to capture infrequent long-range signals. The majority of layers can maintain small attention spans to adequately extract local dependencies and control the attention cost.", + "bbox": [ + 112, + 466, + 487, + 673 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our empirical results demonstrate that, with the same attention cost, MASFormer significantly outperforms sparse attention. Remarkably, MASFormer can achieve comparable performance to full attention while substantially reducing computational cost. Therefore, by mixing different attention spans, MASFormer strikes a better balance between computational cost and model performance.", + "bbox": [ + 112, + 676, + 487, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, MASFormer offers additional implementation advantages. As using the same attention function, MASFormer is easy to implement and compatible with existing pre-trained models. We can build MASFormer upon pre-trained transformers by changing their attention patterns, which does not involve modification on model architectures", + "bbox": [ + 112, + 806, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and pre-trained weights. Meanwhile, acceleration packages, such as FlashAttention (Dao et al., 2022) and xFormers (Lefaudeauux et al., 2022), are applicable to further accelerate the computation of block attention and full attention in MASFormer.", + "bbox": [ + 507, + 466, + 882, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Continual Training with Long Sequences", + "text_level": 1, + "bbox": [ + 507, + 558, + 877, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned, MASFormer can be implemented upon majority of pre-trained transformers by modifying their attention patterns. However, most of publicly available models are pre-trained with sequences shorter than 2048, and often exhibit subpar performance on longer sequences such as $8\\mathrm{k} / 16\\mathrm{k}$ . To bridge this gap, we propose the continual training to adapt the revised model on long sequences and new attention pattern. As such, we can preserve existing pre-trained knowledge and circumvent the intensive overheads of pre-training from scratch. In particular, we first modify the attention pattern of the target model as proposed by MASFormer. If the pre-trained model uses absolute position embeddings, we duplicate them to accommodate long sequences. Subsequently, we provide the revised model with long sequences (e.g., 8k) from pretraining corpus like PILE (Gao et al., 2020). Then we conduct continual pre-training using casual language modeling (CLM) objective. We discuss the effectiveness of continual training in Section 4.3.", + "bbox": [ + 505, + 581, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2778", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 112, + 84, + 258, + 99 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness and efficiency of MASFormer on natural language modeling (ArXiv and PubMed, Cohan et al. (2018)), natural language generation (ArXiv Cohan et al. (2018), QMSUM and GovReport Shaham et al. (2022)). We choose the GPT-3 XL model architecture (Brown et al., 2020) as our base model, which consists of 1.3 billion parameters and 24 layers and is pre-trained on PIL (Gao et al., 2020) for 300 billion tokens. GPT is a general purpose model that can be applied to many tasks instead of tailoring them for specific tasks. As such, it makes easy to control experiments and showcase the difference among various methods.", + "bbox": [ + 112, + 105, + 489, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our base model uses absolute positional embeddings with maximum length 1024. To accommodate longer inputs, we duplicate its positional embeddings to have the maximum length as 8192 such that the model can handle sequences containing up to 8192 tokens. Then, we implement different attention methods by modifying the attention pattern of the base model. We implement all the models with PyTorch (Paszke et al., 2019). All the experiments are conducted on NVIDIA A100 GPUs.", + "bbox": [ + 112, + 332, + 489, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continual Training Details. After changing the attention pattern, we conduct the continual training for MASFormer and baseline methods on PILE corpus (Gao et al., 2020) to adapt the revised models to new attention patterns and long-sequence inputs. We leverage the casual language modeling (CLM) objective to train the model for 50,000 steps with a warmup of 2000 steps. We set the input length as 8192 and use a batch size of 128 such that the models are optimized with 1M tokens per step. We use the constant learning 0.0001 for all methods.", + "bbox": [ + 112, + 511, + 489, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. We compare MASFormer with the following methods:", + "bbox": [ + 112, + 690, + 489, + 721 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- All full attention is to apply full attention at every layer. It has been adopted by most of existing transformer models as default. Although incurring the maximum attention cost, it achieves the best performance for most of our tasks. Hence, it acts as an upper bound for other methods.", + "- All block sparse attention is to apply block attention at every layer, which is an effective method to reduce computational cost when modeling long sequences. Block attention sets the attention span of each layer identical such that it evenly distributes the budget of attention computation across layers." + ], + "bbox": [ + 112, + 724, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- All sliding-window attention is to apply sliding-window attention at every layer, which is another variant of sparse attention. It shares the similar spirits and often performs similarly as block attention.", + "bbox": [ + 507, + 84, + 884, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following experiments, we compare MAS-Former and the baseline methods across different attention cost $\\mathcal{C}$ . That is, for all block sparse attention, we set the block size as $b = \\mathcal{C} / (Ln)$ . For all sliding-window attention, we choose the window size as $w = \\mathcal{C} / (2Ln)$ . For MASFormer, we apply a small block size $m = 1024$ for its block attention and set $l$ as $(\\mathcal{C} - Lmn) / (n^2 - mn)$ . Then we observe how their performance evolves when enhancing the attention cost $\\mathcal{C}$ or input length $n$ .", + "bbox": [ + 507, + 149, + 884, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment Overview. We briefly summarize the experimental contents as follows:", + "bbox": [ + 507, + 311, + 882, + 342 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Section 4.1 presents the perplexity evaluation of all the models on ArXiv and PubMed after continual training.", + "- Section 4.2 compares the summarization performance of the models on ArXiv, QMSUM, and GovReport after fine-tuning. Besides, we also discuss the difference between perplexity and downstream evaluation in reflecting model capacity to capture long-range dependencies.", + "- Section 4.3 provides three crucial analyses: (i) we evaluate the benefits of increasing input length and discuss the requirements to attain these gains; (ii) we analyze the effectiveness of continual training for long-sequence modeling; (iii) we conduct an ablation study to demonstrate that equipping bottom layers with full attention yields the most significant performance gains than other options. We further provide the explanations." + ], + "bbox": [ + 507, + 344, + 885, + 634 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Natural Language Modeling", + "text_level": 1, + "bbox": [ + 507, + 649, + 779, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Datasets and Evaluation Details", + "text_level": 1, + "bbox": [ + 507, + 671, + 818, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the perplexity of the updated GPT-2 for each attention method after continual training. The evaluation is conducted on test sets of ArXiv and PubMed (Cohan et al., 2018). Table 5 presents the statistics of these two datasets. Pubmed consists of scientific documents, with a document's content used as input and its corresponding abstract as the target summary. ArXiv is similar to PubMed, with documents from arXiv.", + "bbox": [ + 507, + 693, + 884, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Details. We conduct the perplexity evaluation under two settings. (i) We calculate the perplexity (ppl.) with all documents from test sets. Table 1 presents the overall perplexity of different models on two datasets. (ii) To showcase", + "bbox": [ + 507, + 838, + 882, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2779", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aa87e531b990436c086b6417eaebac905aedebcd05996b5827a767377912c034.jpg", + "image_caption": [ + "(a) ArXiv ppl." + ], + "image_footnote": [], + "bbox": [ + 215, + 67, + 487, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea1a52f7f7b59921579ce8383a2cd331d6c63cbfdf1ca31e5cfd12ab892255b2.jpg", + "image_caption": [ + "(b) PubMed ppl.", + "Figure 3: Perplexity evaluation on ArXiv and PubMed with examples of different length. Here $x$ -axis is the maximum document length of each subset, i.e., $k \\times 1024$ ( $k = 1, 2, 3, \\ldots$ )." + ], + "image_footnote": [], + "bbox": [ + 507, + 67, + 781, + 214 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the varying behaviors of models on documents of different length, we divide all documents into several subsets according to their length. Each subset consists of examples, whose length is within $((k - 1)\\times 1024,k\\times 1024](k = 1,2,3,\\ldots)$ . Then, we evaluate the perplexity on each subset. Figure 3 presents the perplexity of models on different subsets of examples.", + "bbox": [ + 110, + 266, + 489, + 394 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Results", + "text_level": 1, + "bbox": [ + 112, + 403, + 231, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares the overall perplexity on test sets of ArXiv and PubMed. The results suggest that, with $l = 4$ layers of full attention, MASFormer achieves comparable performance to all full attention, while reducing $72\\%$ of its attention cost. With the similar attention cost $\\mathcal{C}$ , MASFormer outperforms all block attention that evenly distributes the budget of attention computation. For example, MASFormer with $l = 2$ achieves 8.75 ppl. on PubMed, which is 1.37 lower than that of block attention of $b = 2048$ .", + "bbox": [ + 112, + 420, + 489, + 598 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26632b80afb2e7994ce3689c38d4f0d3f887fcfc03e3083b578466dbd64ea0c7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCArXivPubMed
Full attention1,610M8.637.63
Block (b=1024)201M13.1912.19
Block (b=2048)402M10.7510.13
MASFormer (l=2)318M10.258.75
MASFormer (l=4)436M9.318.25
MASFormer (l=8)671M9.638.25
", + "bbox": [ + 119, + 606, + 480, + 749 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Perplexity evaluation on ArXiv and PubMed.", + "bbox": [ + 115, + 752, + 482, + 766 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 illustrates the perplexity variation of each method given examples of different length. We can tell that MASFormer and full attention show better performance on longer documents, suggesting increasing context length can improve their prediction performance. Full attention, though incurring the highest attention cost, always achieves the best performance due to its outstanding capability to handle sophisticated dependencies. Notably,", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $27\\%$ of its attention cost, MASFormer exhibits a curve of ppl. v.s. length that closely resembles to that of full attention. This demonstrates the effectiveness and efficiency of MASFormer to capture long-range dependencies. In contrast, block sparse attention benefits much less from long contexts and underperforms both of them because of its incapability to encode long-range signals. For example, when $b = 1024$ , block attention achieves similar perplexity on PubMed examples of different length.", + "bbox": [ + 507, + 266, + 884, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Natural Language Generation", + "text_level": 1, + "bbox": [ + 507, + 437, + 793, + 453 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Datasets and Training Details", + "text_level": 1, + "bbox": [ + 507, + 458, + 801, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the downstream performance of models on several abstractive summarization tasks to compare their capability of handling long sequences in practice. Specifically, we fine-tune models on ArXiv (Cohan et al., 2018), QM-SUM and GovReport (from SCROLLS benchmark, Shaham et al. (2022)). Their statistics are summarized in Table 5. We mainly use ROUGE-2 (R2) score (Lin, 2004) as the evaluation metric, which is more important and sensitive than R1 and RL.", + "bbox": [ + 507, + 478, + 884, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. After continual training, we fine-tune each model and report R2 scores on validation sets. Specifically, we fine-tune models for 3000 steps on QMSUM, 8000 steps on GovReport, and 12000 steps on ArXiv. We set the batch size as 64 for ArXiv and 32 for QMSUM and GovReport. We pick the learning rates from $\\{1 \\times 10^{-5}, 5 \\times 10^{-5}, 1 \\times 10^{-4}, 5 \\times 10^{-4}\\}$ , and choose the optimal ones to report the performance of each method. Moreover, the input length is fixed as 8192. We apply the greedy decoding for generation. Please see Appendix B for more details.", + "bbox": [ + 507, + 642, + 882, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Results", + "text_level": 1, + "bbox": [ + 507, + 845, + 626, + 858 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table $2^{2}$ and Figure 4, we present the fine-tuning results on QMSUM, ArXiv and GovReport across", + "bbox": [ + 507, + 863, + 880, + 895 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Please see Table 6 in Appendix B for all ROUGE scores.", + "bbox": [ + 529, + 903, + 880, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2780", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7abd031293b0a106fc32b6b3eda383c7fe114cdd09263d0c7c29802ada1078f9.jpg", + "image_caption": [ + "(a) ArXiv" + ], + "image_footnote": [], + "bbox": [ + 134, + 82, + 364, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/06f3fa19b661a64eedbac59b93eed38cff3357da2c5afb8d0725d3b7d2bb4714.jpg", + "image_caption": [ + "(b) QMSUM" + ], + "image_footnote": [], + "bbox": [ + 381, + 80, + 611, + 206 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e37f4b3e56d841f515e8e7c046b7e5e493dd9309d4c550d1c24329e45ed393b.jpg", + "image_caption": [ + "(c) GovReport", + "Figure 4: Given input length as 8192, we compare summarization performance between MASFormer and block/full attention when increasing the attention cost." + ], + "image_footnote": [], + "bbox": [ + 630, + 82, + 860, + 208 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bec286d385d911454869fb10e215deec0e4e7ee116298b8d8c88df70de43888b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsCQMSUMArXivGovReport
Full attention1610M8.0019.3228.83
Window(w=1024)402M4.3213.5117.03
Block(b=2048)402M5.039.6112.31
MASFormer(l=4)436M6.5914.9118.82
Window(w=2048)805M5.0515.2122.79
Block(b=4096)805M5.1514.5023.64
MASFormer(l=6)553M7.1515.7221.20
MASFormer(l=8)671M7.4617.0024.42
MASFormer(l=12)906M8.7018.5826.26
", + "bbox": [ + 115, + 272, + 489, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: Summarization performance of models with different attention methods. The best results are shown in bold.", + "bbox": [ + 112, + 475, + 487, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different attention cost. The results demonstrate that, with the similar attention cost, MASFormer significantly outperforms sparse attention variants. Furthermore, when enhancing attention cost, MASFormer achieves greater performance gains than sparse attention methods. This is evident from the steeper slope of its R2 curve versus attention cost, in contrast to the baseline method. For example, when increasing $\\mathcal{C}$ form 553M to 671M, the R2 score of MASFormer on QMSUM exhibits a substantial improvement, reaching 8.70 from 7.46. Remarkably, this score surpasses even that of full attention. Therefore, MASFormer addresses the trade-off between computational cost and performance gains in a more efficient and effective way.", + "bbox": [ + 112, + 529, + 489, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notice that, in order to achieve comparable summarization performance to full attention, MAS-Former needs at least $l = 8$ layers of full attention, and providing more can lead to more gains. This observation is different from the findings in NLM (Figure 3) that increasing $l$ beyond 4 provides limited improvement in perplexity. Their different capacity requirements arise from the fact that predicting next tokens in NLM primarily relies on lo", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cal dependencies. Capturing infrequent long-range tokens does not significantly improve perplexity. Thus, this discrepancy emphasizes the necessity to evaluate long-range models on downstream tasks.", + "bbox": [ + 507, + 267, + 884, + 331 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Analysis", + "text_level": 1, + "bbox": [ + 507, + 344, + 621, + 360 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 Benefits of Increasing Sequence Length", + "text_level": 1, + "bbox": [ + 507, + 367, + 877, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the benefits of increasing input length for downstream performance. Specifically, we select the input length from $\\{2048, 4096, 6144, 8192\\}$ and present the fine-tuning performance of full attention in Figure 5. The results consistently demonstrate that as the input length increases, the model's performance improves. That is, downstream performance benefits significantly from long-sequence inputs. In contrast, increasing example length beyond $6k$ results in marginal improvements in perplexity (See Figure 3), highlighting again the importance of downstream evaluation.", + "bbox": [ + 505, + 388, + 884, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, when comparing the behaviors of block attention in Figure 2c and 2d, we find that sparse attention often insufficiently capitalize on the benefits offered by longer inputs. For instance, given block size as 4096, its performance on ArXiv remains nearly unchanged when increasing input length from 4096 (R2 = 15.52 in Figure 5a) to 8192 (R2 = 14.49 in Figure 2c). This finding suggests that enhancing input length can only improve model performance if the model possesses the sufficient capability to handle long-range dependencies.", + "bbox": [ + 507, + 598, + 885, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 Effectiveness of Continual Training", + "text_level": 1, + "bbox": [ + 507, + 785, + 847, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the effectiveness of continual training by comparing fine-tuning performance of MAS-Former $(l = 8)$ under the following settings: (i) MASFormer without continual training (w.o. C.T.); (ii) MASFormer continually trained with short inputs (C.T. $(n = 2048)$ ); (iii) MASFormer continually trained with long inputs (C.T. $(n = 8192)$ ). Table 3", + "bbox": [ + 507, + 806, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2781", + "bbox": [ + 480, + 927, + 517, + 940 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34f8b43ff943ff02432a8efcea092d5101ce4681ebe8062ad2ef1f4556cd815e.jpg", + "image_caption": [ + "(a) ArXiv" + ], + "image_footnote": [], + "bbox": [ + 134, + 80, + 364, + 206 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/243cd9585f95f0cec8f02ae7e098f293e5b08b58846d6884995bfa11ff8a6731.jpg", + "image_caption": [ + "(b) GovReport" + ], + "image_footnote": [], + "bbox": [ + 381, + 80, + 611, + 206 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/576c77e2d4168b14669caaa7ef1f13475630024a6b889195642c7a54823bebc1.jpg", + "image_caption": [ + "(c) QMSUM", + "Figure 5: Fine-tuning performance of full attention under different input length." + ], + "image_footnote": [], + "bbox": [ + 630, + 80, + 860, + 206 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "presents fine-tuning performance of these models. We can tell that continual training with long inputs indeed facilitates the revised models to adapt to new structures and long-sequence inputs.", + "bbox": [ + 112, + 253, + 487, + 317 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d0a3b1f47d0535c441bdce266d7577db64fd43903ed73a50a36c651c66a44a83.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
l=8QMSUMGovReport
w.o. C.T.29.33/6.43/25.7153.28/23.61/51.74
C.T. (n=2048)29.87/7.16/26.1552.28/23.01/49.83
C.T. (n=8192)30.91/7.45/27.0254.37/24.42/51.87
", + "bbox": [ + 115, + 331, + 495, + 413 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3 Where to use full attention", + "text_level": 1, + "bbox": [ + 112, + 442, + 383, + 456 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To answer where to apply full attention, we compare fine-tuning performance of MASFormers that apply full attention at (i) bottom layers; (ii) middle layers; (iii) top layers; (iv) every $L / l$ layers. The results in Table 4 demonstrate that equipping bottom layers with full attention yields the best performance. This is because that long-range dependencies can be continually captured and reinforced by bottom layers before propagated to upper layers. As such, these long-range signals can be effectively incorporated into the upper layers with local attention, facilitating their encoding of local information. In contrast, when equipping local attention at bottom layers, long-range tokens are first aggregated with neighboring tokens by local attention, thereby weakening their long-range signals. Moreover, if alternating full and local attention every $L / l$ layers, the long-range signals cannot be continually reinforced nor efficiently captured.", + "bbox": [ + 112, + 462, + 489, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Discussion", + "text_level": 1, + "bbox": [ + 112, + 780, + 240, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-Neo (Black et al., 2021) introduces an attention pattern that alternates full and window attention. However, this models is not tailored for long sequences. It sets the local window size as 256 and has the maximum input length as 2048, unable to handle long sequences. Instead, this attention pattern is applied heuristically in an attempt to re", + "bbox": [ + 112, + 806, + 489, + 919 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/80d5984b750b359aeabe979b9b3ae97bd888cd3191dfedb56389309389bd07d1.jpg", + "table_caption": [ + "Table 3: We report R1/R2/RL for the above results." + ], + "table_footnote": [], + "table_body": "
PositionQMSUMGovReport
Every 328.26/6.94/25.0326.16/12.37/24.82
Top 820.89/4.52/18.37-/-/-
Middle 827.27/5.99/24.0620.80/9.01/19.52
Bottom 830.91/7.45/27.0254.37/24.42/51.87
Every 231.27/8.19/27.4135.34/16.04/33.68
Bottom 1232.53/8.70/28.7556.98/26.26/54.46
", + "bbox": [ + 510, + 256, + 878, + 394 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: Performance comparison of MASFormers that apply full attention at different layers (# layers $L = 24$", + "bbox": [ + 507, + 397, + 880, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "duce computational cost. However, as discussed in Section 4.3.3, this approach is neither effective nor efficient as MASFormer when handling long sequences. As shown in Table 4, applying full attention at every 2 or 3 layers underperforms applying it at bottom 12 or 8 layers. Therefore, alternating between full and block attention results in additional computational cost and performance degradation.", + "bbox": [ + 505, + 437, + 882, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast, MASFormer presents an effective solution for efficient long-sequence modeling. It provides guidance on adapting existing pre-trained transformers to long inputs. Meanwhile, it provides insights for designing large long-range models, especially for deeper models. By equipping only a subset of bottom layer with full attention, we can substantially mitigate computational cost. Additionally, the computation of MASFormer can be further optimized by leveraging system-level acceleration techniques (e.g., FlashAttention and xFormer) that support both block and full attention.", + "bbox": [ + 507, + 565, + 882, + 759 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 771, + 640, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose an efficient long-range transformer - MASFormer that utilizes full attention at a few of bottom layers and employs sparse attention at the remaining layers. Our empirical results on natural language modeling and generation tasks demonstrate that MASFormer can address the trade-off between computational cost and performance gains in a more efficient and effective way.", + "bbox": [ + 507, + 790, + 884, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2782", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Joshua Ainslie, Santiago Ontanon, Chris Alberti, Va-clav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. Etc: Encoding long and structured inputs in transformers. arXiv preprint arXiv:2004.08483.", + "Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150.", + "Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow.", + "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.", + "Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Re. 2021. Scatterbrain: Unifying sparse and low-rank attention approximation. arXiv preprint arXiv:2110.15343.", + "Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. A discourse-aware attention model for abstractive summarization of long documents. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 615-621, New Orleans, Louisiana. Association for Computational Linguistics.", + "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860.", + "Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages" + ], + "bbox": [ + 115, + 105, + 485, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929.", + "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.", + "Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. 2020. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487.", + "Albert Gu, Karan Goel, and Christopher Ré. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396.", + "Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, and Yinfei Yang. 2021. Longt5: Efficient text-to-text transformer for long sequences. arXiv preprint arXiv:2112.07916.", + "Ankit Gupta, Albert Gu, and Jonathan Berant. 2022. Diagonal state spaces are as effective as structured state spaces. In Advances in Neural Information Processing Systems.", + "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543.", + "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations.", + "Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. In International Conference on Learning Representations.", + "Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. 2022. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers.", + "Yuhong Li, Tianle Cai, Yi Zhang, Deming Chen, and Debadepta Dey. 2022. What makes convolutional models great on long sequence modeling? arXiv preprint arXiv:2210.09298.", + "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2783", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. 2023. Mega: Moving average equipped gated attention. In *The Eleventh International Conference on Learning Representations*.", + "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 8024-8035.", + "Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. 2021. Random feature attention. In International Conference on Learning Representations.", + "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.", + "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2021. Recipes for building an open-domain chatbot. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 300-325, Online. Association for Computational Linguistics.", + "Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, et al. 2022. Scrols: Standardized comparison over long language sequences. arXiv preprint arXiv:2201.03533.", + "Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. 2023. Simplified state space layers for sequence modeling. In *The Eleventh International Conference on Learning Representations*.", + "Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke, and Mohit Iyyer. 2021. Do long-range language models actually use long-range context? arXiv preprint arXiv:2109.09115.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768.", + "Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297.", + "Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adaptive budget allocation for parameter-efficient fine-tuning. In The Eleventh International Conference on Learning Representations.", + "Simiao Zuo, Xiaodong Liu, Jian Jiao, Denis Charles, Eren Manavoglu, Tuo Zhao, and Jianfeng Gao. 2022. Efficient long sequence modeling via state space augmented transformer. arXiv preprint arXiv:2212.08136." + ], + "bbox": [ + 510, + 85, + 882, + 363 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2784", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A Dataset Statistics", + "text_level": 1, + "bbox": [ + 114, + 84, + 302, + 99 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In the following table, we provide the detailed statistics of datasets in our experiments, including example splits and length statistics.", + "bbox": [ + 112, + 109, + 882, + 140 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/688c9a73142a43430e42bbc86a2f8ff6d2437c4538e40f8c8650a7f1a8794af0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetExample CountInput Length
TrainValidTestAverageMedian90th percentile
ArXiv203,0376,4366,44010,7208,51920,170
PubMed119,9246,6336,6584,7483,8838,883
QMSUM1,2572722819,49714,19727,761
GovReport17,4579729737,8868,84118,835
", + "bbox": [ + 200, + 148, + 796, + 269 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B Natural Language Generation", + "text_level": 1, + "bbox": [ + 112, + 303, + 413, + 319 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B.1 The Results of All ROUGE Scores", + "text_level": 1, + "bbox": [ + 114, + 328, + 430, + 343 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/0835269c3903855cb3d82f836026867f2603cfd3251e27d638513f881b722650.jpg", + "table_caption": [ + "Table 5: Statistics of datasets. Input length measured in tokens using a SentencePiece Model." + ], + "table_footnote": [], + "table_body": "
MethodsCQMSUMArXivGovReport
Full attention1610M31.50 / 8.00 / 27.8146.13 / 19.32 / 41.8960.53 / 28.83 / 57.88
Window (w=1024)402M23.31 / 4.32 / 20.6235.90 / 13.51 / 32.1949.82 / 17.03 / 47.42
Window (w=2048)805M26.73 / 5.05 / 23.4038.74 / 15.21 / 34.8756.14 / 22.79 / 53.50
Block (b=2048)402M26.24 / 5.03 / 23.1321.85 / 9.61 / 19.8626.37 / 12.31 / 25.18
Block (b=4096)805M26.96 / 5.15 / 23.8535.95 / 14.50 / 32.3749.83 / 23.64 / 47.50
MASFormer (l=4)436M29.86 / 6.59 / 25.8738.85 / 14.91 / 34.9846.67 / 18.82 / 44.39
MASFormer (l=6)553M30.83 / 7.15 / 27.1236.29 / 15.72 / 32.9649.26 / 21.20 / 46.89
MASFormer (l=8)671M30.91 / 8.00 / 27.8143.31 / 17.00 / 39.1254.37 / 24.42 / 51.87
MASFormer (l=12)906M32.53 / 8.70 / 28.7545.19 / 18.58 / 40.7256.98 / 26.26 / 54.46
", + "bbox": [ + 122, + 357, + 877, + 556 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B.2 Training Details", + "text_level": 1, + "bbox": [ + 114, + 614, + 290, + 630 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/970356a9f960a9b5a9d13af0c2587d16a39d5b7dd240f8a0c8e97957a5bb66f5.jpg", + "table_caption": [ + "Table 6: Finetuning performance of different attention methods." + ], + "table_footnote": [], + "table_body": "
MethodsQMSUMArXivGovReport
Full attention1 × 10-51 × 10-41 × 10-4
Window attention (w=1024)5 × 10-45 × 10-55 × 10-4
Window attention (w=2048)5 × 10-55 × 10-51 × 10-4
Block attention (w=2048)1 × 10-41 × 10-51 × 10-5
Block attention (w=4096)5 × 10-55 × 10-41 × 10-5
MASFormer (l=4)5 × 10-51 × 10-35 × 10-4
MASFormer (l=6)5 × 10-55 × 10-55 × 10-4
MASFormer (l=8)5 × 10-55 × 10-45 × 10-4
MASFormer (l=12)1 × 10-51 × 10-41 × 10-4
", + "bbox": [ + 242, + 645, + 754, + 825 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 7: The fine-tuning learning rate of each method on each dataset.", + "bbox": [ + 258, + 834, + 734, + 848 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We conduct continual training for all attention methods with training data form PILE and input length as 8192. After continual training, we obtain the continually trained models for each method and fine-tune them on QMSUM, ArXiv and GovReport to compare their summarization performance. During the", + "bbox": [ + 112, + 871, + 882, + 919 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2785", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "fine-tuning, we set the input length as 8192 for all datasets and all models. We apply the greedy decoding for generation and set the maximum output length as 256 for QMSUM, 1024 for GovReport, and 512 for ArXiv. Table 8 lists the details of these hyperparameters. Besides, we apply the linear learning rate schedule to fine-tune the models and the base learning rates are summarized in Table 7.", + "bbox": [ + 112, + 84, + 884, + 149 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/ef60230bc5c59c64126a05152ac0512baa6d553e79b62d8a95f8de9c214edb6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HyperparameterQMSUMArXivGovReport
Training steps3000120008000
Batch size323264
Input length819281928192
Maximum generation length2565121024
Weight decay0.0010.0010.001
", + "bbox": [ + 255, + 160, + 742, + 273 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 8: The other fine-tuning parameters for each dataset, which remain the same for every method.", + "bbox": [ + 157, + 284, + 838, + 299 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2786", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_model.json b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_model.json new file mode 100644 index 0000000000000000000000000000000000000000..07cfce3262baab09e87e09e65c77120781221c02 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_model.json @@ -0,0 +1,2270 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.173, + 0.09, + 0.833, + 0.131 + ], + "angle": 0, + "content": "Efficient Long-Range Transformers: You Need to Attend More, but Not Necessarily at Every Layer" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.149, + 0.836, + 0.166 + ], + "angle": 0, + "content": "Qingru Zhang†, Dhananjay Ram, Cole Hawkins, Sheng Zha, Tuo Zhao" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.167, + 0.735, + 0.183 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Georgia Institute of Technology \\(\\diamond\\) Amazon Web Service" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.185, + 0.674, + 0.2 + ], + "angle": 0, + "content": "{qingru.zhang,tourzhao}@gatech.edu" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.201, + 0.69, + 0.216 + ], + "angle": 0, + "content": "{radhna,colehawk,zhasheng}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.343, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.275, + 0.461, + 0.744 + ], + "angle": 0, + "content": "Pretrained transformer models have demonstrated remarkable performance across various natural language processing tasks. These models leverage the attention mechanism to capture long- and short-range dependencies in the sequence. However, the (full) attention mechanism incurs high computational cost – quadratic in the sequence length, which is not affordable in tasks with long sequences, e.g., inputs with 8k tokens. Although sparse attention can be used to improve computational efficiency, as suggested in existing work, it has limited modeling capacity and often fails to capture complicated dependencies in long sequences. To tackle this challenge, we propose MASFormer, an easy-to-implement transformer variant with Mixed Attention Spans. Specifically, MASFormer is equipped with full attention to capture long-range dependencies, but only at a small number of layers. For the remaining layers, MASformer only employs sparse attention to capture short-range dependencies. Our experiments on natural language modeling and generation tasks show that a decoder-only MASFormer model of 1.3B parameters can achieve competitive performance to vanilla transformers with full attention while significantly reducing computational cost (up to \\(75\\%\\)). Additionally, we investigate the effectiveness of continual training with long sequence data and how sequence length impacts downstream generation performance, which may be of independent interest." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.752, + 0.262, + 0.767 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.776, + 0.491, + 0.889 + ], + "angle": 0, + "content": "Pre-trained transformer models have manifested superior performance in various natural language processing tasks such as natural language modeling (NLM) (Dai et al., 2019; Radford et al., 2019), natural language generation (NLG) (Brown et al., 2020) and natural language understanding (NU) (Devlin et al., 2019; Liu et al., 2019; He et al., 2021b)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.254, + 0.885, + 0.302 + ], + "angle": 0, + "content": "These models leverage the attention mechanism (Vaswani et al., 2017) to compute the dependency score for each pair of tokens in an input sequence." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.305, + 0.886, + 0.61 + ], + "angle": 0, + "content": "Some practical tasks require these transformer models to handle long-sequence inputs like 8k tokens. For example, chatbot systems gather long-term contexts of user interactions to generate informative texts (Roller et al., 2021). Summarization for news, government reports, and academic papers request models to take inputs of long sequences to generate comprehensive summaries (Shaham et al., 2022), otherwise models often miss important information. Note that typical transformer models apply full attention to capture token dependencies pair-wise. It leads to a quadratic time and space complexity w.r.t. input length. However, such a complexity is prohibitive for long sequences. In particular, it incurs massive memory consumption during the back propagation. For example, a transformer model with 250M parameters consumes over 80G GPU memory when sequence length is 8k (Zuo et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.887, + 0.92 + ], + "angle": 0, + "content": "To address this scalability issue, various approaches have been proposed to reduce the complexity. One approach is sparse attention, which restricts each token to attend a subset of tokens based on predefined sparsity patterns (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020). For instance, block sparse attention (Kitaev et al., 2020; Ma et al., 2023) divides the input sequence into several blocks, and only intra-block attention is performed. Besides, sliding-window attention (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020) allows each token to attend to its neighboring tokens within a sliding window. These methods, though reducing the complexity of full attention, cannot sufficiently capture long-range dependencies. Other variants, such as kernel approximation (Peng et al., 2021) and low-rank approximation (Wang et al., 2020; Chen et al., 2021) methods, share the similar spirit and drawbacks. To com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.488, + 0.918 + ], + "angle": 0, + "content": "* Work was done during Qingru Zhang's internship at Amazon Web Service." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "2775" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.781, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2775-2786 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.198 + ], + "angle": 0, + "content": "pensate for the lack of long-range dependencies, LongT5 (Guo et al., 2021) introduces global tokens that are obtained by average pooling on every block of tokens (Ainslie et al., 2020). However, the block pooling operations can weaken the signal of crucial tokens and prevent the long-range dependencies from being detected." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.205, + 0.49, + 0.493 + ], + "angle": 0, + "content": "In addition to these methods, state space models (SSMs) prespecify global dependency patterns to capture the long-range dependencies only (Gu et al., 2020, 2021; Li et al., 2022; Zuo et al., 2022; Ma et al., 2023; Smith et al., 2023). These models can be regarded as linear recurrent neural networks with specifically designed fixed weights. As tailored for global dependencies, SSMs fail to effectively capture local dependencies. In order to combine both local and global dependencies, SPADE (Zuo et al., 2022) and MEGA (Ma et al., 2023) augment SSM layers into transformer layers equipped with local attention. However, state space methods require sophisticated implementation, and often encounter computational instability during the back propagation, especially when scaling up to large model size (Gupta et al., 2022). SPADE and MEGA hence inherit these drawbacks." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.502, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Note that the aforementioned methods apply same attention mechanism for every layer. We challenge this conventional wisdom and propose a transformer variant - MASFormer (Mixed Attention Span Transformer). MASFormer utilizes full attention only at a subset of layers whereas employs sparse attention at the remaining layers. Our design is motivated by the phenomenon - that most contexts in NLP data display a great deal of locality of reference (Zaheer et al., 2020; Beltagy et al., 2020). That is, most of information about a token can be derived from its neighboring tokens. In contrast, long-range dependencies among tokens are sparse and infrequent. Consider an academic paper as an example. Within a paragraph, there exist numerous short-term dependencies. Neighboring tokens are closely connected to convey meaningful semantics. Across paragraphs, there can be a small number of long-range dependencies. For example, tokens associated to the primary theme of the paper exhibit rare and weak dependencies across a long span. Since long-range dependencies occur much less frequently, a few layers of full attention are adequate to capture them. In stark contrast, short-term dependencies are more frequent, necessitating local attention in the majority of layers to fully extract" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.612, + 0.1 + ], + "angle": 0, + "content": "these signals." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.108, + 0.885, + 0.382 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of MASFormer, We conduct experiments on natural language modeling (ArXiv and PubMed Cohen et al. (2018)) and natural language generation (ArXiv, Cohen et al. (2018) and SCROLLS, Shaham et al. (2022)) tasks. Specifically, we compare the performance of MASFormer to other attention methods using a pretrained GPT-2 model (Radford et al., 2019) of 1.3 billion parameters. Our empirical results demonstrate that MASFormer consistently outperforms baseline methods across different attention cost (i.e. the total number of computed attention scores). In particular, MASFormer can achieve comparable performance to full attention while significantly reducing the computational cost. For example, with \\(27\\%\\) of its attention cost, MASFormer achieves a close R2 score as full attention on QMSUM dataset." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.386, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We also make additional discoveries with MASFormer, which are of independent interest. Firstly, we investigate the effectiveness of continual training for long sequence modeling. Many publicly available models are pre-trained with sequences shorter than 2048, and often fail to perform well on longer sequences (e.g. 8k/16k tokens). To bridge the gap, we explore the option of continual training to adapt these models to long sequences, thereby avoiding pre-training from the scratch. We discuss its effectiveness with MASFormer in Section 4.3. Secondly, we showcase that increasing sequence length can yield more performance gains on downstream tasks than NLM tasks evaluated by perplexity. We are aware of the recent findings by Sun et al. (2021) that increasing context length exhibits limited impact on NLM perplexity. Nevertheless, when applying MASFormer to downstream tasks like long-context summarization, we find that model performance benefits significantly from extending context length. Such a difference arises from the fact that predicting the next tokens in NLM primarily relies on locality of reference. Capturing infrequent long-range tokens can improve perplexity but not significantly. Therefore, we emphasize the necessity to evaluate model performance on downstream tasks that require long-range dependencies. Furthermore, our empirical evidence suggests that increasing the length can improve the performance only if models possess sufficient capability to handle additional long-range information. Local attention, as a counterexample, often fails to capture long-range signals and hence" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.414, + 0.1 + ], + "angle": 0, + "content": "benefits much less from long sequences." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.113, + 0.257, + 0.129 + ], + "angle": 0, + "content": "2 Background" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.139, + 0.39, + 0.155 + ], + "angle": 0, + "content": "2.1 Pretrained Language Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.16, + 0.488, + 0.32 + ], + "angle": 0, + "content": "Pre-trained transformer models (Devlin et al., 2019; Liu et al., 2019; Brown et al., 2020; Dosovitskiy et al., 2020; He et al., 2021b,a) have manifested superior performance in various NLP tasks. These models are often pre-trained on enormous amounts of unlabeled data in a unsupervised/self-supervised manner such that they can learn rich semantic knowledge. By further fine-tuning these pre-trained models, we can effectively transfer such knowledge to benefit downstream tasks (Zhang et al., 2023)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.322, + 0.489, + 0.482 + ], + "angle": 0, + "content": "Existing research on long-range transformers commonly requires pre-training the proposed models from scratch to accommodate new architectures and long inputs (Guo et al., 2021; Zuo et al., 2022). However, the significant training overheads raise a barrier for the widespread utilization of these methods across different language models. Motivated by this, we explore the possibility of leveraging existing pre-trained models and adapting them to long sequences though continual training." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.494, + 0.332, + 0.508 + ], + "angle": 0, + "content": "2.2 Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.514, + 0.487, + 0.562 + ], + "angle": 0, + "content": "Suppose the input to the layer is \\(X\\in \\mathbb{R}^{n\\times d}\\), where \\(n\\) is the input sequence length and \\(d\\) is embedding dimension, then self-attention mechanism outputs" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.57, + 0.488, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} (X) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.613, + 0.489, + 0.773 + ], + "angle": 0, + "content": "where \\( Q = XW_{q}, K = XW_{k}, V = VW_{v} \\) and \\( W_{q}, W_{k}, W_{v} \\) are learnable projection weights. Such full attention can simultaneously evaluate the alignment between any pair of tokens in the sequence. Specifically, denote the attention score \\( A = \\text{softmax}(QK^{\\top} / \\sqrt{d}) \\), then \\( A_{ij} \\) captures the alignment between tokens \\( i \\) and \\( j \\). A typical transformer model applies the full attention at every layer. Denote the number of layers as \\( L \\). Then its attention cost is \\( Ln^{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Sparse attention variants are introduced to mitigate the computational cost of full attention. Figures 1a and 1b illustrates the attention patterns of block sparse attention and sliding-window attention. For instance, block sparse attention divides tokens into blocks of size \\( b \\) and performs intrablock attention only, resulting in an attention cost of \\( bn \\). Sliding-window attention allows each token to attend its left/right neighboring tokens within a" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.882, + 0.133 + ], + "angle": 0, + "content": "local window of size \\( w \\). In most of cases, block sparse attention exhibits similar performance as sliding-window attention (Zuo et al., 2022)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.141, + 0.673, + 0.158 + ], + "angle": 0, + "content": "3 Our Approach" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.163, + 0.884, + 0.211 + ], + "angle": 0, + "content": "We present our method - MASFormer, a long-range transformer variant that mixes different attention spans across layers." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.217, + 0.842, + 0.234 + ], + "angle": 0, + "content": "3.1 MASFormer: Mixed Attention Span" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.238, + 0.884, + 0.495 + ], + "angle": 0, + "content": "MASFormer leverages full attention exclusively at a subset of transformer layers, whereas it employs block sparse attention at the remaining layers. The structure of MASFormer is illustrated in Figure 1c. We choose full attention to encode long-range information due to the following reasons: (i) full attention exhibits superior capability to capture long-range dependencies compared to sparse attention; (ii) full attention does not require sophisticated implementation and hence is computationally stable compared to SSMs (Zuo et al., 2022; Gupta et al., 2022); (iii) full attention is compatible with existing pre-trained transformer models, enabling us to conduct continual training which we elaborate in Section 3.2. To mitigate the computational cost, we restrict the number of layers using full attention." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.496, + 0.884, + 0.72 + ], + "angle": 0, + "content": "MASFormer is motivated by empirical investigations on performance comparison between models that apply the same attention span at every layer. Figure 2 presents the performance of block sparse attention and full attention on language modeling and summarization tasks. We find that, given long-sequence inputs, sparse attention is often insufficient to capture long-range dependencies beyond its attention span. As a result, it shows unsatisfactory performance. To remedy it, one can either increase attention span or switch to full attention to improve model capability of capturing sophisticated dependencies. Though improving model performance, it incurs high computational cost." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.721, + 0.884, + 0.897 + ], + "angle": 0, + "content": "Confronting such a trade-off between computational cost and model performance, we challenge the common practice - that applies the same attention span at every layer. MASFormer provides an alternative solution. Instead of increasing attention span evenly, MASFormer allocates a large portion of attention computations to a subset of \\( l \\) layers by equipping them with full attention. Specifically, equipping bottom layers with full attention can yield the best performance as suggested by our empirical analysis in Section 4.31. At the remain-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.828, + 0.919 + ], + "angle": 0, + "content": "1Please see Section 4.3 for detailed explanations" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2777" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.114, + 0.257, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.239, + 0.247, + 0.252 + ], + "angle": 0, + "content": "(a) Block Attention" + }, + { + "type": "image", + "bbox": [ + 0.266, + 0.114, + 0.408, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.239, + 0.404, + 0.252 + ], + "angle": 0, + "content": "(b) Window Attention" + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.086, + 0.882, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.239, + 0.702, + 0.252 + ], + "angle": 0, + "content": "(c) MASFormer" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.256, + 0.882, + 0.285 + ], + "angle": 0, + "content": "Figure 1: Illustration of attention patterns of (a) block sparse attention with block size \\( b = 3 \\); (b) sliding-window attention with window size \\( w = 1 \\) (on each side); (c) MASFormer that integrates full and sparse attention." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.301, + 0.299, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.398, + 0.272, + 0.41 + ], + "angle": 0, + "content": "(a) NLM for ArXiv" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.301, + 0.49, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.398, + 0.471, + 0.41 + ], + "angle": 0, + "content": "(b) NLM for PubMed" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.301, + 0.681, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.398, + 0.653, + 0.41 + ], + "angle": 0, + "content": "(c) Sum. for ArXiv" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.301, + 0.871, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.398, + 0.86, + 0.412 + ], + "angle": 0, + "content": "(d) Sum. for GovReport" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.415, + 0.882, + 0.459 + ], + "angle": 0, + "content": "Figure 2: (a,b): We evaluate the perplexity of a pre-trained GPT-2 model with block attention of differnet block size after continual training. (c,d): We fine-tune a GPT-2 model with block attention and compare the summarization performance on ArXiv and GovReport under different block size. Here the input length \\( n \\) is 8192." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.467, + 0.489, + 0.674 + ], + "angle": 0, + "content": "ing layers, MASFormer utilizes block attention of small size \\( m \\), resulting in a controlled attention cost of \\( (L - l)mn + ln^2 \\). As mentioned in Section 1, such a design is inspired by the phenomenon that most of contexts in NLP data exhibit a great deal of locality of reference. Long-range dependencies, in contrast, are less frequent. Therefore, it is not necessary to enhance attention span at every layer. Instead, a few layers of full attention are sufficient to capture infrequent long-range signals. The majority of layers can maintain small attention spans to adequately extract local dependencies and control the attention cost." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.677, + 0.489, + 0.806 + ], + "angle": 0, + "content": "Our empirical results demonstrate that, with the same attention cost, MASFormer significantly outperforms sparse attention. Remarkably, MASFormer can achieve comparable performance to full attention while substantially reducing computational cost. Therefore, by mixing different attention spans, MASFormer strikes a better balance between computational cost and model performance." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Moreover, MASFormer offers additional implementation advantages. As using the same attention function, MASFormer is easy to implement and compatible with existing pre-trained models. We can build MASFormer upon pre-trained transformers by changing their attention patterns, which does not involve modification on model architectures" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.467, + 0.884, + 0.545 + ], + "angle": 0, + "content": "and pre-trained weights. Meanwhile, acceleration packages, such as FlashAttention (Dao et al., 2022) and xFormers (Lefaudeauux et al., 2022), are applicable to further accelerate the computation of block attention and full attention in MASFormer." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.56, + 0.878, + 0.575 + ], + "angle": 0, + "content": "3.2 Continual Training with Long Sequences" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.582, + 0.884, + 0.919 + ], + "angle": 0, + "content": "As mentioned, MASFormer can be implemented upon majority of pre-trained transformers by modifying their attention patterns. However, most of publicly available models are pre-trained with sequences shorter than 2048, and often exhibit subpar performance on longer sequences such as \\(8\\mathrm{k} / 16\\mathrm{k}\\). To bridge this gap, we propose the continual training to adapt the revised model on long sequences and new attention pattern. As such, we can preserve existing pre-trained knowledge and circumvent the intensive overheads of pre-training from scratch. In particular, we first modify the attention pattern of the target model as proposed by MASFormer. If the pre-trained model uses absolute position embeddings, we duplicate them to accommodate long sequences. Subsequently, we provide the revised model with long sequences (e.g., 8k) from pretraining corpus like PILE (Gao et al., 2020). Then we conduct continual pre-training using casual language modeling (CLM) objective. We discuss the effectiveness of continual training in Section 4.3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2778" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.26, + 0.101 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.107, + 0.49, + 0.331 + ], + "angle": 0, + "content": "We evaluate the effectiveness and efficiency of MASFormer on natural language modeling (ArXiv and PubMed, Cohan et al. (2018)), natural language generation (ArXiv Cohan et al. (2018), QMSUM and GovReport Shaham et al. (2022)). We choose the GPT-3 XL model architecture (Brown et al., 2020) as our base model, which consists of 1.3 billion parameters and 24 layers and is pre-trained on PIL (Gao et al., 2020) for 300 billion tokens. GPT is a general purpose model that can be applied to many tasks instead of tailoring them for specific tasks. As such, it makes easy to control experiments and showcase the difference among various methods." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.334, + 0.49, + 0.51 + ], + "angle": 0, + "content": "Implementation Details. Our base model uses absolute positional embeddings with maximum length 1024. To accommodate longer inputs, we duplicate its positional embeddings to have the maximum length as 8192 such that the model can handle sequences containing up to 8192 tokens. Then, we implement different attention methods by modifying the attention pattern of the base model. We implement all the models with PyTorch (Paszke et al., 2019). All the experiments are conducted on NVIDIA A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.512, + 0.49, + 0.689 + ], + "angle": 0, + "content": "Continual Training Details. After changing the attention pattern, we conduct the continual training for MASFormer and baseline methods on PILE corpus (Gao et al., 2020) to adapt the revised models to new attention patterns and long-sequence inputs. We leverage the casual language modeling (CLM) objective to train the model for 50,000 steps with a warmup of 2000 steps. We set the input length as 8192 and use a batch size of 128 such that the models are optimized with 1M tokens per step. We use the constant learning 0.0001 for all methods." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.691, + 0.49, + 0.722 + ], + "angle": 0, + "content": "Baseline. We compare MASFormer with the following methods:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.725, + 0.488, + 0.821 + ], + "angle": 0, + "content": "- All full attention is to apply full attention at every layer. It has been adopted by most of existing transformer models as default. Although incurring the maximum attention cost, it achieves the best performance for most of our tasks. Hence, it acts as an upper bound for other methods." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "- All block sparse attention is to apply block attention at every layer, which is an effective method to reduce computational cost when modeling long sequences. Block attention sets the attention span of each layer identical such that it evenly distributes the budget of attention computation across layers." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.725, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.885, + 0.149 + ], + "angle": 0, + "content": "- All sliding-window attention is to apply sliding-window attention at every layer, which is another variant of sparse attention. It shares the similar spirits and often performs similarly as block attention." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.151, + 0.885, + 0.311 + ], + "angle": 0, + "content": "In the following experiments, we compare MAS-Former and the baseline methods across different attention cost \\( \\mathcal{C} \\). That is, for all block sparse attention, we set the block size as \\( b = \\mathcal{C} / (Ln) \\). For all sliding-window attention, we choose the window size as \\( w = \\mathcal{C} / (2Ln) \\). For MASFormer, we apply a small block size \\( m = 1024 \\) for its block attention and set \\( l \\) as \\( (\\mathcal{C} - Lmn) / (n^2 - mn) \\). Then we observe how their performance evolves when enhancing the attention cost \\( \\mathcal{C} \\) or input length \\( n \\)." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.312, + 0.884, + 0.343 + ], + "angle": 0, + "content": "Experiment Overview. We briefly summarize the experimental contents as follows:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.345, + 0.884, + 0.393 + ], + "angle": 0, + "content": "- Section 4.1 presents the perplexity evaluation of all the models on ArXiv and PubMed after continual training." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.395, + 0.885, + 0.49 + ], + "angle": 0, + "content": "- Section 4.2 compares the summarization performance of the models on ArXiv, QMSUM, and GovReport after fine-tuning. Besides, we also discuss the difference between perplexity and downstream evaluation in reflecting model capacity to capture long-range dependencies." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.492, + 0.886, + 0.636 + ], + "angle": 0, + "content": "- Section 4.3 provides three crucial analyses: (i) we evaluate the benefits of increasing input length and discuss the requirements to attain these gains; (ii) we analyze the effectiveness of continual training for long-sequence modeling; (iii) we conduct an ablation study to demonstrate that equipping bottom layers with full attention yields the most significant performance gains than other options. We further provide the explanations." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.345, + 0.886, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.65, + 0.78, + 0.666 + ], + "angle": 0, + "content": "4.1 Natural Language Modeling" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.673, + 0.819, + 0.687 + ], + "angle": 0, + "content": "4.1.1 Datasets and Evaluation Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.694, + 0.885, + 0.837 + ], + "angle": 0, + "content": "Datasets. We evaluate the perplexity of the updated GPT-2 for each attention method after continual training. The evaluation is conducted on test sets of ArXiv and PubMed (Cohan et al., 2018). Table 5 presents the statistics of these two datasets. Pubmed consists of scientific documents, with a document's content used as input and its corresponding abstract as the target summary. ArXiv is similar to PubMed, with documents from arXiv." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Evaluation Details. We conduct the perplexity evaluation under two settings. (i) We calculate the perplexity (ppl.) with all documents from test sets. Table 1 presents the overall perplexity of different models on two datasets. (ii) To showcase" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.216, + 0.068, + 0.488, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.217, + 0.393, + 0.228 + ], + "angle": 0, + "content": "(a) ArXiv ppl." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.068, + 0.782, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.216, + 0.694, + 0.228 + ], + "angle": 0, + "content": "(b) PubMed ppl." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.23, + 0.882, + 0.259 + ], + "angle": 0, + "content": "Figure 3: Perplexity evaluation on ArXiv and PubMed with examples of different length. Here \\( x \\)-axis is the maximum document length of each subset, i.e., \\( k \\times 1024 \\) (\\( k = 1, 2, 3, \\ldots \\))." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.267, + 0.49, + 0.395 + ], + "angle": 0, + "content": "the varying behaviors of models on documents of different length, we divide all documents into several subsets according to their length. Each subset consists of examples, whose length is within \\(((k - 1)\\times 1024,k\\times 1024](k = 1,2,3,\\ldots)\\). Then, we evaluate the perplexity on each subset. Figure 3 presents the perplexity of models on different subsets of examples." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.404, + 0.233, + 0.417 + ], + "angle": 0, + "content": "4.1.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.422, + 0.49, + 0.599 + ], + "angle": 0, + "content": "Table 1 compares the overall perplexity on test sets of ArXiv and PubMed. The results suggest that, with \\( l = 4 \\) layers of full attention, MASFormer achieves comparable performance to all full attention, while reducing \\( 72\\% \\) of its attention cost. With the similar attention cost \\( \\mathcal{C} \\), MASFormer outperforms all block attention that evenly distributes the budget of attention computation. For example, MASFormer with \\( l = 2 \\) achieves 8.75 ppl. on PubMed, which is 1.37 lower than that of block attention of \\( b = 2048 \\)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.607, + 0.481, + 0.75 + ], + "angle": 0, + "content": "
MethodsCArXivPubMed
Full attention1,610M8.637.63
Block (b=1024)201M13.1912.19
Block (b=2048)402M10.7510.13
MASFormer (l=2)318M10.258.75
MASFormer (l=4)436M9.318.25
MASFormer (l=8)671M9.638.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.753, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Table 1: Perplexity evaluation on ArXiv and PubMed." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Figure 3 illustrates the perplexity variation of each method given examples of different length. We can tell that MASFormer and full attention show better performance on longer documents, suggesting increasing context length can improve their prediction performance. Full attention, though incurring the highest attention cost, always achieves the best performance due to its outstanding capability to handle sophisticated dependencies. Notably," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.267, + 0.885, + 0.429 + ], + "angle": 0, + "content": "with \\(27\\%\\) of its attention cost, MASFormer exhibits a curve of ppl. v.s. length that closely resembles to that of full attention. This demonstrates the effectiveness and efficiency of MASFormer to capture long-range dependencies. In contrast, block sparse attention benefits much less from long contexts and underperforms both of them because of its incapability to encode long-range signals. For example, when \\(b = 1024\\), block attention achieves similar perplexity on PubMed examples of different length." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.438, + 0.794, + 0.454 + ], + "angle": 0, + "content": "4.2 Natural Language Generation" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.459, + 0.802, + 0.475 + ], + "angle": 0, + "content": "4.2.1 Datasets and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.479, + 0.885, + 0.639 + ], + "angle": 0, + "content": "Datasets. We evaluate the downstream performance of models on several abstractive summarization tasks to compare their capability of handling long sequences in practice. Specifically, we fine-tune models on ArXiv (Cohan et al., 2018), QM-SUM and GovReport (from SCROLLS benchmark, Shaham et al. (2022)). Their statistics are summarized in Table 5. We mainly use ROUGE-2 (R2) score (Lin, 2004) as the evaluation metric, which is more important and sensitive than R1 and RL." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.643, + 0.884, + 0.835 + ], + "angle": 0, + "content": "Training Details. After continual training, we fine-tune each model and report R2 scores on validation sets. Specifically, we fine-tune models for 3000 steps on QMSUM, 8000 steps on GovReport, and 12000 steps on ArXiv. We set the batch size as 64 for ArXiv and 32 for QMSUM and GovReport. We pick the learning rates from \\(\\{1 \\times 10^{-5}, 5 \\times 10^{-5}, 1 \\times 10^{-4}, 5 \\times 10^{-4}\\}\\), and choose the optimal ones to report the performance of each method. Moreover, the input length is fixed as 8192. We apply the greedy decoding for generation. Please see Appendix B for more details." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.846, + 0.628, + 0.859 + ], + "angle": 0, + "content": "4.2.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.864, + 0.882, + 0.896 + ], + "angle": 0, + "content": "In Table \\(2^{2}\\) and Figure 4, we present the fine-tuning results on QMSUM, ArXiv and GovReport across" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.881, + 0.919 + ], + "angle": 0, + "content": "2Please see Table 6 in Appendix B for all ROUGE scores." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2780" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.083, + 0.365, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.213, + 0.279, + 0.226 + ], + "angle": 0, + "content": "(a) ArXiv" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.081, + 0.613, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.456, + 0.213, + 0.537, + 0.226 + ], + "angle": 0, + "content": "(b) QMSUM" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.083, + 0.862, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.7, + 0.213, + 0.791, + 0.226 + ], + "angle": 0, + "content": "(c) GovReport" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.231, + 0.882, + 0.259 + ], + "angle": 0, + "content": "Figure 4: Given input length as 8192, we compare summarization performance between MASFormer and block/full attention when increasing the attention cost." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.273, + 0.49, + 0.472 + ], + "angle": 0, + "content": "
MethodsCQMSUMArXivGovReport
Full attention1610M8.0019.3228.83
Window(w=1024)402M4.3213.5117.03
Block(b=2048)402M5.039.6112.31
MASFormer(l=4)436M6.5914.9118.82
Window(w=2048)805M5.0515.2122.79
Block(b=4096)805M5.1514.5023.64
MASFormer(l=6)553M7.1515.7221.20
MASFormer(l=8)671M7.4617.0024.42
MASFormer(l=12)906M8.7018.5826.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.476, + 0.489, + 0.518 + ], + "angle": 0, + "content": "Table 2: Summarization performance of models with different attention methods. The best results are shown in bold." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.53, + 0.49, + 0.773 + ], + "angle": 0, + "content": "different attention cost. The results demonstrate that, with the similar attention cost, MASFormer significantly outperforms sparse attention variants. Furthermore, when enhancing attention cost, MASFormer achieves greater performance gains than sparse attention methods. This is evident from the steeper slope of its R2 curve versus attention cost, in contrast to the baseline method. For example, when increasing \\(\\mathcal{C}\\) form 553M to 671M, the R2 score of MASFormer on QMSUM exhibits a substantial improvement, reaching 8.70 from 7.46. Remarkably, this score surpasses even that of full attention. Therefore, MASFormer addresses the trade-off between computational cost and performance gains in a more efficient and effective way." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Notice that, in order to achieve comparable summarization performance to full attention, MAS-Former needs at least \\( l = 8 \\) layers of full attention, and providing more can lead to more gains. This observation is different from the findings in NLM (Figure 3) that increasing \\( l \\) beyond 4 provides limited improvement in perplexity. Their different capacity requirements arise from the fact that predicting next tokens in NLM primarily relies on lo" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.268, + 0.885, + 0.332 + ], + "angle": 0, + "content": "cal dependencies. Capturing infrequent long-range tokens does not significantly improve perplexity. Thus, this discrepancy emphasizes the necessity to evaluate long-range models on downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.345, + 0.623, + 0.361 + ], + "angle": 0, + "content": "4.3 Analysis" + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.368, + 0.878, + 0.384 + ], + "angle": 0, + "content": "4.3.1 Benefits of Increasing Sequence Length" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.389, + 0.885, + 0.596 + ], + "angle": 0, + "content": "In this section, we investigate the benefits of increasing input length for downstream performance. Specifically, we select the input length from \\(\\{2048, 4096, 6144, 8192\\}\\) and present the fine-tuning performance of full attention in Figure 5. The results consistently demonstrate that as the input length increases, the model's performance improves. That is, downstream performance benefits significantly from long-sequence inputs. In contrast, increasing example length beyond \\(6k\\) results in marginal improvements in perplexity (See Figure 3), highlighting again the importance of downstream evaluation." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.599, + 0.887, + 0.777 + ], + "angle": 0, + "content": "In addition, when comparing the behaviors of block attention in Figure 2c and 2d, we find that sparse attention often insufficiently capitalize on the benefits offered by longer inputs. For instance, given block size as 4096, its performance on ArXiv remains nearly unchanged when increasing input length from 4096 (R2 = 15.52 in Figure 5a) to 8192 (R2 = 14.49 in Figure 2c). This finding suggests that enhancing input length can only improve model performance if the model possesses the sufficient capability to handle long-range dependencies." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.787, + 0.848, + 0.802 + ], + "angle": 0, + "content": "4.3.2 Effectiveness of Continual Training" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We analyze the effectiveness of continual training by comparing fine-tuning performance of MAS-Former \\((l = 8)\\) under the following settings: (i) MASFormer without continual training (w.o. C.T.); (ii) MASFormer continually trained with short inputs (C.T. \\((n = 2048)\\)); (iii) MASFormer continually trained with long inputs (C.T. \\((n = 8192)\\)). Table 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.941 + ], + "angle": 0, + "content": "2781" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.082, + 0.365, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.211, + 0.278, + 0.222 + ], + "angle": 0, + "content": "(a) ArXiv" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.081, + 0.613, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.452, + 0.211, + 0.542, + 0.223 + ], + "angle": 0, + "content": "(b) GovReport" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.082, + 0.861, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.705, + 0.211, + 0.785, + 0.222 + ], + "angle": 0, + "content": "(c) QMSUM" + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.224, + 0.768, + 0.24 + ], + "angle": 0, + "content": "Figure 5: Fine-tuning performance of full attention under different input length." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.254, + 0.489, + 0.318 + ], + "angle": 0, + "content": "presents fine-tuning performance of these models. We can tell that continual training with long inputs indeed facilitates the revised models to adapt to new structures and long-sequence inputs." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.332, + 0.496, + 0.414 + ], + "angle": 0, + "content": "
l=8QMSUMGovReport
w.o. C.T.29.33/6.43/25.7153.28/23.61/51.74
C.T. (n=2048)29.87/7.16/26.1552.28/23.01/49.83
C.T. (n=8192)30.91/7.45/27.0254.37/24.42/51.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.417, + 0.475, + 0.432 + ], + "angle": 0, + "content": "Table 3: We report R1/R2/RL for the above results." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.443, + 0.384, + 0.457 + ], + "angle": 0, + "content": "4.3.3 Where to use full attention" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.463, + 0.49, + 0.768 + ], + "angle": 0, + "content": "To answer where to apply full attention, we compare fine-tuning performance of MASFormers that apply full attention at (i) bottom layers; (ii) middle layers; (iii) top layers; (iv) every \\( L / l \\) layers. The results in Table 4 demonstrate that equipping bottom layers with full attention yields the best performance. This is because that long-range dependencies can be continually captured and reinforced by bottom layers before propagated to upper layers. As such, these long-range signals can be effectively incorporated into the upper layers with local attention, facilitating their encoding of local information. In contrast, when equipping local attention at bottom layers, long-range tokens are first aggregated with neighboring tokens by local attention, thereby weakening their long-range signals. Moreover, if alternating full and local attention every \\( L / l \\) layers, the long-range signals cannot be continually reinforced nor efficiently captured." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.781, + 0.241, + 0.795 + ], + "angle": 0, + "content": "5 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.92 + ], + "angle": 0, + "content": "GPT-Neo (Black et al., 2021) introduces an attention pattern that alternates full and window attention. However, this models is not tailored for long sequences. It sets the local window size as 256 and has the maximum input length as 2048, unable to handle long sequences. Instead, this attention pattern is applied heuristically in an attempt to re" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.258, + 0.879, + 0.395 + ], + "angle": 0, + "content": "
PositionQMSUMGovReport
Every 328.26/6.94/25.0326.16/12.37/24.82
Top 820.89/4.52/18.37-/-/-
Middle 827.27/5.99/24.0620.80/9.01/19.52
Bottom 830.91/7.45/27.0254.37/24.42/51.87
Every 231.27/8.19/27.4135.34/16.04/33.68
Bottom 1232.53/8.70/28.7556.98/26.26/54.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.398, + 0.882, + 0.427 + ], + "angle": 0, + "content": "Table 4: Performance comparison of MASFormers that apply full attention at different layers (# layers \\(L = 24\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.438, + 0.884, + 0.567 + ], + "angle": 0, + "content": "duce computational cost. However, as discussed in Section 4.3.3, this approach is neither effective nor efficient as MASFormer when handling long sequences. As shown in Table 4, applying full attention at every 2 or 3 layers underperforms applying it at bottom 12 or 8 layers. Therefore, alternating between full and block attention results in additional computational cost and performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.567, + 0.884, + 0.76 + ], + "angle": 0, + "content": "In contrast, MASFormer presents an effective solution for efficient long-sequence modeling. It provides guidance on adapting existing pre-trained transformers to long inputs. Meanwhile, it provides insights for designing large long-range models, especially for deeper models. By equipping only a subset of bottom layer with full attention, we can substantially mitigate computational cost. Additionally, the computation of MASFormer can be further optimized by leveraging system-level acceleration techniques (e.g., FlashAttention and xFormer) that support both block and full attention." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.772, + 0.642, + 0.787 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We propose an efficient long-range transformer - MASFormer that utilizes full attention at a few of bottom layers and employs sparse attention at the remaining layers. Our empirical results on natural language modeling and generation tasks demonstrate that MASFormer can address the trade-off between computational cost and performance gains in a more efficient and effective way." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2782" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.107, + 0.487, + 0.173 + ], + "angle": 0, + "content": "Joshua Ainslie, Santiago Ontanon, Chris Alberti, Va-clav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. Etc: Encoding long and structured inputs in transformers. arXiv preprint arXiv:2004.08483." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.182, + 0.487, + 0.222 + ], + "angle": 0, + "content": "Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.231, + 0.487, + 0.283 + ], + "angle": 0, + "content": "Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.293, + 0.487, + 0.489 + ], + "angle": 0, + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.499, + 0.487, + 0.552 + ], + "angle": 0, + "content": "Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Re. 2021. Scatterbrain: Unifying sparse and low-rank attention approximation. arXiv preprint arXiv:2110.15343." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.562, + 0.487, + 0.693 + ], + "angle": 0, + "content": "Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. A discourse-aware attention model for abstractive summarization of long documents. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 615-621, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.702, + 0.487, + 0.767 + ], + "angle": 0, + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.487, + 0.817 + ], + "angle": 0, + "content": "Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.107, + 0.487, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.113 + ], + "angle": 0, + "content": "4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.122, + 0.882, + 0.213 + ], + "angle": 0, + "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.222, + 0.882, + 0.289 + ], + "angle": 0, + "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.297, + 0.882, + 0.362 + ], + "angle": 0, + "content": "Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. 2020. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.372, + 0.882, + 0.412 + ], + "angle": 0, + "content": "Albert Gu, Karan Goel, and Christopher Ré. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.421, + 0.882, + 0.474 + ], + "angle": 0, + "content": "Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, and Yinfei Yang. 2021. Longt5: Efficient text-to-text transformer for long sequences. arXiv preprint arXiv:2112.07916." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.482, + 0.882, + 0.535 + ], + "angle": 0, + "content": "Ankit Gupta, Albert Gu, and Jonathan Berant. 2022. Diagonal state spaces are as effective as structured state spaces. In Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.544, + 0.882, + 0.597 + ], + "angle": 0, + "content": "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.606, + 0.882, + 0.659 + ], + "angle": 0, + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.667, + 0.882, + 0.708 + ], + "angle": 0, + "content": "Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.716, + 0.882, + 0.808 + ], + "angle": 0, + "content": "Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. 2022. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.817, + 0.882, + 0.87 + ], + "angle": 0, + "content": "Yuhong Li, Tianle Cai, Yi Zhang, Deming Chen, and Debadepta Dey. 2022. What makes convolutional models great on long sequence modeling? arXiv preprint arXiv:2210.09298." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.878, + 0.882, + 0.919 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2783" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.153 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.161, + 0.488, + 0.227 + ], + "angle": 0, + "content": "Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. 2023. Mega: Moving average equipped gated attention. In *The Eleventh International Conference on Learning Representations*." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.236, + 0.488, + 0.405 + ], + "angle": 0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 8024-8035." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.416, + 0.488, + 0.469 + ], + "angle": 0, + "content": "Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. 2021. Random feature attention. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.478, + 0.488, + 0.531 + ], + "angle": 0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.54, + 0.488, + 0.644 + ], + "angle": 0, + "content": "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2021. Recipes for building an open-domain chatbot. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 300-325, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.654, + 0.488, + 0.72 + ], + "angle": 0, + "content": "Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, et al. 2022. Scrols: Standardized comparison over long language sequences. arXiv preprint arXiv:2201.03533." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.729, + 0.488, + 0.782 + ], + "angle": 0, + "content": "Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. 2023. Simplified state space layers for sequence modeling. In *The Eleventh International Conference on Learning Representations*." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.791, + 0.488, + 0.844 + ], + "angle": 0, + "content": "Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke, and Mohit Iyyer. 2021. Do long-range language models actually use long-range context? arXiv preprint arXiv:2109.09115." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.884, + 0.126 + ], + "angle": 0, + "content": "Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.136, + 0.883, + 0.215 + ], + "angle": 0, + "content": "Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.225, + 0.883, + 0.291 + ], + "angle": 0, + "content": "Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adaptive budget allocation for parameter-efficient fine-tuning. In The Eleventh International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.3, + 0.883, + 0.365 + ], + "angle": 0, + "content": "Simiao Zuo, Xiaodong Liu, Jian Jiao, Denis Charles, Eren Manavoglu, Tuo Zhao, and Jianfeng Gao. 2022. Efficient long sequence modeling via state space augmented transformer. arXiv preprint arXiv:2212.08136." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.884, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2784" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.303, + 0.1 + ], + "angle": 0, + "content": "A Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.11, + 0.884, + 0.141 + ], + "angle": 0, + "content": "In the following table, we provide the detailed statistics of datasets in our experiments, including example splits and length statistics." + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.149, + 0.797, + 0.27 + ], + "angle": 0, + "content": "
DatasetExample CountInput Length
TrainValidTestAverageMedian90th percentile
ArXiv203,0376,4366,44010,7208,51920,170
PubMed119,9246,6336,6584,7483,8838,883
QMSUM1,2572722819,49714,19727,761
GovReport17,4579729737,8868,84118,835
" + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.277, + 0.813, + 0.292 + ], + "angle": 0, + "content": "Table 5: Statistics of datasets. Input length measured in tokens using a SentencePiece Model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.304, + 0.414, + 0.32 + ], + "angle": 0, + "content": "B Natural Language Generation" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.329, + 0.431, + 0.344 + ], + "angle": 0, + "content": "B.1 The Results of All ROUGE Scores" + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.358, + 0.878, + 0.557 + ], + "angle": 0, + "content": "
MethodsCQMSUMArXivGovReport
Full attention1610M31.50 / 8.00 / 27.8146.13 / 19.32 / 41.8960.53 / 28.83 / 57.88
Window (w=1024)402M23.31 / 4.32 / 20.6235.90 / 13.51 / 32.1949.82 / 17.03 / 47.42
Window (w=2048)805M26.73 / 5.05 / 23.4038.74 / 15.21 / 34.8756.14 / 22.79 / 53.50
Block (b=2048)402M26.24 / 5.03 / 23.1321.85 / 9.61 / 19.8626.37 / 12.31 / 25.18
Block (b=4096)805M26.96 / 5.15 / 23.8535.95 / 14.50 / 32.3749.83 / 23.64 / 47.50
MASFormer (l=4)436M29.86 / 6.59 / 25.8738.85 / 14.91 / 34.9846.67 / 18.82 / 44.39
MASFormer (l=6)553M30.83 / 7.15 / 27.1236.29 / 15.72 / 32.9649.26 / 21.20 / 46.89
MASFormer (l=8)671M30.91 / 8.00 / 27.8143.31 / 17.00 / 39.1254.37 / 24.42 / 51.87
MASFormer (l=12)906M32.53 / 8.70 / 28.7545.19 / 18.58 / 40.7256.98 / 26.26 / 54.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.566, + 0.715, + 0.581 + ], + "angle": 0, + "content": "Table 6: Finetuning performance of different attention methods." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.615, + 0.291, + 0.631 + ], + "angle": 0, + "content": "B.2 Training Details" + }, + { + "type": "table", + "bbox": [ + 0.243, + 0.646, + 0.756, + 0.826 + ], + "angle": 0, + "content": "
MethodsQMSUMArXivGovReport
Full attention1 × 10-51 × 10-41 × 10-4
Window attention (w=1024)5 × 10-45 × 10-55 × 10-4
Window attention (w=2048)5 × 10-55 × 10-51 × 10-4
Block attention (w=2048)1 × 10-41 × 10-51 × 10-5
Block attention (w=4096)5 × 10-55 × 10-41 × 10-5
MASFormer (l=4)5 × 10-51 × 10-35 × 10-4
MASFormer (l=6)5 × 10-55 × 10-55 × 10-4
MASFormer (l=8)5 × 10-55 × 10-45 × 10-4
MASFormer (l=12)1 × 10-51 × 10-41 × 10-4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.26, + 0.835, + 0.735, + 0.85 + ], + "angle": 0, + "content": "Table 7: The fine-tuning learning rate of each method on each dataset." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.872, + 0.884, + 0.92 + ], + "angle": 0, + "content": "We conduct continual training for all attention methods with training data form PILE and input length as 8192. After continual training, we obtain the continually trained models for each method and fine-tune them on QMSUM, ArXiv and GovReport to compare their summarization performance. During the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "2785" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.885, + 0.15 + ], + "angle": 0, + "content": "fine-tuning, we set the input length as 8192 for all datasets and all models. We apply the greedy decoding for generation and set the maximum output length as 256 for QMSUM, 1024 for GovReport, and 512 for ArXiv. Table 8 lists the details of these hyperparameters. Besides, we apply the linear learning rate schedule to fine-tune the models and the base learning rates are summarized in Table 7." + }, + { + "type": "table", + "bbox": [ + 0.256, + 0.161, + 0.744, + 0.274 + ], + "angle": 0, + "content": "
HyperparameterQMSUMArXivGovReport
Training steps3000120008000
Batch size323264
Input length819281928192
Maximum generation length2565121024
Weight decay0.0010.0010.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.285, + 0.839, + 0.3 + ], + "angle": 0, + "content": "Table 8: The other fine-tuning parameters for each dataset, which remain the same for every method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "2786" + } + ] +] \ No newline at end of file diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_origin.pdf b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..badac6fae422a65f1c52614ad9dacc4daf6647e5 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/b2cbac77-9108-42c4-9fd3-211395757830_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:249de0c8323cfa8f380acf757c56386a49420dda002d70627fcce27711d99493 +size 794045 diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/full.md b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f2720e858c70d817f6e8ebbf5b73f4bb6acf926 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/full.md @@ -0,0 +1,310 @@ +# Efficient Long-Range Transformers: You Need to Attend More, but Not Necessarily at Every Layer + +Qingru Zhang†, Dhananjay Ram, Cole Hawkins, Sheng Zha, Tuo Zhao + +$^{\dagger}$ Georgia Institute of Technology $\diamond$ Amazon Web Service + +{qingru.zhang,tourzhao}@gatech.edu + +{radhna,colehawk,zhasheng}@amazon.com + +# Abstract + +Pretrained transformer models have demonstrated remarkable performance across various natural language processing tasks. These models leverage the attention mechanism to capture long- and short-range dependencies in the sequence. However, the (full) attention mechanism incurs high computational cost – quadratic in the sequence length, which is not affordable in tasks with long sequences, e.g., inputs with 8k tokens. Although sparse attention can be used to improve computational efficiency, as suggested in existing work, it has limited modeling capacity and often fails to capture complicated dependencies in long sequences. To tackle this challenge, we propose MASFormer, an easy-to-implement transformer variant with Mixed Attention Spans. Specifically, MASFormer is equipped with full attention to capture long-range dependencies, but only at a small number of layers. For the remaining layers, MASformer only employs sparse attention to capture short-range dependencies. Our experiments on natural language modeling and generation tasks show that a decoder-only MASFormer model of 1.3B parameters can achieve competitive performance to vanilla transformers with full attention while significantly reducing computational cost (up to $75\%$ ). Additionally, we investigate the effectiveness of continual training with long sequence data and how sequence length impacts downstream generation performance, which may be of independent interest. + +# 1 Introduction + +Pre-trained transformer models have manifested superior performance in various natural language processing tasks such as natural language modeling (NLM) (Dai et al., 2019; Radford et al., 2019), natural language generation (NLG) (Brown et al., 2020) and natural language understanding (NU) (Devlin et al., 2019; Liu et al., 2019; He et al., 2021b). + +These models leverage the attention mechanism (Vaswani et al., 2017) to compute the dependency score for each pair of tokens in an input sequence. + +Some practical tasks require these transformer models to handle long-sequence inputs like 8k tokens. For example, chatbot systems gather long-term contexts of user interactions to generate informative texts (Roller et al., 2021). Summarization for news, government reports, and academic papers request models to take inputs of long sequences to generate comprehensive summaries (Shaham et al., 2022), otherwise models often miss important information. Note that typical transformer models apply full attention to capture token dependencies pair-wise. It leads to a quadratic time and space complexity w.r.t. input length. However, such a complexity is prohibitive for long sequences. In particular, it incurs massive memory consumption during the back propagation. For example, a transformer model with 250M parameters consumes over 80G GPU memory when sequence length is 8k (Zuo et al., 2022). + +To address this scalability issue, various approaches have been proposed to reduce the complexity. One approach is sparse attention, which restricts each token to attend a subset of tokens based on predefined sparsity patterns (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020). For instance, block sparse attention (Kitaev et al., 2020; Ma et al., 2023) divides the input sequence into several blocks, and only intra-block attention is performed. Besides, sliding-window attention (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020) allows each token to attend to its neighboring tokens within a sliding window. These methods, though reducing the complexity of full attention, cannot sufficiently capture long-range dependencies. Other variants, such as kernel approximation (Peng et al., 2021) and low-rank approximation (Wang et al., 2020; Chen et al., 2021) methods, share the similar spirit and drawbacks. To com + +pensate for the lack of long-range dependencies, LongT5 (Guo et al., 2021) introduces global tokens that are obtained by average pooling on every block of tokens (Ainslie et al., 2020). However, the block pooling operations can weaken the signal of crucial tokens and prevent the long-range dependencies from being detected. + +In addition to these methods, state space models (SSMs) prespecify global dependency patterns to capture the long-range dependencies only (Gu et al., 2020, 2021; Li et al., 2022; Zuo et al., 2022; Ma et al., 2023; Smith et al., 2023). These models can be regarded as linear recurrent neural networks with specifically designed fixed weights. As tailored for global dependencies, SSMs fail to effectively capture local dependencies. In order to combine both local and global dependencies, SPADE (Zuo et al., 2022) and MEGA (Ma et al., 2023) augment SSM layers into transformer layers equipped with local attention. However, state space methods require sophisticated implementation, and often encounter computational instability during the back propagation, especially when scaling up to large model size (Gupta et al., 2022). SPADE and MEGA hence inherit these drawbacks. + +Note that the aforementioned methods apply same attention mechanism for every layer. We challenge this conventional wisdom and propose a transformer variant - MASFormer (Mixed Attention Span Transformer). MASFormer utilizes full attention only at a subset of layers whereas employs sparse attention at the remaining layers. Our design is motivated by the phenomenon - that most contexts in NLP data display a great deal of locality of reference (Zaheer et al., 2020; Beltagy et al., 2020). That is, most of information about a token can be derived from its neighboring tokens. In contrast, long-range dependencies among tokens are sparse and infrequent. Consider an academic paper as an example. Within a paragraph, there exist numerous short-term dependencies. Neighboring tokens are closely connected to convey meaningful semantics. Across paragraphs, there can be a small number of long-range dependencies. For example, tokens associated to the primary theme of the paper exhibit rare and weak dependencies across a long span. Since long-range dependencies occur much less frequently, a few layers of full attention are adequate to capture them. In stark contrast, short-term dependencies are more frequent, necessitating local attention in the majority of layers to fully extract + +these signals. + +To demonstrate the effectiveness of MASFormer, We conduct experiments on natural language modeling (ArXiv and PubMed Cohen et al. (2018)) and natural language generation (ArXiv, Cohen et al. (2018) and SCROLLS, Shaham et al. (2022)) tasks. Specifically, we compare the performance of MASFormer to other attention methods using a pretrained GPT-2 model (Radford et al., 2019) of 1.3 billion parameters. Our empirical results demonstrate that MASFormer consistently outperforms baseline methods across different attention cost (i.e. the total number of computed attention scores). In particular, MASFormer can achieve comparable performance to full attention while significantly reducing the computational cost. For example, with $27\%$ of its attention cost, MASFormer achieves a close R2 score as full attention on QMSUM dataset. + +We also make additional discoveries with MASFormer, which are of independent interest. Firstly, we investigate the effectiveness of continual training for long sequence modeling. Many publicly available models are pre-trained with sequences shorter than 2048, and often fail to perform well on longer sequences (e.g. 8k/16k tokens). To bridge the gap, we explore the option of continual training to adapt these models to long sequences, thereby avoiding pre-training from the scratch. We discuss its effectiveness with MASFormer in Section 4.3. Secondly, we showcase that increasing sequence length can yield more performance gains on downstream tasks than NLM tasks evaluated by perplexity. We are aware of the recent findings by Sun et al. (2021) that increasing context length exhibits limited impact on NLM perplexity. Nevertheless, when applying MASFormer to downstream tasks like long-context summarization, we find that model performance benefits significantly from extending context length. Such a difference arises from the fact that predicting the next tokens in NLM primarily relies on locality of reference. Capturing infrequent long-range tokens can improve perplexity but not significantly. Therefore, we emphasize the necessity to evaluate model performance on downstream tasks that require long-range dependencies. Furthermore, our empirical evidence suggests that increasing the length can improve the performance only if models possess sufficient capability to handle additional long-range information. Local attention, as a counterexample, often fails to capture long-range signals and hence + +benefits much less from long sequences. + +# 2 Background + +# 2.1 Pretrained Language Models + +Pre-trained transformer models (Devlin et al., 2019; Liu et al., 2019; Brown et al., 2020; Dosovitskiy et al., 2020; He et al., 2021b,a) have manifested superior performance in various NLP tasks. These models are often pre-trained on enormous amounts of unlabeled data in a unsupervised/self-supervised manner such that they can learn rich semantic knowledge. By further fine-tuning these pre-trained models, we can effectively transfer such knowledge to benefit downstream tasks (Zhang et al., 2023). + +Existing research on long-range transformers commonly requires pre-training the proposed models from scratch to accommodate new architectures and long inputs (Guo et al., 2021; Zuo et al., 2022). However, the significant training overheads raise a barrier for the widespread utilization of these methods across different language models. Motivated by this, we explore the possibility of leveraging existing pre-trained models and adapting them to long sequences though continual training. + +# 2.2 Attention Mechanism + +Suppose the input to the layer is $X\in \mathbb{R}^{n\times d}$ , where $n$ is the input sequence length and $d$ is embedding dimension, then self-attention mechanism outputs + +$$ +\operatorname {A t t n} (X) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d}}\right) V \tag {1} +$$ + +where $Q = XW_{q}, K = XW_{k}, V = VW_{v}$ and $W_{q}, W_{k}, W_{v}$ are learnable projection weights. Such full attention can simultaneously evaluate the alignment between any pair of tokens in the sequence. Specifically, denote the attention score $A = \text{softmax}(QK^{\top} / \sqrt{d})$ , then $A_{ij}$ captures the alignment between tokens $i$ and $j$ . A typical transformer model applies the full attention at every layer. Denote the number of layers as $L$ . Then its attention cost is $Ln^{2}$ . + +Sparse attention variants are introduced to mitigate the computational cost of full attention. Figures 1a and 1b illustrates the attention patterns of block sparse attention and sliding-window attention. For instance, block sparse attention divides tokens into blocks of size $b$ and performs intrablock attention only, resulting in an attention cost of $bn$ . Sliding-window attention allows each token to attend its left/right neighboring tokens within a + +local window of size $w$ . In most of cases, block sparse attention exhibits similar performance as sliding-window attention (Zuo et al., 2022). + +# 3 Our Approach + +We present our method - MASFormer, a long-range transformer variant that mixes different attention spans across layers. + +# 3.1 MASFormer: Mixed Attention Span + +MASFormer leverages full attention exclusively at a subset of transformer layers, whereas it employs block sparse attention at the remaining layers. The structure of MASFormer is illustrated in Figure 1c. We choose full attention to encode long-range information due to the following reasons: (i) full attention exhibits superior capability to capture long-range dependencies compared to sparse attention; (ii) full attention does not require sophisticated implementation and hence is computationally stable compared to SSMs (Zuo et al., 2022; Gupta et al., 2022); (iii) full attention is compatible with existing pre-trained transformer models, enabling us to conduct continual training which we elaborate in Section 3.2. To mitigate the computational cost, we restrict the number of layers using full attention. + +MASFormer is motivated by empirical investigations on performance comparison between models that apply the same attention span at every layer. Figure 2 presents the performance of block sparse attention and full attention on language modeling and summarization tasks. We find that, given long-sequence inputs, sparse attention is often insufficient to capture long-range dependencies beyond its attention span. As a result, it shows unsatisfactory performance. To remedy it, one can either increase attention span or switch to full attention to improve model capability of capturing sophisticated dependencies. Though improving model performance, it incurs high computational cost. + +Confronting such a trade-off between computational cost and model performance, we challenge the common practice - that applies the same attention span at every layer. MASFormer provides an alternative solution. Instead of increasing attention span evenly, MASFormer allocates a large portion of attention computations to a subset of $l$ layers by equipping them with full attention. Specifically, equipping bottom layers with full attention can yield the best performance as suggested by our empirical analysis in Section 4.31. At the remain- + +![](images/b14e862e159937c0a8d0aba62f3dcf18dc7295a58d45f99d7a2c8abb6f7d2625.jpg) +(a) Block Attention + +![](images/e385a1ca1c74eb15b1f237f6586b185c1f36c971d83bf972e19739324ca47bb6.jpg) +(b) Window Attention + +![](images/e07b6b4719286115e818c89a0512ae2593a10b031ab89993a43a5e73348c59ee.jpg) +(c) MASFormer + +![](images/cdc86d8d963cb9cd155602b3ad9c20143a6c51bbf63bd87fc5c7884176cbac8c.jpg) +Figure 1: Illustration of attention patterns of (a) block sparse attention with block size $b = 3$ ; (b) sliding-window attention with window size $w = 1$ (on each side); (c) MASFormer that integrates full and sparse attention. +(a) NLM for ArXiv +Figure 2: (a,b): We evaluate the perplexity of a pre-trained GPT-2 model with block attention of differnet block size after continual training. (c,d): We fine-tune a GPT-2 model with block attention and compare the summarization performance on ArXiv and GovReport under different block size. Here the input length $n$ is 8192. + +![](images/d9ac5c06d068f9a7c4a4b26666d22d047f491027c96e6ad0643d82ebc258f6e7.jpg) +(b) NLM for PubMed + +![](images/3d353730f2b9c2e78c69ce82fb754ba1bc1c9b583bdf4740544689399a05475b.jpg) +(c) Sum. for ArXiv + +![](images/76170219267e1227827b845500c7f6484232a8edfc99eafc43cbac1dbf902c3d.jpg) +(d) Sum. for GovReport + +ing layers, MASFormer utilizes block attention of small size $m$ , resulting in a controlled attention cost of $(L - l)mn + ln^2$ . As mentioned in Section 1, such a design is inspired by the phenomenon that most of contexts in NLP data exhibit a great deal of locality of reference. Long-range dependencies, in contrast, are less frequent. Therefore, it is not necessary to enhance attention span at every layer. Instead, a few layers of full attention are sufficient to capture infrequent long-range signals. The majority of layers can maintain small attention spans to adequately extract local dependencies and control the attention cost. + +Our empirical results demonstrate that, with the same attention cost, MASFormer significantly outperforms sparse attention. Remarkably, MASFormer can achieve comparable performance to full attention while substantially reducing computational cost. Therefore, by mixing different attention spans, MASFormer strikes a better balance between computational cost and model performance. + +Moreover, MASFormer offers additional implementation advantages. As using the same attention function, MASFormer is easy to implement and compatible with existing pre-trained models. We can build MASFormer upon pre-trained transformers by changing their attention patterns, which does not involve modification on model architectures + +and pre-trained weights. Meanwhile, acceleration packages, such as FlashAttention (Dao et al., 2022) and xFormers (Lefaudeauux et al., 2022), are applicable to further accelerate the computation of block attention and full attention in MASFormer. + +# 3.2 Continual Training with Long Sequences + +As mentioned, MASFormer can be implemented upon majority of pre-trained transformers by modifying their attention patterns. However, most of publicly available models are pre-trained with sequences shorter than 2048, and often exhibit subpar performance on longer sequences such as $8\mathrm{k} / 16\mathrm{k}$ . To bridge this gap, we propose the continual training to adapt the revised model on long sequences and new attention pattern. As such, we can preserve existing pre-trained knowledge and circumvent the intensive overheads of pre-training from scratch. In particular, we first modify the attention pattern of the target model as proposed by MASFormer. If the pre-trained model uses absolute position embeddings, we duplicate them to accommodate long sequences. Subsequently, we provide the revised model with long sequences (e.g., 8k) from pretraining corpus like PILE (Gao et al., 2020). Then we conduct continual pre-training using casual language modeling (CLM) objective. We discuss the effectiveness of continual training in Section 4.3. + +# 4 Experiments + +We evaluate the effectiveness and efficiency of MASFormer on natural language modeling (ArXiv and PubMed, Cohan et al. (2018)), natural language generation (ArXiv Cohan et al. (2018), QMSUM and GovReport Shaham et al. (2022)). We choose the GPT-3 XL model architecture (Brown et al., 2020) as our base model, which consists of 1.3 billion parameters and 24 layers and is pre-trained on PIL (Gao et al., 2020) for 300 billion tokens. GPT is a general purpose model that can be applied to many tasks instead of tailoring them for specific tasks. As such, it makes easy to control experiments and showcase the difference among various methods. + +Implementation Details. Our base model uses absolute positional embeddings with maximum length 1024. To accommodate longer inputs, we duplicate its positional embeddings to have the maximum length as 8192 such that the model can handle sequences containing up to 8192 tokens. Then, we implement different attention methods by modifying the attention pattern of the base model. We implement all the models with PyTorch (Paszke et al., 2019). All the experiments are conducted on NVIDIA A100 GPUs. + +Continual Training Details. After changing the attention pattern, we conduct the continual training for MASFormer and baseline methods on PILE corpus (Gao et al., 2020) to adapt the revised models to new attention patterns and long-sequence inputs. We leverage the casual language modeling (CLM) objective to train the model for 50,000 steps with a warmup of 2000 steps. We set the input length as 8192 and use a batch size of 128 such that the models are optimized with 1M tokens per step. We use the constant learning 0.0001 for all methods. + +Baseline. We compare MASFormer with the following methods: + +- All full attention is to apply full attention at every layer. It has been adopted by most of existing transformer models as default. Although incurring the maximum attention cost, it achieves the best performance for most of our tasks. Hence, it acts as an upper bound for other methods. +- All block sparse attention is to apply block attention at every layer, which is an effective method to reduce computational cost when modeling long sequences. Block attention sets the attention span of each layer identical such that it evenly distributes the budget of attention computation across layers. + +- All sliding-window attention is to apply sliding-window attention at every layer, which is another variant of sparse attention. It shares the similar spirits and often performs similarly as block attention. + +In the following experiments, we compare MAS-Former and the baseline methods across different attention cost $\mathcal{C}$ . That is, for all block sparse attention, we set the block size as $b = \mathcal{C} / (Ln)$ . For all sliding-window attention, we choose the window size as $w = \mathcal{C} / (2Ln)$ . For MASFormer, we apply a small block size $m = 1024$ for its block attention and set $l$ as $(\mathcal{C} - Lmn) / (n^2 - mn)$ . Then we observe how their performance evolves when enhancing the attention cost $\mathcal{C}$ or input length $n$ . + +Experiment Overview. We briefly summarize the experimental contents as follows: + +- Section 4.1 presents the perplexity evaluation of all the models on ArXiv and PubMed after continual training. +- Section 4.2 compares the summarization performance of the models on ArXiv, QMSUM, and GovReport after fine-tuning. Besides, we also discuss the difference between perplexity and downstream evaluation in reflecting model capacity to capture long-range dependencies. +- Section 4.3 provides three crucial analyses: (i) we evaluate the benefits of increasing input length and discuss the requirements to attain these gains; (ii) we analyze the effectiveness of continual training for long-sequence modeling; (iii) we conduct an ablation study to demonstrate that equipping bottom layers with full attention yields the most significant performance gains than other options. We further provide the explanations. + +# 4.1 Natural Language Modeling + +# 4.1.1 Datasets and Evaluation Details + +Datasets. We evaluate the perplexity of the updated GPT-2 for each attention method after continual training. The evaluation is conducted on test sets of ArXiv and PubMed (Cohan et al., 2018). Table 5 presents the statistics of these two datasets. Pubmed consists of scientific documents, with a document's content used as input and its corresponding abstract as the target summary. ArXiv is similar to PubMed, with documents from arXiv. + +Evaluation Details. We conduct the perplexity evaluation under two settings. (i) We calculate the perplexity (ppl.) with all documents from test sets. Table 1 presents the overall perplexity of different models on two datasets. (ii) To showcase + +![](images/aa87e531b990436c086b6417eaebac905aedebcd05996b5827a767377912c034.jpg) +(a) ArXiv ppl. + +![](images/ea1a52f7f7b59921579ce8383a2cd331d6c63cbfdf1ca31e5cfd12ab892255b2.jpg) +(b) PubMed ppl. +Figure 3: Perplexity evaluation on ArXiv and PubMed with examples of different length. Here $x$ -axis is the maximum document length of each subset, i.e., $k \times 1024$ ( $k = 1, 2, 3, \ldots$ ). + +the varying behaviors of models on documents of different length, we divide all documents into several subsets according to their length. Each subset consists of examples, whose length is within $((k - 1)\times 1024,k\times 1024](k = 1,2,3,\ldots)$ . Then, we evaluate the perplexity on each subset. Figure 3 presents the perplexity of models on different subsets of examples. + +# 4.1.2 Results + +Table 1 compares the overall perplexity on test sets of ArXiv and PubMed. The results suggest that, with $l = 4$ layers of full attention, MASFormer achieves comparable performance to all full attention, while reducing $72\%$ of its attention cost. With the similar attention cost $\mathcal{C}$ , MASFormer outperforms all block attention that evenly distributes the budget of attention computation. For example, MASFormer with $l = 2$ achieves 8.75 ppl. on PubMed, which is 1.37 lower than that of block attention of $b = 2048$ . + +
MethodsCArXivPubMed
Full attention1,610M8.637.63
Block (b=1024)201M13.1912.19
Block (b=2048)402M10.7510.13
MASFormer (l=2)318M10.258.75
MASFormer (l=4)436M9.318.25
MASFormer (l=8)671M9.638.25
+ +Table 1: Perplexity evaluation on ArXiv and PubMed. + +Figure 3 illustrates the perplexity variation of each method given examples of different length. We can tell that MASFormer and full attention show better performance on longer documents, suggesting increasing context length can improve their prediction performance. Full attention, though incurring the highest attention cost, always achieves the best performance due to its outstanding capability to handle sophisticated dependencies. Notably, + +with $27\%$ of its attention cost, MASFormer exhibits a curve of ppl. v.s. length that closely resembles to that of full attention. This demonstrates the effectiveness and efficiency of MASFormer to capture long-range dependencies. In contrast, block sparse attention benefits much less from long contexts and underperforms both of them because of its incapability to encode long-range signals. For example, when $b = 1024$ , block attention achieves similar perplexity on PubMed examples of different length. + +# 4.2 Natural Language Generation + +# 4.2.1 Datasets and Training Details + +Datasets. We evaluate the downstream performance of models on several abstractive summarization tasks to compare their capability of handling long sequences in practice. Specifically, we fine-tune models on ArXiv (Cohan et al., 2018), QM-SUM and GovReport (from SCROLLS benchmark, Shaham et al. (2022)). Their statistics are summarized in Table 5. We mainly use ROUGE-2 (R2) score (Lin, 2004) as the evaluation metric, which is more important and sensitive than R1 and RL. + +Training Details. After continual training, we fine-tune each model and report R2 scores on validation sets. Specifically, we fine-tune models for 3000 steps on QMSUM, 8000 steps on GovReport, and 12000 steps on ArXiv. We set the batch size as 64 for ArXiv and 32 for QMSUM and GovReport. We pick the learning rates from $\{1 \times 10^{-5}, 5 \times 10^{-5}, 1 \times 10^{-4}, 5 \times 10^{-4}\}$ , and choose the optimal ones to report the performance of each method. Moreover, the input length is fixed as 8192. We apply the greedy decoding for generation. Please see Appendix B for more details. + +# 4.2.2 Results + +In Table $2^{2}$ and Figure 4, we present the fine-tuning results on QMSUM, ArXiv and GovReport across + +![](images/7abd031293b0a106fc32b6b3eda383c7fe114cdd09263d0c7c29802ada1078f9.jpg) +(a) ArXiv + +![](images/06f3fa19b661a64eedbac59b93eed38cff3357da2c5afb8d0725d3b7d2bb4714.jpg) +(b) QMSUM + +![](images/9e37f4b3e56d841f515e8e7c046b7e5e493dd9309d4c550d1c24329e45ed393b.jpg) +(c) GovReport +Figure 4: Given input length as 8192, we compare summarization performance between MASFormer and block/full attention when increasing the attention cost. + +
MethodsCQMSUMArXivGovReport
Full attention1610M8.0019.3228.83
Window(w=1024)402M4.3213.5117.03
Block(b=2048)402M5.039.6112.31
MASFormer(l=4)436M6.5914.9118.82
Window(w=2048)805M5.0515.2122.79
Block(b=4096)805M5.1514.5023.64
MASFormer(l=6)553M7.1515.7221.20
MASFormer(l=8)671M7.4617.0024.42
MASFormer(l=12)906M8.7018.5826.26
+ +Table 2: Summarization performance of models with different attention methods. The best results are shown in bold. + +different attention cost. The results demonstrate that, with the similar attention cost, MASFormer significantly outperforms sparse attention variants. Furthermore, when enhancing attention cost, MASFormer achieves greater performance gains than sparse attention methods. This is evident from the steeper slope of its R2 curve versus attention cost, in contrast to the baseline method. For example, when increasing $\mathcal{C}$ form 553M to 671M, the R2 score of MASFormer on QMSUM exhibits a substantial improvement, reaching 8.70 from 7.46. Remarkably, this score surpasses even that of full attention. Therefore, MASFormer addresses the trade-off between computational cost and performance gains in a more efficient and effective way. + +Notice that, in order to achieve comparable summarization performance to full attention, MAS-Former needs at least $l = 8$ layers of full attention, and providing more can lead to more gains. This observation is different from the findings in NLM (Figure 3) that increasing $l$ beyond 4 provides limited improvement in perplexity. Their different capacity requirements arise from the fact that predicting next tokens in NLM primarily relies on lo + +cal dependencies. Capturing infrequent long-range tokens does not significantly improve perplexity. Thus, this discrepancy emphasizes the necessity to evaluate long-range models on downstream tasks. + +# 4.3 Analysis + +# 4.3.1 Benefits of Increasing Sequence Length + +In this section, we investigate the benefits of increasing input length for downstream performance. Specifically, we select the input length from $\{2048, 4096, 6144, 8192\}$ and present the fine-tuning performance of full attention in Figure 5. The results consistently demonstrate that as the input length increases, the model's performance improves. That is, downstream performance benefits significantly from long-sequence inputs. In contrast, increasing example length beyond $6k$ results in marginal improvements in perplexity (See Figure 3), highlighting again the importance of downstream evaluation. + +In addition, when comparing the behaviors of block attention in Figure 2c and 2d, we find that sparse attention often insufficiently capitalize on the benefits offered by longer inputs. For instance, given block size as 4096, its performance on ArXiv remains nearly unchanged when increasing input length from 4096 (R2 = 15.52 in Figure 5a) to 8192 (R2 = 14.49 in Figure 2c). This finding suggests that enhancing input length can only improve model performance if the model possesses the sufficient capability to handle long-range dependencies. + +# 4.3.2 Effectiveness of Continual Training + +We analyze the effectiveness of continual training by comparing fine-tuning performance of MAS-Former $(l = 8)$ under the following settings: (i) MASFormer without continual training (w.o. C.T.); (ii) MASFormer continually trained with short inputs (C.T. $(n = 2048)$ ); (iii) MASFormer continually trained with long inputs (C.T. $(n = 8192)$ ). Table 3 + +![](images/34f8b43ff943ff02432a8efcea092d5101ce4681ebe8062ad2ef1f4556cd815e.jpg) +(a) ArXiv + +![](images/243cd9585f95f0cec8f02ae7e098f293e5b08b58846d6884995bfa11ff8a6731.jpg) +(b) GovReport + +![](images/576c77e2d4168b14669caaa7ef1f13475630024a6b889195642c7a54823bebc1.jpg) +(c) QMSUM +Figure 5: Fine-tuning performance of full attention under different input length. + +presents fine-tuning performance of these models. We can tell that continual training with long inputs indeed facilitates the revised models to adapt to new structures and long-sequence inputs. + +
l=8QMSUMGovReport
w.o. C.T.29.33/6.43/25.7153.28/23.61/51.74
C.T. (n=2048)29.87/7.16/26.1552.28/23.01/49.83
C.T. (n=8192)30.91/7.45/27.0254.37/24.42/51.87
+ +# 4.3.3 Where to use full attention + +To answer where to apply full attention, we compare fine-tuning performance of MASFormers that apply full attention at (i) bottom layers; (ii) middle layers; (iii) top layers; (iv) every $L / l$ layers. The results in Table 4 demonstrate that equipping bottom layers with full attention yields the best performance. This is because that long-range dependencies can be continually captured and reinforced by bottom layers before propagated to upper layers. As such, these long-range signals can be effectively incorporated into the upper layers with local attention, facilitating their encoding of local information. In contrast, when equipping local attention at bottom layers, long-range tokens are first aggregated with neighboring tokens by local attention, thereby weakening their long-range signals. Moreover, if alternating full and local attention every $L / l$ layers, the long-range signals cannot be continually reinforced nor efficiently captured. + +# 5 Discussion + +GPT-Neo (Black et al., 2021) introduces an attention pattern that alternates full and window attention. However, this models is not tailored for long sequences. It sets the local window size as 256 and has the maximum input length as 2048, unable to handle long sequences. Instead, this attention pattern is applied heuristically in an attempt to re + +Table 3: We report R1/R2/RL for the above results. + +
PositionQMSUMGovReport
Every 328.26/6.94/25.0326.16/12.37/24.82
Top 820.89/4.52/18.37-/-/-
Middle 827.27/5.99/24.0620.80/9.01/19.52
Bottom 830.91/7.45/27.0254.37/24.42/51.87
Every 231.27/8.19/27.4135.34/16.04/33.68
Bottom 1232.53/8.70/28.7556.98/26.26/54.46
+ +Table 4: Performance comparison of MASFormers that apply full attention at different layers (# layers $L = 24$ + +duce computational cost. However, as discussed in Section 4.3.3, this approach is neither effective nor efficient as MASFormer when handling long sequences. As shown in Table 4, applying full attention at every 2 or 3 layers underperforms applying it at bottom 12 or 8 layers. Therefore, alternating between full and block attention results in additional computational cost and performance degradation. + +In contrast, MASFormer presents an effective solution for efficient long-sequence modeling. It provides guidance on adapting existing pre-trained transformers to long inputs. Meanwhile, it provides insights for designing large long-range models, especially for deeper models. By equipping only a subset of bottom layer with full attention, we can substantially mitigate computational cost. Additionally, the computation of MASFormer can be further optimized by leveraging system-level acceleration techniques (e.g., FlashAttention and xFormer) that support both block and full attention. + +# 6 Conclusion + +We propose an efficient long-range transformer - MASFormer that utilizes full attention at a few of bottom layers and employs sparse attention at the remaining layers. Our empirical results on natural language modeling and generation tasks demonstrate that MASFormer can address the trade-off between computational cost and performance gains in a more efficient and effective way. + +# References + +Joshua Ainslie, Santiago Ontanon, Chris Alberti, Va-clav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. Etc: Encoding long and structured inputs in transformers. arXiv preprint arXiv:2004.08483. +Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150. +Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow. +Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual. +Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Re. 2021. Scatterbrain: Unifying sparse and low-rank attention approximation. arXiv preprint arXiv:2110.15343. +Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. A discourse-aware attention model for abstractive summarization of long documents. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 615-621, New Orleans, Louisiana. Association for Computational Linguistics. +Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860. +Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages + +4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. +Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027. +Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. 2020. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487. +Albert Gu, Karan Goel, and Christopher Ré. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396. +Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, and Yinfei Yang. 2021. Longt5: Efficient text-to-text transformer for long sequences. arXiv preprint arXiv:2112.07916. +Ankit Gupta, Albert Gu, and Jonathan Berant. 2022. Diagonal state spaces are as effective as structured state spaces. In Advances in Neural Information Processing Systems. +Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543. +Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations. +Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. In International Conference on Learning Representations. +Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. 2022. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers. +Yuhong Li, Tianle Cai, Yi Zhang, Deming Chen, and Debadepta Dey. 2022. What makes convolutional models great on long sequence modeling? arXiv preprint arXiv:2210.09298. +Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81. + +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. 2023. Mega: Moving average equipped gated attention. In *The Eleventh International Conference on Learning Representations*. +Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 8024-8035. +Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. 2021. Random feature attention. In International Conference on Learning Representations. +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. +Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2021. Recipes for building an open-domain chatbot. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 300-325, Online. Association for Computational Linguistics. +Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, et al. 2022. Scrols: Standardized comparison over long language sequences. arXiv preprint arXiv:2201.03533. +Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. 2023. Simplified state space layers for sequence modeling. In *The Eleventh International Conference on Learning Representations*. +Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke, and Mohit Iyyer. 2021. Do long-range language models actually use long-range context? arXiv preprint arXiv:2109.09115. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. + +Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768. +Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297. +Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adaptive budget allocation for parameter-efficient fine-tuning. In The Eleventh International Conference on Learning Representations. +Simiao Zuo, Xiaodong Liu, Jian Jiao, Denis Charles, Eren Manavoglu, Tuo Zhao, and Jianfeng Gao. 2022. Efficient long sequence modeling via state space augmented transformer. arXiv preprint arXiv:2212.08136. + +# A Dataset Statistics + +In the following table, we provide the detailed statistics of datasets in our experiments, including example splits and length statistics. + +
DatasetExample CountInput Length
TrainValidTestAverageMedian90th percentile
ArXiv203,0376,4366,44010,7208,51920,170
PubMed119,9246,6336,6584,7483,8838,883
QMSUM1,2572722819,49714,19727,761
GovReport17,4579729737,8868,84118,835
+ +# B Natural Language Generation + +# B.1 The Results of All ROUGE Scores + +Table 5: Statistics of datasets. Input length measured in tokens using a SentencePiece Model. + +
MethodsCQMSUMArXivGovReport
Full attention1610M31.50 / 8.00 / 27.8146.13 / 19.32 / 41.8960.53 / 28.83 / 57.88
Window (w=1024)402M23.31 / 4.32 / 20.6235.90 / 13.51 / 32.1949.82 / 17.03 / 47.42
Window (w=2048)805M26.73 / 5.05 / 23.4038.74 / 15.21 / 34.8756.14 / 22.79 / 53.50
Block (b=2048)402M26.24 / 5.03 / 23.1321.85 / 9.61 / 19.8626.37 / 12.31 / 25.18
Block (b=4096)805M26.96 / 5.15 / 23.8535.95 / 14.50 / 32.3749.83 / 23.64 / 47.50
MASFormer (l=4)436M29.86 / 6.59 / 25.8738.85 / 14.91 / 34.9846.67 / 18.82 / 44.39
MASFormer (l=6)553M30.83 / 7.15 / 27.1236.29 / 15.72 / 32.9649.26 / 21.20 / 46.89
MASFormer (l=8)671M30.91 / 8.00 / 27.8143.31 / 17.00 / 39.1254.37 / 24.42 / 51.87
MASFormer (l=12)906M32.53 / 8.70 / 28.7545.19 / 18.58 / 40.7256.98 / 26.26 / 54.46
+ +# B.2 Training Details + +Table 6: Finetuning performance of different attention methods. + +
MethodsQMSUMArXivGovReport
Full attention1 × 10-51 × 10-41 × 10-4
Window attention (w=1024)5 × 10-45 × 10-55 × 10-4
Window attention (w=2048)5 × 10-55 × 10-51 × 10-4
Block attention (w=2048)1 × 10-41 × 10-51 × 10-5
Block attention (w=4096)5 × 10-55 × 10-41 × 10-5
MASFormer (l=4)5 × 10-51 × 10-35 × 10-4
MASFormer (l=6)5 × 10-55 × 10-55 × 10-4
MASFormer (l=8)5 × 10-55 × 10-45 × 10-4
MASFormer (l=12)1 × 10-51 × 10-41 × 10-4
+ +Table 7: The fine-tuning learning rate of each method on each dataset. + +We conduct continual training for all attention methods with training data form PILE and input length as 8192. After continual training, we obtain the continually trained models for each method and fine-tune them on QMSUM, ArXiv and GovReport to compare their summarization performance. During the + +fine-tuning, we set the input length as 8192 for all datasets and all models. We apply the greedy decoding for generation and set the maximum output length as 256 for QMSUM, 1024 for GovReport, and 512 for ArXiv. Table 8 lists the details of these hyperparameters. Besides, we apply the linear learning rate schedule to fine-tune the models and the base learning rates are summarized in Table 7. + +
HyperparameterQMSUMArXivGovReport
Training steps3000120008000
Batch size323264
Input length819281928192
Maximum generation length2565121024
Weight decay0.0010.0010.001
+ +Table 8: The other fine-tuning parameters for each dataset, which remain the same for every method. \ No newline at end of file diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/images.zip b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8323fc205b04724744f87ca9d314e66cdedfd238 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25fa93f7fbe6283c02d25ada53515b4429ab63535037b1b633d1280d92a74c0c +size 597216 diff --git a/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/layout.json b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a657822f82fd91f3339eeeb905ae957adf3e1828 --- /dev/null +++ b/2023/Efficient Long-Range Transformers_ You Need to Attend More, but Not Necessarily at Every Layer/layout.json @@ -0,0 +1,8161 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 75, + 495, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 75, + 495, + 110 + ], + "spans": [ + { + "bbox": [ + 102, + 75, + 495, + 110 + ], + "type": "text", + "content": "Efficient Long-Range Transformers: You Need to Attend More, but Not Necessarily at Every Layer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 125, + 497, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 125, + 497, + 139 + ], + "spans": [ + { + "bbox": [ + 102, + 125, + 497, + 139 + ], + "type": "text", + "content": "Qingru Zhang†, Dhananjay Ram, Cole Hawkins, Sheng Zha, Tuo Zhao" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "spans": [ + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "type": "text", + "content": "Georgia Institute of Technology " + }, + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "type": "inline_equation", + "content": "\\diamond" + }, + { + "bbox": [ + 160, + 140, + 437, + 153 + ], + "type": "text", + "content": " Amazon Web Service" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 155, + 401, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 155, + 401, + 168 + ], + "spans": [ + { + "bbox": [ + 195, + 155, + 401, + 168 + ], + "type": "text", + "content": "{qingru.zhang,tourzhao}@gatech.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 169, + 410, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 169, + 410, + 181 + ], + "spans": [ + { + "bbox": [ + 187, + 169, + 410, + 181 + ], + "type": "text", + "content": "{radhna,colehawk,zhasheng}@amazon.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 231, + 274, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 231, + 274, + 625 + ], + "spans": [ + { + "bbox": [ + 86, + 231, + 274, + 625 + ], + "type": "text", + "content": "Pretrained transformer models have demonstrated remarkable performance across various natural language processing tasks. These models leverage the attention mechanism to capture long- and short-range dependencies in the sequence. However, the (full) attention mechanism incurs high computational cost – quadratic in the sequence length, which is not affordable in tasks with long sequences, e.g., inputs with 8k tokens. Although sparse attention can be used to improve computational efficiency, as suggested in existing work, it has limited modeling capacity and often fails to capture complicated dependencies in long sequences. To tackle this challenge, we propose MASFormer, an easy-to-implement transformer variant with Mixed Attention Spans. Specifically, MASFormer is equipped with full attention to capture long-range dependencies, but only at a small number of layers. For the remaining layers, MASformer only employs sparse attention to capture short-range dependencies. Our experiments on natural language modeling and generation tasks show that a decoder-only MASFormer model of 1.3B parameters can achieve competitive performance to vanilla transformers with full attention while significantly reducing computational cost (up to " + }, + { + "bbox": [ + 86, + 231, + 274, + 625 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 86, + 231, + 274, + 625 + ], + "type": "text", + "content": "). Additionally, we investigate the effectiveness of continual training with long sequence data and how sequence length impacts downstream generation performance, which may be of independent interest." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 632, + 155, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 632, + 155, + 645 + ], + "spans": [ + { + "bbox": [ + 68, + 632, + 155, + 645 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 652, + 292, + 747 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 292, + 747 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 292, + 747 + ], + "type": "text", + "content": "Pre-trained transformer models have manifested superior performance in various natural language processing tasks such as natural language modeling (NLM) (Dai et al., 2019; Radford et al., 2019), natural language generation (NLG) (Brown et al., 2020) and natural language understanding (NU) (Devlin et al., 2019; Liu et al., 2019; He et al., 2021b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 253 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 253 + ], + "type": "text", + "content": "These models leverage the attention mechanism (Vaswani et al., 2017) to compute the dependency score for each pair of tokens in an input sequence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 256, + 527, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 256, + 527, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 256, + 527, + 513 + ], + "type": "text", + "content": "Some practical tasks require these transformer models to handle long-sequence inputs like 8k tokens. For example, chatbot systems gather long-term contexts of user interactions to generate informative texts (Roller et al., 2021). Summarization for news, government reports, and academic papers request models to take inputs of long sequences to generate comprehensive summaries (Shaham et al., 2022), otherwise models often miss important information. Note that typical transformer models apply full attention to capture token dependencies pair-wise. It leads to a quadratic time and space complexity w.r.t. input length. However, such a complexity is prohibitive for long sequences. In particular, it incurs massive memory consumption during the back propagation. For example, a transformer model with 250M parameters consumes over 80G GPU memory when sequence length is 8k (Zuo et al., 2022)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 516, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 527, + 773 + ], + "type": "text", + "content": "To address this scalability issue, various approaches have been proposed to reduce the complexity. One approach is sparse attention, which restricts each token to attend a subset of tokens based on predefined sparsity patterns (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020). For instance, block sparse attention (Kitaev et al., 2020; Ma et al., 2023) divides the input sequence into several blocks, and only intra-block attention is performed. Besides, sliding-window attention (Beltagy et al., 2020; Zaheer et al., 2020; Ainslie et al., 2020) allows each token to attend to its neighboring tokens within a sliding window. These methods, though reducing the complexity of full attention, cannot sufficiently capture long-range dependencies. Other variants, such as kernel approximation (Peng et al., 2021) and low-rank approximation (Wang et al., 2020; Chen et al., 2021) methods, share the similar spirit and drawbacks. To com" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "text", + "content": "* Work was done during Qingru Zhang's internship at Amazon Web Service." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 285, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 285, + 780, + 310, + 791 + ], + "type": "text", + "content": "2775" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2775-2786 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": "pensate for the lack of long-range dependencies, LongT5 (Guo et al., 2021) introduces global tokens that are obtained by average pooling on every block of tokens (Ainslie et al., 2020). However, the block pooling operations can weaken the signal of crucial tokens and prevent the long-range dependencies from being detected." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 172, + 291, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 172, + 291, + 414 + ], + "spans": [ + { + "bbox": [ + 69, + 172, + 291, + 414 + ], + "type": "text", + "content": "In addition to these methods, state space models (SSMs) prespecify global dependency patterns to capture the long-range dependencies only (Gu et al., 2020, 2021; Li et al., 2022; Zuo et al., 2022; Ma et al., 2023; Smith et al., 2023). These models can be regarded as linear recurrent neural networks with specifically designed fixed weights. As tailored for global dependencies, SSMs fail to effectively capture local dependencies. In order to combine both local and global dependencies, SPADE (Zuo et al., 2022) and MEGA (Ma et al., 2023) augment SSM layers into transformer layers equipped with local attention. However, state space methods require sophisticated implementation, and often encounter computational instability during the back propagation, especially when scaling up to large model size (Gupta et al., 2022). SPADE and MEGA hence inherit these drawbacks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 422, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 291, + 772 + ], + "type": "text", + "content": "Note that the aforementioned methods apply same attention mechanism for every layer. We challenge this conventional wisdom and propose a transformer variant - MASFormer (Mixed Attention Span Transformer). MASFormer utilizes full attention only at a subset of layers whereas employs sparse attention at the remaining layers. Our design is motivated by the phenomenon - that most contexts in NLP data display a great deal of locality of reference (Zaheer et al., 2020; Beltagy et al., 2020). That is, most of information about a token can be derived from its neighboring tokens. In contrast, long-range dependencies among tokens are sparse and infrequent. Consider an academic paper as an example. Within a paragraph, there exist numerous short-term dependencies. Neighboring tokens are closely connected to convey meaningful semantics. Across paragraphs, there can be a small number of long-range dependencies. For example, tokens associated to the primary theme of the paper exhibit rare and weak dependencies across a long span. Since long-range dependencies occur much less frequently, a few layers of full attention are adequate to capture them. In stark contrast, short-term dependencies are more frequent, necessitating local attention in the majority of layers to fully extract" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 303, + 71, + 364, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 364, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 364, + 84 + ], + "type": "text", + "content": "these signals." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 90, + 526, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 90, + 526, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 90, + 526, + 321 + ], + "type": "text", + "content": "To demonstrate the effectiveness of MASFormer, We conduct experiments on natural language modeling (ArXiv and PubMed Cohen et al. (2018)) and natural language generation (ArXiv, Cohen et al. (2018) and SCROLLS, Shaham et al. (2022)) tasks. Specifically, we compare the performance of MASFormer to other attention methods using a pretrained GPT-2 model (Radford et al., 2019) of 1.3 billion parameters. Our empirical results demonstrate that MASFormer consistently outperforms baseline methods across different attention cost (i.e. the total number of computed attention scores). In particular, MASFormer can achieve comparable performance to full attention while significantly reducing the computational cost. For example, with " + }, + { + "bbox": [ + 302, + 90, + 526, + 321 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 302, + 90, + 526, + 321 + ], + "type": "text", + "content": " of its attention cost, MASFormer achieves a close R2 score as full attention on QMSUM dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 324, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 324, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 324, + 526, + 772 + ], + "type": "text", + "content": "We also make additional discoveries with MASFormer, which are of independent interest. Firstly, we investigate the effectiveness of continual training for long sequence modeling. Many publicly available models are pre-trained with sequences shorter than 2048, and often fail to perform well on longer sequences (e.g. 8k/16k tokens). To bridge the gap, we explore the option of continual training to adapt these models to long sequences, thereby avoiding pre-training from the scratch. We discuss its effectiveness with MASFormer in Section 4.3. Secondly, we showcase that increasing sequence length can yield more performance gains on downstream tasks than NLM tasks evaluated by perplexity. We are aware of the recent findings by Sun et al. (2021) that increasing context length exhibits limited impact on NLM perplexity. Nevertheless, when applying MASFormer to downstream tasks like long-context summarization, we find that model performance benefits significantly from extending context length. Such a difference arises from the fact that predicting the next tokens in NLM primarily relies on locality of reference. Capturing infrequent long-range tokens can improve perplexity but not significantly. Therefore, we emphasize the necessity to evaluate model performance on downstream tasks that require long-range dependencies. Furthermore, our empirical evidence suggests that increasing the length can improve the performance only if models possess sufficient capability to handle additional long-range information. Local attention, as a counterexample, often fails to capture long-range signals and hence" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2776" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 246, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 246, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 246, + 84 + ], + "type": "text", + "content": "benefits much less from long sequences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 95, + 152, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 95, + 152, + 108 + ], + "spans": [ + { + "bbox": [ + 67, + 95, + 152, + 108 + ], + "type": "text", + "content": "2 Background" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 116, + 232, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 116, + 232, + 130 + ], + "spans": [ + { + "bbox": [ + 67, + 116, + 232, + 130 + ], + "type": "text", + "content": "2.1 Pretrained Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 134, + 290, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 134, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 134, + 290, + 269 + ], + "type": "text", + "content": "Pre-trained transformer models (Devlin et al., 2019; Liu et al., 2019; Brown et al., 2020; Dosovitskiy et al., 2020; He et al., 2021b,a) have manifested superior performance in various NLP tasks. These models are often pre-trained on enormous amounts of unlabeled data in a unsupervised/self-supervised manner such that they can learn rich semantic knowledge. By further fine-tuning these pre-trained models, we can effectively transfer such knowledge to benefit downstream tasks (Zhang et al., 2023)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 270, + 290, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 270, + 290, + 405 + ], + "spans": [ + { + "bbox": [ + 67, + 270, + 290, + 405 + ], + "type": "text", + "content": "Existing research on long-range transformers commonly requires pre-training the proposed models from scratch to accommodate new architectures and long inputs (Guo et al., 2021; Zuo et al., 2022). However, the significant training overheads raise a barrier for the widespread utilization of these methods across different language models. Motivated by this, we explore the possibility of leveraging existing pre-trained models and adapting them to long sequences though continual training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 415, + 197, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 415, + 197, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 415, + 197, + 427 + ], + "type": "text", + "content": "2.2 Attention Mechanism" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "text", + "content": "Suppose the input to the layer is " + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "inline_equation", + "content": "X\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "text", + "content": " is the input sequence length and " + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 67, + 432, + 289, + 472 + ], + "type": "text", + "content": " is embedding dimension, then self-attention mechanism outputs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 103, + 479, + 290, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 479, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 103, + 479, + 290, + 510 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} (X) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V \\tag {1}", + "image_path": "15af3ba9115b861905864b85ec5944cf54f69c4b3042f6c0f0aa0e54b2c7036a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "Q = XW_{q}, K = XW_{k}, V = VW_{v}" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}, W_{v}" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": " are learnable projection weights. Such full attention can simultaneously evaluate the alignment between any pair of tokens in the sequence. Specifically, denote the attention score " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "A = \\text{softmax}(QK^{\\top} / \\sqrt{d})" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "A_{ij}" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": " captures the alignment between tokens " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": ". A typical transformer model applies the full attention at every layer. Denote the number of layers as " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": ". Then its attention cost is " + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "inline_equation", + "content": "Ln^{2}" + }, + { + "bbox": [ + 67, + 515, + 290, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "text", + "content": "Sparse attention variants are introduced to mitigate the computational cost of full attention. Figures 1a and 1b illustrates the attention patterns of block sparse attention and sliding-window attention. For instance, block sparse attention divides tokens into blocks of size " + }, + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "text", + "content": " and performs intrablock attention only, resulting in an attention cost of " + }, + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "inline_equation", + "content": "bn" + }, + { + "bbox": [ + 67, + 651, + 290, + 772 + ], + "type": "text", + "content": ". Sliding-window attention allows each token to attend its left/right neighboring tokens within a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "content": "local window of size " + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 302, + 71, + 524, + 111 + ], + "type": "text", + "content": ". In most of cases, block sparse attention exhibits similar performance as sliding-window attention (Zuo et al., 2022)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 118, + 400, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 118, + 400, + 132 + ], + "spans": [ + { + "bbox": [ + 302, + 118, + 400, + 132 + ], + "type": "text", + "content": "3 Our Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 137, + 525, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 137, + 525, + 177 + ], + "spans": [ + { + "bbox": [ + 302, + 137, + 525, + 177 + ], + "type": "text", + "content": "We present our method - MASFormer, a long-range transformer variant that mixes different attention spans across layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 182, + 500, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 182, + 500, + 196 + ], + "spans": [ + { + "bbox": [ + 302, + 182, + 500, + 196 + ], + "type": "text", + "content": "3.1 MASFormer: Mixed Attention Span" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 200, + 525, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 200, + 525, + 416 + ], + "spans": [ + { + "bbox": [ + 301, + 200, + 525, + 416 + ], + "type": "text", + "content": "MASFormer leverages full attention exclusively at a subset of transformer layers, whereas it employs block sparse attention at the remaining layers. The structure of MASFormer is illustrated in Figure 1c. We choose full attention to encode long-range information due to the following reasons: (i) full attention exhibits superior capability to capture long-range dependencies compared to sparse attention; (ii) full attention does not require sophisticated implementation and hence is computationally stable compared to SSMs (Zuo et al., 2022; Gupta et al., 2022); (iii) full attention is compatible with existing pre-trained transformer models, enabling us to conduct continual training which we elaborate in Section 3.2. To mitigate the computational cost, we restrict the number of layers using full attention." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 417, + 525, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 417, + 525, + 605 + ], + "spans": [ + { + "bbox": [ + 302, + 417, + 525, + 605 + ], + "type": "text", + "content": "MASFormer is motivated by empirical investigations on performance comparison between models that apply the same attention span at every layer. Figure 2 presents the performance of block sparse attention and full attention on language modeling and summarization tasks. We find that, given long-sequence inputs, sparse attention is often insufficient to capture long-range dependencies beyond its attention span. As a result, it shows unsatisfactory performance. To remedy it, one can either increase attention span or switch to full attention to improve model capability of capturing sophisticated dependencies. Though improving model performance, it incurs high computational cost." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 606, + 525, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 606, + 525, + 754 + ], + "spans": [ + { + "bbox": [ + 302, + 606, + 525, + 754 + ], + "type": "text", + "content": "Confronting such a trade-off between computational cost and model performance, we challenge the common practice - that applies the same attention span at every layer. MASFormer provides an alternative solution. Instead of increasing attention span evenly, MASFormer allocates a large portion of attention computations to a subset of " + }, + { + "bbox": [ + 302, + 606, + 525, + 754 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 302, + 606, + 525, + 754 + ], + "type": "text", + "content": " layers by equipping them with full attention. Specifically, equipping bottom layers with full attention can yield the best performance as suggested by our empirical analysis in Section 4.31. At the remain-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 492, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 492, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 492, + 772 + ], + "type": "text", + "content": "1Please see Section 4.3 for detailed explanations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2777" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 95, + 152, + 178 + ], + "blocks": [ + { + "bbox": [ + 70, + 95, + 152, + 178 + ], + "lines": [ + { + "bbox": [ + 70, + 95, + 152, + 178 + ], + "spans": [ + { + "bbox": [ + 70, + 95, + 152, + 178 + ], + "type": "image", + "image_path": "b14e862e159937c0a8d0aba62f3dcf18dc7295a58d45f99d7a2c8abb6f7d2625.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 200, + 146, + 211 + ], + "lines": [ + { + "bbox": [ + 74, + 200, + 146, + 211 + ], + "spans": [ + { + "bbox": [ + 74, + 200, + 146, + 211 + ], + "type": "text", + "content": "(a) Block Attention" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 158, + 95, + 242, + 178 + ], + "blocks": [ + { + "bbox": [ + 158, + 95, + 242, + 178 + ], + "lines": [ + { + "bbox": [ + 158, + 95, + 242, + 178 + ], + "spans": [ + { + "bbox": [ + 158, + 95, + 242, + 178 + ], + "type": "image", + "image_path": "e385a1ca1c74eb15b1f237f6586b185c1f36c971d83bf972e19739324ca47bb6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 200, + 240, + 211 + ], + "lines": [ + { + "bbox": [ + 159, + 200, + 240, + 211 + ], + "spans": [ + { + "bbox": [ + 159, + 200, + 240, + 211 + ], + "type": "text", + "content": "(b) Window Attention" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 252, + 72, + 524, + 196 + ], + "blocks": [ + { + "bbox": [ + 252, + 72, + 524, + 196 + ], + "lines": [ + { + "bbox": [ + 252, + 72, + 524, + 196 + ], + "spans": [ + { + "bbox": [ + 252, + 72, + 524, + 196 + ], + "type": "image", + "image_path": "e07b6b4719286115e818c89a0512ae2593a10b031ab89993a43a5e73348c59ee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 200, + 417, + 211 + ], + "lines": [ + { + "bbox": [ + 357, + 200, + 417, + 211 + ], + "spans": [ + { + "bbox": [ + 357, + 200, + 417, + 211 + ], + "type": "text", + "content": "(c) MASFormer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 253, + 177, + 332 + ], + "blocks": [ + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "lines": [ + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "spans": [ + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "type": "text", + "content": "Figure 1: Illustration of attention patterns of (a) block sparse attention with block size " + }, + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "type": "inline_equation", + "content": "b = 3" + }, + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "type": "text", + "content": "; (b) sliding-window attention with window size " + }, + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "type": "inline_equation", + "content": "w = 1" + }, + { + "bbox": [ + 67, + 215, + 524, + 239 + ], + "type": "text", + "content": " (on each side); (c) MASFormer that integrates full and sparse attention." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 253, + 177, + 332 + ], + "lines": [ + { + "bbox": [ + 74, + 253, + 177, + 332 + ], + "spans": [ + { + "bbox": [ + 74, + 253, + 177, + 332 + ], + "type": "image", + "image_path": "cdc86d8d963cb9cd155602b3ad9c20143a6c51bbf63bd87fc5c7884176cbac8c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 334, + 161, + 344 + ], + "lines": [ + { + "bbox": [ + 90, + 334, + 161, + 344 + ], + "spans": [ + { + "bbox": [ + 90, + 334, + 161, + 344 + ], + "type": "text", + "content": "(a) NLM for ArXiv" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 349, + 524, + 386 + ], + "lines": [ + { + "bbox": [ + 67, + 349, + 524, + 386 + ], + "spans": [ + { + "bbox": [ + 67, + 349, + 524, + 386 + ], + "type": "text", + "content": "Figure 2: (a,b): We evaluate the perplexity of a pre-trained GPT-2 model with block attention of differnet block size after continual training. (c,d): We fine-tune a GPT-2 model with block attention and compare the summarization performance on ArXiv and GovReport under different block size. Here the input length " + }, + { + "bbox": [ + 67, + 349, + 524, + 386 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 349, + 524, + 386 + ], + "type": "text", + "content": " is 8192." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 189, + 253, + 291, + 332 + ], + "blocks": [ + { + "bbox": [ + 189, + 253, + 291, + 332 + ], + "lines": [ + { + "bbox": [ + 189, + 253, + 291, + 332 + ], + "spans": [ + { + "bbox": [ + 189, + 253, + 291, + 332 + ], + "type": "image", + "image_path": "d9ac5c06d068f9a7c4a4b26666d22d047f491027c96e6ad0643d82ebc258f6e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 334, + 280, + 344 + ], + "lines": [ + { + "bbox": [ + 199, + 334, + 280, + 344 + ], + "spans": [ + { + "bbox": [ + 199, + 334, + 280, + 344 + ], + "type": "text", + "content": "(b) NLM for PubMed" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 302, + 253, + 405, + 332 + ], + "blocks": [ + { + "bbox": [ + 302, + 253, + 405, + 332 + ], + "lines": [ + { + "bbox": [ + 302, + 253, + 405, + 332 + ], + "spans": [ + { + "bbox": [ + 302, + 253, + 405, + 332 + ], + "type": "image", + "image_path": "3d353730f2b9c2e78c69ce82fb754ba1bc1c9b583bdf4740544689399a05475b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 334, + 388, + 344 + ], + "lines": [ + { + "bbox": [ + 318, + 334, + 388, + 344 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 388, + 344 + ], + "type": "text", + "content": "(c) Sum. for ArXiv" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 417, + 253, + 518, + 332 + ], + "blocks": [ + { + "bbox": [ + 417, + 253, + 518, + 332 + ], + "lines": [ + { + "bbox": [ + 417, + 253, + 518, + 332 + ], + "spans": [ + { + "bbox": [ + 417, + 253, + 518, + 332 + ], + "type": "image", + "image_path": "76170219267e1227827b845500c7f6484232a8edfc99eafc43cbac1dbf902c3d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 334, + 511, + 346 + ], + "lines": [ + { + "bbox": [ + 423, + 334, + 511, + 346 + ], + "spans": [ + { + "bbox": [ + 423, + 334, + 511, + 346 + ], + "type": "text", + "content": "(d) Sum. for GovReport" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "spans": [ + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "text", + "content": "ing layers, MASFormer utilizes block attention of small size " + }, + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "text", + "content": ", resulting in a controlled attention cost of " + }, + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "inline_equation", + "content": "(L - l)mn + ln^2" + }, + { + "bbox": [ + 67, + 392, + 290, + 566 + ], + "type": "text", + "content": ". As mentioned in Section 1, such a design is inspired by the phenomenon that most of contexts in NLP data exhibit a great deal of locality of reference. Long-range dependencies, in contrast, are less frequent. Therefore, it is not necessary to enhance attention span at every layer. Instead, a few layers of full attention are sufficient to capture infrequent long-range signals. The majority of layers can maintain small attention spans to adequately extract local dependencies and control the attention cost." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 569, + 290, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 569, + 290, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 569, + 290, + 677 + ], + "type": "text", + "content": "Our empirical results demonstrate that, with the same attention cost, MASFormer significantly outperforms sparse attention. Remarkably, MASFormer can achieve comparable performance to full attention while substantially reducing computational cost. Therefore, by mixing different attention spans, MASFormer strikes a better balance between computational cost and model performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": "Moreover, MASFormer offers additional implementation advantages. As using the same attention function, MASFormer is easy to implement and compatible with existing pre-trained models. We can build MASFormer upon pre-trained transformers by changing their attention patterns, which does not involve modification on model architectures" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 392, + 525, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 392, + 525, + 458 + ], + "spans": [ + { + "bbox": [ + 302, + 392, + 525, + 458 + ], + "type": "text", + "content": "and pre-trained weights. Meanwhile, acceleration packages, such as FlashAttention (Dao et al., 2022) and xFormers (Lefaudeauux et al., 2022), are applicable to further accelerate the computation of block attention and full attention in MASFormer." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 470, + 522, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 470, + 522, + 483 + ], + "spans": [ + { + "bbox": [ + 302, + 470, + 522, + 483 + ], + "type": "text", + "content": "3.2 Continual Training with Long Sequences" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 301, + 489, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 489, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 489, + 525, + 772 + ], + "type": "text", + "content": "As mentioned, MASFormer can be implemented upon majority of pre-trained transformers by modifying their attention patterns. However, most of publicly available models are pre-trained with sequences shorter than 2048, and often exhibit subpar performance on longer sequences such as " + }, + { + "bbox": [ + 301, + 489, + 525, + 772 + ], + "type": "inline_equation", + "content": "8\\mathrm{k} / 16\\mathrm{k}" + }, + { + "bbox": [ + 301, + 489, + 525, + 772 + ], + "type": "text", + "content": ". To bridge this gap, we propose the continual training to adapt the revised model on long sequences and new attention pattern. As such, we can preserve existing pre-trained knowledge and circumvent the intensive overheads of pre-training from scratch. In particular, we first modify the attention pattern of the target model as proposed by MASFormer. If the pre-trained model uses absolute position embeddings, we duplicate them to accommodate long sequences. Subsequently, we provide the revised model with long sequences (e.g., 8k) from pretraining corpus like PILE (Gao et al., 2020). Then we conduct continual pre-training using casual language modeling (CLM) objective. We discuss the effectiveness of continual training in Section 4.3." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2778" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 154, + 84 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 291, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 291, + 278 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 291, + 278 + ], + "type": "text", + "content": "We evaluate the effectiveness and efficiency of MASFormer on natural language modeling (ArXiv and PubMed, Cohan et al. (2018)), natural language generation (ArXiv Cohan et al. (2018), QMSUM and GovReport Shaham et al. (2022)). We choose the GPT-3 XL model architecture (Brown et al., 2020) as our base model, which consists of 1.3 billion parameters and 24 layers and is pre-trained on PIL (Gao et al., 2020) for 300 billion tokens. GPT is a general purpose model that can be applied to many tasks instead of tailoring them for specific tasks. As such, it makes easy to control experiments and showcase the difference among various methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 280, + 291, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 280, + 291, + 428 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 291, + 428 + ], + "type": "text", + "content": "Implementation Details. Our base model uses absolute positional embeddings with maximum length 1024. To accommodate longer inputs, we duplicate its positional embeddings to have the maximum length as 8192 such that the model can handle sequences containing up to 8192 tokens. Then, we implement different attention methods by modifying the attention pattern of the base model. We implement all the models with PyTorch (Paszke et al., 2019). All the experiments are conducted on NVIDIA A100 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 430, + 291, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 430, + 291, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 291, + 579 + ], + "type": "text", + "content": "Continual Training Details. After changing the attention pattern, we conduct the continual training for MASFormer and baseline methods on PILE corpus (Gao et al., 2020) to adapt the revised models to new attention patterns and long-sequence inputs. We leverage the casual language modeling (CLM) objective to train the model for 50,000 steps with a warmup of 2000 steps. We set the input length as 8192 and use a batch size of 128 such that the models are optimized with 1M tokens per step. We use the constant learning 0.0001 for all methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 581, + 291, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 291, + 607 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 291, + 607 + ], + "type": "text", + "content": "Baseline. We compare MASFormer with the following methods:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 609, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 67, + 609, + 290, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 609, + 290, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 609, + 290, + 690 + ], + "type": "text", + "content": "- All full attention is to apply full attention at every layer. It has been adopted by most of existing transformer models as default. Although incurring the maximum attention cost, it achieves the best performance for most of our tasks. Hence, it acts as an upper bound for other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "- All block sparse attention is to apply block attention at every layer, which is an effective method to reduce computational cost when modeling long sequences. Block attention sets the attention span of each layer identical such that it evenly distributes the budget of attention computation across layers." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "type": "text", + "content": "- All sliding-window attention is to apply sliding-window attention at every layer, which is another variant of sparse attention. It shares the similar spirits and often performs similarly as block attention." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": "In the following experiments, we compare MAS-Former and the baseline methods across different attention cost " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": ". That is, for all block sparse attention, we set the block size as " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "b = \\mathcal{C} / (Ln)" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": ". For all sliding-window attention, we choose the window size as " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "w = \\mathcal{C} / (2Ln)" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": ". For MASFormer, we apply a small block size " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "m = 1024" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": " for its block attention and set " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "(\\mathcal{C} - Lmn) / (n^2 - mn)" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": ". Then we observe how their performance evolves when enhancing the attention cost " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": " or input length " + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 126, + 526, + 261 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 262, + 525, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 262, + 525, + 288 + ], + "spans": [ + { + "bbox": [ + 302, + 262, + 525, + 288 + ], + "type": "text", + "content": "Experiment Overview. We briefly summarize the experimental contents as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 290, + 527, + 534 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 302, + 290, + 525, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 290, + 525, + 330 + ], + "spans": [ + { + "bbox": [ + 302, + 290, + 525, + 330 + ], + "type": "text", + "content": "- Section 4.1 presents the perplexity evaluation of all the models on ArXiv and PubMed after continual training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 332, + 526, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 332, + 526, + 412 + ], + "spans": [ + { + "bbox": [ + 302, + 332, + 526, + 412 + ], + "type": "text", + "content": "- Section 4.2 compares the summarization performance of the models on ArXiv, QMSUM, and GovReport after fine-tuning. Besides, we also discuss the difference between perplexity and downstream evaluation in reflecting model capacity to capture long-range dependencies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 413, + 527, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 413, + 527, + 534 + ], + "spans": [ + { + "bbox": [ + 302, + 413, + 527, + 534 + ], + "type": "text", + "content": "- Section 4.3 provides three crucial analyses: (i) we evaluate the benefits of increasing input length and discuss the requirements to attain these gains; (ii) we analyze the effectiveness of continual training for long-sequence modeling; (iii) we conduct an ablation study to demonstrate that equipping bottom layers with full attention yields the most significant performance gains than other options. We further provide the explanations." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 546, + 464, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 546, + 464, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 546, + 464, + 560 + ], + "type": "text", + "content": "4.1 Natural Language Modeling" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 565, + 487, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 565, + 487, + 577 + ], + "spans": [ + { + "bbox": [ + 302, + 565, + 487, + 577 + ], + "type": "text", + "content": "4.1.1 Datasets and Evaluation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 583, + 526, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 583, + 526, + 703 + ], + "spans": [ + { + "bbox": [ + 302, + 583, + 526, + 703 + ], + "type": "text", + "content": "Datasets. We evaluate the perplexity of the updated GPT-2 for each attention method after continual training. The evaluation is conducted on test sets of ArXiv and PubMed (Cohan et al., 2018). Table 5 presents the statistics of these two datasets. Pubmed consists of scientific documents, with a document's content used as input and its corresponding abstract as the target summary. ArXiv is similar to PubMed, with documents from arXiv." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": "Evaluation Details. We conduct the perplexity evaluation under two settings. (i) We calculate the perplexity (ppl.) with all documents from test sets. Table 1 presents the overall perplexity of different models on two datasets. (ii) To showcase" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2779" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 128, + 57, + 290, + 180 + ], + "blocks": [ + { + "bbox": [ + 128, + 57, + 290, + 180 + ], + "lines": [ + { + "bbox": [ + 128, + 57, + 290, + 180 + ], + "spans": [ + { + "bbox": [ + 128, + 57, + 290, + 180 + ], + "type": "image", + "image_path": "aa87e531b990436c086b6417eaebac905aedebcd05996b5827a767377912c034.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 182, + 233, + 191 + ], + "lines": [ + { + "bbox": [ + 182, + 182, + 233, + 191 + ], + "spans": [ + { + "bbox": [ + 182, + 182, + 233, + 191 + ], + "type": "text", + "content": "(a) ArXiv ppl." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 57, + 465, + 180 + ], + "blocks": [ + { + "bbox": [ + 302, + 57, + 465, + 180 + ], + "lines": [ + { + "bbox": [ + 302, + 57, + 465, + 180 + ], + "spans": [ + { + "bbox": [ + 302, + 57, + 465, + 180 + ], + "type": "image", + "image_path": "ea1a52f7f7b59921579ce8383a2cd331d6c63cbfdf1ca31e5cfd12ab892255b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 181, + 412, + 191 + ], + "lines": [ + { + "bbox": [ + 352, + 181, + 412, + 191 + ], + "spans": [ + { + "bbox": [ + 352, + 181, + 412, + 191 + ], + "type": "text", + "content": "(b) PubMed ppl." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "lines": [ + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "text", + "content": "Figure 3: Perplexity evaluation on ArXiv and PubMed with examples of different length. Here " + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "text", + "content": "-axis is the maximum document length of each subset, i.e., " + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "inline_equation", + "content": "k \\times 1024" + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "inline_equation", + "content": "k = 1, 2, 3, \\ldots" + }, + { + "bbox": [ + 67, + 193, + 524, + 217 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 224, + 291, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 224, + 291, + 332 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 291, + 332 + ], + "type": "text", + "content": "the varying behaviors of models on documents of different length, we divide all documents into several subsets according to their length. Each subset consists of examples, whose length is within " + }, + { + "bbox": [ + 66, + 224, + 291, + 332 + ], + "type": "inline_equation", + "content": "((k - 1)\\times 1024,k\\times 1024](k = 1,2,3,\\ldots)" + }, + { + "bbox": [ + 66, + 224, + 291, + 332 + ], + "type": "text", + "content": ". Then, we evaluate the perplexity on each subset. Figure 3 presents the perplexity of models on different subsets of examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 339, + 138, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 339, + 138, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 339, + 138, + 350 + ], + "type": "text", + "content": "4.1.2 Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": "Table 1 compares the overall perplexity on test sets of ArXiv and PubMed. The results suggest that, with " + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "inline_equation", + "content": "l = 4" + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": " layers of full attention, MASFormer achieves comparable performance to all full attention, while reducing " + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": " of its attention cost. With the similar attention cost " + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": ", MASFormer outperforms all block attention that evenly distributes the budget of attention computation. For example, MASFormer with " + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "inline_equation", + "content": "l = 2" + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": " achieves 8.75 ppl. on PubMed, which is 1.37 lower than that of block attention of " + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "inline_equation", + "content": "b = 2048" + }, + { + "bbox": [ + 67, + 354, + 291, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 71, + 510, + 286, + 630 + ], + "blocks": [ + { + "bbox": [ + 71, + 510, + 286, + 630 + ], + "lines": [ + { + "bbox": [ + 71, + 510, + 286, + 630 + ], + "spans": [ + { + "bbox": [ + 71, + 510, + 286, + 630 + ], + "type": "table", + "html": "
MethodsCArXivPubMed
Full attention1,610M8.637.63
Block (b=1024)201M13.1912.19
Block (b=2048)402M10.7510.13
MASFormer (l=2)318M10.258.75
MASFormer (l=4)436M9.318.25
MASFormer (l=8)671M9.638.25
", + "image_path": "26632b80afb2e7994ce3689c38d4f0d3f887fcfc03e3083b578466dbd64ea0c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 633, + 287, + 645 + ], + "lines": [ + { + "bbox": [ + 69, + 633, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 633, + 287, + 645 + ], + "type": "text", + "content": "Table 1: Perplexity evaluation on ArXiv and PubMed." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "Figure 3 illustrates the perplexity variation of each method given examples of different length. We can tell that MASFormer and full attention show better performance on longer documents, suggesting increasing context length can improve their prediction performance. Full attention, though incurring the highest attention cost, always achieves the best performance due to its outstanding capability to handle sophisticated dependencies. Notably," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "spans": [ + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "text", + "content": " of its attention cost, MASFormer exhibits a curve of ppl. v.s. length that closely resembles to that of full attention. This demonstrates the effectiveness and efficiency of MASFormer to capture long-range dependencies. In contrast, block sparse attention benefits much less from long contexts and underperforms both of them because of its incapability to encode long-range signals. For example, when " + }, + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "inline_equation", + "content": "b = 1024" + }, + { + "bbox": [ + 302, + 224, + 526, + 360 + ], + "type": "text", + "content": ", block attention achieves similar perplexity on PubMed examples of different length." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 368, + 472, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 368, + 472, + 381 + ], + "spans": [ + { + "bbox": [ + 302, + 368, + 472, + 381 + ], + "type": "text", + "content": "4.2 Natural Language Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 386, + 477, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 386, + 477, + 399 + ], + "spans": [ + { + "bbox": [ + 302, + 386, + 477, + 399 + ], + "type": "text", + "content": "4.2.1 Datasets and Training Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 402, + 526, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 402, + 526, + 537 + ], + "spans": [ + { + "bbox": [ + 302, + 402, + 526, + 537 + ], + "type": "text", + "content": "Datasets. We evaluate the downstream performance of models on several abstractive summarization tasks to compare their capability of handling long sequences in practice. Specifically, we fine-tune models on ArXiv (Cohan et al., 2018), QM-SUM and GovReport (from SCROLLS benchmark, Shaham et al. (2022)). Their statistics are summarized in Table 5. We mainly use ROUGE-2 (R2) score (Lin, 2004) as the evaluation metric, which is more important and sensitive than R1 and RL." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 540, + 525, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 540, + 525, + 702 + ], + "spans": [ + { + "bbox": [ + 302, + 540, + 525, + 702 + ], + "type": "text", + "content": "Training Details. After continual training, we fine-tune each model and report R2 scores on validation sets. Specifically, we fine-tune models for 3000 steps on QMSUM, 8000 steps on GovReport, and 12000 steps on ArXiv. We set the batch size as 64 for ArXiv and 32 for QMSUM and GovReport. We pick the learning rates from " + }, + { + "bbox": [ + 302, + 540, + 525, + 702 + ], + "type": "inline_equation", + "content": "\\{1 \\times 10^{-5}, 5 \\times 10^{-5}, 1 \\times 10^{-4}, 5 \\times 10^{-4}\\}" + }, + { + "bbox": [ + 302, + 540, + 525, + 702 + ], + "type": "text", + "content": ", and choose the optimal ones to report the performance of each method. Moreover, the input length is fixed as 8192. We apply the greedy decoding for generation. Please see Appendix B for more details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 711, + 373, + 722 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 711, + 373, + 722 + ], + "spans": [ + { + "bbox": [ + 302, + 711, + 373, + 722 + ], + "type": "text", + "content": "4.2.2 Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 726, + 524, + 753 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 726, + 524, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 726, + 524, + 753 + ], + "type": "text", + "content": "In Table " + }, + { + "bbox": [ + 302, + 726, + 524, + 753 + ], + "type": "inline_equation", + "content": "2^{2}" + }, + { + "bbox": [ + 302, + 726, + 524, + 753 + ], + "type": "text", + "content": " and Figure 4, we present the fine-tuning results on QMSUM, ArXiv and GovReport across" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "type": "text", + "content": "2Please see Table 6 in Appendix B for all ROUGE scores." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2780" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 217, + 175 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 217, + 175 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 217, + 175 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 217, + 175 + ], + "type": "image", + "image_path": "7abd031293b0a106fc32b6b3eda383c7fe114cdd09263d0c7c29802ada1078f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 179, + 166, + 190 + ], + "lines": [ + { + "bbox": [ + 129, + 179, + 166, + 190 + ], + "spans": [ + { + "bbox": [ + 129, + 179, + 166, + 190 + ], + "type": "text", + "content": "(a) ArXiv" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 68, + 364, + 174 + ], + "blocks": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "lines": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "spans": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "type": "image", + "image_path": "06f3fa19b661a64eedbac59b93eed38cff3357da2c5afb8d0725d3b7d2bb4714.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 179, + 319, + 190 + ], + "lines": [ + { + "bbox": [ + 271, + 179, + 319, + 190 + ], + "spans": [ + { + "bbox": [ + 271, + 179, + 319, + 190 + ], + "type": "text", + "content": "(b) QMSUM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 375, + 69, + 512, + 175 + ], + "blocks": [ + { + "bbox": [ + 375, + 69, + 512, + 175 + ], + "lines": [ + { + "bbox": [ + 375, + 69, + 512, + 175 + ], + "spans": [ + { + "bbox": [ + 375, + 69, + 512, + 175 + ], + "type": "image", + "image_path": "9e37f4b3e56d841f515e8e7c046b7e5e493dd9309d4c550d1c24329e45ed393b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 179, + 470, + 190 + ], + "lines": [ + { + "bbox": [ + 416, + 179, + 470, + 190 + ], + "spans": [ + { + "bbox": [ + 416, + 179, + 470, + 190 + ], + "type": "text", + "content": "(c) GovReport" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 194, + 524, + 217 + ], + "lines": [ + { + "bbox": [ + 67, + 194, + 524, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 194, + 524, + 217 + ], + "type": "text", + "content": "Figure 4: Given input length as 8192, we compare summarization performance between MASFormer and block/full attention when increasing the attention cost." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 69, + 229, + 291, + 396 + ], + "blocks": [ + { + "bbox": [ + 69, + 229, + 291, + 396 + ], + "lines": [ + { + "bbox": [ + 69, + 229, + 291, + 396 + ], + "spans": [ + { + "bbox": [ + 69, + 229, + 291, + 396 + ], + "type": "table", + "html": "
MethodsCQMSUMArXivGovReport
Full attention1610M8.0019.3228.83
Window(w=1024)402M4.3213.5117.03
Block(b=2048)402M5.039.6112.31
MASFormer(l=4)436M6.5914.9118.82
Window(w=2048)805M5.0515.2122.79
Block(b=4096)805M5.1514.5023.64
MASFormer(l=6)553M7.1515.7221.20
MASFormer(l=8)671M7.4617.0024.42
MASFormer(l=12)906M8.7018.5826.26
", + "image_path": "bec286d385d911454869fb10e215deec0e4e7ee116298b8d8c88df70de43888b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 400, + 290, + 435 + ], + "lines": [ + { + "bbox": [ + 67, + 400, + 290, + 435 + ], + "spans": [ + { + "bbox": [ + 67, + 400, + 290, + 435 + ], + "type": "text", + "content": "Table 2: Summarization performance of models with different attention methods. The best results are shown in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 445, + 291, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 445, + 291, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 445, + 291, + 650 + ], + "type": "text", + "content": "different attention cost. The results demonstrate that, with the similar attention cost, MASFormer significantly outperforms sparse attention variants. Furthermore, when enhancing attention cost, MASFormer achieves greater performance gains than sparse attention methods. This is evident from the steeper slope of its R2 curve versus attention cost, in contrast to the baseline method. For example, when increasing " + }, + { + "bbox": [ + 67, + 445, + 291, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 67, + 445, + 291, + 650 + ], + "type": "text", + "content": " form 553M to 671M, the R2 score of MASFormer on QMSUM exhibits a substantial improvement, reaching 8.70 from 7.46. Remarkably, this score surpasses even that of full attention. Therefore, MASFormer addresses the trade-off between computational cost and performance gains in a more efficient and effective way." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "Notice that, in order to achieve comparable summarization performance to full attention, MAS-Former needs at least " + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "inline_equation", + "content": "l = 8" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": " layers of full attention, and providing more can lead to more gains. This observation is different from the findings in NLM (Figure 3) that increasing " + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": " beyond 4 provides limited improvement in perplexity. Their different capacity requirements arise from the fact that predicting next tokens in NLM primarily relies on lo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 225, + 526, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 225, + 526, + 279 + ], + "spans": [ + { + "bbox": [ + 302, + 225, + 526, + 279 + ], + "type": "text", + "content": "cal dependencies. Capturing infrequent long-range tokens does not significantly improve perplexity. Thus, this discrepancy emphasizes the necessity to evaluate long-range models on downstream tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 290, + 370, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 290, + 370, + 303 + ], + "spans": [ + { + "bbox": [ + 302, + 290, + 370, + 303 + ], + "type": "text", + "content": "4.3 Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 309, + 522, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 309, + 522, + 322 + ], + "spans": [ + { + "bbox": [ + 302, + 309, + 522, + 322 + ], + "type": "text", + "content": "4.3.1 Benefits of Increasing Sequence Length" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "spans": [ + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "text", + "content": "In this section, we investigate the benefits of increasing input length for downstream performance. Specifically, we select the input length from " + }, + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "inline_equation", + "content": "\\{2048, 4096, 6144, 8192\\}" + }, + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "text", + "content": " and present the fine-tuning performance of full attention in Figure 5. The results consistently demonstrate that as the input length increases, the model's performance improves. That is, downstream performance benefits significantly from long-sequence inputs. In contrast, increasing example length beyond " + }, + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "inline_equation", + "content": "6k" + }, + { + "bbox": [ + 301, + 327, + 526, + 501 + ], + "type": "text", + "content": " results in marginal improvements in perplexity (See Figure 3), highlighting again the importance of downstream evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 503, + 527, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 503, + 527, + 653 + ], + "spans": [ + { + "bbox": [ + 302, + 503, + 527, + 653 + ], + "type": "text", + "content": "In addition, when comparing the behaviors of block attention in Figure 2c and 2d, we find that sparse attention often insufficiently capitalize on the benefits offered by longer inputs. For instance, given block size as 4096, its performance on ArXiv remains nearly unchanged when increasing input length from 4096 (R2 = 15.52 in Figure 5a) to 8192 (R2 = 14.49 in Figure 2c). This finding suggests that enhancing input length can only improve model performance if the model possesses the sufficient capability to handle long-range dependencies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 661, + 504, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 661, + 504, + 674 + ], + "spans": [ + { + "bbox": [ + 302, + 661, + 504, + 674 + ], + "type": "text", + "content": "4.3.2 Effectiveness of Continual Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": "We analyze the effectiveness of continual training by comparing fine-tuning performance of MAS-Former " + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "inline_equation", + "content": "(l = 8)" + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": " under the following settings: (i) MASFormer without continual training (w.o. C.T.); (ii) MASFormer continually trained with short inputs (C.T. " + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "inline_equation", + "content": "(n = 2048)" + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": "); (iii) MASFormer continually trained with long inputs (C.T. " + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "inline_equation", + "content": "(n = 8192)" + }, + { + "bbox": [ + 302, + 678, + 526, + 772 + ], + "type": "text", + "content": "). Table 3" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "text", + "content": "2781" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 217, + 174 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 217, + 174 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 217, + 174 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 217, + 174 + ], + "type": "image", + "image_path": "34f8b43ff943ff02432a8efcea092d5101ce4681ebe8062ad2ef1f4556cd815e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 177, + 165, + 186 + ], + "lines": [ + { + "bbox": [ + 129, + 177, + 165, + 186 + ], + "spans": [ + { + "bbox": [ + 129, + 177, + 165, + 186 + ], + "type": "text", + "content": "(a) ArXiv" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 68, + 364, + 174 + ], + "blocks": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "lines": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "spans": [ + { + "bbox": [ + 227, + 68, + 364, + 174 + ], + "type": "image", + "image_path": "243cd9585f95f0cec8f02ae7e098f293e5b08b58846d6884995bfa11ff8a6731.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 177, + 322, + 187 + ], + "lines": [ + { + "bbox": [ + 268, + 177, + 322, + 187 + ], + "spans": [ + { + "bbox": [ + 268, + 177, + 322, + 187 + ], + "type": "text", + "content": "(b) GovReport" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 375, + 68, + 512, + 174 + ], + "blocks": [ + { + "bbox": [ + 375, + 68, + 512, + 174 + ], + "lines": [ + { + "bbox": [ + 375, + 68, + 512, + 174 + ], + "spans": [ + { + "bbox": [ + 375, + 68, + 512, + 174 + ], + "type": "image", + "image_path": "576c77e2d4168b14669caaa7ef1f13475630024a6b889195642c7a54823bebc1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 177, + 467, + 186 + ], + "lines": [ + { + "bbox": [ + 419, + 177, + 467, + 186 + ], + "spans": [ + { + "bbox": [ + 419, + 177, + 467, + 186 + ], + "type": "text", + "content": "(c) QMSUM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 135, + 188, + 456, + 201 + ], + "lines": [ + { + "bbox": [ + 135, + 188, + 456, + 201 + ], + "spans": [ + { + "bbox": [ + 135, + 188, + 456, + 201 + ], + "type": "text", + "content": "Figure 5: Fine-tuning performance of full attention under different input length." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 213, + 290, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 213, + 290, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 213, + 290, + 267 + ], + "type": "text", + "content": "presents fine-tuning performance of these models. We can tell that continual training with long inputs indeed facilitates the revised models to adapt to new structures and long-sequence inputs." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 69, + 279, + 295, + 348 + ], + "blocks": [ + { + "bbox": [ + 69, + 279, + 295, + 348 + ], + "lines": [ + { + "bbox": [ + 69, + 279, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 69, + 279, + 295, + 348 + ], + "type": "table", + "html": "
l=8QMSUMGovReport
w.o. C.T.29.33/6.43/25.7153.28/23.61/51.74
C.T. (n=2048)29.87/7.16/26.1552.28/23.01/49.83
C.T. (n=8192)30.91/7.45/27.0254.37/24.42/51.87
", + "image_path": "d0a3b1f47d0535c441bdce266d7577db64fd43903ed73a50a36c651c66a44a83.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 372, + 228, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 372, + 228, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 372, + 228, + 384 + ], + "type": "text", + "content": "4.3.3 Where to use full attention" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "text", + "content": "To answer where to apply full attention, we compare fine-tuning performance of MASFormers that apply full attention at (i) bottom layers; (ii) middle layers; (iii) top layers; (iv) every " + }, + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "inline_equation", + "content": "L / l" + }, + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "text", + "content": " layers. The results in Table 4 demonstrate that equipping bottom layers with full attention yields the best performance. This is because that long-range dependencies can be continually captured and reinforced by bottom layers before propagated to upper layers. As such, these long-range signals can be effectively incorporated into the upper layers with local attention, facilitating their encoding of local information. In contrast, when equipping local attention at bottom layers, long-range tokens are first aggregated with neighboring tokens by local attention, thereby weakening their long-range signals. Moreover, if alternating full and local attention every " + }, + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "inline_equation", + "content": "L / l" + }, + { + "bbox": [ + 67, + 389, + 291, + 645 + ], + "type": "text", + "content": " layers, the long-range signals cannot be continually reinforced nor efficiently captured." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 656, + 143, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 656, + 143, + 668 + ], + "spans": [ + { + "bbox": [ + 67, + 656, + 143, + 668 + ], + "type": "text", + "content": "5 Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 773 + ], + "type": "text", + "content": "GPT-Neo (Black et al., 2021) introduces an attention pattern that alternates full and window attention. However, this models is not tailored for long sequences. It sets the local window size as 256 and has the maximum input length as 2048, unable to handle long sequences. Instead, this attention pattern is applied heuristically in an attempt to re" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 304, + 216, + 523, + 332 + ], + "blocks": [ + { + "bbox": [ + 74, + 350, + 282, + 363 + ], + "lines": [ + { + "bbox": [ + 74, + 350, + 282, + 363 + ], + "spans": [ + { + "bbox": [ + 74, + 350, + 282, + 363 + ], + "type": "text", + "content": "Table 3: We report R1/R2/RL for the above results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 216, + 523, + 332 + ], + "lines": [ + { + "bbox": [ + 304, + 216, + 523, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 523, + 332 + ], + "type": "table", + "html": "
PositionQMSUMGovReport
Every 328.26/6.94/25.0326.16/12.37/24.82
Top 820.89/4.52/18.37-/-/-
Middle 827.27/5.99/24.0620.80/9.01/19.52
Bottom 830.91/7.45/27.0254.37/24.42/51.87
Every 231.27/8.19/27.4135.34/16.04/33.68
Bottom 1232.53/8.70/28.7556.98/26.26/54.46
", + "image_path": "80d5984b750b359aeabe979b9b3ae97bd888cd3191dfedb56389309389bd07d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 334, + 524, + 359 + ], + "lines": [ + { + "bbox": [ + 302, + 334, + 524, + 359 + ], + "spans": [ + { + "bbox": [ + 302, + 334, + 524, + 359 + ], + "type": "text", + "content": "Table 4: Performance comparison of MASFormers that apply full attention at different layers (# layers " + }, + { + "bbox": [ + 302, + 334, + 524, + 359 + ], + "type": "inline_equation", + "content": "L = 24" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 301, + 368, + 525, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 368, + 525, + 476 + ], + "spans": [ + { + "bbox": [ + 301, + 368, + 525, + 476 + ], + "type": "text", + "content": "duce computational cost. However, as discussed in Section 4.3.3, this approach is neither effective nor efficient as MASFormer when handling long sequences. As shown in Table 4, applying full attention at every 2 or 3 layers underperforms applying it at bottom 12 or 8 layers. Therefore, alternating between full and block attention results in additional computational cost and performance degradation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 476, + 525, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 525, + 639 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 525, + 639 + ], + "type": "text", + "content": "In contrast, MASFormer presents an effective solution for efficient long-sequence modeling. It provides guidance on adapting existing pre-trained transformers to long inputs. Meanwhile, it provides insights for designing large long-range models, especially for deeper models. By equipping only a subset of bottom layer with full attention, we can substantially mitigate computational cost. Additionally, the computation of MASFormer can be further optimized by leveraging system-level acceleration techniques (e.g., FlashAttention and xFormer) that support both block and full attention." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 649, + 381, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 649, + 381, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 649, + 381, + 661 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 526, + 772 + ], + "type": "text", + "content": "We propose an efficient long-range transformer - MASFormer that utilizes full attention at a few of bottom layers and employs sparse attention at the remaining layers. Our empirical results on natural language modeling and generation tasks demonstrate that MASFormer can address the trade-off between computational cost and performance gains in a more efficient and effective way." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "2782" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 127, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 89, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 89, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 89, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 69, + 89, + 289, + 145 + ], + "type": "text", + "content": "Joshua Ainslie, Santiago Ontanon, Chris Alberti, Va-clav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. 2020. Etc: Encoding long and structured inputs in transformers. arXiv preprint arXiv:2004.08483." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 153, + 289, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 153, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 289, + 186 + ], + "type": "text", + "content": "Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 194, + 289, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 194, + 289, + 238 + ], + "spans": [ + { + "bbox": [ + 69, + 194, + 289, + 238 + ], + "type": "text", + "content": "Sid Black, Leo Gao, Phil Wang, Connor Leahy, and Stella Biderman. 2021. GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 246, + 289, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 246, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 69, + 246, + 289, + 411 + ], + "type": "text", + "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 419, + 289, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 419, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 69, + 419, + 289, + 464 + ], + "type": "text", + "content": "Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Re. 2021. Scatterbrain: Unifying sparse and low-rank attention approximation. arXiv preprint arXiv:2110.15343." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 472, + 289, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 472, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 69, + 472, + 289, + 582 + ], + "type": "text", + "content": "Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian. 2018. A discourse-aware attention model for abstractive summarization of long documents. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 615-621, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 590, + 289, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 590, + 289, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 590, + 289, + 645 + ], + "type": "text", + "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 653, + 289, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 289, + 687 + ], + "type": "text", + "content": "Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 95 + ], + "type": "text", + "content": "4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 102, + 524, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 524, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 524, + 179 + ], + "type": "text", + "content": "Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 186, + 524, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 524, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 524, + 243 + ], + "type": "text", + "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, et al. 2020. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 249, + 524, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 524, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 524, + 304 + ], + "type": "text", + "content": "Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. 2020. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 312, + 524, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 524, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 524, + 346 + ], + "type": "text", + "content": "Albert Gu, Karan Goel, and Christopher Ré. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 354, + 524, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 524, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 524, + 398 + ], + "type": "text", + "content": "Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, and Yinfei Yang. 2021. Longt5: Efficient text-to-text transformer for long sequences. arXiv preprint arXiv:2112.07916." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 405, + 524, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 524, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 524, + 449 + ], + "type": "text", + "content": "Ankit Gupta, Albert Gu, and Jonathan Berant. 2022. Diagonal state spaces are as effective as structured state spaces. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 457, + 524, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 524, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 524, + 502 + ], + "type": "text", + "content": "Pengcheng He, Jianfeng Gao, and Weizhu Chen. 2021a. Debertav3: Improving deberta using electra-style pretraining with gradient-disentangled embedding sharing. arXiv preprint arXiv:2111.09543." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 509, + 524, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 524, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 524, + 554 + ], + "type": "text", + "content": "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021b. Deberta: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 560, + 524, + 595 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 524, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 524, + 595 + ], + "type": "text", + "content": "Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer. In International Conference on Learning Representations." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 602, + 524, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 524, + 679 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 524, + 679 + ], + "type": "text", + "content": "Benjamin Lefaudeau, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, and Daniel Haziza. 2022. xformers: A modular and hackable transformer modelling library. https://github.com/facebookresearch/xformers." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 687, + 524, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 687, + 524, + 731 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 524, + 731 + ], + "type": "text", + "content": "Yuhong Li, Tianle Cai, Yi Zhang, Deming Chen, and Debadepta Dey. 2022. What makes convolutional models great on long sequence modeling? arXiv preprint arXiv:2210.09298." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 738, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 738, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 738, + 524, + 772 + ], + "type": "text", + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2783" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 128 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 135, + 290, + 190 + ], + "type": "text", + "content": "Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. 2023. Mega: Moving average equipped gated attention. In *The Eleventh International Conference on Learning Representations*." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 198, + 290, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 198, + 290, + 340 + ], + "spans": [ + { + "bbox": [ + 69, + 198, + 290, + 340 + ], + "type": "text", + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 8024-8035." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 349, + 290, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 349, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 69, + 349, + 290, + 394 + ], + "type": "text", + "content": "Hao Peng, Nikolaos Pappas, Dani Yogatama, Roy Schwartz, Noah Smith, and Lingpeng Kong. 2021. Random feature attention. In International Conference on Learning Representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 401, + 290, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 401, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 69, + 401, + 290, + 446 + ], + "type": "text", + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 454, + 290, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 454, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 454, + 290, + 541 + ], + "type": "text", + "content": "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Eric Michael Smith, Y-Lan Boureau, and Jason Weston. 2021. Recipes for building an open-domain chatbot. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 300-325, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 69, + 550, + 290, + 605 + ], + "type": "text", + "content": "Uri Shaham, Elad Segal, Maor Ivgi, Avia Efrat, Ori Yoran, Adi Haviv, Ankit Gupta, Wenhan Xiong, Mor Geva, Jonathan Berant, et al. 2022. Scrols: Standardized comparison over long language sequences. arXiv preprint arXiv:2201.03533." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 613, + 290, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 613, + 290, + 657 + ], + "spans": [ + { + "bbox": [ + 69, + 613, + 290, + 657 + ], + "type": "text", + "content": "Jimmy T.H. Smith, Andrew Warrington, and Scott Linderman. 2023. Simplified state space layers for sequence modeling. In *The Eleventh International Conference on Learning Representations*." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 665, + 290, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 665, + 290, + 709 + ], + "spans": [ + { + "bbox": [ + 69, + 665, + 290, + 709 + ], + "type": "text", + "content": "Simeng Sun, Kalpesh Krishna, Andrew Mattarella-Micke, and Mohit Iyyer. 2021. Do long-range language models actually use long-range context? arXiv preprint arXiv:2109.09115." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 306 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "type": "text", + "content": "Sinong Wang, Belinda Z Li, Madian Khabsa, Han Fang, and Hao Ma. 2020. Linformer: Self-attention with linear complexity. arXiv preprint arXiv:2006.04768." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 114, + 525, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 525, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 525, + 180 + ], + "type": "text", + "content": "Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. 2020. Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33:17283-17297." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 189, + 525, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 189, + 525, + 244 + ], + "spans": [ + { + "bbox": [ + 304, + 189, + 525, + 244 + ], + "type": "text", + "content": "Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. 2023. Adaptive budget allocation for parameter-efficient fine-tuning. In The Eleventh International Conference on Learning Representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 252, + 525, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 525, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 525, + 306 + ], + "type": "text", + "content": "Simiao Zuo, Xiaodong Liu, Jian Jiao, Denis Charles, Eren Manavoglu, Tuo Zhao, and Jianfeng Gao. 2022. Efficient long sequence modeling via state space augmented transformer. arXiv preprint arXiv:2212.08136." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2784" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 180, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 180, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 180, + 84 + ], + "type": "text", + "content": "A Dataset Statistics" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 525, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 525, + 118 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 525, + 118 + ], + "type": "text", + "content": "In the following table, we provide the detailed statistics of datasets in our experiments, including example splits and length statistics." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 119, + 125, + 474, + 227 + ], + "blocks": [ + { + "bbox": [ + 119, + 125, + 474, + 227 + ], + "lines": [ + { + "bbox": [ + 119, + 125, + 474, + 227 + ], + "spans": [ + { + "bbox": [ + 119, + 125, + 474, + 227 + ], + "type": "table", + "html": "
DatasetExample CountInput Length
TrainValidTestAverageMedian90th percentile
ArXiv203,0376,4366,44010,7208,51920,170
PubMed119,9246,6336,6584,7483,8838,883
QMSUM1,2572722819,49714,19727,761
GovReport17,4579729737,8868,84118,835
", + "image_path": "688c9a73142a43430e42bbc86a2f8ff6d2437c4538e40f8c8650a7f1a8794af0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 255, + 246, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 255, + 246, + 269 + ], + "spans": [ + { + "bbox": [ + 67, + 255, + 246, + 269 + ], + "type": "text", + "content": "B Natural Language Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 276, + 256, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 276, + 256, + 289 + ], + "spans": [ + { + "bbox": [ + 68, + 276, + 256, + 289 + ], + "type": "text", + "content": "B.1 The Results of All ROUGE Scores" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 73, + 301, + 522, + 468 + ], + "blocks": [ + { + "bbox": [ + 108, + 232, + 483, + 245 + ], + "lines": [ + { + "bbox": [ + 108, + 232, + 483, + 245 + ], + "spans": [ + { + "bbox": [ + 108, + 232, + 483, + 245 + ], + "type": "text", + "content": "Table 5: Statistics of datasets. Input length measured in tokens using a SentencePiece Model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 301, + 522, + 468 + ], + "lines": [ + { + "bbox": [ + 73, + 301, + 522, + 468 + ], + "spans": [ + { + "bbox": [ + 73, + 301, + 522, + 468 + ], + "type": "table", + "html": "
MethodsCQMSUMArXivGovReport
Full attention1610M31.50 / 8.00 / 27.8146.13 / 19.32 / 41.8960.53 / 28.83 / 57.88
Window (w=1024)402M23.31 / 4.32 / 20.6235.90 / 13.51 / 32.1949.82 / 17.03 / 47.42
Window (w=2048)805M26.73 / 5.05 / 23.4038.74 / 15.21 / 34.8756.14 / 22.79 / 53.50
Block (b=2048)402M26.24 / 5.03 / 23.1321.85 / 9.61 / 19.8626.37 / 12.31 / 25.18
Block (b=4096)805M26.96 / 5.15 / 23.8535.95 / 14.50 / 32.3749.83 / 23.64 / 47.50
MASFormer (l=4)436M29.86 / 6.59 / 25.8738.85 / 14.91 / 34.9846.67 / 18.82 / 44.39
MASFormer (l=6)553M30.83 / 7.15 / 27.1236.29 / 15.72 / 32.9649.26 / 21.20 / 46.89
MASFormer (l=8)671M30.91 / 8.00 / 27.8143.31 / 17.00 / 39.1254.37 / 24.42 / 51.87
MASFormer (l=12)906M32.53 / 8.70 / 28.7545.19 / 18.58 / 40.7256.98 / 26.26 / 54.46
", + "image_path": "0835269c3903855cb3d82f836026867f2603cfd3251e27d638513f881b722650.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 517, + 173, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 517, + 173, + 530 + ], + "spans": [ + { + "bbox": [ + 68, + 517, + 173, + 530 + ], + "type": "text", + "content": "B.2 Training Details" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 144, + 543, + 449, + 694 + ], + "blocks": [ + { + "bbox": [ + 167, + 476, + 425, + 488 + ], + "lines": [ + { + "bbox": [ + 167, + 476, + 425, + 488 + ], + "spans": [ + { + "bbox": [ + 167, + 476, + 425, + 488 + ], + "type": "text", + "content": "Table 6: Finetuning performance of different attention methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 144, + 543, + 449, + 694 + ], + "lines": [ + { + "bbox": [ + 144, + 543, + 449, + 694 + ], + "spans": [ + { + "bbox": [ + 144, + 543, + 449, + 694 + ], + "type": "table", + "html": "
MethodsQMSUMArXivGovReport
Full attention1 × 10-51 × 10-41 × 10-4
Window attention (w=1024)5 × 10-45 × 10-55 × 10-4
Window attention (w=2048)5 × 10-55 × 10-51 × 10-4
Block attention (w=2048)1 × 10-41 × 10-51 × 10-5
Block attention (w=4096)5 × 10-55 × 10-41 × 10-5
MASFormer (l=4)5 × 10-51 × 10-35 × 10-4
MASFormer (l=6)5 × 10-55 × 10-55 × 10-4
MASFormer (l=8)5 × 10-55 × 10-45 × 10-4
MASFormer (l=12)1 × 10-51 × 10-41 × 10-4
", + "image_path": "970356a9f960a9b5a9d13af0c2587d16a39d5b7dd240f8a0c8e97957a5bb66f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 154, + 702, + 437, + 714 + ], + "lines": [ + { + "bbox": [ + 154, + 702, + 437, + 714 + ], + "spans": [ + { + "bbox": [ + 154, + 702, + 437, + 714 + ], + "type": "text", + "content": "Table 7: The fine-tuning learning rate of each method on each dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 733, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 525, + 773 + ], + "type": "text", + "content": "We conduct continual training for all attention methods with training data form PILE and input length as 8192. After continual training, we obtain the continually trained models for each method and fine-tune them on QMSUM, ArXiv and GovReport to compare their summarization performance. During the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "2785" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 526, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 526, + 126 + ], + "type": "text", + "content": "fine-tuning, we set the input length as 8192 for all datasets and all models. We apply the greedy decoding for generation and set the maximum output length as 256 for QMSUM, 1024 for GovReport, and 512 for ArXiv. Table 8 lists the details of these hyperparameters. Besides, we apply the linear learning rate schedule to fine-tune the models and the base learning rates are summarized in Table 7." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 152, + 135, + 442, + 230 + ], + "blocks": [ + { + "bbox": [ + 152, + 135, + 442, + 230 + ], + "lines": [ + { + "bbox": [ + 152, + 135, + 442, + 230 + ], + "spans": [ + { + "bbox": [ + 152, + 135, + 442, + 230 + ], + "type": "table", + "html": "
HyperparameterQMSUMArXivGovReport
Training steps3000120008000
Batch size323264
Input length819281928192
Maximum generation length2565121024
Weight decay0.0010.0010.001
", + "image_path": "ef60230bc5c59c64126a05152ac0512baa6d553e79b62d8a95f8de9c214edb6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 239, + 499, + 252 + ], + "lines": [ + { + "bbox": [ + 94, + 239, + 499, + 252 + ], + "spans": [ + { + "bbox": [ + 94, + 239, + 499, + 252 + ], + "type": "text", + "content": "Table 8: The other fine-tuning parameters for each dataset, which remain the same for every method." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "2786" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_content_list.json b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa504f20a9ab429cb7423f7a9eca119c7d609915 --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_content_list.json @@ -0,0 +1,1895 @@ +[ + { + "type": "text", + "text": "Efficient Multilingual Language Model Compression through Vocabulary Trimming", + "text_level": 1, + "bbox": [ + 226, + 89, + 771, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Asahi Ushio and Yi Zhou and Jose Camacho-Collados", + "bbox": [ + 258, + 155, + 739, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cardiff NLP, School of Computer Science and Informatics, Cardiff University, UK", + "bbox": [ + 164, + 171, + 835, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{UshioA,Zhouy131,CamachoColladosJ}\\@cardiff.ac.uk", + "bbox": [ + 257, + 189, + 744, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multilingual language models (LMs) have become a powerful tool in NLP, especially for non-English languages. Nevertheless, model parameters of multilingual LMs remain large due to the larger embedding matrix of the vocabulary covering tokens in different languages. Instead, monolingual LMs can be trained in a target language with the language-specific vocabulary only. In this paper, we propose vocabulary-trimming (VT), a method to reduce a multilingual LM vocabulary to a target language by deleting potentially irrelevant tokens from its vocabulary. In theory, VT can compress any existing multilingual LM to any language covered by the original model. In our experiments, we show that VT can retain the original performance of the multilingual LM, while being considerably smaller in size than the original multilingual LM. The evaluation is performed over four NLP tasks (two generative and two classification tasks) among four widely used multilingual LMs in seven languages. The results show that this methodology can keep the best of both monolingual and multilingual worlds by keeping a small size as monolingual models without the need for specifically retraining them, and can even help limit potentially harmful social biases.", + "bbox": [ + 144, + 277, + 460, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 684, + 258, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multilingual language model (LM) pre-training (Devlin et al., 2019; Conneau et al., 2019; Liu et al., 2020; Xue et al., 2021) has been shown to be an efficient mechanism to store information from many languages into a single model, without the need for training multiple language-specific models. Moreover, it has been proven reliable for cross-lingual tasks (Pires et al., 2019; Conneau and Lample, 2019) and can provide competitive performance in most settings, generally similar to its monolingual counterparts (Goyal et al., 2021), while being generally less affected by culturally-dependant biases (Ahn and Oh, 2021). Similarly to monolingual", + "bbox": [ + 112, + 709, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "models, multilingual LMs can be used for zero/few-shot learning (Scao et al., 2022) by increasing the model size and, more frequently, can be specialized to different tasks by fine-tuning to specific data. In practice, however, there are a few practical issues when training multilingual LM such as the curse of multilinguality (Conneau et al., 2019; Pfeiffer et al., 2022), a trade-off between the number of languages and individual performance in a single language, or the multilingual vocabulary construction, which requires a careful design for better generalization (Chung et al., 2020; Zheng et al., 2021; Liang et al., 2023).", + "bbox": [ + 507, + 252, + 884, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Besides such generalization concerns, multilingual LMs usually consist of larger parameters than their monolingual counterparts due to the need for a large vocabulary covering multiple languages. This becomes an important issue in practice when the resources to host models are limited. For instance, while using the same configuration (i.e., same number of layers and hidden units), the parameter size of T5SMALL (Raffel et al., 2020) and mT5SMALL (Xue et al., 2021) are 140M and 300M, respectively. This is only due to their difference in vocabulary size, with T5 being 50k and mT5, 250k. In fact, the embedding matrix stemming from the LM vocabulary can occupy a large portion of the parameter space. For instance, the ratio of the embedding matrix to the full model's parameter size in multilingual LMs can be higher than 80% as T5 (see Figure 1).", + "bbox": [ + 507, + 462, + 884, + 751 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a simple vocabulary trimming (VT) method to remove tokens from the vocabulary of multilingual LMs that may be irrelevant to the target language. This is achieved by automatically identifying language-specific tokens from an underlying text corpus. We consider two VT strategies of pre-FT VT (VT before fine-tuning) and post-FT VT (VT after fine-tuning) and", + "bbox": [ + 507, + 753, + 882, + 881 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We release a Python library for VT at https://github.com/asahi417/lm-vocab-trimmer.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14725", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14725-14739 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 208, + 945, + 786, + 972 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f053d4a449685b0f6f10357a3b054e984dccd01320158524f8cedf511529135b.jpg", + "image_caption": [ + "Figure 1: The ratio of the embedding matrix to the number of parameters for each multilingual LM." + ], + "image_footnote": [], + "bbox": [ + 147, + 84, + 467, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "analyse them by varying the final vocabulary size. We conduct experiments on two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI), across seven different languages. The experimental results show that both pre and post fine-tuning VT can reduce the model size while retaining the original performance in generation tasks (QA and QG), and particularly in classification tasks (sentiment and NLI) where the results are close to being identical despite the significant reduction in vocabulary size. In all tasks, the original performance can be generally maintained with less than $40\\%$ of the full model parameters for all languages.", + "bbox": [ + 112, + 300, + 489, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, even though pre-trained LMs have reported impressive performance on various NLP downstream tasks (Kenton and Toutanova, 2019; Liu et al., 2019; Conneau et al., 2019), such LMs also demonstrate worrying levels of social biases in certain situations (May et al., 2019; Kurita et al., 2019; Kaneko and Bollegala, 2021). One natural question that arises is whether VT can have an influence on the bias level in multilingual LMs, including fine-tuned models. For this purpose, we evaluate social bias in multilingual LMs after applying VT with different settings and compare it against its monolingual counterpart. Experimental results show that the monolingual LM tends to contain more bias than its multilingual versions. Moreover, compared to the original multilingual LM, the bias level has no significant change after applying VT. These results suggest that a monolingual LM can be induced by applying VT to its corresponding multilingual LM, thereby obtaining a less biased monolingual LM compared to its original monolingual counterpart.", + "bbox": [ + 112, + 565, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 509, + 83, + 665, + 98 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several studies have explored the possibility to modify or adapt the vocabulary of LMs. For instance, Artetxe et al. (2020) and Marchisio et al. (2022) adapted a mono-lingual LM into another language by learning the embedding matrix on the new language, while fixing the other weights. Similarly, Wang et al. (2019) augmented the vocabulary of a multilingual LM to new languages with multilingual word alignment (Lample et al., 2018). Zheng et al. (2021) proposed to evaluate the ability of a vocabulary to represent a particular language, and Chung et al. (2020) proposed a multilingual vocabulary construction that balances the tradeoff between optimizing for cross-lingual sub-word sharing and the need for robust representation of individual languages. XLM-V (Liang et al., 2023) combines the idea of Zheng et al. (2021) and Chung et al. (2020) to efficiently enlarge the vocabulary size along with the model size scaling. Ostendorff and Rehm (2023) used a multi-stage fine-tuning to obtain a LM in the target language from other LM in the source language. These prior works modify existing mono/multi-lingual LMs to include new languages, i.e. augmenting the multilinguality of the LMs. In contrast, our study focuses on compressing multilingual LMs into the target language to effectively achieve smaller monolingual LMs, i.e. reducing the multilingual representation of the LMs while retaining the capability in a specific target language.", + "bbox": [ + 507, + 109, + 884, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The work of Abdaoui et al. (2020) is the most relevant to our study as, to the best of our knowledge, they introduced the idea of VT for the first time. However, their analysis is limited to NLI with pre-fine-tuning VT with mBERT (Devlin et al., 2019) only, as well as a fixed vocabulary size after VT. In contrast, our study compares two VT strategies, before and after fine-tuning, and show how this latter strategy, not considered in Abdaoui et al. (2020), can be a more effective compression technique in some settings. Furthermore, we extend the experiments to generation tasks as well as classification tasks with more recent LMs such as mBART (Lewis et al., 2020) and mT5 (Xue et al., 2021), and provide an exhaustive analysis on the effect of VT.", + "bbox": [ + 507, + 592, + 884, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Vocabulary Trimming", + "text_level": 1, + "bbox": [ + 507, + 860, + 736, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To perform vocabulary trimming (VT), we first need a multilingual LM as an input. The idea is to", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14726", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4fc018111e4b3ac5afd7840e323af32af5effe1b4e5969e976b76d85a306414b.jpg", + "image_caption": [ + "Figure 2: An illustration of vocabulary trimming for Korean and French." + ], + "image_footnote": [], + "bbox": [ + 154, + 84, + 448, + 269 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": " tailor model to a particular target language $l$ , which in principle belong to the same set of languages used to trained the input multilingual LM. For the target language $l$ , VT first identifies language-specific tokens on a language-specific corpus $\\mathcal{C}_l$ and remove all the tokens along with their embeddings except for those appeared in $\\mathcal{C}_l$ as described in Figure 2. In our analysis (§ 5), we also consider to keep the top- $n$ most frequent tokens in $\\mathcal{C}_l$ to further reduce the model size by removing less frequent tokens. We consider two VT strategies: (1) before fine-tuning and (2) after fine-tuning.", + "bbox": [ + 112, + 336, + 489, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The difference between these two strategies is whether to perform VT before or after fine-tuning, as shown in Figure 3. Both VTs have advantages and drawbacks: while pre-FT VT can reduce the time of fine-tuning as the trimmed LM is smaller than the original LM, post-FT VT only need a fine-tuned multilingual LM - this way, post-FT VT can be used as a postprocessing step and no additional language-specific training is required.", + "bbox": [ + 112, + 531, + 489, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we release a simple LM vocabulary trimming starting package to apply our proposed technique to any input multilingual transformer-based LM, along with all the models and code needed to reproduce our experiments, at https://github.com/asahi417/lm-vocab-trimmer.", + "bbox": [ + 112, + 677, + 489, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Evaluation", + "text_level": 1, + "bbox": [ + 112, + 790, + 243, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our experimental results to test the reliability of our proposed VT methodology in NLP tasks.", + "bbox": [ + 112, + 818, + 489, + 866 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/91d2550dceac5f3d39f6a60a1cbe3e06c1edc8e29abf72d40849b39523663876.jpg", + "image_caption": [ + "Figure 3: Comparisons of Pre-FT vs Post-FT VT in an example of fine-tuning on a task in French." + ], + "image_footnote": [], + "bbox": [ + 547, + 82, + 843, + 149 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Experimental Setting", + "text_level": 1, + "bbox": [ + 507, + 218, + 722, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tasks and datasets. In order to test the efficacy of VT, we consider two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI). As the datasets for QA, we use SQuAD (Rajpurkar et al., 2016) (English), Spanish SQuAD (Casimiro Pio et al., 2019) (Spanish), FQuAD (d'Hoffschmidt et al., 2020) (French), Italian SQuAD (Croce et al., 2018) (Italian), JAQuAD (So et al., 2022) (Japanese), KorQuAD (Lim et al., 2019) (Korean), and SberQuAD (Efimov et al., 2020) (Russian). For QG, we use the same datasets adapted for QG via QG-Bench (Ushio et al., 2022). For sentiment analysis, we use Twitter-based datasets for English (Rosenthal et al., 2017), Arabic (Rosenthal et al., 2017), French (Benamara et al., 2017), Italian (Barbieri et al., 2016), German (Cieliebak et al., 2017), Portuguese (Brum and Nunes, 2017), and Spanish (Diaz-Galiano et al., 2018) from UMSAB (Unified Multilingual Sentiment Analysis Benchmark) (Barbieri et al., 2022). All the sentiment analysis datasets contain three labels: positive, neutral and negative. For NLI, we use XNLI (Conneau et al., 2018), a multilingual NLI dataset, including English, French, German, Spanish and Arabic, which are the languages included in the sentiment analysis experiment. We fine-tune LMs on the training sets of each language, which were translated automatically from English and released in the original paper.", + "bbox": [ + 507, + 243, + 884, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation metrics. For the evaluation, we use the following standard metrics: answer span F1 score (Ans-F1) and exact match (EM) are used for QA; METEOR (MTR) and BERTScore (BS) for QG, which have been shown to be the most correlated metrics to human judgment (Ushio et al., 2022); macro-F1 score for sentiment following (Barbieri et al., 2022); and accuracy for NLI. As the language-specific corpus $\\mathcal{C}_l$ to extract vocabulary counts for VT, we use mC4 (Xue et al., 2021), one of the largest public multilingual corpora.", + "bbox": [ + 507, + 741, + 884, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2In theory, vocabulary trimming could be applied to any language model, even monolingual, but this analysis is out of the scope of this paper.", + "bbox": [ + 112, + 879, + 487, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14727", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/07fec4db6a433d05003569c769ee5b40771c2d84cb2a8726b46a8aa9aa682381.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Lang.VocabularyParametersQAQG
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
mT5EN209K (83.6%)258M (86.1%)70.1 / 55.570.2 / 55.570.1 / 56.423.8 / 90.023.8 / 90.024.0 / 90.1
ES131K (52.4%)178M (59.4%)55.9 / 34.755.9 / 34.757.8 / 37.522.7 / 84.122.7 / 84.122.3 / 84.2
FR131K (52.4%)178M (59.4%)50.0 / 30.950.0 / 30.948.6 / 29.417.5 / 80.717.5 / 80.716.1 / 79.2
IT111K (44.4%)157M (52.6%)53.2 / 37.653.4 / 37.851.5 / 36.017.6 / 80.817.6 / 80.817.5 / 80.6
JA125K (50.0%)172M (57.6%)65.7 / 65.765.7 / 65.763.0 / 63.029.0 / 80.929.0 / 80.928.6 / 81.0
KO73K (29.2%)119M (39.7%)77.1 / 70.677.1 / 70.574.5 / 67.327.5 / 82.927.5 / 83.028.0 / 83.7
RU147K (58.8%)195M (65.1%)73.7 / 51.473.8 / 51.474.8 / 53.426.4 / 84.326.4 / 84.328.9 / 86.4
mBARTEN173K (69.2%)532M (87.1%)76.9 / 62.677.0 / 62.778.4 / 65.725.1 / 90.425.1 / 90.424.7 / 90.1
ES87K (34.8%)443M (72.7%)64.1 / 42.264.5 / 42.863.7 / 43.922.9 / 83.622.8 / 83.622.8 / 84.0
FR85K (34.0%)442M (72.5%)60.4 / 39.361.0 / 39.866.4 / 45.119.8 / 81.719.8 / 81.718.4 / 79.7
IT67K (26.8%)424M (69.5%)64.7 / 50.064.9 / 50.265.8 / 49.818.0 / 80.617.9 / 80.718.9 / 81.1
JA77K (30.8%)434M (71.1%)68.2 / 68.268.2 / 68.270.6 / 70.630.0 / 82.329.7 / 82.129.1 / 80.8
KO46K (18.4%)402M (65.9%)79.3 / 72.379.2 / 72.183.2 / 77.330.2 / 83.930.3 / 84.030.2 / 83.8
RU99K (39.6%)456M (74.8%)78.7 / 58.079.0 / 58.275.5 / 49.929.3 / 87.228.7 / 87.028.3 / 86.7
", + "bbox": [ + 119, + 80, + 878, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Results on QA (Ans-F1/EM) and QG (MTR/BS), including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the parameter size of the original mT5 and mBART (No-Trim) is 300M and 611M, respectively, both with a vocabulary size of 250K.", + "bbox": [ + 112, + 316, + 882, + 374 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Base language models. As the base LMs, given computational constraints we chose the smallest mT5 and mBART to fine-tune on QA and QG, and XLM-R and XLM-V (Liang et al., 2023) for sentiment analysis and NLI. All these models have a vocabulary size of $250\\mathrm{K}$ , except for XLM-V which has a vocabulary size of $901\\mathrm{K}$ subword tokens. For our experiments, we compare the results of pre/post-FT VT against vanilla LM fine-tuning without VT, which we refer to as No-Trim.", + "bbox": [ + 112, + 399, + 487, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tuning. For model fine-tuning, we rely on lmqg (Ushio et al., 2023) for QA/QG, and Ray Tune3 for sentiment analysis. In both cases, we use the default search space for hyperparameter search. For NLI, we follow the same hyperparameters used in Liang et al. (2023). All the resulting models and code will be released upon acceptance of the paper.", + "bbox": [ + 112, + 571, + 487, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Results", + "text_level": 1, + "bbox": [ + 112, + 696, + 218, + 711 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present the results for the generation and classification tasks in section 4.2.1 and section 4.2.2, respectively.", + "bbox": [ + 112, + 718, + 487, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Generation Tasks: QA & QG", + "text_level": 1, + "bbox": [ + 112, + 777, + 403, + 793 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows the overall results on QA and QG. The results confirm that both of pre/post-FT VT can maintain the original performance in most cases, while being smaller than the original models by significantly reducing the vocabulary size. First, post-FT VT achieves at least the same performance as", + "bbox": [ + 112, + 797, + 487, + 892 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the vanilla fine-tuning for all the languages for both LMs in QA and QG, except for a few cases such as mBART QA in Korean and mBART QG in Russian, although the decrease is no more than $0.5\\%$ . Meanwhile, pre-FT VT outperforms its vanilla fine-tuning model with a relatively important margin in some cases, such as mBART French QA and mT5 Spanish QA. In contrast, there are a few models where pre-FT VT degrades the performance of the original model such as mT5 QA in Korean (2.6% decrease in Ans-F1) or mBART QA in Russian (3.2% decrease in Ans-F1).", + "bbox": [ + 505, + 399, + 882, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we keep all the vocabulary that appeared in the language-specific corpus $\\mathcal{C}_l$ , the percentage of reduced parameter depends on the language, and generally VT can reduce the model size for Asian (Japanese/Korean) and European (Spanish/French/Italian) languages efficiently (50% for mT5 and 70% for mBART), but it remains high in other languages (English-Russian).", + "bbox": [ + 507, + 594, + 882, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2 Classification Tasks: Sentiment & NLI", + "text_level": 1, + "bbox": [ + 507, + 736, + 872, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 shows the results on sentiment analysis and NLI. In this case, post-FT VT can robustly preserve the original performance of the original No-Trim baseline in both tasks for XLM-R and XLM-V, while being no more than $40\\%$ and $60\\%$ in vocabulary and overall parameter size, respectively, of the original XLM-V and XLM-R models in all the non-English datasets. XLM-V PT sentiment model is the only post-FT VT where a slight decrease can be observed $(0.1\\%)$ . On the other hand, the accuracy", + "bbox": [ + 507, + 758, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3https://docs.ray.io/en/latest/tune/index.html", + "bbox": [ + 134, + 903, + 485, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14728", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a62350e762289aa4a4e077c0684f5ce0a08f0b58c8c5d63d47c088735e4b66d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Lang.VocabularyParameterSentimentNLI
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
XLM-RAR49K (19.6%)124M (44.7%)66.366.360.975.775.773.8
DE91K (36.4%)156M (56.3%)73.273.373.579.979.978.3
EN173K (69.2%)219M (78.7%)68.468.567.984.684.670.6
ES87K (34.8%)153M (55.0%)69.069.065.079.879.867.2
FR85K (34.0%)151M (54.6%)71.871.872.180.180.179.6
IT67K (26.8%)138M (49.7%)62.962.970.8---
PT66K (26.4%)137M (49.3%)70.770.870.2---
XLM-VAR92K (11.8%)157M (20.2%)59.859.864.775.575.676.1
DE239K (30.7%)269M (34.7%)73.573.573.078.978.979.0
EN484K (62.2%)458M (58.9%)63.963.961.384.484.484.5
ES243K (31.2%)279M (35.1%)60.760.766.680.780.780.6
FR218K (28.0%)253M (32.6%)68.868.859.578.678.679.0
IT184K (23.7%)227M (29.3%)70.270.274.2---
PT181K (23.3%)225M (28.9%)66.666.552.8---
", + "bbox": [ + 149, + 80, + 845, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2: Results of sentiment analysis (macro F1) and XNLI (accuracy) including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the overall parameter size of the original XLM-R and XLM-V (No-Trim) is 278M and 778M, respectively, with the vocabulary size being 250K and 901K vocabulary in each case.", + "bbox": [ + 112, + 317, + 882, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of Pre-FT VT appears to be sensitive to the language and task, where it improves the performance in some languages such as Italian (XLM-R and XLM-V achieve $7.9\\%$ and $3.8\\%$ increase for sentiment analysis), but it decreases the performance with non-trivial margin in other languages such as Arabic, where XLM-R decreases $5\\%$ for sentiment analysis and $2\\%$ for XNLI. Since XLM-V has a larger vocabulary size, the percentage of reduced parameters at VT is more prominent in XLM-V, as seen in Arabic $(20.2\\%)$ and Portuguese $(28.9\\%)$ for example.", + "bbox": [ + 110, + 399, + 489, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Vocabulary Size Analysis", + "text_level": 1, + "bbox": [ + 112, + 602, + 366, + 620 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our main experimental results (§ 4.2), all the unique tokens that appeared in the monolingual corpus were kept, which resulted in a low compression ratio for some languages such as English and Russian. In this analysis, we constrain the number of vocabulary and choose the top- $n$ vocabulary at VT in terms of the frequency in the corpus (see § 3). For QA and QG, we compare mT5SMALL results with $n$ from [5K, 10K, 15K, 30K, 60K, 90K, 120K], which correspond to an overall parameter size of [49M, 54M, 59M, 74M, 105M, 136M, 166M], respectively. For sentiment analysis and NLI, we experiment with XLM-RBASE with $n$ from [5K, 10K, 15K, 30K, 60K], which correspond to [89M, 93M, 97M, 109M, 132M] of parameter size, respectively4.", + "bbox": [ + 110, + 627, + 489, + 885 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4The full results with top- $n$ can be found at Appendix B and Appendix A.", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 Generation Tasks: QA & QG", + "text_level": 1, + "bbox": [ + 507, + 399, + 786, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 shows the results of mT5 on QA and QG. Noticeably, post-FT VT can reduce the vocabulary size to $60\\mathrm{K}$ for both QA and QG in all the languages with a trivial gap (0.3% decrease of EM in Russian QA and 0.1% decrease of BS in French QG), and that is $35\\%$ of the original mT5 in the parameter size. Furthermore, post-FT VT can further reduce the vocabulary to 5K tokens with no more than $0.4\\%$ decrease in each metric for both QA and QG in English, French, and Italian, which is $16\\%$ of the original mT5 in the parameter size. Meanwhile, pre-FT VT outperforms the No-Trim result in all the languages in QA, and the majority of the languages in QG (English, Italian, Korean, and Russian), but the result is sensitive to the choice of $n$ . For example, Japanese/Korean QA and Russian QG with pre-FT VT for top-5K (16% of the original mT5) outperforms No-Trim as well as post-FT VT, but Japanese QG with pre-FT VT is worse in any choice of $n$ on contrary. This larger variation of results may also be due to the parameter size space, as the optimal parameters for the original multilingual LM (which is the one trained for post-FT VT) may differ. We leave this extended analysis for future work.", + "bbox": [ + 505, + 420, + 885, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2 Classification Tasks: Sentiment & NLI", + "text_level": 1, + "bbox": [ + 507, + 833, + 858, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 5 and Figure 6 show the results of XLM-R on sentiment and NLI. In NLI, we can see that post/pre-FT VT both can reduce the vocabulary to $30\\mathrm{K}$ ( $39\\%$ of the original XLM-R in the parameter", + "bbox": [ + 507, + 854, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14729", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5e6650c638298516651a68d4acd871db90087e97bf0034e116a4a47a4562c1da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 82, + 300, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/214b1cb5de93ffa682717eca311d74b5ddd98860b07b847b9bc40f92e49b2e8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 82, + 482, + 147 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ef4be96790435dcb576decfd929b245981d80679335f72a4804c3fc3200b162.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 151, + 299, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/745560447b5a3de1acb235070cd9744770219f7571f3d1aad6265b035e465ccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 309, + 151, + 482, + 215 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/715ebe8415fbefb1a0097708d1e4901076b019320b6bb99ee83d76accb0d6567.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 219, + 299, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b54ba16a3c7b227762bc66d0dbee6a7528d0368e9a2239f999f2fad2432123eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 219, + 482, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a090bc793a0a1869516bb430ca16bd11094503d3f563323833eda28867efc53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 288, + 299, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75c098b5fc4c7ffe757384015efa7fe788d23b6d0edb6e41874109334c440c08.jpg", + "image_caption": [ + "Figure 5: Sentiment analysis macro-F1 results of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + ], + "image_footnote": [], + "bbox": [ + 304, + 288, + 482, + 351 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/47f32c80e7c2e10c226f0396803a7183389b3f1704257858fc712dacfdb6f84d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 357, + 299, + 420 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24170ba9678d8eac52dabf8ca9c46bde1a4fce44ac8cd021b7043becbd112525.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 357, + 482, + 420 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b757cbcbd91cfa363e40352c2e94bd30605250aa42e034ae9b3550e5c7e42ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 425, + 299, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85215f179d7c7ebb6e1a627c6d011a684c3ba3ba329818ee70d99c65976aacc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 304, + 425, + 482, + 489 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afe7ff6bb4af375fdc3f9374c1aec20d7c7c4ffcf56a69f3ea9a4e923512bd42.jpg", + "image_caption": [ + "Figure 4: QG (METEOR) and QA (Ans-F1) results for mT5 with pre/post-FT VT for different vocabulary sizes compared to the original multilingual LM (No-Trim)." + ], + "image_footnote": [], + "bbox": [ + 124, + 493, + 299, + 565 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b4cf891736fc8564d14bf79275ea364660fedb30dac1ed1cacc0e91411c9fe5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 493, + 482, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size) without any decrease except $0.3\\%$ of pre-FT VT for German, and there is no decrease more than $0.4\\%$ even with top-15K of post-FT VT. In sentiment analysis, pre-FT VT with top-10K ( $33\\%$ of the original XLM-R in the parameter size) can retain the accuracy of the No-Trim baseline in French and Italian. Moreover, post-FT VT with 30K can retain the original F1 score without a major drop in sentiment analysis, yet the decrease in F1 score is slightly more prominent than NLI ( $1.1\\%$ in Arabic sentiment analysis).", + "bbox": [ + 112, + 645, + 487, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The sentiment analysis datasets are collected from Twitter, so one dataset in a single language can contain tokens from other languages ( hashtags or named-entities, or even code-switching). In contrast, XNLI translates English NLI into other languages, so there is less chance for a dataset to", + "bbox": [ + 112, + 822, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/146d9236bbdef860d7b91976f3d8c5f137a736392dfaf7a613addd91ab956bbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 82, + 700, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cfc291a843c11054b28ea043e7f22765b94502110b2ffac101a2776168a0c008.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 82, + 880, + 148 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65b148e2aa17e46779763a15938c58f8f311d869565ff8de7e2718adfc9dbb77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 152, + 700, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccde1b462d6a6e11a8a61bf13ea45d0247ec5eb000e3d85060a92bfcf13aaeb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 152, + 880, + 217 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98d1ff5f5094a977fb4764fafc0ca6667516a7a4ee0f30ca8467823a3b877e6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 222, + 700, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f017747f0f32250fc7097711d2769fc2ff5542f0931cd947827cd75add6337d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 222, + 880, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66c339afbae3af67690ff9ad56f6178f87a765586078dcdfeacdc4bc86401f50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 291, + 700, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ba2036ff80ce83fd6878f1d359b292a508134833619f586f8b62f78aec94c5c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 432, + 695, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4721b40d5c2aeac46f6042c96d91ea63d05e992dba1472f0f01bc7dacc07d33d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 432, + 880, + 493 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee2bf3a8a0cae4e99d677ced953d5341b1fd2ab55ecbfdda0c44388ac3e12256.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 497, + 695, + 558 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/440cd0ff6cd1f351989c873d2f0a9cf73696ba90d62230c2167003f54be7976c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 498, + 880, + 558 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8bf9aa983690495010bd0429eaa962a0f9254995a2795bc0b8126bf7d24aac4a.jpg", + "image_caption": [ + "Figure 6: NLI accuracy of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + ], + "image_footnote": [], + "bbox": [ + 512, + 563, + 695, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "contain tokens from other languages. This can explain the effectiveness of top- $n$ VT in NLI compared with sentiment analysis, as smaller values of $n$ should result in a vocabulary with fewer tokens from the other languages, which limits the ability of the models to handle foreign tokens.", + "bbox": [ + 507, + 689, + 882, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6 Monolingual vs. Multilingual LMs: The Case of Social Bias", + "text_level": 1, + "bbox": [ + 507, + 796, + 882, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There has been extensive literature in NLP comparing monolingual and multilingual LMs (Muller et al., 2021; Goyal et al., 2021). As for the performance, there is no clear consensus on which type is better for certain languages, tasks or settings. How-", + "bbox": [ + 507, + 838, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14730", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ever, there are other important factors that play a role in this comparison. First, monolingual models tend to have a smaller vocabulary size, which makes them more practical. In contrast, a single multilingual model can be used for a large number of languages. Moreover, multilingual LMs are less prone to capture and carry cultural- or language-dependent biases. This is due to the combination of languages and cultures into a single model, which makes it less biased toward specific cultures (Liang et al., 2020; Ahn and Oh, 2021). Prior works have shown that different types of biases consistently appear in language-specific models (Nadeem et al., 2021; Nangia et al., 2020; Blodgett et al., 2021; Dhamala et al., 2021; Kaneko et al., 2022; Zhou et al., 2022). While the comparison of monolingual and multilingual LMs is not the main focus of this paper, this analysis is certainly relevant. Trimming the vocabulary of a multilingual model essentially makes the model smaller, and therefore alleviates one of the main drawbacks of using multilingual language models on language-specific tasks, which is its larger size. On top of that, this strategy enables the usage of monolingual models with potentially less social bias. In the following, we present a comparison of monolingual and multilingual LMs (both trimmed and not trimmed) in terms of social bias and general performance.", + "bbox": [ + 112, + 84, + 492, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1 Experimental setting", + "text_level": 1, + "bbox": [ + 112, + 556, + 326, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Social bias datasets. To study the effect of VT on social bias existing in pre-trained LMs, we first conduct experiments on two commonly used social bias evaluation datasets for masked LMs: StereoSet (SS; Nadeem et al., 2021) $^5$ and crowdsourced stereotype pairs benchmark (CP; Nangia et al., 2020) $^6$ . SS consists of associative contexts covering four types of social biases: race, gender, religion, and profession; while CP is crowdsourced and annotated by workers in the United States, which contains nine types of social biases: race, gender, sexual orientation, religion, age, nationality, disability, physical appearance, and socioeconomic status. In order to further investigate the impact of pre/post-FT VT on LMs, we trim and fine-tune models on sentiment analysis with different orders and evaluate the social bias in such models on the Equity Evaluation Corpus (EEC; Kiritchenko and", + "bbox": [ + 112, + 582, + 489, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mohammad, 2018) considering two bias types: gender and race. The EEC dataset was specifically with the aim to examine social bias for sentiment analysis systems.", + "bbox": [ + 507, + 83, + 884, + 149 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. We compare the pseudolikelihood scores returned by each model for stereotypical and anti-stereotypical sentences using AULA (All Unmasked Likelihood with Attention weights) (Kaneko and Bollegala, 2022).8 AULA has been shown to be robust against the frequency biases of the masked tokens and provides a more reliable assessment in contrast to alternative metrics when evaluating social biases in masked language models (MLMs). Given a sentence pair in the test dataset: \"My mom spent all day cooking for Thanksgiving\" vs. \"My dad spent all day cooking for Thanksgiving\", the first sentence is considered as stereotypical while the second one is anti-stereotypical. AULA computes the percentage of stereotypical sentences preferred by the MLM over anti-stereotypical ones as the bias score. An MLM is considered to be unfairly biased if it returns higher pseudo-loglikelihood scores for stereotypical sentences than the corresponding anti-stereotypical sentences. The AULA score falls within the range of [0,100] and an unbiased model would return a bias score close to 50. On the other hand, a bias score greater than or less than 50 indicates the bias direction towards the stereotype or anti-stereotype, respectively. Since the original AULA is not fitted to evaluate fine-tuned models, we adapt AULA to the EEC dataset obtain the bias score for the LMs fine-tuned on sentiment analysis, and denote this metric as EEC-AULA. Specifically, given a model that predicts sentiment labels (e.g., positive, neutral, negative) to sentences, we consider the percentage of stereotypical test sentences with a more negative label over anti-stereotypical ones as the corresponding bias evaluation measure.", + "bbox": [ + 507, + 160, + 885, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General performance. As a proxy to test the general performance, we use the general language understanding evaluation (GLUE; Wang et al., 2018) benchmark. We acknowledge the limitations of using this benchmark to draw reliable conclusions at large (Ethayarajh and Jurafsky, 2020)", + "bbox": [ + 507, + 736, + 884, + 832 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "7https://saifmohammad.com/WebPages/Biases-SA.html", + "bbox": [ + 507, + 843, + 880, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "https://github.com/kanekomasahiro/evaluate.bias_in_m1m", + "bbox": [ + 507, + 866, + 882, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{9}$ https://gluebenchmark.com/; Models are tested on the development sets of each task.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5https://github.com/moinnadeem/StereoSet", + "bbox": [ + 134, + 890, + 440, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "6https://github.com/nyu-mll/crows-pairs", + "bbox": [ + 134, + 904, + 433, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14731", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5e69a86edc9cc19c07171eef4776f4c03d84a4413a0b7e484acfc5858960d6d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3: Results of pre/post-FT VT models compared with the original monolingual and multilingual models on two social bias analysis benchmarks (AULA for pre-trained masked language models and EEC-AULA for models fine-tuned on sentiment analysis) and the GLUE benchmark. The VT models are trimmed on English vocabulary with different vocabulary sizes: EN (full English vocabulary) and 50K (top 50K subword tokens). Note that for post-FT VT, the results on AULA are exactly the same as the original XLM-R. The green and red colours represent the social bias towards anti-stereotypical sentences (scores lower than 50) and stereotypical sentences (scores higher than 50), respectively. The lighter colour indicates less social bias observed in the LM." + ], + "table_body": "
Base ModelVocab TrimmingVocabularySocial BiasGeneral Performance
Pre-FTPost-FTAULAEEC-AULAGLUE
CPSSGenderRace
Monolingual (RoBERTa)-50K50K58.158.864.885.778.0
Multilingual (XLM-R)-250K250K49.554.944.362.077.9
Pre-FT VT (EN)173K173K49.554.942.556.978.0
Post-FT VT (EN)250K173k49.554.944.362.077.9
Pre-FT VT (50K)50K50K49.355.041.062.476.9
Post-FT VT (50K)250K50K49.554.944.362.077.9
", + "bbox": [ + 119, + 80, + 878, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "but it nevertheless provides a good proxy for understanding the overall performance of comparable models in standard NLP tasks. Moreover, these experiments are only aimed at analysing the effect of vocabulary trimming on general performance.", + "bbox": [ + 112, + 365, + 487, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Models. We compute the bias scores of RoBERTa (Liu et al., 2019) as base monolingual LM, and XLM-R (Conneau et al., 2019) as its multilingual counterpart (they have been trained with the same architecture and in an overlapping corpus. We explore two VT settings to be applied to XLM-R: XLM-R with the standard VT including the full English vocabulary (VT XLM-R) and XLM-R with VT for top-50K English vocabulary (top-50K VT XLM-R), which is the same vocabulary size as the monolingual RoBERTa model. Our experiments are based both on masked language models on AULA (in which the post-VT does not have any effect) and models fine-tuned on the sentiment analysis presented in $\\S 4.1$ on EEC-AULA, as well as on the corresponding GLUE training sets.", + "bbox": [ + 115, + 453, + 489, + 711 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2 Results", + "text_level": 1, + "bbox": [ + 112, + 721, + 218, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the performance of pre-FT and post-FT VT models against the original monolingual and multilingual LMs on social bias evaluation datasets and the GLUE benchmark. Both AULA and GLUE results are computed using the LMs without fine-tuning (i.e., RoBERTa, XLM-R, VT XLM-R, and top-50K VT XLM-R), whereas the EEC-AULA results are computed using the models applying VT and fine-tuning strategies. We observe that the monolingual model contains the highest levels of social biases compared to the multilin", + "bbox": [ + 112, + 741, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gual models with different settings. In particular, RoBERTa obtains the overall highest bias score on the EEC dataset after fine-tuning, with an alarmingly high 85.7 score on race. $^{10}$ On the other hand, compared to the original XLM-R, there is no significant change in performance on social biases and GLUE evaluation tasks for pre-FT VT and post-FT VT models. This is important as we can apply the proposed VT method to any multilingual LM, obtaining a monolingual one with consistent performance on the GLUE benchmark and less social biases than the original monolingual model pretrained in the target language, without using any ad-hoc debiasing methods.", + "bbox": [ + 507, + 365, + 884, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 Discussion", + "text_level": 1, + "bbox": [ + 509, + 601, + 636, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vocabulary trimming before and after finetuning. According to the results, pre-FT VT appears to be generally more effective in classification tasks (see section 4.2.2). For generation tasks, both pre/post-FT VT robustly retain the original performance while being able to considerably reduce the model size (see section 4.2.1). As a guideline to choose the type of VT in such a case, post-FT VT should be more suitable if one already has a finetuned model, as no additional training is needed for this case. Moreover, post-FT is more robust as a compression mechanism as the performance is largely maintained with respect to that of the original multilingual LM. On the other hand, if one needs to fine-tune a model from scratch and the computation resources are limited, we recommend", + "bbox": [ + 507, + 627, + 882, + 883 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "10 Appendix C includes a breakdown by type of the social bias results.", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14732", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "exploring pre-FT VT, as fine-tuning on a trimmed LM should be more efficient due to its smaller vocabulary and parameters and, in some cases, can lead to better overall results. However, this process has to be done carefully as the set of optimal parameters could differ from the original multilingual LM fine-tuning process.", + "bbox": [ + 112, + 84, + 489, + 197 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While in this paper we have not compared monolingual and multilingual models, the question would be whether we need vocabulary trimming strategies in a world where monolingual LMs exist. In this case, a monolingual model may perform similarly to a multilingual model (Goyal et al., 2021). However, the multilingual model is often larger mainly due to larger vocabulary storage requirements. In contrast, our proposed VT technique does not require any extra LM training or computational resources. Indeed, only a multilingual LM is needed and we can induce multiple smaller language-specific monolingual models. This may reduce the carbon footprint overall and especially help with less-resource languages when a high-quality monolingual model does not exist. Finally, our social bias analysis see § 6 shows how monolingual models exhibit larger social biases (especially racial) than a VT-induced multilingual LM. This is consistent with prior work suggesting that a multilingual LM has been trained with more languages, and hence more cultural variety, and these diverging viewpoints can compensate each other (Ahn and Oh, 2021).", + "bbox": [ + 112, + 206, + 489, + 608 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 619, + 247, + 634 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this paper, we proposed vocabulary-trimming (VT), a method to reduce the vocabulary of a multilingual LM to a vocabulary specific to any given target language. VT can induce a monolingual LM in the target language by leveraging an existing multilingual LM. The main advantage of this filtering step is the reduced size, as well as avoiding having to train monolingual LMs from scratch, which would be computationally demanding. Our experiments show how VT can retain the high performance of the original multilingual LM, while largely reducing the model size. For all languages evaluated, a $35\\%$ compression rate proves sufficient to keep the original performance of the larger mT5 multilingual LM in both QA and QG, with a similar $39\\%$ in NLI and $55\\%$ in sentiment analysis with XLM-R. Interestingly, in some cases, the", + "bbox": [ + 112, + 645, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "compressed LM can even achieve better results than the original larger model when trimmed before fine-tuning. Since the main goal of the paper was to compress a multilingual LM while keeping its original performance, we leave the analysis of this behaviour for future work.", + "bbox": [ + 507, + 84, + 884, + 180 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 193, + 615, + 209 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We have not tested our methodology in truly low-resource languages. Because of this, there could be a different behaviour when we apply VT to a language with lower resources or that is poorly represented in the underlying training corpus. The LMs we used in the paper limited their size up to 600M, and we have not considered larger models such as mT5XXL or BLOOM (Scao et al., 2022), due to our limited computational resources. As the language-specific corpus to compute frequency, we employ mC4, which is one of the largest multilingual corpora. Nonetheless, this is used as a proxy and having access to the full multilingual model could give potentially better results.", + "bbox": [ + 507, + 219, + 884, + 444 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Similarly, we acknowledge the limitations of the analysis comparing multilingual and monolingual models in terms of social bias. Due to the evaluation data available and the existence of comparable monolingual and multilingual LMs, the evaluation is focused on English only and the results could differ for other languages. Moreover, there are other types of biases not covered in this evaluation.", + "bbox": [ + 507, + 445, + 884, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 509, + 587, + 660, + 602 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pre-trained LMs are known to contain undesirable biases to generate toxic contents in some edge cases (Schick et al., 2021), so the resulting models could inherit such biases. While we have not analysed in detail the output of all models in the tasks evaluated, in this paper we have made an attempt to study this effect in terms of social biases for both base pretrained LMs and fine-tuned LMs.", + "bbox": [ + 507, + 613, + 884, + 741 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 755, + 684, + 771 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jose Camacho-Collados and Yi Zhou are supported by a UKRI Future Leaders Fellowship.", + "bbox": [ + 507, + 781, + 880, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 841, + 608, + 857 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Amine Abdaoui, Camille Pradel, and Grégoire Sigel. 2020. Load what you need: Smaller versions of mutililingual BERT. In Proceedings of SustainNLP: Workshop on Simple and Efficient Natural Language", + "bbox": [ + 510, + 865, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14733", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Processing, pages 119-123, Online. Association for Computational Linguistics.", + "Jaimeen Ahn and Alice Oh. 2021. Mitigating language-dependent ethnic bias in BERT. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 533-549, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics.", + "Francesco Barbieri, Valerio Basile, Danilo Croce, Malvina Nissim, Nicole Novielli, and Viviana Patti. 2016. Overview of the evalita 2016 sentiment polarity classification task. In Proceedings of third Italian conference on computational linguistics (CLiC-it 2016) & fifth evaluation campaign of natural language processing and speech tools for Italian. Final Workshop (EVALITA 2016).", + "Francesco Barbieri, Luis Espinosa Anke, and Jose Camacho-Collados. 2022. XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 258-266, Marseille, France. European Language Resources Association.", + "Farah Benamara, Cyril Grouin, Jihen Karoui, Véronique Moriceau, and Isabelle Robba. 2017. Analyse d'opinion et langage figuratif dans des tweets: presentation et résultats du début fouille de textes deft2017.", + "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1004-1015.", + "Henrico Bertini Brum and Maria das Graças Volpe Nunes. 2017. Building a sentiment corpus of tweets in brazilian portuguese. arXiv preprint arXiv:1712.08917.", + "Carrino Casimiro Pio, Costa-jussa Marta R., and Fonollosa Jose A. R. 2019. Automatic Spanish Translation of the SQuAD Dataset for Multilingual Question Answering. arXiv e-prints, page arXiv:1912.05200v1.", + "Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Cieliebak, Jan Milan Deriu, Dominic Egger, and Fatih Uzdilli. 2017. A twitter corpus and benchmark resources for german sentiment analysis. In 5th International Workshop on Natural Language Processing for Social Media, Boston MA, USA, 11 December 2017, pages 45-51. Association for Computational Linguistics.", + "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116.", + "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32.", + "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics.", + "Danilo Croce, Alexandra Zelenanska, and Roberto Basili. 2018. Neural learning for question answering in italian. In AI*IA 2018 – Advances in Artificial Intelligence, pages 389–402, Cham. Springer International Publishing.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. Bold: Dataset and metrics for measuring biases in open-ended language generation. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 862-872.", + "Martin d'Hoffschmidt, Wacim Belblidia, Quentin Heinrich, Tom Brendle, and Maxime Vidal. 2020. FQuAD: French question answering dataset. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1193–1208, Online. Association for Computational Linguistics.", + "Manuel C Díaz-Galiano, Eugenio Martínez-Cármara, M Ángel García Cumbreras, Manuel García Vega, and Julio Villena Román. 2018. The democratization of deep learning in tass 2017. Procesamento del Lenguaje Natural, 60:37-44." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14734", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Pavel Efimov, Andrey Chertok, Leonid Boytsov, and Pavel Braslavski. 2020. Sberquad-russian reading comprehension dataset: Description and analysis. In International Conference of the Cross-Language Evaluation Forum for European Languages, pages 3-15. Springer.", + "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics.", + "Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, and Alexis Conneau. 2021. Larger-scale transformers for multilingual masked language modeling. In Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021), pages 29–33, Online. Association for Computational Linguistics.", + "Masahiro Kaneko and Danushka Bollegala. 2021. Debiasing pre-trained contextualised embeddings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1256-1266, Online. Association for Computational Linguistics.", + "Masahiro Kaneko and Danushka Bollegala. 2022. Unmasking the mask-evaluating social biases in masked language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11954-11962.", + "Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, and Naoaki Okazaki. 2022. Gender bias in masked language models for multiple languages. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2740-2750, Seattle, United States. Association for Computational Linguistics.", + "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.", + "Svetlana Kiritchenko and Saif M Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. *NAACL HLT* 2018, page 43.", + "Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166-172.", + "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In International Conference on Learning Representations." + ], + "bbox": [ + 115, + 85, + 485, + 917 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.", + "Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, and Madian Khabsa. 2023. Xlm-v: Overcoming the vocabulary bottleneck in multilingual masked language models. arXiv preprint arXiv:2301.10472.", + "Sheng Liang, Philipp Dufter, and Hinrich Schütze. 2020. Monolingual and multilingual reduction of gender bias in contextualized representations. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5082-5093, Barcelona, Spain (Online). International Committee on Computational Linguistics.", + "Seungyoung Lim, Myungji Kim, and Jooyoul Lee. 2019. Korquad1.0: Korean qa dataset for machine reading comprehension. arXiv preprint arXiv:1909.07005.", + "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Kelly Marchisio, Patrick Lewis, Yihong Chen, and Mikel Artetxe. 2022. Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training. arXiv preprint arXiv:2212.10503.", + "Chandler May, Alex Wang, Shikha Bordia, Samuel Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 622-628.", + "Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14735", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics.", + "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953-1967, Online. Association for Computational Linguistics.", + "Malte Ostendorff and Georg Rehm. 2023. Efficient language model training through cross-lingual and progressive transfer learning. arXiv preprint arXiv:2301.09626.", + "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics.", + "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics.", + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551.", + "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.", + "Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 task 4: Sentiment analysis in Twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502-518, Vancouver, Canada. Association for Computational Linguistics.", + "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP. Transactions of the Association for Computational Linguistics, 9:1408-1424.", + "ByungHoon So, Kyuhong Byun, Kyungwon Kang, and Seongjin Cho. 2022. Jaquad: Japanese question answering dataset for machine reading comprehension. arXiv preprint arXiv:2202.01764.", + "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2022. Generative language models for paragraph-level question generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 670-688, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2023. A practical toolkit for multilingual question and answer generation. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Toronto, Canada. Association for Computational Linguistics.", + "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355.", + "Hai Wang, Dian Yu, Kai Sun, Jianshu Chen, and Dong Yu. 2019. Improving pre-trained multilingual model with vocabulary expansion. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 316-327, Hong Kong, China. Association for Computational Linguistics.", + "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.", + "Bo Zheng, Li Dong, Shaohan Huang, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, and Furu Wei. 2021. Allocating large vocabulary capacity for crosslingual language model pre-training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3203-3215, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Yi Zhou, Masahiro Kaneko, and Danushka Bollegala. 2022. Sense embeddings are also biased - evaluating social biases in static and contextualised sense embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14736", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/e9718ab3ec99ac02fae862b9e5c571a095a84004a3d312e91a8e55e390e70788.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Vocab.No-Trim (250K)5K10K15K30K60K
Param.278M89M93M97M109M132M
Post-FT (Sentiment)AR66.364.564.565.965.9
DE73.270.472.472.173.7
EN68.464.066.567.468.6
ES69.066.267.267.868.4
FR71.868.671.471.771.6
IT62.960.861.963.562.3
PT70.763.665.968.269.4
Pre-FT (Sentiment)AR66.358.661.963.163.2
DE73.270.671.873.171.5
EN68.464.666.667.766.0
ES69.060.465.666.365.9
FR71.870.374.373.872.2
IT62.966.267.168.168.2
PT70.765.567.569.667.9
Post-FT (NLI)AR75.775.075.875.875.7
DE79.976.678.979.679.9
EN84.680.583.684.384.6
ES79.875.378.479.479.8
FR80.177.280.080.180.2
Pre-FT (NLI)AR75.773.075.075.475.7
DE79.977.479.079.079.6
EN84.683.384.784.485.1
ES79.878.979.179.281.0
FR80.177.378.978.280.1
", + "bbox": [ + 117, + 82, + 500, + 444 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4: Results of XLM-R for sentiment analysis (macro F1) and NLI (accuracy) with different top- $n$ vocabulary size at VT, where the best results in each LM and language are in the bold characters.", + "bbox": [ + 112, + 455, + 487, + 513 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "(Volume 1: Long Papers), pages 1924-1935, Dublin, Ireland. Association for Computational Linguistics.", + "bbox": [ + 131, + 539, + 487, + 567 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A Top- $n$ VT of XLM-R", + "text_level": 1, + "bbox": [ + 114, + 579, + 334, + 595 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 4 shows the results of XLM-R fine-tuned on sentiment and NLI with post/pre-VT for different top- $n$ .", + "bbox": [ + 112, + 605, + 487, + 653 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B Top- $n$ VT of mT5", + "text_level": 1, + "bbox": [ + 114, + 664, + 305, + 681 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5 shows the results of mT5 fine-tuned on QA and QG with post/pre-VT for different top- $n$ .", + "bbox": [ + 112, + 690, + 487, + 722 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C Details of Results on Social Bias Evaluation", + "text_level": 1, + "bbox": [ + 114, + 734, + 428, + 765 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 6 shows the details of social bias evaluation (EEC dataset) regarding each emotion type observed in the LMs fine-tuned on sentiment analysis. Table 7 shows the details of social bias regarding each bias type in both CP and SS datasets observed in the comparison LMs.", + "bbox": [ + 112, + 776, + 489, + 872 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "14737", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/babc967533548380adbf296da82690e0accd09809baca74250c8df0cd2eaedc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Vocab. \nParam.No-Trim (250K) \n300M5K \n49M10K \n54M15K \n59M30K \n74M60K \n105M90K \n136M120K \n166M
EN70.1 / 55.572.4 / 59.773.2 / 60.573.7 / 60.972.7 / 58.871.0 / 56.470.5 / 55.870.3 / 55.6
ES55.9 / 34.753.5 / 32.554.4 / 33.154.5 / 33.255.2 / 33.856.0 / 34.855.9 / 34.855.9 / 34.7
FR50.0 / 30.950.0 / 31.548.9 / 29.948.6 / 29.349.9 / 30.149.8 / 30.550.2 / 30.950.0 / 30.8
IT53.2 / 37.654.3 / 39.653.8 / 38.454.3 / 38.854.2 / 38.553.7 / 38.153.4 / 37.8-
JA65.7 / 65.754.4 / 54.460.3 / 60.362.6 / 62.660.4 / 60.464.9 / 64.965.6 / 65.665.7 / 65.7
KO77.1 / 70.675.0 / 67.875.8 / 68.776.2 / 69.376.9 / 70.277.3 / 70.7--
RU73.7 / 51.470.1 / 48.871.4 / 49.570.3 / 47.670.3 / 47.973.8 / 51.173.9 / 51.473.8 / 51.4
EN23.8 / 90.023.6 / 89.824.1 / 89.924.2 / 90.024.2 / 90.023.9 / 90.023.9 / 90.023.9 / 90.0
ES22.7 / 84.121.4 / 83.622.0 / 83.822.2 / 83.922.4 / 84.022.7 / 84.122.7 / 84.122.7 / 84.1
FR17.5 / 80.717.1 / 80.317.1 / 80.417.3 / 80.517.3 / 80.517.4 / 80.617.5 / 80.717.5 / 80.7
IT17.6 / 80.817.3 / 80.717.5 / 80.817.5 / 80.817.6 / 80.717.6 / 80.817.6 / 80.8-
JA29.0 / 80.925.7 / 79.227.4 / 80.128.2 / 80.528.8 / 80.929.0 / 80.929.0 / 80.929.0 / 80.9
KO27.5 / 82.927.3 / 82.827.4 / 82.927.4 / 82.927.4 / 82.927.5 / 83.0--
RU26.4 / 84.326.0 / 84.026.2 / 84.226.3 / 84.226.3 / 84.226.3 / 84.326.4 / 84.326.4 / 84.3
EN70.1 / 55.574.0 / 61.371.8 / 59.173.4 / 60.774.3 / 61.371.3 / 56.568.6 / 54.366.2 / 52.1
ES55.9 / 34.756.7 / 36.758.1 / 37.562.2 / 41.459.8 / 40.158.0 / 37.058.9 / 38.152.0 / 32.7
FR50.0 / 30.949.0 / 32.747.6 / 30.243.3 / 26.944.9 / 27.434.3 / 20.043.1 / 24.140.0 / 23.7
IT53.2 / 37.657.6 / 43.761.5 / 46.861.2 / 45.657.8 / 42.156.5 / 40.555.4 / 39.2-
JA65.7 / 65.755.3 / 55.352.2 / 52.261.3 / 61.360.4 / 60.464.6 / 64.666.9 / 66.967.2 / 67.2
KO77.1 / 70.679.5 / 72.683.7 / 77.782.7 / 76.480.4 / 73.781.6 / 75.1--
RU73.7 / 51.473.5 / 51.776.9 / 56.477.4 / 56.975.9 / 54.275.1 / 53.075.0 / 53.073.3 / 51.4
EN23.8 / 90.024.8 / 90.024.4 / 89.924.3 / 89.924.2 / 90.024.2 / 90.023.3 / 89.923.6 / 89.8
ES22.7 / 84.121.9 / 84.121.9 / 83.922.7 / 83.722.0 / 84.322.6 / 84.420.4 / 79.622.6 / 84.1
FR17.5 / 80.716.1 / 79.016.9 / 79.916.9 / 79.515.8 / 78.814.6 / 78.015.7 / 79.017.0 / 79.2
IT17.6 / 80.817.4 / 80.417.8 / 81.118.0 / 80.817.1 / 80.617.4 / 80.917.4 / 80.8-
JA29.0 / 80.926.6 / 79.827.0 / 79.227.5 / 79.727.3 / 79.927.7 / 80.328.3 / 80.828.2 / 80.2
KO27.5 / 82.928.4 / 83.427.8 / 83.428.2 / 84.128.8 / 83.428.4 / 83.4--
RU26.4 / 84.327.7 / 85.930.0 / 86.629.0 / 86.529.2 / 86.329.0 / 86.629.2 / 86.728.9 / 86.0
", + "bbox": [ + 132, + 105, + 865, + 521 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/70f4930854529e6fc638c0eb2c7ba51f1c07fbb96abbffcfdec04a93f551394b.jpg", + "table_caption": [ + "Table 5: Results of mT5 for QA (Ans-F1/EM) and QG (MTR/BS) with different top- $n$ vocabulary size at VT,where the best results in each LM and language are in the bold characters." + ], + "table_footnote": [], + "table_body": "
ModelAULA-EECAngerSadnessFearJoyNo Emotion Type
Gender BiasFT RoBERTa64.850.067.871.564.360.8
FT XLM-R44.346.939.451.434.439.7
Pre-FT XLM-R (EN)42.533.855.845.140.742.7
Post-FT XLM-R (EN)44.346.939.451.434.439.7
Pre-FT XLM-R (50k)44.346.939.451.434.439.7
Post-FT XLM-R (50k)41.041.245.847.236.937.2
Race BiasFT RoBERTa85.750.088.065.450.0100.0
FT XLM-R62.074.532.147.374.617.4
Pre-FT VT XLM-R (EN)56.933.758.869.260.150.0
Pre-FT VT XLM-R (EN)62.074.532.147.374.617.4
Pre-FT VT XLM-R (50k)62.074.532.147.374.617.4
Post-FT VT XLM-R (50k)62.455.770.370.959.159.0
", + "bbox": [ + 115, + 617, + 882, + 850 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Table 6: Social bias scores of LMs fine-tuned on sentiment analysis on each emotion type w/wo VT on the EEC dataset.", + "bbox": [ + 112, + 860, + 882, + 889 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "14738", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/7814161b730d216654a6dd64a07329188f414f76ecb9dad36c928a61bbd9b922.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RoBERTaXLM-RVT XLM-R (EN)VT XLM-R (50K)
CP DatasetAULA58.149.549.549.3
Age59.849.449.449.4
Disability68.366.766.768.3
Gender53.448.548.548.1
Nationality56.644.044.044.7
Physical-Appearance52.458.758.755.6
Race-Color56.843.843.844.0
Religion53.354.354.353.3
Sexual-Orientation67.954.854.854.8
Socioeconomic66.358.158.157.6
SS DatasetAULA58.854.954.955.0
Gender62.455.355.354.9
Profession60.454.354.354.6
Race57.056.056.056.0
Religion54.446.846.846.8
", + "bbox": [ + 152, + 337, + 845, + 619 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 7: Social bias scores of VT LMs compared to their original ones on both CP and SS datasets for each bias type.", + "bbox": [ + 112, + 627, + 880, + 659 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "14739", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 14 + } +] \ No newline at end of file diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_model.json b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_model.json new file mode 100644 index 0000000000000000000000000000000000000000..571d4a22fa17a2cec1982982f864b9af1d39e08a --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_model.json @@ -0,0 +1,2463 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.227, + 0.09, + 0.773, + 0.131 + ], + "angle": 0, + "content": "Efficient Multilingual Language Model Compression through Vocabulary Trimming" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.156, + 0.741, + 0.171 + ], + "angle": 0, + "content": "Asahi Ushio and Yi Zhou and Jose Camacho-Collados" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.172, + 0.836, + 0.189 + ], + "angle": 0, + "content": "Cardiff NLP, School of Computer Science and Informatics, Cardiff University, UK" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.19, + 0.746, + 0.205 + ], + "angle": 0, + "content": "{UshioA,Zhouy131,CamachoColladosJ}\\@cardiff.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.278, + 0.461, + 0.675 + ], + "angle": 0, + "content": "Multilingual language models (LMs) have become a powerful tool in NLP, especially for non-English languages. Nevertheless, model parameters of multilingual LMs remain large due to the larger embedding matrix of the vocabulary covering tokens in different languages. Instead, monolingual LMs can be trained in a target language with the language-specific vocabulary only. In this paper, we propose vocabulary-trimming (VT), a method to reduce a multilingual LM vocabulary to a target language by deleting potentially irrelevant tokens from its vocabulary. In theory, VT can compress any existing multilingual LM to any language covered by the original model. In our experiments, we show that VT can retain the original performance of the multilingual LM, while being considerably smaller in size than the original multilingual LM. The evaluation is performed over four NLP tasks (two generative and two classification tasks) among four widely used multilingual LMs in seven languages. The results show that this methodology can keep the best of both monolingual and multilingual worlds by keeping a small size as monolingual models without the need for specifically retraining them, and can even help limit potentially harmful social biases." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.685, + 0.26, + 0.7 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Multilingual language model (LM) pre-training (Devlin et al., 2019; Conneau et al., 2019; Liu et al., 2020; Xue et al., 2021) has been shown to be an efficient mechanism to store information from many languages into a single model, without the need for training multiple language-specific models. Moreover, it has been proven reliable for cross-lingual tasks (Pires et al., 2019; Conneau and Lample, 2019) and can provide competitive performance in most settings, generally similar to its monolingual counterparts (Goyal et al., 2021), while being generally less affected by culturally-dependant biases (Ahn and Oh, 2021). Similarly to monolingual" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.885, + 0.461 + ], + "angle": 0, + "content": "models, multilingual LMs can be used for zero/few-shot learning (Scao et al., 2022) by increasing the model size and, more frequently, can be specialized to different tasks by fine-tuning to specific data. In practice, however, there are a few practical issues when training multilingual LM such as the curse of multilinguality (Conneau et al., 2019; Pfeiffer et al., 2022), a trade-off between the number of languages and individual performance in a single language, or the multilingual vocabulary construction, which requires a careful design for better generalization (Chung et al., 2020; Zheng et al., 2021; Liang et al., 2023)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.463, + 0.885, + 0.752 + ], + "angle": 0, + "content": "Besides such generalization concerns, multilingual LMs usually consist of larger parameters than their monolingual counterparts due to the need for a large vocabulary covering multiple languages. This becomes an important issue in practice when the resources to host models are limited. For instance, while using the same configuration (i.e., same number of layers and hidden units), the parameter size of T5SMALL (Raffel et al., 2020) and mT5SMALL (Xue et al., 2021) are 140M and 300M, respectively. This is only due to their difference in vocabulary size, with T5 being 50k and mT5, 250k. In fact, the embedding matrix stemming from the LM vocabulary can occupy a large portion of the parameter space. For instance, the ratio of the embedding matrix to the full model's parameter size in multilingual LMs can be higher than 80% as T5 (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.754, + 0.884, + 0.882 + ], + "angle": 0, + "content": "In this paper, we propose a simple vocabulary trimming (VT) method to remove tokens from the vocabulary of multilingual LMs that may be irrelevant to the target language. This is achieved by automatically identifying language-specific tokens from an underlying text corpus. We consider two VT strategies of pre-FT VT (VT before fine-tuning) and post-FT VT (VT after fine-tuning) and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.884, + 0.918 + ], + "angle": 0, + "content": "1We release a Python library for VT at https://github.com/asahi417/lm-vocab-trimmer." + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14725" + }, + { + "type": "footer", + "bbox": [ + 0.21, + 0.946, + 0.788, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14725-14739 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.085, + 0.468, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.204, + 0.486, + 0.233 + ], + "angle": 0, + "content": "Figure 1: The ratio of the embedding matrix to the number of parameters for each multilingual LM." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.302, + 0.49, + 0.543 + ], + "angle": 0, + "content": "analyse them by varying the final vocabulary size. We conduct experiments on two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI), across seven different languages. The experimental results show that both pre and post fine-tuning VT can reduce the model size while retaining the original performance in generation tasks (QA and QG), and particularly in classification tasks (sentiment and NLI) where the results are close to being identical despite the significant reduction in vocabulary size. In all tasks, the original performance can be generally maintained with less than \\(40\\%\\) of the full model parameters for all languages." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.566, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Finally, even though pre-trained LMs have reported impressive performance on various NLP downstream tasks (Kenton and Toutanova, 2019; Liu et al., 2019; Conneau et al., 2019), such LMs also demonstrate worrying levels of social biases in certain situations (May et al., 2019; Kurita et al., 2019; Kaneko and Bollegala, 2021). One natural question that arises is whether VT can have an influence on the bias level in multilingual LMs, including fine-tuned models. For this purpose, we evaluate social bias in multilingual LMs after applying VT with different settings and compare it against its monolingual counterpart. Experimental results show that the monolingual LM tends to contain more bias than its multilingual versions. Moreover, compared to the original multilingual LM, the bias level has no significant change after applying VT. These results suggest that a monolingual LM can be induced by applying VT to its corresponding multilingual LM, thereby obtaining a less biased monolingual LM compared to its original monolingual counterpart." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.667, + 0.099 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.11, + 0.885, + 0.593 + ], + "angle": 0, + "content": "Several studies have explored the possibility to modify or adapt the vocabulary of LMs. For instance, Artetxe et al. (2020) and Marchisio et al. (2022) adapted a mono-lingual LM into another language by learning the embedding matrix on the new language, while fixing the other weights. Similarly, Wang et al. (2019) augmented the vocabulary of a multilingual LM to new languages with multilingual word alignment (Lample et al., 2018). Zheng et al. (2021) proposed to evaluate the ability of a vocabulary to represent a particular language, and Chung et al. (2020) proposed a multilingual vocabulary construction that balances the tradeoff between optimizing for cross-lingual sub-word sharing and the need for robust representation of individual languages. XLM-V (Liang et al., 2023) combines the idea of Zheng et al. (2021) and Chung et al. (2020) to efficiently enlarge the vocabulary size along with the model size scaling. Ostendorff and Rehm (2023) used a multi-stage fine-tuning to obtain a LM in the target language from other LM in the source language. These prior works modify existing mono/multi-lingual LMs to include new languages, i.e. augmenting the multilinguality of the LMs. In contrast, our study focuses on compressing multilingual LMs into the target language to effectively achieve smaller monolingual LMs, i.e. reducing the multilingual representation of the LMs while retaining the capability in a specific target language." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.593, + 0.885, + 0.849 + ], + "angle": 0, + "content": "The work of Abdaoui et al. (2020) is the most relevant to our study as, to the best of our knowledge, they introduced the idea of VT for the first time. However, their analysis is limited to NLI with pre-fine-tuning VT with mBERT (Devlin et al., 2019) only, as well as a fixed vocabulary size after VT. In contrast, our study compares two VT strategies, before and after fine-tuning, and show how this latter strategy, not considered in Abdaoui et al. (2020), can be a more effective compression technique in some settings. Furthermore, we extend the experiments to generation tasks as well as classification tasks with more recent LMs such as mBART (Lewis et al., 2020) and mT5 (Xue et al., 2021), and provide an exhaustive analysis on the effect of VT." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.862, + 0.737, + 0.879 + ], + "angle": 0, + "content": "3 Vocabulary Trimming" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "To perform vocabulary trimming (VT), we first need a multilingual LM as an input. The idea is to" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.085, + 0.449, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.28, + 0.49, + 0.309 + ], + "angle": 0, + "content": "Figure 2: An illustration of vocabulary trimming for Korean and French." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.337, + 0.49, + 0.53 + ], + "angle": 0, + "content": " tailor model to a particular target language \\(l\\), which in principle belong to the same set of languages used to trained the input multilingual LM. For the target language \\(l\\), VT first identifies language-specific tokens on a language-specific corpus \\(\\mathcal{C}_l\\) and remove all the tokens along with their embeddings except for those appeared in \\(\\mathcal{C}_l\\) as described in Figure 2. In our analysis (§ 5), we also consider to keep the top-\\(n\\) most frequent tokens in \\(\\mathcal{C}_l\\) to further reduce the model size by removing less frequent tokens. We consider two VT strategies: (1) before fine-tuning and (2) after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.532, + 0.49, + 0.676 + ], + "angle": 0, + "content": "The difference between these two strategies is whether to perform VT before or after fine-tuning, as shown in Figure 3. Both VTs have advantages and drawbacks: while pre-FT VT can reduce the time of fine-tuning as the trimmed LM is smaller than the original LM, post-FT VT only need a fine-tuned multilingual LM - this way, post-FT VT can be used as a postprocessing step and no additional language-specific training is required." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.773 + ], + "angle": 0, + "content": "Finally, we release a simple LM vocabulary trimming starting package to apply our proposed technique to any input multilingual transformer-based LM, along with all the models and code needed to reproduce our experiments, at https://github.com/asahi417/lm-vocab-trimmer." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.791, + 0.245, + 0.807 + ], + "angle": 0, + "content": "4 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.819, + 0.49, + 0.868 + ], + "angle": 0, + "content": "In this section, we present our experimental results to test the reliability of our proposed VT methodology in NLP tasks." + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.083, + 0.845, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.161, + 0.882, + 0.19 + ], + "angle": 0, + "content": "Figure 3: Comparisons of Pre-FT vs Post-FT VT in an example of fine-tuning on a task in French." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.219, + 0.724, + 0.234 + ], + "angle": 0, + "content": "4.1 Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.244, + 0.885, + 0.729 + ], + "angle": 0, + "content": "Tasks and datasets. In order to test the efficacy of VT, we consider two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI). As the datasets for QA, we use SQuAD (Rajpurkar et al., 2016) (English), Spanish SQuAD (Casimiro Pio et al., 2019) (Spanish), FQuAD (d'Hoffschmidt et al., 2020) (French), Italian SQuAD (Croce et al., 2018) (Italian), JAQuAD (So et al., 2022) (Japanese), KorQuAD (Lim et al., 2019) (Korean), and SberQuAD (Efimov et al., 2020) (Russian). For QG, we use the same datasets adapted for QG via QG-Bench (Ushio et al., 2022). For sentiment analysis, we use Twitter-based datasets for English (Rosenthal et al., 2017), Arabic (Rosenthal et al., 2017), French (Benamara et al., 2017), Italian (Barbieri et al., 2016), German (Cieliebak et al., 2017), Portuguese (Brum and Nunes, 2017), and Spanish (Diaz-Galiano et al., 2018) from UMSAB (Unified Multilingual Sentiment Analysis Benchmark) (Barbieri et al., 2022). All the sentiment analysis datasets contain three labels: positive, neutral and negative. For NLI, we use XNLI (Conneau et al., 2018), a multilingual NLI dataset, including English, French, German, Spanish and Arabic, which are the languages included in the sentiment analysis experiment. We fine-tune LMs on the training sets of each language, which were translated automatically from English and released in the original paper." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Evaluation metrics. For the evaluation, we use the following standard metrics: answer span F1 score (Ans-F1) and exact match (EM) are used for QA; METEOR (MTR) and BERTScore (BS) for QG, which have been shown to be the most correlated metrics to human judgment (Ushio et al., 2022); macro-F1 score for sentiment following (Barbieri et al., 2022); and accuracy for NLI. As the language-specific corpus \\(\\mathcal{C}_l\\) to extract vocabulary counts for VT, we use mC4 (Xue et al., 2021), one of the largest public multilingual corpora." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.88, + 0.489, + 0.92 + ], + "angle": 0, + "content": "2In theory, vocabulary trimming could be applied to any language model, even monolingual, but this analysis is out of the scope of this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14727" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.88, + 0.306 + ], + "angle": 0, + "content": "
Lang.VocabularyParametersQAQG
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
mT5EN209K (83.6%)258M (86.1%)70.1 / 55.570.2 / 55.570.1 / 56.423.8 / 90.023.8 / 90.024.0 / 90.1
ES131K (52.4%)178M (59.4%)55.9 / 34.755.9 / 34.757.8 / 37.522.7 / 84.122.7 / 84.122.3 / 84.2
FR131K (52.4%)178M (59.4%)50.0 / 30.950.0 / 30.948.6 / 29.417.5 / 80.717.5 / 80.716.1 / 79.2
IT111K (44.4%)157M (52.6%)53.2 / 37.653.4 / 37.851.5 / 36.017.6 / 80.817.6 / 80.817.5 / 80.6
JA125K (50.0%)172M (57.6%)65.7 / 65.765.7 / 65.763.0 / 63.029.0 / 80.929.0 / 80.928.6 / 81.0
KO73K (29.2%)119M (39.7%)77.1 / 70.677.1 / 70.574.5 / 67.327.5 / 82.927.5 / 83.028.0 / 83.7
RU147K (58.8%)195M (65.1%)73.7 / 51.473.8 / 51.474.8 / 53.426.4 / 84.326.4 / 84.328.9 / 86.4
mBARTEN173K (69.2%)532M (87.1%)76.9 / 62.677.0 / 62.778.4 / 65.725.1 / 90.425.1 / 90.424.7 / 90.1
ES87K (34.8%)443M (72.7%)64.1 / 42.264.5 / 42.863.7 / 43.922.9 / 83.622.8 / 83.622.8 / 84.0
FR85K (34.0%)442M (72.5%)60.4 / 39.361.0 / 39.866.4 / 45.119.8 / 81.719.8 / 81.718.4 / 79.7
IT67K (26.8%)424M (69.5%)64.7 / 50.064.9 / 50.265.8 / 49.818.0 / 80.617.9 / 80.718.9 / 81.1
JA77K (30.8%)434M (71.1%)68.2 / 68.268.2 / 68.270.6 / 70.630.0 / 82.329.7 / 82.129.1 / 80.8
KO46K (18.4%)402M (65.9%)79.3 / 72.379.2 / 72.183.2 / 77.330.2 / 83.930.3 / 84.030.2 / 83.8
RU99K (39.6%)456M (74.8%)78.7 / 58.079.0 / 58.275.5 / 49.929.3 / 87.228.7 / 87.028.3 / 86.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.317, + 0.884, + 0.375 + ], + "angle": 0, + "content": "Table 1: Results on QA (Ans-F1/EM) and QG (MTR/BS), including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the parameter size of the original mT5 and mBART (No-Trim) is 300M and 611M, respectively, both with a vocabulary size of 250K." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.4, + 0.489, + 0.56 + ], + "angle": 0, + "content": "Base language models. As the base LMs, given computational constraints we chose the smallest mT5 and mBART to fine-tune on QA and QG, and XLM-R and XLM-V (Liang et al., 2023) for sentiment analysis and NLI. All these models have a vocabulary size of \\(250\\mathrm{K}\\), except for XLM-V which has a vocabulary size of \\(901\\mathrm{K}\\) subword tokens. For our experiments, we compare the results of pre/post-FT VT against vanilla LM fine-tuning without VT, which we refer to as No-Trim." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.572, + 0.489, + 0.685 + ], + "angle": 0, + "content": "Fine-tuning. For model fine-tuning, we rely on lmqg (Ushio et al., 2023) for QA/QG, and Ray Tune3 for sentiment analysis. In both cases, we use the default search space for hyperparameter search. For NLI, we follow the same hyperparameters used in Liang et al. (2023). All the resulting models and code will be released upon acceptance of the paper." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.697, + 0.22, + 0.712 + ], + "angle": 0, + "content": "4.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.719, + 0.489, + 0.768 + ], + "angle": 0, + "content": "We present the results for the generation and classification tasks in section 4.2.1 and section 4.2.2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.778, + 0.405, + 0.794 + ], + "angle": 0, + "content": "4.2.1 Generation Tasks: QA & QG" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.798, + 0.489, + 0.894 + ], + "angle": 0, + "content": "Table 1 shows the overall results on QA and QG. The results confirm that both of pre/post-FT VT can maintain the original performance in most cases, while being smaller than the original models by significantly reducing the vocabulary size. First, post-FT VT achieves at least the same performance as" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.4, + 0.884, + 0.593 + ], + "angle": 0, + "content": "the vanilla fine-tuning for all the languages for both LMs in QA and QG, except for a few cases such as mBART QA in Korean and mBART QG in Russian, although the decrease is no more than \\(0.5\\%\\). Meanwhile, pre-FT VT outperforms its vanilla fine-tuning model with a relatively important margin in some cases, such as mBART French QA and mT5 Spanish QA. In contrast, there are a few models where pre-FT VT degrades the performance of the original model such as mT5 QA in Korean (2.6% decrease in Ans-F1) or mBART QA in Russian (3.2% decrease in Ans-F1)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.595, + 0.884, + 0.723 + ], + "angle": 0, + "content": "Since we keep all the vocabulary that appeared in the language-specific corpus \\(\\mathcal{C}_l\\), the percentage of reduced parameter depends on the language, and generally VT can reduce the model size for Asian (Japanese/Korean) and European (Spanish/French/Italian) languages efficiently (50% for mT5 and 70% for mBART), but it remains high in other languages (English-Russian)." + }, + { + "type": "title", + "bbox": [ + 0.508, + 0.737, + 0.873, + 0.752 + ], + "angle": 0, + "content": "4.2.2 Classification Tasks: Sentiment & NLI" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Table 2 shows the results on sentiment analysis and NLI. In this case, post-FT VT can robustly preserve the original performance of the original No-Trim baseline in both tasks for XLM-R and XLM-V, while being no more than \\(40\\%\\) and \\(60\\%\\) in vocabulary and overall parameter size, respectively, of the original XLM-V and XLM-R models in all the non-English datasets. XLM-V PT sentiment model is the only post-FT VT where a slight decrease can be observed \\((0.1\\%)\\). On the other hand, the accuracy" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.486, + 0.919 + ], + "angle": 0, + "content": "3https://docs.ray.io/en/latest/tune/index.html" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14728" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.082, + 0.847, + 0.306 + ], + "angle": 0, + "content": "
Lang.VocabularyParameterSentimentNLI
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
XLM-RAR49K (19.6%)124M (44.7%)66.366.360.975.775.773.8
DE91K (36.4%)156M (56.3%)73.273.373.579.979.978.3
EN173K (69.2%)219M (78.7%)68.468.567.984.684.670.6
ES87K (34.8%)153M (55.0%)69.069.065.079.879.867.2
FR85K (34.0%)151M (54.6%)71.871.872.180.180.179.6
IT67K (26.8%)138M (49.7%)62.962.970.8---
PT66K (26.4%)137M (49.3%)70.770.870.2---
XLM-VAR92K (11.8%)157M (20.2%)59.859.864.775.575.676.1
DE239K (30.7%)269M (34.7%)73.573.573.078.978.979.0
EN484K (62.2%)458M (58.9%)63.963.961.384.484.484.5
ES243K (31.2%)279M (35.1%)60.760.766.680.780.780.6
FR218K (28.0%)253M (32.6%)68.868.859.578.678.679.0
IT184K (23.7%)227M (29.3%)70.270.274.2---
PT181K (23.3%)225M (28.9%)66.666.552.8---
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.318, + 0.883, + 0.377 + ], + "angle": 0, + "content": "Table 2: Results of sentiment analysis (macro F1) and XNLI (accuracy) including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the overall parameter size of the original XLM-R and XLM-V (No-Trim) is 278M and 778M, respectively, with the vocabulary size being 250K and 901K vocabulary in each case." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.4, + 0.49, + 0.593 + ], + "angle": 0, + "content": "of Pre-FT VT appears to be sensitive to the language and task, where it improves the performance in some languages such as Italian (XLM-R and XLM-V achieve \\(7.9\\%\\) and \\(3.8\\%\\) increase for sentiment analysis), but it decreases the performance with non-trivial margin in other languages such as Arabic, where XLM-R decreases \\(5\\%\\) for sentiment analysis and \\(2\\%\\) for XNLI. Since XLM-V has a larger vocabulary size, the percentage of reduced parameters at VT is more prominent in XLM-V, as seen in Arabic \\((20.2\\%)\\) and Portuguese \\((28.9\\%)\\) for example." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.604, + 0.367, + 0.621 + ], + "angle": 0, + "content": "5 Vocabulary Size Analysis" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.629, + 0.49, + 0.886 + ], + "angle": 0, + "content": "In our main experimental results (§ 4.2), all the unique tokens that appeared in the monolingual corpus were kept, which resulted in a low compression ratio for some languages such as English and Russian. In this analysis, we constrain the number of vocabulary and choose the top-\\(n\\) vocabulary at VT in terms of the frequency in the corpus (see § 3). For QA and QG, we compare mT5SMALL results with \\(n\\) from [5K, 10K, 15K, 30K, 60K, 90K, 120K], which correspond to an overall parameter size of [49M, 54M, 59M, 74M, 105M, 136M, 166M], respectively. For sentiment analysis and NLI, we experiment with XLM-RBASE with \\(n\\) from [5K, 10K, 15K, 30K, 60K], which correspond to [89M, 93M, 97M, 109M, 132M] of parameter size, respectively4." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "4The full results with top- \\(n\\) can be found at Appendix B and Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.4, + 0.787, + 0.416 + ], + "angle": 0, + "content": "5.1 Generation Tasks: QA & QG" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.421, + 0.886, + 0.822 + ], + "angle": 0, + "content": "Figure 4 shows the results of mT5 on QA and QG. Noticeably, post-FT VT can reduce the vocabulary size to \\(60\\mathrm{K}\\) for both QA and QG in all the languages with a trivial gap (0.3% decrease of EM in Russian QA and 0.1% decrease of BS in French QG), and that is \\(35\\%\\) of the original mT5 in the parameter size. Furthermore, post-FT VT can further reduce the vocabulary to 5K tokens with no more than \\(0.4\\%\\) decrease in each metric for both QA and QG in English, French, and Italian, which is \\(16\\%\\) of the original mT5 in the parameter size. Meanwhile, pre-FT VT outperforms the No-Trim result in all the languages in QA, and the majority of the languages in QG (English, Italian, Korean, and Russian), but the result is sensitive to the choice of \\(n\\). For example, Japanese/Korean QA and Russian QG with pre-FT VT for top-5K (16% of the original mT5) outperforms No-Trim as well as post-FT VT, but Japanese QG with pre-FT VT is worse in any choice of \\(n\\) on contrary. This larger variation of results may also be due to the parameter size space, as the optimal parameters for the original multilingual LM (which is the one trained for post-FT VT) may differ. We leave this extended analysis for future work." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.834, + 0.86, + 0.849 + ], + "angle": 0, + "content": "5.2 Classification Tasks: Sentiment & NLI" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Figure 5 and Figure 6 show the results of XLM-R on sentiment and NLI. In NLI, we can see that post/pre-FT VT both can reduce the vocabulary to \\(30\\mathrm{K}\\) (\\(39\\%\\) of the original XLM-R in the parameter" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.083, + 0.301, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.083, + 0.484, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.152, + 0.3, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.152, + 0.484, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.22, + 0.3, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.22, + 0.484, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.289, + 0.3, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.289, + 0.484, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.358, + 0.3, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.358, + 0.484, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.426, + 0.3, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.426, + 0.484, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.494, + 0.3, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.494, + 0.484, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.577, + 0.489, + 0.622 + ], + "angle": 0, + "content": "Figure 4: QG (METEOR) and QA (Ans-F1) results for mT5 with pre/post-FT VT for different vocabulary sizes compared to the original multilingual LM (No-Trim)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.489, + 0.822 + ], + "angle": 0, + "content": "size) without any decrease except \\(0.3\\%\\) of pre-FT VT for German, and there is no decrease more than \\(0.4\\%\\) even with top-15K of post-FT VT. In sentiment analysis, pre-FT VT with top-10K (\\(33\\%\\) of the original XLM-R in the parameter size) can retain the accuracy of the No-Trim baseline in French and Italian. Moreover, post-FT VT with 30K can retain the original F1 score without a major drop in sentiment analysis, yet the decrease in F1 score is slightly more prominent than NLI (\\(1.1\\%\\) in Arabic sentiment analysis)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.489, + 0.919 + ], + "angle": 0, + "content": "The sentiment analysis datasets are collected from Twitter, so one dataset in a single language can contain tokens from other languages ( hashtags or named-entities, or even code-switching). In contrast, XNLI translates English NLI into other languages, so there is less chance for a dataset to" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.083, + 0.701, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.083, + 0.881, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.153, + 0.701, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.153, + 0.881, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.223, + 0.701, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.223, + 0.881, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.292, + 0.701, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.376, + 0.885, + 0.419 + ], + "angle": 0, + "content": "Figure 5: Sentiment analysis macro-F1 results of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.433, + 0.697, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.433, + 0.881, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.498, + 0.697, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.499, + 0.881, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.564, + 0.697, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.636, + 0.883, + 0.666 + ], + "angle": 0, + "content": "Figure 6: NLI accuracy of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.69, + 0.884, + 0.787 + ], + "angle": 0, + "content": "contain tokens from other languages. This can explain the effectiveness of top-\\(n\\) VT in NLI compared with sentiment analysis, as smaller values of \\(n\\) should result in a vocabulary with fewer tokens from the other languages, which limits the ability of the models to handle foreign tokens." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.797, + 0.883, + 0.829 + ], + "angle": 0, + "content": "6 Monolingual vs. Multilingual LMs: The Case of Social Bias" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.885, + 0.919 + ], + "angle": 0, + "content": "There has been extensive literature in NLP comparing monolingual and multilingual LMs (Muller et al., 2021; Goyal et al., 2021). As for the performance, there is no clear consensus on which type is better for certain languages, tasks or settings. How-" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.538 + ], + "angle": 0, + "content": "ever, there are other important factors that play a role in this comparison. First, monolingual models tend to have a smaller vocabulary size, which makes them more practical. In contrast, a single multilingual model can be used for a large number of languages. Moreover, multilingual LMs are less prone to capture and carry cultural- or language-dependent biases. This is due to the combination of languages and cultures into a single model, which makes it less biased toward specific cultures (Liang et al., 2020; Ahn and Oh, 2021). Prior works have shown that different types of biases consistently appear in language-specific models (Nadeem et al., 2021; Nangia et al., 2020; Blodgett et al., 2021; Dhamala et al., 2021; Kaneko et al., 2022; Zhou et al., 2022). While the comparison of monolingual and multilingual LMs is not the main focus of this paper, this analysis is certainly relevant. Trimming the vocabulary of a multilingual model essentially makes the model smaller, and therefore alleviates one of the main drawbacks of using multilingual language models on language-specific tasks, which is its larger size. On top of that, this strategy enables the usage of monolingual models with potentially less social bias. In the following, we present a comparison of monolingual and multilingual LMs (both trimmed and not trimmed) in terms of social bias and general performance." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.557, + 0.327, + 0.572 + ], + "angle": 0, + "content": "6.1 Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.583, + 0.49, + 0.873 + ], + "angle": 0, + "content": "Social bias datasets. To study the effect of VT on social bias existing in pre-trained LMs, we first conduct experiments on two commonly used social bias evaluation datasets for masked LMs: StereoSet (SS; Nadeem et al., 2021)\\(^5\\) and crowdsourced stereotype pairs benchmark (CP; Nangia et al., 2020)\\(^6\\). SS consists of associative contexts covering four types of social biases: race, gender, religion, and profession; while CP is crowdsourced and annotated by workers in the United States, which contains nine types of social biases: race, gender, sexual orientation, religion, age, nationality, disability, physical appearance, and socioeconomic status. In order to further investigate the impact of pre/post-FT VT on LMs, we trim and fine-tune models on sentiment analysis with different orders and evaluate the social bias in such models on the Equity Evaluation Corpus (EEC; Kiritchenko and" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.084, + 0.885, + 0.15 + ], + "angle": 0, + "content": "Mohammad, 2018) considering two bias types: gender and race. The EEC dataset was specifically with the aim to examine social bias for sentiment analysis systems." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.161, + 0.886, + 0.725 + ], + "angle": 0, + "content": "Evaluation metrics. We compare the pseudolikelihood scores returned by each model for stereotypical and anti-stereotypical sentences using AULA (All Unmasked Likelihood with Attention weights) (Kaneko and Bollegala, 2022).8 AULA has been shown to be robust against the frequency biases of the masked tokens and provides a more reliable assessment in contrast to alternative metrics when evaluating social biases in masked language models (MLMs). Given a sentence pair in the test dataset: \"My mom spent all day cooking for Thanksgiving\" vs. \"My dad spent all day cooking for Thanksgiving\", the first sentence is considered as stereotypical while the second one is anti-stereotypical. AULA computes the percentage of stereotypical sentences preferred by the MLM over anti-stereotypical ones as the bias score. An MLM is considered to be unfairly biased if it returns higher pseudo-loglikelihood scores for stereotypical sentences than the corresponding anti-stereotypical sentences. The AULA score falls within the range of [0,100] and an unbiased model would return a bias score close to 50. On the other hand, a bias score greater than or less than 50 indicates the bias direction towards the stereotype or anti-stereotype, respectively. Since the original AULA is not fitted to evaluate fine-tuned models, we adapt AULA to the EEC dataset obtain the bias score for the LMs fine-tuned on sentiment analysis, and denote this metric as EEC-AULA. Specifically, given a model that predicts sentiment labels (e.g., positive, neutral, negative) to sentences, we consider the percentage of stereotypical test sentences with a more negative label over anti-stereotypical ones as the corresponding bias evaluation measure." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.737, + 0.885, + 0.833 + ], + "angle": 0, + "content": "General performance. As a proxy to test the general performance, we use the general language understanding evaluation (GLUE; Wang et al., 2018) benchmark. We acknowledge the limitations of using this benchmark to draw reliable conclusions at large (Ethayarajh and Jurafsky, 2020)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.844, + 0.882, + 0.867 + ], + "angle": 0, + "content": "7https://saifmohammad.com/WebPages/Biases-SA.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.868, + 0.883, + 0.892 + ], + "angle": 0, + "content": "https://github.com/kanekomasahiro/evaluate.bias_in_m1m" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.884, + 0.919 + ], + "angle": 0, + "content": "\\(^{9}\\)https://gluebenchmark.com/; Models are tested on the development sets of each task." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.844, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.891, + 0.441, + 0.905 + ], + "angle": 0, + "content": "5https://github.com/moinnadeem/StereoSet" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.905, + 0.435, + 0.919 + ], + "angle": 0, + "content": "6https://github.com/nyu-mll/crows-pairs" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.891, + 0.441, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "14731" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.88, + 0.23 + ], + "angle": 0, + "content": "
Base ModelVocab TrimmingVocabularySocial BiasGeneral Performance
Pre-FTPost-FTAULAEEC-AULAGLUE
CPSSGenderRace
Monolingual (RoBERTa)-50K50K58.158.864.885.778.0
Multilingual (XLM-R)-250K250K49.554.944.362.077.9
Pre-FT VT (EN)173K173K49.554.942.556.978.0
Post-FT VT (EN)250K173k49.554.944.362.077.9
Pre-FT VT (50K)50K50K49.355.041.062.476.9
Post-FT VT (50K)250K50K49.554.944.362.077.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.113, + 0.24, + 0.885, + 0.342 + ], + "angle": 0, + "content": "Table 3: Results of pre/post-FT VT models compared with the original monolingual and multilingual models on two social bias analysis benchmarks (AULA for pre-trained masked language models and EEC-AULA for models fine-tuned on sentiment analysis) and the GLUE benchmark. The VT models are trimmed on English vocabulary with different vocabulary sizes: EN (full English vocabulary) and 50K (top 50K subword tokens). Note that for post-FT VT, the results on AULA are exactly the same as the original XLM-R. The green and red colours represent the social bias towards anti-stereotypical sentences (scores lower than 50) and stereotypical sentences (scores higher than 50), respectively. The lighter colour indicates less social bias observed in the LM." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.366, + 0.489, + 0.447 + ], + "angle": 0, + "content": "but it nevertheless provides a good proxy for understanding the overall performance of comparable models in standard NLP tasks. Moreover, these experiments are only aimed at analysing the effect of vocabulary trimming on general performance." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.454, + 0.49, + 0.712 + ], + "angle": 0, + "content": "Models. We compute the bias scores of RoBERTa (Liu et al., 2019) as base monolingual LM, and XLM-R (Conneau et al., 2019) as its multilingual counterpart (they have been trained with the same architecture and in an overlapping corpus. We explore two VT settings to be applied to XLM-R: XLM-R with the standard VT including the full English vocabulary (VT XLM-R) and XLM-R with VT for top-50K English vocabulary (top-50K VT XLM-R), which is the same vocabulary size as the monolingual RoBERTa model. Our experiments are based both on masked language models on AULA (in which the post-VT does not have any effect) and models fine-tuned on the sentiment analysis presented in \\(\\S 4.1\\) on EEC-AULA, as well as on the corresponding GLUE training sets." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.722, + 0.22, + 0.737 + ], + "angle": 0, + "content": "6.2 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Table 3 shows the performance of pre-FT and post-FT VT models against the original monolingual and multilingual LMs on social bias evaluation datasets and the GLUE benchmark. Both AULA and GLUE results are computed using the LMs without fine-tuning (i.e., RoBERTa, XLM-R, VT XLM-R, and top-50K VT XLM-R), whereas the EEC-AULA results are computed using the models applying VT and fine-tuning strategies. We observe that the monolingual model contains the highest levels of social biases compared to the multilin" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.366, + 0.885, + 0.591 + ], + "angle": 0, + "content": "gual models with different settings. In particular, RoBERTa obtains the overall highest bias score on the EEC dataset after fine-tuning, with an alarmingly high 85.7 score on race.\\(^{10}\\) On the other hand, compared to the original XLM-R, there is no significant change in performance on social biases and GLUE evaluation tasks for pre-FT VT and post-FT VT models. This is important as we can apply the proposed VT method to any multilingual LM, obtaining a monolingual one with consistent performance on the GLUE benchmark and less social biases than the original monolingual model pretrained in the target language, without using any ad-hoc debiasing methods." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.602, + 0.637, + 0.618 + ], + "angle": 0, + "content": "7 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.628, + 0.884, + 0.884 + ], + "angle": 0, + "content": "Vocabulary trimming before and after finetuning. According to the results, pre-FT VT appears to be generally more effective in classification tasks (see section 4.2.2). For generation tasks, both pre/post-FT VT robustly retain the original performance while being able to considerably reduce the model size (see section 4.2.1). As a guideline to choose the type of VT in such a case, post-FT VT should be more suitable if one already has a finetuned model, as no additional training is needed for this case. Moreover, post-FT is more robust as a compression mechanism as the performance is largely maintained with respect to that of the original multilingual LM. On the other hand, if one needs to fine-tune a model from scratch and the computation resources are limited, we recommend" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.883, + 0.918 + ], + "angle": 0, + "content": "10 Appendix C includes a breakdown by type of the social bias results." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.198 + ], + "angle": 0, + "content": "exploring pre-FT VT, as fine-tuning on a trimmed LM should be more efficient due to its smaller vocabulary and parameters and, in some cases, can lead to better overall results. However, this process has to be done carefully as the set of optimal parameters could differ from the original multilingual LM fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.207, + 0.49, + 0.609 + ], + "angle": 0, + "content": "While in this paper we have not compared monolingual and multilingual models, the question would be whether we need vocabulary trimming strategies in a world where monolingual LMs exist. In this case, a monolingual model may perform similarly to a multilingual model (Goyal et al., 2021). However, the multilingual model is often larger mainly due to larger vocabulary storage requirements. In contrast, our proposed VT technique does not require any extra LM training or computational resources. Indeed, only a multilingual LM is needed and we can induce multiple smaller language-specific monolingual models. This may reduce the carbon footprint overall and especially help with less-resource languages when a high-quality monolingual model does not exist. Finally, our social bias analysis see § 6 shows how monolingual models exhibit larger social biases (especially racial) than a VT-induced multilingual LM. This is consistent with prior work suggesting that a multilingual LM has been trained with more languages, and hence more cultural variety, and these diverging viewpoints can compensate each other (Ahn and Oh, 2021)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.62, + 0.248, + 0.636 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In this paper, we proposed vocabulary-trimming (VT), a method to reduce the vocabulary of a multilingual LM to a vocabulary specific to any given target language. VT can induce a monolingual LM in the target language by leveraging an existing multilingual LM. The main advantage of this filtering step is the reduced size, as well as avoiding having to train monolingual LMs from scratch, which would be computationally demanding. Our experiments show how VT can retain the high performance of the original multilingual LM, while largely reducing the model size. For all languages evaluated, a \\(35\\%\\) compression rate proves sufficient to keep the original performance of the larger mT5 multilingual LM in both QA and QG, with a similar \\(39\\%\\) in NLI and \\(55\\%\\) in sentiment analysis with XLM-R. Interestingly, in some cases, the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.181 + ], + "angle": 0, + "content": "compressed LM can even achieve better results than the original larger model when trimmed before fine-tuning. Since the main goal of the paper was to compress a multilingual LM while keeping its original performance, we leave the analysis of this behaviour for future work." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.194, + 0.616, + 0.21 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.221, + 0.885, + 0.445 + ], + "angle": 0, + "content": "We have not tested our methodology in truly low-resource languages. Because of this, there could be a different behaviour when we apply VT to a language with lower resources or that is poorly represented in the underlying training corpus. The LMs we used in the paper limited their size up to 600M, and we have not considered larger models such as mT5XXL or BLOOM (Scao et al., 2022), due to our limited computational resources. As the language-specific corpus to compute frequency, we employ mC4, which is one of the largest multilingual corpora. Nonetheless, this is used as a proxy and having access to the full multilingual model could give potentially better results." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.447, + 0.885, + 0.575 + ], + "angle": 0, + "content": "Similarly, we acknowledge the limitations of the analysis comparing multilingual and monolingual models in terms of social bias. Due to the evaluation data available and the existence of comparable monolingual and multilingual LMs, the evaluation is focused on English only and the results could differ for other languages. Moreover, there are other types of biases not covered in this evaluation." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.588, + 0.662, + 0.603 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.885, + 0.743 + ], + "angle": 0, + "content": "Pre-trained LMs are known to contain undesirable biases to generate toxic contents in some edge cases (Schick et al., 2021), so the resulting models could inherit such biases. While we have not analysed in detail the output of all models in the tasks evaluated, in this paper we have made an attempt to study this effect in terms of social biases for both base pretrained LMs and fine-tuned LMs." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.756, + 0.685, + 0.772 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.782, + 0.882, + 0.814 + ], + "angle": 0, + "content": "Jose Camacho-Collados and Yi Zhou are supported by a UKRI Future Leaders Fellowship." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.842, + 0.61, + 0.858 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.866, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Amine Abdaoui, Camille Pradel, and Grégoire Sigel. 2020. Load what you need: Smaller versions of mutililingual BERT. In Proceedings of SustainNLP: Workshop on Simple and Efficient Natural Language" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14733" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.113 + ], + "angle": 0, + "content": "Processing, pages 119-123, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.124, + 0.487, + 0.203 + ], + "angle": 0, + "content": "Jaimeen Ahn and Alice Oh. 2021. Mitigating language-dependent ethnic bias in BERT. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 533-549, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.213, + 0.487, + 0.292 + ], + "angle": 0, + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.303, + 0.487, + 0.407 + ], + "angle": 0, + "content": "Francesco Barbieri, Valerio Basile, Danilo Croce, Malvina Nissim, Nicole Novielli, and Viviana Patti. 2016. Overview of the evalita 2016 sentiment polarity classification task. In Proceedings of third Italian conference on computational linguistics (CLiC-it 2016) & fifth evaluation campaign of natural language processing and speech tools for Italian. Final Workshop (EVALITA 2016)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.418, + 0.487, + 0.509 + ], + "angle": 0, + "content": "Francesco Barbieri, Luis Espinosa Anke, and Jose Camacho-Collados. 2022. XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 258-266, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.521, + 0.487, + 0.573 + ], + "angle": 0, + "content": "Farah Benamara, Cyril Grouin, Jihen Karoui, Véronique Moriceau, and Isabelle Robba. 2017. Analyse d'opinion et langage figuratif dans des tweets: presentation et résultats du début fouille de textes deft2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.584, + 0.487, + 0.689 + ], + "angle": 0, + "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1004-1015." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.7, + 0.487, + 0.751 + ], + "angle": 0, + "content": "Henrico Bertini Brum and Maria das Graças Volpe Nunes. 2017. Building a sentiment corpus of tweets in brazilian portuguese. arXiv preprint arXiv:1712.08917." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.763, + 0.487, + 0.816 + ], + "angle": 0, + "content": "Carrino Casimiro Pio, Costa-jussa Marta R., and Fonollosa Jose A. R. 2019. Automatic Spanish Translation of the SQuAD Dataset for Multilingual Question Answering. arXiv e-prints, page arXiv:1912.05200v1." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.178 + ], + "angle": 0, + "content": "Mark Cieliebak, Jan Milan Deriu, Dominic Egger, and Fatih Uzdilli. 2017. A twitter corpus and benchmark resources for german sentiment analysis. In 5th International Workshop on Natural Language Processing for Social Media, Boston MA, USA, 11 December 2017, pages 45-51. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.19, + 0.882, + 0.269 + ], + "angle": 0, + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.281, + 0.882, + 0.32 + ], + "angle": 0, + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.333, + 0.882, + 0.437 + ], + "angle": 0, + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.449, + 0.882, + 0.515 + ], + "angle": 0, + "content": "Danilo Croce, Alexandra Zelenanska, and Roberto Basili. 2018. Neural learning for question answering in italian. In AI*IA 2018 – Advances in Artificial Intelligence, pages 389–402, Cham. Springer International Publishing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.527, + 0.882, + 0.645 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.657, + 0.882, + 0.747 + ], + "angle": 0, + "content": "Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. Bold: Dataset and metrics for measuring biases in open-ended language generation. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 862-872." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.761, + 0.882, + 0.84 + ], + "angle": 0, + "content": "Martin d'Hoffschmidt, Wacim Belblidia, Quentin Heinrich, Tom Brendle, and Maxime Vidal. 2020. FQuAD: French question answering dataset. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1193–1208, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.851, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Manuel C Díaz-Galiano, Eugenio Martínez-Cármara, M Ángel García Cumbreras, Manuel García Vega, and Julio Villena Román. 2018. The democratization of deep learning in tass 2017. Procesamento del Lenguaje Natural, 60:37-44." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.166 + ], + "angle": 0, + "content": "Pavel Efimov, Andrey Chertok, Leonid Boytsov, and Pavel Braslavski. 2020. Sberquad-russian reading comprehension dataset: Description and analysis. In International Conference of the Cross-Language Evaluation Forum for European Languages, pages 3-15. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.178, + 0.487, + 0.258 + ], + "angle": 0, + "content": "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.271, + 0.487, + 0.363 + ], + "angle": 0, + "content": "Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, and Alexis Conneau. 2021. Larger-scale transformers for multilingual masked language modeling. In Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021), pages 29–33, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.376, + 0.487, + 0.455 + ], + "angle": 0, + "content": "Masahiro Kaneko and Danushka Bollegala. 2021. Debiasing pre-trained contextualised embeddings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1256-1266, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.468, + 0.487, + 0.534 + ], + "angle": 0, + "content": "Masahiro Kaneko and Danushka Bollegala. 2022. Unmasking the mask-evaluating social biases in masked language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11954-11962." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.548, + 0.487, + 0.653 + ], + "angle": 0, + "content": "Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, and Naoaki Okazaki. 2022. Gender bias in masked language models for multiple languages. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2740-2750, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.666, + 0.487, + 0.72 + ], + "angle": 0, + "content": "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.733, + 0.487, + 0.773 + ], + "angle": 0, + "content": "Svetlana Kiritchenko and Saif M Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. *NAACL HLT* 2018, page 43." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.786, + 0.487, + 0.853 + ], + "angle": 0, + "content": "Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166-172." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In International Conference on Learning Representations." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.487, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.204 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.216, + 0.882, + 0.283 + ], + "angle": 0, + "content": "Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, and Madian Khabsa. 2023. Xlm-v: Overcoming the vocabulary bottleneck in multilingual masked language models. arXiv preprint arXiv:2301.10472." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.294, + 0.882, + 0.386 + ], + "angle": 0, + "content": "Sheng Liang, Philipp Dufter, and Hinrich Schütze. 2020. Monolingual and multilingual reduction of gender bias in contextualized representations. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5082-5093, Barcelona, Spain (Online). International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.398, + 0.882, + 0.438 + ], + "angle": 0, + "content": "Seungyoung Lim, Myungji Kim, and Jooyoul Lee. 2019. Korquad1.0: Korean qa dataset for machine reading comprehension. arXiv preprint arXiv:1909.07005." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.449, + 0.882, + 0.528 + ], + "angle": 0, + "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.541, + 0.882, + 0.607 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.618, + 0.882, + 0.684 + ], + "angle": 0, + "content": "Kelly Marchisio, Patrick Lewis, Yihong Chen, and Mikel Artetxe. 2022. Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training. arXiv preprint arXiv:2212.10503." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.696, + 0.882, + 0.789 + ], + "angle": 0, + "content": "Chandler May, Alex Wang, Shikha Bordia, Samuel Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 622-628." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.801, + 0.882, + 0.919 + ], + "angle": 0, + "content": "Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.192 + ], + "angle": 0, + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.2, + 0.49, + 0.293 + ], + "angle": 0, + "content": "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953-1967, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.3, + 0.489, + 0.354 + ], + "angle": 0, + "content": "Malte Ostendorff and Georg Rehm. 2023. Efficient language model training through cross-lingual and progressive transfer learning. arXiv preprint arXiv:2301.09626." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.362, + 0.49, + 0.481 + ], + "angle": 0, + "content": "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.488, + 0.49, + 0.568 + ], + "angle": 0, + "content": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.577, + 0.49, + 0.655 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.664, + 0.49, + 0.744 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.752, + 0.49, + 0.831 + ], + "angle": 0, + "content": "Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 task 4: Sentiment analysis in Twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502-518, Vancouver, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.84, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.151 + ], + "angle": 0, + "content": "Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP. Transactions of the Association for Computational Linguistics, 9:1408-1424." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.161, + 0.885, + 0.214 + ], + "angle": 0, + "content": "ByungHoon So, Kyuhong Byun, Kyungwon Kang, and Seongjin Cho. 2022. Jaquad: Japanese question answering dataset for machine reading comprehension. arXiv preprint arXiv:2202.01764." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.222, + 0.885, + 0.315 + ], + "angle": 0, + "content": "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2022. Generative language models for paragraph-level question generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 670-688, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.323, + 0.885, + 0.415 + ], + "angle": 0, + "content": "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2023. A practical toolkit for multilingual question and answer generation. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.424, + 0.885, + 0.515 + ], + "angle": 0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.524, + 0.885, + 0.616 + ], + "angle": 0, + "content": "Hai Wang, Dian Yu, Kai Sun, Jianshu Chen, and Dong Yu. 2019. Improving pre-trained multilingual model with vocabulary expansion. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 316-327, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.625, + 0.885, + 0.73 + ], + "angle": 0, + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.738, + 0.885, + 0.844 + ], + "angle": 0, + "content": "Bo Zheng, Li Dong, Shaohan Huang, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, and Furu Wei. 2021. Allocating large vocabulary capacity for crosslingual language model pre-training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3203-3215, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.852, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Yi Zhou, Masahiro Kaneko, and Danushka Bollegala. 2022. Sense embeddings are also biased - evaluating social biases in static and contextualised sense embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14736" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.083, + 0.501, + 0.445 + ], + "angle": 0, + "content": "
Vocab.No-Trim (250K)5K10K15K30K60K
Param.278M89M93M97M109M132M
Post-FT (Sentiment)AR66.364.564.565.965.9
DE73.270.472.472.173.7
EN68.464.066.567.468.6
ES69.066.267.267.868.4
FR71.868.671.471.771.6
IT62.960.861.963.562.3
PT70.763.665.968.269.4
Pre-FT (Sentiment)AR66.358.661.963.163.2
DE73.270.671.873.171.5
EN68.464.666.667.766.0
ES69.060.465.666.365.9
FR71.870.374.373.872.2
IT62.966.267.168.168.2
PT70.765.567.569.667.9
Post-FT (NLI)AR75.775.075.875.875.7
DE79.976.678.979.679.9
EN84.680.583.684.384.6
ES79.875.378.479.479.8
FR80.177.280.080.180.2
Pre-FT (NLI)AR75.773.075.075.475.7
DE79.977.479.079.079.6
EN84.683.384.784.485.1
ES79.878.979.179.281.0
FR80.177.378.978.280.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.456, + 0.489, + 0.514 + ], + "angle": 0, + "content": "Table 4: Results of XLM-R for sentiment analysis (macro F1) and NLI (accuracy) with different top-\\(n\\) vocabulary size at VT, where the best results in each LM and language are in the bold characters." + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.54, + 0.489, + 0.568 + ], + "angle": 0, + "content": "(Volume 1: Long Papers), pages 1924-1935, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.58, + 0.335, + 0.596 + ], + "angle": 0, + "content": "A Top- \\(n\\) VT of XLM-R" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.606, + 0.488, + 0.655 + ], + "angle": 0, + "content": "Table 4 shows the results of XLM-R fine-tuned on sentiment and NLI with post/pre-VT for different top-\\(n\\)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.665, + 0.306, + 0.682 + ], + "angle": 0, + "content": "B Top- \\(n\\) VT of mT5" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.691, + 0.489, + 0.723 + ], + "angle": 0, + "content": "Table 5 shows the results of mT5 fine-tuned on QA and QG with post/pre-VT for different top- \\( n \\) ." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.735, + 0.429, + 0.766 + ], + "angle": 0, + "content": "C Details of Results on Social Bias Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.777, + 0.49, + 0.873 + ], + "angle": 0, + "content": "Table 6 shows the details of social bias evaluation (EEC dataset) regarding each emotion type observed in the LMs fine-tuned on sentiment analysis. Table 7 shows the details of social bias regarding each bias type in both CP and SS datasets observed in the comparison LMs." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14737" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.106, + 0.867, + 0.523 + ], + "angle": 0, + "content": "
Vocab. \nParam.No-Trim (250K) \n300M5K \n49M10K \n54M15K \n59M30K \n74M60K \n105M90K \n136M120K \n166M
EN70.1 / 55.572.4 / 59.773.2 / 60.573.7 / 60.972.7 / 58.871.0 / 56.470.5 / 55.870.3 / 55.6
ES55.9 / 34.753.5 / 32.554.4 / 33.154.5 / 33.255.2 / 33.856.0 / 34.855.9 / 34.855.9 / 34.7
FR50.0 / 30.950.0 / 31.548.9 / 29.948.6 / 29.349.9 / 30.149.8 / 30.550.2 / 30.950.0 / 30.8
IT53.2 / 37.654.3 / 39.653.8 / 38.454.3 / 38.854.2 / 38.553.7 / 38.153.4 / 37.8-
JA65.7 / 65.754.4 / 54.460.3 / 60.362.6 / 62.660.4 / 60.464.9 / 64.965.6 / 65.665.7 / 65.7
KO77.1 / 70.675.0 / 67.875.8 / 68.776.2 / 69.376.9 / 70.277.3 / 70.7--
RU73.7 / 51.470.1 / 48.871.4 / 49.570.3 / 47.670.3 / 47.973.8 / 51.173.9 / 51.473.8 / 51.4
EN23.8 / 90.023.6 / 89.824.1 / 89.924.2 / 90.024.2 / 90.023.9 / 90.023.9 / 90.023.9 / 90.0
ES22.7 / 84.121.4 / 83.622.0 / 83.822.2 / 83.922.4 / 84.022.7 / 84.122.7 / 84.122.7 / 84.1
FR17.5 / 80.717.1 / 80.317.1 / 80.417.3 / 80.517.3 / 80.517.4 / 80.617.5 / 80.717.5 / 80.7
IT17.6 / 80.817.3 / 80.717.5 / 80.817.5 / 80.817.6 / 80.717.6 / 80.817.6 / 80.8-
JA29.0 / 80.925.7 / 79.227.4 / 80.128.2 / 80.528.8 / 80.929.0 / 80.929.0 / 80.929.0 / 80.9
KO27.5 / 82.927.3 / 82.827.4 / 82.927.4 / 82.927.4 / 82.927.5 / 83.0--
RU26.4 / 84.326.0 / 84.026.2 / 84.226.3 / 84.226.3 / 84.226.3 / 84.326.4 / 84.326.4 / 84.3
EN70.1 / 55.574.0 / 61.371.8 / 59.173.4 / 60.774.3 / 61.371.3 / 56.568.6 / 54.366.2 / 52.1
ES55.9 / 34.756.7 / 36.758.1 / 37.562.2 / 41.459.8 / 40.158.0 / 37.058.9 / 38.152.0 / 32.7
FR50.0 / 30.949.0 / 32.747.6 / 30.243.3 / 26.944.9 / 27.434.3 / 20.043.1 / 24.140.0 / 23.7
IT53.2 / 37.657.6 / 43.761.5 / 46.861.2 / 45.657.8 / 42.156.5 / 40.555.4 / 39.2-
JA65.7 / 65.755.3 / 55.352.2 / 52.261.3 / 61.360.4 / 60.464.6 / 64.666.9 / 66.967.2 / 67.2
KO77.1 / 70.679.5 / 72.683.7 / 77.782.7 / 76.480.4 / 73.781.6 / 75.1--
RU73.7 / 51.473.5 / 51.776.9 / 56.477.4 / 56.975.9 / 54.275.1 / 53.075.0 / 53.073.3 / 51.4
EN23.8 / 90.024.8 / 90.024.4 / 89.924.3 / 89.924.2 / 90.024.2 / 90.023.3 / 89.923.6 / 89.8
ES22.7 / 84.121.9 / 84.121.9 / 83.922.7 / 83.722.0 / 84.322.6 / 84.420.4 / 79.622.6 / 84.1
FR17.5 / 80.716.1 / 79.016.9 / 79.916.9 / 79.515.8 / 78.814.6 / 78.015.7 / 79.017.0 / 79.2
IT17.6 / 80.817.4 / 80.417.8 / 81.118.0 / 80.817.1 / 80.617.4 / 80.917.4 / 80.8-
JA29.0 / 80.926.6 / 79.827.0 / 79.227.5 / 79.727.3 / 79.927.7 / 80.328.3 / 80.828.2 / 80.2
KO27.5 / 82.928.4 / 83.427.8 / 83.428.2 / 84.128.8 / 83.428.4 / 83.4--
RU26.4 / 84.327.7 / 85.930.0 / 86.629.0 / 86.529.2 / 86.329.0 / 86.629.2 / 86.728.9 / 86.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.532, + 0.884, + 0.561 + ], + "angle": 0, + "content": "Table 5: Results of mT5 for QA (Ans-F1/EM) and QG (MTR/BS) with different top- \\( n \\) vocabulary size at VT,where the best results in each LM and language are in the bold characters." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.618, + 0.883, + 0.851 + ], + "angle": 0, + "content": "
ModelAULA-EECAngerSadnessFearJoyNo Emotion Type
Gender BiasFT RoBERTa64.850.067.871.564.360.8
FT XLM-R44.346.939.451.434.439.7
Pre-FT XLM-R (EN)42.533.855.845.140.742.7
Post-FT XLM-R (EN)44.346.939.451.434.439.7
Pre-FT XLM-R (50k)44.346.939.451.434.439.7
Post-FT XLM-R (50k)41.041.245.847.236.937.2
Race BiasFT RoBERTa85.750.088.065.450.0100.0
FT XLM-R62.074.532.147.374.617.4
Pre-FT VT XLM-R (EN)56.933.758.869.260.150.0
Pre-FT VT XLM-R (EN)62.074.532.147.374.617.4
Pre-FT VT XLM-R (50k)62.074.532.147.374.617.4
Post-FT VT XLM-R (50k)62.455.770.370.959.159.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.861, + 0.883, + 0.89 + ], + "angle": 0, + "content": "Table 6: Social bias scores of LMs fine-tuned on sentiment analysis on each emotion type w/wo VT on the EEC dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14738" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.338, + 0.846, + 0.62 + ], + "angle": 0, + "content": "
RoBERTaXLM-RVT XLM-R (EN)VT XLM-R (50K)
CP DatasetAULA58.149.549.549.3
Age59.849.449.449.4
Disability68.366.766.768.3
Gender53.448.548.548.1
Nationality56.644.044.044.7
Physical-Appearance52.458.758.755.6
Race-Color56.843.843.844.0
Religion53.354.354.353.3
Sexual-Orientation67.954.854.854.8
Socioeconomic66.358.158.157.6
SS DatasetAULA58.854.954.955.0
Gender62.455.355.354.9
Profession60.454.354.354.6
Race57.056.056.056.0
Religion54.446.846.846.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.629, + 0.881, + 0.66 + ], + "angle": 0, + "content": "Table 7: Social bias scores of VT LMs compared to their original ones on both CP and SS datasets for each bias type." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14739" + } + ] +] \ No newline at end of file diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_origin.pdf b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..77fa0d6af90c879909dd0a29c36ea71fc6485726 --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/f34b65cf-02f1-40ee-8ff5-d2a0c21b9d93_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f037ef5c1cc21f3798a61d242aaeee2c462072a5e653897a5bbfea129e9e14 +size 2923034 diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/full.md b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f405b858962b9934d35ab671858167f2d6f6bf9c --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/full.md @@ -0,0 +1,328 @@ +# Efficient Multilingual Language Model Compression through Vocabulary Trimming + +Asahi Ushio and Yi Zhou and Jose Camacho-Collados + +Cardiff NLP, School of Computer Science and Informatics, Cardiff University, UK + +{UshioA,Zhouy131,CamachoColladosJ}\@cardiff.ac.uk + +# Abstract + +Multilingual language models (LMs) have become a powerful tool in NLP, especially for non-English languages. Nevertheless, model parameters of multilingual LMs remain large due to the larger embedding matrix of the vocabulary covering tokens in different languages. Instead, monolingual LMs can be trained in a target language with the language-specific vocabulary only. In this paper, we propose vocabulary-trimming (VT), a method to reduce a multilingual LM vocabulary to a target language by deleting potentially irrelevant tokens from its vocabulary. In theory, VT can compress any existing multilingual LM to any language covered by the original model. In our experiments, we show that VT can retain the original performance of the multilingual LM, while being considerably smaller in size than the original multilingual LM. The evaluation is performed over four NLP tasks (two generative and two classification tasks) among four widely used multilingual LMs in seven languages. The results show that this methodology can keep the best of both monolingual and multilingual worlds by keeping a small size as monolingual models without the need for specifically retraining them, and can even help limit potentially harmful social biases. + +# 1 Introduction + +Multilingual language model (LM) pre-training (Devlin et al., 2019; Conneau et al., 2019; Liu et al., 2020; Xue et al., 2021) has been shown to be an efficient mechanism to store information from many languages into a single model, without the need for training multiple language-specific models. Moreover, it has been proven reliable for cross-lingual tasks (Pires et al., 2019; Conneau and Lample, 2019) and can provide competitive performance in most settings, generally similar to its monolingual counterparts (Goyal et al., 2021), while being generally less affected by culturally-dependant biases (Ahn and Oh, 2021). Similarly to monolingual + +models, multilingual LMs can be used for zero/few-shot learning (Scao et al., 2022) by increasing the model size and, more frequently, can be specialized to different tasks by fine-tuning to specific data. In practice, however, there are a few practical issues when training multilingual LM such as the curse of multilinguality (Conneau et al., 2019; Pfeiffer et al., 2022), a trade-off between the number of languages and individual performance in a single language, or the multilingual vocabulary construction, which requires a careful design for better generalization (Chung et al., 2020; Zheng et al., 2021; Liang et al., 2023). + +Besides such generalization concerns, multilingual LMs usually consist of larger parameters than their monolingual counterparts due to the need for a large vocabulary covering multiple languages. This becomes an important issue in practice when the resources to host models are limited. For instance, while using the same configuration (i.e., same number of layers and hidden units), the parameter size of T5SMALL (Raffel et al., 2020) and mT5SMALL (Xue et al., 2021) are 140M and 300M, respectively. This is only due to their difference in vocabulary size, with T5 being 50k and mT5, 250k. In fact, the embedding matrix stemming from the LM vocabulary can occupy a large portion of the parameter space. For instance, the ratio of the embedding matrix to the full model's parameter size in multilingual LMs can be higher than 80% as T5 (see Figure 1). + +In this paper, we propose a simple vocabulary trimming (VT) method to remove tokens from the vocabulary of multilingual LMs that may be irrelevant to the target language. This is achieved by automatically identifying language-specific tokens from an underlying text corpus. We consider two VT strategies of pre-FT VT (VT before fine-tuning) and post-FT VT (VT after fine-tuning) and + +![](images/f053d4a449685b0f6f10357a3b054e984dccd01320158524f8cedf511529135b.jpg) +Figure 1: The ratio of the embedding matrix to the number of parameters for each multilingual LM. + +analyse them by varying the final vocabulary size. We conduct experiments on two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI), across seven different languages. The experimental results show that both pre and post fine-tuning VT can reduce the model size while retaining the original performance in generation tasks (QA and QG), and particularly in classification tasks (sentiment and NLI) where the results are close to being identical despite the significant reduction in vocabulary size. In all tasks, the original performance can be generally maintained with less than $40\%$ of the full model parameters for all languages. + +Finally, even though pre-trained LMs have reported impressive performance on various NLP downstream tasks (Kenton and Toutanova, 2019; Liu et al., 2019; Conneau et al., 2019), such LMs also demonstrate worrying levels of social biases in certain situations (May et al., 2019; Kurita et al., 2019; Kaneko and Bollegala, 2021). One natural question that arises is whether VT can have an influence on the bias level in multilingual LMs, including fine-tuned models. For this purpose, we evaluate social bias in multilingual LMs after applying VT with different settings and compare it against its monolingual counterpart. Experimental results show that the monolingual LM tends to contain more bias than its multilingual versions. Moreover, compared to the original multilingual LM, the bias level has no significant change after applying VT. These results suggest that a monolingual LM can be induced by applying VT to its corresponding multilingual LM, thereby obtaining a less biased monolingual LM compared to its original monolingual counterpart. + +# 2 Related Work + +Several studies have explored the possibility to modify or adapt the vocabulary of LMs. For instance, Artetxe et al. (2020) and Marchisio et al. (2022) adapted a mono-lingual LM into another language by learning the embedding matrix on the new language, while fixing the other weights. Similarly, Wang et al. (2019) augmented the vocabulary of a multilingual LM to new languages with multilingual word alignment (Lample et al., 2018). Zheng et al. (2021) proposed to evaluate the ability of a vocabulary to represent a particular language, and Chung et al. (2020) proposed a multilingual vocabulary construction that balances the tradeoff between optimizing for cross-lingual sub-word sharing and the need for robust representation of individual languages. XLM-V (Liang et al., 2023) combines the idea of Zheng et al. (2021) and Chung et al. (2020) to efficiently enlarge the vocabulary size along with the model size scaling. Ostendorff and Rehm (2023) used a multi-stage fine-tuning to obtain a LM in the target language from other LM in the source language. These prior works modify existing mono/multi-lingual LMs to include new languages, i.e. augmenting the multilinguality of the LMs. In contrast, our study focuses on compressing multilingual LMs into the target language to effectively achieve smaller monolingual LMs, i.e. reducing the multilingual representation of the LMs while retaining the capability in a specific target language. + +The work of Abdaoui et al. (2020) is the most relevant to our study as, to the best of our knowledge, they introduced the idea of VT for the first time. However, their analysis is limited to NLI with pre-fine-tuning VT with mBERT (Devlin et al., 2019) only, as well as a fixed vocabulary size after VT. In contrast, our study compares two VT strategies, before and after fine-tuning, and show how this latter strategy, not considered in Abdaoui et al. (2020), can be a more effective compression technique in some settings. Furthermore, we extend the experiments to generation tasks as well as classification tasks with more recent LMs such as mBART (Lewis et al., 2020) and mT5 (Xue et al., 2021), and provide an exhaustive analysis on the effect of VT. + +# 3 Vocabulary Trimming + +To perform vocabulary trimming (VT), we first need a multilingual LM as an input. The idea is to + +![](images/4fc018111e4b3ac5afd7840e323af32af5effe1b4e5969e976b76d85a306414b.jpg) +Figure 2: An illustration of vocabulary trimming for Korean and French. + +tailor model to a particular target language $l$ , which in principle belong to the same set of languages used to trained the input multilingual LM. For the target language $l$ , VT first identifies language-specific tokens on a language-specific corpus $\mathcal{C}_l$ and remove all the tokens along with their embeddings except for those appeared in $\mathcal{C}_l$ as described in Figure 2. In our analysis (§ 5), we also consider to keep the top- $n$ most frequent tokens in $\mathcal{C}_l$ to further reduce the model size by removing less frequent tokens. We consider two VT strategies: (1) before fine-tuning and (2) after fine-tuning. + +The difference between these two strategies is whether to perform VT before or after fine-tuning, as shown in Figure 3. Both VTs have advantages and drawbacks: while pre-FT VT can reduce the time of fine-tuning as the trimmed LM is smaller than the original LM, post-FT VT only need a fine-tuned multilingual LM - this way, post-FT VT can be used as a postprocessing step and no additional language-specific training is required. + +Finally, we release a simple LM vocabulary trimming starting package to apply our proposed technique to any input multilingual transformer-based LM, along with all the models and code needed to reproduce our experiments, at https://github.com/asahi417/lm-vocab-trimmer. + +# 4 Evaluation + +In this section, we present our experimental results to test the reliability of our proposed VT methodology in NLP tasks. + +![](images/91d2550dceac5f3d39f6a60a1cbe3e06c1edc8e29abf72d40849b39523663876.jpg) +Figure 3: Comparisons of Pre-FT vs Post-FT VT in an example of fine-tuning on a task in French. + +# 4.1 Experimental Setting + +Tasks and datasets. In order to test the efficacy of VT, we consider two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI). As the datasets for QA, we use SQuAD (Rajpurkar et al., 2016) (English), Spanish SQuAD (Casimiro Pio et al., 2019) (Spanish), FQuAD (d'Hoffschmidt et al., 2020) (French), Italian SQuAD (Croce et al., 2018) (Italian), JAQuAD (So et al., 2022) (Japanese), KorQuAD (Lim et al., 2019) (Korean), and SberQuAD (Efimov et al., 2020) (Russian). For QG, we use the same datasets adapted for QG via QG-Bench (Ushio et al., 2022). For sentiment analysis, we use Twitter-based datasets for English (Rosenthal et al., 2017), Arabic (Rosenthal et al., 2017), French (Benamara et al., 2017), Italian (Barbieri et al., 2016), German (Cieliebak et al., 2017), Portuguese (Brum and Nunes, 2017), and Spanish (Diaz-Galiano et al., 2018) from UMSAB (Unified Multilingual Sentiment Analysis Benchmark) (Barbieri et al., 2022). All the sentiment analysis datasets contain three labels: positive, neutral and negative. For NLI, we use XNLI (Conneau et al., 2018), a multilingual NLI dataset, including English, French, German, Spanish and Arabic, which are the languages included in the sentiment analysis experiment. We fine-tune LMs on the training sets of each language, which were translated automatically from English and released in the original paper. + +Evaluation metrics. For the evaluation, we use the following standard metrics: answer span F1 score (Ans-F1) and exact match (EM) are used for QA; METEOR (MTR) and BERTScore (BS) for QG, which have been shown to be the most correlated metrics to human judgment (Ushio et al., 2022); macro-F1 score for sentiment following (Barbieri et al., 2022); and accuracy for NLI. As the language-specific corpus $\mathcal{C}_l$ to extract vocabulary counts for VT, we use mC4 (Xue et al., 2021), one of the largest public multilingual corpora. + +
Lang.VocabularyParametersQAQG
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
mT5EN209K (83.6%)258M (86.1%)70.1 / 55.570.2 / 55.570.1 / 56.423.8 / 90.023.8 / 90.024.0 / 90.1
ES131K (52.4%)178M (59.4%)55.9 / 34.755.9 / 34.757.8 / 37.522.7 / 84.122.7 / 84.122.3 / 84.2
FR131K (52.4%)178M (59.4%)50.0 / 30.950.0 / 30.948.6 / 29.417.5 / 80.717.5 / 80.716.1 / 79.2
IT111K (44.4%)157M (52.6%)53.2 / 37.653.4 / 37.851.5 / 36.017.6 / 80.817.6 / 80.817.5 / 80.6
JA125K (50.0%)172M (57.6%)65.7 / 65.765.7 / 65.763.0 / 63.029.0 / 80.929.0 / 80.928.6 / 81.0
KO73K (29.2%)119M (39.7%)77.1 / 70.677.1 / 70.574.5 / 67.327.5 / 82.927.5 / 83.028.0 / 83.7
RU147K (58.8%)195M (65.1%)73.7 / 51.473.8 / 51.474.8 / 53.426.4 / 84.326.4 / 84.328.9 / 86.4
mBARTEN173K (69.2%)532M (87.1%)76.9 / 62.677.0 / 62.778.4 / 65.725.1 / 90.425.1 / 90.424.7 / 90.1
ES87K (34.8%)443M (72.7%)64.1 / 42.264.5 / 42.863.7 / 43.922.9 / 83.622.8 / 83.622.8 / 84.0
FR85K (34.0%)442M (72.5%)60.4 / 39.361.0 / 39.866.4 / 45.119.8 / 81.719.8 / 81.718.4 / 79.7
IT67K (26.8%)424M (69.5%)64.7 / 50.064.9 / 50.265.8 / 49.818.0 / 80.617.9 / 80.718.9 / 81.1
JA77K (30.8%)434M (71.1%)68.2 / 68.268.2 / 68.270.6 / 70.630.0 / 82.329.7 / 82.129.1 / 80.8
KO46K (18.4%)402M (65.9%)79.3 / 72.379.2 / 72.183.2 / 77.330.2 / 83.930.3 / 84.030.2 / 83.8
RU99K (39.6%)456M (74.8%)78.7 / 58.079.0 / 58.275.5 / 49.929.3 / 87.228.7 / 87.028.3 / 86.7
+ +Table 1: Results on QA (Ans-F1/EM) and QG (MTR/BS), including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\%). The best results in each LM and language are in bold characters. Note that the parameter size of the original mT5 and mBART (No-Trim) is 300M and 611M, respectively, both with a vocabulary size of 250K. + +Base language models. As the base LMs, given computational constraints we chose the smallest mT5 and mBART to fine-tune on QA and QG, and XLM-R and XLM-V (Liang et al., 2023) for sentiment analysis and NLI. All these models have a vocabulary size of $250\mathrm{K}$ , except for XLM-V which has a vocabulary size of $901\mathrm{K}$ subword tokens. For our experiments, we compare the results of pre/post-FT VT against vanilla LM fine-tuning without VT, which we refer to as No-Trim. + +Fine-tuning. For model fine-tuning, we rely on lmqg (Ushio et al., 2023) for QA/QG, and Ray Tune3 for sentiment analysis. In both cases, we use the default search space for hyperparameter search. For NLI, we follow the same hyperparameters used in Liang et al. (2023). All the resulting models and code will be released upon acceptance of the paper. + +# 4.2 Results + +We present the results for the generation and classification tasks in section 4.2.1 and section 4.2.2, respectively. + +# 4.2.1 Generation Tasks: QA & QG + +Table 1 shows the overall results on QA and QG. The results confirm that both of pre/post-FT VT can maintain the original performance in most cases, while being smaller than the original models by significantly reducing the vocabulary size. First, post-FT VT achieves at least the same performance as + +the vanilla fine-tuning for all the languages for both LMs in QA and QG, except for a few cases such as mBART QA in Korean and mBART QG in Russian, although the decrease is no more than $0.5\%$ . Meanwhile, pre-FT VT outperforms its vanilla fine-tuning model with a relatively important margin in some cases, such as mBART French QA and mT5 Spanish QA. In contrast, there are a few models where pre-FT VT degrades the performance of the original model such as mT5 QA in Korean (2.6% decrease in Ans-F1) or mBART QA in Russian (3.2% decrease in Ans-F1). + +Since we keep all the vocabulary that appeared in the language-specific corpus $\mathcal{C}_l$ , the percentage of reduced parameter depends on the language, and generally VT can reduce the model size for Asian (Japanese/Korean) and European (Spanish/French/Italian) languages efficiently (50% for mT5 and 70% for mBART), but it remains high in other languages (English-Russian). + +# 4.2.2 Classification Tasks: Sentiment & NLI + +Table 2 shows the results on sentiment analysis and NLI. In this case, post-FT VT can robustly preserve the original performance of the original No-Trim baseline in both tasks for XLM-R and XLM-V, while being no more than $40\%$ and $60\%$ in vocabulary and overall parameter size, respectively, of the original XLM-V and XLM-R models in all the non-English datasets. XLM-V PT sentiment model is the only post-FT VT where a slight decrease can be observed $(0.1\%)$ . On the other hand, the accuracy + +
Lang.VocabularyParameterSentimentNLI
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
XLM-RAR49K (19.6%)124M (44.7%)66.366.360.975.775.773.8
DE91K (36.4%)156M (56.3%)73.273.373.579.979.978.3
EN173K (69.2%)219M (78.7%)68.468.567.984.684.670.6
ES87K (34.8%)153M (55.0%)69.069.065.079.879.867.2
FR85K (34.0%)151M (54.6%)71.871.872.180.180.179.6
IT67K (26.8%)138M (49.7%)62.962.970.8---
PT66K (26.4%)137M (49.3%)70.770.870.2---
XLM-VAR92K (11.8%)157M (20.2%)59.859.864.775.575.676.1
DE239K (30.7%)269M (34.7%)73.573.573.078.978.979.0
EN484K (62.2%)458M (58.9%)63.963.961.384.484.484.5
ES243K (31.2%)279M (35.1%)60.760.766.680.780.780.6
FR218K (28.0%)253M (32.6%)68.868.859.578.678.679.0
IT184K (23.7%)227M (29.3%)70.270.274.2---
PT181K (23.3%)225M (28.9%)66.666.552.8---
+ +Table 2: Results of sentiment analysis (macro F1) and XNLI (accuracy) including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\%). The best results in each LM and language are in bold characters. Note that the overall parameter size of the original XLM-R and XLM-V (No-Trim) is 278M and 778M, respectively, with the vocabulary size being 250K and 901K vocabulary in each case. + +of Pre-FT VT appears to be sensitive to the language and task, where it improves the performance in some languages such as Italian (XLM-R and XLM-V achieve $7.9\%$ and $3.8\%$ increase for sentiment analysis), but it decreases the performance with non-trivial margin in other languages such as Arabic, where XLM-R decreases $5\%$ for sentiment analysis and $2\%$ for XNLI. Since XLM-V has a larger vocabulary size, the percentage of reduced parameters at VT is more prominent in XLM-V, as seen in Arabic $(20.2\%)$ and Portuguese $(28.9\%)$ for example. + +# 5 Vocabulary Size Analysis + +In our main experimental results (§ 4.2), all the unique tokens that appeared in the monolingual corpus were kept, which resulted in a low compression ratio for some languages such as English and Russian. In this analysis, we constrain the number of vocabulary and choose the top- $n$ vocabulary at VT in terms of the frequency in the corpus (see § 3). For QA and QG, we compare mT5SMALL results with $n$ from [5K, 10K, 15K, 30K, 60K, 90K, 120K], which correspond to an overall parameter size of [49M, 54M, 59M, 74M, 105M, 136M, 166M], respectively. For sentiment analysis and NLI, we experiment with XLM-RBASE with $n$ from [5K, 10K, 15K, 30K, 60K], which correspond to [89M, 93M, 97M, 109M, 132M] of parameter size, respectively4. + +4The full results with top- $n$ can be found at Appendix B and Appendix A. + +# 5.1 Generation Tasks: QA & QG + +Figure 4 shows the results of mT5 on QA and QG. Noticeably, post-FT VT can reduce the vocabulary size to $60\mathrm{K}$ for both QA and QG in all the languages with a trivial gap (0.3% decrease of EM in Russian QA and 0.1% decrease of BS in French QG), and that is $35\%$ of the original mT5 in the parameter size. Furthermore, post-FT VT can further reduce the vocabulary to 5K tokens with no more than $0.4\%$ decrease in each metric for both QA and QG in English, French, and Italian, which is $16\%$ of the original mT5 in the parameter size. Meanwhile, pre-FT VT outperforms the No-Trim result in all the languages in QA, and the majority of the languages in QG (English, Italian, Korean, and Russian), but the result is sensitive to the choice of $n$ . For example, Japanese/Korean QA and Russian QG with pre-FT VT for top-5K (16% of the original mT5) outperforms No-Trim as well as post-FT VT, but Japanese QG with pre-FT VT is worse in any choice of $n$ on contrary. This larger variation of results may also be due to the parameter size space, as the optimal parameters for the original multilingual LM (which is the one trained for post-FT VT) may differ. We leave this extended analysis for future work. + +# 5.2 Classification Tasks: Sentiment & NLI + +Figure 5 and Figure 6 show the results of XLM-R on sentiment and NLI. In NLI, we can see that post/pre-FT VT both can reduce the vocabulary to $30\mathrm{K}$ ( $39\%$ of the original XLM-R in the parameter + +![](images/5e6650c638298516651a68d4acd871db90087e97bf0034e116a4a47a4562c1da.jpg) + +![](images/214b1cb5de93ffa682717eca311d74b5ddd98860b07b847b9bc40f92e49b2e8c.jpg) + +![](images/4ef4be96790435dcb576decfd929b245981d80679335f72a4804c3fc3200b162.jpg) + +![](images/745560447b5a3de1acb235070cd9744770219f7571f3d1aad6265b035e465ccf.jpg) + +![](images/715ebe8415fbefb1a0097708d1e4901076b019320b6bb99ee83d76accb0d6567.jpg) + +![](images/b54ba16a3c7b227762bc66d0dbee6a7528d0368e9a2239f999f2fad2432123eb.jpg) + +![](images/9a090bc793a0a1869516bb430ca16bd11094503d3f563323833eda28867efc53.jpg) + +![](images/75c098b5fc4c7ffe757384015efa7fe788d23b6d0edb6e41874109334c440c08.jpg) +Figure 5: Sentiment analysis macro-F1 results of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim. + +![](images/47f32c80e7c2e10c226f0396803a7183389b3f1704257858fc712dacfdb6f84d.jpg) + +![](images/24170ba9678d8eac52dabf8ca9c46bde1a4fce44ac8cd021b7043becbd112525.jpg) + +![](images/6b757cbcbd91cfa363e40352c2e94bd30605250aa42e034ae9b3550e5c7e42ea.jpg) + +![](images/85215f179d7c7ebb6e1a627c6d011a684c3ba3ba329818ee70d99c65976aacc2.jpg) + +![](images/afe7ff6bb4af375fdc3f9374c1aec20d7c7c4ffcf56a69f3ea9a4e923512bd42.jpg) +Figure 4: QG (METEOR) and QA (Ans-F1) results for mT5 with pre/post-FT VT for different vocabulary sizes compared to the original multilingual LM (No-Trim). + +![](images/b4cf891736fc8564d14bf79275ea364660fedb30dac1ed1cacc0e91411c9fe5b.jpg) + +size) without any decrease except $0.3\%$ of pre-FT VT for German, and there is no decrease more than $0.4\%$ even with top-15K of post-FT VT. In sentiment analysis, pre-FT VT with top-10K ( $33\%$ of the original XLM-R in the parameter size) can retain the accuracy of the No-Trim baseline in French and Italian. Moreover, post-FT VT with 30K can retain the original F1 score without a major drop in sentiment analysis, yet the decrease in F1 score is slightly more prominent than NLI ( $1.1\%$ in Arabic sentiment analysis). + +The sentiment analysis datasets are collected from Twitter, so one dataset in a single language can contain tokens from other languages ( hashtags or named-entities, or even code-switching). In contrast, XNLI translates English NLI into other languages, so there is less chance for a dataset to + +![](images/146d9236bbdef860d7b91976f3d8c5f137a736392dfaf7a613addd91ab956bbc.jpg) + +![](images/cfc291a843c11054b28ea043e7f22765b94502110b2ffac101a2776168a0c008.jpg) + +![](images/65b148e2aa17e46779763a15938c58f8f311d869565ff8de7e2718adfc9dbb77.jpg) + +![](images/ccde1b462d6a6e11a8a61bf13ea45d0247ec5eb000e3d85060a92bfcf13aaeb4.jpg) + +![](images/98d1ff5f5094a977fb4764fafc0ca6667516a7a4ee0f30ca8467823a3b877e6c.jpg) + +![](images/8f017747f0f32250fc7097711d2769fc2ff5542f0931cd947827cd75add6337d.jpg) + +![](images/66c339afbae3af67690ff9ad56f6178f87a765586078dcdfeacdc4bc86401f50.jpg) + +![](images/ba2036ff80ce83fd6878f1d359b292a508134833619f586f8b62f78aec94c5c2.jpg) + +![](images/4721b40d5c2aeac46f6042c96d91ea63d05e992dba1472f0f01bc7dacc07d33d.jpg) + +![](images/ee2bf3a8a0cae4e99d677ced953d5341b1fd2ab55ecbfdda0c44388ac3e12256.jpg) + +![](images/440cd0ff6cd1f351989c873d2f0a9cf73696ba90d62230c2167003f54be7976c.jpg) + +![](images/8bf9aa983690495010bd0429eaa962a0f9254995a2795bc0b8126bf7d24aac4a.jpg) +Figure 6: NLI accuracy of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim. + +contain tokens from other languages. This can explain the effectiveness of top- $n$ VT in NLI compared with sentiment analysis, as smaller values of $n$ should result in a vocabulary with fewer tokens from the other languages, which limits the ability of the models to handle foreign tokens. + +# 6 Monolingual vs. Multilingual LMs: The Case of Social Bias + +There has been extensive literature in NLP comparing monolingual and multilingual LMs (Muller et al., 2021; Goyal et al., 2021). As for the performance, there is no clear consensus on which type is better for certain languages, tasks or settings. How- + +ever, there are other important factors that play a role in this comparison. First, monolingual models tend to have a smaller vocabulary size, which makes them more practical. In contrast, a single multilingual model can be used for a large number of languages. Moreover, multilingual LMs are less prone to capture and carry cultural- or language-dependent biases. This is due to the combination of languages and cultures into a single model, which makes it less biased toward specific cultures (Liang et al., 2020; Ahn and Oh, 2021). Prior works have shown that different types of biases consistently appear in language-specific models (Nadeem et al., 2021; Nangia et al., 2020; Blodgett et al., 2021; Dhamala et al., 2021; Kaneko et al., 2022; Zhou et al., 2022). While the comparison of monolingual and multilingual LMs is not the main focus of this paper, this analysis is certainly relevant. Trimming the vocabulary of a multilingual model essentially makes the model smaller, and therefore alleviates one of the main drawbacks of using multilingual language models on language-specific tasks, which is its larger size. On top of that, this strategy enables the usage of monolingual models with potentially less social bias. In the following, we present a comparison of monolingual and multilingual LMs (both trimmed and not trimmed) in terms of social bias and general performance. + +# 6.1 Experimental setting + +Social bias datasets. To study the effect of VT on social bias existing in pre-trained LMs, we first conduct experiments on two commonly used social bias evaluation datasets for masked LMs: StereoSet (SS; Nadeem et al., 2021) $^5$ and crowdsourced stereotype pairs benchmark (CP; Nangia et al., 2020) $^6$ . SS consists of associative contexts covering four types of social biases: race, gender, religion, and profession; while CP is crowdsourced and annotated by workers in the United States, which contains nine types of social biases: race, gender, sexual orientation, religion, age, nationality, disability, physical appearance, and socioeconomic status. In order to further investigate the impact of pre/post-FT VT on LMs, we trim and fine-tune models on sentiment analysis with different orders and evaluate the social bias in such models on the Equity Evaluation Corpus (EEC; Kiritchenko and + +Mohammad, 2018) considering two bias types: gender and race. The EEC dataset was specifically with the aim to examine social bias for sentiment analysis systems. + +Evaluation metrics. We compare the pseudolikelihood scores returned by each model for stereotypical and anti-stereotypical sentences using AULA (All Unmasked Likelihood with Attention weights) (Kaneko and Bollegala, 2022).8 AULA has been shown to be robust against the frequency biases of the masked tokens and provides a more reliable assessment in contrast to alternative metrics when evaluating social biases in masked language models (MLMs). Given a sentence pair in the test dataset: "My mom spent all day cooking for Thanksgiving" vs. "My dad spent all day cooking for Thanksgiving", the first sentence is considered as stereotypical while the second one is anti-stereotypical. AULA computes the percentage of stereotypical sentences preferred by the MLM over anti-stereotypical ones as the bias score. An MLM is considered to be unfairly biased if it returns higher pseudo-loglikelihood scores for stereotypical sentences than the corresponding anti-stereotypical sentences. The AULA score falls within the range of [0,100] and an unbiased model would return a bias score close to 50. On the other hand, a bias score greater than or less than 50 indicates the bias direction towards the stereotype or anti-stereotype, respectively. Since the original AULA is not fitted to evaluate fine-tuned models, we adapt AULA to the EEC dataset obtain the bias score for the LMs fine-tuned on sentiment analysis, and denote this metric as EEC-AULA. Specifically, given a model that predicts sentiment labels (e.g., positive, neutral, negative) to sentences, we consider the percentage of stereotypical test sentences with a more negative label over anti-stereotypical ones as the corresponding bias evaluation measure. + +General performance. As a proxy to test the general performance, we use the general language understanding evaluation (GLUE; Wang et al., 2018) benchmark. We acknowledge the limitations of using this benchmark to draw reliable conclusions at large (Ethayarajh and Jurafsky, 2020) + +
Base ModelVocab TrimmingVocabularySocial BiasGeneral Performance
Pre-FTPost-FTAULAEEC-AULAGLUE
CPSSGenderRace
Monolingual (RoBERTa)-50K50K58.158.864.885.778.0
Multilingual (XLM-R)-250K250K49.554.944.362.077.9
Pre-FT VT (EN)173K173K49.554.942.556.978.0
Post-FT VT (EN)250K173k49.554.944.362.077.9
Pre-FT VT (50K)50K50K49.355.041.062.476.9
Post-FT VT (50K)250K50K49.554.944.362.077.9
+ +Table 3: Results of pre/post-FT VT models compared with the original monolingual and multilingual models on two social bias analysis benchmarks (AULA for pre-trained masked language models and EEC-AULA for models fine-tuned on sentiment analysis) and the GLUE benchmark. The VT models are trimmed on English vocabulary with different vocabulary sizes: EN (full English vocabulary) and 50K (top 50K subword tokens). Note that for post-FT VT, the results on AULA are exactly the same as the original XLM-R. The green and red colours represent the social bias towards anti-stereotypical sentences (scores lower than 50) and stereotypical sentences (scores higher than 50), respectively. The lighter colour indicates less social bias observed in the LM. + +but it nevertheless provides a good proxy for understanding the overall performance of comparable models in standard NLP tasks. Moreover, these experiments are only aimed at analysing the effect of vocabulary trimming on general performance. + +Models. We compute the bias scores of RoBERTa (Liu et al., 2019) as base monolingual LM, and XLM-R (Conneau et al., 2019) as its multilingual counterpart (they have been trained with the same architecture and in an overlapping corpus. We explore two VT settings to be applied to XLM-R: XLM-R with the standard VT including the full English vocabulary (VT XLM-R) and XLM-R with VT for top-50K English vocabulary (top-50K VT XLM-R), which is the same vocabulary size as the monolingual RoBERTa model. Our experiments are based both on masked language models on AULA (in which the post-VT does not have any effect) and models fine-tuned on the sentiment analysis presented in $\S 4.1$ on EEC-AULA, as well as on the corresponding GLUE training sets. + +# 6.2 Results + +Table 3 shows the performance of pre-FT and post-FT VT models against the original monolingual and multilingual LMs on social bias evaluation datasets and the GLUE benchmark. Both AULA and GLUE results are computed using the LMs without fine-tuning (i.e., RoBERTa, XLM-R, VT XLM-R, and top-50K VT XLM-R), whereas the EEC-AULA results are computed using the models applying VT and fine-tuning strategies. We observe that the monolingual model contains the highest levels of social biases compared to the multilin + +gual models with different settings. In particular, RoBERTa obtains the overall highest bias score on the EEC dataset after fine-tuning, with an alarmingly high 85.7 score on race. $^{10}$ On the other hand, compared to the original XLM-R, there is no significant change in performance on social biases and GLUE evaluation tasks for pre-FT VT and post-FT VT models. This is important as we can apply the proposed VT method to any multilingual LM, obtaining a monolingual one with consistent performance on the GLUE benchmark and less social biases than the original monolingual model pretrained in the target language, without using any ad-hoc debiasing methods. + +# 7 Discussion + +Vocabulary trimming before and after finetuning. According to the results, pre-FT VT appears to be generally more effective in classification tasks (see section 4.2.2). For generation tasks, both pre/post-FT VT robustly retain the original performance while being able to considerably reduce the model size (see section 4.2.1). As a guideline to choose the type of VT in such a case, post-FT VT should be more suitable if one already has a finetuned model, as no additional training is needed for this case. Moreover, post-FT is more robust as a compression mechanism as the performance is largely maintained with respect to that of the original multilingual LM. On the other hand, if one needs to fine-tune a model from scratch and the computation resources are limited, we recommend + +exploring pre-FT VT, as fine-tuning on a trimmed LM should be more efficient due to its smaller vocabulary and parameters and, in some cases, can lead to better overall results. However, this process has to be done carefully as the set of optimal parameters could differ from the original multilingual LM fine-tuning process. + +While in this paper we have not compared monolingual and multilingual models, the question would be whether we need vocabulary trimming strategies in a world where monolingual LMs exist. In this case, a monolingual model may perform similarly to a multilingual model (Goyal et al., 2021). However, the multilingual model is often larger mainly due to larger vocabulary storage requirements. In contrast, our proposed VT technique does not require any extra LM training or computational resources. Indeed, only a multilingual LM is needed and we can induce multiple smaller language-specific monolingual models. This may reduce the carbon footprint overall and especially help with less-resource languages when a high-quality monolingual model does not exist. Finally, our social bias analysis see § 6 shows how monolingual models exhibit larger social biases (especially racial) than a VT-induced multilingual LM. This is consistent with prior work suggesting that a multilingual LM has been trained with more languages, and hence more cultural variety, and these diverging viewpoints can compensate each other (Ahn and Oh, 2021). + +# 8 Conclusion + +In this paper, we proposed vocabulary-trimming (VT), a method to reduce the vocabulary of a multilingual LM to a vocabulary specific to any given target language. VT can induce a monolingual LM in the target language by leveraging an existing multilingual LM. The main advantage of this filtering step is the reduced size, as well as avoiding having to train monolingual LMs from scratch, which would be computationally demanding. Our experiments show how VT can retain the high performance of the original multilingual LM, while largely reducing the model size. For all languages evaluated, a $35\%$ compression rate proves sufficient to keep the original performance of the larger mT5 multilingual LM in both QA and QG, with a similar $39\%$ in NLI and $55\%$ in sentiment analysis with XLM-R. Interestingly, in some cases, the + +compressed LM can even achieve better results than the original larger model when trimmed before fine-tuning. Since the main goal of the paper was to compress a multilingual LM while keeping its original performance, we leave the analysis of this behaviour for future work. + +# Limitations + +We have not tested our methodology in truly low-resource languages. Because of this, there could be a different behaviour when we apply VT to a language with lower resources or that is poorly represented in the underlying training corpus. The LMs we used in the paper limited their size up to 600M, and we have not considered larger models such as mT5XXL or BLOOM (Scao et al., 2022), due to our limited computational resources. As the language-specific corpus to compute frequency, we employ mC4, which is one of the largest multilingual corpora. Nonetheless, this is used as a proxy and having access to the full multilingual model could give potentially better results. + +Similarly, we acknowledge the limitations of the analysis comparing multilingual and monolingual models in terms of social bias. Due to the evaluation data available and the existence of comparable monolingual and multilingual LMs, the evaluation is focused on English only and the results could differ for other languages. Moreover, there are other types of biases not covered in this evaluation. + +# Ethics Statement + +Pre-trained LMs are known to contain undesirable biases to generate toxic contents in some edge cases (Schick et al., 2021), so the resulting models could inherit such biases. While we have not analysed in detail the output of all models in the tasks evaluated, in this paper we have made an attempt to study this effect in terms of social biases for both base pretrained LMs and fine-tuned LMs. + +# Acknowledgements + +Jose Camacho-Collados and Yi Zhou are supported by a UKRI Future Leaders Fellowship. + +# References + +Amine Abdaoui, Camille Pradel, and Grégoire Sigel. 2020. Load what you need: Smaller versions of mutililingual BERT. In Proceedings of SustainNLP: Workshop on Simple and Efficient Natural Language + +Processing, pages 119-123, Online. Association for Computational Linguistics. +Jaimeen Ahn and Alice Oh. 2021. Mitigating language-dependent ethnic bias in BERT. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 533-549, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics. +Francesco Barbieri, Valerio Basile, Danilo Croce, Malvina Nissim, Nicole Novielli, and Viviana Patti. 2016. Overview of the evalita 2016 sentiment polarity classification task. In Proceedings of third Italian conference on computational linguistics (CLiC-it 2016) & fifth evaluation campaign of natural language processing and speech tools for Italian. Final Workshop (EVALITA 2016). +Francesco Barbieri, Luis Espinosa Anke, and Jose Camacho-Collados. 2022. XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 258-266, Marseille, France. European Language Resources Association. +Farah Benamara, Cyril Grouin, Jihen Karoui, Véronique Moriceau, and Isabelle Robba. 2017. Analyse d'opinion et langage figuratif dans des tweets: presentation et résultats du début fouille de textes deft2017. +Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1004-1015. +Henrico Bertini Brum and Maria das Graças Volpe Nunes. 2017. Building a sentiment corpus of tweets in brazilian portuguese. arXiv preprint arXiv:1712.08917. +Carrino Casimiro Pio, Costa-jussa Marta R., and Fonollosa Jose A. R. 2019. Automatic Spanish Translation of the SQuAD Dataset for Multilingual Question Answering. arXiv e-prints, page arXiv:1912.05200v1. +Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics. + +Mark Cieliebak, Jan Milan Deriu, Dominic Egger, and Fatih Uzdilli. 2017. A twitter corpus and benchmark resources for german sentiment analysis. In 5th International Workshop on Natural Language Processing for Social Media, Boston MA, USA, 11 December 2017, pages 45-51. Association for Computational Linguistics. +Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116. +Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32. +Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics. +Danilo Croce, Alexandra Zelenanska, and Roberto Basili. 2018. Neural learning for question answering in italian. In AI*IA 2018 – Advances in Artificial Intelligence, pages 389–402, Cham. Springer International Publishing. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. Bold: Dataset and metrics for measuring biases in open-ended language generation. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 862-872. +Martin d'Hoffschmidt, Wacim Belblidia, Quentin Heinrich, Tom Brendle, and Maxime Vidal. 2020. FQuAD: French question answering dataset. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1193–1208, Online. Association for Computational Linguistics. +Manuel C Díaz-Galiano, Eugenio Martínez-Cármara, M Ángel García Cumbreras, Manuel García Vega, and Julio Villena Román. 2018. The democratization of deep learning in tass 2017. Procesamento del Lenguaje Natural, 60:37-44. + +Pavel Efimov, Andrey Chertok, Leonid Boytsov, and Pavel Braslavski. 2020. Sberquad-russian reading comprehension dataset: Description and analysis. In International Conference of the Cross-Language Evaluation Forum for European Languages, pages 3-15. Springer. +Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics. +Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, and Alexis Conneau. 2021. Larger-scale transformers for multilingual masked language modeling. In Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021), pages 29–33, Online. Association for Computational Linguistics. +Masahiro Kaneko and Danushka Bollegala. 2021. Debiasing pre-trained contextualised embeddings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1256-1266, Online. Association for Computational Linguistics. +Masahiro Kaneko and Danushka Bollegala. 2022. Unmasking the mask-evaluating social biases in masked language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11954-11962. +Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, and Naoaki Okazaki. 2022. Gender bias in masked language models for multiple languages. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2740-2750, Seattle, United States. Association for Computational Linguistics. +Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186. +Svetlana Kiritchenko and Saif M Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. *NAACL HLT* 2018, page 43. +Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166-172. +Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In International Conference on Learning Representations. + +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. +Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, and Madian Khabsa. 2023. Xlm-v: Overcoming the vocabulary bottleneck in multilingual masked language models. arXiv preprint arXiv:2301.10472. +Sheng Liang, Philipp Dufter, and Hinrich Schütze. 2020. Monolingual and multilingual reduction of gender bias in contextualized representations. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5082-5093, Barcelona, Spain (Online). International Committee on Computational Linguistics. +Seungyoung Lim, Myungji Kim, and Jooyoul Lee. 2019. Korquad1.0: Korean qa dataset for machine reading comprehension. arXiv preprint arXiv:1909.07005. +Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Kelly Marchisio, Patrick Lewis, Yihong Chen, and Mikel Artetxe. 2022. Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training. arXiv preprint arXiv:2212.10503. +Chandler May, Alex Wang, Shikha Bordia, Samuel Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 622-628. +Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics. + +Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics. +Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953-1967, Online. Association for Computational Linguistics. +Malte Ostendorff and Georg Rehm. 2023. Efficient language model training through cross-lingual and progressive transfer learning. arXiv preprint arXiv:2301.09626. +Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics. +Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics. +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551. +Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics. +Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 task 4: Sentiment analysis in Twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502-518, Vancouver, Canada. Association for Computational Linguistics. +Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100. + +Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP. Transactions of the Association for Computational Linguistics, 9:1408-1424. +ByungHoon So, Kyuhong Byun, Kyungwon Kang, and Seongjin Cho. 2022. Jaquad: Japanese question answering dataset for machine reading comprehension. arXiv preprint arXiv:2202.01764. +Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2022. Generative language models for paragraph-level question generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 670-688, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2023. A practical toolkit for multilingual question and answer generation. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Toronto, Canada. Association for Computational Linguistics. +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355. +Hai Wang, Dian Yu, Kai Sun, Jianshu Chen, and Dong Yu. 2019. Improving pre-trained multilingual model with vocabulary expansion. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 316-327, Hong Kong, China. Association for Computational Linguistics. +Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics. +Bo Zheng, Li Dong, Shaohan Huang, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, and Furu Wei. 2021. Allocating large vocabulary capacity for crosslingual language model pre-training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3203-3215, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Yi Zhou, Masahiro Kaneko, and Danushka Bollegala. 2022. Sense embeddings are also biased - evaluating social biases in static and contextualised sense embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics + +
Vocab.No-Trim (250K)5K10K15K30K60K
Param.278M89M93M97M109M132M
Post-FT (Sentiment)AR66.364.564.565.965.9
DE73.270.472.472.173.7
EN68.464.066.567.468.6
ES69.066.267.267.868.4
FR71.868.671.471.771.6
IT62.960.861.963.562.3
PT70.763.665.968.269.4
Pre-FT (Sentiment)AR66.358.661.963.163.2
DE73.270.671.873.171.5
EN68.464.666.667.766.0
ES69.060.465.666.365.9
FR71.870.374.373.872.2
IT62.966.267.168.168.2
PT70.765.567.569.667.9
Post-FT (NLI)AR75.775.075.875.875.7
DE79.976.678.979.679.9
EN84.680.583.684.384.6
ES79.875.378.479.479.8
FR80.177.280.080.180.2
Pre-FT (NLI)AR75.773.075.075.475.7
DE79.977.479.079.079.6
EN84.683.384.784.485.1
ES79.878.979.179.281.0
FR80.177.378.978.280.1
+ +Table 4: Results of XLM-R for sentiment analysis (macro F1) and NLI (accuracy) with different top- $n$ vocabulary size at VT, where the best results in each LM and language are in the bold characters. + +(Volume 1: Long Papers), pages 1924-1935, Dublin, Ireland. Association for Computational Linguistics. + +# A Top- $n$ VT of XLM-R + +Table 4 shows the results of XLM-R fine-tuned on sentiment and NLI with post/pre-VT for different top- $n$ . + +# B Top- $n$ VT of mT5 + +Table 5 shows the results of mT5 fine-tuned on QA and QG with post/pre-VT for different top- $n$ . + +# C Details of Results on Social Bias Evaluation + +Table 6 shows the details of social bias evaluation (EEC dataset) regarding each emotion type observed in the LMs fine-tuned on sentiment analysis. Table 7 shows the details of social bias regarding each bias type in both CP and SS datasets observed in the comparison LMs. + +
Vocab. +Param.No-Trim (250K) +300M5K +49M10K +54M15K +59M30K +74M60K +105M90K +136M120K +166M
EN70.1 / 55.572.4 / 59.773.2 / 60.573.7 / 60.972.7 / 58.871.0 / 56.470.5 / 55.870.3 / 55.6
ES55.9 / 34.753.5 / 32.554.4 / 33.154.5 / 33.255.2 / 33.856.0 / 34.855.9 / 34.855.9 / 34.7
FR50.0 / 30.950.0 / 31.548.9 / 29.948.6 / 29.349.9 / 30.149.8 / 30.550.2 / 30.950.0 / 30.8
IT53.2 / 37.654.3 / 39.653.8 / 38.454.3 / 38.854.2 / 38.553.7 / 38.153.4 / 37.8-
JA65.7 / 65.754.4 / 54.460.3 / 60.362.6 / 62.660.4 / 60.464.9 / 64.965.6 / 65.665.7 / 65.7
KO77.1 / 70.675.0 / 67.875.8 / 68.776.2 / 69.376.9 / 70.277.3 / 70.7--
RU73.7 / 51.470.1 / 48.871.4 / 49.570.3 / 47.670.3 / 47.973.8 / 51.173.9 / 51.473.8 / 51.4
EN23.8 / 90.023.6 / 89.824.1 / 89.924.2 / 90.024.2 / 90.023.9 / 90.023.9 / 90.023.9 / 90.0
ES22.7 / 84.121.4 / 83.622.0 / 83.822.2 / 83.922.4 / 84.022.7 / 84.122.7 / 84.122.7 / 84.1
FR17.5 / 80.717.1 / 80.317.1 / 80.417.3 / 80.517.3 / 80.517.4 / 80.617.5 / 80.717.5 / 80.7
IT17.6 / 80.817.3 / 80.717.5 / 80.817.5 / 80.817.6 / 80.717.6 / 80.817.6 / 80.8-
JA29.0 / 80.925.7 / 79.227.4 / 80.128.2 / 80.528.8 / 80.929.0 / 80.929.0 / 80.929.0 / 80.9
KO27.5 / 82.927.3 / 82.827.4 / 82.927.4 / 82.927.4 / 82.927.5 / 83.0--
RU26.4 / 84.326.0 / 84.026.2 / 84.226.3 / 84.226.3 / 84.226.3 / 84.326.4 / 84.326.4 / 84.3
EN70.1 / 55.574.0 / 61.371.8 / 59.173.4 / 60.774.3 / 61.371.3 / 56.568.6 / 54.366.2 / 52.1
ES55.9 / 34.756.7 / 36.758.1 / 37.562.2 / 41.459.8 / 40.158.0 / 37.058.9 / 38.152.0 / 32.7
FR50.0 / 30.949.0 / 32.747.6 / 30.243.3 / 26.944.9 / 27.434.3 / 20.043.1 / 24.140.0 / 23.7
IT53.2 / 37.657.6 / 43.761.5 / 46.861.2 / 45.657.8 / 42.156.5 / 40.555.4 / 39.2-
JA65.7 / 65.755.3 / 55.352.2 / 52.261.3 / 61.360.4 / 60.464.6 / 64.666.9 / 66.967.2 / 67.2
KO77.1 / 70.679.5 / 72.683.7 / 77.782.7 / 76.480.4 / 73.781.6 / 75.1--
RU73.7 / 51.473.5 / 51.776.9 / 56.477.4 / 56.975.9 / 54.275.1 / 53.075.0 / 53.073.3 / 51.4
EN23.8 / 90.024.8 / 90.024.4 / 89.924.3 / 89.924.2 / 90.024.2 / 90.023.3 / 89.923.6 / 89.8
ES22.7 / 84.121.9 / 84.121.9 / 83.922.7 / 83.722.0 / 84.322.6 / 84.420.4 / 79.622.6 / 84.1
FR17.5 / 80.716.1 / 79.016.9 / 79.916.9 / 79.515.8 / 78.814.6 / 78.015.7 / 79.017.0 / 79.2
IT17.6 / 80.817.4 / 80.417.8 / 81.118.0 / 80.817.1 / 80.617.4 / 80.917.4 / 80.8-
JA29.0 / 80.926.6 / 79.827.0 / 79.227.5 / 79.727.3 / 79.927.7 / 80.328.3 / 80.828.2 / 80.2
KO27.5 / 82.928.4 / 83.427.8 / 83.428.2 / 84.128.8 / 83.428.4 / 83.4--
RU26.4 / 84.327.7 / 85.930.0 / 86.629.0 / 86.529.2 / 86.329.0 / 86.629.2 / 86.728.9 / 86.0
+ +Table 5: Results of mT5 for QA (Ans-F1/EM) and QG (MTR/BS) with different top- $n$ vocabulary size at VT,where the best results in each LM and language are in the bold characters. + +
ModelAULA-EECAngerSadnessFearJoyNo Emotion Type
Gender BiasFT RoBERTa64.850.067.871.564.360.8
FT XLM-R44.346.939.451.434.439.7
Pre-FT XLM-R (EN)42.533.855.845.140.742.7
Post-FT XLM-R (EN)44.346.939.451.434.439.7
Pre-FT XLM-R (50k)44.346.939.451.434.439.7
Post-FT XLM-R (50k)41.041.245.847.236.937.2
Race BiasFT RoBERTa85.750.088.065.450.0100.0
FT XLM-R62.074.532.147.374.617.4
Pre-FT VT XLM-R (EN)56.933.758.869.260.150.0
Pre-FT VT XLM-R (EN)62.074.532.147.374.617.4
Pre-FT VT XLM-R (50k)62.074.532.147.374.617.4
Post-FT VT XLM-R (50k)62.455.770.370.959.159.0
+ +Table 6: Social bias scores of LMs fine-tuned on sentiment analysis on each emotion type w/wo VT on the EEC dataset. + +
RoBERTaXLM-RVT XLM-R (EN)VT XLM-R (50K)
CP DatasetAULA58.149.549.549.3
Age59.849.449.449.4
Disability68.366.766.768.3
Gender53.448.548.548.1
Nationality56.644.044.044.7
Physical-Appearance52.458.758.755.6
Race-Color56.843.843.844.0
Religion53.354.354.353.3
Sexual-Orientation67.954.854.854.8
Socioeconomic66.358.158.157.6
SS DatasetAULA58.854.954.955.0
Gender62.455.355.354.9
Profession60.454.354.354.6
Race57.056.056.056.0
Religion54.446.846.846.8
+ +Table 7: Social bias scores of VT LMs compared to their original ones on both CP and SS datasets for each bias type. \ No newline at end of file diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/images.zip b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14661103f035dc3284ca636731f10debf78f229a --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44bc087a259de2fc6d848a2782dd3e70d62683ac6058c8c68d597900d89ac520 +size 1061039 diff --git a/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/layout.json b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7d8fea7d56a2104980845db65265b60fe11a77c --- /dev/null +++ b/2023/Efficient Multilingual Language Model Compression through Vocabulary Trimming/layout.json @@ -0,0 +1,8773 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 75, + 459, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 75, + 459, + 110 + ], + "spans": [ + { + "bbox": [ + 135, + 75, + 459, + 110 + ], + "type": "text", + "content": "Efficient Multilingual Language Model Compression through Vocabulary Trimming" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 154, + 131, + 440, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 131, + 440, + 143 + ], + "spans": [ + { + "bbox": [ + 154, + 131, + 440, + 143 + ], + "type": "text", + "content": "Asahi Ushio and Yi Zhou and Jose Camacho-Collados" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 144, + 497, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 144, + 497, + 158 + ], + "spans": [ + { + "bbox": [ + 98, + 144, + 497, + 158 + ], + "type": "text", + "content": "Cardiff NLP, School of Computer Science and Informatics, Cardiff University, UK" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 159, + 443, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 159, + 443, + 172 + ], + "spans": [ + { + "bbox": [ + 153, + 159, + 443, + 172 + ], + "type": "text", + "content": "{UshioA,Zhouy131,CamachoColladosJ}\\@cardiff.ac.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 233, + 274, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 233, + 274, + 567 + ], + "spans": [ + { + "bbox": [ + 86, + 233, + 274, + 567 + ], + "type": "text", + "content": "Multilingual language models (LMs) have become a powerful tool in NLP, especially for non-English languages. Nevertheless, model parameters of multilingual LMs remain large due to the larger embedding matrix of the vocabulary covering tokens in different languages. Instead, monolingual LMs can be trained in a target language with the language-specific vocabulary only. In this paper, we propose vocabulary-trimming (VT), a method to reduce a multilingual LM vocabulary to a target language by deleting potentially irrelevant tokens from its vocabulary. In theory, VT can compress any existing multilingual LM to any language covered by the original model. In our experiments, we show that VT can retain the original performance of the multilingual LM, while being considerably smaller in size than the original multilingual LM. The evaluation is performed over four NLP tasks (two generative and two classification tasks) among four widely used multilingual LMs in seven languages. The results show that this methodology can keep the best of both monolingual and multilingual worlds by keeping a small size as monolingual models without the need for specifically retraining them, and can even help limit potentially harmful social biases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "spans": [ + { + "bbox": [ + 68, + 576, + 154, + 588 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "content": "Multilingual language model (LM) pre-training (Devlin et al., 2019; Conneau et al., 2019; Liu et al., 2020; Xue et al., 2021) has been shown to be an efficient mechanism to store information from many languages into a single model, without the need for training multiple language-specific models. Moreover, it has been proven reliable for cross-lingual tasks (Pires et al., 2019; Conneau and Lample, 2019) and can provide competitive performance in most settings, generally similar to its monolingual counterparts (Goyal et al., 2021), while being generally less affected by culturally-dependant biases (Ahn and Oh, 2021). Similarly to monolingual" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 212, + 526, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 526, + 387 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 526, + 387 + ], + "type": "text", + "content": "models, multilingual LMs can be used for zero/few-shot learning (Scao et al., 2022) by increasing the model size and, more frequently, can be specialized to different tasks by fine-tuning to specific data. In practice, however, there are a few practical issues when training multilingual LM such as the curse of multilinguality (Conneau et al., 2019; Pfeiffer et al., 2022), a trade-off between the number of languages and individual performance in a single language, or the multilingual vocabulary construction, which requires a careful design for better generalization (Chung et al., 2020; Zheng et al., 2021; Liang et al., 2023)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 389, + 526, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 389, + 526, + 632 + ], + "spans": [ + { + "bbox": [ + 302, + 389, + 526, + 632 + ], + "type": "text", + "content": "Besides such generalization concerns, multilingual LMs usually consist of larger parameters than their monolingual counterparts due to the need for a large vocabulary covering multiple languages. This becomes an important issue in practice when the resources to host models are limited. For instance, while using the same configuration (i.e., same number of layers and hidden units), the parameter size of T5SMALL (Raffel et al., 2020) and mT5SMALL (Xue et al., 2021) are 140M and 300M, respectively. This is only due to their difference in vocabulary size, with T5 being 50k and mT5, 250k. In fact, the embedding matrix stemming from the LM vocabulary can occupy a large portion of the parameter space. For instance, the ratio of the embedding matrix to the full model's parameter size in multilingual LMs can be higher than 80% as T5 (see Figure 1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 634, + 525, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 634, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 302, + 634, + 525, + 741 + ], + "type": "text", + "content": "In this paper, we propose a simple vocabulary trimming (VT) method to remove tokens from the vocabulary of multilingual LMs that may be irrelevant to the target language. This is achieved by automatically identifying language-specific tokens from an underlying text corpus. We consider two VT strategies of pre-FT VT (VT before fine-tuning) and post-FT VT (VT after fine-tuning) and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "1We release a Python library for VT at https://github.com/asahi417/lm-vocab-trimmer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "14725" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "spans": [ + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14725-14739 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 71, + 278, + 163 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 278, + 163 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 278, + 163 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 278, + 163 + ], + "type": "image", + "image_path": "f053d4a449685b0f6f10357a3b054e984dccd01320158524f8cedf511529135b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 171, + 289, + 195 + ], + "lines": [ + { + "bbox": [ + 67, + 171, + 289, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 289, + 195 + ], + "type": "text", + "content": "Figure 1: The ratio of the embedding matrix to the number of parameters for each multilingual LM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 253, + 291, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 253, + 291, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 253, + 291, + 456 + ], + "type": "text", + "content": "analyse them by varying the final vocabulary size. We conduct experiments on two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI), across seven different languages. The experimental results show that both pre and post fine-tuning VT can reduce the model size while retaining the original performance in generation tasks (QA and QG), and particularly in classification tasks (sentiment and NLI) where the results are close to being identical despite the significant reduction in vocabulary size. In all tasks, the original performance can be generally maintained with less than " + }, + { + "bbox": [ + 67, + 253, + 291, + 456 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 67, + 253, + 291, + 456 + ], + "type": "text", + "content": " of the full model parameters for all languages." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 476, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 476, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 476, + 291, + 772 + ], + "type": "text", + "content": "Finally, even though pre-trained LMs have reported impressive performance on various NLP downstream tasks (Kenton and Toutanova, 2019; Liu et al., 2019; Conneau et al., 2019), such LMs also demonstrate worrying levels of social biases in certain situations (May et al., 2019; Kurita et al., 2019; Kaneko and Bollegala, 2021). One natural question that arises is whether VT can have an influence on the bias level in multilingual LMs, including fine-tuned models. For this purpose, we evaluate social bias in multilingual LMs after applying VT with different settings and compare it against its monolingual counterpart. Experimental results show that the monolingual LM tends to contain more bias than its multilingual versions. Moreover, compared to the original multilingual LM, the bias level has no significant change after applying VT. These results suggest that a monolingual LM can be induced by applying VT to its corresponding multilingual LM, thereby obtaining a less biased monolingual LM compared to its original monolingual counterpart." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 303, + 70, + 396, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 396, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 396, + 83 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 92, + 526, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 526, + 498 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 526, + 498 + ], + "type": "text", + "content": "Several studies have explored the possibility to modify or adapt the vocabulary of LMs. For instance, Artetxe et al. (2020) and Marchisio et al. (2022) adapted a mono-lingual LM into another language by learning the embedding matrix on the new language, while fixing the other weights. Similarly, Wang et al. (2019) augmented the vocabulary of a multilingual LM to new languages with multilingual word alignment (Lample et al., 2018). Zheng et al. (2021) proposed to evaluate the ability of a vocabulary to represent a particular language, and Chung et al. (2020) proposed a multilingual vocabulary construction that balances the tradeoff between optimizing for cross-lingual sub-word sharing and the need for robust representation of individual languages. XLM-V (Liang et al., 2023) combines the idea of Zheng et al. (2021) and Chung et al. (2020) to efficiently enlarge the vocabulary size along with the model size scaling. Ostendorff and Rehm (2023) used a multi-stage fine-tuning to obtain a LM in the target language from other LM in the source language. These prior works modify existing mono/multi-lingual LMs to include new languages, i.e. augmenting the multilinguality of the LMs. In contrast, our study focuses on compressing multilingual LMs into the target language to effectively achieve smaller monolingual LMs, i.e. reducing the multilingual representation of the LMs while retaining the capability in a specific target language." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 498, + 526, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 498, + 526, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 526, + 714 + ], + "type": "text", + "content": "The work of Abdaoui et al. (2020) is the most relevant to our study as, to the best of our knowledge, they introduced the idea of VT for the first time. However, their analysis is limited to NLI with pre-fine-tuning VT with mBERT (Devlin et al., 2019) only, as well as a fixed vocabulary size after VT. In contrast, our study compares two VT strategies, before and after fine-tuning, and show how this latter strategy, not considered in Abdaoui et al. (2020), can be a more effective compression technique in some settings. Furthermore, we extend the experiments to generation tasks as well as classification tasks with more recent LMs such as mBART (Lewis et al., 2020) and mT5 (Xue et al., 2021), and provide an exhaustive analysis on the effect of VT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 724, + 438, + 739 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 724, + 438, + 739 + ], + "spans": [ + { + "bbox": [ + 302, + 724, + 438, + 739 + ], + "type": "text", + "content": "3 Vocabulary Trimming" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "To perform vocabulary trimming (VT), we first need a multilingual LM as an input. The idea is to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14726" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 71, + 267, + 227 + ], + "blocks": [ + { + "bbox": [ + 92, + 71, + 267, + 227 + ], + "lines": [ + { + "bbox": [ + 92, + 71, + 267, + 227 + ], + "spans": [ + { + "bbox": [ + 92, + 71, + 267, + 227 + ], + "type": "image", + "image_path": "4fc018111e4b3ac5afd7840e323af32af5effe1b4e5969e976b76d85a306414b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 235, + 291, + 259 + ], + "lines": [ + { + "bbox": [ + 67, + 235, + 291, + 259 + ], + "spans": [ + { + "bbox": [ + 67, + 235, + 291, + 259 + ], + "type": "text", + "content": "Figure 2: An illustration of vocabulary trimming for Korean and French." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "spans": [ + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": " tailor model to a particular target language " + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": ", which in principle belong to the same set of languages used to trained the input multilingual LM. For the target language " + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": ", VT first identifies language-specific tokens on a language-specific corpus " + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": " and remove all the tokens along with their embeddings except for those appeared in " + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": " as described in Figure 2. In our analysis (§ 5), we also consider to keep the top-" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": " most frequent tokens in " + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l" + }, + { + "bbox": [ + 67, + 283, + 291, + 445 + ], + "type": "text", + "content": " to further reduce the model size by removing less frequent tokens. We consider two VT strategies: (1) before fine-tuning and (2) after fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 447, + 291, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 291, + 568 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 291, + 568 + ], + "type": "text", + "content": "The difference between these two strategies is whether to perform VT before or after fine-tuning, as shown in Figure 3. Both VTs have advantages and drawbacks: while pre-FT VT can reduce the time of fine-tuning as the trimmed LM is smaller than the original LM, post-FT VT only need a fine-tuned multilingual LM - this way, post-FT VT can be used as a postprocessing step and no additional language-specific training is required." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 570, + 291, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 650 + ], + "type": "text", + "content": "Finally, we release a simple LM vocabulary trimming starting package to apply our proposed technique to any input multilingual transformer-based LM, along with all the models and code needed to reproduce our experiments, at https://github.com/asahi417/lm-vocab-trimmer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 665, + 145, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 145, + 678 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 145, + 678 + ], + "type": "text", + "content": "4 Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 688, + 291, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 688, + 291, + 729 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 291, + 729 + ], + "type": "text", + "content": "In this section, we present our experimental results to test the reliability of our proposed VT methodology in NLP tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 502, + 126 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 502, + 126 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 502, + 126 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 502, + 126 + ], + "type": "image", + "image_path": "91d2550dceac5f3d39f6a60a1cbe3e06c1edc8e29abf72d40849b39523663876.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 135, + 524, + 159 + ], + "lines": [ + { + "bbox": [ + 302, + 135, + 524, + 159 + ], + "spans": [ + { + "bbox": [ + 302, + 135, + 524, + 159 + ], + "type": "text", + "content": "Figure 3: Comparisons of Pre-FT vs Post-FT VT in an example of fine-tuning on a task in French." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 184, + 430, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 184, + 430, + 196 + ], + "spans": [ + { + "bbox": [ + 302, + 184, + 430, + 196 + ], + "type": "text", + "content": "4.1 Experimental Setting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 205, + 526, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 205, + 526, + 613 + ], + "spans": [ + { + "bbox": [ + 302, + 205, + 526, + 613 + ], + "type": "text", + "content": "Tasks and datasets. In order to test the efficacy of VT, we consider two generation tasks, question answering (QA) and question generation (QG), and two classification tasks, sentiment analysis and natural language inference (NLI). As the datasets for QA, we use SQuAD (Rajpurkar et al., 2016) (English), Spanish SQuAD (Casimiro Pio et al., 2019) (Spanish), FQuAD (d'Hoffschmidt et al., 2020) (French), Italian SQuAD (Croce et al., 2018) (Italian), JAQuAD (So et al., 2022) (Japanese), KorQuAD (Lim et al., 2019) (Korean), and SberQuAD (Efimov et al., 2020) (Russian). For QG, we use the same datasets adapted for QG via QG-Bench (Ushio et al., 2022). For sentiment analysis, we use Twitter-based datasets for English (Rosenthal et al., 2017), Arabic (Rosenthal et al., 2017), French (Benamara et al., 2017), Italian (Barbieri et al., 2016), German (Cieliebak et al., 2017), Portuguese (Brum and Nunes, 2017), and Spanish (Diaz-Galiano et al., 2018) from UMSAB (Unified Multilingual Sentiment Analysis Benchmark) (Barbieri et al., 2022). All the sentiment analysis datasets contain three labels: positive, neutral and negative. For NLI, we use XNLI (Conneau et al., 2018), a multilingual NLI dataset, including English, French, German, Spanish and Arabic, which are the languages included in the sentiment analysis experiment. We fine-tune LMs on the training sets of each language, which were translated automatically from English and released in the original paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 624, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 526, + 773 + ], + "type": "text", + "content": "Evaluation metrics. For the evaluation, we use the following standard metrics: answer span F1 score (Ans-F1) and exact match (EM) are used for QA; METEOR (MTR) and BERTScore (BS) for QG, which have been shown to be the most correlated metrics to human judgment (Ushio et al., 2022); macro-F1 score for sentiment following (Barbieri et al., 2022); and accuracy for NLI. As the language-specific corpus " + }, + { + "bbox": [ + 302, + 624, + 526, + 773 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l" + }, + { + "bbox": [ + 302, + 624, + 526, + 773 + ], + "type": "text", + "content": " to extract vocabulary counts for VT, we use mC4 (Xue et al., 2021), one of the largest public multilingual corpora." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "text", + "content": "2In theory, vocabulary trimming could be applied to any language model, even monolingual, but this analysis is out of the scope of this paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14727" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 523, + 257 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 523, + 257 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 523, + 257 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 523, + 257 + ], + "type": "table", + "html": "
Lang.VocabularyParametersQAQG
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
mT5EN209K (83.6%)258M (86.1%)70.1 / 55.570.2 / 55.570.1 / 56.423.8 / 90.023.8 / 90.024.0 / 90.1
ES131K (52.4%)178M (59.4%)55.9 / 34.755.9 / 34.757.8 / 37.522.7 / 84.122.7 / 84.122.3 / 84.2
FR131K (52.4%)178M (59.4%)50.0 / 30.950.0 / 30.948.6 / 29.417.5 / 80.717.5 / 80.716.1 / 79.2
IT111K (44.4%)157M (52.6%)53.2 / 37.653.4 / 37.851.5 / 36.017.6 / 80.817.6 / 80.817.5 / 80.6
JA125K (50.0%)172M (57.6%)65.7 / 65.765.7 / 65.763.0 / 63.029.0 / 80.929.0 / 80.928.6 / 81.0
KO73K (29.2%)119M (39.7%)77.1 / 70.677.1 / 70.574.5 / 67.327.5 / 82.927.5 / 83.028.0 / 83.7
RU147K (58.8%)195M (65.1%)73.7 / 51.473.8 / 51.474.8 / 53.426.4 / 84.326.4 / 84.328.9 / 86.4
mBARTEN173K (69.2%)532M (87.1%)76.9 / 62.677.0 / 62.778.4 / 65.725.1 / 90.425.1 / 90.424.7 / 90.1
ES87K (34.8%)443M (72.7%)64.1 / 42.264.5 / 42.863.7 / 43.922.9 / 83.622.8 / 83.622.8 / 84.0
FR85K (34.0%)442M (72.5%)60.4 / 39.361.0 / 39.866.4 / 45.119.8 / 81.719.8 / 81.718.4 / 79.7
IT67K (26.8%)424M (69.5%)64.7 / 50.064.9 / 50.265.8 / 49.818.0 / 80.617.9 / 80.718.9 / 81.1
JA77K (30.8%)434M (71.1%)68.2 / 68.268.2 / 68.270.6 / 70.630.0 / 82.329.7 / 82.129.1 / 80.8
KO46K (18.4%)402M (65.9%)79.3 / 72.379.2 / 72.183.2 / 77.330.2 / 83.930.3 / 84.030.2 / 83.8
RU99K (39.6%)456M (74.8%)78.7 / 58.079.0 / 58.275.5 / 49.929.3 / 87.228.7 / 87.028.3 / 86.7
", + "image_path": "07fec4db6a433d05003569c769ee5b40771c2d84cb2a8726b46a8aa9aa682381.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 266, + 525, + 315 + ], + "lines": [ + { + "bbox": [ + 67, + 266, + 525, + 315 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 525, + 315 + ], + "type": "text", + "content": "Table 1: Results on QA (Ans-F1/EM) and QG (MTR/BS), including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the parameter size of the original mT5 and mBART (No-Trim) is 300M and 611M, respectively, both with a vocabulary size of 250K." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "text", + "content": "Base language models. As the base LMs, given computational constraints we chose the smallest mT5 and mBART to fine-tune on QA and QG, and XLM-R and XLM-V (Liang et al., 2023) for sentiment analysis and NLI. All these models have a vocabulary size of " + }, + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "inline_equation", + "content": "250\\mathrm{K}" + }, + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "text", + "content": ", except for XLM-V which has a vocabulary size of " + }, + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "inline_equation", + "content": "901\\mathrm{K}" + }, + { + "bbox": [ + 67, + 336, + 290, + 470 + ], + "type": "text", + "content": " subword tokens. For our experiments, we compare the results of pre/post-FT VT against vanilla LM fine-tuning without VT, which we refer to as No-Trim." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 481, + 290, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 290, + 576 + ], + "type": "text", + "content": "Fine-tuning. For model fine-tuning, we rely on lmqg (Ushio et al., 2023) for QA/QG, and Ray Tune3 for sentiment analysis. In both cases, we use the default search space for hyperparameter search. For NLI, we follow the same hyperparameters used in Liang et al. (2023). All the resulting models and code will be released upon acceptance of the paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 586, + 130, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 586, + 130, + 598 + ], + "spans": [ + { + "bbox": [ + 67, + 586, + 130, + 598 + ], + "type": "text", + "content": "4.2 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 604, + 290, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 604, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 604, + 290, + 645 + ], + "type": "text", + "content": "We present the results for the generation and classification tasks in section 4.2.1 and section 4.2.2, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 654, + 240, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 240, + 667 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 240, + 667 + ], + "type": "text", + "content": "4.2.1 Generation Tasks: QA & QG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 671, + 290, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 290, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 290, + 751 + ], + "type": "text", + "content": "Table 1 shows the overall results on QA and QG. The results confirm that both of pre/post-FT VT can maintain the original performance in most cases, while being smaller than the original models by significantly reducing the vocabulary size. First, post-FT VT achieves at least the same performance as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 336, + 525, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 336, + 525, + 498 + ], + "spans": [ + { + "bbox": [ + 301, + 336, + 525, + 498 + ], + "type": "text", + "content": "the vanilla fine-tuning for all the languages for both LMs in QA and QG, except for a few cases such as mBART QA in Korean and mBART QG in Russian, although the decrease is no more than " + }, + { + "bbox": [ + 301, + 336, + 525, + 498 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 301, + 336, + 525, + 498 + ], + "type": "text", + "content": ". Meanwhile, pre-FT VT outperforms its vanilla fine-tuning model with a relatively important margin in some cases, such as mBART French QA and mT5 Spanish QA. In contrast, there are a few models where pre-FT VT degrades the performance of the original model such as mT5 QA in Korean (2.6% decrease in Ans-F1) or mBART QA in Russian (3.2% decrease in Ans-F1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 500, + 525, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 500, + 525, + 608 + ], + "spans": [ + { + "bbox": [ + 302, + 500, + 525, + 608 + ], + "type": "text", + "content": "Since we keep all the vocabulary that appeared in the language-specific corpus " + }, + { + "bbox": [ + 302, + 500, + 525, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_l" + }, + { + "bbox": [ + 302, + 500, + 525, + 608 + ], + "type": "text", + "content": ", the percentage of reduced parameter depends on the language, and generally VT can reduce the model size for Asian (Japanese/Korean) and European (Spanish/French/Italian) languages efficiently (50% for mT5 and 70% for mBART), but it remains high in other languages (English-Russian)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 619, + 519, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 619, + 519, + 632 + ], + "spans": [ + { + "bbox": [ + 302, + 619, + 519, + 632 + ], + "type": "text", + "content": "4.2.2 Classification Tasks: Sentiment & NLI" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "content": "Table 2 shows the results on sentiment analysis and NLI. In this case, post-FT VT can robustly preserve the original performance of the original No-Trim baseline in both tasks for XLM-R and XLM-V, while being no more than " + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "content": " in vocabulary and overall parameter size, respectively, of the original XLM-V and XLM-R models in all the non-English datasets. XLM-V PT sentiment model is the only post-FT VT where a slight decrease can be observed " + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "inline_equation", + "content": "(0.1\\%)" + }, + { + "bbox": [ + 302, + 638, + 525, + 772 + ], + "type": "text", + "content": ". On the other hand, the accuracy" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 289, + 772 + ], + "type": "text", + "content": "3https://docs.ray.io/en/latest/tune/index.html" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14728" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 68, + 503, + 257 + ], + "blocks": [ + { + "bbox": [ + 89, + 68, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 89, + 68, + 503, + 257 + ], + "spans": [ + { + "bbox": [ + 89, + 68, + 503, + 257 + ], + "type": "table", + "html": "
Lang.VocabularyParameterSentimentNLI
No-TrimPost-FTPre-FTNo-TrimPost-FTPre-FT
XLM-RAR49K (19.6%)124M (44.7%)66.366.360.975.775.773.8
DE91K (36.4%)156M (56.3%)73.273.373.579.979.978.3
EN173K (69.2%)219M (78.7%)68.468.567.984.684.670.6
ES87K (34.8%)153M (55.0%)69.069.065.079.879.867.2
FR85K (34.0%)151M (54.6%)71.871.872.180.180.179.6
IT67K (26.8%)138M (49.7%)62.962.970.8---
PT66K (26.4%)137M (49.3%)70.770.870.2---
XLM-VAR92K (11.8%)157M (20.2%)59.859.864.775.575.676.1
DE239K (30.7%)269M (34.7%)73.573.573.078.978.979.0
EN484K (62.2%)458M (58.9%)63.963.961.384.484.484.5
ES243K (31.2%)279M (35.1%)60.760.766.680.780.780.6
FR218K (28.0%)253M (32.6%)68.868.859.578.678.679.0
IT184K (23.7%)227M (29.3%)70.270.274.2---
PT181K (23.3%)225M (28.9%)66.666.552.8---
", + "image_path": "a62350e762289aa4a4e077c0684f5ce0a08f0b58c8c5d63d47c088735e4b66d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 267, + 525, + 317 + ], + "lines": [ + { + "bbox": [ + 67, + 267, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 525, + 317 + ], + "type": "text", + "content": "Table 2: Results of sentiment analysis (macro F1) and XNLI (accuracy) including both the vocabulary size and the number of parameters after VT with the ratio to the original model (\\%). The best results in each LM and language are in bold characters. Note that the overall parameter size of the original XLM-R and XLM-V (No-Trim) is 278M and 778M, respectively, with the vocabulary size being 250K and 901K vocabulary in each case." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "spans": [ + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": "of Pre-FT VT appears to be sensitive to the language and task, where it improves the performance in some languages such as Italian (XLM-R and XLM-V achieve " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " increase for sentiment analysis), but it decreases the performance with non-trivial margin in other languages such as Arabic, where XLM-R decreases " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " for sentiment analysis and " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " for XNLI. Since XLM-V has a larger vocabulary size, the percentage of reduced parameters at VT is more prominent in XLM-V, as seen in Arabic " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "(20.2\\%)" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " and Portuguese " + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "inline_equation", + "content": "(28.9\\%)" + }, + { + "bbox": [ + 66, + 336, + 291, + 498 + ], + "type": "text", + "content": " for example." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 507, + 218, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 507, + 218, + 522 + ], + "spans": [ + { + "bbox": [ + 67, + 507, + 218, + 522 + ], + "type": "text", + "content": "5 Vocabulary Size Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "text", + "content": "In our main experimental results (§ 4.2), all the unique tokens that appeared in the monolingual corpus were kept, which resulted in a low compression ratio for some languages such as English and Russian. In this analysis, we constrain the number of vocabulary and choose the top-" + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "text", + "content": " vocabulary at VT in terms of the frequency in the corpus (see § 3). For QA and QG, we compare mT5SMALL results with " + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "text", + "content": " from [5K, 10K, 15K, 30K, 60K, 90K, 120K], which correspond to an overall parameter size of [49M, 54M, 59M, 74M, 105M, 136M, 166M], respectively. For sentiment analysis and NLI, we experiment with XLM-RBASE with " + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 66, + 528, + 291, + 745 + ], + "type": "text", + "content": " from [5K, 10K, 15K, 30K, 60K], which correspond to [89M, 93M, 97M, 109M, 132M] of parameter size, respectively4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "4The full results with top- " + }, + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": " can be found at Appendix B and Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 336, + 468, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 336, + 468, + 349 + ], + "spans": [ + { + "bbox": [ + 302, + 336, + 468, + 349 + ], + "type": "text", + "content": "5.1 Generation Tasks: QA & QG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "spans": [ + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": "Figure 4 shows the results of mT5 on QA and QG. Noticeably, post-FT VT can reduce the vocabulary size to " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "60\\mathrm{K}" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": " for both QA and QG in all the languages with a trivial gap (0.3% decrease of EM in Russian QA and 0.1% decrease of BS in French QG), and that is " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": " of the original mT5 in the parameter size. Furthermore, post-FT VT can further reduce the vocabulary to 5K tokens with no more than " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": " decrease in each metric for both QA and QG in English, French, and Italian, which is " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": " of the original mT5 in the parameter size. Meanwhile, pre-FT VT outperforms the No-Trim result in all the languages in QA, and the majority of the languages in QG (English, Italian, Korean, and Russian), but the result is sensitive to the choice of " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": ". For example, Japanese/Korean QA and Russian QG with pre-FT VT for top-5K (16% of the original mT5) outperforms No-Trim as well as post-FT VT, but Japanese QG with pre-FT VT is worse in any choice of " + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 301, + 354, + 527, + 691 + ], + "type": "text", + "content": " on contrary. This larger variation of results may also be due to the parameter size space, as the optimal parameters for the original multilingual LM (which is the one trained for post-FT VT) may differ. We leave this extended analysis for future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 701, + 511, + 714 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 701, + 511, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 701, + 511, + 714 + ], + "type": "text", + "content": "5.2 Classification Tasks: Sentiment & NLI" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "content": "Figure 5 and Figure 6 show the results of XLM-R on sentiment and NLI. In NLI, we can see that post/pre-FT VT both can reduce the vocabulary to " + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "inline_equation", + "content": "30\\mathrm{K}" + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 302, + 719, + 526, + 772 + ], + "type": "text", + "content": " of the original XLM-R in the parameter" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14729" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 179, + 124 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 179, + 124 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 179, + 124 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 179, + 124 + ], + "type": "image", + "image_path": "5e6650c638298516651a68d4acd871db90087e97bf0034e116a4a47a4562c1da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 180, + 69, + 287, + 124 + ], + "blocks": [ + { + "bbox": [ + 180, + 69, + 287, + 124 + ], + "lines": [ + { + "bbox": [ + 180, + 69, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 180, + 69, + 287, + 124 + ], + "type": "image", + "image_path": "214b1cb5de93ffa682717eca311d74b5ddd98860b07b847b9bc40f92e49b2e8c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 74, + 127, + 178, + 181 + ], + "blocks": [ + { + "bbox": [ + 74, + 127, + 178, + 181 + ], + "lines": [ + { + "bbox": [ + 74, + 127, + 178, + 181 + ], + "spans": [ + { + "bbox": [ + 74, + 127, + 178, + 181 + ], + "type": "image", + "image_path": "4ef4be96790435dcb576decfd929b245981d80679335f72a4804c3fc3200b162.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 184, + 127, + 287, + 181 + ], + "blocks": [ + { + "bbox": [ + 184, + 127, + 287, + 181 + ], + "lines": [ + { + "bbox": [ + 184, + 127, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 184, + 127, + 287, + 181 + ], + "type": "image", + "image_path": "745560447b5a3de1acb235070cd9744770219f7571f3d1aad6265b035e465ccf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 74, + 185, + 178, + 238 + ], + "blocks": [ + { + "bbox": [ + 74, + 185, + 178, + 238 + ], + "lines": [ + { + "bbox": [ + 74, + 185, + 178, + 238 + ], + "spans": [ + { + "bbox": [ + 74, + 185, + 178, + 238 + ], + "type": "image", + "image_path": "715ebe8415fbefb1a0097708d1e4901076b019320b6bb99ee83d76accb0d6567.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 185, + 287, + 238 + ], + "blocks": [ + { + "bbox": [ + 185, + 185, + 287, + 238 + ], + "lines": [ + { + "bbox": [ + 185, + 185, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 185, + 185, + 287, + 238 + ], + "type": "image", + "image_path": "b54ba16a3c7b227762bc66d0dbee6a7528d0368e9a2239f999f2fad2432123eb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 74, + 243, + 178, + 296 + ], + "blocks": [ + { + "bbox": [ + 74, + 243, + 178, + 296 + ], + "lines": [ + { + "bbox": [ + 74, + 243, + 178, + 296 + ], + "spans": [ + { + "bbox": [ + 74, + 243, + 178, + 296 + ], + "type": "image", + "image_path": "9a090bc793a0a1869516bb430ca16bd11094503d3f563323833eda28867efc53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 243, + 287, + 296 + ], + "blocks": [ + { + "bbox": [ + 181, + 243, + 287, + 296 + ], + "lines": [ + { + "bbox": [ + 181, + 243, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 181, + 243, + 287, + 296 + ], + "type": "image", + "image_path": "75c098b5fc4c7ffe757384015efa7fe788d23b6d0edb6e41874109334c440c08.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 316, + 526, + 352 + ], + "lines": [ + { + "bbox": [ + 302, + 316, + 526, + 352 + ], + "spans": [ + { + "bbox": [ + 302, + 316, + 526, + 352 + ], + "type": "text", + "content": "Figure 5: Sentiment analysis macro-F1 results of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 74, + 301, + 178, + 354 + ], + "blocks": [ + { + "bbox": [ + 74, + 301, + 178, + 354 + ], + "lines": [ + { + "bbox": [ + 74, + 301, + 178, + 354 + ], + "spans": [ + { + "bbox": [ + 74, + 301, + 178, + 354 + ], + "type": "image", + "image_path": "47f32c80e7c2e10c226f0396803a7183389b3f1704257858fc712dacfdb6f84d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 185, + 301, + 287, + 354 + ], + "blocks": [ + { + "bbox": [ + 185, + 301, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 185, + 301, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 185, + 301, + 287, + 354 + ], + "type": "image", + "image_path": "24170ba9678d8eac52dabf8ca9c46bde1a4fce44ac8cd021b7043becbd112525.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 74, + 358, + 178, + 411 + ], + "blocks": [ + { + "bbox": [ + 74, + 358, + 178, + 411 + ], + "lines": [ + { + "bbox": [ + 74, + 358, + 178, + 411 + ], + "spans": [ + { + "bbox": [ + 74, + 358, + 178, + 411 + ], + "type": "image", + "image_path": "6b757cbcbd91cfa363e40352c2e94bd30605250aa42e034ae9b3550e5c7e42ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 181, + 358, + 287, + 412 + ], + "blocks": [ + { + "bbox": [ + 181, + 358, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 181, + 358, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 181, + 358, + 287, + 412 + ], + "type": "image", + "image_path": "85215f179d7c7ebb6e1a627c6d011a684c3ba3ba329818ee70d99c65976aacc2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 74, + 415, + 178, + 476 + ], + "blocks": [ + { + "bbox": [ + 74, + 415, + 178, + 476 + ], + "lines": [ + { + "bbox": [ + 74, + 415, + 178, + 476 + ], + "spans": [ + { + "bbox": [ + 74, + 415, + 178, + 476 + ], + "type": "image", + "image_path": "afe7ff6bb4af375fdc3f9374c1aec20d7c7c4ffcf56a69f3ea9a4e923512bd42.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 485, + 290, + 523 + ], + "lines": [ + { + "bbox": [ + 67, + 485, + 290, + 523 + ], + "spans": [ + { + "bbox": [ + 67, + 485, + 290, + 523 + ], + "type": "text", + "content": "Figure 4: QG (METEOR) and QA (Ans-F1) results for mT5 with pre/post-FT VT for different vocabulary sizes compared to the original multilingual LM (No-Trim)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 185, + 415, + 287, + 476 + ], + "blocks": [ + { + "bbox": [ + 185, + 415, + 287, + 476 + ], + "lines": [ + { + "bbox": [ + 185, + 415, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 185, + 415, + 287, + 476 + ], + "type": "image", + "image_path": "b4cf891736fc8564d14bf79275ea364660fedb30dac1ed1cacc0e91411c9fe5b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "content": "size) without any decrease except " + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "content": " of pre-FT VT for German, and there is no decrease more than " + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "content": " even with top-15K of post-FT VT. In sentiment analysis, pre-FT VT with top-10K (" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "content": " of the original XLM-R in the parameter size) can retain the accuracy of the No-Trim baseline in French and Italian. Moreover, post-FT VT with 30K can retain the original F1 score without a major drop in sentiment analysis, yet the decrease in F1 score is slightly more prominent than NLI (" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 67, + 543, + 290, + 691 + ], + "type": "text", + "content": " in Arabic sentiment analysis)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "type": "text", + "content": "The sentiment analysis datasets are collected from Twitter, so one dataset in a single language can contain tokens from other languages ( hashtags or named-entities, or even code-switching). In contrast, XNLI translates English NLI into other languages, so there is less chance for a dataset to" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 417, + 125 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 417, + 125 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 417, + 125 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 417, + 125 + ], + "type": "image", + "image_path": "146d9236bbdef860d7b91976f3d8c5f137a736392dfaf7a613addd91ab956bbc.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 419, + 69, + 524, + 125 + ], + "blocks": [ + { + "bbox": [ + 419, + 69, + 524, + 125 + ], + "lines": [ + { + "bbox": [ + 419, + 69, + 524, + 125 + ], + "spans": [ + { + "bbox": [ + 419, + 69, + 524, + 125 + ], + "type": "image", + "image_path": "cfc291a843c11054b28ea043e7f22765b94502110b2ffac101a2776168a0c008.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 305, + 128, + 417, + 184 + ], + "blocks": [ + { + "bbox": [ + 305, + 128, + 417, + 184 + ], + "lines": [ + { + "bbox": [ + 305, + 128, + 417, + 184 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 417, + 184 + ], + "type": "image", + "image_path": "65b148e2aa17e46779763a15938c58f8f311d869565ff8de7e2718adfc9dbb77.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 419, + 128, + 524, + 183 + ], + "blocks": [ + { + "bbox": [ + 419, + 128, + 524, + 183 + ], + "lines": [ + { + "bbox": [ + 419, + 128, + 524, + 183 + ], + "spans": [ + { + "bbox": [ + 419, + 128, + 524, + 183 + ], + "type": "image", + "image_path": "ccde1b462d6a6e11a8a61bf13ea45d0247ec5eb000e3d85060a92bfcf13aaeb4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 305, + 187, + 417, + 242 + ], + "blocks": [ + { + "bbox": [ + 305, + 187, + 417, + 242 + ], + "lines": [ + { + "bbox": [ + 305, + 187, + 417, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 417, + 242 + ], + "type": "image", + "image_path": "98d1ff5f5094a977fb4764fafc0ca6667516a7a4ee0f30ca8467823a3b877e6c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 419, + 187, + 524, + 241 + ], + "blocks": [ + { + "bbox": [ + 419, + 187, + 524, + 241 + ], + "lines": [ + { + "bbox": [ + 419, + 187, + 524, + 241 + ], + "spans": [ + { + "bbox": [ + 419, + 187, + 524, + 241 + ], + "type": "image", + "image_path": "8f017747f0f32250fc7097711d2769fc2ff5542f0931cd947827cd75add6337d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 305, + 245, + 417, + 306 + ], + "blocks": [ + { + "bbox": [ + 305, + 245, + 417, + 306 + ], + "lines": [ + { + "bbox": [ + 305, + 245, + 417, + 306 + ], + "spans": [ + { + "bbox": [ + 305, + 245, + 417, + 306 + ], + "type": "image", + "image_path": "66c339afbae3af67690ff9ad56f6178f87a765586078dcdfeacdc4bc86401f50.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 305, + 364, + 414, + 415 + ], + "blocks": [ + { + "bbox": [ + 305, + 364, + 414, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 364, + 414, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 414, + 415 + ], + "type": "image", + "image_path": "ba2036ff80ce83fd6878f1d359b292a508134833619f586f8b62f78aec94c5c2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 421, + 364, + 524, + 415 + ], + "blocks": [ + { + "bbox": [ + 421, + 364, + 524, + 415 + ], + "lines": [ + { + "bbox": [ + 421, + 364, + 524, + 415 + ], + "spans": [ + { + "bbox": [ + 421, + 364, + 524, + 415 + ], + "type": "image", + "image_path": "4721b40d5c2aeac46f6042c96d91ea63d05e992dba1472f0f01bc7dacc07d33d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 305, + 418, + 414, + 470 + ], + "blocks": [ + { + "bbox": [ + 305, + 418, + 414, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 418, + 414, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 414, + 470 + ], + "type": "image", + "image_path": "ee2bf3a8a0cae4e99d677ced953d5341b1fd2ab55ecbfdda0c44388ac3e12256.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 417, + 419, + 524, + 470 + ], + "blocks": [ + { + "bbox": [ + 417, + 419, + 524, + 470 + ], + "lines": [ + { + "bbox": [ + 417, + 419, + 524, + 470 + ], + "spans": [ + { + "bbox": [ + 417, + 419, + 524, + 470 + ], + "type": "image", + "image_path": "440cd0ff6cd1f351989c873d2f0a9cf73696ba90d62230c2167003f54be7976c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 305, + 474, + 414, + 524 + ], + "blocks": [ + { + "bbox": [ + 305, + 474, + 414, + 524 + ], + "lines": [ + { + "bbox": [ + 305, + 474, + 414, + 524 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 414, + 524 + ], + "type": "image", + "image_path": "8bf9aa983690495010bd0429eaa962a0f9254995a2795bc0b8126bf7d24aac4a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 534, + 525, + 560 + ], + "lines": [ + { + "bbox": [ + 302, + 534, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 302, + 534, + 525, + 560 + ], + "type": "text", + "content": "Figure 6: NLI accuracy of XLM-R with pre/post-FT VT for different vocabulary sizes compared to No-Trim." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "text", + "content": "contain tokens from other languages. This can explain the effectiveness of top-" + }, + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "text", + "content": " VT in NLI compared with sentiment analysis, as smaller values of " + }, + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 580, + 525, + 661 + ], + "type": "text", + "content": " should result in a vocabulary with fewer tokens from the other languages, which limits the ability of the models to handle foreign tokens." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 302, + 670, + 525, + 697 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 670, + 525, + 697 + ], + "spans": [ + { + "bbox": [ + 302, + 670, + 525, + 697 + ], + "type": "text", + "content": "6 Monolingual vs. Multilingual LMs: The Case of Social Bias" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 302, + 705, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 526, + 772 + ], + "type": "text", + "content": "There has been extensive literature in NLP comparing monolingual and multilingual LMs (Muller et al., 2021; Goyal et al., 2021). As for the performance, there is no clear consensus on which type is better for certain languages, tasks or settings. How-" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14730" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 452 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 452 + ], + "type": "text", + "content": "ever, there are other important factors that play a role in this comparison. First, monolingual models tend to have a smaller vocabulary size, which makes them more practical. In contrast, a single multilingual model can be used for a large number of languages. Moreover, multilingual LMs are less prone to capture and carry cultural- or language-dependent biases. This is due to the combination of languages and cultures into a single model, which makes it less biased toward specific cultures (Liang et al., 2020; Ahn and Oh, 2021). Prior works have shown that different types of biases consistently appear in language-specific models (Nadeem et al., 2021; Nangia et al., 2020; Blodgett et al., 2021; Dhamala et al., 2021; Kaneko et al., 2022; Zhou et al., 2022). While the comparison of monolingual and multilingual LMs is not the main focus of this paper, this analysis is certainly relevant. Trimming the vocabulary of a multilingual model essentially makes the model smaller, and therefore alleviates one of the main drawbacks of using multilingual language models on language-specific tasks, which is its larger size. On top of that, this strategy enables the usage of monolingual models with potentially less social bias. In the following, we present a comparison of monolingual and multilingual LMs (both trimmed and not trimmed) in terms of social bias and general performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 468, + 194, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 468, + 194, + 481 + ], + "spans": [ + { + "bbox": [ + 67, + 468, + 194, + 481 + ], + "type": "text", + "content": "6.1 Experimental setting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "text", + "content": "Social bias datasets. To study the effect of VT on social bias existing in pre-trained LMs, we first conduct experiments on two commonly used social bias evaluation datasets for masked LMs: StereoSet (SS; Nadeem et al., 2021)" + }, + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "inline_equation", + "content": "^5" + }, + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "text", + "content": " and crowdsourced stereotype pairs benchmark (CP; Nangia et al., 2020)" + }, + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "inline_equation", + "content": "^6" + }, + { + "bbox": [ + 67, + 490, + 291, + 734 + ], + "type": "text", + "content": ". SS consists of associative contexts covering four types of social biases: race, gender, religion, and profession; while CP is crowdsourced and annotated by workers in the United States, which contains nine types of social biases: race, gender, sexual orientation, religion, age, nationality, disability, physical appearance, and socioeconomic status. In order to further investigate the impact of pre/post-FT VT on LMs, we trim and fine-tune models on sentiment analysis with different orders and evaluate the social bias in such models on the Equity Evaluation Corpus (EEC; Kiritchenko and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 302, + 70, + 526, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 70, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 526, + 126 + ], + "type": "text", + "content": "Mohammad, 2018) considering two bias types: gender and race. The EEC dataset was specifically with the aim to examine social bias for sentiment analysis systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 135, + 527, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 135, + 527, + 609 + ], + "spans": [ + { + "bbox": [ + 302, + 135, + 527, + 609 + ], + "type": "text", + "content": "Evaluation metrics. We compare the pseudolikelihood scores returned by each model for stereotypical and anti-stereotypical sentences using AULA (All Unmasked Likelihood with Attention weights) (Kaneko and Bollegala, 2022).8 AULA has been shown to be robust against the frequency biases of the masked tokens and provides a more reliable assessment in contrast to alternative metrics when evaluating social biases in masked language models (MLMs). Given a sentence pair in the test dataset: \"My mom spent all day cooking for Thanksgiving\" vs. \"My dad spent all day cooking for Thanksgiving\", the first sentence is considered as stereotypical while the second one is anti-stereotypical. AULA computes the percentage of stereotypical sentences preferred by the MLM over anti-stereotypical ones as the bias score. An MLM is considered to be unfairly biased if it returns higher pseudo-loglikelihood scores for stereotypical sentences than the corresponding anti-stereotypical sentences. The AULA score falls within the range of [0,100] and an unbiased model would return a bias score close to 50. On the other hand, a bias score greater than or less than 50 indicates the bias direction towards the stereotype or anti-stereotype, respectively. Since the original AULA is not fitted to evaluate fine-tuned models, we adapt AULA to the EEC dataset obtain the bias score for the LMs fine-tuned on sentiment analysis, and denote this metric as EEC-AULA. Specifically, given a model that predicts sentiment labels (e.g., positive, neutral, negative) to sentences, we consider the percentage of stereotypical test sentences with a more negative label over anti-stereotypical ones as the corresponding bias evaluation measure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 619, + 526, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 619, + 526, + 700 + ], + "spans": [ + { + "bbox": [ + 302, + 619, + 526, + 700 + ], + "type": "text", + "content": "General performance. As a proxy to test the general performance, we use the general language understanding evaluation (GLUE; Wang et al., 2018) benchmark. We acknowledge the limitations of using this benchmark to draw reliable conclusions at large (Ethayarajh and Jurafsky, 2020)" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 709, + 524, + 729 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 709, + 524, + 729 + ], + "spans": [ + { + "bbox": [ + 302, + 709, + 524, + 729 + ], + "type": "text", + "content": "7https://saifmohammad.com/WebPages/Biases-SA.html" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "spans": [ + { + "bbox": [ + 302, + 729, + 525, + 750 + ], + "type": "text", + "content": "https://github.com/kanekomasahiro/evaluate.bias_in_m1m" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "inline_equation", + "content": "^{9}" + }, + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "https://gluebenchmark.com/; Models are tested on the development sets of each task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 749, + 262, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 749, + 262, + 761 + ], + "spans": [ + { + "bbox": [ + 80, + 749, + 262, + 761 + ], + "type": "text", + "content": "5https://github.com/moinnadeem/StereoSet" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 761, + 258, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 761, + 258, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 761, + 258, + 772 + ], + "type": "text", + "content": "6https://github.com/nyu-mll/crows-pairs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "14731" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 523, + 193 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 523, + 193 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 523, + 193 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 523, + 193 + ], + "type": "table", + "html": "
Base ModelVocab TrimmingVocabularySocial BiasGeneral Performance
Pre-FTPost-FTAULAEEC-AULAGLUE
CPSSGenderRace
Monolingual (RoBERTa)-50K50K58.158.864.885.778.0
Multilingual (XLM-R)-250K250K49.554.944.362.077.9
Pre-FT VT (EN)173K173K49.554.942.556.978.0
Post-FT VT (EN)250K173k49.554.944.362.077.9
Pre-FT VT (50K)50K50K49.355.041.062.476.9
Post-FT VT (50K)250K50K49.554.944.362.077.9
", + "image_path": "e5e69a86edc9cc19c07171eef4776f4c03d84a4413a0b7e484acfc5858960d6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 201, + 526, + 287 + ], + "lines": [ + { + "bbox": [ + 67, + 201, + 526, + 287 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 526, + 287 + ], + "type": "text", + "content": "Table 3: Results of pre/post-FT VT models compared with the original monolingual and multilingual models on two social bias analysis benchmarks (AULA for pre-trained masked language models and EEC-AULA for models fine-tuned on sentiment analysis) and the GLUE benchmark. The VT models are trimmed on English vocabulary with different vocabulary sizes: EN (full English vocabulary) and 50K (top 50K subword tokens). Note that for post-FT VT, the results on AULA are exactly the same as the original XLM-R. The green and red colours represent the social bias towards anti-stereotypical sentences (scores lower than 50) and stereotypical sentences (scores higher than 50), respectively. The lighter colour indicates less social bias observed in the LM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 307, + 290, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 307, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 307, + 290, + 375 + ], + "type": "text", + "content": "but it nevertheless provides a good proxy for understanding the overall performance of comparable models in standard NLP tasks. Moreover, these experiments are only aimed at analysing the effect of vocabulary trimming on general performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 381, + 291, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 381, + 291, + 598 + ], + "spans": [ + { + "bbox": [ + 69, + 381, + 291, + 598 + ], + "type": "text", + "content": "Models. We compute the bias scores of RoBERTa (Liu et al., 2019) as base monolingual LM, and XLM-R (Conneau et al., 2019) as its multilingual counterpart (they have been trained with the same architecture and in an overlapping corpus. We explore two VT settings to be applied to XLM-R: XLM-R with the standard VT including the full English vocabulary (VT XLM-R) and XLM-R with VT for top-50K English vocabulary (top-50K VT XLM-R), which is the same vocabulary size as the monolingual RoBERTa model. Our experiments are based both on masked language models on AULA (in which the post-VT does not have any effect) and models fine-tuned on the sentiment analysis presented in " + }, + { + "bbox": [ + 69, + 381, + 291, + 598 + ], + "type": "inline_equation", + "content": "\\S 4.1" + }, + { + "bbox": [ + 69, + 381, + 291, + 598 + ], + "type": "text", + "content": " on EEC-AULA, as well as on the corresponding GLUE training sets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 607, + 130, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 607, + 130, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 607, + 130, + 619 + ], + "type": "text", + "content": "6.2 Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "content": "Table 3 shows the performance of pre-FT and post-FT VT models against the original monolingual and multilingual LMs on social bias evaluation datasets and the GLUE benchmark. Both AULA and GLUE results are computed using the LMs without fine-tuning (i.e., RoBERTa, XLM-R, VT XLM-R, and top-50K VT XLM-R), whereas the EEC-AULA results are computed using the models applying VT and fine-tuning strategies. We observe that the monolingual model contains the highest levels of social biases compared to the multilin" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 307, + 526, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 307, + 526, + 497 + ], + "spans": [ + { + "bbox": [ + 302, + 307, + 526, + 497 + ], + "type": "text", + "content": "gual models with different settings. In particular, RoBERTa obtains the overall highest bias score on the EEC dataset after fine-tuning, with an alarmingly high 85.7 score on race." + }, + { + "bbox": [ + 302, + 307, + 526, + 497 + ], + "type": "inline_equation", + "content": "^{10}" + }, + { + "bbox": [ + 302, + 307, + 526, + 497 + ], + "type": "text", + "content": " On the other hand, compared to the original XLM-R, there is no significant change in performance on social biases and GLUE evaluation tasks for pre-FT VT and post-FT VT models. This is important as we can apply the proposed VT method to any multilingual LM, obtaining a monolingual one with consistent performance on the GLUE benchmark and less social biases than the original monolingual model pretrained in the target language, without using any ad-hoc debiasing methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 506, + 379, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 506, + 379, + 519 + ], + "spans": [ + { + "bbox": [ + 303, + 506, + 379, + 519 + ], + "type": "text", + "content": "7 Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 528, + 525, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 528, + 525, + 743 + ], + "spans": [ + { + "bbox": [ + 302, + 528, + 525, + 743 + ], + "type": "text", + "content": "Vocabulary trimming before and after finetuning. According to the results, pre-FT VT appears to be generally more effective in classification tasks (see section 4.2.2). For generation tasks, both pre/post-FT VT robustly retain the original performance while being able to considerably reduce the model size (see section 4.2.1). As a guideline to choose the type of VT in such a case, post-FT VT should be more suitable if one already has a finetuned model, as no additional training is needed for this case. Moreover, post-FT is more robust as a compression mechanism as the performance is largely maintained with respect to that of the original multilingual LM. On the other hand, if one needs to fine-tune a model from scratch and the computation resources are limited, we recommend" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "10 Appendix C includes a breakdown by type of the social bias results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14732" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 166 + ], + "type": "text", + "content": "exploring pre-FT VT, as fine-tuning on a trimmed LM should be more efficient due to its smaller vocabulary and parameters and, in some cases, can lead to better overall results. However, this process has to be done carefully as the set of optimal parameters could differ from the original multilingual LM fine-tuning process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 174, + 291, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 174, + 291, + 512 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 291, + 512 + ], + "type": "text", + "content": "While in this paper we have not compared monolingual and multilingual models, the question would be whether we need vocabulary trimming strategies in a world where monolingual LMs exist. In this case, a monolingual model may perform similarly to a multilingual model (Goyal et al., 2021). However, the multilingual model is often larger mainly due to larger vocabulary storage requirements. In contrast, our proposed VT technique does not require any extra LM training or computational resources. Indeed, only a multilingual LM is needed and we can induce multiple smaller language-specific monolingual models. This may reduce the carbon footprint overall and especially help with less-resource languages when a high-quality monolingual model does not exist. Finally, our social bias analysis see § 6 shows how monolingual models exhibit larger social biases (especially racial) than a VT-induced multilingual LM. This is consistent with prior work suggesting that a multilingual LM has been trained with more languages, and hence more cultural variety, and these diverging viewpoints can compensate each other (Ahn and Oh, 2021)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 521, + 147, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 521, + 147, + 534 + ], + "spans": [ + { + "bbox": [ + 67, + 521, + 147, + 534 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": "In this paper, we proposed vocabulary-trimming (VT), a method to reduce the vocabulary of a multilingual LM to a vocabulary specific to any given target language. VT can induce a monolingual LM in the target language by leveraging an existing multilingual LM. The main advantage of this filtering step is the reduced size, as well as avoiding having to train monolingual LMs from scratch, which would be computationally demanding. Our experiments show how VT can retain the high performance of the original multilingual LM, while largely reducing the model size. For all languages evaluated, a " + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": " compression rate proves sufficient to keep the original performance of the larger mT5 multilingual LM in both QA and QG, with a similar " + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": " in NLI and " + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 67, + 543, + 291, + 772 + ], + "type": "text", + "content": " in sentiment analysis with XLM-R. Interestingly, in some cases, the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 152 + ], + "type": "text", + "content": "compressed LM can even achieve better results than the original larger model when trimmed before fine-tuning. Since the main goal of the paper was to compress a multilingual LM while keeping its original performance, we leave the analysis of this behaviour for future work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 163, + 366, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 163, + 366, + 176 + ], + "spans": [ + { + "bbox": [ + 303, + 163, + 366, + 176 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 185, + 526, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 185, + 526, + 374 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 526, + 374 + ], + "type": "text", + "content": "We have not tested our methodology in truly low-resource languages. Because of this, there could be a different behaviour when we apply VT to a language with lower resources or that is poorly represented in the underlying training corpus. The LMs we used in the paper limited their size up to 600M, and we have not considered larger models such as mT5XXL or BLOOM (Scao et al., 2022), due to our limited computational resources. As the language-specific corpus to compute frequency, we employ mC4, which is one of the largest multilingual corpora. Nonetheless, this is used as a proxy and having access to the full multilingual model could give potentially better results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 375, + 526, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 375, + 526, + 483 + ], + "spans": [ + { + "bbox": [ + 302, + 375, + 526, + 483 + ], + "type": "text", + "content": "Similarly, we acknowledge the limitations of the analysis comparing multilingual and monolingual models in terms of social bias. Due to the evaluation data available and the existence of comparable monolingual and multilingual LMs, the evaluation is focused on English only and the results could differ for other languages. Moreover, there are other types of biases not covered in this evaluation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 494, + 393, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 494, + 393, + 507 + ], + "spans": [ + { + "bbox": [ + 303, + 494, + 393, + 507 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 516, + 526, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 526, + 624 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 526, + 624 + ], + "type": "text", + "content": "Pre-trained LMs are known to contain undesirable biases to generate toxic contents in some edge cases (Schick et al., 2021), so the resulting models could inherit such biases. While we have not analysed in detail the output of all models in the tasks evaluated, in this paper we have made an attempt to study this effect in terms of social biases for both base pretrained LMs and fine-tuned LMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 635, + 407, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 635, + 407, + 649 + ], + "spans": [ + { + "bbox": [ + 303, + 635, + 407, + 649 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 657, + 524, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 657, + 524, + 684 + ], + "spans": [ + { + "bbox": [ + 302, + 657, + 524, + 684 + ], + "type": "text", + "content": "Jose Camacho-Collados and Yi Zhou are supported by a UKRI Future Leaders Fellowship." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 708, + 362, + 721 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 708, + 362, + 721 + ], + "spans": [ + { + "bbox": [ + 304, + 708, + 362, + 721 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 526, + 772 + ], + "type": "text", + "content": "Amine Abdaoui, Camille Pradel, and Grégoire Sigel. 2020. Load what you need: Smaller versions of mutililingual BERT. In Proceedings of SustainNLP: Workshop on Simple and Efficient Natural Language" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14733" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 95 + ], + "type": "text", + "content": "Processing, pages 119-123, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 69, + 104, + 289, + 170 + ], + "type": "text", + "content": "Jaimeen Ahn and Alice Oh. 2021. Mitigating language-dependent ethnic bias in BERT. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 533-549, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 179, + 289, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 179, + 289, + 245 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 289, + 245 + ], + "type": "text", + "content": "Mikel Artetxe, Sebastian Ruder, and Dani Yogatama. 2020. On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4623–4637, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 254, + 289, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 254, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 254, + 289, + 342 + ], + "type": "text", + "content": "Francesco Barbieri, Valerio Basile, Danilo Croce, Malvina Nissim, Nicole Novielli, and Viviana Patti. 2016. Overview of the evalita 2016 sentiment polarity classification task. In Proceedings of third Italian conference on computational linguistics (CLiC-it 2016) & fifth evaluation campaign of natural language processing and speech tools for Italian. Final Workshop (EVALITA 2016)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 351, + 289, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 351, + 289, + 428 + ], + "spans": [ + { + "bbox": [ + 69, + 351, + 289, + 428 + ], + "type": "text", + "content": "Francesco Barbieri, Luis Espinosa Anke, and Jose Camacho-Collados. 2022. XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 258-266, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 438, + 289, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 438, + 289, + 481 + ], + "spans": [ + { + "bbox": [ + 69, + 438, + 289, + 481 + ], + "type": "text", + "content": "Farah Benamara, Cyril Grouin, Jihen Karoui, Véronique Moriceau, and Isabelle Robba. 2017. Analyse d'opinion et langage figuratif dans des tweets: presentation et résultats du début fouille de textes deft2017." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 289, + 579 + ], + "type": "text", + "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. 2021. Stereotyping norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1004-1015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 588, + 289, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 588, + 289, + 631 + ], + "spans": [ + { + "bbox": [ + 69, + 588, + 289, + 631 + ], + "type": "text", + "content": "Henrico Bertini Brum and Maria das Graças Volpe Nunes. 2017. Building a sentiment corpus of tweets in brazilian portuguese. arXiv preprint arXiv:1712.08917." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 641, + 289, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 641, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 69, + 641, + 289, + 686 + ], + "type": "text", + "content": "Carrino Casimiro Pio, Costa-jussa Marta R., and Fonollosa Jose A. R. 2019. Automatic Spanish Translation of the SQuAD Dataset for Multilingual Question Answering. arXiv e-prints, page arXiv:1912.05200v1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 289, + 772 + ], + "type": "text", + "content": "Hyung Won Chung, Dan Garrette, Kiat Chuan Tan, and Jason Riesa. 2020. Improving multilingual models with language-clustered vocabularies. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4536-4546, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 149 + ], + "type": "text", + "content": "Mark Cieliebak, Jan Milan Deriu, Dominic Egger, and Fatih Uzdilli. 2017. A twitter corpus and benchmark resources for german sentiment analysis. In 5th International Workshop on Natural Language Processing for Social Media, Boston MA, USA, 11 December 2017, pages 45-51. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 524, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 524, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 524, + 226 + ], + "type": "text", + "content": "Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. arXiv preprint arXiv:1911.02116." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 236, + 524, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 524, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 524, + 269 + ], + "type": "text", + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 280, + 524, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 524, + 367 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 524, + 367 + ], + "type": "text", + "content": "Alexis Conneau, Rudy Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating crosslingual sentence representations. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2475-2485, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 377, + 524, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 524, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 524, + 433 + ], + "type": "text", + "content": "Danilo Croce, Alexandra Zelenanska, and Roberto Basili. 2018. Neural learning for question answering in italian. In AI*IA 2018 – Advances in Artificial Intelligence, pages 389–402, Cham. Springer International Publishing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 443, + 524, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 524, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 524, + 542 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 552, + 524, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 524, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 524, + 628 + ], + "type": "text", + "content": "Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. 2021. Bold: Dataset and metrics for measuring biases in open-ended language generation. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 862-872." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 640, + 524, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 524, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 524, + 706 + ], + "type": "text", + "content": "Martin d'Hoffschmidt, Wacim Belblidia, Quentin Heinrich, Tom Brendle, and Maxime Vidal. 2020. FQuAD: French question answering dataset. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1193–1208, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 715, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 715, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 715, + 524, + 772 + ], + "type": "text", + "content": "Manuel C Díaz-Galiano, Eugenio Martínez-Cármara, M Ángel García Cumbreras, Manuel García Vega, and Julio Villena Román. 2018. The democratization of deep learning in tass 2017. Procesamento del Lenguaje Natural, 60:37-44." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14734" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 289, + 139 + ], + "type": "text", + "content": "Pavel Efimov, Andrey Chertok, Leonid Boytsov, and Pavel Braslavski. 2020. Sberquad-russian reading comprehension dataset: Description and analysis. In International Conference of the Cross-Language Evaluation Forum for European Languages, pages 3-15. Springer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 149, + 289, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 149, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 149, + 289, + 216 + ], + "type": "text", + "content": "Kawin Ethayarajh and Dan Jurafsky. 2020. Utility is in the eye of the user: A critique of NLP leaderboards. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4846-4853, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 227, + 289, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 227, + 289, + 305 + ], + "spans": [ + { + "bbox": [ + 69, + 227, + 289, + 305 + ], + "type": "text", + "content": "Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, and Alexis Conneau. 2021. Larger-scale transformers for multilingual masked language modeling. In Proceedings of the 6th Workshop on Representation Learning for NLP (RepL4NLP-2021), pages 29–33, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 316, + 289, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 316, + 289, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 316, + 289, + 382 + ], + "type": "text", + "content": "Masahiro Kaneko and Danushka Bollegala. 2021. Debiasing pre-trained contextualised embeddings. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1256-1266, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 393, + 289, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 393, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 393, + 289, + 449 + ], + "type": "text", + "content": "Masahiro Kaneko and Danushka Bollegala. 2022. Unmasking the mask-evaluating social biases in masked language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11954-11962." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 460, + 289, + 549 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 460, + 289, + 549 + ], + "spans": [ + { + "bbox": [ + 69, + 460, + 289, + 549 + ], + "type": "text", + "content": "Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, and Naoaki Okazaki. 2022. Gender bias in masked language models for multiple languages. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2740-2750, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 560, + 289, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 560, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 69, + 560, + 289, + 605 + ], + "type": "text", + "content": "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 616, + 289, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 289, + 650 + ], + "type": "text", + "content": "Svetlana Kiritchenko and Saif M Mohammad. 2018. Examining gender and race bias in two hundred sentiment analysis systems. *NAACL HLT* 2018, page 43." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 661, + 289, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 661, + 289, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 661, + 289, + 717 + ], + "type": "text", + "content": "Keita Kurita, Nidhi Vyas, Ayush Pareek, Alan W Black, and Yulia Tsvetkov. 2019. Measuring bias in contextualized word representations. In Proceedings of the First Workshop on Gender Bias in Natural Language Processing, pages 166-172." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018. Word translation without parallel data. In International Conference on Learning Representations." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 171 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 181, + 524, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 524, + 238 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 524, + 238 + ], + "type": "text", + "content": "Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, and Madian Khabsa. 2023. Xlm-v: Overcoming the vocabulary bottleneck in multilingual masked language models. arXiv preprint arXiv:2301.10472." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 524, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 524, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 524, + 324 + ], + "type": "text", + "content": "Sheng Liang, Philipp Dufter, and Hinrich Schütze. 2020. Monolingual and multilingual reduction of gender bias in contextualized representations. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5082-5093, Barcelona, Spain (Online). International Committee on Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 334, + 524, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 524, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 524, + 368 + ], + "type": "text", + "content": "Seungyoung Lim, Myungji Kim, and Jooyoul Lee. 2019. Korquad1.0: Korean qa dataset for machine reading comprehension. arXiv preprint arXiv:1909.07005." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 377, + 524, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 524, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 524, + 444 + ], + "type": "text", + "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 454, + 524, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 524, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 524, + 510 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 519, + 524, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 524, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 524, + 575 + ], + "type": "text", + "content": "Kelly Marchisio, Patrick Lewis, Yihong Chen, and Mikel Artetxe. 2022. Mini-model adaptation: Efficiently extending pretrained models to new languages via aligned shallow training. arXiv preprint arXiv:2212.10503." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 585, + 524, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 524, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 524, + 663 + ], + "type": "text", + "content": "Chandler May, Alex Wang, Shikha Bordia, Samuel Bowman, and Rachel Rudinger. 2019. On measuring social biases in sentence encoders. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 622-628." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 673, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 524, + 772 + ], + "type": "text", + "content": "Benjamin Muller, Antonios Anastasopoulos, Benoit Sagot, and Djamé Seddah. 2021. When being unseen from mBERT is just the beginning: Handling new languages with multilingual language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 448-462, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14735" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 161 + ], + "type": "text", + "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. StereoSet: Measuring stereotypical bias in pretrained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 5356-5371, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 168, + 291, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 168, + 291, + 246 + ], + "spans": [ + { + "bbox": [ + 69, + 168, + 291, + 246 + ], + "type": "text", + "content": "Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. 2020. CrowS-pairs: A challenge dataset for measuring social biases in masked language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1953-1967, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 252, + 290, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 252, + 290, + 297 + ], + "spans": [ + { + "bbox": [ + 69, + 252, + 290, + 297 + ], + "type": "text", + "content": "Malte Ostendorff and Georg Rehm. 2023. Efficient language model training through cross-lingual and progressive transfer learning. arXiv preprint arXiv:2301.09626." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 304, + 291, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 304, + 291, + 404 + ], + "spans": [ + { + "bbox": [ + 69, + 304, + 291, + 404 + ], + "type": "text", + "content": "Jonas Pfeiffer, Naman Goyal, Xi Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. 2022. Lifting the curse of multilinguality by pre-training modular transformers. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3479-3495, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 410, + 291, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 410, + 291, + 477 + ], + "spans": [ + { + "bbox": [ + 69, + 410, + 291, + 477 + ], + "type": "text", + "content": "Telmo Pires, Eva Schlinger, and Dan Garrette. 2019. How multilingual is multilingual BERT? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4996-5001, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 485, + 291, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 485, + 291, + 550 + ], + "spans": [ + { + "bbox": [ + 69, + 485, + 291, + 550 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 558, + 291, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 558, + 291, + 625 + ], + "spans": [ + { + "bbox": [ + 69, + 558, + 291, + 625 + ], + "type": "text", + "content": "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 632, + 291, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 632, + 291, + 698 + ], + "spans": [ + { + "bbox": [ + 69, + 632, + 291, + 698 + ], + "type": "text", + "content": "Sara Rosenthal, Noura Farra, and Preslav Nakov. 2017. SemEval-2017 task 4: Sentiment analysis in Twitter. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 502-518, Vancouver, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "text", + "content": "Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Galle, et al. 2022. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 126 + ], + "type": "text", + "content": "Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in NLP. Transactions of the Association for Computational Linguistics, 9:1408-1424." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 135, + 526, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 526, + 179 + ], + "type": "text", + "content": "ByungHoon So, Kyuhong Byun, Kyungwon Kang, and Seongjin Cho. 2022. Jaquad: Japanese question answering dataset for machine reading comprehension. arXiv preprint arXiv:2202.01764." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 186, + 526, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 526, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 526, + 264 + ], + "type": "text", + "content": "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2022. Generative language models for paragraph-level question generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 670-688, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 271, + 526, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 526, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 526, + 349 + ], + "type": "text", + "content": "Asahi Ushio, Fernando Alva-Manchego, and Jose Camacho-Collados. 2023. A practical toolkit for multilingual question and answer generation. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 356, + 526, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 526, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 526, + 433 + ], + "type": "text", + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 440, + 526, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 526, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 526, + 518 + ], + "type": "text", + "content": "Hai Wang, Dian Yu, Kai Sun, Jianshu Chen, and Dong Yu. 2019. Improving pre-trained multilingual model with vocabulary expansion. In Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL), pages 316-327, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 525, + 526, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 526, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 526, + 613 + ], + "type": "text", + "content": "Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 620, + 526, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 526, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 526, + 709 + ], + "type": "text", + "content": "Bo Zheng, Li Dong, Shaohan Huang, Saksham Singhal, Wanxiang Che, Ting Liu, Xia Song, and Furu Wei. 2021. Allocating large vocabulary capacity for crosslingual language model pre-training. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3203-3215, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 716, + 526, + 772 + ], + "type": "text", + "content": "Yi Zhou, Masahiro Kaneko, and Danushka Bollegala. 2022. Sense embeddings are also biased - evaluating social biases in static and contextualised sense embeddings. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14736" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 69, + 298, + 374 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 298, + 374 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 298, + 374 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 298, + 374 + ], + "type": "table", + "html": "
Vocab.No-Trim (250K)5K10K15K30K60K
Param.278M89M93M97M109M132M
Post-FT (Sentiment)AR66.364.564.565.965.9
DE73.270.472.472.173.7
EN68.464.066.567.468.6
ES69.066.267.267.868.4
FR71.868.671.471.771.6
IT62.960.861.963.562.3
PT70.763.665.968.269.4
Pre-FT (Sentiment)AR66.358.661.963.163.2
DE73.270.671.873.171.5
EN68.464.666.667.766.0
ES69.060.465.666.365.9
FR71.870.374.373.872.2
IT62.966.267.168.168.2
PT70.765.567.569.667.9
Post-FT (NLI)AR75.775.075.875.875.7
DE79.976.678.979.679.9
EN84.680.583.684.384.6
ES79.875.378.479.479.8
FR80.177.280.080.180.2
Pre-FT (NLI)AR75.773.075.075.475.7
DE79.977.479.079.079.6
EN84.683.384.784.485.1
ES79.878.979.179.281.0
FR80.177.378.978.280.1
", + "image_path": "e9718ab3ec99ac02fae862b9e5c571a095a84004a3d312e91a8e55e390e70788.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 383, + 290, + 432 + ], + "lines": [ + { + "bbox": [ + 67, + 383, + 290, + 432 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 290, + 432 + ], + "type": "text", + "content": "Table 4: Results of XLM-R for sentiment analysis (macro F1) and NLI (accuracy) with different top-" + }, + { + "bbox": [ + 67, + 383, + 290, + 432 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 383, + 290, + 432 + ], + "type": "text", + "content": " vocabulary size at VT, where the best results in each LM and language are in the bold characters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 78, + 454, + 290, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 454, + 290, + 477 + ], + "spans": [ + { + "bbox": [ + 78, + 454, + 290, + 477 + ], + "type": "text", + "content": "(Volume 1: Long Papers), pages 1924-1935, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 487, + 199, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 487, + 199, + 501 + ], + "spans": [ + { + "bbox": [ + 68, + 487, + 199, + 501 + ], + "type": "text", + "content": "A Top- " + }, + { + "bbox": [ + 68, + 487, + 199, + 501 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 68, + 487, + 199, + 501 + ], + "type": "text", + "content": " VT of XLM-R" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 509, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 509, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 509, + 290, + 550 + ], + "type": "text", + "content": "Table 4 shows the results of XLM-R fine-tuned on sentiment and NLI with post/pre-VT for different top-" + }, + { + "bbox": [ + 67, + 509, + 290, + 550 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 509, + 290, + 550 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 559, + 182, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 559, + 182, + 573 + ], + "spans": [ + { + "bbox": [ + 68, + 559, + 182, + 573 + ], + "type": "text", + "content": "B Top- " + }, + { + "bbox": [ + 68, + 559, + 182, + 573 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 68, + 559, + 182, + 573 + ], + "type": "text", + "content": " VT of mT5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 581, + 290, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 290, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 290, + 608 + ], + "type": "text", + "content": "Table 5 shows the results of mT5 fine-tuned on QA and QG with post/pre-VT for different top- " + }, + { + "bbox": [ + 67, + 581, + 290, + 608 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 581, + 290, + 608 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 618, + 255, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 618, + 255, + 644 + ], + "spans": [ + { + "bbox": [ + 68, + 618, + 255, + 644 + ], + "type": "text", + "content": "C Details of Results on Social Bias Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 653, + 291, + 734 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 653, + 291, + 734 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 291, + 734 + ], + "type": "text", + "content": "Table 6 shows the details of social bias evaluation (EEC dataset) regarding each emotion type observed in the LMs fine-tuned on sentiment analysis. Table 7 shows the details of social bias regarding each bias type in both CP and SS datasets observed in the comparison LMs." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14737" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 79, + 89, + 515, + 439 + ], + "blocks": [ + { + "bbox": [ + 79, + 89, + 515, + 439 + ], + "lines": [ + { + "bbox": [ + 79, + 89, + 515, + 439 + ], + "spans": [ + { + "bbox": [ + 79, + 89, + 515, + 439 + ], + "type": "table", + "html": "
Vocab. \nParam.No-Trim (250K) \n300M5K \n49M10K \n54M15K \n59M30K \n74M60K \n105M90K \n136M120K \n166M
EN70.1 / 55.572.4 / 59.773.2 / 60.573.7 / 60.972.7 / 58.871.0 / 56.470.5 / 55.870.3 / 55.6
ES55.9 / 34.753.5 / 32.554.4 / 33.154.5 / 33.255.2 / 33.856.0 / 34.855.9 / 34.855.9 / 34.7
FR50.0 / 30.950.0 / 31.548.9 / 29.948.6 / 29.349.9 / 30.149.8 / 30.550.2 / 30.950.0 / 30.8
IT53.2 / 37.654.3 / 39.653.8 / 38.454.3 / 38.854.2 / 38.553.7 / 38.153.4 / 37.8-
JA65.7 / 65.754.4 / 54.460.3 / 60.362.6 / 62.660.4 / 60.464.9 / 64.965.6 / 65.665.7 / 65.7
KO77.1 / 70.675.0 / 67.875.8 / 68.776.2 / 69.376.9 / 70.277.3 / 70.7--
RU73.7 / 51.470.1 / 48.871.4 / 49.570.3 / 47.670.3 / 47.973.8 / 51.173.9 / 51.473.8 / 51.4
EN23.8 / 90.023.6 / 89.824.1 / 89.924.2 / 90.024.2 / 90.023.9 / 90.023.9 / 90.023.9 / 90.0
ES22.7 / 84.121.4 / 83.622.0 / 83.822.2 / 83.922.4 / 84.022.7 / 84.122.7 / 84.122.7 / 84.1
FR17.5 / 80.717.1 / 80.317.1 / 80.417.3 / 80.517.3 / 80.517.4 / 80.617.5 / 80.717.5 / 80.7
IT17.6 / 80.817.3 / 80.717.5 / 80.817.5 / 80.817.6 / 80.717.6 / 80.817.6 / 80.8-
JA29.0 / 80.925.7 / 79.227.4 / 80.128.2 / 80.528.8 / 80.929.0 / 80.929.0 / 80.929.0 / 80.9
KO27.5 / 82.927.3 / 82.827.4 / 82.927.4 / 82.927.4 / 82.927.5 / 83.0--
RU26.4 / 84.326.0 / 84.026.2 / 84.226.3 / 84.226.3 / 84.226.3 / 84.326.4 / 84.326.4 / 84.3
EN70.1 / 55.574.0 / 61.371.8 / 59.173.4 / 60.774.3 / 61.371.3 / 56.568.6 / 54.366.2 / 52.1
ES55.9 / 34.756.7 / 36.758.1 / 37.562.2 / 41.459.8 / 40.158.0 / 37.058.9 / 38.152.0 / 32.7
FR50.0 / 30.949.0 / 32.747.6 / 30.243.3 / 26.944.9 / 27.434.3 / 20.043.1 / 24.140.0 / 23.7
IT53.2 / 37.657.6 / 43.761.5 / 46.861.2 / 45.657.8 / 42.156.5 / 40.555.4 / 39.2-
JA65.7 / 65.755.3 / 55.352.2 / 52.261.3 / 61.360.4 / 60.464.6 / 64.666.9 / 66.967.2 / 67.2
KO77.1 / 70.679.5 / 72.683.7 / 77.782.7 / 76.480.4 / 73.781.6 / 75.1--
RU73.7 / 51.473.5 / 51.776.9 / 56.477.4 / 56.975.9 / 54.275.1 / 53.075.0 / 53.073.3 / 51.4
EN23.8 / 90.024.8 / 90.024.4 / 89.924.3 / 89.924.2 / 90.024.2 / 90.023.3 / 89.923.6 / 89.8
ES22.7 / 84.121.9 / 84.121.9 / 83.922.7 / 83.722.0 / 84.322.6 / 84.420.4 / 79.622.6 / 84.1
FR17.5 / 80.716.1 / 79.016.9 / 79.916.9 / 79.515.8 / 78.814.6 / 78.015.7 / 79.017.0 / 79.2
IT17.6 / 80.817.4 / 80.417.8 / 81.118.0 / 80.817.1 / 80.617.4 / 80.917.4 / 80.8-
JA29.0 / 80.926.6 / 79.827.0 / 79.227.5 / 79.727.3 / 79.927.7 / 80.328.3 / 80.828.2 / 80.2
KO27.5 / 82.928.4 / 83.427.8 / 83.428.2 / 84.128.8 / 83.428.4 / 83.4--
RU26.4 / 84.327.7 / 85.930.0 / 86.629.0 / 86.529.2 / 86.329.0 / 86.629.2 / 86.728.9 / 86.0
", + "image_path": "babc967533548380adbf296da82690e0accd09809baca74250c8df0cd2eaedc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 519, + 525, + 715 + ], + "blocks": [ + { + "bbox": [ + 67, + 447, + 525, + 471 + ], + "lines": [ + { + "bbox": [ + 67, + 447, + 525, + 471 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 525, + 471 + ], + "type": "text", + "content": "Table 5: Results of mT5 for QA (Ans-F1/EM) and QG (MTR/BS) with different top- " + }, + { + "bbox": [ + 67, + 447, + 525, + 471 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 67, + 447, + 525, + 471 + ], + "type": "text", + "content": " vocabulary size at VT,where the best results in each LM and language are in the bold characters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 519, + 525, + 715 + ], + "lines": [ + { + "bbox": [ + 69, + 519, + 525, + 715 + ], + "spans": [ + { + "bbox": [ + 69, + 519, + 525, + 715 + ], + "type": "table", + "html": "
ModelAULA-EECAngerSadnessFearJoyNo Emotion Type
Gender BiasFT RoBERTa64.850.067.871.564.360.8
FT XLM-R44.346.939.451.434.439.7
Pre-FT XLM-R (EN)42.533.855.845.140.742.7
Post-FT XLM-R (EN)44.346.939.451.434.439.7
Pre-FT XLM-R (50k)44.346.939.451.434.439.7
Post-FT XLM-R (50k)41.041.245.847.236.937.2
Race BiasFT RoBERTa85.750.088.065.450.0100.0
FT XLM-R62.074.532.147.374.617.4
Pre-FT VT XLM-R (EN)56.933.758.869.260.150.0
Pre-FT VT XLM-R (EN)62.074.532.147.374.617.4
Pre-FT VT XLM-R (50k)62.074.532.147.374.617.4
Post-FT VT XLM-R (50k)62.455.770.370.959.159.0
", + "image_path": "70f4930854529e6fc638c0eb2c7ba51f1c07fbb96abbffcfdec04a93f551394b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 724, + 525, + 748 + ], + "lines": [ + { + "bbox": [ + 67, + 724, + 525, + 748 + ], + "spans": [ + { + "bbox": [ + 67, + 724, + 525, + 748 + ], + "type": "text", + "content": "Table 6: Social bias scores of LMs fine-tuned on sentiment analysis on each emotion type w/wo VT on the EEC dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14738" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 284, + 503, + 521 + ], + "blocks": [ + { + "bbox": [ + 91, + 284, + 503, + 521 + ], + "lines": [ + { + "bbox": [ + 91, + 284, + 503, + 521 + ], + "spans": [ + { + "bbox": [ + 91, + 284, + 503, + 521 + ], + "type": "table", + "html": "
RoBERTaXLM-RVT XLM-R (EN)VT XLM-R (50K)
CP DatasetAULA58.149.549.549.3
Age59.849.449.449.4
Disability68.366.766.768.3
Gender53.448.548.548.1
Nationality56.644.044.044.7
Physical-Appearance52.458.758.755.6
Race-Color56.843.843.844.0
Religion53.354.354.353.3
Sexual-Orientation67.954.854.854.8
Socioeconomic66.358.158.157.6
SS DatasetAULA58.854.954.955.0
Gender62.455.355.354.9
Profession60.454.354.354.6
Race57.056.056.056.0
Religion54.446.846.846.8
", + "image_path": "7814161b730d216654a6dd64a07329188f414f76ecb9dad36c928a61bbd9b922.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 528, + 524, + 555 + ], + "lines": [ + { + "bbox": [ + 67, + 528, + 524, + 555 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 524, + 555 + ], + "type": "text", + "content": "Table 7: Social bias scores of VT LMs compared to their original ones on both CP and SS datasets for each bias type." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14739" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_content_list.json b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..54559b08940e3a8309395c8b248c8252166d5335 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_content_list.json @@ -0,0 +1,2797 @@ +[ + { + "type": "text", + "text": "Efficient $k$ -NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition", + "text_level": 1, + "bbox": [ + 238, + 90, + 757, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nishant Yadav†, Nicholas Monath‡, Manzil Zaheer§, and Andrew McCallum†", + "bbox": [ + 161, + 153, + 836, + 170 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "† University of Massachusetts Amherst ‡Google Research", + "bbox": [ + 253, + 171, + 747, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{nishantyavad, mccallum}@cs.umass.edu {nmonath,manzilzaheer}@google.com", + "bbox": [ + 134, + 189, + 867, + 204 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cross-encoder models, which jointly encode and score a query-item pair, are prohibitively expensive for direct $k$ -nearest neighbor ( $k$ -NN) search. Consequently, $k$ -NN search typically employs a fast approximate retrieval (e.g. using BM25 or dual-encoder vectors), followed by reranking with a cross-encoder; however, the retrieval approximation often has detrimental recall regret. This problem is tackled by ANNCUR (Yadav et al., 2022), a recent work that employs a cross-encoder only, making search efficient using a relatively small number of anchor items, and a CUR matrix factorization. While ANNCUR's one-time selection of anchors tends to approximate the cross-encoder distances on average, doing so forfeits the capacity to accurately estimate distances to items near the query, leading to regret in the crucial end-task: recall of top- $k$ items. In this paper, we propose ADACUR, a method that adaptively, iteratively, and efficiently minimizes the approximation error for the practically important top- $k$ neighbors. It does so by iteratively performing $k$ -NN search using the anchors available so far, then adding these retrieved nearest neighbors to the anchor set for the next round. Empirically, on multiple datasets, in comparison to previous traditional and state-of-the-art methods such as ANNCUR and dual-encoder-based retrieval-and-erank, our proposed approach ADACUR consistently reduces recall error—by up to $70\\%$ on the important $k = 1$ setting—while using no more compute than its competitors.", + "bbox": [ + 144, + 279, + 460, + 740 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 751, + 258, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$k$ -nearest neighbor ( $k$ -NN) search is a core subroutine of a variety of tasks in NLP such as entity linking (Wu et al., 2020), passage retrieval for QA (Karpukhin et al., 2020), and more generally, in retrieval-augmented machine learning models (Guu et al., 2020; Izacard et al., 2023). For many of these applications, the state-of-the-art similarity function is a cross-encoder that directly outputs a scalar similarity score after jointly encoding a", + "bbox": [ + 112, + 776, + 489, + 919 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7841db05b54c8187f64cb341ff50b33be4113076ed8294d41719b564dc366f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 250, + 852, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd6f11d141d922b6e9cae84b134ed0f204a697693f8ce9a7c1b7873bf21364b7.jpg", + "image_caption": [ + "(a) Uniformly at random" + ], + "image_footnote": [], + "bbox": [ + 512, + 290, + 687, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d01a79b72270176b45e252c19a0483397eee9525be8a52a8670ce78aeb31fc8c.jpg", + "image_caption": [ + "(b) Biased sampling", + "Figure 1: Exact versus approximate cross-encoder scores (computed using ANNCUR) of all items for a test-query in domain=YuGi0h. ANNCUR incurs high approximation error on $k$ -NN items wrt exact scores when using 50 anchor items sampled uniformly at random (Fig. 1a). In contrast, sampling 50 anchor items with probability proportional to exact cross-encoder scores (Fig. 1b) significantly improves approximation of top-scoring items." + ], + "image_footnote": [], + "bbox": [ + 702, + 290, + 880, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "given query-item pair. However, computing a single query-item score using a cross-encoder requires a forward pass of the model which can be computationally expensive as cross-encoders are typically parameterized using deep neural models such as transformers (Vaswani et al., 2017). For this reason, $k$ -NN search with a cross-encoder typically involves retrieving candidate items using additional models such as dual-encoders or BM25 (Robertson et al., 2009), followed by re-ranking items using the cross-encoder (Logeswaran et al., 2019; Zhang and Stratos, 2021; Qu et al., 2021). However, the accuracy of such retrieve-and-erank approaches is upperbound by the recall of first-stage retrieval and may require computationally expensive distillation-based training of dual-encoders to improve recall.", + "bbox": [ + 507, + 552, + 882, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent work by Yadav et al. (2022) proposed ANNCUR, a CUR factorization (Mahoney and Drineas, 2009) based method, that approximates cross-encoder using dot-product of latent query and item embeddings and performs $k$ -NN retrieval using approximate scores followed by optionally re-ranking retrieved items using exact cross", + "bbox": [ + 507, + 808, + 884, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8088", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8088-8103", + "bbox": [ + 216, + 945, + 779, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 277, + 958, + 719, + 971 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "encoder scores. The latent item embeddings are computed by comparing each item against a set of anchor queries, and the latent query embedding is computed using the query's cross-encoder scores against a fixed set of anchor items. As shown in Figure 1, when ANNCUR selects the anchor items uniformly at random (Fig 1a), it incurs higher approximation error on the top-scoring items than the rest of the items, resulting in poor $k$ -NN recall, and including some $k$ -NN items as part of anchor items (Fig. 1b) can significantly improve approximation error for top-scoring items.", + "bbox": [ + 112, + 84, + 489, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose ADACUR, a search strategy that improves $k$ -NN search recall by improving the approximation of top-scoring items. ADACUR performs retrieval over multiple rounds, retrieving the first batch of items either uniformly at random or using heuristic or auxiliary models such as dual-encoder or BM25 to get a first coarse approximation of item scores for the test query. In subsequent rounds, it alternates between a) performing retrieval using approximate scores and scoring retrieved items using cross-encoder and b) using all retrieved items as anchor items to improve the approximation and hence retrieval of relevant items in the subsequent rounds. Our proposed approach provides significant improvements in $k$ -NN search recall over ANNCUR and dual-encoder based retrieve-and-erank approaches when performing $k$ -NN search with cross-encoder models trained for the task of entity linking and information retrieval.", + "bbox": [ + 115, + 275, + 489, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Proposed Method: ADACUR", + "text_level": 1, + "bbox": [ + 112, + 599, + 394, + 615 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Given a scoring function $f_{\\theta} \\colon \\mathcal{Q} \\times \\mathcal{I} \\to \\mathbb{R}$ that maps a query-item pair to a scalar score, and a query $q \\in \\mathcal{Q}$ , the $k$ -nearest neighbor search task is to retrieve top- $k$ scoring items from a fixed item set $\\mathcal{I}$ according to the given scoring function $f_{\\theta}$ .", + "bbox": [ + 112, + 623, + 487, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 ADACUR: Offline Indexing of Items", + "text_level": 1, + "bbox": [ + 112, + 712, + 443, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The indexing step of ADACUR involves using the cross-encoder model $(f_{\\theta})$ to score each item against a fixed set of $k_q$ anchor/train queries $(Q_{train})$ , to get score matrix $R_{\\mathrm{anc}}$", + "bbox": [ + 112, + 732, + 485, + 795 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {a n c}} (q, i) = f _ {\\theta} (q, i), \\forall (q, i) \\in \\mathcal {Q} _ {\\text {t r a i n}} \\times \\mathcal {I}\n$$\n", + "text_format": "latex", + "bbox": [ + 144, + 801, + 455, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Each column of $E^{\\mathcal{I}} \\coloneqq R_{\\mathrm{anc}} \\in \\mathbb{R}^{k_q \\times |\\mathcal{I}|}$ corresponds to a $k_q$ -dimensional latent item embedding.", + "bbox": [ + 112, + 824, + 487, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 ADACUR: Test-time inference", + "text_level": 1, + "bbox": [ + 112, + 866, + 396, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The baseline method ANNCUR computes the latent test-query embedding $e_{q_{\\mathrm{test}}}$ using $C_{q_{\\mathrm{test}}}$ , a", + "bbox": [ + 112, + 887, + 487, + 922 + ], + "page_idx": 1 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 ADACUR $k$ -NN Search" + ], + "code_body": "1: Input: $(q_{\\mathrm{test}}, R_{\\mathrm{anc}}, N_{\\mathcal{R}}, \\mathcal{B}_{\\mathrm{CE}}, \\mathcal{A})$ \n2: $q_{\\mathrm{test}}$ : Test query \n3: $R_{\\mathrm{anc}}$ : Matrix containing CE scores between $\\mathcal{Q}_{\\mathrm{train}}$ and $\\mathcal{I}$ \n4: $\\mathcal{B}_{\\mathrm{CE}}$ : Total cross-encoder (CE) call budget. \n5: $\\mathcal{A}$ : Algorithm to use for selecting (anchor) items. \n6: $N_{\\mathcal{R}}$ : Number of iterative search rounds \n7: Output: $\\hat{S}$ : Approximate scores of $q_{\\mathrm{test}}$ with all items, $\\mathcal{I}_{\\mathrm{anc}}$ : Retrieved (anchor) items with CE scores in $C_{\\mathrm{test}}$ . \n8: $k_s \\gets \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}} \\triangleright$ Num. of items to sample per round \n9: $\\mathcal{I}_{\\mathrm{anc}} \\gets \\text{INIT}(\\mathcal{I}, k_s)$ $\\triangleright$ Initial set of anchor items \n10: $C_{\\mathrm{test}} \\gets [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}} \\triangleright$ CE scores of $\\mathcal{I}_{\\mathrm{anc}}$ for $q_{\\mathrm{test}}$ \n11: for $j \\gets 2$ to $N_{\\mathcal{R}}$ do \n12: $U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}$ $\\triangleright$ $U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}$ \n13: $\\hat{S}^{(j)} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright$ Update approx. scores \n14: $\\mathcal{I}_{\\mathrm{anc}}^{(j)} \\gets \\text{SAMPLEITEMS}(\\mathcal{A}, k_s, \\mathcal{I}_{\\mathrm{anc}}, \\hat{S}^{(j)})$ \n15: $\\mathcal{I}_{\\mathrm{anc}} \\gets \\mathcal{I}_{\\mathrm{anc}} \\cup \\mathcal{I}_{\\mathrm{anc}}^{(j)}$ \n16: $C_{\\mathrm{test}} \\gets C_{\\mathrm{test}} \\oplus [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}^{(j)}} \\triangleright$ Update $C_{\\mathrm{test}}$ \n17: $U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}$ $\\triangleright$ $U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}$ \n18: $\\hat{S} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright$ Compute approx. scores \n19: return $\\hat{S}, \\mathcal{I}_{\\mathrm{anc}}, C_{\\mathrm{test}}$", + "bbox": [ + 519, + 103, + 882, + 370 + ], + "page_idx": 1 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 SAMPLEITEMS" + ], + "code_body": "1: Input: $(\\mathcal{A}, k_s, \\mathcal{I}_{\\text{mask}}, S)$ \n2: $\\mathcal{A}$ : Algorithm for sampling items \n3: $k_s$ : Number of items to sample \n4: $\\mathcal{I}_{\\text{mask}}$ : Set of items to mask \n5: $S$ : (Approximate) Scores for all items \n6: Output: $\\mathcal{I}_{\\text{select}}$ : Set of sampled $k_s$ items \n7: $\\bar{S} \\gets \\text{SOFTMAX}(S)$ \n8: $\\bar{S}[\\mathcal{I}_{\\text{mask}}] \\gets 0$ > Mask items in $\\mathcal{I}_{\\text{mask}}$ \n9: if $\\mathcal{A} = \\text{TopK}$ then \n10: $\\mathcal{I}_{\\text{select}} \\gets \\text{TOPK}(\\bar{S}, k_s)$ \n11: else if $\\mathcal{A} = \\text{SoftMax}$ then \n12: $\\mathcal{I}_{\\text{select}} \\gets k_s$ items sampled using $\\bar{S}$ \n13: else if $\\mathcal{A} = \\text{Random}$ then \n14: $\\mathcal{I}_{\\text{select}} \\gets k_s$ items uniformly sampled from $\\mathcal{I} \\setminus \\mathcal{I}_{\\text{mask}}$ \n15: return $\\mathcal{I}_{\\text{select}}$", + "bbox": [ + 515, + 404, + 880, + 596 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$|\\mathcal{I}_{\\mathrm{anc}}|$ -dimensional vector containing cross-encoder scores of $q_{\\mathrm{test}}$ with a set of anchor items $(\\mathcal{I}_{\\mathrm{anc}})$ as:", + "bbox": [ + 507, + 624, + 882, + 657 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nC _ {q _ {\\text {t e s t}}} = \\left[ f _ {\\theta} \\left(q _ {\\text {t e s t}}, i\\right) \\right] _ {i \\in \\mathcal {I} _ {\\text {a n c}}}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 669, + 800, + 688 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ne _ {q _ {\\text {t e s t}}} = C _ {q _ {\\text {t e s t}}} \\times U\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 692, + 752, + 711 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $U\\in \\mathbb{R}^{|\\mathcal{I}_{\\mathrm{anc}}|\\times |\\mathcal{Q}_{\\mathrm{train}}|}$ is the pseudo-inverse of $R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]$ , the subset of columns of $R_{\\mathrm{anc}}$ corresponding to (anchor) items $\\mathcal{I}_{\\mathrm{anc}}$ . Finally, ANCUR approximates the score for a query-item pair $(q_{\\mathrm{test}},i)$ using dot-product of the query embedding $e_{q_{\\mathrm{test}}}$ and item embedding $E^{\\mathcal{I}}[:,i]$ as", + "bbox": [ + 507, + 722, + 882, + 822 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {\\theta} (q _ {\\mathrm {t e s t}}, i) = e _ {q _ {\\mathrm {t e s t}}} ^ {\\top} E ^ {\\mathcal {I}} [:, i ]\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 828, + 800, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main bottleneck at test-time inference is the number of items scored using the cross-encoder for the given test-query. For a given budget of cross-encoder calls, ANNCUR splits the budget $(\\mathcal{B}_{\\mathrm{CE}})$", + "bbox": [ + 507, + 854, + 882, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8089", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into two parts - it uses $k_{i}$ cross-encoder calls to compare against anchor items (chosen uniformly at random or using heuristic methods) and retrieves $\\mathcal{B}_{\\mathrm{CE}} - k_{i}$ items using approximate scores and re-ranks them using exact cross-encoder scores.", + "bbox": [ + 112, + 84, + 487, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, our proposed approach ADACUR uses the cross-encoder call budget to adaptively retrieve and score items over $N_{\\mathcal{R}}$ rounds, repurposing the retrieved items as anchor items as shown in Algorithm 1. ADACUR begins by sampling the first batch of $k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}$ (anchor) items uniformly at random. The first batch of items can also be selected using baseline retrieval methods such as BM25 and dual-encoders. In the $j^{\\mathrm{th}}$ round, the items retrieved up to round $j - 1$ are used as anchor items to revise the test-query embedding, which in turn is used to update the approximate scores (line 13 in Algorithm 1). Finally, the items selected so far are masked out and the next batch of $k_{s}$ items in round $j$ is retrieved using the updated approximate scores in the following two ways:", + "bbox": [ + 115, + 164, + 489, + 414 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TopK: Greedily pick top- $k_{s}$ items according to approximate scores.", + "- SoftMax: Convert approximate item scores into probability using softmax and sample $k_{s}$ items without replacement." + ], + "bbox": [ + 134, + 416, + 485, + 499 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, ADACUR returns top- $k$ items based on exact cross-encoder scores $^{1}$ from the set of retrieved (anchor) items as approximate $k$ -NN items. We refer interested readers to Appendix B.5 for a discussion on the time complexity of ADACUR.", + "bbox": [ + 112, + 500, + 489, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Experiments", + "text_level": 1, + "bbox": [ + 112, + 590, + 260, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we evaluate the proposed approach and baselines on the task of finding $k$ -nearest neighbor items as per a given cross-encoder. We experiment with two cross-encoders - one trained for the task of zero-shot entity linking, and another trained on information retrieval datasets.", + "bbox": [ + 112, + 615, + 489, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental Setup We run evaluation on domains YuGiOh, StarTrek, and Military from ZESHEL-a zero-shot entity linking dataset (Logeswaran et al., 2019), and domains SciDocs and HotpotQA from BEIR-a zero-shot information retrieval benchmark (Thakur et al., 2021). We use two cross-encoder models trained on labeled training data from the corresponding benchmark and evaluate separately on each domain on the task of find", + "bbox": [ + 112, + 719, + 489, + 860 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing $k$ -NN cross-encoder items. For each ZESHEL domain, we randomly split the query set into a train/anchor set $(Q_{train})$ and a test set $(Q_{test})$ while for BEIR domains, we use pseudo-queries released as part of the benchmark as train/anchor queries and evaluate on queries in the official test split. We refer the reader to Table 1 for additional details.", + "bbox": [ + 507, + 84, + 882, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baselines We compare our proposed approach with the following baseline retrieval methods.", + "bbox": [ + 507, + 212, + 880, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual-Encoders (DE): Query-item scores are computed using dot-product of embeddings produced by a learned deep encoder model. DE is used for initial retrieval followed by re-ranking using the cross-encoder. We report results for $\\mathrm{DE}_{\\mathrm{BASE}}$ , a dual-encoder trained on training domains in the corresponding dataset, and the following two dual-encoder models trained on the target domain via distillation using the cross-encoder.", + "bbox": [ + 507, + 248, + 882, + 388 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}$ : DE initialized with BERT (Devlin et al., 2019) and trained only on the target domain via distillation using the cross-encoder.", + "- $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ : $\\mathrm{DE}_{\\mathrm{BASE}}$ model further fine-tuned on the target domain via distillation." + ], + "bbox": [ + 531, + 393, + 882, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ANNCUR: $k$ -NN search method proposed by Yadav et al. (2022) where anchor items are chosen uniformly at random. We additionally experiment with ANNCUR $_{\\text{DE}_{\\text{BASE}}}$ which uses top-scoring items retrieved using DE $_{\\text{BASE}}$ as anchor items.", + "bbox": [ + 507, + 486, + 882, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation Metric Following the precedent set by previous work (Yadav et al., 2022), we evaluate all approaches using Top- $k$ -Recall@ $\\mathcal{B}_{\\mathrm{CE}}$ which is defined as the fraction of $k$ -NN items retrieved under test-time cost budget $\\mathcal{B}_{\\mathrm{CE}}$ where the cost is defined as the number of cross-encoder calls made during inference. DE baselines will use the entire budget of $\\mathcal{B}_{\\mathrm{CE}}$ calls for re-ranking retrieved items using exact cross-encoder scores, ANNCUR splits the budget between scoring anchor $k_{i}$ items and using exact cross-encoder scores for re-ranking $\\mathcal{B}_{\\mathrm{CE}} - k_{i}$ retrieved items, and ADACUR use the budget to adaptively search for $k$ -NN items.", + "bbox": [ + 507, + 583, + 882, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For ADACUR, unless noted otherwise, we use $N_{\\mathcal{R}} = 5$ with TopK method for retrieving items using approximate scores, and retrieve the first batch of items uniformly at random (ADACUR) or using $\\mathrm{DE}_{\\mathrm{BASE}}$ (ADACUR $\\mathrm{DE}_{\\mathrm{BASE}}$ ). We refer readers to Appendix B for implementation details and details on training and parameterization of cross-encoder and dual-encoder models used in our experiments.", + "bbox": [ + 507, + 793, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "${}^{1}$ Sorting retrieved items based on exact cross-encoder scores does not require any additional cross-encoder calls as cross-encoder scores for these items have already been computed (see line 16 in Algorithm 1).", + "bbox": [ + 112, + 869, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8090", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/468480c5d18e58a6266e7f86040cda08c8fe2863a85b408fd2c723b680c3d722.jpg", + "image_caption": [ + "Figure 2: Top- $k$ -Recall for ADACUR and baselines for domain=YuGi0h, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ . ADACUR consistently outperforms the corresponding ANNCUR variant and $\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}$ outperforms all DE-based retrieve-and-erank approaches including $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ , a DE model trained via distillation on the target domain using the cross-encoder." + ], + "image_footnote": [], + "bbox": [ + 117, + 84, + 843, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Results", + "text_level": 1, + "bbox": [ + 112, + 316, + 218, + 331 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ADACUR versus baselines Figure 2 shows Top- $k$ -Recall for ADACUR and baselines on domain $\\equiv$ YuGiOh. ADACUR, which uses adaptively retrieved items as anchor items over $N_{\\mathcal{R}} = 5$ rounds, consistently outperforms ANNCUR which selects all anchor items uniformly at random. ADACUR also outperforms strong DE baseline $\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}$ for all values of $k$ and outperforms $\\mathrm{DE}_{\\mathrm{BASE}}$ & $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ for large values of $k = 10,100$", + "bbox": [ + 112, + 341, + 489, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling anchor items using $\\mathbf{DE}_{\\mathrm{BASE}}$ For $k = 1, 10$ , Top- $k$ -Recall for both ANNCUR and ADACUR can be further improved by leveraging baseline retrieval models such as $\\mathbf{DE}_{\\mathrm{BASE}}$ for retrieving all and the first batch of anchor items respectively instead of sampling them uniformly at random. ADACUR $_{\\mathsf{DE}_{\\mathrm{BASE}}}$ always performs better than ANNCUR $_{\\mathsf{DE}_{\\mathrm{BASE}}}$ which in turn performs better than re-ranking items retrieved using $\\mathbf{DE}_{\\mathrm{BASE}}$ . Thus, for a given cost budget $(\\mathcal{B}_{\\mathrm{CE}})$ , even when a strong baseline retrieval model such as $\\mathbf{DE}_{\\mathrm{BASE}}$ is available, using the baseline retrieval model to select the first batch of $\\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}$ items followed by using the proposed approach to adaptively retrieve more items over remaining $N_{\\mathcal{R}} - 1$ rounds performs better than merely re-ranking $\\mathcal{B}_{\\mathrm{CE}}$ top-scoring items retrieved using $\\mathbf{DE}_{\\mathrm{BASE}}$ . Finally, note that ADACUR $_{\\mathsf{DE}_{\\mathrm{BASE}}}$ outperforms all baselines including $\\mathbf{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ which requires additional compute- and resource-intensive distillation-based fine-tuning of $\\mathbf{DE}_{\\mathrm{BASE}}$ on the target domain.", + "bbox": [ + 112, + 504, + 489, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We refer the reader to Appendix C for results on other domains, training data size $(|\\mathcal{Q}_{\\mathrm{train}}|)$ values, oracle-based sampling experiments (§C.2), and additional results for multi-vector encoder (§C.3) and TF-IDF (§C.4) baselines.", + "bbox": [ + 112, + 840, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fcec76acae6b794158eba27856dfc360ee84cbac3b85b68afc4194cd0a01c3ef.jpg", + "image_caption": [ + "Figure 3: Top- $k$ -Recall for ADACUR for different number of rounds for domain=YuGiOh, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$" + ], + "image_footnote": [], + "bbox": [ + 524, + 316, + 867, + 429 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/08f54e341fd0cf65dec2446d644726baaca859380610577c62cb2d64470196dc.jpg", + "image_caption": [ + "Figure 4: ADACUR inference latency versus number of rounds for domain $= \\text{YuGi0h}$ , $|\\mathcal{Q}_{\\text{train}}| = 500$ ." + ], + "image_footnote": [], + "bbox": [ + 512, + 486, + 885, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Effect of number of rounds Figure 3 shows Top- $k$ -Recall and Figure 4 shows total inference latency of ADACUR on primary y-axis for various values of the test-time cross-encoder call budget $(\\mathcal{B}_{\\mathrm{CE}})$ . The secondary y-axis in Figure 4 shows the fraction of total inference time spent on each one of the three main steps in Algorithm 1 - (a) computing $U$ as pseudo-inverse of $R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]$ (line 12), (b) updating approximate scores for all items (line 13), and (c) computing exact cross-encoder scores for retrieved items (line 16). Since ADACUR selects items uniformly at random in the first round, ADACUR with $N_{\\mathcal{R}} = 1$ performs poorly as it simply returns a subset of items sampled uniformly at random. As expected, Top- $k$ -Recall for ADACUR increases with the number of rounds and saturates", + "bbox": [ + 507, + 665, + 884, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8091", + "bbox": [ + 480, + 927, + 517, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "at around 5-10 rounds while incurring negligible overhead in inference latency. As shown in Figure 4, cross-encoder score computation is the main bottleneck in test-time inference, taking $\\sim 7$ ms per score2. Increasing $N_{\\mathcal{R}}$ to large values such as 100 can incur up to $25\\%$ overhead with step (a) contributing significantly to this overhead. Although the time taken by matrix multiplication in step (b) is linear in the number of items in the domain, we observe that it is a negligible fraction of overall latency on GPUs even for domain=HotpotQA with 5 million items (see Figure 5) as GPUs can enable significant speedup even for brute-force computation of this step.", + "bbox": [ + 112, + 84, + 492, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Conclusion", + "text_level": 1, + "bbox": [ + 112, + 319, + 247, + 336 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we presented an adaptive search strategy that incrementally builds a query embedding to approximate cross-encoder scores and performs $k$ -NN search using approximate scores over several rounds. Our approach is designed to reduce approximation error for the top-scoring items and hence improves $k$ -NN search recall when retrieving items based on the approximate scores. We perform an in-depth empirical analysis of the proposed approach in terms of both retrieval quality and efficiency.", + "bbox": [ + 112, + 348, + 489, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 521, + 220, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building the index for the ADACUR is more expensive than the traditional dual-encoder index due to the computation of dense cross-encoder scores matrix (see §2.1). We have successfully run our approach on up to 5 million items, but scaling to billions of items is an interesting direction for future work. Dual-encoder-based retrieve-and-erank baseline approaches can benefit from training the dual-encoder on multiple domains. It is not clear if data from multiple domains can be leveraged to improve performance of the proposed approach on a given target domain; although in any case, cross-encoders tend to be more robust to domain shift than using only dual-encoders for retrieval.", + "bbox": [ + 112, + 549, + 489, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 785, + 265, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper we tackle the task of finding $k$ -nearest neighbor items for a given query when query-items scores are computed using a black-box similarity function such as a cross-encoder model. The cross-encoder scoring function may have certain biases", + "bbox": [ + 112, + 812, + 489, + 892 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and error tendencies, and it is unclear if our proposed method to approximate cross-encoder scores exacerbates or mitigates such biases.", + "bbox": [ + 507, + 84, + 884, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 509, + 143, + 672, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thank members of UMass IESL for helpful discussions and feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the National Science Foundation under Grant No. NSF1763618, in part by the Chan Zuckerberg Initiative under the project \"Scientific Knowledge Base Construction\", in part by International Business Machines Corporation Cognitive Horizons Network agreement number W1668553, in part by Amazon Digital Services, and in part using high-performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative. Any opinions, findings, conclusions, and recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor(s).", + "bbox": [ + 507, + 168, + 885, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 481, + 608, + 495 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. MS-MARCO: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Realm: retrieval-augmented language model pre-training. In International Conference on Machine Learning.", + "Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2020. Improving efficient neural ranking models with cross-architecture knowledge distillation. *ArXiv*, abs/2010.02666.", + "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24." + ], + "bbox": [ + 509, + 505, + 885, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "For a 12-layer transformer model on an Nvidia 2080ti", + "bbox": [ + 134, + 903, + 470, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8092", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient and effective passage search via contextualized late interaction over bert. In ACM SIGIR Conference on Research and Development in Information Retrieval.", + "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations, ICLR.", + "Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.", + "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations, ICLR.", + "Xinyin Ma, Yong Jiang, Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei Huang, and Weiming Lu. 2021. MuVER: Improving first-stage entity retrieval with multi-view entity representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.", + "Michael W Mahoney and Petros Drineas. 2009. Cur matrix decompositions for improved data analysis. Proceedings of the National Academy of Sciences, 106:697-702.", + "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.", + "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval, 3(4).", + "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.", + "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems.", + "Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems.", + "Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).", + "Nishant Yadav, Nicholas Monath, Rico Angell, Manzil Zaheer, and Andrew McCallum. 2022. Efficient nearest neighbor search for cross-encoder models using matrix factorization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing.", + "Wenzheng Zhang and Karl Stratos. 2021. Understanding hard negatives in noise contrastive estimation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + ], + "bbox": [ + 510, + 85, + 884, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Appendix A", + "text_level": 1, + "bbox": [ + 510, + 475, + 650, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B Training and Implementation Details", + "text_level": 1, + "bbox": [ + 510, + 499, + 867, + 514 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the code for reproducing experiments is available at https://github.com/iesl/anncur.", + "bbox": [ + 510, + 524, + 882, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B.1 Training Cross-Encoder Models", + "text_level": 1, + "bbox": [ + 510, + 565, + 808, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, we use [EMB]-CE, a cross-encoder model variant proposed by Yadav et al. (2022) that jointly encodes a query-item pair and computes the final score using dot-product of contextualized query and item embeddings extracted after joint encoding.", + "bbox": [ + 510, + 586, + 882, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "B.1.1 ZESHEL Dataset", + "text_level": 1, + "bbox": [ + 510, + 689, + 710, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ZESHEL dataset is a zero-shot entity linking containing a set of 16 domains, each containing a disjoint set of items (entities). Each domain contains a set of queries (mention) paired with their ground-truth items (entities). For ZESHEL, we use the cross-encoder model checkpoint released by Yadav et al. (2022). The cross-encoder model was trained by first training a dual-encoder model on ZESHEL training data using hard negatives, and then training a cross-encoder model for the task of zero-shot entity-linking on all eight training domains using cross-entropy loss with ground-truth", + "bbox": [ + 510, + 708, + 882, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://huggingface.co/nishantyadav/emb_crossenc_zeshel", + "bbox": [ + 529, + 904, + 885, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8093", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d5184a9ad74ea34768e26193fcb2a40ce5f567429656c9ce4848800852aba096.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDomain|I|(|Qtrain|, |Qtest|) SplitsTrain Query (Qtrain) Type
ZESHELYuGiOh10,031(100/3274), (500/2874), (2000/1374)Real Queries
ZESHELStarTrek34,430(100/4127), (500/3727), (2000/2227)Real Queries
ZESHELMilitary104,520(100/2300), (500/1900), (2000/0400)Real Queries
BEIRSciDocs25,657(1000/1000)Pseudo-Queries
BEIRHotpotQA5,233,329(1000/1000)Pseudo-Queries
", + "bbox": [ + 168, + 80, + 828, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Statistics on the number of items $(\\mathcal{I})$ and the number of queries in train and test splits for each domain. The train-query $(Q_{train})$ split refers to queries used for distilling dual-encoder models or for indexing items using ADACUR and ANNCUR. For ZESHEL domains, we create train-test splits by splitting the queries in each domain uniformly at random and test with three different splits by putting 100, 500, or 2000 queries in train split. For BEIR domains, we use pseudo-queries released as part of the benchmark as train queries $(Q_{train})$ and run $k$ -NN evaluation on test queries from the official test split (as per BEIR benchmark) of these domains. For HotpotQA, we use the first 1K queries out of a total of 7K test queries and we use all 1K test queries for SciDocs.", + "bbox": [ + 110, + 185, + 884, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "entity and negative entities mined using the dual-encoder. We refer readers to Yadav et al. (2022) for further details on cross-encoder training.", + "bbox": [ + 112, + 309, + 487, + 355 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform $k$ -NN experiments on domains YuGiOh, StarTrek, and Military from ZESHEL of which only Military was part of the training data used to train the cross-encoder model and YuGiOh and StarTrek are part of the original ZESHEL test domains and the cross-encoder model was not trained on these domains.", + "bbox": [ + 112, + 356, + 487, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.1.2 BEIR", + "text_level": 1, + "bbox": [ + 112, + 476, + 223, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow the training setup used by Hofstätter et al. (2020). We first train three teacher cross-encoders initialized with albert-large-v2 (Lan et al., 2020), bert-large-whole-word-masking, and bert-base-uncased (Devlin et al., 2019), and compute soft labels on 40 million (query, positive item, negative item) triplets in MS-MARCO dataset (Bajaj et al., 2016). We then train our cross-encoder model parameterized using a 6-layer MINI-LM model (Wang et al., 2020) via distillation using average scores of the three teacher models as target signal and minimizing mean-square-error between predicted and target scores. We use training scripts available as part of sentence-transformer $^{4}$ repository to train the cross-encoder model and use a dot-product based scoring mechanism for cross-encoders proposed by Yadav et al. (2022).", + "bbox": [ + 112, + 495, + 489, + 778 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.2 Training Dual-Encoder Models", + "text_level": 1, + "bbox": [ + 112, + 791, + 408, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.2.1 ZESHEL dataset", + "text_level": 1, + "bbox": [ + 112, + 812, + 312, + 826 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report results for DE baselines as reported in Yadav et al. (2022). The DE models were initialized using bert-base-uncased and contain separate query and item encoders, thus containing a total of", + "bbox": [ + 112, + 831, + 487, + 894 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$2 \\times 110M$ parameters. We refer readers to Yadav et al. (2022) for details related to the training of all DE model variants on ZESHEL dataset.", + "bbox": [ + 507, + 309, + 882, + 355 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.2.2 BEIR benchmark", + "text_level": 1, + "bbox": [ + 507, + 367, + 714, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For BEIR domains, we use a dual-encoder model released as part of sentence-transformer repository as $\\mathrm{DE}_{\\mathrm{BASE}}$ . This dual-encoder model was initialized using distillbert-base (Sanh et al., 2019) and trained on MS-MARCO dataset. This $\\mathrm{DE}_{\\mathrm{BASE}}$ is not trained on target domains SciDocs and HotpotQA used for running $k$ -NN experiments.", + "bbox": [ + 507, + 386, + 882, + 497 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We finetune $\\mathrm{DE}_{\\mathrm{BASE}}$ via distillation on the target domain to get $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ model. Given a set of training queries $\\mathcal{Q}_{\\mathrm{train}}$ from the target domain, we retrieve top-100 or top-1000 items for each query, score the items with the cross-encoder model, and train the dual-encoder by minimizing cross-entropy loss between predicted query-item scores (using DE) and target query-item scores (obtained using cross-encoder). Training a $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ with 1K queries and 100 or 1000 items per query takes around 2 hrs and 10 hrs respectively on an Nvidia RTX8000 GPU with 48GB memory. We train $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ for 10 epochs when using top-100 items per query and for 4 epochs when using top-1000 items per query using AdamW (Loshchilov and Hutter, 2019) optimizer with learning rate 1e-5.", + "bbox": [ + 507, + 499, + 882, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B.3 ANNCUR Implementation details", + "text_level": 1, + "bbox": [ + 507, + 766, + 821, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ANNCUR, we report results for the optimal split of cross-encoder call budget $(\\mathcal{B}_{\\mathrm{CE}})$ between scoring $k_{i}$ anchor items followed by retrieving $\\mathcal{B}_{\\mathrm{CE}} - k_{i}$ items for re-ranking. We experiment with $k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}$ . If the retrieved items contain a subset of anchor items for which exact", + "bbox": [ + 507, + 789, + 880, + 883 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "msmarco-distilroberta-base-v2:", + "bbox": [ + 534, + 891, + 764, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "www.sbert.net/docs/pretrained-models/msmarco-v2.html", + "bbox": [ + 510, + 904, + 857, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4https://github.com/UKPLab/sentence-transformers", + "bbox": [ + 134, + 903, + 450, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8094", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/29c2c49f811e0b2b46b1be1bcb8659e203c48f49683aae2ede5bc8bf2d365d87.jpg", + "image_caption": [ + "(a) Domain $=$ Military (100K items), $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ . Each CE call takes $\\sim 7\\mathrm{ms}$ on an Nvidia 2080ti GPU for a cross-encoder parameterized using 12-layered transformer model." + ], + "image_footnote": [], + "bbox": [ + 117, + 84, + 490, + 197 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91413bc34b2a9a2d551191352b6f7b59461108ea5396ebb81fa13c00eded5eea.jpg", + "image_caption": [ + "(b) Domain=HotpotQA (5 Million items), $|\\mathcal{Q}_{\\mathrm{train}}| = 1000$ . Each CE call takes $\\sim 2$ ms on an Nvidia RTX8000 GPU for a cross-encoder parameterized using 6-layered transformer model.", + "Figure 5: ADACUR inference latency versus number of rounds for two different domains. The primary bottleneck at inference time is the time taken to compute cross-encoder (CE) scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. See §B.5 for detailed discussion." + ], + "image_footnote": [], + "bbox": [ + 505, + 83, + 878, + 197 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cross-encoder score has already been computed, we retrieve more than $\\mathcal{B}_{\\mathrm{CE}} - k_{i}$ items using approximate scores and compute exact cross-encoder scores for them until we have exhausted the entire cross-encoder call budget for the re-ranking step.", + "bbox": [ + 112, + 319, + 487, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B.4 ADACUR Implementation details", + "text_level": 1, + "bbox": [ + 112, + 412, + 425, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For all our $k$ -NN search experiments, we used Nvidia 2080ti GPUs with 12GB memory for domains YuGiOh (10K items), StarTrek (34K items), Military (100K items), and SciDocs (25K items), and we used Nvidia RTX8000 GPUs with 48GB memory for HotpotQA (5 million items).", + "bbox": [ + 112, + 434, + 487, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For HotpotQA, we restrict our $k$ -NN search to top-10K items wrt $\\mathrm{DE}_{\\mathrm{BASE}}$ for ADACUR $_{\\mathrm{DE}_{\\mathrm{BASE}}}$ . For ZESHEL domains and SciDocs, we do not use any such heuristic and search over all the items in the corresponding domain.", + "bbox": [ + 112, + 531, + 487, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "B.5 Time Complexity of ADACUR", + "text_level": 1, + "bbox": [ + 112, + 624, + 400, + 639 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The offline indexing step for ADACUR takes $\\mathcal{O}(k_q|\\mathcal{I}|\\mathcal{C}_{f_\\theta})$ time as it involves computing exact cross-encoder scores for all $|\\mathcal{I}|$ items in the target domain against $k_q$ anchor queries, and computing each cross-encoder score takes $\\mathcal{C}_{f_\\theta}$ units of time.", + "bbox": [ + 112, + 646, + 487, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "At test time, we are given a budget $\\mathcal{B}_{\\mathrm{CE}}$ on the number of cross-encoder calls. Each one of the $N_{\\mathcal{R}}$ rounds during inference with ADACUR involves approximating all item scores for the test query $(q_{\\mathrm{test}})$ followed by sampling the next batch of $k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}$ items using the updated approximate scores. In the $j^{th}$ round, the score approximation step involves computing the pseudo-inverse of a $k_{q} \\times jk_{s}$ -dimensional matrix (line 12 in Algo. 1), which takes $\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s})$ time, followed by a matrix multiplication step to compute updated approximate scores (line 13 in Algo. 1)", + "bbox": [ + 112, + 727, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which takes $\\mathcal{O}(\\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})$ time. The time taken to update the approximate scores in each round is $\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s} + \\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})$ , and the time taken to compute cross-encoder scores for the next batch of $k_{s}$ items is $\\mathcal{O}(k_s\\mathcal{C}_{f_\\theta})$ . Thus, the total inference latency for retrieving items over $N_{\\mathcal{R}}$ rounds under a given budget of $\\mathcal{B}_{\\mathrm{CE}}$ cross-encoder calls is", + "bbox": [ + 507, + 316, + 882, + 434 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {O} \\Big (\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(k _ {s} \\mathcal {C} _ {f _ {\\theta}} + \\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q}, j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q}, j k _ {s}, | \\mathcal {I} |}\\right) \\Big) \\\\ = \\mathcal {O} \\left(\\mathcal {B} _ {\\mathrm {C E}} \\mathcal {C} _ {f _ {\\theta}} + \\underbrace {\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(\\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q} , j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q} , j k _ {s} , | \\mathcal {I} |}\\right)} _ {\\text {O v e r h e a d o f A D A C U R}}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 445, + 855, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 shows the breakdown of ADACUR's inference latency in terms of time spent on computing cross-encoder scores, and the overhead of computing matrix inverse in line 12 and updating approximate scores by multiplying matrices in line 13 of Algorithm 1. Empirically, we observe that the primary bottleneck at inference time is the time taken to compute cross-encoder scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. The overhead is mostly dominated by computing pseudo-inverse (see line 12 in Algorithm 1) and this step is independent of the target domain size. The matrix multiplication step (line 13 in Algorithm 1) has a linear dependence on the number of items in the target domain but it is a negligible fraction of the overall running time as it can be significantly sped up using GPUs.", + "bbox": [ + 507, + 571, + 882, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For ZESHEL domains, we use a cross-encoder parameterized using bert-base (Devlin et al., 2019), and observe that each cross-encoder call takes amortized time of $\\sim 7\\mathrm{ms}$ on an Nvidia 2080ti", + "bbox": [ + 507, + 856, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8095", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPU when the scores are computed in batches of size 50. Computing each cross-encoder score sequentially i.e. with batch-size $= 1$ takes $\\sim 13\\mathrm{ms}$ per score. We did not observe any further reduction in amortized time to compute each score when increasing the batch size beyond 50.", + "bbox": [ + 112, + 84, + 487, + 178 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The amortized time per cross-encoder call is approximately 6ms and 2ms for SciDocs and HotpotQA respectively when using batch size=50 and MINI-LM-based (Wang et al., 2020) cross-encoder. The difference in time per cross-encoder score for SciDocs and HotpotQA is due to the difference in average query-item pair sequence length.", + "bbox": [ + 112, + 180, + 489, + 291 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C Additional Results and Analysis", + "text_level": 1, + "bbox": [ + 112, + 304, + 428, + 322 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C.1 TopK vs SoftMax for ADACUR", + "text_level": 1, + "bbox": [ + 112, + 331, + 400, + 346 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 6 shows Top- $k$ -Recall for ADACUR on domain=YuGiOh, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ , when using TopK and SoftMax strategies for sampling items based on approximate scores (see §2.2 for details). TopK sampling strategy which greedily picks top- $k$ items based on approximate scores results in superior recall as compared to sampling items using softmax over approximate scores.", + "bbox": [ + 112, + 353, + 489, + 479 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/88ec95ef8657532e0fba2ace46f8d303891aeece3b7da3a9729ac6d7f87a43d7.jpg", + "image_caption": [ + "Figure 6: Top- $k$ -Recall for ADACUR on YuGi0h, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ for different strategies for sampling items based on approximate scores described in §2.2." + ], + "image_footnote": [], + "bbox": [ + 117, + 494, + 482, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "C.2 Anchor Item Sampling with Oracle", + "text_level": 1, + "bbox": [ + 112, + 739, + 440, + 755 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ADACUR performs retrieval over multiple rounds using approximate cross-encoder scores and uses the items retrieved based on the approximate scores as anchor items to improve the approximation and hence retrieval in subsequent rounds. In this section, we run experiments where the anchor item sampling method has oracle access to exact cross-encoder scores of all items for the given test query to better understand the effect of anchor items on the approximation of cross-encoder scores and", + "bbox": [ + 112, + 760, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "hence subsequent retrieval based on the approximate scores. We experiment with the following strategies for sampling $k_{i}$ anchor items for a given test query:", + "bbox": [ + 507, + 84, + 884, + 146 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- TopK $_{k_m, \\epsilon}^{\\mathcal{O}}$ : Mask out top- $k_m$ items wrt exact cross-encoder scores and select $k_i$ anchor items by greedily picking $(1 - \\epsilon) k_i$ items starting from rank $k_m + 1$ , and sample remaining $\\epsilon k_i$ anchor items uniformly at random.", + "- SoftMax $\\mathcal{O}_{k_m,\\epsilon}$ : Mask out top- $k_m$ items wrt exact cross-encoder scores and select $k_i$ anchor items by sampling $(1 - \\epsilon)k_i$ anchor items using softmax over exact cross-encoder scores, and sample remaining $\\epsilon k_i$ anchor items uniformly at random." + ], + "bbox": [ + 531, + 148, + 884, + 326 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For a given test-time cross-encoder call budget $\\mathcal{B}_{\\mathrm{CE}}$ , we select $k_{i}$ anchor items, compute approximate cross-encoder scores using the chosen anchor items, and then retrieve $\\mathcal{B}_{\\mathrm{CE}} - k_{i}$ items based on the approximate scores. We experiment with $k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}$ and report results for the best budget split.", + "bbox": [ + 507, + 329, + 882, + 439 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/46192b36f9a83ae7f884c44a35c09ca0e972ff14238cb37c0e0e652fb35ecd76.jpg", + "image_caption": [ + "Figure 7: Average approximation error for CUR matrix factorization on test-queries for domain=YuGiOh and $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ when choosing $k_{i} = 50$ anchor items uniformly at random (ANNCUR), using oracle strategies from §C.2 and for ADACUR when sampling anchor items over five rounds. Approximation error is computed as the average of absolute difference between approximate and exact item scores." + ], + "image_footnote": [], + "bbox": [ + 521, + 453, + 865, + 592 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Figure 8a shows Top- $k$ -Recall of anchor item sampling strategies $\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}$ and SoftMax for $k_{m}\\in \\{0,k\\}$ ,domain $\\equiv$ YuGiOh. Sampling strategies SoftMax and TopK,0, which mask out top-k items, perform significantly worse than SoftMax0,0 and TopK,0 respectively when searching for $k =$ 1, 10 nearest neighbors. This indicates that the significant improvement in Top-1-Recall and Top-10Recall for TopK,and SoftMax sampling strategies can be attributed to the presence of top- $k$ items in the anchor item set. This is because CUR matrix", + "bbox": [ + 507, + 725, + 882, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8096", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/8f7f1770612cfcd55f960a07171c6abbce5cd7b7c03130a041f38b0c9cbdeaee.jpg", + "image_caption": [ + "(a) Sampling with and without ground-truth top- $k$ items" + ], + "image_footnote": [], + "bbox": [ + 117, + 84, + 484, + 225 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/731f63720dd9f479c9ffd9510a11934778c93d5043e7b4cabf77468ce995826d.jpg", + "image_caption": [ + "(b) Sampling using TopK strategy with exact CE scores while varying $\\epsilon$ , the fraction of items sampled uniformly at random" + ], + "image_footnote": [], + "bbox": [ + 117, + 249, + 480, + 359 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/069c5b8fd34201c93a5345d64aec6cec9eaa9cf7437043e7da6070c95ef90aa7.jpg", + "image_caption": [ + "(c) Sampling using SoftMax of exact CE scores while varying $\\epsilon$ , the fraction of items sampled uniformly at random", + "Figure 8: Top- $k$ -Recall for ADACUR, ANNCUR, and oracle sampling strategies (§C.2) that have oracle access to exact cross-encoder scores for all items for domain=YuGiOh, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ .", + "Figure 9: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGiOh, when choosing $k_{i} = 50$ anchor items using oracle strategies from §C.2 and $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ . Top- $k$ for $k = 1,10,100$ wrt exact cross-encoder scores are annotated with text along with vertical lines, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue." + ], + "image_footnote": [], + "bbox": [ + 117, + 394, + 480, + 505 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "factorization which is used to compute the approximate scores incurs negligible approximation error on anchor items, and hence on top- $k$ items when these items are part of the anchor set as shown in figures 7 and 9. For $\\mathsf{TopK}_{k,0}^{\\mathcal{O}}$ and SoftMax $_{k,0}$ sampling strategies, since the top- $k$ items are not part of the anchor set, CUR incurs a much higher approximation error for the top- $k$ items (see examples in Figures 9a and 9b), thus resulting in poor Top $k$ -Recall as shown in Figure 8a.", + "bbox": [ + 112, + 613, + 489, + 772 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Effect of diversity in anchor items Figure 8a shows that sampling items based on softmax of exact cross-encoder scores $(\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}})$ performs better than greedily picking top-scoring items $(\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}})$ , for both $k_{m} = 0,k$ . The reason behind $\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}}$ performing better than $\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}}$ is that sampling based on softmax of exact scores yields an anchor set with a more diverse", + "bbox": [ + 112, + 787, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/d4edc159cafaf082455b5d7a4a6aaf880489564b4ac572d5de20c7eb97be2d5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 80, + 687, + 197 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3b184eae35c46035b21e18a832ec8bbe02e620fb6eb15ece27a3b7e65f6a9c7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 203, + 687, + 312 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/34fc54cb2a1665ab542402350f2ff32cb35feb88c4fcd09a8fe610bf43c6b5f4.jpg", + "image_caption": [ + "(e) $\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}},\\epsilon = 0.75$" + ], + "image_footnote": [], + "bbox": [ + 512, + 317, + 687, + 426 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2f77a9f69c2ead25abd460c0df00537461618319f565f7e2b9279727e2cf9988.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 80, + 880, + 197 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/339b6daec2e6a8ca22674231d1241050aad8810a8688fa8da4bd92ea03412f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 203, + 880, + 312 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/6c73fd025cc0b7b43528bd16db8bbe1ad8fd52d35e6db6f2a06649cdf5dd95fb.jpg", + "image_caption": [ + "(f) SoftMax0,ε, ε = 0.75" + ], + "image_footnote": [], + "bbox": [ + 704, + 317, + 880, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "score distribution whereas greedily selecting topscoring items using exact scores results in an anchor set with items having similar cross-encoder scores. However, as shown in Figures 9c and 9d, both of these sampling strategies can result in overestimating scores for all items, even the irrelevant ones (i.e. items beyond top- $k$ items) due to insufficient representation of the irrelevant items in the anchor set. Thus retrieving based on approximated scores may struggle to retrieve relevant $k$ -NN items, especially for larger values of $k$ such as $k = 100$ when the anchor items are chosen using oracle strategies such as $\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}$ .", + "bbox": [ + 505, + 583, + 882, + 790 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figures 9e and 9f, where $\\epsilon = 75\\%$ of 50 items are sampled uniformly at random, show that overestimating scores of irrelevant items can be avoided by sampling a fraction of anchor items uniformly at random to increase the diversity of the anchor item set. As shown in Figures 8b and 8c, Top- $k$ -Recall for both SoftMax $_{0,\\epsilon}$ and TopK $_{0,\\epsilon}$ generally improves with an increase in $\\epsilon$ , the fraction of ran", + "bbox": [ + 507, + 791, + 884, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8097", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "dom items in the anchor set, due to increased diversity in the anchor item set. Since $\\mathrm{SoftMax}_{0,\\epsilon}^{\\mathcal{O}}$ already samples a diverse set of anchor items, increasing $\\epsilon$ yields only marginal improvement while for $\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}}$ , increasing $\\epsilon$ yields significant improvements due to increased diversity of the anchor set. A small drop in performance is observed for larger values of $\\epsilon$ as increasing $\\epsilon$ beyond a threshold results in some of the top- $k$ items to be excluded from the anchor item set. This results in a poorer approximation of scores for the missing top- $k$ items and hence poor retrieval recall as the retrieval is done using the approximate scores.", + "bbox": [ + 112, + 84, + 489, + 288 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Finally, the optimal strategy for choosing the set of anchor items is the one that strikes a balance between selecting anchor items with diverse cross-encoder scores and greedily picking top- $k$ items. Our proposed strategy ADACUR improves over ANNCUR as greedily picking top-scoring items according to approximate scores to expand set of anchor items increases the likelihood of picking ground-truth $k$ -NN items to be part of the anchor set, with this likelihood improving after each round with improvement in the score approximation, and ADACUR achieves diversity in the anchor items as a result of sampling items uniformly at random in the first round and due to error in the approximate scores, as shown in Figure 11.", + "bbox": [ + 115, + 291, + 489, + 527 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C.3 Comparison with Multi-Vector Models", + "text_level": 1, + "bbox": [ + 112, + 543, + 468, + 558 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Multi-vector models (Khattab and Zaharia, 2020; Ma et al., 2021) produce multiple embeddings for each query and item. For a given query $q$ and item $i$ , the query-item score is computed using simple functions such as average similarity or sum-of-maximum similarities between the set of embeddings for query $q$ and item $i$ .", + "bbox": [ + 112, + 565, + 487, + 677 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 10 shows Top- $k$ -Recall for $\\mathrm{DE}_{\\mathrm{BASE}}$ , $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ , $\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}$ , and MUVER (Ma et al., 2021), a recent multi-vector model trained on ZESHEL dataset. For MUVER, we use the pretrained checkpoint released by Ma et al. (2021) with the view-merging inference strategy as described in Ma et al. (2021). While MUVER can be more accurate than $\\mathrm{DE}_{\\mathrm{BASE}}$ , $\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}$ obtained by finetuning $\\mathrm{DE}_{\\mathrm{BASE}}$ model on the target domain outperforms MUVER and our proposed method $\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}$ yields the best Top- $k$ -Recall versus inference cost trade-offs for all values of $k$ .", + "bbox": [ + 112, + 678, + 487, + 869 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We would also like to note that while multi-vector models such as MUVER can be more accurate than single-embedding models such as $\\mathrm{DE}_{\\mathrm{BASE}}$", + "bbox": [ + 112, + 871, + 489, + 919 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/075c8aae0b2e00531a9aba5d734b5499cca0bc9ef307b92b50e25fa87d081f63.jpg", + "image_caption": [ + "(a) YuGiOh" + ], + "image_footnote": [], + "bbox": [ + 517, + 84, + 875, + 246 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/1f34f8dda5414af9c3860fa48b5405b56577d490818f162f157277aebc39529b.jpg", + "image_caption": [ + "(b) StarTrek", + "Figure 10: Top- $k$ -Recall for $\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}$ and baselines including multi-vector models on ZESHEL domains, $|\\mathcal{Q}_{\\mathrm{train}}| = 2000$ . See §C.3 for discussion." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 875, + 403 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "such multi-vector models incur significant memory overhead for storing query/item embeddings. For instance, using 15 embeddings per item with 768-dimensional embeddings would take around 250GB space for 5 million items for HotpotQA.", + "bbox": [ + 507, + 485, + 882, + 564 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "C.4 Results for TF-IDF baseline", + "text_level": 1, + "bbox": [ + 507, + 574, + 773, + 589 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "TF-IDF: All queries and items are embedded using a TF-IDF vectorizer trained on item descriptions and top- $k$ items are retrieved using the dot-product of sparse query and item embeddings.", + "bbox": [ + 507, + 595, + 880, + 657 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "For domains in ZESHEL, we report results for TF-IDF baseline, for ANNCUR when anchor items are chosen using TF-IDF baseline (ANNCUR $_{TF-IDF}$ ), and for ADACUR when the first batch of anchor items is chosen using TF-IDF baseline (ADACUR $_{TF-IDF}$ ). Figures 13, 14, and 15 show Top- $k$ -Recall for domains YuGiOh, StarTrek, and Military respectively for $Q_{train} \\in \\{100, 500, 2000\\}$ . For each baseline retrieval method, ADACUR always performs better than ANNCUR which in turn generally performs better than merely re-ranking items retrieved using the corresponding baseline retrieval method. In most cases, Top- $k$ -Recall for ADACUR $_{DE_BASE}$ > ANNCUR $_{DE_BASE}$ > DE_BASE, and ADACUR $_{TF-IDF}$ > ANNCUR $_{TF-IDF}$ > TF-IDF.", + "bbox": [ + 507, + 658, + 882, + 910 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8098", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/763b9c48814b7745250edc69497dfb3c38126d61b02e4957e735caa83a0b6642.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 80, + 690, + 112 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/7518ad6ae0a574130bd4080f38f196955d7a2a1c1d79ca4d8ce2109833a17779.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 128, + 347, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/96a47a66eb2f2b4ef6a38afb09f68b8d6853aff1bb8474e5ae63e9e0df010b7c.jpg", + "image_caption": [ + "(a-2) ADACUR Round 2. Total 20 anchor items." + ], + "image_footnote": [], + "bbox": [ + 383, + 128, + 613, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1f379d1691fb3eb6a3681ba1fda992b299e2fb50bfaa076d10e2c519db05df39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 128, + 880, + 247 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/5151c29ddb39fc6530ecfd730fc03df4943f463421ddff978a4042133c404e0d.jpg", + "image_caption": [ + "(a-1) ADACUR Round 1. Total 10 anchor items.", + "(a-4) ADACUR Round 4. Total 40 anchor items." + ], + "image_footnote": [], + "bbox": [ + 117, + 281, + 344, + 401 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/162b5f10276b0571ecf8f0a62fcc9a82a0483204fc7bd1d8c8d96e580a90a811.jpg", + "image_caption": [ + "(a-5) ADACUR Round 5. Total 50 anchor items." + ], + "image_footnote": [], + "bbox": [ + 381, + 281, + 610, + 401 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/6c4e7c6366495e8ef3caaab2830694afa6ce04e3478481dfa4202ca424d13b0a.jpg", + "image_caption": [ + "(a-3) ADACUR Round 3. Total 30 anchor items.", + "(a-6) ANNCUR- Sampling all 50 anchor items uniformly at random." + ], + "image_footnote": [], + "bbox": [ + 645, + 281, + 873, + 400 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/f69a884dabd28ca8be0280a730f6964371c86fb8107c6a1d53b6a4588dd43b4c.jpg", + "image_caption": [ + "(a) Sampling 50 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random)." + ], + "image_footnote": [], + "bbox": [ + 117, + 453, + 347, + 571 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/3b376c8bdf6b620d1392a9d0bc11e0a9b80dcb52d4ac0b52a27ca3344912ad93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 384, + 453, + 611, + 571 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/c82a498bdf536b39bf5dbc1cebe2cb8b68d3d8e77d55a239959ee38bb2e74ae3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 453, + 878, + 571 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/36ffbad975499429d4b0e9fbe49510eddb56bf9700b28b9e3ae37d3120f3b6e9.jpg", + "image_caption": [ + "(b-1) ADACUR Round 1. Total 40 anchor items.", + "(b-4) ADACUR Round 4. Total 160 anchor items.", + "(b) Sampling 200 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random).", + "Figure 11: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGi0h, $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ when choosing $k_{i} = 50$ and 200 anchor items with ADACUR over five rounds, and uniformly at random with ANNCUR. Top- $k$ for $k = 1,10,100$ wrt exact cross-encoder scores are annotated with text, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue. With ADACUR, the first batch containing anchor items in Figure 11a-1 and 11b-1 is chosen uniformly at random and in subsequent rounds, items with highest approximate scores are chosen. Note that the approximation error for top-scoring items improves significantly when the 50 anchor items are chosen adaptively (see Figure 11a-5) with the improvement being much more significant than merely increasing the number of anchor items sampled uniformly at random from 50 in Figure 11a-6 to 200 in Figure 11b-6." + ], + "image_footnote": [], + "bbox": [ + 117, + 605, + 346, + 724 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/91cb2168fd0e13c5f46a97a1cf8b351be75c6bef9bb087d07f8c1b307b09695e.jpg", + "image_caption": [ + "(b-2) ADACUR Round 2. Total 80 anchor items.", + "(b-5) ADACUR Round 5. Total 200 anchor items." + ], + "image_footnote": [], + "bbox": [ + 381, + 605, + 610, + 724 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/1c713cf03417e4fb66d9c05c60dfa6c63b995cc5ed78cb34af3aec20da7ad11f.jpg", + "image_caption": [ + "(b-3) ADACUR Round 3. Total 120 anchor items.", + "(b-6) ANNCUR- Sampling all 200 anchor items uniformly at random." + ], + "image_footnote": [], + "bbox": [ + 645, + 605, + 873, + 724 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8099", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/548aaf1aef290dfdacc82ec3453e94d19934334b277e2c81e44aa9a524cae0d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 275, + 297, + 724, + 321 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/5cf1e21353df666ec04e7cb036f03c5d321ba5d2f3e051d562d6f3b7b30942d5.jpg", + "image_caption": [ + "(a) SciDocs" + ], + "image_footnote": [], + "bbox": [ + 115, + 325, + 880, + 494 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/45dcacc22c42b998ee9475d79959dc13350f58d2d3abc69aab19a7a49d6e1f3d.jpg", + "image_caption": [ + "(b) HotpotQA", + "Figure 12: Top- $k$ -Recall for ADACUR (using ten rounds) and baselines for SciDocs and HotpotQA, $|\\mathcal{Q}_{\\mathrm{train}}| = 1000$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 529, + 880, + 674 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "8100", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/26c367d591eafc828e8ec3fa8bb01cd893bbee9e643b5dc7c63da4191b66d281.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 183, + 394, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/dc1c1e07649c52381c6155f65fb7df2f8875ce3804e9d64c99eed8b365d16e0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 505, + 184, + 858, + 203 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7caefe3aca3ccd9ae6909d884802e7775f61ed061657b1062175088d48d7c369.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 260, + 212, + 737, + 230 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/818abc361746554d903cacdd8f337f48e3f46664230fc529991d037c924a878f.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 249, + 406, + 379 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d85e84b09a763817d072b1320fc6f98d929a283c552f7a2d373aade6d32f5a8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 250, + 672, + 379 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/718beb75cf87b581b7282721bb0e92e396ecfb0cc0c05c3d2c521272c71fc0c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 250, + 878, + 379 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d6e18a9b73534f56dda5f095c391c89e3f4f6e3866314e916c6bf26f78afbd68.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 431, + 406, + 558 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/b52b342d648f393aa2e434f9775554dc850e0ae0e26e91aa834841b00fdd1e30.jpg", + "image_caption": [ + "(a) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 100$ ." + ], + "image_footnote": [], + "bbox": [ + 418, + 431, + 673, + 558 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3d070858727996c8cd8553c21c7b517d98f2b9454e6e322df036bb0080c755d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 431, + 878, + 558 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a6a2b60dedf4b8e1d94ed34aa0e62b2ab92873291cb31172520e3e25183e3266.jpg", + "image_caption": [ + "(b) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ .", + "Inference Cost (Total Number of Cross-Encoder Calls)", + "(c) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 2000$ .", + "Figure 13: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=YuGiOh. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\\mathcal{Q}_{\\text {train }}|)$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 612, + 406, + 740 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/2df33e476e3c134a616b3611da7f4056eabe4e5274a16dc9280a4ebbc2f20314.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 612, + 673, + 740 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/af420a45981dc937114ec3b8c1f2a0d3ed76f0faf7cc6bf350cac5551de96d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 612, + 878, + 740 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "8101", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/8fc6aa09b47562a568286c43a5f0751d4fb8bc7f5291a76d9b167ab6e089face.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 184, + 394, + 203 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/193b5ce2fb71560cc198e7b315a9220074d6fc389adc07e07aae0978a957e76c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 505, + 184, + 858, + 203 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/3ccf1511fc31c8bffc0db6503a9a735c57c7b1c983c5aa3a265f67efb26a1192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 260, + 212, + 737, + 231 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/fb751c61a7f6d3a7e7cf0ad1ecb4df49b04148a5c343609e6abc21aebb322ceb.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 249, + 408, + 379 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/828d4053b0ba61338871c695d331ff7aca3f1b66d92a776827f31a029b8335f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 249, + 675, + 379 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/3daf09084dc3fd2f6ee54d843d7c2232e38e8c1031de72eaf0ea1706244f0a15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 250, + 880, + 379 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/4d5d66fe6e128567aa6fa9720657c0dd8bbcfdc4d1170bcc4b12bc5016a5e669.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 431, + 408, + 558 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/0a4caec315ba0439acceea4f6791efb451cff55db854029ea953990aa1d7f449.jpg", + "image_caption": [ + "(a) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 100$" + ], + "image_footnote": [], + "bbox": [ + 418, + 431, + 675, + 558 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/7eb4b41eff8084365b22a320c07bb9cfb45ff4ad494d696bebf723d3a0898d45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 431, + 880, + 558 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/4d44ed728ec4c5448ce05911874c41e131047aa70bac2767856963857438068d.jpg", + "image_caption": [ + "(b) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ .", + "Inference Cost (Total Number of Cross-Encoder Calls)", + "(c) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 2000$ .", + "Figure 14: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=StarTrek. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\\mathcal{Q}_{\\mathrm{train}}|)$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 612, + 408, + 740 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/3cfe699bb790ada1a2c1459da3fb3442c5b77988509b09690d430cd92b4758a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 612, + 675, + 740 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/a4c791d390c0c11ac9c0a7b5bd8fce0a25e27d766bbc18098428af1f976724a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 612, + 880, + 740 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "8102", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/c206c7baf8658830cb6aca78021beeabcfdaf3716eb97a831b6d877f8126306c.jpg", + "image_caption": [ + "Proposed Approaches" + ], + "image_footnote": [], + "bbox": [ + 122, + 171, + 393, + 189 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/ad691f5806c7f6d529484dade0d1c51d78bbef1f6ddc415fd2c4692dcbee40ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 505, + 171, + 858, + 189 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/8de34fc979ea5757c0a473b97caa7406fb3c16e48de28e3fe577910e597b8c5f.jpg", + "image_caption": [ + "DACUR" + ], + "image_footnote": [], + "bbox": [ + 389, + 200, + 425, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/9eb61d65e7a5b6446841ec364adc76aea5fcdb36c74fc808337d25f9d337e7a4.jpg", + "image_caption": [ + "ADA" + ], + "image_footnote": [], + "bbox": [ + 490, + 200, + 519, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/4951cb502194006a395ca30868bb9fa756a46dc61b61cb996d4cf40edb4e2d37.jpg", + "image_caption": [ + "R" + ], + "image_footnote": [], + "bbox": [ + 551, + 200, + 573, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/1f78ac0be2fd48a2926be69ec03e15a9ca7470dd5981c96d479d6ddee346f069.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 608, + 200, + 638, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/4fc45c793e7d2ade24038526682a09f9f4d7be1585cbe99bfeefcb500e4b18ba.jpg", + "image_caption": [ + "AD", + "\"" + ], + "image_footnote": [], + "bbox": [ + 673, + 200, + 702, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/6415aa68c1a31c13c8a53ba8cdb651ddcd9e37c1f185196b2992612b787c365c.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 737, + 200, + 766, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/7f87e08e8f83543348c1f535aba8374d15184fa31a41d7861c771b6f1813d7cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 200, + 855, + 211 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/452c47ed38e428ebd5a84e12d9c181c4689ea1fd63e8c6ccf9b8522184a24444.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 236, + 406, + 365 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/a88d57a0c64c3b47a74ad1044c677bb43b2b6a059d6c42c3492d9f663714474e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 237, + 673, + 363 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/5fcea90f2cff7cbce75777cb63906624f696c247e9cf6b32c6d458ba0b940280.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 237, + 878, + 363 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/103460186e44d634f0671448bba078f34bebad947e0c8d9725279e21e224a9ae.jpg", + "image_caption": [ + "Inference Cost (Total Number of Cross-Encoder Calls)" + ], + "image_footnote": [], + "bbox": [ + 117, + 417, + 406, + 545 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/1b208771490ef453d2aad2814f1eb473e6b6534912dc4c840f768069b16f62e8.jpg", + "image_caption": [ + "(a) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 100$ ." + ], + "image_footnote": [], + "bbox": [ + 418, + 417, + 673, + 545 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/c982832636752109ab23ef035405faab9f68d52bac0b76386fab6c27176d98b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 417, + 878, + 545 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/203513f44c2614a26a273351a33a23b528c2a3ec0bd168a0e9ce411312851fa0.jpg", + "image_caption": [ + "(b) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 500$ .", + "Inference Cost (Total Number of Cross-Encoder Calls)", + "(c) Number of train/anchor queries $|\\mathcal{Q}_{\\mathrm{train}}| = 2000$ .", + "Figure 15: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=Military. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\\mathcal{Q}_{train}|)$ . Note that $\\mathrm{DE}_{\\mathrm{BASE}}$ has high Top-1-Recall values as domain=Military is included in the set of train domains in ZESHEL which are used to train both $\\mathrm{DE}_{\\mathrm{BASE}}$ and the cross-encoder model." + ], + "image_footnote": [], + "bbox": [ + 117, + 598, + 406, + 726 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/c70668bc95c575f3ab747d5b95f4885237498be6a8b9e037a6f116d7ef56b89b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 418, + 598, + 673, + 726 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/1c3751bc076387a24ad37899f74b4a3f7f59313542929440e5c428a7dd1965d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 685, + 598, + 878, + 726 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "8103", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_model.json b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2f0fc1c994a88bbefebbc44cfcdb121fc91cc584 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_model.json @@ -0,0 +1,3543 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.24, + 0.091, + 0.758, + 0.131 + ], + "angle": 0, + "content": "Efficient \\(k\\)-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.154, + 0.838, + 0.171 + ], + "angle": 0, + "content": "Nishant Yadav†, Nicholas Monath‡, Manzil Zaheer§, and Andrew McCallum†" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.172, + 0.749, + 0.189 + ], + "angle": 0, + "content": "† University of Massachusetts Amherst ‡Google Research" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.19, + 0.868, + 0.205 + ], + "angle": 0, + "content": "{nishantyavad, mccallum}@cs.umass.edu {nmonath,manzilzaheer}@google.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.28, + 0.461, + 0.741 + ], + "angle": 0, + "content": "Cross-encoder models, which jointly encode and score a query-item pair, are prohibitively expensive for direct \\( k \\)-nearest neighbor (\\( k \\)-NN) search. Consequently, \\( k \\)-NN search typically employs a fast approximate retrieval (e.g. using BM25 or dual-encoder vectors), followed by reranking with a cross-encoder; however, the retrieval approximation often has detrimental recall regret. This problem is tackled by ANNCUR (Yadav et al., 2022), a recent work that employs a cross-encoder only, making search efficient using a relatively small number of anchor items, and a CUR matrix factorization. While ANNCUR's one-time selection of anchors tends to approximate the cross-encoder distances on average, doing so forfeits the capacity to accurately estimate distances to items near the query, leading to regret in the crucial end-task: recall of top-\\( k \\) items. In this paper, we propose ADACUR, a method that adaptively, iteratively, and efficiently minimizes the approximation error for the practically important top-\\( k \\) neighbors. It does so by iteratively performing \\( k \\)-NN search using the anchors available so far, then adding these retrieved nearest neighbors to the anchor set for the next round. Empirically, on multiple datasets, in comparison to previous traditional and state-of-the-art methods such as ANNCUR and dual-encoder-based retrieval-and-erank, our proposed approach ADACUR consistently reduces recall error—by up to \\( 70\\% \\) on the important \\( k = 1 \\) setting—while using no more compute than its competitors." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.752, + 0.26, + 0.767 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.777, + 0.49, + 0.92 + ], + "angle": 0, + "content": "\\(k\\)-nearest neighbor (\\(k\\)-NN) search is a core subroutine of a variety of tasks in NLP such as entity linking (Wu et al., 2020), passage retrieval for QA (Karpukhin et al., 2020), and more generally, in retrieval-augmented machine learning models (Guu et al., 2020; Izacard et al., 2023). For many of these applications, the state-of-the-art similarity function is a cross-encoder that directly outputs a scalar similarity score after jointly encoding a" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.251, + 0.853, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.291, + 0.688, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.39, + 0.678, + 0.404 + ], + "angle": 0, + "content": "(a) Uniformly at random" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.291, + 0.882, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.39, + 0.855, + 0.404 + ], + "angle": 0, + "content": "(b) Biased sampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.415, + 0.885, + 0.541 + ], + "angle": 0, + "content": "Figure 1: Exact versus approximate cross-encoder scores (computed using ANNCUR) of all items for a test-query in domain=YuGi0h. ANNCUR incurs high approximation error on \\( k \\)-NN items wrt exact scores when using 50 anchor items sampled uniformly at random (Fig. 1a). In contrast, sampling 50 anchor items with probability proportional to exact cross-encoder scores (Fig. 1b) significantly improves approximation of top-scoring items." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.553, + 0.884, + 0.805 + ], + "angle": 0, + "content": "given query-item pair. However, computing a single query-item score using a cross-encoder requires a forward pass of the model which can be computationally expensive as cross-encoders are typically parameterized using deep neural models such as transformers (Vaswani et al., 2017). For this reason, \\( k \\)-NN search with a cross-encoder typically involves retrieving candidate items using additional models such as dual-encoders or BM25 (Robertson et al., 2009), followed by re-ranking items using the cross-encoder (Logeswaran et al., 2019; Zhang and Stratos, 2021; Qu et al., 2021). However, the accuracy of such retrieve-and-erank approaches is upperbound by the recall of first-stage retrieval and may require computationally expensive distillation-based training of dual-encoders to improve recall." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.809, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Recent work by Yadav et al. (2022) proposed ANNCUR, a CUR factorization (Mahoney and Drineas, 2009) based method, that approximates cross-encoder using dot-product of latent query and item embeddings and performs \\(k\\)-NN retrieval using approximate scores followed by optionally re-ranking retrieved items using exact cross" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8088" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.78, + 0.958 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8088-8103" + }, + { + "type": "footer", + "bbox": [ + 0.278, + 0.959, + 0.72, + 0.972 + ], + "angle": 0, + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.273 + ], + "angle": 0, + "content": "encoder scores. The latent item embeddings are computed by comparing each item against a set of anchor queries, and the latent query embedding is computed using the query's cross-encoder scores against a fixed set of anchor items. As shown in Figure 1, when ANNCUR selects the anchor items uniformly at random (Fig 1a), it incurs higher approximation error on the top-scoring items than the rest of the items, resulting in poor \\( k \\)-NN recall, and including some \\( k \\)-NN items as part of anchor items (Fig. 1b) can significantly improve approximation error for top-scoring items." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.276, + 0.49, + 0.59 + ], + "angle": 0, + "content": "In this work, we propose ADACUR, a search strategy that improves \\( k \\)-NN search recall by improving the approximation of top-scoring items. ADACUR performs retrieval over multiple rounds, retrieving the first batch of items either uniformly at random or using heuristic or auxiliary models such as dual-encoder or BM25 to get a first coarse approximation of item scores for the test query. In subsequent rounds, it alternates between a) performing retrieval using approximate scores and scoring retrieved items using cross-encoder and b) using all retrieved items as anchor items to improve the approximation and hence retrieval of relevant items in the subsequent rounds. Our proposed approach provides significant improvements in \\( k \\)-NN search recall over ANNCUR and dual-encoder based retrieve-and-erank approaches when performing \\( k \\)-NN search with cross-encoder models trained for the task of entity linking and information retrieval." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.6, + 0.396, + 0.617 + ], + "angle": 0, + "content": "2 Proposed Method: ADACUR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.624, + 0.488, + 0.704 + ], + "angle": 0, + "content": "Task Given a scoring function \\(f_{\\theta} \\colon \\mathcal{Q} \\times \\mathcal{I} \\to \\mathbb{R}\\) that maps a query-item pair to a scalar score, and a query \\(q \\in \\mathcal{Q}\\), the \\(k\\)-nearest neighbor search task is to retrieve top-\\(k\\) scoring items from a fixed item set \\(\\mathcal{I}\\) according to the given scoring function \\(f_{\\theta}\\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.713, + 0.445, + 0.728 + ], + "angle": 0, + "content": "2.1 ADACUR: Offline Indexing of Items" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.733, + 0.487, + 0.796 + ], + "angle": 0, + "content": "The indexing step of ADACUR involves using the cross-encoder model \\((f_{\\theta})\\) to score each item against a fixed set of \\(k_q\\) anchor/train queries \\((Q_{train})\\), to get score matrix \\(R_{\\mathrm{anc}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.802, + 0.457, + 0.819 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {a n c}} (q, i) = f _ {\\theta} (q, i), \\forall (q, i) \\in \\mathcal {Q} _ {\\text {t r a i n}} \\times \\mathcal {I}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.825, + 0.489, + 0.859 + ], + "angle": 0, + "content": "Each column of \\( E^{\\mathcal{I}} \\coloneqq R_{\\mathrm{anc}} \\in \\mathbb{R}^{k_q \\times |\\mathcal{I}|} \\) corresponds to a \\( k_q \\)-dimensional latent item embedding." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.867, + 0.398, + 0.882 + ], + "angle": 0, + "content": "2.2 ADACUR: Test-time inference" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.923 + ], + "angle": 0, + "content": "The baseline method ANNCUR computes the latent test-query embedding \\( e_{q_{\\mathrm{test}}} \\) using \\( C_{q_{\\mathrm{test}}} \\), a" + }, + { + "type": "code_caption", + "bbox": [ + 0.512, + 0.085, + 0.79, + 0.1 + ], + "angle": 0, + "content": "Algorithm 1 ADACUR \\(k\\)-NN Search" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.104, + 0.884, + 0.371 + ], + "angle": 0, + "content": "1: Input: \\((q_{\\mathrm{test}}, R_{\\mathrm{anc}}, N_{\\mathcal{R}}, \\mathcal{B}_{\\mathrm{CE}}, \\mathcal{A})\\) \n2: \\(q_{\\mathrm{test}}\\): Test query \n3: \\(R_{\\mathrm{anc}}\\): Matrix containing CE scores between \\(\\mathcal{Q}_{\\mathrm{train}}\\) and \\(\\mathcal{I}\\) \n4: \\(\\mathcal{B}_{\\mathrm{CE}}\\): Total cross-encoder (CE) call budget. \n5: \\(\\mathcal{A}\\): Algorithm to use for selecting (anchor) items. \n6: \\(N_{\\mathcal{R}}\\): Number of iterative search rounds \n7: Output: \\(\\hat{S}\\): Approximate scores of \\(q_{\\mathrm{test}}\\) with all items, \\(\\mathcal{I}_{\\mathrm{anc}}\\): Retrieved (anchor) items with CE scores in \\(C_{\\mathrm{test}}\\). \n8: \\(k_s \\gets \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}} \\triangleright\\) Num. of items to sample per round \n9: \\(\\mathcal{I}_{\\mathrm{anc}} \\gets \\text{INIT}(\\mathcal{I}, k_s)\\) \\(\\triangleright\\) Initial set of anchor items \n10: \\(C_{\\mathrm{test}} \\gets [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}} \\triangleright\\) CE scores of \\(\\mathcal{I}_{\\mathrm{anc}}\\) for \\(q_{\\mathrm{test}}\\) \n11: for \\(j \\gets 2\\) to \\(N_{\\mathcal{R}}\\) do \n12: \\(U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}\\) \\(\\triangleright\\) \\(U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}\\) \n13: \\(\\hat{S}^{(j)} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright\\) Update approx. scores \n14: \\(\\mathcal{I}_{\\mathrm{anc}}^{(j)} \\gets \\text{SAMPLEITEMS}(\\mathcal{A}, k_s, \\mathcal{I}_{\\mathrm{anc}}, \\hat{S}^{(j)})\\) \n15: \\(\\mathcal{I}_{\\mathrm{anc}} \\gets \\mathcal{I}_{\\mathrm{anc}} \\cup \\mathcal{I}_{\\mathrm{anc}}^{(j)}\\) \n16: \\(C_{\\mathrm{test}} \\gets C_{\\mathrm{test}} \\oplus [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}^{(j)}} \\triangleright\\) Update \\(C_{\\mathrm{test}}\\) \n17: \\(U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}\\) \\(\\triangleright\\) \\(U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}\\) \n18: \\(\\hat{S} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright\\) Compute approx. scores \n19: return \\(\\hat{S}, \\mathcal{I}_{\\mathrm{anc}}, C_{\\mathrm{test}}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.511, + 0.389, + 0.729, + 0.404 + ], + "angle": 0, + "content": "Algorithm 2 SAMPLEITEMS" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.405, + 0.882, + 0.598 + ], + "angle": 0, + "content": "1: Input: \\((\\mathcal{A}, k_s, \\mathcal{I}_{\\text{mask}}, S)\\) \n2: \\(\\mathcal{A}\\): Algorithm for sampling items \n3: \\(k_s\\): Number of items to sample \n4: \\(\\mathcal{I}_{\\text{mask}}\\): Set of items to mask \n5: \\(S\\): (Approximate) Scores for all items \n6: Output: \\(\\mathcal{I}_{\\text{select}}\\): Set of sampled \\(k_s\\) items \n7: \\(\\bar{S} \\gets \\text{SOFTMAX}(S)\\) \n8: \\(\\bar{S}[\\mathcal{I}_{\\text{mask}}] \\gets 0\\) > Mask items in \\(\\mathcal{I}_{\\text{mask}}\\) \n9: if \\(\\mathcal{A} = \\text{TopK}\\) then \n10: \\(\\mathcal{I}_{\\text{select}} \\gets \\text{TOPK}(\\bar{S}, k_s)\\) \n11: else if \\(\\mathcal{A} = \\text{SoftMax}\\) then \n12: \\(\\mathcal{I}_{\\text{select}} \\gets k_s\\) items sampled using \\(\\bar{S}\\) \n13: else if \\(\\mathcal{A} = \\text{Random}\\) then \n14: \\(\\mathcal{I}_{\\text{select}} \\gets k_s\\) items uniformly sampled from \\(\\mathcal{I} \\setminus \\mathcal{I}_{\\text{mask}}\\) \n15: return \\(\\mathcal{I}_{\\text{select}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.625, + 0.884, + 0.658 + ], + "angle": 0, + "content": "\\(|\\mathcal{I}_{\\mathrm{anc}}|\\)-dimensional vector containing cross-encoder scores of \\(q_{\\mathrm{test}}\\) with a set of anchor items \\((\\mathcal{I}_{\\mathrm{anc}})\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.67, + 0.801, + 0.689 + ], + "angle": 0, + "content": "\\[\nC _ {q _ {\\text {t e s t}}} = \\left[ f _ {\\theta} \\left(q _ {\\text {t e s t}}, i\\right) \\right] _ {i \\in \\mathcal {I} _ {\\text {a n c}}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.693, + 0.754, + 0.712 + ], + "angle": 0, + "content": "\\[\ne _ {q _ {\\text {t e s t}}} = C _ {q _ {\\text {t e s t}}} \\times U\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.724, + 0.884, + 0.823 + ], + "angle": 0, + "content": "where \\(U\\in \\mathbb{R}^{|\\mathcal{I}_{\\mathrm{anc}}|\\times |\\mathcal{Q}_{\\mathrm{train}}|}\\) is the pseudo-inverse of \\(R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]\\), the subset of columns of \\(R_{\\mathrm{anc}}\\) corresponding to (anchor) items \\(\\mathcal{I}_{\\mathrm{anc}}\\). Finally, ANCUR approximates the score for a query-item pair \\((q_{\\mathrm{test}},i)\\) using dot-product of the query embedding \\(e_{q_{\\mathrm{test}}}\\) and item embedding \\(E^{\\mathcal{I}}[:,i]\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.829, + 0.8, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {\\theta} (q _ {\\mathrm {t e s t}}, i) = e _ {q _ {\\mathrm {t e s t}}} ^ {\\top} E ^ {\\mathcal {I}} [:, i ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.92 + ], + "angle": 0, + "content": "The main bottleneck at test-time inference is the number of items scored using the cross-encoder for the given test-query. For a given budget of cross-encoder calls, ANNCUR splits the budget \\((\\mathcal{B}_{\\mathrm{CE}})\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "8089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.163 + ], + "angle": 0, + "content": "into two parts - it uses \\( k_{i} \\) cross-encoder calls to compare against anchor items (chosen uniformly at random or using heuristic methods) and retrieves \\( \\mathcal{B}_{\\mathrm{CE}} - k_{i} \\) items using approximate scores and re-ranks them using exact cross-encoder scores." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.165, + 0.49, + 0.416 + ], + "angle": 0, + "content": "In contrast, our proposed approach ADACUR uses the cross-encoder call budget to adaptively retrieve and score items over \\(N_{\\mathcal{R}}\\) rounds, repurposing the retrieved items as anchor items as shown in Algorithm 1. ADACUR begins by sampling the first batch of \\(k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}\\) (anchor) items uniformly at random. The first batch of items can also be selected using baseline retrieval methods such as BM25 and dual-encoders. In the \\(j^{\\mathrm{th}}\\) round, the items retrieved up to round \\(j - 1\\) are used as anchor items to revise the test-query embedding, which in turn is used to update the approximate scores (line 13 in Algorithm 1). Finally, the items selected so far are masked out and the next batch of \\(k_{s}\\) items in round \\(j\\) is retrieved using the updated approximate scores in the following two ways:" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.417, + 0.486, + 0.448 + ], + "angle": 0, + "content": "- TopK: Greedily pick top-\\(k_{s}\\) items according to approximate scores." + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.454, + 0.486, + 0.5 + ], + "angle": 0, + "content": "- SoftMax: Convert approximate item scores into probability using softmax and sample \\(k_{s}\\) items without replacement." + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.417, + 0.486, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.581 + ], + "angle": 0, + "content": "Finally, ADACUR returns top-\\(k\\) items based on exact cross-encoder scores\\(^{1}\\) from the set of retrieved (anchor) items as approximate \\(k\\)-NN items. We refer interested readers to Appendix B.5 for a discussion on the time complexity of ADACUR." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.591, + 0.262, + 0.608 + ], + "angle": 0, + "content": "3 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.616, + 0.49, + 0.711 + ], + "angle": 0, + "content": "In our experiments, we evaluate the proposed approach and baselines on the task of finding \\( k \\)-nearest neighbor items as per a given cross-encoder. We experiment with two cross-encoders - one trained for the task of zero-shot entity linking, and another trained on information retrieval datasets." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.72, + 0.49, + 0.862 + ], + "angle": 0, + "content": "Experimental Setup We run evaluation on domains YuGiOh, StarTrek, and Military from ZESHEL-a zero-shot entity linking dataset (Logeswaran et al., 2019), and domains SciDocs and HotpotQA from BEIR-a zero-shot information retrieval benchmark (Thakur et al., 2021). We use two cross-encoder models trained on labeled training data from the corresponding benchmark and evaluate separately on each domain on the task of find" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.196 + ], + "angle": 0, + "content": "ing \\(k\\)-NN cross-encoder items. For each ZESHEL domain, we randomly split the query set into a train/anchor set \\((Q_{train})\\) and a test set \\((Q_{test})\\) while for BEIR domains, we use pseudo-queries released as part of the benchmark as train/anchor queries and evaluate on queries in the official test split. We refer the reader to Table 1 for additional details." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.214, + 0.882, + 0.246 + ], + "angle": 0, + "content": "Baselines We compare our proposed approach with the following baseline retrieval methods." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.249, + 0.884, + 0.39 + ], + "angle": 0, + "content": "Dual-Encoders (DE): Query-item scores are computed using dot-product of embeddings produced by a learned deep encoder model. DE is used for initial retrieval followed by re-ranking using the cross-encoder. We report results for \\(\\mathrm{DE}_{\\mathrm{BASE}}\\), a dual-encoder trained on training domains in the corresponding dataset, and the following two dual-encoder models trained on the target domain via distillation using the cross-encoder." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.394, + 0.884, + 0.445 + ], + "angle": 0, + "content": "- \\(\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}\\): DE initialized with BERT (Devlin et al., 2019) and trained only on the target domain via distillation using the cross-encoder." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.448, + 0.882, + 0.483 + ], + "angle": 0, + "content": "- \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\): \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) model further fine-tuned on the target domain via distillation." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.394, + 0.884, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.487, + 0.884, + 0.567 + ], + "angle": 0, + "content": "ANNCUR: \\(k\\)-NN search method proposed by Yadav et al. (2022) where anchor items are chosen uniformly at random. We additionally experiment with ANNCUR\\(_{\\text{DE}_{\\text{BASE}}}\\) which uses top-scoring items retrieved using DE\\(_{\\text{BASE}}\\) as anchor items." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.585, + 0.883, + 0.79 + ], + "angle": 0, + "content": "Evaluation Metric Following the precedent set by previous work (Yadav et al., 2022), we evaluate all approaches using Top-\\(k\\)-Recall@ \\(\\mathcal{B}_{\\mathrm{CE}}\\) which is defined as the fraction of \\(k\\)-NN items retrieved under test-time cost budget \\(\\mathcal{B}_{\\mathrm{CE}}\\) where the cost is defined as the number of cross-encoder calls made during inference. DE baselines will use the entire budget of \\(\\mathcal{B}_{\\mathrm{CE}}\\) calls for re-ranking retrieved items using exact cross-encoder scores, ANNCUR splits the budget between scoring anchor \\(k_{i}\\) items and using exact cross-encoder scores for re-ranking \\(\\mathcal{B}_{\\mathrm{CE}} - k_{i}\\) retrieved items, and ADACUR use the budget to adaptively search for \\(k\\)-NN items." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.794, + 0.884, + 0.919 + ], + "angle": 0, + "content": "For ADACUR, unless noted otherwise, we use \\( N_{\\mathcal{R}} = 5 \\) with TopK method for retrieving items using approximate scores, and retrieve the first batch of items uniformly at random (ADACUR) or using \\( \\mathrm{DE}_{\\mathrm{BASE}} \\) (ADACUR \\( \\mathrm{DE}_{\\mathrm{BASE}} \\)). We refer readers to Appendix B for implementation details and details on training and parameterization of cross-encoder and dual-encoder models used in our experiments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.87, + 0.488, + 0.919 + ], + "angle": 0, + "content": "\\( {}^{1} \\) Sorting retrieved items based on exact cross-encoder scores does not require any additional cross-encoder calls as cross-encoder scores for these items have already been computed (see line 16 in Algorithm 1)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.085, + 0.844, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.252, + 0.885, + 0.3 + ], + "angle": 0, + "content": "Figure 2: Top-\\(k\\)-Recall for ADACUR and baselines for domain=YuGi0h, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\). ADACUR consistently outperforms the corresponding ANNCUR variant and \\(\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}\\) outperforms all DE-based retrieve-and-erank approaches including \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\), a DE model trained via distillation on the target domain using the cross-encoder." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.317, + 0.22, + 0.332 + ], + "angle": 0, + "content": "3.1 Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.342, + 0.49, + 0.488 + ], + "angle": 0, + "content": "ADACUR versus baselines Figure 2 shows Top- \\(k\\) -Recall for ADACUR and baselines on domain \\(\\equiv\\) YuGiOh. ADACUR, which uses adaptively retrieved items as anchor items over \\(N_{\\mathcal{R}} = 5\\) rounds, consistently outperforms ANNCUR which selects all anchor items uniformly at random. ADACUR also outperforms strong DE baseline \\(\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}\\) for all values of \\(k\\) and outperforms \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) & \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) for large values of \\(k = 10,100\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.505, + 0.49, + 0.839 + ], + "angle": 0, + "content": "Sampling anchor items using \\(\\mathbf{DE}_{\\mathrm{BASE}}\\) For \\(k = 1, 10\\), Top-\\(k\\)-Recall for both ANNCUR and ADACUR can be further improved by leveraging baseline retrieval models such as \\(\\mathbf{DE}_{\\mathrm{BASE}}\\) for retrieving all and the first batch of anchor items respectively instead of sampling them uniformly at random. ADACUR\\(_{\\mathsf{DE}_{\\mathrm{BASE}}}\\) always performs better than ANNCUR\\(_{\\mathsf{DE}_{\\mathrm{BASE}}}\\) which in turn performs better than re-ranking items retrieved using \\(\\mathbf{DE}_{\\mathrm{BASE}}\\). Thus, for a given cost budget \\((\\mathcal{B}_{\\mathrm{CE}})\\), even when a strong baseline retrieval model such as \\(\\mathbf{DE}_{\\mathrm{BASE}}\\) is available, using the baseline retrieval model to select the first batch of \\(\\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}\\) items followed by using the proposed approach to adaptively retrieve more items over remaining \\(N_{\\mathcal{R}} - 1\\) rounds performs better than merely re-ranking \\(\\mathcal{B}_{\\mathrm{CE}}\\) top-scoring items retrieved using \\(\\mathbf{DE}_{\\mathrm{BASE}}\\). Finally, note that ADACUR\\(_{\\mathsf{DE}_{\\mathrm{BASE}}}\\) outperforms all baselines including \\(\\mathbf{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) which requires additional compute- and resource-intensive distillation-based fine-tuning of \\(\\mathbf{DE}_{\\mathrm{BASE}}\\) on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.841, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We refer the reader to Appendix C for results on other domains, training data size \\((|\\mathcal{Q}_{\\mathrm{train}}|)\\) values, oracle-based sampling experiments (§C.2), and additional results for multi-vector encoder (§C.3) and TF-IDF (§C.4) baselines." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.317, + 0.868, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.442, + 0.885, + 0.472 + ], + "angle": 0, + "content": "Figure 3: Top-\\(k\\)-Recall for ADACUR for different number of rounds for domain=YuGiOh, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.487, + 0.887, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.611, + 0.884, + 0.641 + ], + "angle": 0, + "content": "Figure 4: ADACUR inference latency versus number of rounds for domain \\(= \\text{YuGi0h}\\), \\(|\\mathcal{Q}_{\\text{train}}| = 500\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.667, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Effect of number of rounds Figure 3 shows Top- \\(k\\)-Recall and Figure 4 shows total inference latency of ADACUR on primary y-axis for various values of the test-time cross-encoder call budget \\((\\mathcal{B}_{\\mathrm{CE}})\\). The secondary y-axis in Figure 4 shows the fraction of total inference time spent on each one of the three main steps in Algorithm 1 - (a) computing \\(U\\) as pseudo-inverse of \\(R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]\\) (line 12), (b) updating approximate scores for all items (line 13), and (c) computing exact cross-encoder scores for retrieved items (line 16). Since ADACUR selects items uniformly at random in the first round, ADACUR with \\(N_{\\mathcal{R}} = 1\\) performs poorly as it simply returns a subset of items sampled uniformly at random. As expected, Top-\\(k\\)-Recall for ADACUR increases with the number of rounds and saturates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.941 + ], + "angle": 0, + "content": "8091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.306 + ], + "angle": 0, + "content": "at around 5-10 rounds while incurring negligible overhead in inference latency. As shown in Figure 4, cross-encoder score computation is the main bottleneck in test-time inference, taking \\(\\sim 7\\) ms per score2. Increasing \\(N_{\\mathcal{R}}\\) to large values such as 100 can incur up to \\(25\\%\\) overhead with step (a) contributing significantly to this overhead. Although the time taken by matrix multiplication in step (b) is linear in the number of items in the domain, we observe that it is a negligible fraction of overall latency on GPUs even for domain=HotpotQA with 5 million items (see Figure 5) as GPUs can enable significant speedup even for brute-force computation of this step." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.321, + 0.248, + 0.337 + ], + "angle": 0, + "content": "4 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.349, + 0.49, + 0.507 + ], + "angle": 0, + "content": "In this paper, we presented an adaptive search strategy that incrementally builds a query embedding to approximate cross-encoder scores and performs \\( k \\)-NN search using approximate scores over several rounds. Our approach is designed to reduce approximation error for the top-scoring items and hence improves \\( k \\)-NN search recall when retrieving items based on the approximate scores. We perform an in-depth empirical analysis of the proposed approach in terms of both retrieval quality and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.522, + 0.221, + 0.537 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.55, + 0.49, + 0.771 + ], + "angle": 0, + "content": "Building the index for the ADACUR is more expensive than the traditional dual-encoder index due to the computation of dense cross-encoder scores matrix (see §2.1). We have successfully run our approach on up to 5 million items, but scaling to billions of items is an interesting direction for future work. Dual-encoder-based retrieve-and-erank baseline approaches can benefit from training the dual-encoder on multiple domains. It is not clear if data from multiple domains can be leveraged to improve performance of the proposed approach on a given target domain; although in any case, cross-encoders tend to be more robust to domain shift than using only dual-encoders for retrieval." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.786, + 0.267, + 0.801 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.813, + 0.49, + 0.893 + ], + "angle": 0, + "content": "In this paper we tackle the task of finding \\( k \\)-nearest neighbor items for a given query when query-items scores are computed using a black-box similarity function such as a cross-encoder model. The cross-encoder scoring function may have certain biases" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.885, + 0.133 + ], + "angle": 0, + "content": "and error tendencies, and it is unclear if our proposed method to approximate cross-encoder scores exacerbates or mitigates such biases." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.145, + 0.673, + 0.162 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.17, + 0.886, + 0.456 + ], + "angle": 0, + "content": "We thank members of UMass IESL for helpful discussions and feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the National Science Foundation under Grant No. NSF1763618, in part by the Chan Zuckerberg Initiative under the project \"Scientific Knowledge Base Construction\", in part by International Business Machines Corporation Cognitive Horizons Network agreement number W1668553, in part by Amazon Digital Services, and in part using high-performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative. Any opinions, findings, conclusions, and recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor(s)." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.482, + 0.61, + 0.497 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.506, + 0.886, + 0.584 + ], + "angle": 0, + "content": "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. MS-MARCO: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.595, + 0.887, + 0.688 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.699, + 0.887, + 0.753 + ], + "angle": 0, + "content": "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Realm: retrieval-augmented language model pre-training. In International Conference on Machine Learning." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.763, + 0.885, + 0.829 + ], + "angle": 0, + "content": "Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2020. Improving efficient neural ranking models with cross-architecture knowledge distillation. *ArXiv*, abs/2010.02666." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.84, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24." + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.506, + 0.887, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.472, + 0.919 + ], + "angle": 0, + "content": "For a 12-layer transformer model on an Nvidia 2080ti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8092" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.086, + 0.489, + 0.166 + ], + "angle": 0, + "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.174, + 0.49, + 0.24 + ], + "angle": 0, + "content": "Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient and effective passage search via contextualized late interaction over bert. In ACM SIGIR Conference on Research and Development in Information Retrieval." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.25, + 0.49, + 0.316 + ], + "angle": 0, + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations, ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.325, + 0.49, + 0.392 + ], + "angle": 0, + "content": "Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.401, + 0.489, + 0.441 + ], + "angle": 0, + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations, ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.45, + 0.489, + 0.529 + ], + "angle": 0, + "content": "Xinyin Ma, Yong Jiang, Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei Huang, and Weiming Lu. 2021. MuVER: Improving first-stage entity retrieval with multi-view entity representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.538, + 0.489, + 0.59 + ], + "angle": 0, + "content": "Michael W Mahoney and Petros Drineas. 2009. Cur matrix decompositions for improved data analysis. Proceedings of the National Academy of Sciences, 106:697-702." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.601, + 0.49, + 0.706 + ], + "angle": 0, + "content": "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.715, + 0.49, + 0.768 + ], + "angle": 0, + "content": "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval, 3(4)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.489, + 0.83 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.84, + 0.49, + 0.918 + ], + "angle": 0, + "content": "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.152 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.161, + 0.885, + 0.228 + ], + "angle": 0, + "content": "Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.236, + 0.885, + 0.303 + ], + "angle": 0, + "content": "Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.311, + 0.883, + 0.39 + ], + "angle": 0, + "content": "Nishant Yadav, Nicholas Monath, Rico Angell, Manzil Zaheer, and Andrew McCallum. 2022. Efficient nearest neighbor search for cross-encoder models using matrix factorization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.399, + 0.883, + 0.466 + ], + "angle": 0, + "content": "Wenzheng Zhang and Karl Stratos. 2021. Understanding hard negatives in noise contrastive estimation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.476, + 0.652, + 0.492 + ], + "angle": 0, + "content": "A Appendix A" + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.5, + 0.868, + 0.516 + ], + "angle": 0, + "content": "B Training and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.883, + 0.556 + ], + "angle": 0, + "content": "All the code for reproducing experiments is available at https://github.com/iesl/anncur." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.566, + 0.81, + 0.581 + ], + "angle": 0, + "content": "B.1 Training Cross-Encoder Models" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.587, + 0.884, + 0.681 + ], + "angle": 0, + "content": "In our experiments, we use [EMB]-CE, a cross-encoder model variant proposed by Yadav et al. (2022) that jointly encodes a query-item pair and computes the final score using dot-product of contextualized query and item embeddings extracted after joint encoding." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.69, + 0.711, + 0.703 + ], + "angle": 0, + "content": "B.1.1 ZESHEL Dataset" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.709, + 0.884, + 0.897 + ], + "angle": 0, + "content": "ZESHEL dataset is a zero-shot entity linking containing a set of 16 domains, each containing a disjoint set of items (entities). Each domain contains a set of queries (mention) paired with their ground-truth items (entities). For ZESHEL, we use the cross-encoder model checkpoint released by Yadav et al. (2022). The cross-encoder model was trained by first training a dual-encoder model on ZESHEL training data using hard negatives, and then training a cross-encoder model for the task of zero-shot entity-linking on all eight training domains using cross-entropy loss with ground-truth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.887, + 0.919 + ], + "angle": 0, + "content": "3https://huggingface.co/nishantyadav/emb_crossenc_zeshel" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8093" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.082, + 0.83, + 0.177 + ], + "angle": 0, + "content": "
DatasetDomain|I|(|Qtrain|, |Qtest|) SplitsTrain Query (Qtrain) Type
ZESHELYuGiOh10,031(100/3274), (500/2874), (2000/1374)Real Queries
ZESHELStarTrek34,430(100/4127), (500/3727), (2000/2227)Real Queries
ZESHELMilitary104,520(100/2300), (500/1900), (2000/0400)Real Queries
BEIRSciDocs25,657(1000/1000)Pseudo-Queries
BEIRHotpotQA5,233,329(1000/1000)Pseudo-Queries
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.186, + 0.885, + 0.286 + ], + "angle": 0, + "content": "Table 1: Statistics on the number of items \\((\\mathcal{I})\\) and the number of queries in train and test splits for each domain. The train-query \\((Q_{train})\\) split refers to queries used for distilling dual-encoder models or for indexing items using ADACUR and ANNCUR. For ZESHEL domains, we create train-test splits by splitting the queries in each domain uniformly at random and test with three different splits by putting 100, 500, or 2000 queries in train split. For BEIR domains, we use pseudo-queries released as part of the benchmark as train queries \\((Q_{train})\\) and run \\(k\\)-NN evaluation on test queries from the official test split (as per BEIR benchmark) of these domains. For HotpotQA, we use the first 1K queries out of a total of 7K test queries and we use all 1K test queries for SciDocs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.31, + 0.489, + 0.356 + ], + "angle": 0, + "content": "entity and negative entities mined using the dual-encoder. We refer readers to Yadav et al. (2022) for further details on cross-encoder training." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.357, + 0.489, + 0.466 + ], + "angle": 0, + "content": "We perform \\( k \\)-NN experiments on domains YuGiOh, StarTrek, and Military from ZESHEL of which only Military was part of the training data used to train the cross-encoder model and YuGiOh and StarTrek are part of the original ZESHEL test domains and the cross-encoder model was not trained on these domains." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.477, + 0.225, + 0.491 + ], + "angle": 0, + "content": "B.1.2 BEIR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.496, + 0.49, + 0.78 + ], + "angle": 0, + "content": "We follow the training setup used by Hofstätter et al. (2020). We first train three teacher cross-encoders initialized with albert-large-v2 (Lan et al., 2020), bert-large-whole-word-masking, and bert-base-uncased (Devlin et al., 2019), and compute soft labels on 40 million (query, positive item, negative item) triplets in MS-MARCO dataset (Bajaj et al., 2016). We then train our cross-encoder model parameterized using a 6-layer MINI-LM model (Wang et al., 2020) via distillation using average scores of the three teacher models as target signal and minimizing mean-square-error between predicted and target scores. We use training scripts available as part of sentence-transformer\\(^{4}\\) repository to train the cross-encoder model and use a dot-product based scoring mechanism for cross-encoders proposed by Yadav et al. (2022)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.792, + 0.409, + 0.807 + ], + "angle": 0, + "content": "B.2 Training Dual-Encoder Models" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.813, + 0.313, + 0.827 + ], + "angle": 0, + "content": "B.2.1 ZESHEL dataset" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.832, + 0.489, + 0.895 + ], + "angle": 0, + "content": "We report results for DE baselines as reported in Yadav et al. (2022). The DE models were initialized using bert-base-uncased and contain separate query and item encoders, thus containing a total of" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.31, + 0.883, + 0.356 + ], + "angle": 0, + "content": "\\(2 \\times 110M\\) parameters. We refer readers to Yadav et al. (2022) for details related to the training of all DE model variants on ZESHEL dataset." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.368, + 0.715, + 0.382 + ], + "angle": 0, + "content": "B.2.2 BEIR benchmark" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.387, + 0.884, + 0.498 + ], + "angle": 0, + "content": "For BEIR domains, we use a dual-encoder model released as part of sentence-transformer repository as \\(\\mathrm{DE}_{\\mathrm{BASE}}\\). This dual-encoder model was initialized using distillbert-base (Sanh et al., 2019) and trained on MS-MARCO dataset. This \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) is not trained on target domains SciDocs and HotpotQA used for running \\(k\\)-NN experiments." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.5, + 0.884, + 0.755 + ], + "angle": 0, + "content": "We finetune \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) via distillation on the target domain to get \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) model. Given a set of training queries \\(\\mathcal{Q}_{\\mathrm{train}}\\) from the target domain, we retrieve top-100 or top-1000 items for each query, score the items with the cross-encoder model, and train the dual-encoder by minimizing cross-entropy loss between predicted query-item scores (using DE) and target query-item scores (obtained using cross-encoder). Training a \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) with 1K queries and 100 or 1000 items per query takes around 2 hrs and 10 hrs respectively on an Nvidia RTX8000 GPU with 48GB memory. We train \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) for 10 epochs when using top-100 items per query and for 4 epochs when using top-1000 items per query using AdamW (Loshchilov and Hutter, 2019) optimizer with learning rate 1e-5." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.768, + 0.822, + 0.783 + ], + "angle": 0, + "content": "B.3 ANNCUR Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.79, + 0.882, + 0.884 + ], + "angle": 0, + "content": "For ANNCUR, we report results for the optimal split of cross-encoder call budget \\((\\mathcal{B}_{\\mathrm{CE}})\\) between scoring \\(k_{i}\\) anchor items followed by retrieving \\(\\mathcal{B}_{\\mathrm{CE}} - k_{i}\\) items for re-ranking. We experiment with \\(k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}\\). If the retrieved items contain a subset of anchor items for which exact" + }, + { + "type": "page_footnote", + "bbox": [ + 0.536, + 0.892, + 0.765, + 0.905 + ], + "angle": 0, + "content": "msmarco-distilroberta-base-v2:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.906, + 0.858, + 0.918 + ], + "angle": 0, + "content": "www.sbert.net/docs/pretrained-models/msmarco-v2.html" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.892, + 0.858, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.452, + 0.919 + ], + "angle": 0, + "content": "4https://github.com/UKPLab/sentence-transformers" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8094" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.085, + 0.491, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.204, + 0.495, + 0.241 + ], + "angle": 0, + "content": "(a) Domain \\(=\\) Military (100K items), \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\) . Each CE call takes \\(\\sim 7\\mathrm{ms}\\) on an Nvidia 2080ti GPU for a cross-encoder parameterized using 12-layered transformer model." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.084, + 0.88, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.204, + 0.884, + 0.241 + ], + "angle": 0, + "content": "(b) Domain=HotpotQA (5 Million items), \\( |\\mathcal{Q}_{\\mathrm{train}}| = 1000 \\). Each CE call takes \\(\\sim 2\\) ms on an Nvidia RTX8000 GPU for a cross-encoder parameterized using 6-layered transformer model." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.251, + 0.884, + 0.294 + ], + "angle": 0, + "content": "Figure 5: ADACUR inference latency versus number of rounds for two different domains. The primary bottleneck at inference time is the time taken to compute cross-encoder (CE) scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. See §B.5 for detailed discussion." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.32, + 0.489, + 0.399 + ], + "angle": 0, + "content": "cross-encoder score has already been computed, we retrieve more than \\(\\mathcal{B}_{\\mathrm{CE}} - k_{i}\\) items using approximate scores and compute exact cross-encoder scores for them until we have exhausted the entire cross-encoder call budget for the re-ranking step." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.413, + 0.426, + 0.428 + ], + "angle": 0, + "content": "B.4 ADACUR Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.435, + 0.489, + 0.53 + ], + "angle": 0, + "content": "For all our \\(k\\)-NN search experiments, we used Nvidia 2080ti GPUs with 12GB memory for domains YuGiOh (10K items), StarTrek (34K items), Military (100K items), and SciDocs (25K items), and we used Nvidia RTX8000 GPUs with 48GB memory for HotpotQA (5 million items)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.532, + 0.488, + 0.61 + ], + "angle": 0, + "content": "For HotpotQA, we restrict our \\(k\\)-NN search to top-10K items wrt \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) for ADACUR\\(_{\\mathrm{DE}_{\\mathrm{BASE}}}\\). For ZESHEL domains and SciDocs, we do not use any such heuristic and search over all the items in the corresponding domain." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.625, + 0.4, + 0.64 + ], + "angle": 0, + "content": "B.5 Time Complexity of ADACUR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.648, + 0.488, + 0.727 + ], + "angle": 0, + "content": "The offline indexing step for ADACUR takes \\(\\mathcal{O}(k_q|\\mathcal{I}|\\mathcal{C}_{f_\\theta})\\) time as it involves computing exact cross-encoder scores for all \\(|\\mathcal{I}|\\) items in the target domain against \\(k_q\\) anchor queries, and computing each cross-encoder score takes \\(\\mathcal{C}_{f_\\theta}\\) units of time." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.728, + 0.49, + 0.919 + ], + "angle": 0, + "content": "At test time, we are given a budget \\(\\mathcal{B}_{\\mathrm{CE}}\\) on the number of cross-encoder calls. Each one of the \\(N_{\\mathcal{R}}\\) rounds during inference with ADACUR involves approximating all item scores for the test query \\((q_{\\mathrm{test}})\\) followed by sampling the next batch of \\(k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}\\) items using the updated approximate scores. In the \\(j^{th}\\) round, the score approximation step involves computing the pseudo-inverse of a \\(k_{q} \\times jk_{s}\\) -dimensional matrix (line 12 in Algo. 1), which takes \\(\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s})\\) time, followed by a matrix multiplication step to compute updated approximate scores (line 13 in Algo. 1)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.317, + 0.884, + 0.435 + ], + "angle": 0, + "content": "which takes \\(\\mathcal{O}(\\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})\\) time. The time taken to update the approximate scores in each round is \\(\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s} + \\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})\\), and the time taken to compute cross-encoder scores for the next batch of \\(k_{s}\\) items is \\(\\mathcal{O}(k_s\\mathcal{C}_{f_\\theta})\\). Thus, the total inference latency for retrieving items over \\(N_{\\mathcal{R}}\\) rounds under a given budget of \\(\\mathcal{B}_{\\mathrm{CE}}\\) cross-encoder calls is" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.447, + 0.857, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {O} \\Big (\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(k _ {s} \\mathcal {C} _ {f _ {\\theta}} + \\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q}, j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q}, j k _ {s}, | \\mathcal {I} |}\\right) \\Big) \\\\ = \\mathcal {O} \\left(\\mathcal {B} _ {\\mathrm {C E}} \\mathcal {C} _ {f _ {\\theta}} + \\underbrace {\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(\\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q} , j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q} , j k _ {s} , | \\mathcal {I} |}\\right)} _ {\\text {O v e r h e a d o f A D A C U R}}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.572, + 0.884, + 0.856 + ], + "angle": 0, + "content": "Figure 5 shows the breakdown of ADACUR's inference latency in terms of time spent on computing cross-encoder scores, and the overhead of computing matrix inverse in line 12 and updating approximate scores by multiplying matrices in line 13 of Algorithm 1. Empirically, we observe that the primary bottleneck at inference time is the time taken to compute cross-encoder scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. The overhead is mostly dominated by computing pseudo-inverse (see line 12 in Algorithm 1) and this step is independent of the target domain size. The matrix multiplication step (line 13 in Algorithm 1) has a linear dependence on the number of items in the target domain but it is a negligible fraction of the overall running time as it can be significantly sped up using GPUs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.857, + 0.884, + 0.919 + ], + "angle": 0, + "content": "For ZESHEL domains, we use a cross-encoder parameterized using bert-base (Devlin et al., 2019), and observe that each cross-encoder call takes amortized time of \\(\\sim 7\\mathrm{ms}\\) on an Nvidia 2080ti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.52, + 0.941 + ], + "angle": 0, + "content": "8095" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.179 + ], + "angle": 0, + "content": "GPU when the scores are computed in batches of size 50. Computing each cross-encoder score sequentially i.e. with batch-size \\(= 1\\) takes \\(\\sim 13\\mathrm{ms}\\) per score. We did not observe any further reduction in amortized time to compute each score when increasing the batch size beyond 50." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.181, + 0.49, + 0.292 + ], + "angle": 0, + "content": "The amortized time per cross-encoder call is approximately 6ms and 2ms for SciDocs and HotpotQA respectively when using batch size=50 and MINI-LM-based (Wang et al., 2020) cross-encoder. The difference in time per cross-encoder score for SciDocs and HotpotQA is due to the difference in average query-item pair sequence length." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.305, + 0.43, + 0.323 + ], + "angle": 0, + "content": "C Additional Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.332, + 0.4, + 0.347 + ], + "angle": 0, + "content": "C.1 TopK vs SoftMax for ADACUR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.354, + 0.49, + 0.48 + ], + "angle": 0, + "content": "Figure 6 shows Top-\\(k\\)-Recall for ADACUR on domain=YuGiOh, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\), when using TopK and SoftMax strategies for sampling items based on approximate scores (see §2.2 for details). TopK sampling strategy which greedily picks top-\\(k\\) items based on approximate scores results in superior recall as compared to sampling items using softmax over approximate scores." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.495, + 0.484, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.684, + 0.489, + 0.727 + ], + "angle": 0, + "content": "Figure 6: Top-\\(k\\)-Recall for ADACUR on YuGi0h, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\) for different strategies for sampling items based on approximate scores described in §2.2." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.74, + 0.442, + 0.756 + ], + "angle": 0, + "content": "C.2 Anchor Item Sampling with Oracle" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.762, + 0.49, + 0.919 + ], + "angle": 0, + "content": "ADACUR performs retrieval over multiple rounds using approximate cross-encoder scores and uses the items retrieved based on the approximate scores as anchor items to improve the approximation and hence retrieval in subsequent rounds. In this section, we run experiments where the anchor item sampling method has oracle access to exact cross-encoder scores of all items for the given test query to better understand the effect of anchor items on the approximation of cross-encoder scores and" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.147 + ], + "angle": 0, + "content": "hence subsequent retrieval based on the approximate scores. We experiment with the following strategies for sampling \\( k_{i} \\) anchor items for a given test query:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.149, + 0.883, + 0.228 + ], + "angle": 0, + "content": "- TopK\\(_{k_m, \\epsilon}^{\\mathcal{O}}\\): Mask out top-\\(k_m\\) items wrt exact cross-encoder scores and select \\(k_i\\) anchor items by greedily picking \\((1 - \\epsilon) k_i\\) items starting from rank \\(k_m + 1\\), and sample remaining \\(\\epsilon k_i\\) anchor items uniformly at random." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.233, + 0.885, + 0.327 + ], + "angle": 0, + "content": "- SoftMax\\(\\mathcal{O}_{k_m,\\epsilon}\\): Mask out top-\\(k_m\\) items wrt exact cross-encoder scores and select \\(k_i\\) anchor items by sampling \\((1 - \\epsilon)k_i\\) anchor items using softmax over exact cross-encoder scores, and sample remaining \\(\\epsilon k_i\\) anchor items uniformly at random." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.149, + 0.885, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.33, + 0.884, + 0.44 + ], + "angle": 0, + "content": "For a given test-time cross-encoder call budget \\(\\mathcal{B}_{\\mathrm{CE}}\\), we select \\(k_{i}\\) anchor items, compute approximate cross-encoder scores using the chosen anchor items, and then retrieve \\(\\mathcal{B}_{\\mathrm{CE}} - k_{i}\\) items based on the approximate scores. We experiment with \\(k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}\\) and report results for the best budget split." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.454, + 0.866, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.593, + 0.884, + 0.704 + ], + "angle": 0, + "content": "Figure 7: Average approximation error for CUR matrix factorization on test-queries for domain=YuGiOh and \\( |\\mathcal{Q}_{\\mathrm{train}}| = 500 \\) when choosing \\( k_{i} = 50 \\) anchor items uniformly at random (ANNCUR), using oracle strategies from §C.2 and for ADACUR when sampling anchor items over five rounds. Approximation error is computed as the average of absolute difference between approximate and exact item scores." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.726, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Figure 8a shows Top- \\(k\\) -Recall of anchor item sampling strategies \\(\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}\\) and SoftMax for \\(k_{m}\\in \\{0,k\\}\\) ,domain \\(\\equiv\\) YuGiOh. Sampling strategies SoftMax and TopK,0, which mask out top-k items, perform significantly worse than SoftMax0,0 and TopK,0 respectively when searching for \\(k =\\) 1, 10 nearest neighbors. This indicates that the significant improvement in Top-1-Recall and Top-10Recall for TopK,and SoftMax sampling strategies can be attributed to the presence of top- \\(k\\) items in the anchor item set. This is because CUR matrix" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8096" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.085, + 0.485, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.233, + 0.468, + 0.247 + ], + "angle": 0, + "content": "(a) Sampling with and without ground-truth top-\\(k\\) items" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.25, + 0.482, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.367, + 0.486, + 0.392 + ], + "angle": 0, + "content": "(b) Sampling using TopK strategy with exact CE scores while varying \\(\\epsilon\\), the fraction of items sampled uniformly at random" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.395, + 0.482, + 0.506 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.512, + 0.486, + 0.537 + ], + "angle": 0, + "content": "(c) Sampling using SoftMax of exact CE scores while varying \\(\\epsilon\\), the fraction of items sampled uniformly at random" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.548, + 0.49, + 0.605 + ], + "angle": 0, + "content": "Figure 8: Top-\\(k\\)-Recall for ADACUR, ANNCUR, and oracle sampling strategies (§C.2) that have oracle access to exact cross-encoder scores for all items for domain=YuGiOh, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.614, + 0.49, + 0.773 + ], + "angle": 0, + "content": "factorization which is used to compute the approximate scores incurs negligible approximation error on anchor items, and hence on top- \\(k\\) items when these items are part of the anchor set as shown in figures 7 and 9. For \\(\\mathsf{TopK}_{k,0}^{\\mathcal{O}}\\) and SoftMax\\(_{k,0}\\) sampling strategies, since the top- \\(k\\) items are not part of the anchor set, CUR incurs a much higher approximation error for the top- \\(k\\) items (see examples in Figures 9a and 9b), thus resulting in poor Top\\(k\\)-Recall as shown in Figure 8a." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.788, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Effect of diversity in anchor items Figure 8a shows that sampling items based on softmax of exact cross-encoder scores \\((\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}})\\) performs better than greedily picking top-scoring items \\((\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}})\\), for both \\(k_{m} = 0,k\\). The reason behind \\(\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}}\\) performing better than \\(\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}}\\) is that sampling based on softmax of exact scores yields an anchor set with a more diverse" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.081, + 0.689, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.204, + 0.688, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.318, + 0.688, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.432, + 0.667, + 0.448 + ], + "angle": 0, + "content": "(e) \\(\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}},\\epsilon = 0.75\\)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.081, + 0.881, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.204, + 0.881, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.318, + 0.881, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.432, + 0.868, + 0.448 + ], + "angle": 0, + "content": "(f) SoftMax0,ε, ε = 0.75" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.459, + 0.884, + 0.57 + ], + "angle": 0, + "content": "Figure 9: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGiOh, when choosing \\( k_{i} = 50 \\) anchor items using oracle strategies from §C.2 and \\( |\\mathcal{Q}_{\\mathrm{train}}| = 500 \\). Top-\\( k \\) for \\( k = 1,10,100 \\) wrt exact cross-encoder scores are annotated with text along with vertical lines, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.585, + 0.884, + 0.791 + ], + "angle": 0, + "content": "score distribution whereas greedily selecting topscoring items using exact scores results in an anchor set with items having similar cross-encoder scores. However, as shown in Figures 9c and 9d, both of these sampling strategies can result in overestimating scores for all items, even the irrelevant ones (i.e. items beyond top- \\(k\\) items) due to insufficient representation of the irrelevant items in the anchor set. Thus retrieving based on approximated scores may struggle to retrieve relevant \\(k\\)-NN items, especially for larger values of \\(k\\) such as \\(k = 100\\) when the anchor items are chosen using oracle strategies such as \\(\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.793, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Figures 9e and 9f, where \\(\\epsilon = 75\\%\\) of 50 items are sampled uniformly at random, show that overestimating scores of irrelevant items can be avoided by sampling a fraction of anchor items uniformly at random to increase the diversity of the anchor item set. As shown in Figures 8b and 8c, Top-\\(k\\)-Recall for both SoftMax\\(_{0,\\epsilon}\\) and TopK\\(_{0,\\epsilon}\\) generally improves with an increase in \\(\\epsilon\\), the fraction of ran" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.52, + 0.941 + ], + "angle": 0, + "content": "8097" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.29 + ], + "angle": 0, + "content": "dom items in the anchor set, due to increased diversity in the anchor item set. Since \\(\\mathrm{SoftMax}_{0,\\epsilon}^{\\mathcal{O}}\\) already samples a diverse set of anchor items, increasing \\(\\epsilon\\) yields only marginal improvement while for \\(\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}}\\), increasing \\(\\epsilon\\) yields significant improvements due to increased diversity of the anchor set. A small drop in performance is observed for larger values of \\(\\epsilon\\) as increasing \\(\\epsilon\\) beyond a threshold results in some of the top-\\(k\\) items to be excluded from the anchor item set. This results in a poorer approximation of scores for the missing top-\\(k\\) items and hence poor retrieval recall as the retrieval is done using the approximate scores." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.292, + 0.49, + 0.528 + ], + "angle": 0, + "content": "Finally, the optimal strategy for choosing the set of anchor items is the one that strikes a balance between selecting anchor items with diverse cross-encoder scores and greedily picking top-\\(k\\) items. Our proposed strategy ADACUR improves over ANNCUR as greedily picking top-scoring items according to approximate scores to expand set of anchor items increases the likelihood of picking ground-truth \\(k\\)-NN items to be part of the anchor set, with this likelihood improving after each round with improvement in the score approximation, and ADACUR achieves diversity in the anchor items as a result of sampling items uniformly at random in the first round and due to error in the approximate scores, as shown in Figure 11." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.544, + 0.469, + 0.56 + ], + "angle": 0, + "content": "C.3 Comparison with Multi-Vector Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.567, + 0.489, + 0.678 + ], + "angle": 0, + "content": "Multi-vector models (Khattab and Zaharia, 2020; Ma et al., 2021) produce multiple embeddings for each query and item. For a given query \\( q \\) and item \\( i \\), the query-item score is computed using simple functions such as average similarity or sum-of-maximum similarities between the set of embeddings for query \\( q \\) and item \\( i \\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.68, + 0.489, + 0.87 + ], + "angle": 0, + "content": "Figure 10 shows Top- \\(k\\)-Recall for \\(\\mathrm{DE}_{\\mathrm{BASE}}\\), \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\), \\(\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}\\), and MUVER (Ma et al., 2021), a recent multi-vector model trained on ZESHEL dataset. For MUVER, we use the pretrained checkpoint released by Ma et al. (2021) with the view-merging inference strategy as described in Ma et al. (2021). While MUVER can be more accurate than \\(\\mathrm{DE}_{\\mathrm{BASE}}\\), \\(\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}\\) obtained by finetuning \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) model on the target domain outperforms MUVER and our proposed method \\(\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}\\) yields the best Top- \\(k\\)-Recall versus inference cost trade-offs for all values of \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We would also like to note that while multi-vector models such as MUVER can be more accurate than single-embedding models such as \\(\\mathrm{DE}_{\\mathrm{BASE}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.085, + 0.876, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.254, + 0.731, + 0.266 + ], + "angle": 0, + "content": "(a) YuGiOh" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.877, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.411, + 0.739, + 0.424 + ], + "angle": 0, + "content": "(b) StarTrek" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.434, + 0.885, + 0.478 + ], + "angle": 0, + "content": "Figure 10: Top-\\(k\\)-Recall for \\(\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}\\) and baselines including multi-vector models on ZESHEL domains, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 2000\\). See §C.3 for discussion." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.486, + 0.884, + 0.565 + ], + "angle": 0, + "content": "such multi-vector models incur significant memory overhead for storing query/item embeddings. For instance, using 15 embeddings per item with 768-dimensional embeddings would take around 250GB space for 5 million items for HotpotQA." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.575, + 0.774, + 0.59 + ], + "angle": 0, + "content": "C.4 Results for TF-IDF baseline" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.596, + 0.882, + 0.658 + ], + "angle": 0, + "content": "TF-IDF: All queries and items are embedded using a TF-IDF vectorizer trained on item descriptions and top-\\(k\\) items are retrieved using the dot-product of sparse query and item embeddings." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.659, + 0.884, + 0.911 + ], + "angle": 0, + "content": "For domains in ZESHEL, we report results for TF-IDF baseline, for ANNCUR when anchor items are chosen using TF-IDF baseline (ANNCUR\\(_{TF-IDF}\\)), and for ADACUR when the first batch of anchor items is chosen using TF-IDF baseline (ADACUR\\(_{TF-IDF}\\)). Figures 13, 14, and 15 show Top-\\(k\\)-Recall for domains YuGiOh, StarTrek, and Military respectively for \\(Q_{train} \\in \\{100, 500, 2000\\}\\). For each baseline retrieval method, ADACUR always performs better than ANNCUR which in turn generally performs better than merely re-ranking items retrieved using the corresponding baseline retrieval method. In most cases, Top-\\(k\\)-Recall for ADACUR\\(_{DE_BASE}\\) > ANNCUR\\(_{DE_BASE}\\) > DE_BASE, and ADACUR\\(_{TF-IDF}\\) > ANNCUR\\(_{TF-IDF}\\) > TF-IDF." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8098" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.307, + 0.082, + 0.692, + 0.113 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.129, + 0.348, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.256, + 0.301, + 0.28 + ], + "angle": 0, + "content": "(a-1) ADACUR Round 1. Total 10 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.129, + 0.614, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.256, + 0.568, + 0.279 + ], + "angle": 0, + "content": "(a-2) ADACUR Round 2. Total 20 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.129, + 0.882, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.256, + 0.834, + 0.279 + ], + "angle": 0, + "content": "(a-3) ADACUR Round 3. Total 30 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.282, + 0.346, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.409, + 0.301, + 0.432 + ], + "angle": 0, + "content": "(a-4) ADACUR Round 4. Total 40 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.282, + 0.611, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.409, + 0.565, + 0.432 + ], + "angle": 0, + "content": "(a-5) ADACUR Round 5. Total 50 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.282, + 0.875, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.409, + 0.88, + 0.433 + ], + "angle": 0, + "content": "(a-6) ANNCUR- Sampling all 50 anchor items uniformly at random." + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.439, + 0.847, + 0.452 + ], + "angle": 0, + "content": "(a) Sampling 50 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.454, + 0.348, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.581, + 0.301, + 0.604 + ], + "angle": 0, + "content": "(b-1) ADACUR Round 1. Total 40 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.454, + 0.613, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.58, + 0.568, + 0.603 + ], + "angle": 0, + "content": "(b-2) ADACUR Round 2. Total 80 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.454, + 0.88, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.581, + 0.838, + 0.603 + ], + "angle": 0, + "content": "(b-3) ADACUR Round 3. Total 120 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.606, + 0.347, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.733, + 0.305, + 0.756 + ], + "angle": 0, + "content": "(b-4) ADACUR Round 4. Total 160 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.383, + 0.606, + 0.611, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.733, + 0.569, + 0.756 + ], + "angle": 0, + "content": "(b-5) ADACUR Round 5. Total 200 anchor items." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.606, + 0.875, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.733, + 0.88, + 0.757 + ], + "angle": 0, + "content": "(b-6) ANNCUR- Sampling all 200 anchor items uniformly at random." + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.763, + 0.851, + 0.777 + ], + "angle": 0, + "content": "(b) Sampling 200 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random)." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.787, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Figure 11: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGi0h, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 500\\) when choosing \\(k_{i} = 50\\) and 200 anchor items with ADACUR over five rounds, and uniformly at random with ANNCUR. Top-\\(k\\) for \\(k = 1,10,100\\) wrt exact cross-encoder scores are annotated with text, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue. With ADACUR, the first batch containing anchor items in Figure 11a-1 and 11b-1 is chosen uniformly at random and in subsequent rounds, items with highest approximate scores are chosen. Note that the approximation error for top-scoring items improves significantly when the 50 anchor items are chosen adaptively (see Figure 11a-5) with the improvement being much more significant than merely increasing the number of anchor items sampled uniformly at random from 50 in Figure 11a-6 to 200 in Figure 11b-6." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.276, + 0.298, + 0.725, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.326, + 0.882, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.503, + 0.537, + 0.514 + ], + "angle": 0, + "content": "(a) SciDocs" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.53, + 0.882, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.683, + 0.542, + 0.695 + ], + "angle": 0, + "content": "(b) HotpotQA" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.706, + 0.882, + 0.721 + ], + "angle": 0, + "content": "Figure 12: Top-\\(k\\)-Recall for ADACUR (using ten rounds) and baselines for SciDocs and HotpotQA, \\(|\\mathcal{Q}_{\\mathrm{train}}| = 1000\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "8100" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.184, + 0.395, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.185, + 0.86, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.214, + 0.738, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.25, + 0.408, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.251, + 0.673, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.251, + 0.88, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.38, + 0.724, + 0.394 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.401, + 0.652, + 0.415 + ], + "angle": 0, + "content": "(a) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 100 \\)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.432, + 0.408, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.432, + 0.674, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.432, + 0.88, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.56, + 0.724, + 0.575 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.582, + 0.652, + 0.596 + ], + "angle": 0, + "content": "(b) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 500 \\)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.613, + 0.408, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.613, + 0.674, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.613, + 0.88, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.742, + 0.724, + 0.756 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.763, + 0.655, + 0.777 + ], + "angle": 0, + "content": "(c) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 2000 \\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.787, + 0.882, + 0.816 + ], + "angle": 0, + "content": "Figure 13: Top-\\(k\\)-Recall for ADACUR (using five rounds) and baselines for domain=YuGiOh. Each subfigure corresponds to a different value of the number of train/anchor queries \\((|\\mathcal{Q}_{\\text {train }}|)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "8101" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.185, + 0.395, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.185, + 0.86, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.214, + 0.738, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.25, + 0.41, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.25, + 0.676, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.251, + 0.881, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.38, + 0.724, + 0.394 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.401, + 0.652, + 0.415 + ], + "angle": 0, + "content": "(a) Number of train/anchor queries \\(|\\mathcal{Q}_{\\mathrm{train}}| = 100\\)" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.432, + 0.41, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.432, + 0.676, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.432, + 0.881, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.56, + 0.724, + 0.575 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.582, + 0.652, + 0.596 + ], + "angle": 0, + "content": "(b) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 500 \\)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.613, + 0.41, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.613, + 0.676, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.613, + 0.881, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.742, + 0.724, + 0.756 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.763, + 0.655, + 0.777 + ], + "angle": 0, + "content": "(c) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 2000 \\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.787, + 0.883, + 0.817 + ], + "angle": 0, + "content": "Figure 14: Top-\\(k\\)-Recall for ADACUR (using five rounds) and baselines for domain=StarTrek. Each subfigure corresponds to a different value of the number of train/anchor queries \\((|\\mathcal{Q}_{\\mathrm{train}}|)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "8102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.172, + 0.394, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.172, + 0.86, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.201, + 0.38, + 0.213 + ], + "angle": 0, + "content": "Proposed Approaches" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.201, + 0.426, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.438, + 0.201, + 0.481, + 0.212 + ], + "angle": 0, + "content": "DACUR" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.201, + 0.521, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.525, + 0.201, + 0.549, + 0.212 + ], + "angle": 0, + "content": "ADA" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.201, + 0.574, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.201, + 0.602, + 0.212 + ], + "angle": 0, + "content": "R" + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.201, + 0.64, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.201, + 0.67, + 0.212 + ], + "angle": 0, + "content": "AD" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.201, + 0.703, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.201, + 0.731, + 0.212 + ], + "angle": 0, + "content": "\"" + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.201, + 0.768, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.201, + 0.805, + 0.212 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.201, + 0.856, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.237, + 0.408, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.238, + 0.675, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.238, + 0.88, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.366, + 0.724, + 0.379 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.387, + 0.652, + 0.401 + ], + "angle": 0, + "content": "(a) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 100 \\)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.418, + 0.408, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.418, + 0.675, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.418, + 0.88, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.547, + 0.724, + 0.561 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.568, + 0.652, + 0.582 + ], + "angle": 0, + "content": "(b) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 500 \\)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.599, + 0.408, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.599, + 0.675, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.599, + 0.88, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.383, + 0.728, + 0.724, + 0.741 + ], + "angle": 0, + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.749, + 0.655, + 0.763 + ], + "angle": 0, + "content": "(c) Number of train/anchor queries \\( |\\mathcal{Q}_{\\mathrm{train}}| = 2000 \\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.773, + 0.885, + 0.83 + ], + "angle": 0, + "content": "Figure 15: Top-\\(k\\)-Recall for ADACUR (using five rounds) and baselines for domain=Military. Each subfigure corresponds to a different value of the number of train/anchor queries \\((|\\mathcal{Q}_{train}|)\\). Note that \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) has high Top-1-Recall values as domain=Military is included in the set of train domains in ZESHEL which are used to train both \\(\\mathrm{DE}_{\\mathrm{BASE}}\\) and the cross-encoder model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.52, + 0.941 + ], + "angle": 0, + "content": "8103" + } + ] +] \ No newline at end of file diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_origin.pdf b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..01ccf2f7d1da72f7fafb06706882696530f97e33 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/06755a7b-5456-4010-a183-a7ae8a6d5cd1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e79b5dd5e20f94f5ef978eaf5483292b78aaf9a16feae22fd19492a3391cf965 +size 9985824 diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/full.md b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..01b45c44b5fc2d165c84245de4216430a30f7b69 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/full.md @@ -0,0 +1,526 @@ +# Efficient $k$ -NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition + +Nishant Yadav†, Nicholas Monath‡, Manzil Zaheer§, and Andrew McCallum† + +† University of Massachusetts Amherst ‡Google Research + +{nishantyavad, mccallum}@cs.umass.edu {nmonath,manzilzaheer}@google.com + +# Abstract + +Cross-encoder models, which jointly encode and score a query-item pair, are prohibitively expensive for direct $k$ -nearest neighbor ( $k$ -NN) search. Consequently, $k$ -NN search typically employs a fast approximate retrieval (e.g. using BM25 or dual-encoder vectors), followed by reranking with a cross-encoder; however, the retrieval approximation often has detrimental recall regret. This problem is tackled by ANNCUR (Yadav et al., 2022), a recent work that employs a cross-encoder only, making search efficient using a relatively small number of anchor items, and a CUR matrix factorization. While ANNCUR's one-time selection of anchors tends to approximate the cross-encoder distances on average, doing so forfeits the capacity to accurately estimate distances to items near the query, leading to regret in the crucial end-task: recall of top- $k$ items. In this paper, we propose ADACUR, a method that adaptively, iteratively, and efficiently minimizes the approximation error for the practically important top- $k$ neighbors. It does so by iteratively performing $k$ -NN search using the anchors available so far, then adding these retrieved nearest neighbors to the anchor set for the next round. Empirically, on multiple datasets, in comparison to previous traditional and state-of-the-art methods such as ANNCUR and dual-encoder-based retrieval-and-erank, our proposed approach ADACUR consistently reduces recall error—by up to $70\%$ on the important $k = 1$ setting—while using no more compute than its competitors. + +# 1 Introduction + +$k$ -nearest neighbor ( $k$ -NN) search is a core subroutine of a variety of tasks in NLP such as entity linking (Wu et al., 2020), passage retrieval for QA (Karpukhin et al., 2020), and more generally, in retrieval-augmented machine learning models (Guu et al., 2020; Izacard et al., 2023). For many of these applications, the state-of-the-art similarity function is a cross-encoder that directly outputs a scalar similarity score after jointly encoding a + +![](images/d7841db05b54c8187f64cb341ff50b33be4113076ed8294d41719b564dc366f7.jpg) + +![](images/bd6f11d141d922b6e9cae84b134ed0f204a697693f8ce9a7c1b7873bf21364b7.jpg) +(a) Uniformly at random + +![](images/d01a79b72270176b45e252c19a0483397eee9525be8a52a8670ce78aeb31fc8c.jpg) +(b) Biased sampling +Figure 1: Exact versus approximate cross-encoder scores (computed using ANNCUR) of all items for a test-query in domain=YuGi0h. ANNCUR incurs high approximation error on $k$ -NN items wrt exact scores when using 50 anchor items sampled uniformly at random (Fig. 1a). In contrast, sampling 50 anchor items with probability proportional to exact cross-encoder scores (Fig. 1b) significantly improves approximation of top-scoring items. + +given query-item pair. However, computing a single query-item score using a cross-encoder requires a forward pass of the model which can be computationally expensive as cross-encoders are typically parameterized using deep neural models such as transformers (Vaswani et al., 2017). For this reason, $k$ -NN search with a cross-encoder typically involves retrieving candidate items using additional models such as dual-encoders or BM25 (Robertson et al., 2009), followed by re-ranking items using the cross-encoder (Logeswaran et al., 2019; Zhang and Stratos, 2021; Qu et al., 2021). However, the accuracy of such retrieve-and-erank approaches is upperbound by the recall of first-stage retrieval and may require computationally expensive distillation-based training of dual-encoders to improve recall. + +Recent work by Yadav et al. (2022) proposed ANNCUR, a CUR factorization (Mahoney and Drineas, 2009) based method, that approximates cross-encoder using dot-product of latent query and item embeddings and performs $k$ -NN retrieval using approximate scores followed by optionally re-ranking retrieved items using exact cross + +encoder scores. The latent item embeddings are computed by comparing each item against a set of anchor queries, and the latent query embedding is computed using the query's cross-encoder scores against a fixed set of anchor items. As shown in Figure 1, when ANNCUR selects the anchor items uniformly at random (Fig 1a), it incurs higher approximation error on the top-scoring items than the rest of the items, resulting in poor $k$ -NN recall, and including some $k$ -NN items as part of anchor items (Fig. 1b) can significantly improve approximation error for top-scoring items. + +In this work, we propose ADACUR, a search strategy that improves $k$ -NN search recall by improving the approximation of top-scoring items. ADACUR performs retrieval over multiple rounds, retrieving the first batch of items either uniformly at random or using heuristic or auxiliary models such as dual-encoder or BM25 to get a first coarse approximation of item scores for the test query. In subsequent rounds, it alternates between a) performing retrieval using approximate scores and scoring retrieved items using cross-encoder and b) using all retrieved items as anchor items to improve the approximation and hence retrieval of relevant items in the subsequent rounds. Our proposed approach provides significant improvements in $k$ -NN search recall over ANNCUR and dual-encoder based retrieve-and-erank approaches when performing $k$ -NN search with cross-encoder models trained for the task of entity linking and information retrieval. + +# 2 Proposed Method: ADACUR + +Task Given a scoring function $f_{\theta} \colon \mathcal{Q} \times \mathcal{I} \to \mathbb{R}$ that maps a query-item pair to a scalar score, and a query $q \in \mathcal{Q}$ , the $k$ -nearest neighbor search task is to retrieve top- $k$ scoring items from a fixed item set $\mathcal{I}$ according to the given scoring function $f_{\theta}$ . + +# 2.1 ADACUR: Offline Indexing of Items + +The indexing step of ADACUR involves using the cross-encoder model $(f_{\theta})$ to score each item against a fixed set of $k_q$ anchor/train queries $(Q_{train})$ , to get score matrix $R_{\mathrm{anc}}$ + +$$ +R _ {\mathrm {a n c}} (q, i) = f _ {\theta} (q, i), \forall (q, i) \in \mathcal {Q} _ {\text {t r a i n}} \times \mathcal {I} +$$ + +Each column of $E^{\mathcal{I}} \coloneqq R_{\mathrm{anc}} \in \mathbb{R}^{k_q \times |\mathcal{I}|}$ corresponds to a $k_q$ -dimensional latent item embedding. + +# 2.2 ADACUR: Test-time inference + +The baseline method ANNCUR computes the latent test-query embedding $e_{q_{\mathrm{test}}}$ using $C_{q_{\mathrm{test}}}$ , a + +Algorithm 1 ADACUR $k$ -NN Search +1: Input: $(q_{\mathrm{test}}, R_{\mathrm{anc}}, N_{\mathcal{R}}, \mathcal{B}_{\mathrm{CE}}, \mathcal{A})$ +2: $q_{\mathrm{test}}$ : Test query +3: $R_{\mathrm{anc}}$ : Matrix containing CE scores between $\mathcal{Q}_{\mathrm{train}}$ and $\mathcal{I}$ +4: $\mathcal{B}_{\mathrm{CE}}$ : Total cross-encoder (CE) call budget. +5: $\mathcal{A}$ : Algorithm to use for selecting (anchor) items. +6: $N_{\mathcal{R}}$ : Number of iterative search rounds +7: Output: $\hat{S}$ : Approximate scores of $q_{\mathrm{test}}$ with all items, $\mathcal{I}_{\mathrm{anc}}$ : Retrieved (anchor) items with CE scores in $C_{\mathrm{test}}$ . +8: $k_s \gets \mathcal{B}_{\mathrm{CE}} / N_{\mathcal{R}} \triangleright$ Num. of items to sample per round +9: $\mathcal{I}_{\mathrm{anc}} \gets \text{INIT}(\mathcal{I}, k_s)$ $\triangleright$ Initial set of anchor items +10: $C_{\mathrm{test}} \gets [f_\theta(q_{\mathrm{test}}, i)]_{i \in \mathcal{I}_{\mathrm{anc}}} \triangleright$ CE scores of $\mathcal{I}_{\mathrm{anc}}$ for $q_{\mathrm{test}}$ +11: for $j \gets 2$ to $N_{\mathcal{R}}$ do +12: $U \gets R_{\mathrm{anc}}[\mathcal{I}_{\mathrm{anc}}]^{\dagger}$ $\triangleright$ $U \in \mathbb{R}^{|I_{\mathrm{anc}}| \times |\mathcal{Q}_{\mathrm{train}}|}$ +13: $\hat{S}^{(j)} \gets C_{\mathrm{test}} \times U \times R_{\mathrm{anc}} \triangleright$ Update approx. scores +14: $\mathcal{I}_{\mathrm{anc}}^{(j)} \gets \text{SAMPLEITEMS}(\mathcal{A}, k_s, \mathcal{I}_{\mathrm{anc}}, \hat{S}^{(j)})$ +15: $\mathcal{I}_{\mathrm{anc}} \gets \mathcal{I}_{\mathrm{anc}} \cup \mathcal{I}_{\mathrm{anc}}^{(j)}$ +16: $C_{\mathrm{test}} \gets C_{\mathrm{test}} \oplus [f_\theta(q_{\mathrm{test}}, i)]_{i \in \mathcal{I}_{\mathrm{anc}}^{(j)}} \triangleright$ Update $C_{\mathrm{test}}$ +17: $U \gets R_{\mathrm{anc}}[\mathcal{I}_{\mathrm{anc}}]^{\dagger}$ $\triangleright$ $U \in \mathbb{R}^{|I_{\mathrm{anc}}| \times |\mathcal{Q}_{\mathrm{train}}|}$ +18: $\hat{S} \gets C_{\mathrm{test}} \times U \times R_{\mathrm{anc}} \triangleright$ Compute approx. scores +19: return $\hat{S}, \mathcal{I}_{\mathrm{anc}}, C_{\mathrm{test}}$ + +Algorithm 2 SAMPLEITEMS +1: Input: $(\mathcal{A}, k_s, \mathcal{I}_{\text{mask}}, S)$ +2: $\mathcal{A}$ : Algorithm for sampling items +3: $k_s$ : Number of items to sample +4: $\mathcal{I}_{\text{mask}}$ : Set of items to mask +5: $S$ : (Approximate) Scores for all items +6: Output: $\mathcal{I}_{\text{select}}$ : Set of sampled $k_s$ items +7: $\bar{S} \gets \text{SOFTMAX}(S)$ +8: $\bar{S}[\mathcal{I}_{\text{mask}}] \gets 0$ > Mask items in $\mathcal{I}_{\text{mask}}$ +9: if $\mathcal{A} = \text{TopK}$ then +10: $\mathcal{I}_{\text{select}} \gets \text{TOPK}(\bar{S}, k_s)$ +11: else if $\mathcal{A} = \text{SoftMax}$ then +12: $\mathcal{I}_{\text{select}} \gets k_s$ items sampled using $\bar{S}$ +13: else if $\mathcal{A} = \text{Random}$ then +14: $\mathcal{I}_{\text{select}} \gets k_s$ items uniformly sampled from $\mathcal{I} \setminus \mathcal{I}_{\text{mask}}$ +15: return $\mathcal{I}_{\text{select}}$ + +$|\mathcal{I}_{\mathrm{anc}}|$ -dimensional vector containing cross-encoder scores of $q_{\mathrm{test}}$ with a set of anchor items $(\mathcal{I}_{\mathrm{anc}})$ as: + +$$ +C _ {q _ {\text {t e s t}}} = \left[ f _ {\theta} \left(q _ {\text {t e s t}}, i\right) \right] _ {i \in \mathcal {I} _ {\text {a n c}}} +$$ + +$$ +e _ {q _ {\text {t e s t}}} = C _ {q _ {\text {t e s t}}} \times U +$$ + +where $U\in \mathbb{R}^{|\mathcal{I}_{\mathrm{anc}}|\times |\mathcal{Q}_{\mathrm{train}}|}$ is the pseudo-inverse of $R_{\mathrm{anc}}[\mathcal{I}_{\mathrm{anc}}]$ , the subset of columns of $R_{\mathrm{anc}}$ corresponding to (anchor) items $\mathcal{I}_{\mathrm{anc}}$ . Finally, ANCUR approximates the score for a query-item pair $(q_{\mathrm{test}},i)$ using dot-product of the query embedding $e_{q_{\mathrm{test}}}$ and item embedding $E^{\mathcal{I}}[:,i]$ as + +$$ +\hat {f} _ {\theta} (q _ {\mathrm {t e s t}}, i) = e _ {q _ {\mathrm {t e s t}}} ^ {\top} E ^ {\mathcal {I}} [:, i ] +$$ + +The main bottleneck at test-time inference is the number of items scored using the cross-encoder for the given test-query. For a given budget of cross-encoder calls, ANNCUR splits the budget $(\mathcal{B}_{\mathrm{CE}})$ + +into two parts - it uses $k_{i}$ cross-encoder calls to compare against anchor items (chosen uniformly at random or using heuristic methods) and retrieves $\mathcal{B}_{\mathrm{CE}} - k_{i}$ items using approximate scores and re-ranks them using exact cross-encoder scores. + +In contrast, our proposed approach ADACUR uses the cross-encoder call budget to adaptively retrieve and score items over $N_{\mathcal{R}}$ rounds, repurposing the retrieved items as anchor items as shown in Algorithm 1. ADACUR begins by sampling the first batch of $k_{s} = \mathcal{B}_{\mathrm{CE}} / N_{\mathcal{R}}$ (anchor) items uniformly at random. The first batch of items can also be selected using baseline retrieval methods such as BM25 and dual-encoders. In the $j^{\mathrm{th}}$ round, the items retrieved up to round $j - 1$ are used as anchor items to revise the test-query embedding, which in turn is used to update the approximate scores (line 13 in Algorithm 1). Finally, the items selected so far are masked out and the next batch of $k_{s}$ items in round $j$ is retrieved using the updated approximate scores in the following two ways: + +- TopK: Greedily pick top- $k_{s}$ items according to approximate scores. +- SoftMax: Convert approximate item scores into probability using softmax and sample $k_{s}$ items without replacement. + +Finally, ADACUR returns top- $k$ items based on exact cross-encoder scores $^{1}$ from the set of retrieved (anchor) items as approximate $k$ -NN items. We refer interested readers to Appendix B.5 for a discussion on the time complexity of ADACUR. + +# 3 Experiments + +In our experiments, we evaluate the proposed approach and baselines on the task of finding $k$ -nearest neighbor items as per a given cross-encoder. We experiment with two cross-encoders - one trained for the task of zero-shot entity linking, and another trained on information retrieval datasets. + +Experimental Setup We run evaluation on domains YuGiOh, StarTrek, and Military from ZESHEL-a zero-shot entity linking dataset (Logeswaran et al., 2019), and domains SciDocs and HotpotQA from BEIR-a zero-shot information retrieval benchmark (Thakur et al., 2021). We use two cross-encoder models trained on labeled training data from the corresponding benchmark and evaluate separately on each domain on the task of find + +ing $k$ -NN cross-encoder items. For each ZESHEL domain, we randomly split the query set into a train/anchor set $(Q_{train})$ and a test set $(Q_{test})$ while for BEIR domains, we use pseudo-queries released as part of the benchmark as train/anchor queries and evaluate on queries in the official test split. We refer the reader to Table 1 for additional details. + +Baselines We compare our proposed approach with the following baseline retrieval methods. + +Dual-Encoders (DE): Query-item scores are computed using dot-product of embeddings produced by a learned deep encoder model. DE is used for initial retrieval followed by re-ranking using the cross-encoder. We report results for $\mathrm{DE}_{\mathrm{BASE}}$ , a dual-encoder trained on training domains in the corresponding dataset, and the following two dual-encoder models trained on the target domain via distillation using the cross-encoder. + +- $\mathrm{DE}_{\mathrm{BERT}}^{\mathrm{CE}}$ : DE initialized with BERT (Devlin et al., 2019) and trained only on the target domain via distillation using the cross-encoder. +- $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ : $\mathrm{DE}_{\mathrm{BASE}}$ model further fine-tuned on the target domain via distillation. + +ANNCUR: $k$ -NN search method proposed by Yadav et al. (2022) where anchor items are chosen uniformly at random. We additionally experiment with ANNCUR $_{\text{DE}_{\text{BASE}}}$ which uses top-scoring items retrieved using DE $_{\text{BASE}}$ as anchor items. + +Evaluation Metric Following the precedent set by previous work (Yadav et al., 2022), we evaluate all approaches using Top- $k$ -Recall@ $\mathcal{B}_{\mathrm{CE}}$ which is defined as the fraction of $k$ -NN items retrieved under test-time cost budget $\mathcal{B}_{\mathrm{CE}}$ where the cost is defined as the number of cross-encoder calls made during inference. DE baselines will use the entire budget of $\mathcal{B}_{\mathrm{CE}}$ calls for re-ranking retrieved items using exact cross-encoder scores, ANNCUR splits the budget between scoring anchor $k_{i}$ items and using exact cross-encoder scores for re-ranking $\mathcal{B}_{\mathrm{CE}} - k_{i}$ retrieved items, and ADACUR use the budget to adaptively search for $k$ -NN items. + +For ADACUR, unless noted otherwise, we use $N_{\mathcal{R}} = 5$ with TopK method for retrieving items using approximate scores, and retrieve the first batch of items uniformly at random (ADACUR) or using $\mathrm{DE}_{\mathrm{BASE}}$ (ADACUR $\mathrm{DE}_{\mathrm{BASE}}$ ). We refer readers to Appendix B for implementation details and details on training and parameterization of cross-encoder and dual-encoder models used in our experiments. + +![](images/468480c5d18e58a6266e7f86040cda08c8fe2863a85b408fd2c723b680c3d722.jpg) +Figure 2: Top- $k$ -Recall for ADACUR and baselines for domain=YuGi0h, $|\mathcal{Q}_{\mathrm{train}}| = 500$ . ADACUR consistently outperforms the corresponding ANNCUR variant and $\mathrm{ADACUR}_{\mathrm{DE}_{\mathrm{BASE}}}$ outperforms all DE-based retrieve-and-erank approaches including $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ , a DE model trained via distillation on the target domain using the cross-encoder. + +# 3.1 Results + +ADACUR versus baselines Figure 2 shows Top- $k$ -Recall for ADACUR and baselines on domain $\equiv$ YuGiOh. ADACUR, which uses adaptively retrieved items as anchor items over $N_{\mathcal{R}} = 5$ rounds, consistently outperforms ANNCUR which selects all anchor items uniformly at random. ADACUR also outperforms strong DE baseline $\mathrm{DE}_{\mathrm{BERT}}^{\mathrm{CE}}$ for all values of $k$ and outperforms $\mathrm{DE}_{\mathrm{BASE}}$ & $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ for large values of $k = 10,100$ + +Sampling anchor items using $\mathbf{DE}_{\mathrm{BASE}}$ For $k = 1, 10$ , Top- $k$ -Recall for both ANNCUR and ADACUR can be further improved by leveraging baseline retrieval models such as $\mathbf{DE}_{\mathrm{BASE}}$ for retrieving all and the first batch of anchor items respectively instead of sampling them uniformly at random. ADACUR $_{\mathsf{DE}_{\mathrm{BASE}}}$ always performs better than ANNCUR $_{\mathsf{DE}_{\mathrm{BASE}}}$ which in turn performs better than re-ranking items retrieved using $\mathbf{DE}_{\mathrm{BASE}}$ . Thus, for a given cost budget $(\mathcal{B}_{\mathrm{CE}})$ , even when a strong baseline retrieval model such as $\mathbf{DE}_{\mathrm{BASE}}$ is available, using the baseline retrieval model to select the first batch of $\mathcal{B}_{\mathrm{CE}} / N_{\mathcal{R}}$ items followed by using the proposed approach to adaptively retrieve more items over remaining $N_{\mathcal{R}} - 1$ rounds performs better than merely re-ranking $\mathcal{B}_{\mathrm{CE}}$ top-scoring items retrieved using $\mathbf{DE}_{\mathrm{BASE}}$ . Finally, note that ADACUR $_{\mathsf{DE}_{\mathrm{BASE}}}$ outperforms all baselines including $\mathbf{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ which requires additional compute- and resource-intensive distillation-based fine-tuning of $\mathbf{DE}_{\mathrm{BASE}}$ on the target domain. + +We refer the reader to Appendix C for results on other domains, training data size $(|\mathcal{Q}_{\mathrm{train}}|)$ values, oracle-based sampling experiments (§C.2), and additional results for multi-vector encoder (§C.3) and TF-IDF (§C.4) baselines. + +![](images/fcec76acae6b794158eba27856dfc360ee84cbac3b85b68afc4194cd0a01c3ef.jpg) +Figure 3: Top- $k$ -Recall for ADACUR for different number of rounds for domain=YuGiOh, $|\mathcal{Q}_{\mathrm{train}}| = 500$ + +![](images/08f54e341fd0cf65dec2446d644726baaca859380610577c62cb2d64470196dc.jpg) +Figure 4: ADACUR inference latency versus number of rounds for domain $= \text{YuGi0h}$ , $|\mathcal{Q}_{\text{train}}| = 500$ . + +Effect of number of rounds Figure 3 shows Top- $k$ -Recall and Figure 4 shows total inference latency of ADACUR on primary y-axis for various values of the test-time cross-encoder call budget $(\mathcal{B}_{\mathrm{CE}})$ . The secondary y-axis in Figure 4 shows the fraction of total inference time spent on each one of the three main steps in Algorithm 1 - (a) computing $U$ as pseudo-inverse of $R_{\mathrm{anc}}[\mathcal{I}_{\mathrm{anc}}]$ (line 12), (b) updating approximate scores for all items (line 13), and (c) computing exact cross-encoder scores for retrieved items (line 16). Since ADACUR selects items uniformly at random in the first round, ADACUR with $N_{\mathcal{R}} = 1$ performs poorly as it simply returns a subset of items sampled uniformly at random. As expected, Top- $k$ -Recall for ADACUR increases with the number of rounds and saturates + +at around 5-10 rounds while incurring negligible overhead in inference latency. As shown in Figure 4, cross-encoder score computation is the main bottleneck in test-time inference, taking $\sim 7$ ms per score2. Increasing $N_{\mathcal{R}}$ to large values such as 100 can incur up to $25\%$ overhead with step (a) contributing significantly to this overhead. Although the time taken by matrix multiplication in step (b) is linear in the number of items in the domain, we observe that it is a negligible fraction of overall latency on GPUs even for domain=HotpotQA with 5 million items (see Figure 5) as GPUs can enable significant speedup even for brute-force computation of this step. + +# 4 Conclusion + +In this paper, we presented an adaptive search strategy that incrementally builds a query embedding to approximate cross-encoder scores and performs $k$ -NN search using approximate scores over several rounds. Our approach is designed to reduce approximation error for the top-scoring items and hence improves $k$ -NN search recall when retrieving items based on the approximate scores. We perform an in-depth empirical analysis of the proposed approach in terms of both retrieval quality and efficiency. + +# Limitations + +Building the index for the ADACUR is more expensive than the traditional dual-encoder index due to the computation of dense cross-encoder scores matrix (see §2.1). We have successfully run our approach on up to 5 million items, but scaling to billions of items is an interesting direction for future work. Dual-encoder-based retrieve-and-erank baseline approaches can benefit from training the dual-encoder on multiple domains. It is not clear if data from multiple domains can be leveraged to improve performance of the proposed approach on a given target domain; although in any case, cross-encoders tend to be more robust to domain shift than using only dual-encoders for retrieval. + +# Ethics Statement + +In this paper we tackle the task of finding $k$ -nearest neighbor items for a given query when query-items scores are computed using a black-box similarity function such as a cross-encoder model. The cross-encoder scoring function may have certain biases + +and error tendencies, and it is unclear if our proposed method to approximate cross-encoder scores exacerbates or mitigates such biases. + +# Acknowledgments + +We thank members of UMass IESL for helpful discussions and feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the National Science Foundation under Grant No. NSF1763618, in part by the Chan Zuckerberg Initiative under the project "Scientific Knowledge Base Construction", in part by International Business Machines Corporation Cognitive Horizons Network agreement number W1668553, in part by Amazon Digital Services, and in part using high-performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative. Any opinions, findings, conclusions, and recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor(s). + +# References + +Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. MS-MARCO: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Realm: retrieval-augmented language model pre-training. In International Conference on Machine Learning. +Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2020. Improving efficient neural ranking models with cross-architecture knowledge distillation. *ArXiv*, abs/2010.02666. +Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24. + +Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). +Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient and effective passage search via contextualized late interaction over bert. In ACM SIGIR Conference on Research and Development in Information Retrieval. +Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations, ICLR. +Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. +Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations, ICLR. +Xinyin Ma, Yong Jiang, Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei Huang, and Weiming Lu. 2021. MuVER: Improving first-stage entity retrieval with multi-view entity representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. +Michael W Mahoney and Petros Drineas. 2009. Cur matrix decompositions for improved data analysis. Proceedings of the National Academy of Sciences, 106:697-702. +Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. +Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval, 3(4). +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. +Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems. +Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems. +Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). +Nishant Yadav, Nicholas Monath, Rico Angell, Manzil Zaheer, and Andrew McCallum. 2022. Efficient nearest neighbor search for cross-encoder models using matrix factorization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. +Wenzheng Zhang and Karl Stratos. 2021. Understanding hard negatives in noise contrastive estimation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. + +# A Appendix A + +# B Training and Implementation Details + +All the code for reproducing experiments is available at https://github.com/iesl/anncur. + +# B.1 Training Cross-Encoder Models + +In our experiments, we use [EMB]-CE, a cross-encoder model variant proposed by Yadav et al. (2022) that jointly encodes a query-item pair and computes the final score using dot-product of contextualized query and item embeddings extracted after joint encoding. + +# B.1.1 ZESHEL Dataset + +ZESHEL dataset is a zero-shot entity linking containing a set of 16 domains, each containing a disjoint set of items (entities). Each domain contains a set of queries (mention) paired with their ground-truth items (entities). For ZESHEL, we use the cross-encoder model checkpoint released by Yadav et al. (2022). The cross-encoder model was trained by first training a dual-encoder model on ZESHEL training data using hard negatives, and then training a cross-encoder model for the task of zero-shot entity-linking on all eight training domains using cross-entropy loss with ground-truth + +
DatasetDomain|I|(|Qtrain|, |Qtest|) SplitsTrain Query (Qtrain) Type
ZESHELYuGiOh10,031(100/3274), (500/2874), (2000/1374)Real Queries
ZESHELStarTrek34,430(100/4127), (500/3727), (2000/2227)Real Queries
ZESHELMilitary104,520(100/2300), (500/1900), (2000/0400)Real Queries
BEIRSciDocs25,657(1000/1000)Pseudo-Queries
BEIRHotpotQA5,233,329(1000/1000)Pseudo-Queries
+ +Table 1: Statistics on the number of items $(\mathcal{I})$ and the number of queries in train and test splits for each domain. The train-query $(Q_{train})$ split refers to queries used for distilling dual-encoder models or for indexing items using ADACUR and ANNCUR. For ZESHEL domains, we create train-test splits by splitting the queries in each domain uniformly at random and test with three different splits by putting 100, 500, or 2000 queries in train split. For BEIR domains, we use pseudo-queries released as part of the benchmark as train queries $(Q_{train})$ and run $k$ -NN evaluation on test queries from the official test split (as per BEIR benchmark) of these domains. For HotpotQA, we use the first 1K queries out of a total of 7K test queries and we use all 1K test queries for SciDocs. + +entity and negative entities mined using the dual-encoder. We refer readers to Yadav et al. (2022) for further details on cross-encoder training. + +We perform $k$ -NN experiments on domains YuGiOh, StarTrek, and Military from ZESHEL of which only Military was part of the training data used to train the cross-encoder model and YuGiOh and StarTrek are part of the original ZESHEL test domains and the cross-encoder model was not trained on these domains. + +# B.1.2 BEIR + +We follow the training setup used by Hofstätter et al. (2020). We first train three teacher cross-encoders initialized with albert-large-v2 (Lan et al., 2020), bert-large-whole-word-masking, and bert-base-uncased (Devlin et al., 2019), and compute soft labels on 40 million (query, positive item, negative item) triplets in MS-MARCO dataset (Bajaj et al., 2016). We then train our cross-encoder model parameterized using a 6-layer MINI-LM model (Wang et al., 2020) via distillation using average scores of the three teacher models as target signal and minimizing mean-square-error between predicted and target scores. We use training scripts available as part of sentence-transformer $^{4}$ repository to train the cross-encoder model and use a dot-product based scoring mechanism for cross-encoders proposed by Yadav et al. (2022). + +# B.2 Training Dual-Encoder Models + +# B.2.1 ZESHEL dataset + +We report results for DE baselines as reported in Yadav et al. (2022). The DE models were initialized using bert-base-uncased and contain separate query and item encoders, thus containing a total of + +$2 \times 110M$ parameters. We refer readers to Yadav et al. (2022) for details related to the training of all DE model variants on ZESHEL dataset. + +# B.2.2 BEIR benchmark + +For BEIR domains, we use a dual-encoder model released as part of sentence-transformer repository as $\mathrm{DE}_{\mathrm{BASE}}$ . This dual-encoder model was initialized using distillbert-base (Sanh et al., 2019) and trained on MS-MARCO dataset. This $\mathrm{DE}_{\mathrm{BASE}}$ is not trained on target domains SciDocs and HotpotQA used for running $k$ -NN experiments. + +We finetune $\mathrm{DE}_{\mathrm{BASE}}$ via distillation on the target domain to get $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ model. Given a set of training queries $\mathcal{Q}_{\mathrm{train}}$ from the target domain, we retrieve top-100 or top-1000 items for each query, score the items with the cross-encoder model, and train the dual-encoder by minimizing cross-entropy loss between predicted query-item scores (using DE) and target query-item scores (obtained using cross-encoder). Training a $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ with 1K queries and 100 or 1000 items per query takes around 2 hrs and 10 hrs respectively on an Nvidia RTX8000 GPU with 48GB memory. We train $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ for 10 epochs when using top-100 items per query and for 4 epochs when using top-1000 items per query using AdamW (Loshchilov and Hutter, 2019) optimizer with learning rate 1e-5. + +# B.3 ANNCUR Implementation details + +For ANNCUR, we report results for the optimal split of cross-encoder call budget $(\mathcal{B}_{\mathrm{CE}})$ between scoring $k_{i}$ anchor items followed by retrieving $\mathcal{B}_{\mathrm{CE}} - k_{i}$ items for re-ranking. We experiment with $k_{i} \in \{i\mathcal{B}_{\mathrm{CE}} / 10 : 1 \leq i \leq 9\}$ . If the retrieved items contain a subset of anchor items for which exact + +![](images/29c2c49f811e0b2b46b1be1bcb8659e203c48f49683aae2ede5bc8bf2d365d87.jpg) +(a) Domain $=$ Military (100K items), $|\mathcal{Q}_{\mathrm{train}}| = 500$ . Each CE call takes $\sim 7\mathrm{ms}$ on an Nvidia 2080ti GPU for a cross-encoder parameterized using 12-layered transformer model. + +![](images/91413bc34b2a9a2d551191352b6f7b59461108ea5396ebb81fa13c00eded5eea.jpg) +(b) Domain=HotpotQA (5 Million items), $|\mathcal{Q}_{\mathrm{train}}| = 1000$ . Each CE call takes $\sim 2$ ms on an Nvidia RTX8000 GPU for a cross-encoder parameterized using 6-layered transformer model. +Figure 5: ADACUR inference latency versus number of rounds for two different domains. The primary bottleneck at inference time is the time taken to compute cross-encoder (CE) scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. See §B.5 for detailed discussion. + +cross-encoder score has already been computed, we retrieve more than $\mathcal{B}_{\mathrm{CE}} - k_{i}$ items using approximate scores and compute exact cross-encoder scores for them until we have exhausted the entire cross-encoder call budget for the re-ranking step. + +# B.4 ADACUR Implementation details + +For all our $k$ -NN search experiments, we used Nvidia 2080ti GPUs with 12GB memory for domains YuGiOh (10K items), StarTrek (34K items), Military (100K items), and SciDocs (25K items), and we used Nvidia RTX8000 GPUs with 48GB memory for HotpotQA (5 million items). + +For HotpotQA, we restrict our $k$ -NN search to top-10K items wrt $\mathrm{DE}_{\mathrm{BASE}}$ for ADACUR $_{\mathrm{DE}_{\mathrm{BASE}}}$ . For ZESHEL domains and SciDocs, we do not use any such heuristic and search over all the items in the corresponding domain. + +# B.5 Time Complexity of ADACUR + +The offline indexing step for ADACUR takes $\mathcal{O}(k_q|\mathcal{I}|\mathcal{C}_{f_\theta})$ time as it involves computing exact cross-encoder scores for all $|\mathcal{I}|$ items in the target domain against $k_q$ anchor queries, and computing each cross-encoder score takes $\mathcal{C}_{f_\theta}$ units of time. + +At test time, we are given a budget $\mathcal{B}_{\mathrm{CE}}$ on the number of cross-encoder calls. Each one of the $N_{\mathcal{R}}$ rounds during inference with ADACUR involves approximating all item scores for the test query $(q_{\mathrm{test}})$ followed by sampling the next batch of $k_{s} = \mathcal{B}_{\mathrm{CE}} / N_{\mathcal{R}}$ items using the updated approximate scores. In the $j^{th}$ round, the score approximation step involves computing the pseudo-inverse of a $k_{q} \times jk_{s}$ -dimensional matrix (line 12 in Algo. 1), which takes $\mathcal{O}(\mathcal{C}_{\mathrm{inv}}^{k_q,jk_s})$ time, followed by a matrix multiplication step to compute updated approximate scores (line 13 in Algo. 1) + +which takes $\mathcal{O}(\mathcal{C}_{\mathrm{mul}}^{k_q,jk_s,|\mathcal{I}|})$ time. The time taken to update the approximate scores in each round is $\mathcal{O}(\mathcal{C}_{\mathrm{inv}}^{k_q,jk_s} + \mathcal{C}_{\mathrm{mul}}^{k_q,jk_s,|\mathcal{I}|})$ , and the time taken to compute cross-encoder scores for the next batch of $k_{s}$ items is $\mathcal{O}(k_s\mathcal{C}_{f_\theta})$ . Thus, the total inference latency for retrieving items over $N_{\mathcal{R}}$ rounds under a given budget of $\mathcal{B}_{\mathrm{CE}}$ cross-encoder calls is + +$$ +\begin{array}{l} \mathcal {O} \Big (\sum_ {j = 1} ^ {N _ {\mathcal {R}}} \left(k _ {s} \mathcal {C} _ {f _ {\theta}} + \mathcal {C} _ {\mathrm {i n v}} ^ {k _ {q}, j k _ {s}} + \mathcal {C} _ {\mathrm {m u l}} ^ {k _ {q}, j k _ {s}, | \mathcal {I} |}\right) \Big) \\ = \mathcal {O} \left(\mathcal {B} _ {\mathrm {C E}} \mathcal {C} _ {f _ {\theta}} + \underbrace {\sum_ {j = 1} ^ {N _ {\mathcal {R}}} \left(\mathcal {C} _ {\mathrm {i n v}} ^ {k _ {q} , j k _ {s}} + \mathcal {C} _ {\mathrm {m u l}} ^ {k _ {q} , j k _ {s} , | \mathcal {I} |}\right)} _ {\text {O v e r h e a d o f A D A C U R}}\right) \\ \end{array} +$$ + +Figure 5 shows the breakdown of ADACUR's inference latency in terms of time spent on computing cross-encoder scores, and the overhead of computing matrix inverse in line 12 and updating approximate scores by multiplying matrices in line 13 of Algorithm 1. Empirically, we observe that the primary bottleneck at inference time is the time taken to compute cross-encoder scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. The overhead is mostly dominated by computing pseudo-inverse (see line 12 in Algorithm 1) and this step is independent of the target domain size. The matrix multiplication step (line 13 in Algorithm 1) has a linear dependence on the number of items in the target domain but it is a negligible fraction of the overall running time as it can be significantly sped up using GPUs. + +For ZESHEL domains, we use a cross-encoder parameterized using bert-base (Devlin et al., 2019), and observe that each cross-encoder call takes amortized time of $\sim 7\mathrm{ms}$ on an Nvidia 2080ti + +GPU when the scores are computed in batches of size 50. Computing each cross-encoder score sequentially i.e. with batch-size $= 1$ takes $\sim 13\mathrm{ms}$ per score. We did not observe any further reduction in amortized time to compute each score when increasing the batch size beyond 50. + +The amortized time per cross-encoder call is approximately 6ms and 2ms for SciDocs and HotpotQA respectively when using batch size=50 and MINI-LM-based (Wang et al., 2020) cross-encoder. The difference in time per cross-encoder score for SciDocs and HotpotQA is due to the difference in average query-item pair sequence length. + +# C Additional Results and Analysis + +# C.1 TopK vs SoftMax for ADACUR + +Figure 6 shows Top- $k$ -Recall for ADACUR on domain=YuGiOh, $|\mathcal{Q}_{\mathrm{train}}| = 500$ , when using TopK and SoftMax strategies for sampling items based on approximate scores (see §2.2 for details). TopK sampling strategy which greedily picks top- $k$ items based on approximate scores results in superior recall as compared to sampling items using softmax over approximate scores. + +![](images/88ec95ef8657532e0fba2ace46f8d303891aeece3b7da3a9729ac6d7f87a43d7.jpg) +Figure 6: Top- $k$ -Recall for ADACUR on YuGi0h, $|\mathcal{Q}_{\mathrm{train}}| = 500$ for different strategies for sampling items based on approximate scores described in §2.2. + +# C.2 Anchor Item Sampling with Oracle + +ADACUR performs retrieval over multiple rounds using approximate cross-encoder scores and uses the items retrieved based on the approximate scores as anchor items to improve the approximation and hence retrieval in subsequent rounds. In this section, we run experiments where the anchor item sampling method has oracle access to exact cross-encoder scores of all items for the given test query to better understand the effect of anchor items on the approximation of cross-encoder scores and + +hence subsequent retrieval based on the approximate scores. We experiment with the following strategies for sampling $k_{i}$ anchor items for a given test query: + +- TopK $_{k_m, \epsilon}^{\mathcal{O}}$ : Mask out top- $k_m$ items wrt exact cross-encoder scores and select $k_i$ anchor items by greedily picking $(1 - \epsilon) k_i$ items starting from rank $k_m + 1$ , and sample remaining $\epsilon k_i$ anchor items uniformly at random. +- SoftMax $\mathcal{O}_{k_m,\epsilon}$ : Mask out top- $k_m$ items wrt exact cross-encoder scores and select $k_i$ anchor items by sampling $(1 - \epsilon)k_i$ anchor items using softmax over exact cross-encoder scores, and sample remaining $\epsilon k_i$ anchor items uniformly at random. + +For a given test-time cross-encoder call budget $\mathcal{B}_{\mathrm{CE}}$ , we select $k_{i}$ anchor items, compute approximate cross-encoder scores using the chosen anchor items, and then retrieve $\mathcal{B}_{\mathrm{CE}} - k_{i}$ items based on the approximate scores. We experiment with $k_{i} \in \{i\mathcal{B}_{\mathrm{CE}} / 10 : 1 \leq i \leq 9\}$ and report results for the best budget split. + +![](images/46192b36f9a83ae7f884c44a35c09ca0e972ff14238cb37c0e0e652fb35ecd76.jpg) +Figure 7: Average approximation error for CUR matrix factorization on test-queries for domain=YuGiOh and $|\mathcal{Q}_{\mathrm{train}}| = 500$ when choosing $k_{i} = 50$ anchor items uniformly at random (ANNCUR), using oracle strategies from §C.2 and for ADACUR when sampling anchor items over five rounds. Approximation error is computed as the average of absolute difference between approximate and exact item scores. + +Figure 8a shows Top- $k$ -Recall of anchor item sampling strategies $\mathsf{TopK}_{k_m,0}^{\mathcal{O}}$ and SoftMax for $k_{m}\in \{0,k\}$ ,domain $\equiv$ YuGiOh. Sampling strategies SoftMax and TopK,0, which mask out top-k items, perform significantly worse than SoftMax0,0 and TopK,0 respectively when searching for $k =$ 1, 10 nearest neighbors. This indicates that the significant improvement in Top-1-Recall and Top-10Recall for TopK,and SoftMax sampling strategies can be attributed to the presence of top- $k$ items in the anchor item set. This is because CUR matrix + +![](images/8f7f1770612cfcd55f960a07171c6abbce5cd7b7c03130a041f38b0c9cbdeaee.jpg) +(a) Sampling with and without ground-truth top- $k$ items + +![](images/731f63720dd9f479c9ffd9510a11934778c93d5043e7b4cabf77468ce995826d.jpg) +(b) Sampling using TopK strategy with exact CE scores while varying $\epsilon$ , the fraction of items sampled uniformly at random + +![](images/069c5b8fd34201c93a5345d64aec6cec9eaa9cf7437043e7da6070c95ef90aa7.jpg) +(c) Sampling using SoftMax of exact CE scores while varying $\epsilon$ , the fraction of items sampled uniformly at random +Figure 8: Top- $k$ -Recall for ADACUR, ANNCUR, and oracle sampling strategies (§C.2) that have oracle access to exact cross-encoder scores for all items for domain=YuGiOh, $|\mathcal{Q}_{\mathrm{train}}| = 500$ . +Figure 9: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGiOh, when choosing $k_{i} = 50$ anchor items using oracle strategies from §C.2 and $|\mathcal{Q}_{\mathrm{train}}| = 500$ . Top- $k$ for $k = 1,10,100$ wrt exact cross-encoder scores are annotated with text along with vertical lines, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue. + +factorization which is used to compute the approximate scores incurs negligible approximation error on anchor items, and hence on top- $k$ items when these items are part of the anchor set as shown in figures 7 and 9. For $\mathsf{TopK}_{k,0}^{\mathcal{O}}$ and SoftMax $_{k,0}$ sampling strategies, since the top- $k$ items are not part of the anchor set, CUR incurs a much higher approximation error for the top- $k$ items (see examples in Figures 9a and 9b), thus resulting in poor Top $k$ -Recall as shown in Figure 8a. + +Effect of diversity in anchor items Figure 8a shows that sampling items based on softmax of exact cross-encoder scores $(\mathrm{SoftMax}_{k_m,0}^{\mathcal{O}})$ performs better than greedily picking top-scoring items $(\mathrm{TopK}_{k_m,0}^{\mathcal{O}})$ , for both $k_{m} = 0,k$ . The reason behind $\mathrm{SoftMax}_{k_m,0}^{\mathcal{O}}$ performing better than $\mathrm{TopK}_{k_m,0}^{\mathcal{O}}$ is that sampling based on softmax of exact scores yields an anchor set with a more diverse + +![](images/d4edc159cafaf082455b5d7a4a6aaf880489564b4ac572d5de20c7eb97be2d5e.jpg) + +![](images/3b184eae35c46035b21e18a832ec8bbe02e620fb6eb15ece27a3b7e65f6a9c7f.jpg) + +![](images/34fc54cb2a1665ab542402350f2ff32cb35feb88c4fcd09a8fe610bf43c6b5f4.jpg) +(e) $\mathrm{TopK}_{0,\epsilon}^{\mathcal{O}},\epsilon = 0.75$ + +![](images/2f77a9f69c2ead25abd460c0df00537461618319f565f7e2b9279727e2cf9988.jpg) + +![](images/339b6daec2e6a8ca22674231d1241050aad8810a8688fa8da4bd92ea03412f5d.jpg) + +![](images/6c73fd025cc0b7b43528bd16db8bbe1ad8fd52d35e6db6f2a06649cdf5dd95fb.jpg) +(f) SoftMax0,ε, ε = 0.75 + +score distribution whereas greedily selecting topscoring items using exact scores results in an anchor set with items having similar cross-encoder scores. However, as shown in Figures 9c and 9d, both of these sampling strategies can result in overestimating scores for all items, even the irrelevant ones (i.e. items beyond top- $k$ items) due to insufficient representation of the irrelevant items in the anchor set. Thus retrieving based on approximated scores may struggle to retrieve relevant $k$ -NN items, especially for larger values of $k$ such as $k = 100$ when the anchor items are chosen using oracle strategies such as $\mathsf{TopK}_{k_m,0}^{\mathcal{O}}$ . + +Figures 9e and 9f, where $\epsilon = 75\%$ of 50 items are sampled uniformly at random, show that overestimating scores of irrelevant items can be avoided by sampling a fraction of anchor items uniformly at random to increase the diversity of the anchor item set. As shown in Figures 8b and 8c, Top- $k$ -Recall for both SoftMax $_{0,\epsilon}$ and TopK $_{0,\epsilon}$ generally improves with an increase in $\epsilon$ , the fraction of ran + +dom items in the anchor set, due to increased diversity in the anchor item set. Since $\mathrm{SoftMax}_{0,\epsilon}^{\mathcal{O}}$ already samples a diverse set of anchor items, increasing $\epsilon$ yields only marginal improvement while for $\mathrm{TopK}_{0,\epsilon}^{\mathcal{O}}$ , increasing $\epsilon$ yields significant improvements due to increased diversity of the anchor set. A small drop in performance is observed for larger values of $\epsilon$ as increasing $\epsilon$ beyond a threshold results in some of the top- $k$ items to be excluded from the anchor item set. This results in a poorer approximation of scores for the missing top- $k$ items and hence poor retrieval recall as the retrieval is done using the approximate scores. + +Finally, the optimal strategy for choosing the set of anchor items is the one that strikes a balance between selecting anchor items with diverse cross-encoder scores and greedily picking top- $k$ items. Our proposed strategy ADACUR improves over ANNCUR as greedily picking top-scoring items according to approximate scores to expand set of anchor items increases the likelihood of picking ground-truth $k$ -NN items to be part of the anchor set, with this likelihood improving after each round with improvement in the score approximation, and ADACUR achieves diversity in the anchor items as a result of sampling items uniformly at random in the first round and due to error in the approximate scores, as shown in Figure 11. + +# C.3 Comparison with Multi-Vector Models + +Multi-vector models (Khattab and Zaharia, 2020; Ma et al., 2021) produce multiple embeddings for each query and item. For a given query $q$ and item $i$ , the query-item score is computed using simple functions such as average similarity or sum-of-maximum similarities between the set of embeddings for query $q$ and item $i$ . + +Figure 10 shows Top- $k$ -Recall for $\mathrm{DE}_{\mathrm{BASE}}$ , $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ , $\mathrm{ADACUR}_{\mathrm{DE}_{\mathrm{BASE}}}$ , and MUVER (Ma et al., 2021), a recent multi-vector model trained on ZESHEL dataset. For MUVER, we use the pretrained checkpoint released by Ma et al. (2021) with the view-merging inference strategy as described in Ma et al. (2021). While MUVER can be more accurate than $\mathrm{DE}_{\mathrm{BASE}}$ , $\mathrm{DE}_{\mathrm{BASE}}^{\mathrm{CE}}$ obtained by finetuning $\mathrm{DE}_{\mathrm{BASE}}$ model on the target domain outperforms MUVER and our proposed method $\mathrm{ADACUR}_{\mathrm{DE}_{\mathrm{BASE}}}$ yields the best Top- $k$ -Recall versus inference cost trade-offs for all values of $k$ . + +We would also like to note that while multi-vector models such as MUVER can be more accurate than single-embedding models such as $\mathrm{DE}_{\mathrm{BASE}}$ + +![](images/075c8aae0b2e00531a9aba5d734b5499cca0bc9ef307b92b50e25fa87d081f63.jpg) +(a) YuGiOh + +![](images/1f34f8dda5414af9c3860fa48b5405b56577d490818f162f157277aebc39529b.jpg) +(b) StarTrek +Figure 10: Top- $k$ -Recall for $\mathrm{ADACUR}_{\mathrm{DE}_{\mathrm{BASE}}}$ and baselines including multi-vector models on ZESHEL domains, $|\mathcal{Q}_{\mathrm{train}}| = 2000$ . See §C.3 for discussion. + +such multi-vector models incur significant memory overhead for storing query/item embeddings. For instance, using 15 embeddings per item with 768-dimensional embeddings would take around 250GB space for 5 million items for HotpotQA. + +# C.4 Results for TF-IDF baseline + +TF-IDF: All queries and items are embedded using a TF-IDF vectorizer trained on item descriptions and top- $k$ items are retrieved using the dot-product of sparse query and item embeddings. + +For domains in ZESHEL, we report results for TF-IDF baseline, for ANNCUR when anchor items are chosen using TF-IDF baseline (ANNCUR $_{TF-IDF}$ ), and for ADACUR when the first batch of anchor items is chosen using TF-IDF baseline (ADACUR $_{TF-IDF}$ ). Figures 13, 14, and 15 show Top- $k$ -Recall for domains YuGiOh, StarTrek, and Military respectively for $Q_{train} \in \{100, 500, 2000\}$ . For each baseline retrieval method, ADACUR always performs better than ANNCUR which in turn generally performs better than merely re-ranking items retrieved using the corresponding baseline retrieval method. In most cases, Top- $k$ -Recall for ADACUR $_{DE_BASE}$ > ANNCUR $_{DE_BASE}$ > DE_BASE, and ADACUR $_{TF-IDF}$ > ANNCUR $_{TF-IDF}$ > TF-IDF. + +![](images/763b9c48814b7745250edc69497dfb3c38126d61b02e4957e735caa83a0b6642.jpg) + +![](images/7518ad6ae0a574130bd4080f38f196955d7a2a1c1d79ca4d8ce2109833a17779.jpg) + +![](images/96a47a66eb2f2b4ef6a38afb09f68b8d6853aff1bb8474e5ae63e9e0df010b7c.jpg) +(a-2) ADACUR Round 2. Total 20 anchor items. + +![](images/1f379d1691fb3eb6a3681ba1fda992b299e2fb50bfaa076d10e2c519db05df39.jpg) + +![](images/5151c29ddb39fc6530ecfd730fc03df4943f463421ddff978a4042133c404e0d.jpg) +(a-1) ADACUR Round 1. Total 10 anchor items. +(a-4) ADACUR Round 4. Total 40 anchor items. + +![](images/162b5f10276b0571ecf8f0a62fcc9a82a0483204fc7bd1d8c8d96e580a90a811.jpg) +(a-5) ADACUR Round 5. Total 50 anchor items. + +![](images/6c4e7c6366495e8ef3caaab2830694afa6ce04e3478481dfa4202ca424d13b0a.jpg) +(a-3) ADACUR Round 3. Total 30 anchor items. +(a-6) ANNCUR- Sampling all 50 anchor items uniformly at random. + +![](images/f69a884dabd28ca8be0280a730f6964371c86fb8107c6a1d53b6a4588dd43b4c.jpg) +(a) Sampling 50 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random). + +![](images/3b376c8bdf6b620d1392a9d0bc11e0a9b80dcb52d4ac0b52a27ca3344912ad93.jpg) + +![](images/c82a498bdf536b39bf5dbc1cebe2cb8b68d3d8e77d55a239959ee38bb2e74ae3.jpg) + +![](images/36ffbad975499429d4b0e9fbe49510eddb56bf9700b28b9e3ae37d3120f3b6e9.jpg) +(b-1) ADACUR Round 1. Total 40 anchor items. +(b-4) ADACUR Round 4. Total 160 anchor items. +(b) Sampling 200 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random). +Figure 11: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGi0h, $|\mathcal{Q}_{\mathrm{train}}| = 500$ when choosing $k_{i} = 50$ and 200 anchor items with ADACUR over five rounds, and uniformly at random with ANNCUR. Top- $k$ for $k = 1,10,100$ wrt exact cross-encoder scores are annotated with text, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue. With ADACUR, the first batch containing anchor items in Figure 11a-1 and 11b-1 is chosen uniformly at random and in subsequent rounds, items with highest approximate scores are chosen. Note that the approximation error for top-scoring items improves significantly when the 50 anchor items are chosen adaptively (see Figure 11a-5) with the improvement being much more significant than merely increasing the number of anchor items sampled uniformly at random from 50 in Figure 11a-6 to 200 in Figure 11b-6. + +![](images/91cb2168fd0e13c5f46a97a1cf8b351be75c6bef9bb087d07f8c1b307b09695e.jpg) +(b-2) ADACUR Round 2. Total 80 anchor items. +(b-5) ADACUR Round 5. Total 200 anchor items. + +![](images/1c713cf03417e4fb66d9c05c60dfa6c63b995cc5ed78cb34af3aec20da7ad11f.jpg) +(b-3) ADACUR Round 3. Total 120 anchor items. +(b-6) ANNCUR- Sampling all 200 anchor items uniformly at random. + +![](images/548aaf1aef290dfdacc82ec3453e94d19934334b277e2c81e44aa9a524cae0d7.jpg) + +![](images/5cf1e21353df666ec04e7cb036f03c5d321ba5d2f3e051d562d6f3b7b30942d5.jpg) +(a) SciDocs + +![](images/45dcacc22c42b998ee9475d79959dc13350f58d2d3abc69aab19a7a49d6e1f3d.jpg) +(b) HotpotQA +Figure 12: Top- $k$ -Recall for ADACUR (using ten rounds) and baselines for SciDocs and HotpotQA, $|\mathcal{Q}_{\mathrm{train}}| = 1000$ . + +![](images/26c367d591eafc828e8ec3fa8bb01cd893bbee9e643b5dc7c63da4191b66d281.jpg) + +![](images/dc1c1e07649c52381c6155f65fb7df2f8875ce3804e9d64c99eed8b365d16e0a.jpg) + +![](images/7caefe3aca3ccd9ae6909d884802e7775f61ed061657b1062175088d48d7c369.jpg) + +![](images/818abc361746554d903cacdd8f337f48e3f46664230fc529991d037c924a878f.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/d85e84b09a763817d072b1320fc6f98d929a283c552f7a2d373aade6d32f5a8b.jpg) + +![](images/718beb75cf87b581b7282721bb0e92e396ecfb0cc0c05c3d2c521272c71fc0c0.jpg) + +![](images/d6e18a9b73534f56dda5f095c391c89e3f4f6e3866314e916c6bf26f78afbd68.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/b52b342d648f393aa2e434f9775554dc850e0ae0e26e91aa834841b00fdd1e30.jpg) +(a) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 100$ . + +![](images/3d070858727996c8cd8553c21c7b517d98f2b9454e6e322df036bb0080c755d1.jpg) + +![](images/a6a2b60dedf4b8e1d94ed34aa0e62b2ab92873291cb31172520e3e25183e3266.jpg) +(b) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 500$ . +Inference Cost (Total Number of Cross-Encoder Calls) +(c) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 2000$ . +Figure 13: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=YuGiOh. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\mathcal{Q}_{\text {train }}|)$ . + +![](images/2df33e476e3c134a616b3611da7f4056eabe4e5274a16dc9280a4ebbc2f20314.jpg) + +![](images/af420a45981dc937114ec3b8c1f2a0d3ed76f0faf7cc6bf350cac5551de96d7f.jpg) + +![](images/8fc6aa09b47562a568286c43a5f0751d4fb8bc7f5291a76d9b167ab6e089face.jpg) + +![](images/193b5ce2fb71560cc198e7b315a9220074d6fc389adc07e07aae0978a957e76c.jpg) + +![](images/3ccf1511fc31c8bffc0db6503a9a735c57c7b1c983c5aa3a265f67efb26a1192.jpg) + +![](images/fb751c61a7f6d3a7e7cf0ad1ecb4df49b04148a5c343609e6abc21aebb322ceb.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/828d4053b0ba61338871c695d331ff7aca3f1b66d92a776827f31a029b8335f5.jpg) + +![](images/3daf09084dc3fd2f6ee54d843d7c2232e38e8c1031de72eaf0ea1706244f0a15.jpg) + +![](images/4d5d66fe6e128567aa6fa9720657c0dd8bbcfdc4d1170bcc4b12bc5016a5e669.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/0a4caec315ba0439acceea4f6791efb451cff55db854029ea953990aa1d7f449.jpg) +(a) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 100$ + +![](images/7eb4b41eff8084365b22a320c07bb9cfb45ff4ad494d696bebf723d3a0898d45.jpg) + +![](images/4d44ed728ec4c5448ce05911874c41e131047aa70bac2767856963857438068d.jpg) +(b) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 500$ . +Inference Cost (Total Number of Cross-Encoder Calls) +(c) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 2000$ . +Figure 14: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=StarTrek. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\mathcal{Q}_{\mathrm{train}}|)$ . + +![](images/3cfe699bb790ada1a2c1459da3fb3442c5b77988509b09690d430cd92b4758a0.jpg) + +![](images/a4c791d390c0c11ac9c0a7b5bd8fce0a25e27d766bbc18098428af1f976724a2.jpg) + +![](images/c206c7baf8658830cb6aca78021beeabcfdaf3716eb97a831b6d877f8126306c.jpg) +Proposed Approaches + +![](images/ad691f5806c7f6d529484dade0d1c51d78bbef1f6ddc415fd2c4692dcbee40ac.jpg) + +![](images/8de34fc979ea5757c0a473b97caa7406fb3c16e48de28e3fe577910e597b8c5f.jpg) +DACUR + +![](images/9eb61d65e7a5b6446841ec364adc76aea5fcdb36c74fc808337d25f9d337e7a4.jpg) +ADA + +![](images/4951cb502194006a395ca30868bb9fa756a46dc61b61cb996d4cf40edb4e2d37.jpg) +R + +![](images/1f78ac0be2fd48a2926be69ec03e15a9ca7470dd5981c96d479d6ddee346f069.jpg) + +![](images/4fc45c793e7d2ade24038526682a09f9f4d7be1585cbe99bfeefcb500e4b18ba.jpg) +AD +" + +![](images/6415aa68c1a31c13c8a53ba8cdb651ddcd9e37c1f185196b2992612b787c365c.jpg) + +![](images/7f87e08e8f83543348c1f535aba8374d15184fa31a41d7861c771b6f1813d7cb.jpg) + +![](images/452c47ed38e428ebd5a84e12d9c181c4689ea1fd63e8c6ccf9b8522184a24444.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/a88d57a0c64c3b47a74ad1044c677bb43b2b6a059d6c42c3492d9f663714474e.jpg) + +![](images/5fcea90f2cff7cbce75777cb63906624f696c247e9cf6b32c6d458ba0b940280.jpg) + +![](images/103460186e44d634f0671448bba078f34bebad947e0c8d9725279e21e224a9ae.jpg) +Inference Cost (Total Number of Cross-Encoder Calls) + +![](images/1b208771490ef453d2aad2814f1eb473e6b6534912dc4c840f768069b16f62e8.jpg) +(a) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 100$ . + +![](images/c982832636752109ab23ef035405faab9f68d52bac0b76386fab6c27176d98b8.jpg) + +![](images/203513f44c2614a26a273351a33a23b528c2a3ec0bd168a0e9ce411312851fa0.jpg) +(b) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 500$ . +Inference Cost (Total Number of Cross-Encoder Calls) +(c) Number of train/anchor queries $|\mathcal{Q}_{\mathrm{train}}| = 2000$ . +Figure 15: Top- $k$ -Recall for ADACUR (using five rounds) and baselines for domain=Military. Each subfigure corresponds to a different value of the number of train/anchor queries $(|\mathcal{Q}_{train}|)$ . Note that $\mathrm{DE}_{\mathrm{BASE}}$ has high Top-1-Recall values as domain=Military is included in the set of train domains in ZESHEL which are used to train both $\mathrm{DE}_{\mathrm{BASE}}$ and the cross-encoder model. + +![](images/c70668bc95c575f3ab747d5b95f4885237498be6a8b9e037a6f116d7ef56b89b.jpg) + +![](images/1c3751bc076387a24ad37899f74b4a3f7f59313542929440e5c428a7dd1965d2.jpg) \ No newline at end of file diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/images.zip b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8db0eb4243644447ca21f2aabba940cebc46b3e2 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ca4198ce7c4e8593041545144f9ee372d372857c613d44f1d4df8f744dc6738 +size 1897755 diff --git a/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/layout.json b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..735496e8048deab85b8e8599ea66ac47090b5058 --- /dev/null +++ b/2023/Efficient k-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition/layout.json @@ -0,0 +1,18334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 142, + 76, + 451, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 76, + 451, + 110 + ], + "spans": [ + { + "bbox": [ + 142, + 76, + 451, + 110 + ], + "type": "text", + "content": "Efficient " + }, + { + "bbox": [ + 142, + 76, + 451, + 110 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 142, + 76, + 451, + 110 + ], + "type": "text", + "content": "-NN Search with Cross-Encoders using Adaptive Multi-Round CUR Decomposition" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 129, + 498, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 129, + 498, + 143 + ], + "spans": [ + { + "bbox": [ + 96, + 129, + 498, + 143 + ], + "type": "text", + "content": "Nishant Yadav†, Nicholas Monath‡, Manzil Zaheer§, and Andrew McCallum†" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 151, + 144, + 445, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 144, + 445, + 158 + ], + "spans": [ + { + "bbox": [ + 151, + 144, + 445, + 158 + ], + "type": "text", + "content": "† University of Massachusetts Amherst ‡Google Research" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 159, + 516, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 159, + 516, + 172 + ], + "spans": [ + { + "bbox": [ + 80, + 159, + 516, + 172 + ], + "type": "text", + "content": "{nishantyavad, mccallum}@cs.umass.edu {nmonath,manzilzaheer}@google.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "spans": [ + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": "Cross-encoder models, which jointly encode and score a query-item pair, are prohibitively expensive for direct " + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": "-nearest neighbor (" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": "-NN) search. Consequently, " + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": "-NN search typically employs a fast approximate retrieval (e.g. using BM25 or dual-encoder vectors), followed by reranking with a cross-encoder; however, the retrieval approximation often has detrimental recall regret. This problem is tackled by ANNCUR (Yadav et al., 2022), a recent work that employs a cross-encoder only, making search efficient using a relatively small number of anchor items, and a CUR matrix factorization. While ANNCUR's one-time selection of anchors tends to approximate the cross-encoder distances on average, doing so forfeits the capacity to accurately estimate distances to items near the query, leading to regret in the crucial end-task: recall of top-" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": " items. In this paper, we propose ADACUR, a method that adaptively, iteratively, and efficiently minimizes the approximation error for the practically important top-" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": " neighbors. It does so by iteratively performing " + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": "-NN search using the anchors available so far, then adding these retrieved nearest neighbors to the anchor set for the next round. Empirically, on multiple datasets, in comparison to previous traditional and state-of-the-art methods such as ANNCUR and dual-encoder-based retrieval-and-erank, our proposed approach ADACUR consistently reduces recall error—by up to " + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": " on the important " + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 86, + 235, + 274, + 623 + ], + "type": "text", + "content": " setting—while using no more compute than its competitors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 632, + 154, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 632, + 154, + 645 + ], + "spans": [ + { + "bbox": [ + 68, + 632, + 154, + 645 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "type": "text", + "content": "-nearest neighbor (" + }, + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 653, + 291, + 773 + ], + "type": "text", + "content": "-NN) search is a core subroutine of a variety of tasks in NLP such as entity linking (Wu et al., 2020), passage retrieval for QA (Karpukhin et al., 2020), and more generally, in retrieval-augmented machine learning models (Guu et al., 2020; Izacard et al., 2023). For many of these applications, the state-of-the-art similarity function is a cross-encoder that directly outputs a scalar similarity score after jointly encoding a" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 211, + 507, + 233 + ], + "blocks": [ + { + "bbox": [ + 323, + 211, + 507, + 233 + ], + "lines": [ + { + "bbox": [ + 323, + 211, + 507, + 233 + ], + "spans": [ + { + "bbox": [ + 323, + 211, + 507, + 233 + ], + "type": "image", + "image_path": "d7841db05b54c8187f64cb341ff50b33be4113076ed8294d41719b564dc366f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 244, + 409, + 322 + ], + "blocks": [ + { + "bbox": [ + 305, + 244, + 409, + 322 + ], + "lines": [ + { + "bbox": [ + 305, + 244, + 409, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 409, + 322 + ], + "type": "image", + "image_path": "bd6f11d141d922b6e9cae84b134ed0f204a697693f8ce9a7c1b7873bf21364b7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 327, + 403, + 339 + ], + "lines": [ + { + "bbox": [ + 312, + 327, + 403, + 339 + ], + "spans": [ + { + "bbox": [ + 312, + 327, + 403, + 339 + ], + "type": "text", + "content": "(a) Uniformly at random" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 418, + 244, + 524, + 322 + ], + "blocks": [ + { + "bbox": [ + 418, + 244, + 524, + 322 + ], + "lines": [ + { + "bbox": [ + 418, + 244, + 524, + 322 + ], + "spans": [ + { + "bbox": [ + 418, + 244, + 524, + 322 + ], + "type": "image", + "image_path": "d01a79b72270176b45e252c19a0483397eee9525be8a52a8670ce78aeb31fc8c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 327, + 508, + 339 + ], + "lines": [ + { + "bbox": [ + 434, + 327, + 508, + 339 + ], + "spans": [ + { + "bbox": [ + 434, + 327, + 508, + 339 + ], + "type": "text", + "content": "(b) Biased sampling" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 349, + 526, + 454 + ], + "lines": [ + { + "bbox": [ + 302, + 349, + 526, + 454 + ], + "spans": [ + { + "bbox": [ + 302, + 349, + 526, + 454 + ], + "type": "text", + "content": "Figure 1: Exact versus approximate cross-encoder scores (computed using ANNCUR) of all items for a test-query in domain=YuGi0h. ANNCUR incurs high approximation error on " + }, + { + "bbox": [ + 302, + 349, + 526, + 454 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 349, + 526, + 454 + ], + "type": "text", + "content": "-NN items wrt exact scores when using 50 anchor items sampled uniformly at random (Fig. 1a). In contrast, sampling 50 anchor items with probability proportional to exact cross-encoder scores (Fig. 1b) significantly improves approximation of top-scoring items." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 465, + 525, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 465, + 525, + 677 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 525, + 677 + ], + "type": "text", + "content": "given query-item pair. However, computing a single query-item score using a cross-encoder requires a forward pass of the model which can be computationally expensive as cross-encoders are typically parameterized using deep neural models such as transformers (Vaswani et al., 2017). For this reason, " + }, + { + "bbox": [ + 302, + 465, + 525, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 465, + 525, + 677 + ], + "type": "text", + "content": "-NN search with a cross-encoder typically involves retrieving candidate items using additional models such as dual-encoders or BM25 (Robertson et al., 2009), followed by re-ranking items using the cross-encoder (Logeswaran et al., 2019; Zhang and Stratos, 2021; Qu et al., 2021). However, the accuracy of such retrieve-and-erank approaches is upperbound by the recall of first-stage retrieval and may require computationally expensive distillation-based training of dual-encoders to improve recall." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 680, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 680, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 680, + 526, + 773 + ], + "type": "text", + "content": "Recent work by Yadav et al. (2022) proposed ANNCUR, a CUR factorization (Mahoney and Drineas, 2009) based method, that approximates cross-encoder using dot-product of latent query and item embeddings and performs " + }, + { + "bbox": [ + 302, + 680, + 526, + 773 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 680, + 526, + 773 + ], + "type": "text", + "content": "-NN retrieval using approximate scores followed by optionally re-ranking retrieved items using exact cross" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8088" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8088-8103" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "spans": [ + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "type": "text", + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "text", + "content": "encoder scores. The latent item embeddings are computed by comparing each item against a set of anchor queries, and the latent query embedding is computed using the query's cross-encoder scores against a fixed set of anchor items. As shown in Figure 1, when ANNCUR selects the anchor items uniformly at random (Fig 1a), it incurs higher approximation error on the top-scoring items than the rest of the items, resulting in poor " + }, + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "text", + "content": "-NN recall, and including some " + }, + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 71, + 291, + 229 + ], + "type": "text", + "content": "-NN items as part of anchor items (Fig. 1b) can significantly improve approximation error for top-scoring items." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "text", + "content": "In this work, we propose ADACUR, a search strategy that improves " + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "text", + "content": "-NN search recall by improving the approximation of top-scoring items. ADACUR performs retrieval over multiple rounds, retrieving the first batch of items either uniformly at random or using heuristic or auxiliary models such as dual-encoder or BM25 to get a first coarse approximation of item scores for the test query. In subsequent rounds, it alternates between a) performing retrieval using approximate scores and scoring retrieved items using cross-encoder and b) using all retrieved items as anchor items to improve the approximation and hence retrieval of relevant items in the subsequent rounds. Our proposed approach provides significant improvements in " + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "text", + "content": "-NN search recall over ANNCUR and dual-encoder based retrieve-and-erank approaches when performing " + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 69, + 232, + 291, + 496 + ], + "type": "text", + "content": "-NN search with cross-encoder models trained for the task of entity linking and information retrieval." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 504, + 235, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 504, + 235, + 518 + ], + "spans": [ + { + "bbox": [ + 67, + 504, + 235, + 518 + ], + "type": "text", + "content": "2 Proposed Method: ADACUR" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": "Task Given a scoring function " + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "f_{\\theta} \\colon \\mathcal{Q} \\times \\mathcal{I} \\to \\mathbb{R}" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": " that maps a query-item pair to a scalar score, and a query " + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "q \\in \\mathcal{Q}" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": "-nearest neighbor search task is to retrieve top-" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": " scoring items from a fixed item set " + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": " according to the given scoring function " + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 67, + 524, + 290, + 592 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 599, + 264, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 599, + 264, + 612 + ], + "spans": [ + { + "bbox": [ + 67, + 599, + 264, + 612 + ], + "type": "text", + "content": "2.1 ADACUR: Offline Indexing of Items" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "text", + "content": "The indexing step of ADACUR involves using the cross-encoder model " + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "inline_equation", + "content": "(f_{\\theta})" + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "text", + "content": " to score each item against a fixed set of " + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "inline_equation", + "content": "k_q" + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "text", + "content": " anchor/train queries " + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "inline_equation", + "content": "(Q_{train})" + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "text", + "content": ", to get score matrix " + }, + { + "bbox": [ + 67, + 616, + 289, + 669 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{anc}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 674, + 271, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 674, + 271, + 688 + ], + "spans": [ + { + "bbox": [ + 86, + 674, + 271, + 688 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {a n c}} (q, i) = f _ {\\theta} (q, i), \\forall (q, i) \\in \\mathcal {Q} _ {\\text {t r a i n}} \\times \\mathcal {I}", + "image_path": "14a0440005b608f9f9225cafc72f47f7605a30655f126117856bac40bad26f29.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "text", + "content": "Each column of " + }, + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "inline_equation", + "content": "E^{\\mathcal{I}} \\coloneqq R_{\\mathrm{anc}} \\in \\mathbb{R}^{k_q \\times |\\mathcal{I}|}" + }, + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "text", + "content": " corresponds to a " + }, + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "inline_equation", + "content": "k_q" + }, + { + "bbox": [ + 67, + 693, + 290, + 722 + ], + "type": "text", + "content": "-dimensional latent item embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 729, + 236, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 236, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 236, + 741 + ], + "type": "text", + "content": "2.2 ADACUR: Test-time inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "text", + "content": "The baseline method ANNCUR computes the latent test-query embedding " + }, + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "inline_equation", + "content": "e_{q_{\\mathrm{test}}}" + }, + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "inline_equation", + "content": "C_{q_{\\mathrm{test}}}" + }, + { + "bbox": [ + 67, + 746, + 290, + 776 + ], + "type": "text", + "content": ", a" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 309, + 87, + 525, + 312 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 470, + 84 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 470, + 84 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 470, + 84 + ], + "type": "text", + "content": "Algorithm 1 ADACUR " + }, + { + "bbox": [ + 304, + 71, + 470, + 84 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 71, + 470, + 84 + ], + "type": "text", + "content": "-NN Search" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": "1: Input: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "(q_{\\mathrm{test}}, R_{\\mathrm{anc}}, N_{\\mathcal{R}}, \\mathcal{B}_{\\mathrm{CE}}, \\mathcal{A})" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{test}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Test query \n3: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{anc}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Matrix containing CE scores between " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{train}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Total cross-encoder (CE) call budget. \n5: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Algorithm to use for selecting (anchor) items. \n6: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Number of iterative search rounds \n7: Output: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Approximate scores of " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{test}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " with all items, " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ": Retrieved (anchor) items with CE scores in " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{test}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": ". \n8: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "k_s \\gets \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}} \\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " Num. of items to sample per round \n9: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}} \\gets \\text{INIT}(\\mathcal{I}, k_s)" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " Initial set of anchor items \n10: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{test}} \\gets [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}} \\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " CE scores of " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{test}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n11: for " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "j \\gets 2" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " do \n12: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\hat{S}^{(j)} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " Update approx. scores \n14: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}}^{(j)} \\gets \\text{SAMPLEITEMS}(\\mathcal{A}, k_s, \\mathcal{I}_{\\mathrm{anc}}, \\hat{S}^{(j)})" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n15: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}} \\gets \\mathcal{I}_{\\mathrm{anc}} \\cup \\mathcal{I}_{\\mathrm{anc}}^{(j)}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n16: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{test}} \\gets C_{\\mathrm{test}} \\oplus [f_\\theta(q_{\\mathrm{test}}, i)]_{i \\in \\mathcal{I}_{\\mathrm{anc}}^{(j)}} \\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " Update " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{test}}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n17: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "U \\gets R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]^{\\dagger}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "U \\in \\mathbb{R}^{|I_{\\mathrm{anc}}| \\times |\\mathcal{Q}_{\\mathrm{train}}|}" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " \n18: " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\hat{S} \\gets C_{\\mathrm{test}} \\times U \\times R_{\\mathrm{anc}} \\triangleright" + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "text", + "content": " Compute approx. scores \n19: return " + }, + { + "bbox": [ + 309, + 87, + 525, + 312 + ], + "type": "inline_equation", + "content": "\\hat{S}, \\mathcal{I}_{\\mathrm{anc}}, C_{\\mathrm{test}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 307, + 340, + 524, + 502 + ], + "blocks": [ + { + "bbox": [ + 304, + 327, + 433, + 339 + ], + "lines": [ + { + "bbox": [ + 304, + 327, + 433, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 433, + 339 + ], + "type": "text", + "content": "Algorithm 2 SAMPLEITEMS" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "lines": [ + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": "1: Input: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}, k_s, \\mathcal{I}_{\\text{mask}}, S)" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": ": Algorithm for sampling items \n3: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "k_s" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": ": Number of items to sample \n4: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{mask}}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": ": Set of items to mask \n5: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": ": (Approximate) Scores for all items \n6: Output: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{select}}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": ": Set of sampled " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "k_s" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " items \n7: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\bar{S} \\gets \\text{SOFTMAX}(S)" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\bar{S}[\\mathcal{I}_{\\text{mask}}] \\gets 0" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " > Mask items in " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{mask}}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n9: if " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\text{TopK}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " then \n10: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{select}} \\gets \\text{TOPK}(\\bar{S}, k_s)" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n11: else if " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\text{SoftMax}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " then \n12: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{select}} \\gets k_s" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " items sampled using " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\bar{S}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n13: else if " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\text{Random}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " then \n14: " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{select}} \\gets k_s" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " items uniformly sampled from " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\setminus \\mathcal{I}_{\\text{mask}}" + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "text", + "content": " \n15: return " + }, + { + "bbox": [ + 307, + 340, + 524, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{select}}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "spans": [ + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}_{\\mathrm{anc}}|" + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "text", + "content": "-dimensional vector containing cross-encoder scores of " + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "inline_equation", + "content": "q_{\\mathrm{test}}" + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "text", + "content": " with a set of anchor items " + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}_{\\mathrm{anc}})" + }, + { + "bbox": [ + 302, + 525, + 525, + 553 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 563, + 476, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 563, + 476, + 579 + ], + "spans": [ + { + "bbox": [ + 351, + 563, + 476, + 579 + ], + "type": "interline_equation", + "content": "C _ {q _ {\\text {t e s t}}} = \\left[ f _ {\\theta} \\left(q _ {\\text {t e s t}}, i\\right) \\right] _ {i \\in \\mathcal {I} _ {\\text {a n c}}}", + "image_path": "bce565f4ae964bc2d08e8c42c2969ca10ce4d48ce749b7e92d6bf208e75054aa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 582, + 448, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 582, + 448, + 598 + ], + "spans": [ + { + "bbox": [ + 355, + 582, + 448, + 598 + ], + "type": "interline_equation", + "content": "e _ {q _ {\\text {t e s t}}} = C _ {q _ {\\text {t e s t}}} \\times U", + "image_path": "d6900dd3c0496a7b5a705e5d1071f930d73d557ef94597e84cc613036840f2de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "spans": [ + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "U\\in \\mathbb{R}^{|\\mathcal{I}_{\\mathrm{anc}}|\\times |\\mathcal{Q}_{\\mathrm{train}}|}" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": " is the pseudo-inverse of " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": ", the subset of columns of " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{anc}}" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": " corresponding to (anchor) items " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\mathrm{anc}}" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": ". Finally, ANCUR approximates the score for a query-item pair " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "(q_{\\mathrm{test}},i)" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": " using dot-product of the query embedding " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "e_{q_{\\mathrm{test}}}" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": " and item embedding " + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "inline_equation", + "content": "E^{\\mathcal{I}}[:,i]" + }, + { + "bbox": [ + 302, + 608, + 525, + 692 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 351, + 697, + 476, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 697, + 476, + 716 + ], + "spans": [ + { + "bbox": [ + 351, + 697, + 476, + 716 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {\\theta} (q _ {\\mathrm {t e s t}}, i) = e _ {q _ {\\mathrm {t e s t}}} ^ {\\top} E ^ {\\mathcal {I}} [:, i ]", + "image_path": "7147fb6cc1150d24303dbfc007f00eecea3287dc9cead4ad9e465f9b6fec8a4d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "content": "The main bottleneck at test-time inference is the number of items scored using the cross-encoder for the given test-query. For a given budget of cross-encoder calls, ANNCUR splits the budget " + }, + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathrm{CE}})" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8089" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "text", + "content": "into two parts - it uses " + }, + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "text", + "content": " cross-encoder calls to compare against anchor items (chosen uniformly at random or using heuristic methods) and retrieves " + }, + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} - k_{i}" + }, + { + "bbox": [ + 67, + 71, + 290, + 137 + ], + "type": "text", + "content": " items using approximate scores and re-ranks them using exact cross-encoder scores." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": "In contrast, our proposed approach ADACUR uses the cross-encoder call budget to adaptively retrieve and score items over " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " rounds, repurposing the retrieved items as anchor items as shown in Algorithm 1. ADACUR begins by sampling the first batch of " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " (anchor) items uniformly at random. The first batch of items can also be selected using baseline retrieval methods such as BM25 and dual-encoders. In the " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " round, the items retrieved up to round " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "j - 1" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " are used as anchor items to revise the test-query embedding, which in turn is used to update the approximate scores (line 13 in Algorithm 1). Finally, the items selected so far are masked out and the next batch of " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "k_{s}" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " items in round " + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 69, + 138, + 291, + 349 + ], + "type": "text", + "content": " is retrieved using the updated approximate scores in the following two ways:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 350, + 289, + 420 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 80, + 350, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 350, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 80, + 350, + 289, + 376 + ], + "type": "text", + "content": "- TopK: Greedily pick top-" + }, + { + "bbox": [ + 80, + 350, + 289, + 376 + ], + "type": "inline_equation", + "content": "k_{s}" + }, + { + "bbox": [ + 80, + 350, + 289, + 376 + ], + "type": "text", + "content": " items according to approximate scores." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 381, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 381, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 80, + 381, + 289, + 420 + ], + "type": "text", + "content": "- SoftMax: Convert approximate item scores into probability using softmax and sample " + }, + { + "bbox": [ + 80, + 381, + 289, + 420 + ], + "type": "inline_equation", + "content": "k_{s}" + }, + { + "bbox": [ + 80, + 381, + 289, + 420 + ], + "type": "text", + "content": " items without replacement." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "text", + "content": "Finally, ADACUR returns top-" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "text", + "content": " items based on exact cross-encoder scores" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "text", + "content": " from the set of retrieved (anchor) items as approximate " + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 421, + 291, + 488 + ], + "type": "text", + "content": "-NN items. We refer interested readers to Appendix B.5 for a discussion on the time complexity of ADACUR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 497, + 155, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 497, + 155, + 511 + ], + "spans": [ + { + "bbox": [ + 67, + 497, + 155, + 511 + ], + "type": "text", + "content": "3 Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 518, + 291, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 518, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 518, + 291, + 597 + ], + "type": "text", + "content": "In our experiments, we evaluate the proposed approach and baselines on the task of finding " + }, + { + "bbox": [ + 67, + 518, + 291, + 597 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 518, + 291, + 597 + ], + "type": "text", + "content": "-nearest neighbor items as per a given cross-encoder. We experiment with two cross-encoders - one trained for the task of zero-shot entity linking, and another trained on information retrieval datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 605, + 291, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 605, + 291, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 605, + 291, + 724 + ], + "type": "text", + "content": "Experimental Setup We run evaluation on domains YuGiOh, StarTrek, and Military from ZESHEL-a zero-shot entity linking dataset (Logeswaran et al., 2019), and domains SciDocs and HotpotQA from BEIR-a zero-shot information retrieval benchmark (Thakur et al., 2021). We use two cross-encoder models trained on labeled training data from the corresponding benchmark and evaluate separately on each domain on the task of find" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "text", + "content": "ing " + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "text", + "content": "-NN cross-encoder items. For each ZESHEL domain, we randomly split the query set into a train/anchor set " + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "inline_equation", + "content": "(Q_{train})" + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "text", + "content": " and a test set " + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "inline_equation", + "content": "(Q_{test})" + }, + { + "bbox": [ + 302, + 71, + 525, + 164 + ], + "type": "text", + "content": " while for BEIR domains, we use pseudo-queries released as part of the benchmark as train/anchor queries and evaluate on queries in the official test split. We refer the reader to Table 1 for additional details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 179, + 524, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 179, + 524, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 179, + 524, + 206 + ], + "type": "text", + "content": "Baselines We compare our proposed approach with the following baseline retrieval methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 209, + 525, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 209, + 525, + 327 + ], + "spans": [ + { + "bbox": [ + 302, + 209, + 525, + 327 + ], + "type": "text", + "content": "Dual-Encoders (DE): Query-item scores are computed using dot-product of embeddings produced by a learned deep encoder model. DE is used for initial retrieval followed by re-ranking using the cross-encoder. We report results for " + }, + { + "bbox": [ + 302, + 209, + 525, + 327 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 209, + 525, + 327 + ], + "type": "text", + "content": ", a dual-encoder trained on training domains in the corresponding dataset, and the following two dual-encoder models trained on the target domain via distillation using the cross-encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 331, + 525, + 406 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 316, + 331, + 525, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 525, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 525, + 374 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 316, + 331, + 525, + 374 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 316, + 331, + 525, + 374 + ], + "type": "text", + "content": ": DE initialized with BERT (Devlin et al., 2019) and trained only on the target domain via distillation using the cross-encoder." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 316, + 376, + 524, + 406 + ], + "type": "text", + "content": " model further fine-tuned on the target domain via distillation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "spans": [ + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "text", + "content": "ANNCUR: " + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "text", + "content": "-NN search method proposed by Yadav et al. (2022) where anchor items are chosen uniformly at random. We additionally experiment with ANNCUR" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "inline_equation", + "content": "_{\\text{DE}_{\\text{BASE}}}" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "text", + "content": " which uses top-scoring items retrieved using DE" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "inline_equation", + "content": "_{\\text{BASE}}" + }, + { + "bbox": [ + 302, + 409, + 525, + 476 + ], + "type": "text", + "content": " as anchor items." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "spans": [ + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": "Evaluation Metric Following the precedent set by previous work (Yadav et al., 2022), we evaluate all approaches using Top-" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": "-Recall@ " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": " which is defined as the fraction of " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": "-NN items retrieved under test-time cost budget " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": " where the cost is defined as the number of cross-encoder calls made during inference. DE baselines will use the entire budget of " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": " calls for re-ranking retrieved items using exact cross-encoder scores, ANNCUR splits the budget between scoring anchor " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": " items and using exact cross-encoder scores for re-ranking " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} - k_{i}" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": " retrieved items, and ADACUR use the budget to adaptively search for " + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 491, + 525, + 664 + ], + "type": "text", + "content": "-NN items." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "text", + "content": "For ADACUR, unless noted otherwise, we use " + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}} = 5" + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "text", + "content": " with TopK method for retrieving items using approximate scores, and retrieve the first batch of items uniformly at random (ADACUR) or using " + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "text", + "content": " (ADACUR " + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 667, + 525, + 772 + ], + "type": "text", + "content": "). We refer readers to Appendix B for implementation details and details on training and parameterization of cross-encoder and dual-encoder models used in our experiments." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 67, + 731, + 290, + 772 + ], + "type": "text", + "content": " Sorting retrieved items based on exact cross-encoder scores does not require any additional cross-encoder calls as cross-encoder scores for these items have already been computed (see line 16 in Algorithm 1)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8090" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 502, + 201 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 502, + 201 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 502, + 201 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 502, + 201 + ], + "type": "image", + "image_path": "468480c5d18e58a6266e7f86040cda08c8fe2863a85b408fd2c723b680c3d722.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "lines": [ + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "spans": [ + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "text", + "content": "Figure 2: Top-" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "text", + "content": "-Recall for ADACUR and baselines for domain=YuGi0h, " + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "text", + "content": ". ADACUR consistently outperforms the corresponding ANNCUR variant and " + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "inline_equation", + "content": "\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "text", + "content": " outperforms all DE-based retrieve-and-erank approaches including " + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 66, + 211, + 526, + 252 + ], + "type": "text", + "content": ", a DE model trained via distillation on the target domain using the cross-encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 266, + 130, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 266, + 130, + 279 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 130, + 279 + ], + "type": "text", + "content": "3.1 Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": "ADACUR versus baselines Figure 2 shows Top- " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " -Recall for ADACUR and baselines on domain " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " YuGiOh. ADACUR, which uses adaptively retrieved items as anchor items over " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}} = 5" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " rounds, consistently outperforms ANNCUR which selects all anchor items uniformly at random. ADACUR also outperforms strong DE baseline " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BERT}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " for all values of " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " and outperforms " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " & " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "text", + "content": " for large values of " + }, + { + "bbox": [ + 67, + 287, + 291, + 410 + ], + "type": "inline_equation", + "content": "k = 10,100" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "spans": [ + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": "Sampling anchor items using " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " For " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "k = 1, 10" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": ", Top-" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": "-Recall for both ANNCUR and ADACUR can be further improved by leveraging baseline retrieval models such as " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " for retrieving all and the first batch of anchor items respectively instead of sampling them uniformly at random. ADACUR" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "_{\\mathsf{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " always performs better than ANNCUR" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "_{\\mathsf{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " which in turn performs better than re-ranking items retrieved using " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": ". Thus, for a given cost budget " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathrm{CE}})" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": ", even when a strong baseline retrieval model such as " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " is available, using the baseline retrieval model to select the first batch of " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " items followed by using the proposed approach to adaptively retrieve more items over remaining " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}} - 1" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " rounds performs better than merely re-ranking " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " top-scoring items retrieved using " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": ". Finally, note that ADACUR" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "_{\\mathsf{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " outperforms all baselines including " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " which requires additional compute- and resource-intensive distillation-based fine-tuning of " + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 424, + 291, + 705 + ], + "type": "text", + "content": " on the target domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 707, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 707, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 707, + 291, + 772 + ], + "type": "text", + "content": "We refer the reader to Appendix C for results on other domains, training data size " + }, + { + "bbox": [ + 67, + 707, + 291, + 772 + ], + "type": "inline_equation", + "content": "(|\\mathcal{Q}_{\\mathrm{train}}|)" + }, + { + "bbox": [ + 67, + 707, + 291, + 772 + ], + "type": "text", + "content": " values, oracle-based sampling experiments (§C.2), and additional results for multi-vector encoder (§C.3) and TF-IDF (§C.4) baselines." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 266, + 516, + 361 + ], + "blocks": [ + { + "bbox": [ + 312, + 266, + 516, + 361 + ], + "lines": [ + { + "bbox": [ + 312, + 266, + 516, + 361 + ], + "spans": [ + { + "bbox": [ + 312, + 266, + 516, + 361 + ], + "type": "image", + "image_path": "fcec76acae6b794158eba27856dfc360ee84cbac3b85b68afc4194cd0a01c3ef.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "lines": [ + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "spans": [ + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "type": "text", + "content": "Figure 3: Top-" + }, + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "type": "text", + "content": "-Recall for ADACUR for different number of rounds for domain=YuGiOh, " + }, + { + "bbox": [ + 302, + 371, + 526, + 396 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 305, + 409, + 527, + 506 + ], + "blocks": [ + { + "bbox": [ + 305, + 409, + 527, + 506 + ], + "lines": [ + { + "bbox": [ + 305, + 409, + 527, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 409, + 527, + 506 + ], + "type": "image", + "image_path": "08f54e341fd0cf65dec2446d644726baaca859380610577c62cb2d64470196dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "lines": [ + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "type": "text", + "content": "Figure 4: ADACUR inference latency versus number of rounds for domain " + }, + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "type": "inline_equation", + "content": "= \\text{YuGi0h}" + }, + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\text{train}}| = 500" + }, + { + "bbox": [ + 302, + 513, + 525, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": "Effect of number of rounds Figure 3 shows Top- " + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": "-Recall and Figure 4 shows total inference latency of ADACUR on primary y-axis for various values of the test-time cross-encoder call budget " + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathrm{CE}})" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": ". The secondary y-axis in Figure 4 shows the fraction of total inference time spent on each one of the three main steps in Algorithm 1 - (a) computing " + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": " as pseudo-inverse of " + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{anc}}[\\mathcal{I}_{\\mathrm{anc}}]" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": " (line 12), (b) updating approximate scores for all items (line 13), and (c) computing exact cross-encoder scores for retrieved items (line 16). Since ADACUR selects items uniformly at random in the first round, ADACUR with " + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}} = 1" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": " performs poorly as it simply returns a subset of items sampled uniformly at random. As expected, Top-" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 560, + 526, + 772 + ], + "type": "text", + "content": "-Recall for ADACUR increases with the number of rounds and saturates" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 308, + 791 + ], + "type": "text", + "content": "8091" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "text", + "content": "at around 5-10 rounds while incurring negligible overhead in inference latency. As shown in Figure 4, cross-encoder score computation is the main bottleneck in test-time inference, taking " + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "inline_equation", + "content": "\\sim 7" + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "text", + "content": " ms per score2. Increasing " + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "text", + "content": " to large values such as 100 can incur up to " + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 67, + 71, + 293, + 257 + ], + "type": "text", + "content": " overhead with step (a) contributing significantly to this overhead. Although the time taken by matrix multiplication in step (b) is linear in the number of items in the domain, we observe that it is a negligible fraction of overall latency on GPUs even for domain=HotpotQA with 5 million items (see Figure 5) as GPUs can enable significant speedup even for brute-force computation of this step." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 269, + 147, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 269, + 147, + 283 + ], + "spans": [ + { + "bbox": [ + 67, + 269, + 147, + 283 + ], + "type": "text", + "content": "4 Conclusion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "text", + "content": "In this paper, we presented an adaptive search strategy that incrementally builds a query embedding to approximate cross-encoder scores and performs " + }, + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "text", + "content": "-NN search using approximate scores over several rounds. Our approach is designed to reduce approximation error for the top-scoring items and hence improves " + }, + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 293, + 291, + 426 + ], + "type": "text", + "content": "-NN search recall when retrieving items based on the approximate scores. We perform an in-depth empirical analysis of the proposed approach in terms of both retrieval quality and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 439, + 131, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 439, + 131, + 451 + ], + "spans": [ + { + "bbox": [ + 68, + 439, + 131, + 451 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 462, + 291, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 462, + 291, + 648 + ], + "spans": [ + { + "bbox": [ + 67, + 462, + 291, + 648 + ], + "type": "text", + "content": "Building the index for the ADACUR is more expensive than the traditional dual-encoder index due to the computation of dense cross-encoder scores matrix (see §2.1). We have successfully run our approach on up to 5 million items, but scaling to billions of items is an interesting direction for future work. Dual-encoder-based retrieve-and-erank baseline approaches can benefit from training the dual-encoder on multiple domains. It is not clear if data from multiple domains can be leveraged to improve performance of the proposed approach on a given target domain; although in any case, cross-encoders tend to be more robust to domain shift than using only dual-encoders for retrieval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 661, + 158, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 661, + 158, + 673 + ], + "spans": [ + { + "bbox": [ + 68, + 661, + 158, + 673 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 683, + 291, + 751 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 291, + 751 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 291, + 751 + ], + "type": "text", + "content": "In this paper we tackle the task of finding " + }, + { + "bbox": [ + 67, + 683, + 291, + 751 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 683, + 291, + 751 + ], + "type": "text", + "content": "-nearest neighbor items for a given query when query-items scores are computed using a black-box similarity function such as a cross-encoder model. The cross-encoder scoring function may have certain biases" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 111 + ], + "type": "text", + "content": "and error tendencies, and it is unclear if our proposed method to approximate cross-encoder scores exacerbates or mitigates such biases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 121, + 400, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 121, + 400, + 136 + ], + "spans": [ + { + "bbox": [ + 303, + 121, + 400, + 136 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 142, + 527, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 142, + 527, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 142, + 527, + 383 + ], + "type": "text", + "content": "We thank members of UMass IESL for helpful discussions and feedback. This work was supported in part by the Center for Data Science and the Center for Intelligent Information Retrieval, in part by the National Science Foundation under Grant No. NSF1763618, in part by the Chan Zuckerberg Initiative under the project \"Scientific Knowledge Base Construction\", in part by International Business Machines Corporation Cognitive Horizons Network agreement number W1668553, in part by Amazon Digital Services, and in part using high-performance computing equipment obtained under a grant from the Collaborative R&D Fund managed by the Massachusetts Technology Collaborative. Any opinions, findings, conclusions, and recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor(s)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 405, + 362, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 362, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 362, + 417 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 303, + 425, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 425, + 527, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 527, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 527, + 491 + ], + "type": "text", + "content": "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. MS-MARCO: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 500, + 527, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 527, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 527, + 578 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 587, + 527, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 527, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 527, + 633 + ], + "type": "text", + "content": "Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Realm: retrieval-augmented language model pre-training. In International Conference on Machine Learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 641, + 526, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 641, + 526, + 697 + ], + "spans": [ + { + "bbox": [ + 303, + 641, + 526, + 697 + ], + "type": "text", + "content": "Sebastian Hofstätter, Sophia Althammer, Michael Schröder, Mete Sertkan, and Allan Hanbury. 2020. Improving efficient neural ranking models with cross-architecture knowledge distillation. *ArXiv*, abs/2010.02666." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 706, + 526, + 772 + ], + "type": "text", + "content": "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 280, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 280, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 280, + 772 + ], + "type": "text", + "content": "For a 12-layer transformer model on an Nvidia 2080ti" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8092" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "text", + "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 146, + 291, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 146, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 146, + 291, + 201 + ], + "type": "text", + "content": "Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient and effective passage search via contextualized late interaction over bert. In ACM SIGIR Conference on Research and Development in Information Retrieval." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 210, + 291, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 210, + 291, + 265 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 291, + 265 + ], + "type": "text", + "content": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2020. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations, ICLR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 273, + 291, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 273, + 291, + 329 + ], + "spans": [ + { + "bbox": [ + 69, + 273, + 291, + 329 + ], + "type": "text", + "content": "Lajanugen Logeswaran, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, Jacob Devlin, and Honglak Lee. 2019. Zero-shot entity linking by reading entity descriptions. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 337, + 290, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 337, + 290, + 370 + ], + "spans": [ + { + "bbox": [ + 69, + 337, + 290, + 370 + ], + "type": "text", + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations, ICLR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 378, + 290, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 378, + 290, + 444 + ], + "spans": [ + { + "bbox": [ + 69, + 378, + 290, + 444 + ], + "type": "text", + "content": "Xinyin Ma, Yong Jiang, Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei Huang, and Weiming Lu. 2021. MuVER: Improving first-stage entity retrieval with multi-view entity representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 452, + 290, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 452, + 290, + 496 + ], + "spans": [ + { + "bbox": [ + 69, + 452, + 290, + 496 + ], + "type": "text", + "content": "Michael W Mahoney and Petros Drineas. 2009. Cur matrix decompositions for improved data analysis. Proceedings of the National Academy of Sciences, 106:697-702." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 505, + 291, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 505, + 291, + 593 + ], + "spans": [ + { + "bbox": [ + 69, + 505, + 291, + 593 + ], + "type": "text", + "content": "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 601, + 291, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 601, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 601, + 291, + 645 + ], + "type": "text", + "content": "Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval, 3(4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 653, + 290, + 698 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 290, + 698 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 698 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 706, + 291, + 772 + ], + "type": "text", + "content": "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 391 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 127 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 135, + 526, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 526, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 526, + 191 + ], + "type": "text", + "content": "Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. 2020. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers. Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 198, + 526, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 526, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 526, + 254 + ], + "type": "text", + "content": "Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2020. Scalable zero-shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 261, + 525, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 261, + 525, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 525, + 327 + ], + "type": "text", + "content": "Nishant Yadav, Nicholas Monath, Rico Angell, Manzil Zaheer, and Andrew McCallum. 2022. Efficient nearest neighbor search for cross-encoder models using matrix factorization. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 335, + 525, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 525, + 391 + ], + "type": "text", + "content": "Wenzheng Zhang and Karl Stratos. 2021. Understanding hard negatives in noise contrastive estimation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 400, + 387, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 387, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 387, + 413 + ], + "type": "text", + "content": "A Appendix A" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 420, + 516, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 516, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 516, + 433 + ], + "type": "text", + "content": "B Training and Implementation Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 441, + 525, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 525, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 525, + 467 + ], + "type": "text", + "content": "All the code for reproducing experiments is available at https://github.com/iesl/anncur." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 476, + 481, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 481, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 481, + 488 + ], + "type": "text", + "content": "B.1 Training Cross-Encoder Models" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 493, + 525, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 525, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 525, + 572 + ], + "type": "text", + "content": "In our experiments, we use [EMB]-CE, a cross-encoder model variant proposed by Yadav et al. (2022) that jointly encodes a query-item pair and computes the final score using dot-product of contextualized query and item embeddings extracted after joint encoding." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 580, + 423, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 423, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 423, + 591 + ], + "type": "text", + "content": "B.1.1 ZESHEL Dataset" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 596, + 525, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 525, + 754 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 525, + 754 + ], + "type": "text", + "content": "ZESHEL dataset is a zero-shot entity linking containing a set of 16 domains, each containing a disjoint set of items (entities). Each domain contains a set of queries (mention) paired with their ground-truth items (entities). For ZESHEL, we use the cross-encoder model checkpoint released by Yadav et al. (2022). The cross-encoder model was trained by first training a dual-encoder model on ZESHEL training data using hard negatives, and then training a cross-encoder model for the task of zero-shot entity-linking on all eight training domains using cross-entropy loss with ground-truth" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 527, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 527, + 772 + ], + "type": "text", + "content": "3https://huggingface.co/nishantyadav/emb_crossenc_zeshel" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8093" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 68, + 493, + 148 + ], + "blocks": [ + { + "bbox": [ + 100, + 68, + 493, + 148 + ], + "lines": [ + { + "bbox": [ + 100, + 68, + 493, + 148 + ], + "spans": [ + { + "bbox": [ + 100, + 68, + 493, + 148 + ], + "type": "table", + "html": "
DatasetDomain|I|(|Qtrain|, |Qtest|) SplitsTrain Query (Qtrain) Type
ZESHELYuGiOh10,031(100/3274), (500/2874), (2000/1374)Real Queries
ZESHELStarTrek34,430(100/4127), (500/3727), (2000/2227)Real Queries
ZESHELMilitary104,520(100/2300), (500/1900), (2000/0400)Real Queries
BEIRSciDocs25,657(1000/1000)Pseudo-Queries
BEIRHotpotQA5,233,329(1000/1000)Pseudo-Queries
", + "image_path": "d5184a9ad74ea34768e26193fcb2a40ce5f567429656c9ce4848800852aba096.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "lines": [ + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "spans": [ + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "text", + "content": "Table 1: Statistics on the number of items " + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "inline_equation", + "content": "(\\mathcal{I})" + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "text", + "content": " and the number of queries in train and test splits for each domain. The train-query " + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "inline_equation", + "content": "(Q_{train})" + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "text", + "content": " split refers to queries used for distilling dual-encoder models or for indexing items using ADACUR and ANNCUR. For ZESHEL domains, we create train-test splits by splitting the queries in each domain uniformly at random and test with three different splits by putting 100, 500, or 2000 queries in train split. For BEIR domains, we use pseudo-queries released as part of the benchmark as train queries " + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "inline_equation", + "content": "(Q_{train})" + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "text", + "content": " and run " + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 66, + 156, + 526, + 240 + ], + "type": "text", + "content": "-NN evaluation on test queries from the official test split (as per BEIR benchmark) of these domains. For HotpotQA, we use the first 1K queries out of a total of 7K test queries and we use all 1K test queries for SciDocs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 260, + 290, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 260, + 290, + 299 + ], + "spans": [ + { + "bbox": [ + 67, + 260, + 290, + 299 + ], + "type": "text", + "content": "entity and negative entities mined using the dual-encoder. We refer readers to Yadav et al. (2022) for further details on cross-encoder training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 300, + 290, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 300, + 290, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 300, + 290, + 391 + ], + "type": "text", + "content": "We perform " + }, + { + "bbox": [ + 67, + 300, + 290, + 391 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 300, + 290, + 391 + ], + "type": "text", + "content": "-NN experiments on domains YuGiOh, StarTrek, and Military from ZESHEL of which only Military was part of the training data used to train the cross-encoder model and YuGiOh and StarTrek are part of the original ZESHEL test domains and the cross-encoder model was not trained on these domains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 401, + 133, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 401, + 133, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 401, + 133, + 412 + ], + "type": "text", + "content": "B.1.2 BEIR" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 417, + 291, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 417, + 291, + 655 + ], + "spans": [ + { + "bbox": [ + 67, + 417, + 291, + 655 + ], + "type": "text", + "content": "We follow the training setup used by Hofstätter et al. (2020). We first train three teacher cross-encoders initialized with albert-large-v2 (Lan et al., 2020), bert-large-whole-word-masking, and bert-base-uncased (Devlin et al., 2019), and compute soft labels on 40 million (query, positive item, negative item) triplets in MS-MARCO dataset (Bajaj et al., 2016). We then train our cross-encoder model parameterized using a 6-layer MINI-LM model (Wang et al., 2020) via distillation using average scores of the three teacher models as target signal and minimizing mean-square-error between predicted and target scores. We use training scripts available as part of sentence-transformer" + }, + { + "bbox": [ + 67, + 417, + 291, + 655 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 417, + 291, + 655 + ], + "type": "text", + "content": " repository to train the cross-encoder model and use a dot-product based scoring mechanism for cross-encoders proposed by Yadav et al. (2022)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 666, + 243, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 666, + 243, + 678 + ], + "spans": [ + { + "bbox": [ + 67, + 666, + 243, + 678 + ], + "type": "text", + "content": "B.2 Training Dual-Encoder Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 683, + 186, + 695 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 186, + 695 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 186, + 695 + ], + "type": "text", + "content": "B.2.1 ZESHEL dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 699, + 290, + 752 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 699, + 290, + 752 + ], + "spans": [ + { + "bbox": [ + 67, + 699, + 290, + 752 + ], + "type": "text", + "content": "We report results for DE baselines as reported in Yadav et al. (2022). The DE models were initialized using bert-base-uncased and contain separate query and item encoders, thus containing a total of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 260, + 525, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 260, + 525, + 299 + ], + "spans": [ + { + "bbox": [ + 302, + 260, + 525, + 299 + ], + "type": "inline_equation", + "content": "2 \\times 110M" + }, + { + "bbox": [ + 302, + 260, + 525, + 299 + ], + "type": "text", + "content": " parameters. We refer readers to Yadav et al. (2022) for details related to the training of all DE model variants on ZESHEL dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 309, + 425, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 309, + 425, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 309, + 425, + 321 + ], + "type": "text", + "content": "B.2.2 BEIR benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "text", + "content": "For BEIR domains, we use a dual-encoder model released as part of sentence-transformer repository as " + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "text", + "content": ". This dual-encoder model was initialized using distillbert-base (Sanh et al., 2019) and trained on MS-MARCO dataset. This " + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "text", + "content": " is not trained on target domains SciDocs and HotpotQA used for running " + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 325, + 525, + 418 + ], + "type": "text", + "content": "-NN experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "spans": [ + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": "We finetune " + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": " via distillation on the target domain to get " + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": " model. Given a set of training queries " + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_{\\mathrm{train}}" + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": " from the target domain, we retrieve top-100 or top-1000 items for each query, score the items with the cross-encoder model, and train the dual-encoder by minimizing cross-entropy loss between predicted query-item scores (using DE) and target query-item scores (obtained using cross-encoder). Training a " + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": " with 1K queries and 100 or 1000 items per query takes around 2 hrs and 10 hrs respectively on an Nvidia RTX8000 GPU with 48GB memory. We train " + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 420, + 525, + 634 + ], + "type": "text", + "content": " for 10 epochs when using top-100 items per query and for 4 epochs when using top-1000 items per query using AdamW (Loshchilov and Hutter, 2019) optimizer with learning rate 1e-5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 645, + 489, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 645, + 489, + 658 + ], + "spans": [ + { + "bbox": [ + 302, + 645, + 489, + 658 + ], + "type": "text", + "content": "B.3 ANNCUR Implementation details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "spans": [ + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "content": "For ANNCUR, we report results for the optimal split of cross-encoder call budget " + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}_{\\mathrm{CE}})" + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "content": " between scoring " + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "content": " anchor items followed by retrieving " + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} - k_{i}" + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "content": " items for re-ranking. We experiment with " + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "inline_equation", + "content": "k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}" + }, + { + "bbox": [ + 302, + 664, + 524, + 743 + ], + "type": "text", + "content": ". If the retrieved items contain a subset of anchor items for which exact" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 318, + 750, + 455, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 750, + 455, + 761 + ], + "spans": [ + { + "bbox": [ + 318, + 750, + 455, + 761 + ], + "type": "text", + "content": "msmarco-distilroberta-base-v2:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 761, + 510, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 761, + 510, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 761, + 510, + 772 + ], + "type": "text", + "content": "www.sbert.net/docs/pretrained-models/msmarco-v2.html" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 80, + 760, + 268, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 268, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 268, + 772 + ], + "type": "text", + "content": "4https://github.com/UKPLab/sentence-transformers" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8094" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 292, + 166 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 292, + 166 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 292, + 166 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 292, + 166 + ], + "type": "image", + "image_path": "29c2c49f811e0b2b46b1be1bcb8659e203c48f49683aae2ede5bc8bf2d365d87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "text", + "content": "(a) Domain " + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "text", + "content": " Military (100K items), " + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "text", + "content": " . Each CE call takes " + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{ms}" + }, + { + "bbox": [ + 67, + 171, + 294, + 202 + ], + "type": "text", + "content": " on an Nvidia 2080ti GPU for a cross-encoder parameterized using 12-layered transformer model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 523, + 166 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 523, + 166 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 523, + 166 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 523, + 166 + ], + "type": "image", + "image_path": "91413bc34b2a9a2d551191352b6f7b59461108ea5396ebb81fa13c00eded5eea.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "lines": [ + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "spans": [ + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "type": "text", + "content": "(b) Domain=HotpotQA (5 Million items), " + }, + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 1000" + }, + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "type": "text", + "content": ". Each CE call takes " + }, + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "type": "inline_equation", + "content": "\\sim 2" + }, + { + "bbox": [ + 298, + 171, + 525, + 202 + ], + "type": "text", + "content": " ms on an Nvidia RTX8000 GPU for a cross-encoder parameterized using 6-layered transformer model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 211, + 525, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 211, + 525, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 211, + 525, + 247 + ], + "type": "text", + "content": "Figure 5: ADACUR inference latency versus number of rounds for two different domains. The primary bottleneck at inference time is the time taken to compute cross-encoder (CE) scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. See §B.5 for detailed discussion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 269, + 290, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 269, + 290, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 269, + 290, + 335 + ], + "type": "text", + "content": "cross-encoder score has already been computed, we retrieve more than " + }, + { + "bbox": [ + 67, + 269, + 290, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} - k_{i}" + }, + { + "bbox": [ + 67, + 269, + 290, + 335 + ], + "type": "text", + "content": " items using approximate scores and compute exact cross-encoder scores for them until we have exhausted the entire cross-encoder call budget for the re-ranking step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 347, + 253, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 347, + 253, + 359 + ], + "spans": [ + { + "bbox": [ + 67, + 347, + 253, + 359 + ], + "type": "text", + "content": "B.4 ADACUR Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 365, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 365, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 67, + 365, + 290, + 445 + ], + "type": "text", + "content": "For all our " + }, + { + "bbox": [ + 67, + 365, + 290, + 445 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 365, + 290, + 445 + ], + "type": "text", + "content": "-NN search experiments, we used Nvidia 2080ti GPUs with 12GB memory for domains YuGiOh (10K items), StarTrek (34K items), Military (100K items), and SciDocs (25K items), and we used Nvidia RTX8000 GPUs with 48GB memory for HotpotQA (5 million items)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "text", + "content": "For HotpotQA, we restrict our " + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "text", + "content": "-NN search to top-10K items wrt " + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "text", + "content": " for ADACUR" + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "inline_equation", + "content": "_{\\mathrm{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 447, + 290, + 513 + ], + "type": "text", + "content": ". For ZESHEL domains and SciDocs, we do not use any such heuristic and search over all the items in the corresponding domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 525, + 238, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 525, + 238, + 538 + ], + "spans": [ + { + "bbox": [ + 67, + 525, + 238, + 538 + ], + "type": "text", + "content": "B.5 Time Complexity of ADACUR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "content": "The offline indexing step for ADACUR takes " + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(k_q|\\mathcal{I}|\\mathcal{C}_{f_\\theta})" + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "content": " time as it involves computing exact cross-encoder scores for all " + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "inline_equation", + "content": "|\\mathcal{I}|" + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "content": " items in the target domain against " + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "inline_equation", + "content": "k_q" + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "content": " anchor queries, and computing each cross-encoder score takes " + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{f_\\theta}" + }, + { + "bbox": [ + 67, + 544, + 290, + 611 + ], + "type": "text", + "content": " units of time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": "At test time, we are given a budget " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " on the number of cross-encoder calls. Each one of the " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " rounds during inference with ADACUR involves approximating all item scores for the test query " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "(q_{\\mathrm{test}})" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " followed by sampling the next batch of " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "k_{s} = \\mathcal{B}_{\\mathrm{CE}} / N_{\\mathcal{R}}" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " items using the updated approximate scores. In the " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " round, the score approximation step involves computing the pseudo-inverse of a " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "k_{q} \\times jk_{s}" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " -dimensional matrix (line 12 in Algo. 1), which takes " + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s})" + }, + { + "bbox": [ + 67, + 612, + 291, + 772 + ], + "type": "text", + "content": " time, followed by a matrix multiplication step to compute updated approximate scores (line 13 in Algo. 1)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": "which takes " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": " time. The time taken to update the approximate scores in each round is " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathcal{C}_{\\mathrm{inv}}^{k_q,jk_s} + \\mathcal{C}_{\\mathrm{mul}}^{k_q,jk_s,|\\mathcal{I}|})" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": ", and the time taken to compute cross-encoder scores for the next batch of " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "k_{s}" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": " items is " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(k_s\\mathcal{C}_{f_\\theta})" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": ". Thus, the total inference latency for retrieving items over " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{R}}" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": " rounds under a given budget of " + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 266, + 525, + 365 + ], + "type": "text", + "content": " cross-encoder calls is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 375, + 509, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 375, + 509, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 375, + 509, + 470 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {O} \\Big (\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(k _ {s} \\mathcal {C} _ {f _ {\\theta}} + \\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q}, j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q}, j k _ {s}, | \\mathcal {I} |}\\right) \\Big) \\\\ = \\mathcal {O} \\left(\\mathcal {B} _ {\\mathrm {C E}} \\mathcal {C} _ {f _ {\\theta}} + \\underbrace {\\sum_ {j = 1} ^ {N _ {\\mathcal {R}}} \\left(\\mathcal {C} _ {\\mathrm {i n v}} ^ {k _ {q} , j k _ {s}} + \\mathcal {C} _ {\\mathrm {m u l}} ^ {k _ {q} , j k _ {s} , | \\mathcal {I} |}\\right)} _ {\\text {O v e r h e a d o f A D A C U R}}\\right) \\\\ \\end{array}", + "image_path": "d26e3e10126de3ae44c158e01723496d2ec9ec3bf673fab154189b9890f99c5f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 481, + 525, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 481, + 525, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 481, + 525, + 719 + ], + "type": "text", + "content": "Figure 5 shows the breakdown of ADACUR's inference latency in terms of time spent on computing cross-encoder scores, and the overhead of computing matrix inverse in line 12 and updating approximate scores by multiplying matrices in line 13 of Algorithm 1. Empirically, we observe that the primary bottleneck at inference time is the time taken to compute cross-encoder scores for query-item pairs at test time, and the overhead for ADACUR accumulates linearly as the number of rounds increases. The overhead is mostly dominated by computing pseudo-inverse (see line 12 in Algorithm 1) and this step is independent of the target domain size. The matrix multiplication step (line 13 in Algorithm 1) has a linear dependence on the number of items in the target domain but it is a negligible fraction of the overall running time as it can be significantly sped up using GPUs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 720, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 720, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 720, + 525, + 772 + ], + "type": "text", + "content": "For ZESHEL domains, we use a cross-encoder parameterized using bert-base (Devlin et al., 2019), and observe that each cross-encoder call takes amortized time of " + }, + { + "bbox": [ + 302, + 720, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\sim 7\\mathrm{ms}" + }, + { + "bbox": [ + 302, + 720, + 525, + 772 + ], + "type": "text", + "content": " on an Nvidia 2080ti" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8095" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "text", + "content": "GPU when the scores are computed in batches of size 50. Computing each cross-encoder score sequentially i.e. with batch-size " + }, + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "inline_equation", + "content": "= 1" + }, + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "inline_equation", + "content": "\\sim 13\\mathrm{ms}" + }, + { + "bbox": [ + 67, + 71, + 290, + 150 + ], + "type": "text", + "content": " per score. We did not observe any further reduction in amortized time to compute each score when increasing the batch size beyond 50." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 152, + 291, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 152, + 291, + 245 + ], + "spans": [ + { + "bbox": [ + 67, + 152, + 291, + 245 + ], + "type": "text", + "content": "The amortized time per cross-encoder call is approximately 6ms and 2ms for SciDocs and HotpotQA respectively when using batch size=50 and MINI-LM-based (Wang et al., 2020) cross-encoder. The difference in time per cross-encoder score for SciDocs and HotpotQA is due to the difference in average query-item pair sequence length." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 256, + 255, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 256, + 255, + 271 + ], + "spans": [ + { + "bbox": [ + 67, + 256, + 255, + 271 + ], + "type": "text", + "content": "C Additional Results and Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 279, + 238, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 279, + 238, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 238, + 291 + ], + "type": "text", + "content": "C.1 TopK vs SoftMax for ADACUR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "text", + "content": "Figure 6 shows Top-" + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "text", + "content": "-Recall for ADACUR on domain=YuGiOh, " + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "text", + "content": ", when using TopK and SoftMax strategies for sampling items based on approximate scores (see §2.2 for details). TopK sampling strategy which greedily picks top-" + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 297, + 291, + 403 + ], + "type": "text", + "content": " items based on approximate scores results in superior recall as compared to sampling items using softmax over approximate scores." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 416, + 287, + 565 + ], + "blocks": [ + { + "bbox": [ + 70, + 416, + 287, + 565 + ], + "lines": [ + { + "bbox": [ + 70, + 416, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 70, + 416, + 287, + 565 + ], + "type": "image", + "image_path": "88ec95ef8657532e0fba2ace46f8d303891aeece3b7da3a9729ac6d7f87a43d7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "lines": [ + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "type": "text", + "content": "Figure 6: Top-" + }, + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "type": "text", + "content": "-Recall for ADACUR on YuGi0h, " + }, + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 67, + 575, + 290, + 611 + ], + "type": "text", + "content": " for different strategies for sampling items based on approximate scores described in §2.2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 622, + 262, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 622, + 262, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 622, + 262, + 635 + ], + "type": "text", + "content": "C.2 Anchor Item Sampling with Oracle" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 640, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 640, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 640, + 291, + 772 + ], + "type": "text", + "content": "ADACUR performs retrieval over multiple rounds using approximate cross-encoder scores and uses the items retrieved based on the approximate scores as anchor items to improve the approximation and hence retrieval in subsequent rounds. In this section, we run experiments where the anchor item sampling method has oracle access to exact cross-encoder scores of all items for the given test query to better understand the effect of anchor items on the approximation of cross-encoder scores and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 123 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 123 + ], + "type": "text", + "content": "hence subsequent retrieval based on the approximate scores. We experiment with the following strategies for sampling " + }, + { + "bbox": [ + 302, + 71, + 526, + 123 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 302, + 71, + 526, + 123 + ], + "type": "text", + "content": " anchor items for a given test query:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 125, + 526, + 275 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": "- TopK" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "_{k_m, \\epsilon}^{\\mathcal{O}}" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": ": Mask out top-" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "k_m" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": " items wrt exact cross-encoder scores and select " + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": " anchor items by greedily picking " + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "(1 - \\epsilon) k_i" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": " items starting from rank " + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "k_m + 1" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": ", and sample remaining " + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "inline_equation", + "content": "\\epsilon k_i" + }, + { + "bbox": [ + 316, + 125, + 525, + 191 + ], + "type": "text", + "content": " anchor items uniformly at random." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": "- SoftMax" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{k_m,\\epsilon}" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": ": Mask out top-" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "inline_equation", + "content": "k_m" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": " items wrt exact cross-encoder scores and select " + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": " anchor items by sampling " + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "inline_equation", + "content": "(1 - \\epsilon)k_i" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": " anchor items using softmax over exact cross-encoder scores, and sample remaining " + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "inline_equation", + "content": "\\epsilon k_i" + }, + { + "bbox": [ + 316, + 195, + 526, + 275 + ], + "type": "text", + "content": " anchor items uniformly at random." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "spans": [ + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "content": "For a given test-time cross-encoder call budget " + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "content": ", we select " + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "content": " anchor items, compute approximate cross-encoder scores using the chosen anchor items, and then retrieve " + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\mathrm{CE}} - k_{i}" + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "content": " items based on the approximate scores. We experiment with " + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "inline_equation", + "content": "k_{i} \\in \\{i\\mathcal{B}_{\\mathrm{CE}} / 10 : 1 \\leq i \\leq 9\\}" + }, + { + "bbox": [ + 302, + 277, + 525, + 370 + ], + "type": "text", + "content": " and report results for the best budget split." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 381, + 515, + 498 + ], + "blocks": [ + { + "bbox": [ + 310, + 381, + 515, + 498 + ], + "lines": [ + { + "bbox": [ + 310, + 381, + 515, + 498 + ], + "spans": [ + { + "bbox": [ + 310, + 381, + 515, + 498 + ], + "type": "image", + "image_path": "46192b36f9a83ae7f884c44a35c09ca0e972ff14238cb37c0e0e652fb35ecd76.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "lines": [ + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "type": "text", + "content": "Figure 7: Average approximation error for CUR matrix factorization on test-queries for domain=YuGiOh and " + }, + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "type": "text", + "content": " when choosing " + }, + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "type": "inline_equation", + "content": "k_{i} = 50" + }, + { + "bbox": [ + 302, + 498, + 525, + 592 + ], + "type": "text", + "content": " anchor items uniformly at random (ANNCUR), using oracle strategies from §C.2 and for ADACUR when sampling anchor items over five rounds. Approximation error is computed as the average of absolute difference between approximate and exact item scores." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": "Figure 8a shows Top- " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " -Recall of anchor item sampling strategies " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " and SoftMax for " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "k_{m}\\in \\{0,k\\}" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " ,domain " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " YuGiOh. Sampling strategies SoftMax and TopK,0, which mask out top-k items, perform significantly worse than SoftMax0,0 and TopK,0 respectively when searching for " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "k =" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " 1, 10 nearest neighbors. This indicates that the significant improvement in Top-1-Recall and Top-10Recall for TopK,and SoftMax sampling strategies can be attributed to the presence of top- " + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 610, + 525, + 772 + ], + "type": "text", + "content": " items in the anchor item set. This is because CUR matrix" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8096" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 288, + 190 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 288, + 190 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 288, + 190 + ], + "type": "image", + "image_path": "8f7f1770612cfcd55f960a07171c6abbce5cd7b7c03130a041f38b0c9cbdeaee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 195, + 278, + 207 + ], + "lines": [ + { + "bbox": [ + 77, + 195, + 278, + 207 + ], + "spans": [ + { + "bbox": [ + 77, + 195, + 278, + 207 + ], + "type": "text", + "content": "(a) Sampling with and without ground-truth top-" + }, + { + "bbox": [ + 77, + 195, + 278, + 207 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 77, + 195, + 278, + 207 + ], + "type": "text", + "content": " items" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 210, + 286, + 302 + ], + "blocks": [ + { + "bbox": [ + 70, + 210, + 286, + 302 + ], + "lines": [ + { + "bbox": [ + 70, + 210, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 70, + 210, + 286, + 302 + ], + "type": "image", + "image_path": "731f63720dd9f479c9ffd9510a11934778c93d5043e7b4cabf77468ce995826d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 308, + 289, + 329 + ], + "lines": [ + { + "bbox": [ + 68, + 308, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 68, + 308, + 289, + 329 + ], + "type": "text", + "content": "(b) Sampling using TopK strategy with exact CE scores while varying " + }, + { + "bbox": [ + 68, + 308, + 289, + 329 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 68, + 308, + 289, + 329 + ], + "type": "text", + "content": ", the fraction of items sampled uniformly at random" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 332, + 286, + 425 + ], + "blocks": [ + { + "bbox": [ + 70, + 332, + 286, + 425 + ], + "lines": [ + { + "bbox": [ + 70, + 332, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 70, + 332, + 286, + 425 + ], + "type": "image", + "image_path": "069c5b8fd34201c93a5345d64aec6cec9eaa9cf7437043e7da6070c95ef90aa7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 430, + 289, + 451 + ], + "lines": [ + { + "bbox": [ + 67, + 430, + 289, + 451 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 289, + 451 + ], + "type": "text", + "content": "(c) Sampling using SoftMax of exact CE scores while varying " + }, + { + "bbox": [ + 67, + 430, + 289, + 451 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 67, + 430, + 289, + 451 + ], + "type": "text", + "content": ", the fraction of items sampled uniformly at random" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "lines": [ + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "type": "text", + "content": "Figure 8: Top-" + }, + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "type": "text", + "content": "-Recall for ADACUR, ANNCUR, and oracle sampling strategies (§C.2) that have oracle access to exact cross-encoder scores for all items for domain=YuGiOh, " + }, + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 67, + 460, + 291, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "lines": [ + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "spans": [ + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "text", + "content": "Figure 9: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGiOh, when choosing " + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "inline_equation", + "content": "k_{i} = 50" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "text", + "content": " anchor items using oracle strategies from §C.2 and " + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "text", + "content": ". Top-" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "inline_equation", + "content": "k = 1,10,100" + }, + { + "bbox": [ + 302, + 386, + 525, + 479 + ], + "type": "text", + "content": " wrt exact cross-encoder scores are annotated with text along with vertical lines, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": "factorization which is used to compute the approximate scores incurs negligible approximation error on anchor items, and hence on top- " + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": " items when these items are part of the anchor set as shown in figures 7 and 9. For " + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "\\mathsf{TopK}_{k,0}^{\\mathcal{O}}" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": " and SoftMax" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "_{k,0}" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": " sampling strategies, since the top- " + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": " items are not part of the anchor set, CUR incurs a much higher approximation error for the top- " + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": " items (see examples in Figures 9a and 9b), thus resulting in poor Top" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 516, + 291, + 650 + ], + "type": "text", + "content": "-Recall as shown in Figure 8a." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": "Effect of diversity in anchor items Figure 8a shows that sampling items based on softmax of exact cross-encoder scores " + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "inline_equation", + "content": "(\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}})" + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": " performs better than greedily picking top-scoring items " + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "inline_equation", + "content": "(\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}})" + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": ", for both " + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "inline_equation", + "content": "k_{m} = 0,k" + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": ". The reason behind " + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{SoftMax}_{k_m,0}^{\\mathcal{O}}" + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": " performing better than " + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{TopK}_{k_m,0}^{\\mathcal{O}}" + }, + { + "bbox": [ + 67, + 662, + 291, + 772 + ], + "type": "text", + "content": " is that sampling based on softmax of exact scores yields an anchor set with a more diverse" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 409, + 166 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 409, + 166 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 409, + 166 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 409, + 166 + ], + "type": "image", + "image_path": "d4edc159cafaf082455b5d7a4a6aaf880489564b4ac572d5de20c7eb97be2d5e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 305, + 171, + 409, + 263 + ], + "blocks": [ + { + "bbox": [ + 305, + 171, + 409, + 263 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 409, + 263 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 409, + 263 + ], + "type": "image", + "image_path": "3b184eae35c46035b21e18a832ec8bbe02e620fb6eb15ece27a3b7e65f6a9c7f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 305, + 267, + 409, + 359 + ], + "blocks": [ + { + "bbox": [ + 305, + 267, + 409, + 359 + ], + "lines": [ + { + "bbox": [ + 305, + 267, + 409, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 409, + 359 + ], + "type": "image", + "image_path": "34fc54cb2a1665ab542402350f2ff32cb35feb88c4fcd09a8fe610bf43c6b5f4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 363, + 396, + 376 + ], + "lines": [ + { + "bbox": [ + 317, + 363, + 396, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 363, + 396, + 376 + ], + "type": "text", + "content": "(e) " + }, + { + "bbox": [ + 317, + 363, + 396, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}},\\epsilon = 0.75" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 418, + 68, + 524, + 166 + ], + "blocks": [ + { + "bbox": [ + 418, + 68, + 524, + 166 + ], + "lines": [ + { + "bbox": [ + 418, + 68, + 524, + 166 + ], + "spans": [ + { + "bbox": [ + 418, + 68, + 524, + 166 + ], + "type": "image", + "image_path": "2f77a9f69c2ead25abd460c0df00537461618319f565f7e2b9279727e2cf9988.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 419, + 171, + 524, + 263 + ], + "blocks": [ + { + "bbox": [ + 419, + 171, + 524, + 263 + ], + "lines": [ + { + "bbox": [ + 419, + 171, + 524, + 263 + ], + "spans": [ + { + "bbox": [ + 419, + 171, + 524, + 263 + ], + "type": "image", + "image_path": "339b6daec2e6a8ca22674231d1241050aad8810a8688fa8da4bd92ea03412f5d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 419, + 267, + 524, + 359 + ], + "blocks": [ + { + "bbox": [ + 419, + 267, + 524, + 359 + ], + "lines": [ + { + "bbox": [ + 419, + 267, + 524, + 359 + ], + "spans": [ + { + "bbox": [ + 419, + 267, + 524, + 359 + ], + "type": "image", + "image_path": "6c73fd025cc0b7b43528bd16db8bbe1ad8fd52d35e6db6f2a06649cdf5dd95fb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 363, + 516, + 376 + ], + "lines": [ + { + "bbox": [ + 425, + 363, + 516, + 376 + ], + "spans": [ + { + "bbox": [ + 425, + 363, + 516, + 376 + ], + "type": "text", + "content": "(f) SoftMax0,ε, ε = 0.75" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": "score distribution whereas greedily selecting topscoring items using exact scores results in an anchor set with items having similar cross-encoder scores. However, as shown in Figures 9c and 9d, both of these sampling strategies can result in overestimating scores for all items, even the irrelevant ones (i.e. items beyond top- " + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": " items) due to insufficient representation of the irrelevant items in the anchor set. Thus retrieving based on approximated scores may struggle to retrieve relevant " + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": "-NN items, especially for larger values of " + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": " such as " + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "inline_equation", + "content": "k = 100" + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": " when the anchor items are chosen using oracle strategies such as " + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "inline_equation", + "content": "\\mathsf{TopK}_{k_m,0}^{\\mathcal{O}}" + }, + { + "bbox": [ + 301, + 491, + 525, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": "Figures 9e and 9f, where " + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "inline_equation", + "content": "\\epsilon = 75\\%" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": " of 50 items are sampled uniformly at random, show that overestimating scores of irrelevant items can be avoided by sampling a fraction of anchor items uniformly at random to increase the diversity of the anchor item set. As shown in Figures 8b and 8c, Top-" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": "-Recall for both SoftMax" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "inline_equation", + "content": "_{0,\\epsilon}" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": " and TopK" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "inline_equation", + "content": "_{0,\\epsilon}" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": " generally improves with an increase in " + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 302, + 666, + 526, + 772 + ], + "type": "text", + "content": ", the fraction of ran" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8097" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": "dom items in the anchor set, due to increased diversity in the anchor item set. Since " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{SoftMax}_{0,\\epsilon}^{\\mathcal{O}}" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " already samples a diverse set of anchor items, increasing " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " yields only marginal improvement while for " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\mathrm{TopK}_{0,\\epsilon}^{\\mathcal{O}}" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": ", increasing " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " yields significant improvements due to increased diversity of the anchor set. A small drop in performance is observed for larger values of " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " as increasing " + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " beyond a threshold results in some of the top-" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " items to be excluded from the anchor item set. This results in a poorer approximation of scores for the missing top-" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 71, + 291, + 243 + ], + "type": "text", + "content": " items and hence poor retrieval recall as the retrieval is done using the approximate scores." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "spans": [ + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "text", + "content": "Finally, the optimal strategy for choosing the set of anchor items is the one that strikes a balance between selecting anchor items with diverse cross-encoder scores and greedily picking top-" + }, + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "text", + "content": " items. Our proposed strategy ADACUR improves over ANNCUR as greedily picking top-scoring items according to approximate scores to expand set of anchor items increases the likelihood of picking ground-truth " + }, + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 69, + 245, + 291, + 444 + ], + "type": "text", + "content": "-NN items to be part of the anchor set, with this likelihood improving after each round with improvement in the score approximation, and ADACUR achieves diversity in the anchor items as a result of sampling items uniformly at random in the first round and due to error in the approximate scores, as shown in Figure 11." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 457, + 279, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 279, + 470 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 279, + 470 + ], + "type": "text", + "content": "C.3 Comparison with Multi-Vector Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "spans": [ + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "content": "Multi-vector models (Khattab and Zaharia, 2020; Ma et al., 2021) produce multiple embeddings for each query and item. For a given query " + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "content": " and item " + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "content": ", the query-item score is computed using simple functions such as average similarity or sum-of-maximum similarities between the set of embeddings for query " + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "content": " and item " + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 67, + 476, + 290, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": "Figure 10 shows Top- " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": "-Recall for " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": ", and MUVER (Ma et al., 2021), a recent multi-vector model trained on ZESHEL dataset. For MUVER, we use the pretrained checkpoint released by Ma et al. (2021) with the view-merging inference strategy as described in Ma et al. (2021). While MUVER can be more accurate than " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}^{\\mathrm{CE}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": " obtained by finetuning " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": " model on the target domain outperforms MUVER and our proposed method " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": " yields the best Top- " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": "-Recall versus inference cost trade-offs for all values of " + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 571, + 290, + 731 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "text", + "content": "We would also like to note that while multi-vector models such as MUVER can be more accurate than single-embedding models such as " + }, + { + "bbox": [ + 67, + 733, + 291, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 521, + 207 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 521, + 207 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 521, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 521, + 207 + ], + "type": "image", + "image_path": "075c8aae0b2e00531a9aba5d734b5499cca0bc9ef307b92b50e25fa87d081f63.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 213, + 434, + 223 + ], + "lines": [ + { + "bbox": [ + 394, + 213, + 434, + 223 + ], + "spans": [ + { + "bbox": [ + 394, + 213, + 434, + 223 + ], + "type": "text", + "content": "(a) YuGiOh" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 226, + 521, + 339 + ], + "blocks": [ + { + "bbox": [ + 308, + 226, + 521, + 339 + ], + "lines": [ + { + "bbox": [ + 308, + 226, + 521, + 339 + ], + "spans": [ + { + "bbox": [ + 308, + 226, + 521, + 339 + ], + "type": "image", + "image_path": "1f34f8dda5414af9c3860fa48b5405b56577d490818f162f157277aebc39529b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 345, + 439, + 356 + ], + "lines": [ + { + "bbox": [ + 389, + 345, + 439, + 356 + ], + "spans": [ + { + "bbox": [ + 389, + 345, + 439, + 356 + ], + "type": "text", + "content": "(b) StarTrek" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "lines": [ + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "spans": [ + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "text", + "content": "Figure 10: Top-" + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "text", + "content": "-Recall for " + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "inline_equation", + "content": "\\mathrm{ADACUR}_{\\mathrm{DE}_{\\mathrm{BASE}}}" + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "text", + "content": " and baselines including multi-vector models on ZESHEL domains, " + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 2000" + }, + { + "bbox": [ + 302, + 364, + 526, + 401 + ], + "type": "text", + "content": ". See §C.3 for discussion." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 408, + 525, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 408, + 525, + 475 + ], + "spans": [ + { + "bbox": [ + 302, + 408, + 525, + 475 + ], + "type": "text", + "content": "such multi-vector models incur significant memory overhead for storing query/item embeddings. For instance, using 15 embeddings per item with 768-dimensional embeddings would take around 250GB space for 5 million items for HotpotQA." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 483, + 460, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 483, + 460, + 496 + ], + "spans": [ + { + "bbox": [ + 302, + 483, + 460, + 496 + ], + "type": "text", + "content": "C.4 Results for TF-IDF baseline" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 501, + 524, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 501, + 524, + 553 + ], + "spans": [ + { + "bbox": [ + 302, + 501, + 524, + 553 + ], + "type": "text", + "content": "TF-IDF: All queries and items are embedded using a TF-IDF vectorizer trained on item descriptions and top-" + }, + { + "bbox": [ + 302, + 501, + 524, + 553 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 501, + 524, + 553 + ], + "type": "text", + "content": " items are retrieved using the dot-product of sparse query and item embeddings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "spans": [ + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": "For domains in ZESHEL, we report results for TF-IDF baseline, for ANNCUR when anchor items are chosen using TF-IDF baseline (ANNCUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{TF-IDF}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": "), and for ADACUR when the first batch of anchor items is chosen using TF-IDF baseline (ADACUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{TF-IDF}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": "). Figures 13, 14, and 15 show Top-" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": "-Recall for domains YuGiOh, StarTrek, and Military respectively for " + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "Q_{train} \\in \\{100, 500, 2000\\}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": ". For each baseline retrieval method, ADACUR always performs better than ANNCUR which in turn generally performs better than merely re-ranking items retrieved using the corresponding baseline retrieval method. In most cases, Top-" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": "-Recall for ADACUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{DE_BASE}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": " > ANNCUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{DE_BASE}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": " > DE_BASE, and ADACUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{TF-IDF}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": " > ANNCUR" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "inline_equation", + "content": "_{TF-IDF}" + }, + { + "bbox": [ + 302, + 554, + 525, + 766 + ], + "type": "text", + "content": " > TF-IDF." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "8098" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 182, + 68, + 411, + 95 + ], + "blocks": [ + { + "bbox": [ + 182, + 68, + 411, + 95 + ], + "lines": [ + { + "bbox": [ + 182, + 68, + 411, + 95 + ], + "spans": [ + { + "bbox": [ + 182, + 68, + 411, + 95 + ], + "type": "image", + "image_path": "763b9c48814b7745250edc69497dfb3c38126d61b02e4957e735caa83a0b6642.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 108, + 207, + 208 + ], + "blocks": [ + { + "bbox": [ + 70, + 108, + 207, + 208 + ], + "lines": [ + { + "bbox": [ + 70, + 108, + 207, + 208 + ], + "spans": [ + { + "bbox": [ + 70, + 108, + 207, + 208 + ], + "type": "image", + "image_path": "7518ad6ae0a574130bd4080f38f196955d7a2a1c1d79ca4d8ce2109833a17779.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 228, + 108, + 365, + 208 + ], + "blocks": [ + { + "bbox": [ + 228, + 108, + 365, + 208 + ], + "lines": [ + { + "bbox": [ + 228, + 108, + 365, + 208 + ], + "spans": [ + { + "bbox": [ + 228, + 108, + 365, + 208 + ], + "type": "image", + "image_path": "96a47a66eb2f2b4ef6a38afb09f68b8d6853aff1bb8474e5ae63e9e0df010b7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 215, + 337, + 234 + ], + "lines": [ + { + "bbox": [ + 242, + 215, + 337, + 234 + ], + "spans": [ + { + "bbox": [ + 242, + 215, + 337, + 234 + ], + "type": "text", + "content": "(a-2) ADACUR Round 2. Total 20 anchor items." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 387, + 108, + 524, + 208 + ], + "blocks": [ + { + "bbox": [ + 387, + 108, + 524, + 208 + ], + "lines": [ + { + "bbox": [ + 387, + 108, + 524, + 208 + ], + "spans": [ + { + "bbox": [ + 387, + 108, + 524, + 208 + ], + "type": "image", + "image_path": "1f379d1691fb3eb6a3681ba1fda992b299e2fb50bfaa076d10e2c519db05df39.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 237, + 205, + 338 + ], + "blocks": [ + { + "bbox": [ + 84, + 215, + 179, + 235 + ], + "lines": [ + { + "bbox": [ + 84, + 215, + 179, + 235 + ], + "spans": [ + { + "bbox": [ + 84, + 215, + 179, + 235 + ], + "type": "text", + "content": "(a-1) ADACUR Round 1. Total 10 anchor items." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 237, + 205, + 338 + ], + "lines": [ + { + "bbox": [ + 70, + 237, + 205, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 237, + 205, + 338 + ], + "type": "image", + "image_path": "5151c29ddb39fc6530ecfd730fc03df4943f463421ddff978a4042133c404e0d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 343, + 179, + 363 + ], + "lines": [ + { + "bbox": [ + 83, + 343, + 179, + 363 + ], + "spans": [ + { + "bbox": [ + 83, + 343, + 179, + 363 + ], + "type": "text", + "content": "(a-4) ADACUR Round 4. Total 40 anchor items." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 227, + 237, + 363, + 338 + ], + "blocks": [ + { + "bbox": [ + 227, + 237, + 363, + 338 + ], + "lines": [ + { + "bbox": [ + 227, + 237, + 363, + 338 + ], + "spans": [ + { + "bbox": [ + 227, + 237, + 363, + 338 + ], + "type": "image", + "image_path": "162b5f10276b0571ecf8f0a62fcc9a82a0483204fc7bd1d8c8d96e580a90a811.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 343, + 336, + 363 + ], + "lines": [ + { + "bbox": [ + 240, + 343, + 336, + 363 + ], + "spans": [ + { + "bbox": [ + 240, + 343, + 336, + 363 + ], + "type": "text", + "content": "(a-5) ADACUR Round 5. Total 50 anchor items." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 384, + 237, + 520, + 337 + ], + "blocks": [ + { + "bbox": [ + 401, + 215, + 496, + 234 + ], + "lines": [ + { + "bbox": [ + 401, + 215, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 401, + 215, + 496, + 234 + ], + "type": "text", + "content": "(a-3) ADACUR Round 3. Total 30 anchor items." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 237, + 520, + 337 + ], + "lines": [ + { + "bbox": [ + 384, + 237, + 520, + 337 + ], + "spans": [ + { + "bbox": [ + 384, + 237, + 520, + 337 + ], + "type": "image", + "image_path": "6c4e7c6366495e8ef3caaab2830694afa6ce04e3478481dfa4202ca424d13b0a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 343, + 523, + 364 + ], + "lines": [ + { + "bbox": [ + 383, + 343, + 523, + 364 + ], + "spans": [ + { + "bbox": [ + 383, + 343, + 523, + 364 + ], + "type": "text", + "content": "(a-6) ANNCUR- Sampling all 50 anchor items uniformly at random." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 70, + 381, + 207, + 481 + ], + "blocks": [ + { + "bbox": [ + 88, + 369, + 503, + 380 + ], + "lines": [ + { + "bbox": [ + 88, + 369, + 503, + 380 + ], + "spans": [ + { + "bbox": [ + 88, + 369, + 503, + 380 + ], + "type": "text", + "content": "(a) Sampling 50 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 381, + 207, + 481 + ], + "lines": [ + { + "bbox": [ + 70, + 381, + 207, + 481 + ], + "spans": [ + { + "bbox": [ + 70, + 381, + 207, + 481 + ], + "type": "image", + "image_path": "f69a884dabd28ca8be0280a730f6964371c86fb8107c6a1d53b6a4588dd43b4c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 229, + 381, + 364, + 481 + ], + "blocks": [ + { + "bbox": [ + 229, + 381, + 364, + 481 + ], + "lines": [ + { + "bbox": [ + 229, + 381, + 364, + 481 + ], + "spans": [ + { + "bbox": [ + 229, + 381, + 364, + 481 + ], + "type": "image", + "image_path": "3b376c8bdf6b620d1392a9d0bc11e0a9b80dcb52d4ac0b52a27ca3344912ad93.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 387, + 381, + 523, + 481 + ], + "blocks": [ + { + "bbox": [ + 387, + 381, + 523, + 481 + ], + "lines": [ + { + "bbox": [ + 387, + 381, + 523, + 481 + ], + "spans": [ + { + "bbox": [ + 387, + 381, + 523, + 481 + ], + "type": "image", + "image_path": "c82a498bdf536b39bf5dbc1cebe2cb8b68d3d8e77d55a239959ee38bb2e74ae3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 70, + 509, + 206, + 609 + ], + "blocks": [ + { + "bbox": [ + 84, + 488, + 179, + 507 + ], + "lines": [ + { + "bbox": [ + 84, + 488, + 179, + 507 + ], + "spans": [ + { + "bbox": [ + 84, + 488, + 179, + 507 + ], + "type": "text", + "content": "(b-1) ADACUR Round 1. Total 40 anchor items." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 509, + 206, + 609 + ], + "lines": [ + { + "bbox": [ + 70, + 509, + 206, + 609 + ], + "spans": [ + { + "bbox": [ + 70, + 509, + 206, + 609 + ], + "type": "image", + "image_path": "36ffbad975499429d4b0e9fbe49510eddb56bf9700b28b9e3ae37d3120f3b6e9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 616, + 181, + 635 + ], + "lines": [ + { + "bbox": [ + 83, + 616, + 181, + 635 + ], + "spans": [ + { + "bbox": [ + 83, + 616, + 181, + 635 + ], + "type": "text", + "content": "(b-4) ADACUR Round 4. Total 160 anchor items." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 641, + 506, + 653 + ], + "lines": [ + { + "bbox": [ + 85, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 85, + 641, + 506, + 653 + ], + "type": "text", + "content": "(b) Sampling 200 anchor items adaptively for ADACUR (over five rounds) and for ANNCUR (uniformly at random)." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "lines": [ + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "text", + "content": "Figure 11: Scatter plot showing approximate versus exact cross-encoder scores for a query from domain=YuGi0h, " + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "text", + "content": " when choosing " + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "inline_equation", + "content": "k_{i} = 50" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "text", + "content": " and 200 anchor items with ADACUR over five rounds, and uniformly at random with ANNCUR. Top-" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "inline_equation", + "content": "k = 1,10,100" + }, + { + "bbox": [ + 67, + 661, + 526, + 767 + ], + "type": "text", + "content": " wrt exact cross-encoder scores are annotated with text, different color bands indicate the ordering of items wrt approximate scores, and anchor items are shown in blue. With ADACUR, the first batch containing anchor items in Figure 11a-1 and 11b-1 is chosen uniformly at random and in subsequent rounds, items with highest approximate scores are chosen. Note that the approximation error for top-scoring items improves significantly when the 50 anchor items are chosen adaptively (see Figure 11a-5) with the improvement being much more significant than merely increasing the number of anchor items sampled uniformly at random from 50 in Figure 11a-6 to 200 in Figure 11b-6." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 227, + 509, + 363, + 609 + ], + "blocks": [ + { + "bbox": [ + 242, + 487, + 337, + 507 + ], + "lines": [ + { + "bbox": [ + 242, + 487, + 337, + 507 + ], + "spans": [ + { + "bbox": [ + 242, + 487, + 337, + 507 + ], + "type": "text", + "content": "(b-2) ADACUR Round 2. Total 80 anchor items." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 509, + 363, + 609 + ], + "lines": [ + { + "bbox": [ + 227, + 509, + 363, + 609 + ], + "spans": [ + { + "bbox": [ + 227, + 509, + 363, + 609 + ], + "type": "image", + "image_path": "91cb2168fd0e13c5f46a97a1cf8b351be75c6bef9bb087d07f8c1b307b09695e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 616, + 338, + 635 + ], + "lines": [ + { + "bbox": [ + 240, + 616, + 338, + 635 + ], + "spans": [ + { + "bbox": [ + 240, + 616, + 338, + 635 + ], + "type": "text", + "content": "(b-5) ADACUR Round 5. Total 200 anchor items." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 384, + 509, + 520, + 609 + ], + "blocks": [ + { + "bbox": [ + 401, + 488, + 498, + 507 + ], + "lines": [ + { + "bbox": [ + 401, + 488, + 498, + 507 + ], + "spans": [ + { + "bbox": [ + 401, + 488, + 498, + 507 + ], + "type": "text", + "content": "(b-3) ADACUR Round 3. Total 120 anchor items." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 384, + 509, + 520, + 609 + ], + "lines": [ + { + "bbox": [ + 384, + 509, + 520, + 609 + ], + "spans": [ + { + "bbox": [ + 384, + 509, + 520, + 609 + ], + "type": "image", + "image_path": "1c713cf03417e4fb66d9c05c60dfa6c63b995cc5ed78cb34af3aec20da7ad11f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 616, + 523, + 636 + ], + "lines": [ + { + "bbox": [ + 383, + 616, + 523, + 636 + ], + "spans": [ + { + "bbox": [ + 383, + 616, + 523, + 636 + ], + "type": "text", + "content": "(b-6) ANNCUR- Sampling all 200 anchor items uniformly at random." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8099" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 164, + 250, + 431, + 270 + ], + "blocks": [ + { + "bbox": [ + 164, + 250, + 431, + 270 + ], + "lines": [ + { + "bbox": [ + 164, + 250, + 431, + 270 + ], + "spans": [ + { + "bbox": [ + 164, + 250, + 431, + 270 + ], + "type": "image", + "image_path": "548aaf1aef290dfdacc82ec3453e94d19934334b277e2c81e44aa9a524cae0d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 274, + 524, + 416 + ], + "blocks": [ + { + "bbox": [ + 69, + 274, + 524, + 416 + ], + "lines": [ + { + "bbox": [ + 69, + 274, + 524, + 416 + ], + "spans": [ + { + "bbox": [ + 69, + 274, + 524, + 416 + ], + "type": "image", + "image_path": "5cf1e21353df666ec04e7cb036f03c5d321ba5d2f3e051d562d6f3b7b30942d5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 423, + 319, + 432 + ], + "lines": [ + { + "bbox": [ + 274, + 423, + 319, + 432 + ], + "spans": [ + { + "bbox": [ + 274, + 423, + 319, + 432 + ], + "type": "text", + "content": "(a) SciDocs" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 445, + 524, + 567 + ], + "blocks": [ + { + "bbox": [ + 70, + 445, + 524, + 567 + ], + "lines": [ + { + "bbox": [ + 70, + 445, + 524, + 567 + ], + "spans": [ + { + "bbox": [ + 70, + 445, + 524, + 567 + ], + "type": "image", + "image_path": "45dcacc22c42b998ee9475d79959dc13350f58d2d3abc69aab19a7a49d6e1f3d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 574, + 322, + 584 + ], + "lines": [ + { + "bbox": [ + 271, + 574, + 322, + 584 + ], + "spans": [ + { + "bbox": [ + 271, + 574, + 322, + 584 + ], + "type": "text", + "content": "(b) HotpotQA" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "lines": [ + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "type": "text", + "content": "Figure 12: Top-" + }, + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "type": "text", + "content": "-Recall for ADACUR (using ten rounds) and baselines for SciDocs and HotpotQA, " + }, + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 1000" + }, + { + "bbox": [ + 67, + 593, + 524, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8100" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 154, + 235, + 171 + ], + "blocks": [ + { + "bbox": [ + 71, + 154, + 235, + 171 + ], + "lines": [ + { + "bbox": [ + 71, + 154, + 235, + 171 + ], + "spans": [ + { + "bbox": [ + 71, + 154, + 235, + 171 + ], + "type": "image", + "image_path": "26c367d591eafc828e8ec3fa8bb01cd893bbee9e643b5dc7c63da4191b66d281.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 155, + 511, + 171 + ], + "blocks": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "type": "image", + "image_path": "dc1c1e07649c52381c6155f65fb7df2f8875ce3804e9d64c99eed8b365d16e0a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 155, + 179, + 439, + 194 + ], + "blocks": [ + { + "bbox": [ + 155, + 179, + 439, + 194 + ], + "lines": [ + { + "bbox": [ + 155, + 179, + 439, + 194 + ], + "spans": [ + { + "bbox": [ + 155, + 179, + 439, + 194 + ], + "type": "image", + "image_path": "7caefe3aca3ccd9ae6909d884802e7775f61ed061657b1062175088d48d7c369.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 210, + 242, + 319 + ], + "blocks": [ + { + "bbox": [ + 70, + 210, + 242, + 319 + ], + "lines": [ + { + "bbox": [ + 70, + 210, + 242, + 319 + ], + "spans": [ + { + "bbox": [ + 70, + 210, + 242, + 319 + ], + "type": "image", + "image_path": "818abc361746554d903cacdd8f337f48e3f46664230fc529991d037c924a878f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "lines": [ + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "spans": [ + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 211, + 400, + 319 + ], + "blocks": [ + { + "bbox": [ + 249, + 211, + 400, + 319 + ], + "lines": [ + { + "bbox": [ + 249, + 211, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 249, + 211, + 400, + 319 + ], + "type": "image", + "image_path": "d85e84b09a763817d072b1320fc6f98d929a283c552f7a2d373aade6d32f5a8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 211, + 523, + 319 + ], + "blocks": [ + { + "bbox": [ + 408, + 211, + 523, + 319 + ], + "lines": [ + { + "bbox": [ + 408, + 211, + 523, + 319 + ], + "spans": [ + { + "bbox": [ + 408, + 211, + 523, + 319 + ], + "type": "image", + "image_path": "718beb75cf87b581b7282721bb0e92e396ecfb0cc0c05c3d2c521272c71fc0c0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 363, + 242, + 470 + ], + "blocks": [ + { + "bbox": [ + 70, + 363, + 242, + 470 + ], + "lines": [ + { + "bbox": [ + 70, + 363, + 242, + 470 + ], + "spans": [ + { + "bbox": [ + 70, + 363, + 242, + 470 + ], + "type": "image", + "image_path": "d6e18a9b73534f56dda5f095c391c89e3f4f6e3866314e916c6bf26f78afbd68.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "lines": [ + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "spans": [ + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 363, + 401, + 470 + ], + "blocks": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "lines": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "spans": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "type": "text", + "content": "(a) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 100" + }, + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 363, + 401, + 470 + ], + "lines": [ + { + "bbox": [ + 249, + 363, + 401, + 470 + ], + "spans": [ + { + "bbox": [ + 249, + 363, + 401, + 470 + ], + "type": "image", + "image_path": "b52b342d648f393aa2e434f9775554dc850e0ae0e26e91aa834841b00fdd1e30.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 363, + 523, + 470 + ], + "blocks": [ + { + "bbox": [ + 408, + 363, + 523, + 470 + ], + "lines": [ + { + "bbox": [ + 408, + 363, + 523, + 470 + ], + "spans": [ + { + "bbox": [ + 408, + 363, + 523, + 470 + ], + "type": "image", + "image_path": "3d070858727996c8cd8553c21c7b517d98f2b9454e6e322df036bb0080c755d1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 70, + 515, + 242, + 623 + ], + "blocks": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "lines": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "spans": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "text", + "content": "(b) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 515, + 242, + 623 + ], + "lines": [ + { + "bbox": [ + 70, + 515, + 242, + 623 + ], + "spans": [ + { + "bbox": [ + 70, + 515, + 242, + 623 + ], + "type": "image", + "image_path": "a6a2b60dedf4b8e1d94ed34aa0e62b2ab92873291cb31172520e3e25183e3266.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "lines": [ + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "spans": [ + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "lines": [ + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "spans": [ + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "text", + "content": "(c) Number of train/anchor queries " + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 2000" + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "lines": [ + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "type": "text", + "content": "Figure 13: Top-" + }, + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "type": "text", + "content": "-Recall for ADACUR (using five rounds) and baselines for domain=YuGiOh. Each subfigure corresponds to a different value of the number of train/anchor queries " + }, + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "type": "inline_equation", + "content": "(|\\mathcal{Q}_{\\text {train }}|)" + }, + { + "bbox": [ + 67, + 661, + 524, + 686 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 249, + 515, + 401, + 623 + ], + "blocks": [ + { + "bbox": [ + 249, + 515, + 401, + 623 + ], + "lines": [ + { + "bbox": [ + 249, + 515, + 401, + 623 + ], + "spans": [ + { + "bbox": [ + 249, + 515, + 401, + 623 + ], + "type": "image", + "image_path": "2df33e476e3c134a616b3611da7f4056eabe4e5274a16dc9280a4ebbc2f20314.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 408, + 515, + 523, + 623 + ], + "blocks": [ + { + "bbox": [ + 408, + 515, + 523, + 623 + ], + "lines": [ + { + "bbox": [ + 408, + 515, + 523, + 623 + ], + "spans": [ + { + "bbox": [ + 408, + 515, + 523, + 623 + ], + "type": "image", + "image_path": "af420a45981dc937114ec3b8c1f2a0d3ed76f0faf7cc6bf350cac5551de96d7f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "8101" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 155, + 235, + 171 + ], + "blocks": [ + { + "bbox": [ + 71, + 155, + 235, + 171 + ], + "lines": [ + { + "bbox": [ + 71, + 155, + 235, + 171 + ], + "spans": [ + { + "bbox": [ + 71, + 155, + 235, + 171 + ], + "type": "image", + "image_path": "8fc6aa09b47562a568286c43a5f0751d4fb8bc7f5291a76d9b167ab6e089face.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 155, + 511, + 171 + ], + "blocks": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 301, + 155, + 511, + 171 + ], + "type": "image", + "image_path": "193b5ce2fb71560cc198e7b315a9220074d6fc389adc07e07aae0978a957e76c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 155, + 179, + 439, + 195 + ], + "blocks": [ + { + "bbox": [ + 155, + 179, + 439, + 195 + ], + "lines": [ + { + "bbox": [ + 155, + 179, + 439, + 195 + ], + "spans": [ + { + "bbox": [ + 155, + 179, + 439, + 195 + ], + "type": "image", + "image_path": "3ccf1511fc31c8bffc0db6503a9a735c57c7b1c983c5aa3a265f67efb26a1192.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 210, + 243, + 319 + ], + "blocks": [ + { + "bbox": [ + 70, + 210, + 243, + 319 + ], + "lines": [ + { + "bbox": [ + 70, + 210, + 243, + 319 + ], + "spans": [ + { + "bbox": [ + 70, + 210, + 243, + 319 + ], + "type": "image", + "image_path": "fb751c61a7f6d3a7e7cf0ad1ecb4df49b04148a5c343609e6abc21aebb322ceb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "lines": [ + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "spans": [ + { + "bbox": [ + 227, + 319, + 430, + 331 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 249, + 210, + 402, + 319 + ], + "blocks": [ + { + "bbox": [ + 249, + 210, + 402, + 319 + ], + "lines": [ + { + "bbox": [ + 249, + 210, + 402, + 319 + ], + "spans": [ + { + "bbox": [ + 249, + 210, + 402, + 319 + ], + "type": "image", + "image_path": "828d4053b0ba61338871c695d331ff7aca3f1b66d92a776827f31a029b8335f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 211, + 524, + 319 + ], + "blocks": [ + { + "bbox": [ + 408, + 211, + 524, + 319 + ], + "lines": [ + { + "bbox": [ + 408, + 211, + 524, + 319 + ], + "spans": [ + { + "bbox": [ + 408, + 211, + 524, + 319 + ], + "type": "image", + "image_path": "3daf09084dc3fd2f6ee54d843d7c2232e38e8c1031de72eaf0ea1706244f0a15.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 363, + 243, + 470 + ], + "blocks": [ + { + "bbox": [ + 70, + 363, + 243, + 470 + ], + "lines": [ + { + "bbox": [ + 70, + 363, + 243, + 470 + ], + "spans": [ + { + "bbox": [ + 70, + 363, + 243, + 470 + ], + "type": "image", + "image_path": "4d5d66fe6e128567aa6fa9720657c0dd8bbcfdc4d1170bcc4b12bc5016a5e669.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "lines": [ + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "spans": [ + { + "bbox": [ + 227, + 470, + 430, + 483 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 363, + 402, + 470 + ], + "blocks": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "lines": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "spans": [ + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "type": "text", + "content": "(a) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 337, + 387, + 349 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 100" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 363, + 402, + 470 + ], + "lines": [ + { + "bbox": [ + 249, + 363, + 402, + 470 + ], + "spans": [ + { + "bbox": [ + 249, + 363, + 402, + 470 + ], + "type": "image", + "image_path": "0a4caec315ba0439acceea4f6791efb451cff55db854029ea953990aa1d7f449.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 363, + 524, + 470 + ], + "blocks": [ + { + "bbox": [ + 408, + 363, + 524, + 470 + ], + "lines": [ + { + "bbox": [ + 408, + 363, + 524, + 470 + ], + "spans": [ + { + "bbox": [ + 408, + 363, + 524, + 470 + ], + "type": "image", + "image_path": "7eb4b41eff8084365b22a320c07bb9cfb45ff4ad494d696bebf723d3a0898d45.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 70, + 515, + 243, + 623 + ], + "blocks": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "lines": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "spans": [ + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "text", + "content": "(b) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 205, + 489, + 387, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 515, + 243, + 623 + ], + "lines": [ + { + "bbox": [ + 70, + 515, + 243, + 623 + ], + "spans": [ + { + "bbox": [ + 70, + 515, + 243, + 623 + ], + "type": "image", + "image_path": "4d44ed728ec4c5448ce05911874c41e131047aa70bac2767856963857438068d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "lines": [ + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "spans": [ + { + "bbox": [ + 227, + 624, + 430, + 635 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "lines": [ + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "spans": [ + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "text", + "content": "(c) Number of train/anchor queries " + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 2000" + }, + { + "bbox": [ + 203, + 641, + 389, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "lines": [ + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "type": "text", + "content": "Figure 14: Top-" + }, + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "type": "text", + "content": "-Recall for ADACUR (using five rounds) and baselines for domain=StarTrek. Each subfigure corresponds to a different value of the number of train/anchor queries " + }, + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "type": "inline_equation", + "content": "(|\\mathcal{Q}_{\\mathrm{train}}|)" + }, + { + "bbox": [ + 67, + 661, + 525, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 249, + 515, + 402, + 623 + ], + "blocks": [ + { + "bbox": [ + 249, + 515, + 402, + 623 + ], + "lines": [ + { + "bbox": [ + 249, + 515, + 402, + 623 + ], + "spans": [ + { + "bbox": [ + 249, + 515, + 402, + 623 + ], + "type": "image", + "image_path": "3cfe699bb790ada1a2c1459da3fb3442c5b77988509b09690d430cd92b4758a0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 408, + 515, + 524, + 623 + ], + "blocks": [ + { + "bbox": [ + 408, + 515, + 524, + 623 + ], + "lines": [ + { + "bbox": [ + 408, + 515, + 524, + 623 + ], + "spans": [ + { + "bbox": [ + 408, + 515, + 524, + 623 + ], + "type": "image", + "image_path": "a4c791d390c0c11ac9c0a7b5bd8fce0a25e27d766bbc18098428af1f976724a2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "8102" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 144, + 234, + 159 + ], + "blocks": [ + { + "bbox": [ + 73, + 144, + 234, + 159 + ], + "lines": [ + { + "bbox": [ + 73, + 144, + 234, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 144, + 234, + 159 + ], + "type": "image", + "image_path": "c206c7baf8658830cb6aca78021beeabcfdaf3716eb97a831b6d877f8126306c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 169, + 226, + 179 + ], + "lines": [ + { + "bbox": [ + 155, + 169, + 226, + 179 + ], + "spans": [ + { + "bbox": [ + 155, + 169, + 226, + 179 + ], + "type": "text", + "content": "Proposed Approaches" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 144, + 511, + 159 + ], + "blocks": [ + { + "bbox": [ + 301, + 144, + 511, + 159 + ], + "lines": [ + { + "bbox": [ + 301, + 144, + 511, + 159 + ], + "spans": [ + { + "bbox": [ + 301, + 144, + 511, + 159 + ], + "type": "image", + "image_path": "ad691f5806c7f6d529484dade0d1c51d78bbef1f6ddc415fd2c4692dcbee40ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 232, + 169, + 253, + 178 + ], + "blocks": [ + { + "bbox": [ + 232, + 169, + 253, + 178 + ], + "lines": [ + { + "bbox": [ + 232, + 169, + 253, + 178 + ], + "spans": [ + { + "bbox": [ + 232, + 169, + 253, + 178 + ], + "type": "image", + "image_path": "8de34fc979ea5757c0a473b97caa7406fb3c16e48de28e3fe577910e597b8c5f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 169, + 286, + 178 + ], + "lines": [ + { + "bbox": [ + 260, + 169, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 260, + 169, + 286, + 178 + ], + "type": "text", + "content": "DACUR" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 292, + 169, + 309, + 178 + ], + "blocks": [ + { + "bbox": [ + 292, + 169, + 309, + 178 + ], + "lines": [ + { + "bbox": [ + 292, + 169, + 309, + 178 + ], + "spans": [ + { + "bbox": [ + 292, + 169, + 309, + 178 + ], + "type": "image", + "image_path": "9eb61d65e7a5b6446841ec364adc76aea5fcdb36c74fc808337d25f9d337e7a4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 169, + 326, + 178 + ], + "lines": [ + { + "bbox": [ + 312, + 169, + 326, + 178 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 326, + 178 + ], + "type": "text", + "content": "ADA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 169, + 341, + 178 + ], + "blocks": [ + { + "bbox": [ + 328, + 169, + 341, + 178 + ], + "lines": [ + { + "bbox": [ + 328, + 169, + 341, + 178 + ], + "spans": [ + { + "bbox": [ + 328, + 169, + 341, + 178 + ], + "type": "image", + "image_path": "4951cb502194006a395ca30868bb9fa756a46dc61b61cb996d4cf40edb4e2d37.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 169, + 358, + 178 + ], + "lines": [ + { + "bbox": [ + 343, + 169, + 358, + 178 + ], + "spans": [ + { + "bbox": [ + 343, + 169, + 358, + 178 + ], + "type": "text", + "content": "R" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 362, + 169, + 380, + 178 + ], + "blocks": [ + { + "bbox": [ + 362, + 169, + 380, + 178 + ], + "lines": [ + { + "bbox": [ + 362, + 169, + 380, + 178 + ], + "spans": [ + { + "bbox": [ + 362, + 169, + 380, + 178 + ], + "type": "image", + "image_path": "1f78ac0be2fd48a2926be69ec03e15a9ca7470dd5981c96d479d6ddee346f069.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 401, + 169, + 418, + 178 + ], + "blocks": [ + { + "bbox": [ + 384, + 169, + 398, + 178 + ], + "lines": [ + { + "bbox": [ + 384, + 169, + 398, + 178 + ], + "spans": [ + { + "bbox": [ + 384, + 169, + 398, + 178 + ], + "type": "text", + "content": "AD" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 401, + 169, + 418, + 178 + ], + "lines": [ + { + "bbox": [ + 401, + 169, + 418, + 178 + ], + "spans": [ + { + "bbox": [ + 401, + 169, + 418, + 178 + ], + "type": "image", + "image_path": "4fc45c793e7d2ade24038526682a09f9f4d7be1585cbe99bfeefcb500e4b18ba.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 169, + 434, + 178 + ], + "lines": [ + { + "bbox": [ + 421, + 169, + 434, + 178 + ], + "spans": [ + { + "bbox": [ + 421, + 169, + 434, + 178 + ], + "type": "text", + "content": "\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 439, + 169, + 456, + 178 + ], + "blocks": [ + { + "bbox": [ + 439, + 169, + 456, + 178 + ], + "lines": [ + { + "bbox": [ + 439, + 169, + 456, + 178 + ], + "spans": [ + { + "bbox": [ + 439, + 169, + 456, + 178 + ], + "type": "image", + "image_path": "6415aa68c1a31c13c8a53ba8cdb651ddcd9e37c1f185196b2992612b787c365c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 169, + 478, + 178 + ], + "lines": [ + { + "bbox": [ + 459, + 169, + 478, + 178 + ], + "spans": [ + { + "bbox": [ + 459, + 169, + 478, + 178 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 481, + 169, + 509, + 178 + ], + "blocks": [ + { + "bbox": [ + 481, + 169, + 509, + 178 + ], + "lines": [ + { + "bbox": [ + 481, + 169, + 509, + 178 + ], + "spans": [ + { + "bbox": [ + 481, + 169, + 509, + 178 + ], + "type": "image", + "image_path": "7f87e08e8f83543348c1f535aba8374d15184fa31a41d7861c771b6f1813d7cb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 70, + 199, + 242, + 307 + ], + "blocks": [ + { + "bbox": [ + 70, + 199, + 242, + 307 + ], + "lines": [ + { + "bbox": [ + 70, + 199, + 242, + 307 + ], + "spans": [ + { + "bbox": [ + 70, + 199, + 242, + 307 + ], + "type": "image", + "image_path": "452c47ed38e428ebd5a84e12d9c181c4689ea1fd63e8c6ccf9b8522184a24444.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 307, + 430, + 318 + ], + "lines": [ + { + "bbox": [ + 227, + 307, + 430, + 318 + ], + "spans": [ + { + "bbox": [ + 227, + 307, + 430, + 318 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 249, + 200, + 401, + 306 + ], + "blocks": [ + { + "bbox": [ + 249, + 200, + 401, + 306 + ], + "lines": [ + { + "bbox": [ + 249, + 200, + 401, + 306 + ], + "spans": [ + { + "bbox": [ + 249, + 200, + 401, + 306 + ], + "type": "image", + "image_path": "a88d57a0c64c3b47a74ad1044c677bb43b2b6a059d6c42c3492d9f663714474e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 408, + 200, + 523, + 306 + ], + "blocks": [ + { + "bbox": [ + 408, + 200, + 523, + 306 + ], + "lines": [ + { + "bbox": [ + 408, + 200, + 523, + 306 + ], + "spans": [ + { + "bbox": [ + 408, + 200, + 523, + 306 + ], + "type": "image", + "image_path": "5fcea90f2cff7cbce75777cb63906624f696c247e9cf6b32c6d458ba0b940280.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 70, + 351, + 242, + 459 + ], + "blocks": [ + { + "bbox": [ + 70, + 351, + 242, + 459 + ], + "lines": [ + { + "bbox": [ + 70, + 351, + 242, + 459 + ], + "spans": [ + { + "bbox": [ + 70, + 351, + 242, + 459 + ], + "type": "image", + "image_path": "103460186e44d634f0671448bba078f34bebad947e0c8d9725279e21e224a9ae.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 460, + 430, + 471 + ], + "lines": [ + { + "bbox": [ + 227, + 460, + 430, + 471 + ], + "spans": [ + { + "bbox": [ + 227, + 460, + 430, + 471 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 249, + 351, + 401, + 459 + ], + "blocks": [ + { + "bbox": [ + 205, + 325, + 387, + 337 + ], + "lines": [ + { + "bbox": [ + 205, + 325, + 387, + 337 + ], + "spans": [ + { + "bbox": [ + 205, + 325, + 387, + 337 + ], + "type": "text", + "content": "(a) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 325, + 387, + 337 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 100" + }, + { + "bbox": [ + 205, + 325, + 387, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 351, + 401, + 459 + ], + "lines": [ + { + "bbox": [ + 249, + 351, + 401, + 459 + ], + "spans": [ + { + "bbox": [ + 249, + 351, + 401, + 459 + ], + "type": "image", + "image_path": "1b208771490ef453d2aad2814f1eb473e6b6534912dc4c840f768069b16f62e8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 408, + 351, + 523, + 459 + ], + "blocks": [ + { + "bbox": [ + 408, + 351, + 523, + 459 + ], + "lines": [ + { + "bbox": [ + 408, + 351, + 523, + 459 + ], + "spans": [ + { + "bbox": [ + 408, + 351, + 523, + 459 + ], + "type": "image", + "image_path": "c982832636752109ab23ef035405faab9f68d52bac0b76386fab6c27176d98b8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 70, + 503, + 242, + 611 + ], + "blocks": [ + { + "bbox": [ + 205, + 477, + 387, + 489 + ], + "lines": [ + { + "bbox": [ + 205, + 477, + 387, + 489 + ], + "spans": [ + { + "bbox": [ + 205, + 477, + 387, + 489 + ], + "type": "text", + "content": "(b) Number of train/anchor queries " + }, + { + "bbox": [ + 205, + 477, + 387, + 489 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 500" + }, + { + "bbox": [ + 205, + 477, + 387, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 503, + 242, + 611 + ], + "lines": [ + { + "bbox": [ + 70, + 503, + 242, + 611 + ], + "spans": [ + { + "bbox": [ + 70, + 503, + 242, + 611 + ], + "type": "image", + "image_path": "203513f44c2614a26a273351a33a23b528c2a3ec0bd168a0e9ce411312851fa0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 612, + 430, + 623 + ], + "lines": [ + { + "bbox": [ + 227, + 612, + 430, + 623 + ], + "spans": [ + { + "bbox": [ + 227, + 612, + 430, + 623 + ], + "type": "text", + "content": "Inference Cost (Total Number of Cross-Encoder Calls)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 629, + 389, + 641 + ], + "lines": [ + { + "bbox": [ + 202, + 629, + 389, + 641 + ], + "spans": [ + { + "bbox": [ + 202, + 629, + 389, + 641 + ], + "type": "text", + "content": "(c) Number of train/anchor queries " + }, + { + "bbox": [ + 202, + 629, + 389, + 641 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_{\\mathrm{train}}| = 2000" + }, + { + "bbox": [ + 202, + 629, + 389, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "lines": [ + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "text", + "content": "Figure 15: Top-" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "text", + "content": "-Recall for ADACUR (using five rounds) and baselines for domain=Military. Each subfigure corresponds to a different value of the number of train/anchor queries " + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "inline_equation", + "content": "(|\\mathcal{Q}_{train}|)" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "text", + "content": " has high Top-1-Recall values as domain=Military is included in the set of train domains in ZESHEL which are used to train both " + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "inline_equation", + "content": "\\mathrm{DE}_{\\mathrm{BASE}}" + }, + { + "bbox": [ + 67, + 650, + 526, + 698 + ], + "type": "text", + "content": " and the cross-encoder model." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 249, + 503, + 401, + 611 + ], + "blocks": [ + { + "bbox": [ + 249, + 503, + 401, + 611 + ], + "lines": [ + { + "bbox": [ + 249, + 503, + 401, + 611 + ], + "spans": [ + { + "bbox": [ + 249, + 503, + 401, + 611 + ], + "type": "image", + "image_path": "c70668bc95c575f3ab747d5b95f4885237498be6a8b9e037a6f116d7ef56b89b.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 408, + 503, + 523, + 611 + ], + "blocks": [ + { + "bbox": [ + 408, + 503, + 523, + 611 + ], + "lines": [ + { + "bbox": [ + 408, + 503, + 523, + 611 + ], + "spans": [ + { + "bbox": [ + 408, + 503, + 523, + 611 + ], + "type": "image", + "image_path": "1c3751bc076387a24ad37899f74b4a3f7f59313542929440e5c428a7dd1965d2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "8103" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_content_list.json b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7281e076f048a34383cc99570422b8410bec0b62 --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_content_list.json @@ -0,0 +1,2702 @@ +[ + { + "type": "text", + "text": "Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt", + "text_level": 1, + "bbox": [ + 134, + 89, + 867, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seonghyeon Ye Joel Jang Doyoung Kim Yongrae Jo Minjoon Seo KAIST", + "bbox": [ + 181, + 154, + 821, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{seonghyeon.ye,joeljang,ikevin98,yongrae,minjoon}@kaist.ac.kr", + "bbox": [ + 201, + 189, + 801, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 342, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enhancing the zero-shot performance of instruction-following models requires heavy computation, either by scaling the total number of training datasets or the model size. In this work, we explore how retrieval of soft prompts obtained through prompt tuning can efficiently assist hard prompts in zero-shot task generalization. Specifically, we train soft prompt embeddings for each prompt through prompt tuning, store the samples of the training instances mapped with the prompt embeddings, and retrieve the corresponding prompt embedding of the training instance closest to the query instance during inference. While only adding $0.007\\%$ additional parameters, retrieval of soft prompt enhances the performance of T0 on unseen tasks by outperforming it on 10 out of 11 datasets as well as improving the mean accuracy of T0 on BIG-bench benchmark by $2.39\\%$ points. Also, we report an interesting finding that retrieving source embeddings trained on similar answer choice formats is more important than those on similar task types.", + "bbox": [ + 141, + 284, + 460, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 627, + 260, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training Large Language Models (LLMs) on huge amounts of data has enabled LMs to perform downstream tasks without any fine-tuning with the aid of natural prompts or concatenation of a few demonstration instances (Brown et al., 2020; Rae et al., 2021; Kojima et al., 2022; Chowdhery et al., 2022). Additionally, recent works have shown that adding a instruction tuning stage, an additional training step that helps pretrained LMs understand prompts and demonstrations results in a significant performance boost on zero-shot task generalization even for moderate-sized LMs (Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Ye et al., 2022; Chung et al., 2022). This extra", + "bbox": [ + 112, + 655, + 489, + 881 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f317a13ed48bf7b1b98fcbabeec569fbb45be92a01ee06ccd631a599e8f6e40.jpg", + "image_caption": [ + "Figure 1: During zero-shot inference, ROSPR selects similar training instances with the given input from the Source Prompt Library and retrieves the prompt embeddings corresponding to the selected training instances." + ], + "image_footnote": [], + "bbox": [ + 510, + 250, + 884, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instruction-tuning stage involves explicit, multitask prompted learning on various tasks, enabling LMs to quickly adapt to unseen tasks at inference.", + "bbox": [ + 507, + 419, + 882, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To maximize the effect of instruction-tuning, two approaches have been widely explored: (1) scaling the number of training datasets, and (2) scaling the model size (Wang et al., 2022b; Chung et al., 2022). However, both approaches require heavy computation, not applicable with an academic budget. Specifically, the first approach requires updating the whole parameters of the model every time a training dataset is added, showing limitations in terms of scalability. On the other hand, the second approach requires heavy memory requirements to load and train a massive LLM.", + "bbox": [ + 507, + 468, + 884, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance the zero-shot performance of instruction-following model efficiently, we introduce Retrieval of Soft Prompt (ROSPR), which is easily scalable and requires minimal computation by only adding $0.007\\%$ parameters to the main model during inference. As shown in Figure 1, by training prompt embeddings (soft prompt) for each given hard prompt through prompt tuning, we construct a Source Prompt Library consisting of samples of training instances mapped with their corresponding prompt embeddings. Then, during inference, by using a simple, off-the-shelf dense retriever model, we search for training instances similar to the given query instances and retrieve their corresponding prompt embeddings. Because the backbone LM is frozen, it allows the retrieved", + "bbox": [ + 507, + 661, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Model checkpoints and code implementation are available at github.com/seonghyeonye/RoSPr.", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12288", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12288-12309 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 208, + 945, + 786, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "embeddings to serve as adapters assisting hard prompts. While RoSPR can be applied to any LM, in this work, we use T0 (Sanh et al., 2021) as our initial backbone LM and perform prompt tuning on the tasks used during the instruction-tuning stage.", + "bbox": [ + 112, + 84, + 487, + 164 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While adding only $0.007\\%$ additional parameters, ROSPR outperforms T0 on 10 out of 11 evaluation datasets and outperforms efficient fine-tuning baselines without any target task fine-tuning. ROSPR is also effective for challenging tasks such as tasks from BIG-bench (Srivastava et al., 2022), outperforming T0 by $2.39\\%$ mean accuracy. Furthermore, we provide several interesting findings: (1) Variants of ROSPR that include interpolation of multiple prompt embeddings and scoring method that considers the answer choice distribution during retrieval further increases the effect of ROSPR (2) Also, we provide analysis of which factors attribute to the performance of ROSPR and show that, similarly to the role of demonstrations in in-context learning (Min et al., 2022), heuristic features such as answer choice format are more important than the similarity of the source task.", + "bbox": [ + 115, + 165, + 489, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 112, + 467, + 270, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Task Generalization with Instruction-Tuning", + "text_level": 1, + "bbox": [ + 112, + 494, + 357, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompts and demonstrations are essential for task generalization since proper explanations are required for LMs to understand an unseen task (Kojima et al., 2022; Wei et al., 2022; Lampinen et al., 2022). Instruction-tuning, which is explicit multi-task prompted training on various downstream tasks, is a simple but effective way to achieve this, resulting in improved zero-shot capabilities. Zhong et al. (2021) first introduced the method of instruction-tuning by converting various tasks into a question-answering format and finetuning the model on the aggregated dataset. Following works (Mishra et al., 2022; Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Xu et al., 2022; Ouyang et al., 2022; Ye et al., 2022; Chung et al., 2022) extended this approach on a larger scale and show that zero-shot task generalization could be enhanced with more diverse prompts, a larger number of training downstream tasks, and a larger LM.", + "bbox": [ + 112, + 532, + 489, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Source Task Retrieval", + "text_level": 1, + "bbox": [ + 112, + 865, + 332, + 879 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Retrieving a source task that is relevant to the target task has shown to result in faster and better task", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adaptation. For parameter-efficient fine-tuning, Vu et al. (2022); Su et al. (2022) retrieve source prompt embedding that is similar to the target prompt embedding and obtain a better initialization point for prompt tuning. Instead of utilizing a single prompt embedding, recent works show a mixture of multiple prompt embeddings to be effective (Asai et al., 2022; Qin and Eisner, 2021).", + "bbox": [ + 507, + 84, + 884, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For instruction-tuning, Lin et al. (2022) retrieve training instances that are similar to the query through a dense retriever and fine-tune the model using the retrieved examples. For in-context learning, Rubin et al. (2021); Liu et al. (2022b); Wang et al. (2023) retrieve training data that could be used for demonstrations. Wang et al. (2022c) show the effect of retrieving prompt embeddings in a continual learning setting. Although our proposed method is related to these works, the novelty of our work lies in applying source task retrieval in the zero-shot setting and retrieving soft prompts instead of training instances.", + "bbox": [ + 507, + 212, + 884, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Method", + "text_level": 1, + "bbox": [ + 509, + 434, + 613, + 449 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce Retrieval of Prompt Tuning (ROSPR) for zero-shot task generalization. A detailed overview is shown in Figure 2. We first train source prompt embeddings of LM for each hard prompt given a source task using prompt tuning (Section 3.1). Then, we save training instance samples along with their prompt embeddings in the Source Prompt Library and use it to retrieve embeddings at inference to perform tasks in a zero-shot manner (Section 3.2). We additionally introduce interpolation of multiple source prompt embeddings (Section 3.3) and variance-based ranking (Section 3.4) to increase robustness and accuracy.", + "bbox": [ + 507, + 460, + 882, + 668 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Training Source Prompt Embeddings", + "text_level": 1, + "bbox": [ + 507, + 680, + 850, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even though RoSPR may be used to augment any type of LM, we use T0 (Sanh et al., 2021) as the backbone LM for this paper. For training of soft prompts, we utilize the source tasks and prompts used for the instruction-tuning phase of T0. While T0 was trained in a multi-task learning manner, we freeze the initial T0 parameters and train only soft prompts (source prompt embeddings) for each hard prompt of the source task.", + "bbox": [ + 507, + 701, + 882, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Tuning Among various parameter-efficient fine-tuning methods, we follow prompt tuning proposed by Lester et al. (2021) because the number of trainable parameters is extremely small", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12289", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d9561a1f59a0b1caf4a6392d148c33d0c316d8439c27366ca7da20fc3a899fb.jpg", + "image_caption": [ + "Figure 2: An overview of RoSPR. For each hard prompt of the source datasets, soft prompts are trained via prompt tuning. After storing training instances as keys and corresponding prompt embedding as values, RoSPR searches training instances similar to query set $Q$ , retrieves the corresponding prompt embeddings, and selects the most frequently retrieved candidate for inference. Variants of selection strategy are also shown: RoSPR+INTER interpolates between multiple related source embeddings and RoSPR+VAR ranks candidate embeddings considering both frequency and variance." + ], + "image_footnote": [], + "bbox": [ + 171, + 84, + 843, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(\\sim 204\\mathrm{K}$ parameters per prompt), which implies that the memory overhead of parameter retrieval at inference is negligible.", + "bbox": [ + 112, + 338, + 487, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each source training dataset $D_{i}$ $(i = 1,\\dots,T)$ where $T$ is the total number of source datasets, we train source embeddings $E_{ij}$ $(j = 1,\\dots,M_i)$ where $M_{i}$ is the number of hard prompts in $D_{i}$ , making soft prompt embeddings for each individual hard prompts. Specifically, given a training instance $\\{x_{ik},y_{ik}\\} (k = 1,\\dots,K)$ from $D_{i}$ where $K$ is the number of sampled training instances per dataset, we first convert it into its hard prompted version $\\{h_j(x_{ik}),h_j(y_{ik})\\}$ where $h_j(\\cdot)$ denotes adding the $j$ -th hard prompt. Next, we train the LM with the following objective:", + "bbox": [ + 112, + 388, + 489, + 583 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {E _ {i j}} P \\left(h _ {j} \\left(y _ {i k}\\right) \\mid E _ {i j}; h _ {j} \\left(x _ {i k}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 598, + 487, + 624 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where all the parameters of the underlying backbone LM are frozen and only $E_{ij}$ is trainable. In short, given $D_{i}$ , we perform $M_{i}$ number of prompt tunings for each hard prompts, resulting in $\\sum_{i=1}^{T} M_{i}$ total number of source prompt embeddings. For training efficiency, we only train on $K = 5000$ training instances for a single epoch for each source prompt embedding.", + "bbox": [ + 112, + 637, + 489, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Zero-Shot Embedding Retrieval", + "text_level": 1, + "bbox": [ + 112, + 780, + 413, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After source prompt embedding training, we retrieve the most related source embeddings and select one from the retrieved candidates to be used at inference (right part of Figure 2).", + "bbox": [ + 112, + 804, + 489, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first construct a Source Prompt Library, consisting of sentence-level representations of training instance inputs as keys and the corresponding source prompt embedding as the values. For each available source prompt embedding, $n$ number of samples are stored in the library. The sentence-level representations are obtained by getting the mean representation of hidden states of the last layer of the dense retriever. We use a T0-small encoder as a dense retriever, replicated based on Sanh et al. (2021) with smaller model size.", + "bbox": [ + 507, + 340, + 884, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At inference, we first randomly sample $Q$ query instances from the target task, following Lin et al. (2022). After obtaining sentence-level representations for each query through our T0-small encoder, we retrieve top- $N$ examples for each query instance using MIPS (maximum inner product search) operation on our Source Prompt Library, retrieving a total of $Q*N$ prompt embeddings. As the default methodology, among the retrieved embedding candidates, we select the most frequently retrieved prompt embedding as our designated soft prompt for the given target task and concatenate the embedding with each of the target task instances before feeding it to our backbone LM. In the next two subsections, we explain different strategies for calculating the target embedding from the $Q*N$ prompt embedding candidates.", + "bbox": [ + 507, + 517, + 885, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Interpolation of Prompt Embeddings", + "text_level": 1, + "bbox": [ + 507, + 800, + 847, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When retrieving only a single prompt embedding for a given task (Section 3.2), it may result in high variance across evaluation prompts when the selected prompt embedding does not fit well with the", + "bbox": [ + 507, + 820, + 882, + 885 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3For all indexing and searching, we use FAISS (Johnson et al., 2019) for fast source prompt embedding retrieval.", + "bbox": [ + 507, + 891, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For each instances, the input and output are converted into its prompted version using the promptsource toolkit (Bach et al., 2022). An example is given in Appendix G.", + "bbox": [ + 112, + 879, + 487, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12290", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "given task. Recent works on prompt embedding retrieval have shown that the interpolation of prompt embeddings effectively transfers to the target task (Asai et al., 2022; Vu et al., 2022). We also explore calculating the target embedding through interpolation of multiple source embeddings instead of just using a single embedding. Among $Q * N$ prompt candidates searched in Section 3.2, we select top- $N'$ candidate embeddings based on the frequency of the search. Then, we calculate the weighted sum of the candidate embeddings, where the interpolation weight for each source embedding is based on the proportion of frequency. While Asai et al. (2022); Vu et al. (2022) require fine-tuning the target embeddings on the target task to calculate the weights for interpolation, our approach does not require any target task fine-tuning, enabling zero-shot task transfer.", + "bbox": [ + 112, + 84, + 492, + 374 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4 Variance-based Ranking", + "text_level": 1, + "bbox": [ + 112, + 387, + 354, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the scoring and calibration method of Lu et al. (2022); Zhao et al. (2021), we introduce a scoring method applicable to zero-shot classification tasks that allows ranking the $Q * N$ retrieved prompt embedding candidates by considering the answer choice distribution of the given target task as extra cues together with the original frequency cues. To accomplish this, we perform a forward pass with the concatenation of each candidate prompt embeddings together with the given hard prompt (only including the instruction, excluding the input instance) of the target task and give a higher score to the embedding candidate that results in lower variance. Ideally, the combination of soft and hard prompts should result in equal probability among the answer choices because the actual context of the task is not included.", + "bbox": [ + 112, + 407, + 489, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, when given a target task with $k$ -th hard prompt $h_k$ , for each candidate embedding $E_{ij}$ we calculate the scoring as follows:", + "bbox": [ + 112, + 683, + 489, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S c o r e} \\left(h _ {k}, E _ {i j}\\right) = \\frac {\\operatorname {f r e q} \\left(h _ {k} , E _ {i j}\\right)}{\\sqrt {\\operatorname {V a r} \\left[ P \\left(y \\mid E _ {i j} , h _ {k}\\right) \\right]}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 141, + 743, + 487, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y$ refers to the available output options of the target task.", + "bbox": [ + 112, + 797, + 487, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Experimental Settings", + "text_level": 1, + "bbox": [ + 112, + 844, + 341, + 860 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we explain the experimental settings of training of source prompt embedding and construction of our Source Prompt Library. We also", + "bbox": [ + 112, + 871, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "explain our evaluation setting during zero-shot inference and baseline models. We provide detailed experiment configurations in Appendix F.", + "bbox": [ + 507, + 84, + 884, + 133 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 Source Tasks", + "text_level": 1, + "bbox": [ + 507, + 145, + 660, + 158 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For training soft prompts through prompt tuning, we use the subset of source tasks used for the initial T0 instruction-tuning (Sanh et al., 2021) $^4$ . For each source task, we use the prompts for each dataset in T0, resulting in a total of 230 prompts. For Source Prompt Library construction, we sample only $n = 100$ training instances per source embedding to minimize the inference latency. We show a variation of $n$ and different methods to sample $n$ training instances in Appendix D.", + "bbox": [ + 507, + 166, + 884, + 326 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Evaluation Tasks", + "text_level": 1, + "bbox": [ + 507, + 338, + 690, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following Sanh et al. (2021), we evaluate on the validation set of 4 held-out tasks (natural language inference, sentence completion, coreference resolution, word sense disambiguation) resulting in a total of 11 evaluation datasets. We also follow Sanh et al. (2021) and evaluate on 14 different datasets from the BIG-bench benchmark (Srivastava et al., 2022). We use rank classification evaluation method by selecting the output option with higher log-likelihood following Brown et al. (2020); Sanh et al. (2021). For all evaluation tasks, we use accuracy as an evaluation metric and report the mean accuracy and standard deviation of all of the evaluation prompts per given dataset (average of $\\sim 10$ prompts per evaluation dataset). For BIG-bench tasks, we do not report standard deviation because only one prompt is provided per task.", + "bbox": [ + 507, + 360, + 884, + 634 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 Baseline Models", + "text_level": 1, + "bbox": [ + 507, + 645, + 684, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero-shot Baseline For zero-shot baseline models, we show the results of T0 (3B) together with a 4 times larger T0 (11B) instruction-tuned model. We also compare with GPT-3 (175B) model which is 60 times larger than T0 (3B).", + "bbox": [ + 507, + 667, + 882, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tuning Baseline We also compare with efficient fine-tuning baseline models that utilize prompt tuning. These models require target task prompt tuning, indicating that zero-shot transfer", + "bbox": [ + 507, + 757, + 882, + 822 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "4We use 29 out of 38 datasets that are used to train T0. We explain the training task selection rationale in Appendix H.1", + "bbox": [ + 507, + 831, + 882, + 854 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5We provide the full list of evaluation datasets in Appendix H.", + "bbox": [ + 510, + 856, + 882, + 879 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For methods based on RoSPR, we report the performance average of 3 runs with different random seeds for the sampling of evaluation queries used for the prompt retrieval.", + "bbox": [ + 510, + 881, + 880, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12291", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1393e1b953f1ef5f4db6e8065fc94b4d7459c55de8397bfe54e16aa16a79fee8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method# of Param\n(Base/Trainable)NLISentence CompletionCoreference Resolut.WSDTotal Avg.
RTECBAN. R1AN. R2AN. R3COPAHellasw.StoryC.Winogr.WSCWiCMeanSTD
T03B / 064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.223.62
PT (FT)3B / 204K63.1444.5233.0731.4432.9473.0026.39-50.6746.7350.02--
ATTEMPT (FT)3B / 614K68.9544.8836.1934.7334.8174.3826.76-51.3364.9051.05--
T0+RoSPR3B / 204K71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.243.62
w/ INTER3B / 204K70.7152.3037.3034.3433.8978.2526.9485.6251.1064.5250.7353.243.30
w/ VAR3B / 204K71.7850.3637.0734.5833.9078.8827.0185.5251.4564.9451.9453.383.38
w/ VAR & INTER3B / 204K72.6051.9837.2534.3133.9577.8326.8485.5850.9364.9751.1853.403.47
ORACLE3B / 204K73.7958.1037.6534.9234.9181.1327.7587.5752.3668.1755.2655.603.07
T011B / 080.8370.1243.5638.6841.2690.0233.5892.4059.9461.4556.5860.76
GPT-3175B / 063.5046.4034.6035.4034.5091.0078.9083.2070.2065.4045.9259.00-
", + "bbox": [ + 114, + 80, + 882, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: RoSPR refers to our main proposed method, W/ INTER refers to applying interpolation of multiple source embedding candidates, W/ VAR refers to retrieval through variance-based ranking, W/ VAR & INTER refers to applying both interpolation and variance-based ranking where the interpolation weight is based on the variance-based ranking score, and ORACLE refers performance when the most optimal source embedding is retrieved from the candidates, acting as an upper bound performance for retrieval. FT refers to fine-tuned models on the target tasks. For FT models, we exclude StoryCloze due to the absence of training instances. The best and second-best performance is shown in bold and underline respectively. Comparison with hard prompt optimization techniques and visualization of the results is shown in Appendix A and Appendix E, respectively.", + "bbox": [ + 110, + 230, + 882, + 317 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is infeasible. Similar to our source prompt tuning process, we train each target prompt for a single epoch with a maximum of 5,000 training instances. The first baseline model is naive prompt tuning on the target tasks without any prompt retrieval, referred to as PT (Lester et al., 2021). The second model is ATTEMPT (Asai et al., 2022), which trains the target soft prompts through attentional mixtures of source prompts. Because StoryCloze (Mostafazadeh et al., 2016) does not contain training instances, we exclude the dataset for fine-tuning. More training details of fine-tuning baseline are specified in Appendix C.", + "bbox": [ + 110, + 329, + 489, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 112, + 552, + 334, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RoSPR enhances the performance of T0 efficiently. Table 1 shows the performance of the 11 evaluation datasets. T0+RoSPR outperforms T0 on 10 datasets among the 11 evaluation datasets. Specifically, T0+RoSPR outperforms T0 on RTE (+6.99% points), CB (+4.12% points), ANLI R1 (+3.24% points), and COPA (+2.87% points). This shows that soft prompt retrieval assists hard prompts for zero-shot task generalization with a negligible number of additional parameters (0.007%). Also, while T0 outperforms GPT-3 on 3 datasets (RTE, StoryCloze, WiC), T0+RoSPR additionally outperforms GPT-3 on 2 datasets (ANLI R1 and CB) and enlarging the score gap for RTE, StoryCloze and WiC.", + "bbox": [ + 110, + 580, + 489, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ROSPR also outperforms finetuning baselines even without utilizing any training instances of the target task. We first observe that PT harms the performance of the backbone model, which aligns with the result of Liu et al. (2022a); Gu et al. (2022) that prompt tuning is unstable when the training instances or the training steps are small. By comparing ATTEMPT with ROSPR, ROSPR outperforms ATTEMPT on 7 out of 10 tasks and $1.21\\%$ points on the mean accuracy of 10 tasks. This shows that ROSPR is more applicable for efficient adaptation because it requires 3 times fewer additional parameters compared to ATTEMPT and also does not require any further fine-tuning of the target task.", + "bbox": [ + 507, + 329, + 884, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "INTER and VAR enhance the performance of ROSPR. We also analyze the effect of introducing variants of ROSPR: interpolation of soft prompts (INTER) and variance-based ranking (VAR) in Table 1. First, applying INTER shows similar accuracy compared to ROSPR. However, as shown in the last column of Table 1, INTER reduces the standard deviation of T0 and ROSPR by $8.84\\%$ while improving the mean accuracy of T0, indicating increased robustness to different surface forms of evaluation prompts. This indicates that interpolation of multiple source embeddings outperforms a single source embedding retrieval, aligning with the result of Asai et al. (2022). Applying VAR with $\\mathrm{T0 + ROSPR}$ improves both zero-shot accuracy and robustness of $\\mathrm{T0 + ROSPR}$ , showing that considering the answer choice distribution is beneficial for zero-shot setting, aligned with results from Zhao et al. (2021); Shi et al. (2022). Moreover, applying both VAR+INTER results in the highest overall average accuracy, outperforming T0 by $2.18\\%$ points by largely reducing the gap between larger LLMs.", + "bbox": [ + 507, + 565, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "One exception is WSC, which is a binary classification task (yes/no) predicting whether the reference of the pronoun is correct. We observed that the evaluation data of this dataset has unbalanced labels, containing over $60\\%$ of \"No\" labels. This might be the reason why T0-11B underperforms T0-3B only on this dataset (Sanh et al., 2021). Indeed, predicting only \"No\" on this dataset outperforms T0-11B $(63.46 > 61.45)$ .", + "bbox": [ + 112, + 832, + 487, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12292", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ac012d05c905fb7395d0a8393846221f227d53d80a09f3bab98b34d3d8e99bf.jpg", + "image_caption": [ + "Figure 3: Mean accuracy of 14 datasets of BIG-bench. We evaluate on a single prompt following Sanh et al. (2021). By only adding $0.007\\%$ parameters to T0-3B, $\\mathrm{T0 + ROSPR}$ largely reduces the performance gap between 4 times larger T0-11B. The full result is provided in Appendix B." + ], + "image_footnote": [], + "bbox": [ + 136, + 83, + 467, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The effect of ROSPR generalizes to challenging tasks. ROSPR is also effective for challenging tasks such as tasks from BIG-bench benchmark. As shown in Figure 3, T0+ROSPR improves the mean accuracy performance of T0-3B by $2.39\\%$ points while only adding $0.007\\%$ additional parameters. T0+ROSPR also outperforms 60 times larger zero-shot and 1-shot GPT-3 and largely reduces the performance gap between 4 times larger T0-11B (1.84% points) or 60 times larger 3-shot GPT-3 (0.53% points). Applying INTER with T0+ROSPR results in additional mean accuracy enhancement, outperforming T0-3B by $2.67\\%$ points8.", + "bbox": [ + 112, + 298, + 489, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6 Analysis of ROSPR", + "text_level": 1, + "bbox": [ + 112, + 519, + 315, + 533 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot task adaptation of LMs is often seen as a problem of task location, locating the target task to where the model can solve it using the intrinsic ability obtained at pretraining stage with the aid of prompts and demonstrations (Reynolds and McDonell, 2021). In this section, we analyze which factors contribute to the performance enhancement in the perspective of identifying better task location. We find that although the target task performance depends on the source task types, heuristic features such as the answer choice format are more important. This agrees with previous findings that a met-trained LM focuses on simple features such as the label space, the input distribution, and sequence format, instead of complex semantics (Webson and Pavlick, 2021; Min et al., 2022).", + "bbox": [ + 112, + 543, + 489, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Target task performance depends on source task types. To analyze the effect of different source task types on each target task, we measure the frequency ratio of each source task type that results", + "bbox": [ + 112, + 809, + 489, + 873 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f31b568c9169f5ed17a16f3d6271df6eb735e813c92b9945d226e18762868bc8.jpg", + "image_caption": [ + "Figure 4: Frequency of source task types (x-axis) that maximizes (i.e. ORACLE) the accuracy of each target task (y-axis)." + ], + "image_footnote": [], + "bbox": [ + 529, + 82, + 865, + 221 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in the best performance (ORACLE) for the given prompts of the target tasks (visualized in Figure 4). From this figure, we can observe a few patterns: paraphrase task assists NLI and word sense disambiguation task while multi-choice QA (MQA) task assists sentence completion task. For coreference resolution task, various source task types (paraphrase, summarization, multi-choice QA) assist the target task.", + "bbox": [ + 507, + 266, + 884, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e471b75fad8028d974258ce484f8e07e0b8f08de5928b1e837b9038f781e0359.jpg", + "image_caption": [ + "Figure 5: Effect of answer choice format alignment across different target datasets (RTE, COPA, WiC). We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + ], + "image_footnote": [], + "bbox": [ + 515, + 419, + 870, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Answer choice format is important for task location. We also analyze the effect of using different answer choice formats with the same source task. Answer choice format decides how the available answer choices are given to the LM through the input. For example, a prompt that requires classifying a movie review into good/bad has a different answer choice format from classifying it into positive/negative.", + "bbox": [ + 505, + 598, + 884, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment on 3 datasets (RTE, COPA, WiC) which correspond to different tasks (NLI, sentence completion, word sense disambiguation) respectively. For each dataset, we select a source dataset that is retrieved the most for ORACLE. Among the source prompts of the selected source dataset, we select a prompt that has the same answer choice format as the target task (ALIGNED) and another prompt that has a different answer choice format (MISALIGNED). Figure 5 shows the effect of answer choice format alignment on the target task", + "bbox": [ + 507, + 741, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "8We observe that applying VAR results in the same performance as T0+ROSPR because frequency of retrieval has much more influence than variance for BIG-bench tasks.", + "bbox": [ + 112, + 879, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12293", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "performance by comparing ALIGNED and MIS-ALIGNED. The result shows that for all 3 datasets, ALIGNED significantly outperforms MISALIGNED. This result is non-trivial considering that the two prompt embeddings are trained on the same source training dataset and the same training configuration, with the only difference in the given answer choice format, implying that how the answer choices are given to solve a specific task is more important than the content of the training data for task location. ROSPR is mostly comparable to ALIGNED embedding, implying that retrieving a source prompt embedding by searching for similar input instances results in retrieving a source embedding with similar answer choice formats.", + "bbox": [ + 112, + 84, + 492, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/273c898d6a37ac95ab5400e985e36458d79a5b3199a888f1e549dd322b7aa5ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 332, + 430, + 443 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a0360ca8488b3b4ce587546cde7130424806ceb3a518be9d3e8b43cdc6ad70e.jpg", + "image_caption": [ + "Figure 6: RTE (Top) and WiC (Bottom) evaluation result by retrieving MISALIGNED and ALIGNED answer choice format across various source tasks. We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + ], + "image_footnote": [], + "bbox": [ + 183, + 443, + 430, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We additionally analyze the effect of answer choice formats on RTE and WiC datasets by retrieving prompt embeddings trained on various source tasks. Both target datasets have the answer choices of either yes/no. Similar to the previous experiments, we retrieve ALIGNED (yes/no format) and MISALIGNED prompt embeddings across three source tasks: paraphrase, sentiment classification, and multi-choice QA. As shown in Figure 6, for both target datasets, ALIGNED outperforms MISALIGNED across all three source tasks. This shows that aligning to the answer choice format of the target task is crucial regardless of the retrieved source task.", + "bbox": [ + 112, + 621, + 489, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Answer choice format is more important than task similarity. From Figure 6, we can see that all three source tasks benefit from aligning to the target task answer choice format. One may think", + "bbox": [ + 112, + 854, + 489, + 920 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that embeddings from source tasks requiring similar knowledge to the target task may be important. Counterintuitively, for both RTE and WiC target tasks, when the answer choice format is aligned, the task source embedding of sentiment classification, which is known to be irrelevant to RTE and WiC (Pruksachatkun et al., 2020), outperforms other embeddings sourced from datasets that are more relevant to the target datasets (paraphrase and multi-choice QA) (Appendix G). This implies that for retrieval of source embedding for task location, answer choice format is more important than containing similar knowledge required to solve the target task.", + "bbox": [ + 507, + 84, + 884, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Role of ROSPR is similar to in-context learning. From the findings explained in previous paragraphs, we can conclude that although the source task types influence the target task performance, retrieving a similar answer choice format is more important for task location. Indeed, source tasks containing similar knowledge can help target tasks only if the answer choice formats are aligned to the target task. These findings support Min et al. (2022); Webson and Pavlick (2021) that a meta-trained LM \"takes less effort\" to understand the input: models exploit the simple aspects of prompts and demonstrations such as the format and distribution instead of complex semantics. Especially, for in-context learning, Xie et al. (2021); Min et al. (2022) show that the role of demonstrations lies in providing the shared concept and distribution hints of the target task. From this aspect, the role of ROSPR is similar to demonstrations. However, it is more efficient than including demonstrations because it avoids heavy computation at inference from long sequence lengths (Liu et al., 2022a; Choi et al., 2022) since ROSPR prepends a fixed length of prefix tokens regardless of the task. Also, ROSPR is free from the instability of in-context learning coming from different orderings of demonstrations (Lu et al., 2022; Zhao et al., 2021). Lastly, we conjecture that ROSPR also has the benefits of soft prompts (Li and Liang, 2021) such as having more expressiveness.", + "bbox": [ + 507, + 318, + 884, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7 Ablation Studies", + "text_level": 1, + "bbox": [ + 507, + 813, + 687, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze the effect of the number of (1) prompts, (2) source datasets, and (3) queries sampled for evaluation. We evaluate variations of our proposed methods on 4 datasets: RTE for NLI, COPA for sentence completion, Winogrande", + "bbox": [ + 507, + 839, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12294", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1defe8e68d4ed49a58d8f3c1201254db1ffa3563b1d5ecf359d4894f1cdb96b9.jpg", + "image_caption": [ + "(a) Datasets vs Prompts", + "Figure 7: Result of various ablation settings of ROSPR. (a) compares the effect of scaling the number of datasets with scaling the number of prompts and (b) shows the effect of the number of sampled queries at inference. Additional ablation results are shown in Appendix D." + ], + "image_footnote": [], + "bbox": [ + 131, + 84, + 292, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f71eb27c7fedf771f1e614bb3dba7a2722033716596e727b6f5cae0fcd0d36bf.jpg", + "image_caption": [ + "(b) Query variation" + ], + "image_footnote": [], + "bbox": [ + 307, + 80, + 470, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for coreference, and WiC for word sense disambiguation task. We report the average of the mean accuracy of all evaluation prompts for each dataset by running 3 different runs.", + "bbox": [ + 112, + 296, + 489, + 360 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling number of prompts vs. number of datasets. Recent works on meta-trained LMs show that the number of source datasets and prompts is an important factor for zero-shot task generalization (Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Chung et al., 2022). We also show ablations for ROSPR and measure how the zero-shot generalization performance changes when we vary the number of prompts and datasets available during the prompt tuning stage (shown in Figure 7a). First, we vary (1) the total number of source prompts by 60, 120, and 230 by increasing the number of prompts per dataset and (2) the number of datasets by 8, 16, and 30 by increasing the number of datasets per task cluster. Note that in (1), the total number of datasets is fixed while in (2), we use all available prompts for each dataset while varying the number of datasets per task cluster.", + "bbox": [ + 115, + 368, + 487, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In contrast to (1), (2) does not always lead to a linearly increasing performance boost; the performance saturates as more source datasets are included. By comparing the effect of scaling datasets and scaling prompts for similar Source Prompt Library sizes, we observe that the number of prompts has more impact on the accuracy of the target task (Figure 7a).", + "bbox": [ + 112, + 658, + 487, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This ablation study also supports the analysis of the previous section; diverse answer choice formats of prompts, which are mostly influenced by the total number of source prompts, are more important than source task types which are influenced by the number of source datasets.[10] Therefore, if", + "bbox": [ + 112, + 787, + 487, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the number of task clusters is sufficient to some extent, scaling the number of source prompts per dataset is more crucial than scaling the number of source datasets per task cluster.", + "bbox": [ + 507, + 84, + 882, + 149 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More sampled queries improve the performance. We also analyze the effect of the number of query instances sampled at inference for retrieval. As seen in Figure 7b, increasing the number of queries results in higher mean accuracy. This is different from the analysis of Lin et al. (2022) that sampling more queries leads to better performance only to some point. Because we use the frequency of each prompt embedding candidate as the default metric for retrieval, utilizing more query instances would represent the evaluation data more accurately, resulting in a reduced number of wrong retrievals.", + "bbox": [ + 507, + 158, + 884, + 350 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 363, + 640, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce ROSPR, a method that efficiently enhances zero-shot generalization capabilities of a meta-trained LM by retrieving prompt-specific source prompt embeddings (soft prompts) for a given target task. We accomplish this by first training the soft prompts for hard prompt of the source tasks. After training source prompt embeddings, we construct the Source Prompt Library by storing the mean representation of training instances as keys and the corresponding prompt embeddings as values. At inference, we search for training instances stored in the library similar to sample instances from the target task, retrieve the corresponding prompt embedding, select the most frequently retrieved embedding, and append it to each of the target task instances for prediction. Our results show that ROSPR efficiently enhances the zero-shot performance of the backbone model while introducing minimal additional parameters during inference. We additionally provide analysis of which factors attribute to the performance of ROSPR and find that heuristic cues such as the answer choice format are critical for generalization performance, implying that it may play a role similar to demonstrations in in-context learning.", + "bbox": [ + 507, + 388, + 882, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9 Limitations", + "text_level": 1, + "bbox": [ + 507, + 803, + 645, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although we show the effectiveness of ROSPR by applying it on T0-3B (Sanh et al., 2021), we did", + "bbox": [ + 507, + 829, + 882, + 862 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "the number of datasets increases, we find that answer choice formats are similar across the same task type, meaning that the diversity of answer choice formats is not increased by increasing the number of datasets per task clusters.", + "bbox": [ + 507, + 869, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "${}^{9}$ Task cluster is defined as a cluster of the same task types.", + "bbox": [ + 132, + 891, + 485, + 904 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "10 Although the total number of prompts also increases as", + "bbox": [ + 132, + 904, + 485, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12295", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not evaluate our method on different model scales such as the T0-11B variant and other LM architectures such as decoder-only LMs due to limited computational resources. This leaves future works on applying ROSPR to even larger LMs and diverse LM architectures (Wang et al., 2022a). Moreover, it is hard to apply VAR to target tasks without answer choices such as free-form generation because variance among options cannot be obtained. However, ROSPR and ROSPR+INTER can still be utilized and we leave applying ROSPR on zero-shot task location of free-form generation as future work (Scialom et al., 2022).", + "bbox": [ + 112, + 84, + 492, + 294 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 304, + 285, + 322 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Sohee Yang and Eunbi Choi for helpful feedback on the paper. This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00264, Comprehensive Video Understanding and Generation with Knowledge-based Deep Logic Neural Network, $80\\%$ ; No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST), $20\\%$ ).", + "bbox": [ + 112, + 329, + 490, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 514, + 213, + 531 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Akari Asai, Mohammadreza Salehi, Matthew E Peters, and Hannaneh Hajishirzi. 2022. Attentional mixtures of soft prompt tuning for parameter-efficient multi-task knowledge sharing. arXiv preprint arXiv:2205.11961.", + "Stephen Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-david, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Fries, Maged Alshaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang, Dragomir Radev, Mike Tian-jian Jiang, and Alexander Rush. 2022. Prompt-Source: An integrated development environment and repository for natural language prompts. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 93-104, Dublin, Ireland. Association for Computational Linguistics.", + "Max Bartolo, Alastair Roberts, Johannes Welbl, Sebastian Riedel, and Pontus Stenetorp. 2020. Beat the AI: Investigating adversarial human annotation for reading comprehension. Transactions of the Association for Computational Linguistics, 8:662-678.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind" + ], + "bbox": [ + 115, + 537, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.", + "Eunbi Choi, Yongrae Jo, Joel Jang, and Minjoon Seo. 2022. Prompt injection: Parameterization of fixed inputs.", + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311.", + "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416.", + "Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine learning challenges workshop*, pages 177-190. Springer.", + "Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107-124.", + "Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 70–79, Hong Kong, China. Association for Computational Linguistics.", + "David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2003. English gigaword. Linguistic Data Consortium, Philadelphia, 4(1):34.", + "Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. 2022. PPT: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, Dublin, Ireland. Association for Computational Linguistics.", + "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2391-2401, Hong Kong, China. Association for Computational Linguistics.", + "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12296", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916.", + "Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. 2022. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329.", + "Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1203-1213, Austin, Texas. Association for Computational Linguistics.", + "Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, D. Kontokostas, Pablo N. Mendes, Sebastian Hellmann, M. Morsey, Patrick van Kleef, S. Auer, and C. Bizer. 2015. Dbpedia - a large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6:167-195.", + "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning.", + "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierric Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chonghua Liao, Yanan Zheng, and Zhilin Yang. 2022. Zero-label prompt selection.", + "Bill Yuchen Lin, Kangmin Tan, Chris Miller, Beiwen Tian, and Xiang Ren. 2022. Unsupervised cross-task generalization via retrieval augmentation. arXiv preprint arXiv:2204.07937.", + "Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823-1840, Online. Association for Computational Linguistics.", + "Kevin Lin, Oyvind Tafjord, Peter Clark, and Matt Gardner. 2019. Reasoning over paragraph effects in situations. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 58-62, Hong Kong, China. Association for Computational Linguistics.", + "Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022a. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. arXiv preprint arXiv:2205.05638.", + "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022b. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics.", + "Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics.", + "Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.", + "Julian J. McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In Seventh ACM Conference on Recommender Systems, RecSys '13, Hong Kong, China, October 12-16, 2013, pages 165-172. ACM.", + "Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12297", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettle-moyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837.", + "Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470-3487, Dublin, Ireland. Association for Computational Linguistics.", + "Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849, San Diego, California. Association for Computational Linguistics.", + "Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797-1807, Brussels, Belgium. Association for Computational Linguistics.", + "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics.", + "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155.", + "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05), pages 115-124, Ann Arbor, Michigan. Association for Computational Linguistics.", + "Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Roktaschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544, Online. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, and Samuel R. Bowman. 2020. Intermediate-task transfer learning with pretrained language models: When and why does it work? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5231-5247, Online. Association for Computational Linguistics.", + "Guanghui Qin and Jason Eisner. 2021. Learning how to ask: Querying LMs with mixtures of soft prompts. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5203-5212, Online. Association for Computational Linguistics.", + "Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446.", + "Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932-4942, Florence, Italy. Association for Computational Linguistics.", + "Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In *Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems*, pages 1-7.", + "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95.", + "Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. 2020. Getting closer to ai complete question answering: A set of prerequisite real tasks. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8722-8731.", + "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12298", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amrita Saha, Rahul Aralikatte, Mitesh M. Khapra, and Karthik Sankaranarayanan. 2018. DuoRC: Towards complex language understanding with paraphrased reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683-1693, Melbourne, Australia. Association for Computational Linguistics.", + "Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106.", + "Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207.", + "Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social IQa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4463-4473, Hong Kong, China. Association for Computational Linguistics.", + "Thomas Scialom, Tuhin Chakrabarty, and Smaranda Muresan. 2022. Continual-t0: Progressively instructing $50+$ tasks to language models without forgetting. arXiv preprint arXiv:2205.12393.", + "Weijia Shi, Julian Michael, Suchin Gururangan, and Luke Zettlemoyer. 2022. Nearest neighbor zero-shot inference. arXiv preprint arXiv:2205.13792.", + "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615.", + "Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Huadong Wang, Kaiyue Wen, Zhiyuan Liu, Peng Li, Juanzi Li, Lei Hou, Maosong Sun, and Jie Zhou. 2022. On transferability of prompt tuning for natural language processing. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3949-3969, Seattle, United States. Association for Computational Linguistics.", + "Niket Tandon, Bhavana Dalvi, Keisuke Sakaguchi, Peter Clark, and Antoine Bosselut. 2019. WIQA: A dataset for \"what if...\" reasoning over procedural text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6076-6085, Hong Kong, China. Association for Computational Linguistics.", + "Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou', and Daniel Cer. 2022. SPoT: Better frozen model adaptation through soft prompt transfer. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5039-5059, Dublin, Ireland. Association for Computational Linguistics.", + "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics.", + "Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. 2022a. What language model architecture and pretraining objective work best for zero-shot generalization? arXiv preprint arXiv:2204.05832.", + "Xinyi Wang, Wanrong Zhu, and William Yang Wang. 2023. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. arXiv preprint arXiv:2301.11916.", + "Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. 2022b. Benchmarking generalization via in-context instructions on 1,600+ language tasks. arXiv preprint arXiv:2204.07705.", + "Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022c. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149.", + "Albert Webson and Ellie Pavlick. 2021. Do prompt-based models really understand the meaning of their prompts? arXiv preprint arXiv:2109.01247.", + "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652.", + "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "12299", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 94-106, Copenhagen, Denmark. Association for Computational Linguistics.", + "Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing datasets for multi-hop reading comprehension across documents. Transactions of the Association for Computational Linguistics, 6:287-302.", + "Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080.", + "Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. 2022. Zero-prompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot generalization. arXiv preprint arXiv:2201.06910.", + "Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013-2018, Lisbon, Portugal. Association for Computational Linguistics.", + "Seonghyeon Ye, Doyoung Kim, Joel Jang, Joongbo Shin, and Minjoon Seo. 2022. Guess the instruction! making language models stronger zero-shot learners. arXiv preprint arXiv:2210.02969.", + "Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics.", + "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015a. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc.", + "Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015b. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 649-657.", + "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR.", + "Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2856-2878, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers." + ], + "bbox": [ + 510, + 85, + 880, + 316 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "12300", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A Comparison with Hard Prompt Optimization", + "text_level": 1, + "bbox": [ + 114, + 83, + 425, + 118 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We have conducted additional experiments to compare our method with 3 different hard prompt optimization techniques including (1) RoHPr which utilizes the same search technique as RoSPr but retrieves 4 corresponding training instances for incontext learning instead of the corresponding soft prompt, and (2) APE (Zhou et al., 2023) which is an automatic prompt engineering method that utilizes the generation results of an LLM (Ouyang et al., 2022), known to show better performance than human instructions, and (3) ZPS (Liao et al., 2022) which selects an optimal hard prompt from a prompt pool using prompt ensemble and pseudolabels. As shown in the result of Table 2, on average, RoSPr performs the best on average, showing the benefits of using soft prompts over hard prompts for task generalization. While RoSPr shows improvement on 10 out of 11 datasets, other hard prompt optimization methods do not show consistent improvements.", + "bbox": [ + 115, + 124, + 489, + 445 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B Full Result of BIG-bench Evaluation", + "text_level": 1, + "bbox": [ + 112, + 457, + 468, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We provide the task generalization performance result of 14 tasks from BIG-bench, shown in Table 3. Applying RoSPR largely improves the performance for 3 datasets (Hindu Knowledge, Novel Concepts, Misconceptions): $(+17.72\\%, +3.12\\%, +1.82\\%)$ compared to T0-3B and $(+13.72\\%, +3.12\\%, +2.05\\%)$ compared to T0-11B. For mean accuracy of 14 tasks, T0+RoSPR outperforms T0-3B by $2.39\\%$ points, reducing the gap between 4 times larger T0-11B to $1.84\\%$ points. Moreover, applying INTER to T0+RoSPR enhances the performance of T0+RoSPR for most tasks, indicating that interpolation of multiple embeddings is effective for challenging tasks.", + "bbox": [ + 112, + 482, + 489, + 706 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "C Fine-tuning Baseline Details", + "text_level": 1, + "bbox": [ + 112, + 717, + 394, + 734 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "For fine-tuning baseline models (PT and ATTEMPT), we follow the training configuration of source prompt tuning. We train each target prompt with a single epoch using a maximum of 5,000 training instances. For fine-tuning, we use a batch size of 32 and a learning rate of 1e-3. Also, we randomly select hard prompts (templates) of the training dataset during fine-tuning. For ATTEMPT, we randomly sample one source prompt per task cluster, resulting in interpolation between 8 soft prompts.", + "bbox": [ + 112, + 741, + 489, + 919 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D Additional Ablation Studies", + "text_level": 1, + "bbox": [ + 509, + 83, + 789, + 99 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/7c6ab7b6dea7392b5cf3dd29e262296a2f1dd12b367774760af9b83d72bb884f.jpg", + "image_caption": [ + "Figure 8: Variation of number of prompts by increasing the number of prompts per dataset." + ], + "image_footnote": [], + "bbox": [ + 568, + 122, + 823, + 288 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3a71c461961e3e29507a80cc7186c63a26d336cb5ea8305c590954fc1f19f36b.jpg", + "image_caption": [ + "Figure 9: Variation of number of datasets by increasing the number of datasets per task cluster." + ], + "image_footnote": [], + "bbox": [ + 569, + 362, + 821, + 530 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We provide detailed result of variation of the number of prompts (Figure 8) and the number of datasets (Figure 9). We additionally analyze the effect of (1) different sampling methods for constructing the Source Prompt Library, (2) the number of instances sampled for constructing the Source Prompt Library, (3) the number of top-N retrieval for embedding retrieval, and (4) the number of multiple source embeddings to interpolate. Same as Section 7, we report the mean accuracy of 4 evaluation datasets: RTE, COPA, Winogrande, and WiC with 3 runs with different random seeds for the sampling of evaluation queries.", + "bbox": [ + 507, + 592, + 882, + 800 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "D.1 Sampling Methods for Source Prompt Library", + "text_level": 1, + "bbox": [ + 507, + 816, + 858, + 848 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We experiment three different methods to sample instances for constructing Source Prompt Library and analyze the effect of each method. By default, we choose RANDOM method, where we sample", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "12301", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/41b1fe2d2f5050236c422b05d0f62242a4d46df7221f212b558cfba016a35fc9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RTECBAN.R1AN.R2AN.R3COPAHellasw.StoryC.Winogr.WSCWiCAVG
T064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.22
RoHPr47.1045.0033.5333.3033.0973.3826.6586.9350.5164.4250.0049.45
APE (Zhou et al., 2023)68.1948.3335.8235.3734.2768.0025.7678.7350.4358.7550.8550.41
ZPS (Liao et al., 2022)58.4860.7136.6034.4033.3376.0028.4987.3951.7864.4250.6352.93
RoSPr71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.24
", + "bbox": [ + 119, + 80, + 885, + 164 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/29b67b33e7c0060b76579ce4e0d640616762d28fb771d14a43881287aef50a97.jpg", + "table_caption": [ + "Table 2: Comparison result of RoSPr with different hard prompt optimization techniques." + ], + "table_footnote": [], + "table_body": "
T0-3BROSPRINTERT0-11B
Strategy.52.7952.0552.2352.75
Movie D.52.8551.4552.2353.69
Known Un.47.8347.8347.8358.70
Logic Grid41.1037.0037.4038.30
Hindu Kn.25.7143.4345.7129.71
Code D.46.6745.0040.0043.33
Concept45.5267.6167.6169.29
Language14.8413.6814.4020.20
Vitamin58.8853.7154.5364.73
Syllogism52.9450.6451.3451.81
Misconcept.50.2352.0552.0550.00
Logical46.6454.8654.8654.86
Winowhy44.2944.3344.2952.11
Novel Con.15.6315.6318.7515.63
AVG42.5644.9545.2346.79
", + "bbox": [ + 124, + 210, + 478, + 458 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 3: Evaluation result of 14 tasks from BIG-bench (Srivastava et al., 2022).", + "bbox": [ + 112, + 468, + 487, + 495 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/a4fe1ea67a1f90c1d4b419df853c5a552f683205a91e522c4d57494a989f37af.jpg", + "image_caption": [ + "Figure 10: Different instance sampling methods for constructing Source Prompt Library." + ], + "image_footnote": [], + "bbox": [ + 117, + 512, + 482, + 627 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "100 instances by random for each prompt by assuming that each random 100 queries of instances can represent the whole prompt. Second, we experiment CLUSTERING method, where we sample top 100 instances which has the closest mean representation from its overall average for each prompt. If we say each instance has a distance of mean representation from its overall average as $d_{i}$ in increasing order, we sample $\\{d_1,d_2,\\dots,d_{100}\\}$ . The last method we use is DISTRIBUTed method, where we sample 100 instances in a distributed way with respect to its distance of mean representation from its overall average. If we say each instance has a distance of mean representation from its over", + "bbox": [ + 112, + 694, + 490, + 917 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "all average as $d_{i}$ in increasing order, we sample $\\{d_1,d_{1 + N / 100},d_{1 + 2*N / 100},\\dots ,d_{1 + 99*N / 100}\\}$ , assuming there are total $N$ training instances in a dataset.", + "bbox": [ + 507, + 212, + 884, + 275 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "As shown in Figure 10, RANDOM method outperforms CLUSTERING and DISTRIBUTED methods. Interestingly, CLUSTERING method significantly hurts the performance on all 4 proposed methods, suggesting that storing similar instances per prompt results in retrieval failures more often. Also for DISTRIBUTED method, most of the methods significantly underperform RANDOM, except INTER and ROSPR+VAR+INTER. From these results, we can conclude that random sampling represents the source dataset most effectively.", + "bbox": [ + 507, + 278, + 885, + 455 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/771e75538b33e4966cec3169ff4caa82aaf3409a9384edf93e1349e8c0eff70d.jpg", + "image_caption": [ + "Figure 11: Variation of number of instances sampled for constructing Source Prompt Library. Default setting is $n = 100$ ." + ], + "image_footnote": [], + "bbox": [ + 568, + 469, + 823, + 636 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "D.2 Number of Instances Sampled for Constructing Source Prompt Library", + "text_level": 1, + "bbox": [ + 507, + 720, + 850, + 753 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We analyze the effect of size of the Source Prompt Library by varying the number of instances $n$ to sample for each hard prompt by 100, 300, 500. Therefore, $n \\times$ (number of total hard prompts) would be the size of the Source Prompt Library. As shown in Figure 11, increasing the number of sampled instances does not increase the performance; it hurts the performance for most cases. This suggests that only a few number of training instances are enough to represent the distribution of prompted", + "bbox": [ + 505, + 757, + 884, + 919 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "12302", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 14 + }, + { + "type": "image", + "img_path": "images/8925a804240b359526e812738a5b7a830639cc6bf88c8f1706c4e4f2b9aa6341.jpg", + "image_caption": [ + "Figure 12: Variation of number of Top $N$ instances for embedding retrieval. Default setting is $N = 10$ ." + ], + "image_footnote": [], + "bbox": [ + 174, + 84, + 433, + 250 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "input (hard prompt + input instances) for each hard prompt and increasing the number sometimes hurt the performance by adding noise to the distribution. This also supports the importance of heuristic cues in Section 6 by showing that adding more training instances per hard prompt does not increase the performance. Instead, adding hard prompts with diverse answer choice format is more important.", + "bbox": [ + 112, + 318, + 489, + 448 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D.3 Number of Top N instances for Embedding Retrieval", + "text_level": 1, + "bbox": [ + 112, + 460, + 406, + 493 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We vary the number of top- $N$ instances that are retrieved given each query through MIPS search. As shown in Figure 12, varying the number of top- $N$ instances does not have much effect compared to increasing the number of sampled queries (Figure 7b). This implies that if the size of the evaluation set of the target task is large, sampling more queries is effective than searching for more similar instances per query. This is important for variance-based methods because the number of forward passes needed before evaluation is proportional to $Q*N$ . Therefore, we can reduce the time latency by reducing the number of instances retrieved per query without hurting the performance much.", + "bbox": [ + 112, + 498, + 489, + 724 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D.4 Number of Source Embeddings for Interpolation", + "text_level": 1, + "bbox": [ + 112, + 736, + 438, + 766 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We analyze the effect of number of source embeddings for interpolation by varying top- $N^{\\prime}$ interpolation from 1 (no interpolation) to 5 shown in Figure 13. By comparing between single prompt embedding retrieval $(N^{\\prime} = 1)$ and the interpolation of multiple embeddings $(N^{\\prime} > 1)$ , the mean accuracy drops by adding multiple source embeddings for retrieval because interpolation-based methods underperform on tasks such as COPA as shown", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/62f18d4144a73705322ce9eb6db64e09537df79b85f9c48ef1656b5ea55cc28a.jpg", + "image_caption": [ + "Figure 13: Variation of number of source embeddings for interpolation based methods. Default setting is $K' = 3$ ." + ], + "image_footnote": [], + "bbox": [ + 568, + 82, + 823, + 250 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "in Table 1. Mean accuracy would increase if we add other datasets for evaluation that benefit from interpolation such as WSC and CB.", + "bbox": [ + 507, + 331, + 882, + 378 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "By comparing among various $N'$ values, we find that for ROSPR+INTER, the accuracy substantially decreases for $K' = 2$ , implying that the possibility of a wrong retrieval varies depending on the value of $N'$ . In contrast, ROSPR+VAR+INTER is more robust to the value of $K'$ , showing that variance-based ranking increases robustness to different numbers of source embeddings for interpolation as well.", + "bbox": [ + 507, + 379, + 884, + 523 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "E Visualization of Results", + "text_level": 1, + "bbox": [ + 507, + 535, + 752, + 550 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We show the visualization of the evaluation result on 11 datasets in Figure 14. Methods based on ROSPR not only show higher accuracy, but lower variance for many datasets.", + "bbox": [ + 507, + 560, + 882, + 625 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "F Experimental Configurations", + "text_level": 1, + "bbox": [ + 507, + 636, + 800, + 653 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "As mentioned in the previous sections, we use T0-3B as our backbone meta-trained LM. For prompt tuning, we fix the prefix length as 100 and the embeddings are initialized from 5,000 most common vocabulary following Lester et al. (2021). We train each source embedding for a single epoch with a learning rate of 0.1 and a batch size of 32. We use the Adam optimizer with weight decay of 1e-5. For retrieval, we randomly sample $Q = 32$ query instances and retrieve top $K = 10$ examples for each query. We train a T0-small variant ( $\\sim 35\\mathrm{M}$ params) as our dense retriever by multitask prompted training on T5+LM model (Lester et al., 2021), replicating the original training setting of T0 by training T5+LM for 8 epochs using the same training instances of Sanh et al. (2021) with", + "bbox": [ + 507, + 661, + 884, + 919 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "12303", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 15 + }, + { + "type": "image", + "img_path": "images/f6b681d8de7ac8cdac12bf6b353e78bbe2965768c33bc4d4baf40abff101026f.jpg", + "image_caption": [ + "(a) RTE" + ], + "image_footnote": [], + "bbox": [ + 119, + 93, + 285, + 181 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/8edc524f24162a48e4833bbdad69b57897357326563c1f4eaea143f56bf8032a.jpg", + "image_caption": [ + "(b) CB" + ], + "image_footnote": [], + "bbox": [ + 317, + 93, + 475, + 180 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/9490c4e09c316164a435fde9b295f7e6dfc9926d451292a2424456bfba8aac34.jpg", + "image_caption": [ + "(c) ANLI" + ], + "image_footnote": [], + "bbox": [ + 512, + 93, + 672, + 181 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/7d237e9658defe93969abaf5410e969848d5bb88f1f9c16f61f17a5c2304e433.jpg", + "image_caption": [ + "(d) ANLI R2" + ], + "image_footnote": [], + "bbox": [ + 707, + 93, + 863, + 177 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/94345d78f9ba02784fc5758aa5298af1f471f707f89243831aa1f69ca9103cf2.jpg", + "image_caption": [ + "(e) ANLI R3" + ], + "image_footnote": [], + "bbox": [ + 119, + 221, + 282, + 303 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/cd031f39a39d528fdc225d388dc7060ef6c76d6425f60064f8c685701d44b79a.jpg", + "image_caption": [ + "(f) COPA" + ], + "image_footnote": [], + "bbox": [ + 319, + 218, + 475, + 306 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/481e804c5f6c417b8768efd7ece1919d29d27c091101a3ea125c3802c04137de.jpg", + "image_caption": [ + "(g) Hellaswag" + ], + "image_footnote": [], + "bbox": [ + 509, + 218, + 670, + 304 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/bbe34813cb710dde0b1ac3ca17b659e72ce568c5a4a751273e353e0fdbdb0e0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 221, + 863, + 303 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/7173cd83ab347f6daf8a9358825658e1162d4e03c2d2b1a72b54cbe060be00b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 347, + 394, + 429 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/7ba60fd0fd27431dc310a232c0695e5978d4f3914e454d39f98b1802390d3ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 347, + 571, + 429 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/7de310496739df397c37b7563fdd412cfe08fce4901f5ce1608dcbcdceebcf4a.jpg", + "image_caption": [ + "(h) StoryCloze" + ], + "image_footnote": [], + "bbox": [ + 594, + 347, + 752, + 429 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/2f1b160c84f15db68dbe6a93ecf996f65a7657a7b825daf242a039ccb2777701.jpg", + "image_caption": [ + "(i) Winogrande", + "(j) WSC", + "(k)WiC", + "Figure 14: Visualization of evaluation results of 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 189, + 462, + 803, + 480 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "a learning rate of 1e-3, input sequence length 512, output sequence 128, and batch size of 1024. We select model checkpoint by early stopping based on validation accuracy. We use a meta-trained LM instead of a naive pretrained model (e.g. SentenceBERT) because meta-trained LM is shown to be more effective for retrieval (Lin et al., 2022). For the interpolation experiment, we set $K' = 3$ for top- $K'$ prompt embedding candidates. For training source prompt embeddings, we used 8 V100 GPUs.", + "bbox": [ + 112, + 537, + 489, + 699 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "G Examples of Applying Prompts, Answer Choice Format and Source Task Types", + "text_level": 1, + "bbox": [ + 112, + 712, + 455, + 763 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Figure 15 shows an example of applying prompt through Promptsource (Bach et al., 2022) as mentioned in Section 3.1.", + "bbox": [ + 112, + 772, + 489, + 820 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We assert that answer choice format is more important than task similarity in Section 6. We further provide details of the input instances of the mentioned tasks: paraphrase, NLI, word sense disambiguation, and sentiment classification in Figure 16. As supported in Pruksachatkun et al. (2020),", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "intuitively, paraphrase task is more similar to the task of word sense disambiguation task or NLI, implying their task similarity, while the task of sentiment classification is very different. However, our counterintuitive result shows the soft prompt to show the best performance in Figure 6 Section 6, bolstering the claim that similar source task types are not a major factor for evaluation performance.", + "bbox": [ + 507, + 537, + 884, + 667 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "H Full List of Source Training and Evaluation Datasets", + "text_level": 1, + "bbox": [ + 507, + 678, + 826, + 712 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "All of our training and evaluation datasets are a subset of datasets used in Sanh et al. (2021). We use Huggingface version of each dataset (Lhoest et al., 2021).", + "bbox": [ + 507, + 722, + 882, + 788 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "H.1 Training Datasets", + "text_level": 1, + "bbox": [ + 507, + 800, + 700, + 816 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Following Sanh et al. (2021), we use 8 task clusters for training of source prompt embedding: sentiment classification, paraphrase, topic classification, summarization, struc-to-text, multiple-choice QA, extractive QA, and closed book QA. We use imdb (Maas et al., 2011), amazon_polarity", + "bbox": [ + 507, + 822, + 884, + 919 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "12304", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/c50949b1eb4152b355b40437041c77e0619ed61eb9b9188201d943658e560a22.jpg", + "image_caption": [ + "Figure 15: Example of applying prompt to a given instance through Promptsource (Bach et al., 2022)." + ], + "image_footnote": [], + "bbox": [ + 230, + 174, + 763, + 336 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/7e0c78c2453749ae0067e28a53ad919ec36b7398bd2909971dd6c3f175995f20.jpg", + "image_caption": [ + "Figure 16: Examples of instances of different source tasks." + ], + "image_footnote": [], + "bbox": [ + 152, + 545, + 848, + 803 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "12305", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "(McAuley and Leskovec, 2013), rotten_tomatoes (Pang and Lee, 2005), yelp_review_full (Zhang et al., 2015b) for sentiment, glue/qqp (Wang et al., 2018), paws/labeled_final (Zhang et al., 2019) for paraphrase, ag_news (Zhang et al., 2015a), dbpedia_14 (Lehmann et al., 2015) for topic classification, gigaword (Graff et al., 2003), multi_news citefabbri-etal-2019-multi, samsum (Gliwa et al., 2019), xsum (Narayan et al., 2018) for summarization, common_gen (Lin et al., 2020), wiki.bio (Lebret et al., 2016) for struct-to-text, cos_e/v1.11 (Rajani et al., 2019), quail (Rogers et al., 2020), social_i_qa (Sap et al., 2019), wiqa (Tandon et al., 2019), cosmos_qa (Huang et al., 2019), sciq (Welbl et al., 2017), wiki_hop/original (Welbl et al., 2018) for multi-choice QA, adversarial_qa/dbidaf, adversarial_qa/dbert, adversarial_qa/droberta, quoref (Bartolo et al., 2020), ropes (Lin et al., 2019), duorc/SelfRC, duorc/Paraphrase IdentificationRC (Saha et al., 2018) for extractive QA, and kilt_tasks/hotpotqa (Petroni et al., 2021), wiki_qa (Yang et al., 2015) for closed book QA.", + "bbox": [ + 115, + 84, + 490, + 437 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We exclude 6 datasets (MRPC, TREC, DREAM, QuaRTz, QASC, QuaRel) that have small training sets because it leads to task imbalance, which is critical for training our small dense retriever ( $\\sim$ 35M params). We also exclude CNN Daily Mail, App Reviews, and WikiQA dataset due to dataset download issues, absence of any test or validation data, and unbalanced label distribution, respectively.", + "bbox": [ + 112, + 444, + 489, + 574 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "H.2 Evaluation Datasets", + "text_level": 1, + "bbox": [ + 114, + 606, + 321, + 620 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Following Sanh et al. (2021), we include 11 evaluation datasets as follows: RTE (Dagan et al., 2005), CB (De Marneffe et al., 2019), ANLI (Nie et al., 2020) for natural language inference task, COPA (Roemmele et al., 2011), Hellaswag (Zellers et al., 2019), Storycloze (Mostafazadeh et al., 2016) for sentence completion task, Winogrande (Sakaguchi et al., 2021), WSC (Levesque et al., 2012) for coreference resolution task, and WiC (Pilehvar and Camacho-Collados, 2019) for word sense disambiguation task.", + "bbox": [ + 112, + 639, + 489, + 816 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "For BIG-bench tasks, we evaluate on 14 tasks, following Sanh et al. (2021): Known Unknown, Logic Grid, StrategyQA, Hindu Knowledge, Movie Dialog, Code Description, Conceptual, Language ID, Vitamin C, Syllogisms, Misconceptions, Logical Deduction, Winowhy and Novel Concepts.", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "I Full List of Retrieved Prompt Embeddings", + "text_level": 1, + "bbox": [ + 509, + 83, + 796, + 117 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We provide a full list of retrieved prompt embeddings of ROSPR and ORACLE for all prompts of 11 evaluation datasets. We report retrieval results of a single random seed (Table $4\\sim$ Table 14).", + "bbox": [ + 507, + 124, + 884, + 190 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "12306", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 18 + }, + { + "type": "table", + "img_path": "images/6889feb02c7342f0e1037fd18ccb7d8c882011b4709fe9a97e3fcd648bc77858.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
GPT-3 style61.3774.01paws/labeled_final/PAWS-ANLI GPT374.01paws/labeled_final/context-question
MNLI crowdsourced63.5370.04paws/labeled_final/context-question72.20paws/labeled_final/context-question
based on the previous passage68.2376.53paws/labeled_final/context-question76.53paws/labeled_final/PAWS-ANLI GPT3
can we infer59.5773.29glue/qqp/meaning73.29paws/labeled_final/PAWS-ANLI GPT3
does it follow that61.7371.84paws/labeled_final/context-question71.84paws/labeled_final/context-question-no-label
does this imply64.6268.59paws/labeled_final/context-question71.48paws/labeled_final/context-question
guaranteed true68.9575.09paws/labeled_final-context-question75.81paws/labeled_final/PAWS-ANLI GPT3
should assume66.4371.12glue/qqp/meaning76.53paws/labeled_final/PAWS-ANLI GPT3
justified in saying61.0158.12paws/labeled_final/paraphrase-task71.12paws/labeled_final/context-question
must be true70.0474.37paws/labeled_final/context-question75.09paws/labeled_final/PAWS-ANLI GPT3-no-label
Avg.64.5571.3073.79
", + "bbox": [ + 115, + 86, + 789, + 218 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/a673f8bf16cbfdeef8e9366728aa5a91936f54aa8bf28a6b939feea55c69f890.jpg", + "table_caption": [ + "Table 4: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of RTE." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer55.3651.79social_i_qa/Show choices and generate index67.86samsum/Write a dialogue that match this summary
based on the previous passage44.6458.93social_i_qa/Show choices and generate index69.64samsum/Write a dialogue that match this summary
claim true/false/inconclusive50.0067.86social_i_qa/Show choices and generate index69.64samsum/Summarize:
does it follow that64.2950.00social_i_qa/Show choices and generate index67.86samsum/Summarize:
justified in saying53.5750.00social_i_qa/Show choices and generate index62.50samsum/Write a dialogue that match this summary
always/sometimes/never39.2941.07social_i_qa/Show choices and generate index41.07social_i_qa/Show choices and generate index
GPT-3 style51.7967.86social_i_qa/Show choices and generate index69.64social_i_qa/Show choices and generate answer
consider always/sometimes/never35.7135.71social_i_qa/Show choices and generate index39.29social_i_qa/Generate answer
guaranteed true48.2150.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
must be true53.5750.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
guaranteed/possible/impossible8.938.93social_i_qa/Show choices and generate index8.93-(all same)
does this imply58.9351.79social_i_qa/Show choices and generate index66.07glue/qqp/duplicate
MNLI crowdsourced8.9339.29social_i_qa/Show choices and generate index42.86cos_e/v1.11/question_option_description_text
should assume57.1450.00social_i_qa/Show choices and generate index66.07cos_e/v1.11/aligned_with_common Sense
take the following as truth50.0067.86social_i_qa/Show choices and generate index71.43cos_e/v1.11/description_question_option_id
Avg.45.3649.4058.10
", + "bbox": [ + 115, + 261, + 838, + 439 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/bb06910fd9d02e5f41c4072a5eca0b888792e4bc193beb7b3b345c9da1c2bd9b.jpg", + "table_caption": [ + "Table 5: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of CB." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.9038.90paws/labeled_final/context-question39.40paws/labeled_final/PAWS-ANLI GPT3
based on the previous passage33.9038.50paws/labeled_final/context-question38.60paws/labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive35.6036.70paws/labeled_final/PAWS-ANLI GPT339.10paws/labeled_final/context-question
does it follow that36.0040.50paws/labeled_final/context-question40.50paws/labeled_final/PAWS-ANLI GPT3
justified in saying33.1038.10paws/labeled_final-context-question38.80paws/labeled_final/context-question-no-label
always/sometimes/never33.4033.40paws/labeled_final/paragraph-task33.40-(all 33.4)
GPT-3 style33.8037.30paws/labeled_final/PAWS-ANLI GPT338.50paws/labeled_final/PAWS-ANLI GPT3-no-label
consider always/sometimes/never33.2033.40paws/labeled_final/PAWS-ANLI GPT333.50-(all 33.4)
guaranteed true33.7038.50paws/labeled_final/context-question38.70paws/labeled_final/PAWS-ANLI GPT3
must be true34.4039.60paws/labeled_final CONTEXT-question39.70paws/labeled_final-context-question
guaranteed/possible/impossible33.3033.30paws/labeled_final/PAWS-ANLI GPT333.30imdb/Text Expressed Sentiment
does this imply33.6038.20paws/labeled_final contextual-question38.20paws/labeled_final contextual-question-no-label
MNLI crowdsourced33.6033.80paws/labeled_final contextual-question35.40dbpedia_14/given_list WHAT_category DOES_theParagraph_belong_to
should assume33.2038.80paws/labeled_final contextual-question39.00paws/labeled_final/PAWS-ANLI GPT3-no-label
take the following as truth32.4037.10paws/labeled_final/PAWS-ANLI GPT338.70paws/labeled_final contextual-question-no-label
Avg.33.8137.0737.65
", + "bbox": [ + 115, + 483, + 890, + 661 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/7a77a88e13784d8134bd4986d6cc9f6ae11c5ba03fda7fcea5b0cf751a649941.jpg", + "table_caption": [ + "Table 6: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R1." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer30.4035.30paws/ labeled_final/context-question35.30paws/ labeled_final/PAWS-ANLI GPT3
based on the previous passage31.4035.40paws/ labeled_final/context-question35.40paws/ labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive34.9035.10paws/ labeled_final/PAWS-ANLI GPT335.90paws/ labeled_final/context-question
does it follow that34.5036.00paws/ labeled_final/context-question36.00paws/ labeled_final/PAWS-ANLI GPT3
justified in saying33.5035.70paws/ labeled_final/context-question35.70rotten_tomatoes/Reviewer Enjoyment Yes No
always/sometimes/never33.4033.40paws/ labeled_final/paragraph-task33.50adversarial_qa/droberta/based_on,
GPT-3 style33.5034.90paws/ labeled_final/PAWS-ANLI GPT335.00paws/ labeled_final/context-question-no-label
consider always/sometimes/never33.7033.40dbpedia_14/given_a_choice_of categories34.50paws/ labeled_final/PAWS-ANLI GPT3-no-label
guaranteed true32.9034.00paws/ labeled_final/context-question34.30paws/ labeled_final/PAWS-ANLI GPT3
must be true35.1034.60paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
guaranteed/possible/impossible33.3033.30paws/ labeled_final/context-question-no-label33.30imdb/Writer Expressed Sentiment
does this imply32.7033.90paws/ labeled_final/context-question34.10paws/ labeled_final/context-question
MNLI crowdsourced33.4034.70dbpedia_14/given_a_choice_ofcategories34.90dbpedia_14/given_a_choice_ofcategories
should assume32.4035.10paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
take the following as truth31.6035.70paws/ labeled_final/paragraph-task35.70paws/ labeled_final/PAWS-ANLI GPT3
Avg.33.1134.7034.92
", + "bbox": [ + 115, + 706, + 825, + 883 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Table 7: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R2.", + "bbox": [ + 132, + 892, + 858, + 908 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "12307", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 19 + }, + { + "type": "table", + "img_path": "images/79fba8fb4e872b5d4c25eedf328d02cf77f5c3c550f822adfdeb5f728ca7ea68.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.0034.75glue/qqp/meaning34.75paws/labeled_final/context-question-no-label
based on the previous passage33.3334.08paws/labeled_final/context-question35.33paws/labeled_final/context-question-no-label
claim true/false/inconclusive32.8335.58paws/labeled_final/paraphrase-task35.92cos_e/v1.11/aligned_with_common Sense
does it follow that34.0834.67paws/labeled_final/context-question35.33amazon_polarity/User_recommend.this_product
justified in saying33.5833.00paws/labeled_final/paraphrase-task35.42amazon_polarity/User Recommend.this_product
always/sometimes/never33.4233.42paws/labeled_final/paraphrase-task33.50paws/labeled_final/paraphrase-task
GPT-3 style33.3334.00paws/labeled_final/PAWS-ANLI GPT334.92paws/labeled_final/PAWS-ANLI GPT3
consider always/sometimes/never33.0832.42ropes/plain_nobackground33.67cos_e/v1.11/question_description_option_text
guaranteed true32.5834.08paws/labeled_final/context-question34.83paws/labeled_final/context-question
must be true33.8333.75paws/labeled_final/paraphrase-task35.42rotten_tomatoes/Reviewer Enjoy Yes No
guaranteed/possible/impossible33.5033.50paws/labeled_final/paraphrase-task33.58social_i_qa/Show choices and generate answer
does this imply32.9233.58glue/qqp/meaning34.50paws/labeled_final/context-question
MNLI crowdsourced33.7533.67rotten_tomatoes/Text Expressed Sentiment34.00dbpedia_14/given_a_choice_ofcategories
should assume33.2534.67paws/labeled_final/context-question34.92paws/labeled_final/context-question-no-label
take the following as truth33.4232.75social_i_qa/Show choices and generate index36.67paws/labeled_final/context-question-no-label
Avg.33.3333.8634.91
", + "bbox": [ + 115, + 93, + 828, + 273 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/151bd3efd99a1c7b1e832a20503c80b33827d3652ed5e03b692c5d742a39338b.jpg", + "table_caption": [ + "Table 8: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R3." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
exercise80.0079.00cos_e/v1.11/question_option_description_text80.00cos_e/v1.11/question_option_description_text
plausible_alternatives84.0083.00cos_e/v1.11/question_option_description_text83.00cos_e/v1.11/question_option_description_text
"C1 or C2? premise, so/because..."61.0067.00social_i_qa/Check if a random answer is valid or not75.00cos_e/v1.11/descriptionQuestion_option_text
best_option70.0076.00social_i_qa/Show choices and generate answer79.00cos_e/v1.11/question_option_description_text
more likely79.0083.00cos_e/v1.11/question_option_description_text85.00cos_e/v1.11/descriptionQuestion_option_text
cause-effect74.0078.00cos_e/v1.11/question_option_description_text83.00common_gen/random task template prompt
choose80.0082.00cos_e/v1.11/question_option_description_text82.00cosmos_qa/no_prompt_text
i_am_hesitating79.0082.00cos_e/v1.11/question_option_description_text82.00cos_e/v1.11/question_option_description_text
Avg.75.8878.7581.13
", + "bbox": [ + 115, + 332, + 858, + 445 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/a2229f2b60afef3815db3c973515c0ff1fdc90d65d885db4e2a71bab817f9870.jpg", + "table_caption": [ + "Table 9: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of COPA." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Predict ending with hint26.8327.70social_i_qa/Show choices and generate index29.11cos_e/v1.11/question_option_description_text
Randomized prompts template26.8726.83social_i_qa/Show choices and generate index27.84wiqa/what_is_the MISSING_first_step
complete_first THEN28.3727.35ropes/prompt_bottom_no_hint28.35cos_e/v1.11/question_option_description_text
if_beginnerscontinue25.2826.58social_i_qa/Show choices and generate index26.58cos_e/v1.11/question_option_description_text
how Ends25.6726.86social_i_qa/Show choices and generate index26.86cos_e/v1.11/question_option_description_text
Avg.26.6027.0627.75
", + "bbox": [ + 115, + 504, + 806, + 589 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/78e3c6fb6dbe3e08100c23bfaa4295246a6c01c2447e4c5e21f4d1e59aa5d944.jpg", + "table_caption": [ + "Table 10: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Hellaswag." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Answer Given options86.4887.76social_i_qa/Show choices and generate answer87.76social_i_qa/Show choices and generate answer
Choose Story Ending86.9188.24social_i_qa/Show choices and generate answer88.24cos_e/v1.11/question_option_description_text
Movie What Happens Next78.0978.03social_i_qa/Show choices and generate answer87.33cosmos_qa/no_prompt_text
Story Continuation and Options83.5985.89social_i_qa/Show choices and generate answer86.53social_i_qa/Show choices and generate answer
Novel Correct Ending85.0987.65social_i_qa/Show choices and generate answer87.97social_i_qa/Show choices and generate index
Avg.84.0385.5287.57
", + "bbox": [ + 115, + 646, + 821, + 732 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/3117bd0163f6efbc8eb00d222521288cc3ab43c17075ad7f5cb8cea1e6a35b42.jpg", + "table_caption": [ + "Table 11: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of StoryCloze." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does underscore refer to stand for49.2551.54cos_e/v1.11/question_option_description_text51.78paws/ labeled_final/ PAWS-ANLI GPT3
50.5150.59cos_e/v1.11/question_option_description_text52.09ropes/plain_nobackground
underscore refer to fill in the blank50.4350.59cos_e/v1.11/question_option_description_text52.33ropes/prompt_bottom_no_hint
52.1751.38cos_e/v1.11/question_option_description_text52.01cos_e/v1.11/question_option_description_text
Replace52.4953.59cos_e/v1.11/question_option_description_text53.59paws/ labeled_final/ paraphrase-task
Avg.50.9751.5452.36
", + "bbox": [ + 115, + 791, + 784, + 876 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Table 12: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Winogrande.", + "bbox": [ + 119, + 885, + 873, + 901 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "12308", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 20 + }, + { + "type": "table", + "img_path": "images/3842c60d39f9d34d30029f5364dd7f7097e4c2f87aa2b157f6450b6572e3bcfd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does the pronoun refer to by p they mean68.2766.34social_i_qa/Check if a random answer is valid or not66.35paws/labeled_final/PAWS-ANLI GPT3
62.5066.35social_i_qa/Check if a random answer is valid or not69.23glue/qqp/duplicate or not
in other words67.3167.31social_i_qa/Show choices and generate answer72.12samsum/Write a dialogue that match this summary
I think they mean69.2368.27social_i_qa/Show choices and generate answer78.08social_i_qa/Generate answer
replaced with64.4259.62social_i_qa/Check if a random answer is valid or not66.35social_i_qa/Show choices and generate index
p is/are r62.5064.42social_i_qa/Show choices and generate index65.38social_i_qa/Show choices and generate index
the pronoun refers to64.4264.42social_i_qa/Show choices and generate index67.31social_i_qa/Check if a random answer is valid or not
Who or what is/are64.4263.46social_i_qa/Show choices and generate index65.38samsum/To sum up this dialog
does p stand for67.3163.46social_i_qa/I was wondering69.23samsum/Write a dialogue that match this summary
GPT-3 Style60.5867.31social_i_qa/Show choices and generate index67.31rotten_tomatoes/Reviewer Expressed Sentiment
Avg.65.1065.1068.17
", + "bbox": [ + 115, + 209, + 840, + 340 + ], + "page_idx": 21 + }, + { + "type": "table", + "img_path": "images/1de7d1e38a875316b71af801502dcf5a6895fb65e9a7f26419df25755ace72ef.jpg", + "table_caption": [ + "Table 13: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WSC." + ], + "table_footnote": [], + "table_body": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
question-context-meaning-with-label50.3153.76glue/qqp/meaning53.76paws/labeled_final/context-question-no-label
question-context-meaning50.6350.16ropes/prompt_bottom_no_hint57.05paws/labeled_final/PAWS-ANLI GPT3
grammar Homework49.8450.16ropes/prompt_bottom_no_hint57.99rotten_tomatoes/Reviewer Enjoyment Yes No
affirmation_true_orfalse49.6951.57paws/labeled_final/Rewrite-no-label53.92paws/labeled_final/Meaning-no-label
GPT-3-prompt51.7250.00social_i_qa/Show choices and generate index55.64paws/labeled_final/PAWS-ANLI GPT3
same_sense49.8452.82paws/labeled_final/Rewrite55.17paws/labeled_final/task_description-no-label
question-context51.8853.29social_i_qa/Check if a random answer is valid or not57.05amazon_polarity/Is.this_product_review_positive
GPT-3-prompt-with-label50.4752.19glue/qqp/meaning52.66glue/qqp/meaning
polysemous50.0052.82paws/labeled_final/Rewrite53.29paws/labeled_final/Rewrite-no-label
similar-sense52.5150.00social_i_qa/Show choices and generate index56.11paws/labeled_final/task_description-no-label
Avg.50.6951.6855.26
", + "bbox": [ + 115, + 631, + 875, + 762 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Table 14: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WiC.", + "bbox": [ + 144, + 771, + 847, + 785 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "12309", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 21 + } +] \ No newline at end of file diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_model.json b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b2b7a44d59e4290c9162f8d724c962b0aab5d3e --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_model.json @@ -0,0 +1,3764 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.135, + 0.09, + 0.868, + 0.131 + ], + "angle": 0, + "content": "Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.155, + 0.823, + 0.188 + ], + "angle": 0, + "content": "Seonghyeon Ye Joel Jang Doyoung Kim Yongrae Jo Minjoon Seo KAIST" + }, + { + "type": "text", + "bbox": [ + 0.202, + 0.19, + 0.802, + 0.206 + ], + "angle": 0, + "content": "{seonghyeon.ye,joeljang,ikevin98,yongrae,minjoon}@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.343, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.285, + 0.462, + 0.612 + ], + "angle": 0, + "content": "Enhancing the zero-shot performance of instruction-following models requires heavy computation, either by scaling the total number of training datasets or the model size. In this work, we explore how retrieval of soft prompts obtained through prompt tuning can efficiently assist hard prompts in zero-shot task generalization. Specifically, we train soft prompt embeddings for each prompt through prompt tuning, store the samples of the training instances mapped with the prompt embeddings, and retrieve the corresponding prompt embedding of the training instance closest to the query instance during inference. While only adding \\(0.007\\%\\) additional parameters, retrieval of soft prompt enhances the performance of T0 on unseen tasks by outperforming it on 10 out of 11 datasets as well as improving the mean accuracy of T0 on BIG-bench benchmark by \\(2.39\\%\\) points. Also, we report an interesting finding that retrieving source embeddings trained on similar answer choice formats is more important than those on similar task types." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.628, + 0.262, + 0.643 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.656, + 0.49, + 0.882 + ], + "angle": 0, + "content": "Training Large Language Models (LLMs) on huge amounts of data has enabled LMs to perform downstream tasks without any fine-tuning with the aid of natural prompts or concatenation of a few demonstration instances (Brown et al., 2020; Rae et al., 2021; Kojima et al., 2022; Chowdhery et al., 2022). Additionally, recent works have shown that adding a instruction tuning stage, an additional training step that helps pretrained LMs understand prompts and demonstrations results in a significant performance boost on zero-shot task generalization even for moderate-sized LMs (Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Ye et al., 2022; Chung et al., 2022). This extra" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.252, + 0.885, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.362, + 0.884, + 0.412 + ], + "angle": 0, + "content": "Figure 1: During zero-shot inference, ROSPR selects similar training instances with the given input from the Source Prompt Library and retrieves the prompt embeddings corresponding to the selected training instances." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.42, + 0.884, + 0.468 + ], + "angle": 0, + "content": "instruction-tuning stage involves explicit, multitask prompted learning on various tasks, enabling LMs to quickly adapt to unseen tasks at inference." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.469, + 0.885, + 0.66 + ], + "angle": 0, + "content": "To maximize the effect of instruction-tuning, two approaches have been widely explored: (1) scaling the number of training datasets, and (2) scaling the model size (Wang et al., 2022b; Chung et al., 2022). However, both approaches require heavy computation, not applicable with an academic budget. Specifically, the first approach requires updating the whole parameters of the model every time a training dataset is added, showing limitations in terms of scalability. On the other hand, the second approach requires heavy memory requirements to load and train a massive LLM." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.884, + 0.919 + ], + "angle": 0, + "content": "To enhance the zero-shot performance of instruction-following model efficiently, we introduce Retrieval of Soft Prompt (ROSPR), which is easily scalable and requires minimal computation by only adding \\(0.007\\%\\) parameters to the main model during inference. As shown in Figure 1, by training prompt embeddings (soft prompt) for each given hard prompt through prompt tuning, we construct a Source Prompt Library consisting of samples of training instances mapped with their corresponding prompt embeddings. Then, during inference, by using a simple, off-the-shelf dense retriever model, we search for training instances similar to the given query instances and retrieve their corresponding prompt embeddings. Because the backbone LM is frozen, it allows the retrieved" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "1Model checkpoints and code implementation are available at github.com/seonghyeonye/RoSPr." + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12288" + }, + { + "type": "footer", + "bbox": [ + 0.21, + 0.946, + 0.788, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12288-12309 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.165 + ], + "angle": 0, + "content": "embeddings to serve as adapters assisting hard prompts. While RoSPR can be applied to any LM, in this work, we use T0 (Sanh et al., 2021) as our initial backbone LM and perform prompt tuning on the tasks used during the instruction-tuning stage." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.166, + 0.49, + 0.455 + ], + "angle": 0, + "content": "While adding only \\(0.007\\%\\) additional parameters, ROSPR outperforms T0 on 10 out of 11 evaluation datasets and outperforms efficient fine-tuning baselines without any target task fine-tuning. ROSPR is also effective for challenging tasks such as tasks from BIG-bench (Srivastava et al., 2022), outperforming T0 by \\(2.39\\%\\) mean accuracy. Furthermore, we provide several interesting findings: (1) Variants of ROSPR that include interpolation of multiple prompt embeddings and scoring method that considers the answer choice distribution during retrieval further increases the effect of ROSPR (2) Also, we provide analysis of which factors attribute to the performance of ROSPR and show that, similarly to the role of demonstrations in in-context learning (Min et al., 2022), heuristic features such as answer choice format are more important than the similarity of the source task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.468, + 0.271, + 0.483 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.495, + 0.358, + 0.527 + ], + "angle": 0, + "content": "2.1 Task Generalization with Instruction-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.533, + 0.49, + 0.854 + ], + "angle": 0, + "content": "Prompts and demonstrations are essential for task generalization since proper explanations are required for LMs to understand an unseen task (Kojima et al., 2022; Wei et al., 2022; Lampinen et al., 2022). Instruction-tuning, which is explicit multi-task prompted training on various downstream tasks, is a simple but effective way to achieve this, resulting in improved zero-shot capabilities. Zhong et al. (2021) first introduced the method of instruction-tuning by converting various tasks into a question-answering format and finetuning the model on the aggregated dataset. Following works (Mishra et al., 2022; Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Xu et al., 2022; Ouyang et al., 2022; Ye et al., 2022; Chung et al., 2022) extended this approach on a larger scale and show that zero-shot task generalization could be enhanced with more diverse prompts, a larger number of training downstream tasks, and a larger LM." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.866, + 0.334, + 0.88 + ], + "angle": 0, + "content": "2.2 Source Task Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.918 + ], + "angle": 0, + "content": "Retrieving a source task that is relevant to the target task has shown to result in faster and better task" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.213 + ], + "angle": 0, + "content": "adaptation. For parameter-efficient fine-tuning, Vu et al. (2022); Su et al. (2022) retrieve source prompt embedding that is similar to the target prompt embedding and obtain a better initialization point for prompt tuning. Instead of utilizing a single prompt embedding, recent works show a mixture of multiple prompt embeddings to be effective (Asai et al., 2022; Qin and Eisner, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.214, + 0.885, + 0.422 + ], + "angle": 0, + "content": "For instruction-tuning, Lin et al. (2022) retrieve training instances that are similar to the query through a dense retriever and fine-tune the model using the retrieved examples. For in-context learning, Rubin et al. (2021); Liu et al. (2022b); Wang et al. (2023) retrieve training data that could be used for demonstrations. Wang et al. (2022c) show the effect of retrieving prompt embeddings in a continual learning setting. Although our proposed method is related to these works, the novelty of our work lies in applying source task retrieval in the zero-shot setting and retrieving soft prompts instead of training instances." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.435, + 0.614, + 0.45 + ], + "angle": 0, + "content": "3 Method" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.461, + 0.884, + 0.669 + ], + "angle": 0, + "content": "In this section, we introduce Retrieval of Prompt Tuning (ROSPR) for zero-shot task generalization. A detailed overview is shown in Figure 2. We first train source prompt embeddings of LM for each hard prompt given a source task using prompt tuning (Section 3.1). Then, we save training instance samples along with their prompt embeddings in the Source Prompt Library and use it to retrieve embeddings at inference to perform tasks in a zero-shot manner (Section 3.2). We additionally introduce interpolation of multiple source prompt embeddings (Section 3.3) and variance-based ranking (Section 3.4) to increase robustness and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.681, + 0.851, + 0.696 + ], + "angle": 0, + "content": "3.1 Training Source Prompt Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.702, + 0.883, + 0.846 + ], + "angle": 0, + "content": "Even though RoSPR may be used to augment any type of LM, we use T0 (Sanh et al., 2021) as the backbone LM for this paper. For training of soft prompts, we utilize the source tasks and prompts used for the instruction-tuning phase of T0. While T0 was trained in a multi-task learning manner, we freeze the initial T0 parameters and train only soft prompts (source prompt embeddings) for each hard prompt of the source task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.856, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Prompt Tuning Among various parameter-efficient fine-tuning methods, we follow prompt tuning proposed by Lester et al. (2021) because the number of trainable parameters is extremely small" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.172, + 0.085, + 0.844, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.271, + 0.885, + 0.334 + ], + "angle": 0, + "content": "Figure 2: An overview of RoSPR. For each hard prompt of the source datasets, soft prompts are trained via prompt tuning. After storing training instances as keys and corresponding prompt embedding as values, RoSPR searches training instances similar to query set \\( Q \\), retrieves the corresponding prompt embeddings, and selects the most frequently retrieved candidate for inference. Variants of selection strategy are also shown: RoSPR+INTER interpolates between multiple related source embeddings and RoSPR+VAR ranks candidate embeddings considering both frequency and variance." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.34, + 0.489, + 0.388 + ], + "angle": 0, + "content": "\\((\\sim 204\\mathrm{K}\\) parameters per prompt), which implies that the memory overhead of parameter retrieval at inference is negligible." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.39, + 0.49, + 0.584 + ], + "angle": 0, + "content": "For each source training dataset \\(D_{i}\\) \\((i = 1,\\dots,T)\\) where \\(T\\) is the total number of source datasets, we train source embeddings \\(E_{ij}\\) \\((j = 1,\\dots,M_i)\\) where \\(M_{i}\\) is the number of hard prompts in \\(D_{i}\\), making soft prompt embeddings for each individual hard prompts. Specifically, given a training instance \\(\\{x_{ik},y_{ik}\\} (k = 1,\\dots,K)\\) from \\(D_{i}\\) where \\(K\\) is the number of sampled training instances per dataset, we first convert it into its hard prompted version \\(\\{h_j(x_{ik}),h_j(y_{ik})\\}\\) where \\(h_j(\\cdot)\\) denotes adding the \\(j\\)-th hard prompt. Next, we train the LM with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.599, + 0.489, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\max _ {E _ {i j}} P \\left(h _ {j} \\left(y _ {i k}\\right) \\mid E _ {i j}; h _ {j} \\left(x _ {i k}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.638, + 0.49, + 0.768 + ], + "angle": 0, + "content": "where all the parameters of the underlying backbone LM are frozen and only \\(E_{ij}\\) is trainable. In short, given \\(D_{i}\\), we perform \\(M_{i}\\) number of prompt tunings for each hard prompts, resulting in \\(\\sum_{i=1}^{T} M_{i}\\) total number of source prompt embeddings. For training efficiency, we only train on \\(K = 5000\\) training instances for a single epoch for each source prompt embedding." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.781, + 0.414, + 0.797 + ], + "angle": 0, + "content": "3.2 Zero-Shot Embedding Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.805, + 0.49, + 0.87 + ], + "angle": 0, + "content": "After source prompt embedding training, we retrieve the most related source embeddings and select one from the retrieved candidates to be used at inference (right part of Figure 2)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.341, + 0.885, + 0.517 + ], + "angle": 0, + "content": "We first construct a Source Prompt Library, consisting of sentence-level representations of training instance inputs as keys and the corresponding source prompt embedding as the values. For each available source prompt embedding, \\( n \\) number of samples are stored in the library. The sentence-level representations are obtained by getting the mean representation of hidden states of the last layer of the dense retriever. We use a T0-small encoder as a dense retriever, replicated based on Sanh et al. (2021) with smaller model size." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.518, + 0.886, + 0.791 + ], + "angle": 0, + "content": "At inference, we first randomly sample \\( Q \\) query instances from the target task, following Lin et al. (2022). After obtaining sentence-level representations for each query through our T0-small encoder, we retrieve top- \\( N \\) examples for each query instance using MIPS (maximum inner product search) operation on our Source Prompt Library, retrieving a total of \\( Q*N \\) prompt embeddings. As the default methodology, among the retrieved embedding candidates, we select the most frequently retrieved prompt embedding as our designated soft prompt for the given target task and concatenate the embedding with each of the target task instances before feeding it to our backbone LM. In the next two subsections, we explain different strategies for calculating the target embedding from the \\( Q*N \\) prompt embedding candidates." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.801, + 0.848, + 0.817 + ], + "angle": 0, + "content": "3.3 Interpolation of Prompt Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.821, + 0.884, + 0.886 + ], + "angle": 0, + "content": "When retrieving only a single prompt embedding for a given task (Section 3.2), it may result in high variance across evaluation prompts when the selected prompt embedding does not fit well with the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.892, + 0.883, + 0.919 + ], + "angle": 0, + "content": "3For all indexing and searching, we use FAISS (Johnson et al., 2019) for fast source prompt embedding retrieval." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.489, + 0.92 + ], + "angle": 0, + "content": "For each instances, the input and output are converted into its prompted version using the promptsource toolkit (Bach et al., 2022). An example is given in Appendix G." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12290" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.375 + ], + "angle": 0, + "content": "given task. Recent works on prompt embedding retrieval have shown that the interpolation of prompt embeddings effectively transfers to the target task (Asai et al., 2022; Vu et al., 2022). We also explore calculating the target embedding through interpolation of multiple source embeddings instead of just using a single embedding. Among \\( Q * N \\) prompt candidates searched in Section 3.2, we select top-\\( N' \\) candidate embeddings based on the frequency of the search. Then, we calculate the weighted sum of the candidate embeddings, where the interpolation weight for each source embedding is based on the proportion of frequency. While Asai et al. (2022); Vu et al. (2022) require fine-tuning the target embeddings on the target task to calculate the weights for interpolation, our approach does not require any target task fine-tuning, enabling zero-shot task transfer." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.388, + 0.356, + 0.403 + ], + "angle": 0, + "content": "3.4 Variance-based Ranking" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.409, + 0.49, + 0.681 + ], + "angle": 0, + "content": "Similar to the scoring and calibration method of Lu et al. (2022); Zhao et al. (2021), we introduce a scoring method applicable to zero-shot classification tasks that allows ranking the \\( Q * N \\) retrieved prompt embedding candidates by considering the answer choice distribution of the given target task as extra cues together with the original frequency cues. To accomplish this, we perform a forward pass with the concatenation of each candidate prompt embeddings together with the given hard prompt (only including the instruction, excluding the input instance) of the target task and give a higher score to the embedding candidate that results in lower variance. Ideally, the combination of soft and hard prompts should result in equal probability among the answer choices because the actual context of the task is not included." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.684, + 0.49, + 0.732 + ], + "angle": 0, + "content": "Specifically, when given a target task with \\( k \\)-th hard prompt \\( h_k \\), for each candidate embedding \\( E_{ij} \\) we calculate the scoring as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.744, + 0.488, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S c o r e} \\left(h _ {k}, E _ {i j}\\right) = \\frac {\\operatorname {f r e q} \\left(h _ {k} , E _ {i j}\\right)}{\\sqrt {\\operatorname {V a r} \\left[ P \\left(y \\mid E _ {i j} , h _ {k}\\right) \\right]}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.799, + 0.489, + 0.83 + ], + "angle": 0, + "content": "where \\(y\\) refers to the available output options of the target task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.845, + 0.342, + 0.861 + ], + "angle": 0, + "content": "4 Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In this section, we explain the experimental settings of training of source prompt embedding and construction of our Source Prompt Library. We also" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.134 + ], + "angle": 0, + "content": "explain our evaluation setting during zero-shot inference and baseline models. We provide detailed experiment configurations in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.146, + 0.661, + 0.159 + ], + "angle": 0, + "content": "4.1 Source Tasks" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.167, + 0.885, + 0.328 + ], + "angle": 0, + "content": "For training soft prompts through prompt tuning, we use the subset of source tasks used for the initial T0 instruction-tuning (Sanh et al., 2021)\\(^4\\). For each source task, we use the prompts for each dataset in T0, resulting in a total of 230 prompts. For Source Prompt Library construction, we sample only \\( n = 100 \\) training instances per source embedding to minimize the inference latency. We show a variation of \\( n \\) and different methods to sample \\( n \\) training instances in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.34, + 0.692, + 0.354 + ], + "angle": 0, + "content": "4.2 Evaluation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.361, + 0.885, + 0.635 + ], + "angle": 0, + "content": "Following Sanh et al. (2021), we evaluate on the validation set of 4 held-out tasks (natural language inference, sentence completion, coreference resolution, word sense disambiguation) resulting in a total of 11 evaluation datasets. We also follow Sanh et al. (2021) and evaluate on 14 different datasets from the BIG-bench benchmark (Srivastava et al., 2022). We use rank classification evaluation method by selecting the output option with higher log-likelihood following Brown et al. (2020); Sanh et al. (2021). For all evaluation tasks, we use accuracy as an evaluation metric and report the mean accuracy and standard deviation of all of the evaluation prompts per given dataset (average of \\(\\sim 10\\) prompts per evaluation dataset). For BIG-bench tasks, we do not report standard deviation because only one prompt is provided per task." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.646, + 0.685, + 0.66 + ], + "angle": 0, + "content": "4.3 Baseline Models" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.668, + 0.884, + 0.747 + ], + "angle": 0, + "content": "Zero-shot Baseline For zero-shot baseline models, we show the results of T0 (3B) together with a 4 times larger T0 (11B) instruction-tuned model. We also compare with GPT-3 (175B) model which is 60 times larger than T0 (3B)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.758, + 0.884, + 0.823 + ], + "angle": 0, + "content": "Fine-tuning Baseline We also compare with efficient fine-tuning baseline models that utilize prompt tuning. These models require target task prompt tuning, indicating that zero-shot transfer" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.832, + 0.883, + 0.856 + ], + "angle": 0, + "content": "4We use 29 out of 38 datasets that are used to train T0. We explain the training task selection rationale in Appendix H.1" + }, + { + "type": "page_footnote", + "bbox": [ + 0.511, + 0.857, + 0.883, + 0.88 + ], + "angle": 0, + "content": "5We provide the full list of evaluation datasets in Appendix H." + }, + { + "type": "page_footnote", + "bbox": [ + 0.511, + 0.882, + 0.882, + 0.919 + ], + "angle": 0, + "content": "For methods based on RoSPR, we report the performance average of 3 runs with different random seeds for the sampling of evaluation queries used for the prompt retrieval." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.832, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "12291" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.082, + 0.884, + 0.221 + ], + "angle": 0, + "content": "
Method# of Param\n(Base/Trainable)NLISentence CompletionCoreference Resolut.WSDTotal Avg.
RTECBAN. R1AN. R2AN. R3COPAHellasw.StoryC.Winogr.WSCWiCMeanSTD
T03B / 064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.223.62
PT (FT)3B / 204K63.1444.5233.0731.4432.9473.0026.39-50.6746.7350.02--
ATTEMPT (FT)3B / 614K68.9544.8836.1934.7334.8174.3826.76-51.3364.9051.05--
T0+RoSPR3B / 204K71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.243.62
w/ INTER3B / 204K70.7152.3037.3034.3433.8978.2526.9485.6251.1064.5250.7353.243.30
w/ VAR3B / 204K71.7850.3637.0734.5833.9078.8827.0185.5251.4564.9451.9453.383.38
w/ VAR & INTER3B / 204K72.6051.9837.2534.3133.9577.8326.8485.5850.9364.9751.1853.403.47
ORACLE3B / 204K73.7958.1037.6534.9234.9181.1327.7587.5752.3668.1755.2655.603.07
T011B / 080.8370.1243.5638.6841.2690.0233.5892.4059.9461.4556.5860.76
GPT-3175B / 063.5046.4034.6035.4034.5091.0078.9083.2070.2065.4045.9259.00-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.231, + 0.883, + 0.318 + ], + "angle": 0, + "content": "Table 1: RoSPR refers to our main proposed method, W/ INTER refers to applying interpolation of multiple source embedding candidates, W/ VAR refers to retrieval through variance-based ranking, W/ VAR & INTER refers to applying both interpolation and variance-based ranking where the interpolation weight is based on the variance-based ranking score, and ORACLE refers performance when the most optimal source embedding is retrieved from the candidates, acting as an upper bound performance for retrieval. FT refers to fine-tuned models on the target tasks. For FT models, we exclude StoryCloze due to the absence of training instances. The best and second-best performance is shown in bold and underline respectively. Comparison with hard prompt optimization techniques and visualization of the results is shown in Appendix A and Appendix E, respectively." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.33, + 0.49, + 0.541 + ], + "angle": 0, + "content": "is infeasible. Similar to our source prompt tuning process, we train each target prompt for a single epoch with a maximum of 5,000 training instances. The first baseline model is naive prompt tuning on the target tasks without any prompt retrieval, referred to as PT (Lester et al., 2021). The second model is ATTEMPT (Asai et al., 2022), which trains the target soft prompts through attentional mixtures of source prompts. Because StoryCloze (Mostafazadeh et al., 2016) does not contain training instances, we exclude the dataset for fine-tuning. More training details of fine-tuning baseline are specified in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.554, + 0.336, + 0.57 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.581, + 0.49, + 0.822 + ], + "angle": 0, + "content": "RoSPR enhances the performance of T0 efficiently. Table 1 shows the performance of the 11 evaluation datasets. T0+RoSPR outperforms T0 on 10 datasets among the 11 evaluation datasets. Specifically, T0+RoSPR outperforms T0 on RTE (+6.99% points), CB (+4.12% points), ANLI R1 (+3.24% points), and COPA (+2.87% points). This shows that soft prompt retrieval assists hard prompts for zero-shot task generalization with a negligible number of additional parameters (0.007%). Also, while T0 outperforms GPT-3 on 3 datasets (RTE, StoryCloze, WiC), T0+RoSPR additionally outperforms GPT-3 on 2 datasets (ANLI R1 and CB) and enlarging the score gap for RTE, StoryCloze and WiC." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.33, + 0.885, + 0.556 + ], + "angle": 0, + "content": "ROSPR also outperforms finetuning baselines even without utilizing any training instances of the target task. We first observe that PT harms the performance of the backbone model, which aligns with the result of Liu et al. (2022a); Gu et al. (2022) that prompt tuning is unstable when the training instances or the training steps are small. By comparing ATTEMPT with ROSPR, ROSPR outperforms ATTEMPT on 7 out of 10 tasks and \\(1.21\\%\\) points on the mean accuracy of 10 tasks. This shows that ROSPR is more applicable for efficient adaptation because it requires 3 times fewer additional parameters compared to ATTEMPT and also does not require any further fine-tuning of the target task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.566, + 0.885, + 0.92 + ], + "angle": 0, + "content": "INTER and VAR enhance the performance of ROSPR. We also analyze the effect of introducing variants of ROSPR: interpolation of soft prompts (INTER) and variance-based ranking (VAR) in Table 1. First, applying INTER shows similar accuracy compared to ROSPR. However, as shown in the last column of Table 1, INTER reduces the standard deviation of T0 and ROSPR by \\(8.84\\%\\) while improving the mean accuracy of T0, indicating increased robustness to different surface forms of evaluation prompts. This indicates that interpolation of multiple source embeddings outperforms a single source embedding retrieval, aligning with the result of Asai et al. (2022). Applying VAR with \\(\\mathrm{T0 + ROSPR}\\) improves both zero-shot accuracy and robustness of \\(\\mathrm{T0 + ROSPR}\\), showing that considering the answer choice distribution is beneficial for zero-shot setting, aligned with results from Zhao et al. (2021); Shi et al. (2022). Moreover, applying both VAR+INTER results in the highest overall average accuracy, outperforming T0 by \\(2.18\\%\\) points by largely reducing the gap between larger LLMs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.833, + 0.488, + 0.919 + ], + "angle": 0, + "content": "One exception is WSC, which is a binary classification task (yes/no) predicting whether the reference of the pronoun is correct. We observed that the evaluation data of this dataset has unbalanced labels, containing over \\(60\\%\\) of \"No\" labels. This might be the reason why T0-11B underperforms T0-3B only on this dataset (Sanh et al., 2021). Indeed, predicting only \"No\" on this dataset outperforms T0-11B \\((63.46 > 61.45)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.084, + 0.468, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.231, + 0.49, + 0.294 + ], + "angle": 0, + "content": "Figure 3: Mean accuracy of 14 datasets of BIG-bench. We evaluate on a single prompt following Sanh et al. (2021). By only adding \\(0.007\\%\\) parameters to T0-3B, \\(\\mathrm{T0 + ROSPR}\\) largely reduces the performance gap between 4 times larger T0-11B. The full result is provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.299, + 0.49, + 0.508 + ], + "angle": 0, + "content": "The effect of ROSPR generalizes to challenging tasks. ROSPR is also effective for challenging tasks such as tasks from BIG-bench benchmark. As shown in Figure 3, T0+ROSPR improves the mean accuracy performance of T0-3B by \\(2.39\\%\\) points while only adding \\(0.007\\%\\) additional parameters. T0+ROSPR also outperforms 60 times larger zero-shot and 1-shot GPT-3 and largely reduces the performance gap between 4 times larger T0-11B (1.84% points) or 60 times larger 3-shot GPT-3 (0.53% points). Applying INTER with T0+ROSPR results in additional mean accuracy enhancement, outperforming T0-3B by \\(2.67\\%\\) points8." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.52, + 0.317, + 0.535 + ], + "angle": 0, + "content": "6 Analysis of ROSPR" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.544, + 0.49, + 0.801 + ], + "angle": 0, + "content": "Zero-shot task adaptation of LMs is often seen as a problem of task location, locating the target task to where the model can solve it using the intrinsic ability obtained at pretraining stage with the aid of prompts and demonstrations (Reynolds and McDonell, 2021). In this section, we analyze which factors contribute to the performance enhancement in the perspective of identifying better task location. We find that although the target task performance depends on the source task types, heuristic features such as the answer choice format are more important. This agrees with previous findings that a met-trained LM focuses on simple features such as the label space, the input distribution, and sequence format, instead of complex semantics (Webson and Pavlick, 2021; Min et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.81, + 0.49, + 0.874 + ], + "angle": 0, + "content": "Target task performance depends on source task types. To analyze the effect of different source task types on each target task, we measure the frequency ratio of each source task type that results" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.083, + 0.867, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.233, + 0.885, + 0.261 + ], + "angle": 0, + "content": "Figure 4: Frequency of source task types (x-axis) that maximizes (i.e. ORACLE) the accuracy of each target task (y-axis)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.267, + 0.885, + 0.412 + ], + "angle": 0, + "content": "in the best performance (ORACLE) for the given prompts of the target tasks (visualized in Figure 4). From this figure, we can observe a few patterns: paraphrase task assists NLI and word sense disambiguation task while multi-choice QA (MQA) task assists sentence completion task. For coreference resolution task, various source task types (paraphrase, summarization, multi-choice QA) assist the target task." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.42, + 0.871, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.542, + 0.883, + 0.592 + ], + "angle": 0, + "content": "Figure 5: Effect of answer choice format alignment across different target datasets (RTE, COPA, WiC). We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.599, + 0.885, + 0.742 + ], + "angle": 0, + "content": "Answer choice format is important for task location. We also analyze the effect of using different answer choice formats with the same source task. Answer choice format decides how the available answer choices are given to the LM through the input. For example, a prompt that requires classifying a movie review into good/bad has a different answer choice format from classifying it into positive/negative." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We experiment on 3 datasets (RTE, COPA, WiC) which correspond to different tasks (NLI, sentence completion, word sense disambiguation) respectively. For each dataset, we select a source dataset that is retrieved the most for ORACLE. Among the source prompts of the selected source dataset, we select a prompt that has the same answer choice format as the target task (ALIGNED) and another prompt that has a different answer choice format (MISALIGNED). Figure 5 shows the effect of answer choice format alignment on the target task" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.881, + 0.489, + 0.918 + ], + "angle": 0, + "content": "8We observe that applying VAR results in the same performance as T0+ROSPR because frequency of retrieval has much more influence than variance for BIG-bench tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12293" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.326 + ], + "angle": 0, + "content": "performance by comparing ALIGNED and MIS-ALIGNED. The result shows that for all 3 datasets, ALIGNED significantly outperforms MISALIGNED. This result is non-trivial considering that the two prompt embeddings are trained on the same source training dataset and the same training configuration, with the only difference in the given answer choice format, implying that how the answer choices are given to solve a specific task is more important than the content of the training data for task location. ROSPR is mostly comparable to ALIGNED embedding, implying that retrieving a source prompt embedding by searching for similar input instances results in retrieving a source embedding with similar answer choice formats." + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.333, + 0.431, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.444, + 0.431, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.561, + 0.49, + 0.621 + ], + "angle": 0, + "content": "Figure 6: RTE (Top) and WiC (Bottom) evaluation result by retrieving MISALIGNED and ALIGNED answer choice format across various source tasks. We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.622, + 0.49, + 0.846 + ], + "angle": 0, + "content": "We additionally analyze the effect of answer choice formats on RTE and WiC datasets by retrieving prompt embeddings trained on various source tasks. Both target datasets have the answer choices of either yes/no. Similar to the previous experiments, we retrieve ALIGNED (yes/no format) and MISALIGNED prompt embeddings across three source tasks: paraphrase, sentiment classification, and multi-choice QA. As shown in Figure 6, for both target datasets, ALIGNED outperforms MISALIGNED across all three source tasks. This shows that aligning to the answer choice format of the target task is crucial regardless of the retrieved source task." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.856, + 0.49, + 0.921 + ], + "angle": 0, + "content": "Answer choice format is more important than task similarity. From Figure 6, we can see that all three source tasks benefit from aligning to the target task answer choice format. One may think" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.31 + ], + "angle": 0, + "content": "that embeddings from source tasks requiring similar knowledge to the target task may be important. Counterintuitively, for both RTE and WiC target tasks, when the answer choice format is aligned, the task source embedding of sentiment classification, which is known to be irrelevant to RTE and WiC (Pruksachatkun et al., 2020), outperforms other embeddings sourced from datasets that are more relevant to the target datasets (paraphrase and multi-choice QA) (Appendix G). This implies that for retrieval of source embedding for task location, answer choice format is more important than containing similar knowledge required to solve the target task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.319, + 0.885, + 0.802 + ], + "angle": 0, + "content": "Role of ROSPR is similar to in-context learning. From the findings explained in previous paragraphs, we can conclude that although the source task types influence the target task performance, retrieving a similar answer choice format is more important for task location. Indeed, source tasks containing similar knowledge can help target tasks only if the answer choice formats are aligned to the target task. These findings support Min et al. (2022); Webson and Pavlick (2021) that a meta-trained LM \"takes less effort\" to understand the input: models exploit the simple aspects of prompts and demonstrations such as the format and distribution instead of complex semantics. Especially, for in-context learning, Xie et al. (2021); Min et al. (2022) show that the role of demonstrations lies in providing the shared concept and distribution hints of the target task. From this aspect, the role of ROSPR is similar to demonstrations. However, it is more efficient than including demonstrations because it avoids heavy computation at inference from long sequence lengths (Liu et al., 2022a; Choi et al., 2022) since ROSPR prepends a fixed length of prefix tokens regardless of the task. Also, ROSPR is free from the instability of in-context learning coming from different orderings of demonstrations (Lu et al., 2022; Zhao et al., 2021). Lastly, we conjecture that ROSPR also has the benefits of soft prompts (Li and Liang, 2021) such as having more expressiveness." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.814, + 0.689, + 0.829 + ], + "angle": 0, + "content": "7 Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "In this section, we analyze the effect of the number of (1) prompts, (2) source datasets, and (3) queries sampled for evaluation. We evaluate variations of our proposed methods on 4 datasets: RTE for NLI, COPA for sentence completion, Winogrande" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12294" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.085, + 0.293, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.203, + 0.285, + 0.216 + ], + "angle": 0, + "content": "(a) Datasets vs Prompts" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.081, + 0.471, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.331, + 0.203, + 0.45, + 0.215 + ], + "angle": 0, + "content": "(b) Query variation" + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.228, + 0.489, + 0.289 + ], + "angle": 0, + "content": "Figure 7: Result of various ablation settings of ROSPR. (a) compares the effect of scaling the number of datasets with scaling the number of prompts and (b) shows the effect of the number of sampled queries at inference. Additional ablation results are shown in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.297, + 0.49, + 0.361 + ], + "angle": 0, + "content": "for coreference, and WiC for word sense disambiguation task. We report the average of the mean accuracy of all evaluation prompts for each dataset by running 3 different runs." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.369, + 0.489, + 0.658 + ], + "angle": 0, + "content": "Scaling number of prompts vs. number of datasets. Recent works on meta-trained LMs show that the number of source datasets and prompts is an important factor for zero-shot task generalization (Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Chung et al., 2022). We also show ablations for ROSPR and measure how the zero-shot generalization performance changes when we vary the number of prompts and datasets available during the prompt tuning stage (shown in Figure 7a). First, we vary (1) the total number of source prompts by 60, 120, and 230 by increasing the number of prompts per dataset and (2) the number of datasets by 8, 16, and 30 by increasing the number of datasets per task cluster. Note that in (1), the total number of datasets is fixed while in (2), we use all available prompts for each dataset while varying the number of datasets per task cluster." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.659, + 0.489, + 0.787 + ], + "angle": 0, + "content": "In contrast to (1), (2) does not always lead to a linearly increasing performance boost; the performance saturates as more source datasets are included. By comparing the effect of scaling datasets and scaling prompts for similar Source Prompt Library sizes, we observe that the number of prompts has more impact on the accuracy of the target task (Figure 7a)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.788, + 0.489, + 0.884 + ], + "angle": 0, + "content": "This ablation study also supports the analysis of the previous section; diverse answer choice formats of prompts, which are mostly influenced by the total number of source prompts, are more important than source task types which are influenced by the number of source datasets.[10] Therefore, if" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.15 + ], + "angle": 0, + "content": "the number of task clusters is sufficient to some extent, scaling the number of source prompts per dataset is more crucial than scaling the number of source datasets per task cluster." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.159, + 0.885, + 0.351 + ], + "angle": 0, + "content": "More sampled queries improve the performance. We also analyze the effect of the number of query instances sampled at inference for retrieval. As seen in Figure 7b, increasing the number of queries results in higher mean accuracy. This is different from the analysis of Lin et al. (2022) that sampling more queries leads to better performance only to some point. Because we use the frequency of each prompt embedding candidate as the default metric for retrieval, utilizing more query instances would represent the evaluation data more accurately, resulting in a reduced number of wrong retrievals." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.364, + 0.642, + 0.379 + ], + "angle": 0, + "content": "8 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.39, + 0.884, + 0.791 + ], + "angle": 0, + "content": "In this paper, we introduce ROSPR, a method that efficiently enhances zero-shot generalization capabilities of a meta-trained LM by retrieving prompt-specific source prompt embeddings (soft prompts) for a given target task. We accomplish this by first training the soft prompts for hard prompt of the source tasks. After training source prompt embeddings, we construct the Source Prompt Library by storing the mean representation of training instances as keys and the corresponding prompt embeddings as values. At inference, we search for training instances stored in the library similar to sample instances from the target task, retrieve the corresponding prompt embedding, select the most frequently retrieved embedding, and append it to each of the target task instances for prediction. Our results show that ROSPR efficiently enhances the zero-shot performance of the backbone model while introducing minimal additional parameters during inference. We additionally provide analysis of which factors attribute to the performance of ROSPR and find that heuristic cues such as the answer choice format are critical for generalization performance, implying that it may play a role similar to demonstrations in in-context learning." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.804, + 0.646, + 0.819 + ], + "angle": 0, + "content": "9 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.83, + 0.883, + 0.863 + ], + "angle": 0, + "content": "Although we show the effectiveness of ROSPR by applying it on T0-3B (Sanh et al., 2021), we did" + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.87, + 0.883, + 0.919 + ], + "angle": 0, + "content": "the number of datasets increases, we find that answer choice formats are similar across the same task type, meaning that the diversity of answer choice formats is not increased by increasing the number of datasets per task clusters." + }, + { + "type": "page_footnote", + "bbox": [ + 0.133, + 0.892, + 0.486, + 0.905 + ], + "angle": 0, + "content": "\\( {}^{9} \\) Task cluster is defined as a cluster of the same task types." + }, + { + "type": "page_footnote", + "bbox": [ + 0.133, + 0.905, + 0.486, + 0.918 + ], + "angle": 0, + "content": "10 Although the total number of prompts also increases as" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.295 + ], + "angle": 0, + "content": "not evaluate our method on different model scales such as the T0-11B variant and other LM architectures such as decoder-only LMs due to limited computational resources. This leaves future works on applying ROSPR to even larger LMs and diverse LM architectures (Wang et al., 2022a). Moreover, it is hard to apply VAR to target tasks without answer choices such as free-form generation because variance among options cannot be obtained. However, ROSPR and ROSPR+INTER can still be utilized and we leave applying ROSPR on zero-shot task location of free-form generation as future work (Scialom et al., 2022)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.305, + 0.287, + 0.323 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.33, + 0.491, + 0.492 + ], + "angle": 0, + "content": "We thank Sohee Yang and Eunbi Choi for helpful feedback on the paper. This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00264, Comprehensive Video Understanding and Generation with Knowledge-based Deep Logic Neural Network, \\(80\\%\\); No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST), \\(20\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.516, + 0.214, + 0.532 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.538, + 0.49, + 0.606 + ], + "angle": 0, + "content": "Akari Asai, Mohammadreza Salehi, Matthew E Peters, and Hannaneh Hajishirzi. 2022. Attentional mixtures of soft prompt tuning for parameter-efficient multi-task knowledge sharing. arXiv preprint arXiv:2205.11961." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.613, + 0.49, + 0.81 + ], + "angle": 0, + "content": "Stephen Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-david, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Fries, Maged Alshaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang, Dragomir Radev, Mike Tian-jian Jiang, and Alexander Rush. 2022. Prompt-Source: An integrated development environment and repository for natural language prompts. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 93-104, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.818, + 0.49, + 0.884 + ], + "angle": 0, + "content": "Max Bartolo, Alastair Roberts, Johannes Welbl, Sebastian Riedel, and Pontus Stenetorp. 2020. Beat the AI: Investigating adversarial human annotation for reading comprehension. Transactions of the Association for Computational Linguistics, 8:662-678." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.892, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind" + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.538, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.883, + 0.14 + ], + "angle": 0, + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.149, + 0.885, + 0.19 + ], + "angle": 0, + "content": "Eunbi Choi, Yongrae Jo, Joel Jang, and Minjoon Seo. 2022. Prompt injection: Parameterization of fixed inputs." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.2, + 0.885, + 0.279 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.29, + 0.885, + 0.357 + ], + "angle": 0, + "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.367, + 0.885, + 0.421 + ], + "angle": 0, + "content": "Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine learning challenges workshop*, pages 177-190. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.43, + 0.885, + 0.497 + ], + "angle": 0, + "content": "Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107-124." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.507, + 0.885, + 0.598 + ], + "angle": 0, + "content": "Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 70–79, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.609, + 0.885, + 0.651 + ], + "angle": 0, + "content": "David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2003. English gigaword. Linguistic Data Consortium, Philadelphia, 4(1):34." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.66, + 0.885, + 0.741 + ], + "angle": 0, + "content": "Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. 2022. PPT: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.75, + 0.885, + 0.87 + ], + "angle": 0, + "content": "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2391-2401, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.879, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12296" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.138 + ], + "angle": 0, + "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.149, + 0.488, + 0.226 + ], + "angle": 0, + "content": "Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. 2022. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.238, + 0.488, + 0.317 + ], + "angle": 0, + "content": "Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1203-1213, Austin, Texas. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.327, + 0.488, + 0.405 + ], + "angle": 0, + "content": "Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, D. Kontokostas, Pablo N. Mendes, Sebastian Hellmann, M. Morsey, Patrick van Kleef, S. Auer, and C. Bizer. 2015. Dbpedia - a large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6:167-195." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.416, + 0.488, + 0.508 + ], + "angle": 0, + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.518, + 0.488, + 0.572 + ], + "angle": 0, + "content": "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.582, + 0.488, + 0.803 + ], + "angle": 0, + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierric Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.813, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.113 + ], + "angle": 0, + "content": "Chonghua Liao, Yanan Zheng, and Zhilin Yang. 2022. Zero-label prompt selection." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.122, + 0.883, + 0.175 + ], + "angle": 0, + "content": "Bill Yuchen Lin, Kangmin Tan, Chris Miller, Beiwen Tian, and Xiang Ren. 2022. Unsupervised cross-task generalization via retrieval augmentation. arXiv preprint arXiv:2204.07937." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.184, + 0.883, + 0.276 + ], + "angle": 0, + "content": "Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823-1840, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.285, + 0.883, + 0.364 + ], + "angle": 0, + "content": "Kevin Lin, Oyvind Tafjord, Peter Clark, and Matt Gardner. 2019. Reasoning over paragraph effects in situations. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 58-62, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.373, + 0.883, + 0.439 + ], + "angle": 0, + "content": "Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022a. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. arXiv preprint arXiv:2205.05638." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.448, + 0.883, + 0.553 + ], + "angle": 0, + "content": "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022b. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.562, + 0.883, + 0.667 + ], + "angle": 0, + "content": "Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.676, + 0.883, + 0.782 + ], + "angle": 0, + "content": "Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.791, + 0.883, + 0.869 + ], + "angle": 0, + "content": "Julian J. McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In Seventh ACM Conference on Recommender Systems, RecSys '13, Hong Kong, China, October 12-16, 2013, pages 165-172. ACM." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.878, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "12297" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettle-moyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.16, + 0.49, + 0.252 + ], + "angle": 0, + "content": "Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470-3487, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.26, + 0.49, + 0.391 + ], + "angle": 0, + "content": "Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849, San Diego, California. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.4, + 0.49, + 0.492 + ], + "angle": 0, + "content": "Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797-1807, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.5, + 0.49, + 0.592 + ], + "angle": 0, + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.601, + 0.49, + 0.678 + ], + "angle": 0, + "content": "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.687, + 0.49, + 0.779 + ], + "angle": 0, + "content": "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05), pages 115-124, Ann Arbor, Michigan. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.788, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Roktaschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544, Online. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.204 + ], + "angle": 0, + "content": "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.216, + 0.885, + 0.334 + ], + "angle": 0, + "content": "Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, and Samuel R. Bowman. 2020. Intermediate-task transfer learning with pretrained language models: When and why does it work? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5231-5247, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.346, + 0.885, + 0.438 + ], + "angle": 0, + "content": "Guanghui Qin and Jason Eisner. 2021. Learning how to ask: Querying LMs with mixtures of soft prompts. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5203-5212, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.45, + 0.885, + 0.529 + ], + "angle": 0, + "content": "Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.541, + 0.885, + 0.633 + ], + "angle": 0, + "content": "Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932-4942, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.645, + 0.885, + 0.711 + ], + "angle": 0, + "content": "Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In *Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems*, pages 1-7." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.723, + 0.885, + 0.789 + ], + "angle": 0, + "content": "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.801, + 0.885, + 0.867 + ], + "angle": 0, + "content": "Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. 2020. Getting closer to ai complete question answering: A set of prerequisite real tasks. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8722-8731." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.878, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "12298" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.191 + ], + "angle": 0, + "content": "Amrita Saha, Rahul Aralikatte, Mitesh M. Khapra, and Karthik Sankaranarayanan. 2018. DuoRC: Towards complex language understanding with paraphrased reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683-1693, Melbourne, Australia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.203, + 0.49, + 0.256 + ], + "angle": 0, + "content": "Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.268, + 0.488, + 0.346 + ], + "angle": 0, + "content": "Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.359, + 0.488, + 0.477 + ], + "angle": 0, + "content": "Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social IQa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4463-4473, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.489, + 0.488, + 0.542 + ], + "angle": 0, + "content": "Thomas Scialom, Tuhin Chakrabarty, and Smaranda Muresan. 2022. Continual-t0: Progressively instructing \\(50+\\) tasks to language models without forgetting. arXiv preprint arXiv:2205.12393." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.553, + 0.488, + 0.594 + ], + "angle": 0, + "content": "Weijia Shi, Julian Michael, Suchin Gururangan, and Luke Zettlemoyer. 2022. Nearest neighbor zero-shot inference. arXiv preprint arXiv:2205.13792." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.606, + 0.488, + 0.696 + ], + "angle": 0, + "content": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.71, + 0.49, + 0.841 + ], + "angle": 0, + "content": "Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Huadong Wang, Kaiyue Wen, Zhiyuan Liu, Peng Li, Juanzi Li, Lei Hou, Maosong Sun, and Jie Zhou. 2022. On transferability of prompt tuning for natural language processing. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3949-3969, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.852, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Niket Tandon, Bhavana Dalvi, Keisuke Sakaguchi, Peter Clark, and Antoine Bosselut. 2019. WIQA: A dataset for \"what if...\" reasoning over procedural text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.139 + ], + "angle": 0, + "content": "9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6076-6085, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.151, + 0.884, + 0.243 + ], + "angle": 0, + "content": "Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou', and Daniel Cer. 2022. SPoT: Better frozen model adaptation through soft prompt transfer. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5039-5059, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.255, + 0.884, + 0.36 + ], + "angle": 0, + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.372, + 0.883, + 0.45 + ], + "angle": 0, + "content": "Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. 2022a. What language model architecture and pretraining objective work best for zero-shot generalization? arXiv preprint arXiv:2204.05832." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.463, + 0.883, + 0.528 + ], + "angle": 0, + "content": "Xinyi Wang, Wanrong Zhu, and William Yang Wang. 2023. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. arXiv preprint arXiv:2301.11916." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.541, + 0.883, + 0.632 + ], + "angle": 0, + "content": "Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. 2022b. Benchmarking generalization via in-context instructions on 1,600+ language tasks. arXiv preprint arXiv:2204.07705." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.645, + 0.883, + 0.724 + ], + "angle": 0, + "content": "Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022c. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.736, + 0.883, + 0.776 + ], + "angle": 0, + "content": "Albert Webson and Ellie Pavlick. 2021. Do prompt-based models really understand the meaning of their prompts? arXiv preprint arXiv:2109.01247." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.788, + 0.883, + 0.853 + ], + "angle": 0, + "content": "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "12299" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 94-106, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.167, + 0.49, + 0.232 + ], + "angle": 0, + "content": "Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing datasets for multi-hop reading comprehension across documents. Transactions of the Association for Computational Linguistics, 6:287-302." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.248, + 0.488, + 0.3 + ], + "angle": 0, + "content": "Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.316, + 0.488, + 0.381 + ], + "angle": 0, + "content": "Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. 2022. Zero-prompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot generalization. arXiv preprint arXiv:2201.06910." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.396, + 0.488, + 0.475 + ], + "angle": 0, + "content": "Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013-2018, Lisbon, Portugal. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.491, + 0.488, + 0.543 + ], + "angle": 0, + "content": "Seonghyeon Ye, Doyoung Kim, Joel Jang, Joongbo Shin, and Minjoon Seo. 2022. Guess the instruction! making language models stronger zero-shot learners. arXiv preprint arXiv:2210.02969." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.558, + 0.488, + 0.637 + ], + "angle": 0, + "content": "Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.652, + 0.488, + 0.704 + ], + "angle": 0, + "content": "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015a. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.719, + 0.488, + 0.799 + ], + "angle": 0, + "content": "Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015b. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 649-657." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.813, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.15 + ], + "angle": 0, + "content": "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.162, + 0.882, + 0.254 + ], + "angle": 0, + "content": "Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2856-2878, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.263, + 0.882, + 0.317 + ], + "angle": 0, + "content": "Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12300" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.426, + 0.119 + ], + "angle": 0, + "content": "A Comparison with Hard Prompt Optimization" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.126, + 0.49, + 0.447 + ], + "angle": 0, + "content": "We have conducted additional experiments to compare our method with 3 different hard prompt optimization techniques including (1) RoHPr which utilizes the same search technique as RoSPr but retrieves 4 corresponding training instances for incontext learning instead of the corresponding soft prompt, and (2) APE (Zhou et al., 2023) which is an automatic prompt engineering method that utilizes the generation results of an LLM (Ouyang et al., 2022), known to show better performance than human instructions, and (3) ZPS (Liao et al., 2022) which selects an optimal hard prompt from a prompt pool using prompt ensemble and pseudolabels. As shown in the result of Table 2, on average, RoSPr performs the best on average, showing the benefits of using soft prompts over hard prompts for task generalization. While RoSPr shows improvement on 10 out of 11 datasets, other hard prompt optimization methods do not show consistent improvements." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.458, + 0.469, + 0.473 + ], + "angle": 0, + "content": "B Full Result of BIG-bench Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.483, + 0.49, + 0.707 + ], + "angle": 0, + "content": "We provide the task generalization performance result of 14 tasks from BIG-bench, shown in Table 3. Applying RoSPR largely improves the performance for 3 datasets (Hindu Knowledge, Novel Concepts, Misconceptions): \\((+17.72\\%, +3.12\\%, +1.82\\%)\\) compared to T0-3B and \\((+13.72\\%, +3.12\\%, +2.05\\%)\\) compared to T0-11B. For mean accuracy of 14 tasks, T0+RoSPR outperforms T0-3B by \\(2.39\\%\\) points, reducing the gap between 4 times larger T0-11B to \\(1.84\\%\\) points. Moreover, applying INTER to T0+RoSPR enhances the performance of T0+RoSPR for most tasks, indicating that interpolation of multiple embeddings is effective for challenging tasks." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.718, + 0.396, + 0.735 + ], + "angle": 0, + "content": "C Fine-tuning Baseline Details" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.92 + ], + "angle": 0, + "content": "For fine-tuning baseline models (PT and ATTEMPT), we follow the training configuration of source prompt tuning. We train each target prompt with a single epoch using a maximum of 5,000 training instances. For fine-tuning, we use a batch size of 32 and a learning rate of 1e-3. Also, we randomly select hard prompts (templates) of the training dataset during fine-tuning. For ATTEMPT, we randomly sample one source prompt per task cluster, resulting in interpolation between 8 soft prompts." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.791, + 0.1 + ], + "angle": 0, + "content": "D Additional Ablation Studies" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.123, + 0.825, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.302, + 0.883, + 0.333 + ], + "angle": 0, + "content": "Figure 8: Variation of number of prompts by increasing the number of prompts per dataset." + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.363, + 0.822, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.543, + 0.883, + 0.573 + ], + "angle": 0, + "content": "Figure 9: Variation of number of datasets by increasing the number of datasets per task cluster." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.593, + 0.884, + 0.801 + ], + "angle": 0, + "content": "We provide detailed result of variation of the number of prompts (Figure 8) and the number of datasets (Figure 9). We additionally analyze the effect of (1) different sampling methods for constructing the Source Prompt Library, (2) the number of instances sampled for constructing the Source Prompt Library, (3) the number of top-N retrieval for embedding retrieval, and (4) the number of multiple source embeddings to interpolate. Same as Section 7, we report the mean accuracy of 4 evaluation datasets: RTE, COPA, Winogrande, and WiC with 3 runs with different random seeds for the sampling of evaluation queries." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.817, + 0.859, + 0.849 + ], + "angle": 0, + "content": "D.1 Sampling Methods for Source Prompt Library" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We experiment three different methods to sample instances for constructing Source Prompt Library and analyze the effect of each method. By default, we choose RANDOM method, where we sample" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "12301" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.082, + 0.886, + 0.165 + ], + "angle": 0, + "content": "
RTECBAN.R1AN.R2AN.R3COPAHellasw.StoryC.Winogr.WSCWiCAVG
T064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.22
RoHPr47.1045.0033.5333.3033.0973.3826.6586.9350.5164.4250.0049.45
APE (Zhou et al., 2023)68.1948.3335.8235.3734.2768.0025.7678.7350.4358.7550.8550.41
ZPS (Liao et al., 2022)58.4860.7136.6034.4033.3376.0028.4987.3951.7864.4250.6352.93
RoSPr71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.173, + 0.8, + 0.188 + ], + "angle": 0, + "content": "Table 2: Comparison result of RoSPr with different hard prompt optimization techniques." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.211, + 0.48, + 0.46 + ], + "angle": 0, + "content": "
T0-3BROSPRINTERT0-11B
Strategy.52.7952.0552.2352.75
Movie D.52.8551.4552.2353.69
Known Un.47.8347.8347.8358.70
Logic Grid41.1037.0037.4038.30
Hindu Kn.25.7143.4345.7129.71
Code D.46.6745.0040.0043.33
Concept45.5267.6167.6169.29
Language14.8413.6814.4020.20
Vitamin58.8853.7154.5364.73
Syllogism52.9450.6451.3451.81
Misconcept.50.2352.0552.0550.00
Logical46.6454.8654.8654.86
Winowhy44.2944.3344.2952.11
Novel Con.15.6315.6318.7515.63
AVG42.5644.9545.2346.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.469, + 0.489, + 0.497 + ], + "angle": 0, + "content": "Table 3: Evaluation result of 14 tasks from BIG-bench (Srivastava et al., 2022)." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.513, + 0.484, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.637, + 0.49, + 0.667 + ], + "angle": 0, + "content": "Figure 10: Different instance sampling methods for constructing Source Prompt Library." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.491, + 0.919 + ], + "angle": 0, + "content": "100 instances by random for each prompt by assuming that each random 100 queries of instances can represent the whole prompt. Second, we experiment CLUSTERING method, where we sample top 100 instances which has the closest mean representation from its overall average for each prompt. If we say each instance has a distance of mean representation from its overall average as \\( d_{i} \\) in increasing order, we sample \\( \\{d_1,d_2,\\dots,d_{100}\\} \\). The last method we use is DISTRIBUTed method, where we sample 100 instances in a distributed way with respect to its distance of mean representation from its overall average. If we say each instance has a distance of mean representation from its over" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.214, + 0.885, + 0.276 + ], + "angle": 0, + "content": "all average as \\( d_{i} \\) in increasing order, we sample \\( \\{d_1,d_{1 + N / 100},d_{1 + 2*N / 100},\\dots ,d_{1 + 99*N / 100}\\} \\), assuming there are total \\( N \\) training instances in a dataset." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.279, + 0.886, + 0.456 + ], + "angle": 0, + "content": "As shown in Figure 10, RANDOM method outperforms CLUSTERING and DISTRIBUTED methods. Interestingly, CLUSTERING method significantly hurts the performance on all 4 proposed methods, suggesting that storing similar instances per prompt results in retrieval failures more often. Also for DISTRIBUTED method, most of the methods significantly underperform RANDOM, except INTER and ROSPR+VAR+INTER. From these results, we can conclude that random sampling represents the source dataset most effectively." + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.47, + 0.825, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.648, + 0.884, + 0.69 + ], + "angle": 0, + "content": "Figure 11: Variation of number of instances sampled for constructing Source Prompt Library. Default setting is \\( n = 100 \\)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.721, + 0.852, + 0.754 + ], + "angle": 0, + "content": "D.2 Number of Instances Sampled for Constructing Source Prompt Library" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.758, + 0.885, + 0.92 + ], + "angle": 0, + "content": "We analyze the effect of size of the Source Prompt Library by varying the number of instances \\( n \\) to sample for each hard prompt by 100, 300, 500. Therefore, \\( n \\times \\) (number of total hard prompts) would be the size of the Source Prompt Library. As shown in Figure 11, increasing the number of sampled instances does not increase the performance; it hurts the performance for most cases. This suggests that only a few number of training instances are enough to represent the distribution of prompted" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.085, + 0.434, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.263, + 0.489, + 0.294 + ], + "angle": 0, + "content": "Figure 12: Variation of number of Top \\(N\\) instances for embedding retrieval. Default setting is \\(N = 10\\)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.319, + 0.49, + 0.449 + ], + "angle": 0, + "content": "input (hard prompt + input instances) for each hard prompt and increasing the number sometimes hurt the performance by adding noise to the distribution. This also supports the importance of heuristic cues in Section 6 by showing that adding more training instances per hard prompt does not increase the performance. Instead, adding hard prompts with diverse answer choice format is more important." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.461, + 0.407, + 0.494 + ], + "angle": 0, + "content": "D.3 Number of Top N instances for Embedding Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.499, + 0.49, + 0.725 + ], + "angle": 0, + "content": "We vary the number of top- \\(N\\) instances that are retrieved given each query through MIPS search. As shown in Figure 12, varying the number of top- \\(N\\) instances does not have much effect compared to increasing the number of sampled queries (Figure 7b). This implies that if the size of the evaluation set of the target task is large, sampling more queries is effective than searching for more similar instances per query. This is important for variance-based methods because the number of forward passes needed before evaluation is proportional to \\(Q*N\\). Therefore, we can reduce the time latency by reducing the number of instances retrieved per query without hurting the performance much." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.737, + 0.44, + 0.768 + ], + "angle": 0, + "content": "D.4 Number of Source Embeddings for Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "We analyze the effect of number of source embeddings for interpolation by varying top- \\(N^{\\prime}\\) interpolation from 1 (no interpolation) to 5 shown in Figure 13. By comparing between single prompt embedding retrieval \\((N^{\\prime} = 1)\\) and the interpolation of multiple embeddings \\((N^{\\prime} > 1)\\), the mean accuracy drops by adding multiple source embeddings for retrieval because interpolation-based methods underperform on tasks such as COPA as shown" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.083, + 0.825, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.263, + 0.883, + 0.306 + ], + "angle": 0, + "content": "Figure 13: Variation of number of source embeddings for interpolation based methods. Default setting is \\( K' = 3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.332, + 0.883, + 0.379 + ], + "angle": 0, + "content": "in Table 1. Mean accuracy would increase if we add other datasets for evaluation that benefit from interpolation such as WSC and CB." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.38, + 0.885, + 0.524 + ], + "angle": 0, + "content": "By comparing among various \\( N' \\) values, we find that for ROSPR+INTER, the accuracy substantially decreases for \\( K' = 2 \\), implying that the possibility of a wrong retrieval varies depending on the value of \\( N' \\). In contrast, ROSPR+VAR+INTER is more robust to the value of \\( K' \\), showing that variance-based ranking increases robustness to different numbers of source embeddings for interpolation as well." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.536, + 0.753, + 0.551 + ], + "angle": 0, + "content": "E Visualization of Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.561, + 0.884, + 0.626 + ], + "angle": 0, + "content": "We show the visualization of the evaluation result on 11 datasets in Figure 14. Methods based on ROSPR not only show higher accuracy, but lower variance for many datasets." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.637, + 0.8, + 0.654 + ], + "angle": 0, + "content": "F Experimental Configurations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.92 + ], + "angle": 0, + "content": "As mentioned in the previous sections, we use T0-3B as our backbone meta-trained LM. For prompt tuning, we fix the prefix length as 100 and the embeddings are initialized from 5,000 most common vocabulary following Lester et al. (2021). We train each source embedding for a single epoch with a learning rate of 0.1 and a batch size of 32. We use the Adam optimizer with weight decay of 1e-5. For retrieval, we randomly sample \\( Q = 32 \\) query instances and retrieve top \\( K = 10 \\) examples for each query. We train a T0-small variant (\\( \\sim 35\\mathrm{M} \\) params) as our dense retriever by multitask prompted training on T5+LM model (Lester et al., 2021), replicating the original training setting of T0 by training T5+LM for 8 epochs using the same training instances of Sanh et al. (2021) with" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.094, + 0.286, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.195, + 0.235, + 0.209 + ], + "angle": 0, + "content": "(a) RTE" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.094, + 0.477, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.379, + 0.195, + 0.424, + 0.209 + ], + "angle": 0, + "content": "(b) CB" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.095, + 0.673, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.195, + 0.626, + 0.209 + ], + "angle": 0, + "content": "(c) ANLI" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.095, + 0.865, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.195, + 0.831, + 0.209 + ], + "angle": 0, + "content": "(d) ANLI R2" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.222, + 0.283, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.322, + 0.249, + 0.335 + ], + "angle": 0, + "content": "(e) ANLI R3" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.219, + 0.477, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.322, + 0.432, + 0.335 + ], + "angle": 0, + "content": "(f) COPA" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.219, + 0.672, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.322, + 0.64, + 0.336 + ], + "angle": 0, + "content": "(g) Hellaswag" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.222, + 0.865, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.322, + 0.836, + 0.336 + ], + "angle": 0, + "content": "(h) StoryCloze" + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.348, + 0.396, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.448, + 0.367, + 0.461 + ], + "angle": 0, + "content": "(i) Winogrande" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.348, + 0.573, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.448, + 0.526, + 0.461 + ], + "angle": 0, + "content": "(j) WSC" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.348, + 0.753, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.448, + 0.706, + 0.461 + ], + "angle": 0, + "content": "(k)WiC" + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.463, + 0.804, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.498, + 0.702, + 0.512 + ], + "angle": 0, + "content": "Figure 14: Visualization of evaluation results of 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.538, + 0.49, + 0.7 + ], + "angle": 0, + "content": "a learning rate of 1e-3, input sequence length 512, output sequence 128, and batch size of 1024. We select model checkpoint by early stopping based on validation accuracy. We use a meta-trained LM instead of a naive pretrained model (e.g. SentenceBERT) because meta-trained LM is shown to be more effective for retrieval (Lin et al., 2022). For the interpolation experiment, we set \\( K' = 3 \\) for top-\\( K' \\) prompt embedding candidates. For training source prompt embeddings, we used 8 V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.713, + 0.456, + 0.764 + ], + "angle": 0, + "content": "G Examples of Applying Prompts, Answer Choice Format and Source Task Types" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.774, + 0.49, + 0.821 + ], + "angle": 0, + "content": "Figure 15 shows an example of applying prompt through Promptsource (Bach et al., 2022) as mentioned in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.824, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We assert that answer choice format is more important than task similarity in Section 6. We further provide details of the input instances of the mentioned tasks: paraphrase, NLI, word sense disambiguation, and sentiment classification in Figure 16. As supported in Pruksachatkun et al. (2020)," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.538, + 0.885, + 0.668 + ], + "angle": 0, + "content": "intuitively, paraphrase task is more similar to the task of word sense disambiguation task or NLI, implying their task similarity, while the task of sentiment classification is very different. However, our counterintuitive result shows the soft prompt to show the best performance in Figure 6 Section 6, bolstering the claim that similar source task types are not a major factor for evaluation performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.68, + 0.828, + 0.713 + ], + "angle": 0, + "content": "H Full List of Source Training and Evaluation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.724, + 0.883, + 0.789 + ], + "angle": 0, + "content": "All of our training and evaluation datasets are a subset of datasets used in Sanh et al. (2021). We use Huggingface version of each dataset (Lhoest et al., 2021)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.801, + 0.702, + 0.817 + ], + "angle": 0, + "content": "H.1 Training Datasets" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Following Sanh et al. (2021), we use 8 task clusters for training of source prompt embedding: sentiment classification, paraphrase, topic classification, summarization, struc-to-text, multiple-choice QA, extractive QA, and closed book QA. We use imdb (Maas et al., 2011), amazon_polarity" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.231, + 0.175, + 0.764, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.347, + 0.841, + 0.362 + ], + "angle": 0, + "content": "Figure 15: Example of applying prompt to a given instance through Promptsource (Bach et al., 2022)." + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.546, + 0.849, + 0.804 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.811, + 0.698, + 0.827 + ], + "angle": 0, + "content": "Figure 16: Examples of instances of different source tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12305" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.491, + 0.438 + ], + "angle": 0, + "content": "(McAuley and Leskovec, 2013), rotten_tomatoes (Pang and Lee, 2005), yelp_review_full (Zhang et al., 2015b) for sentiment, glue/qqp (Wang et al., 2018), paws/labeled_final (Zhang et al., 2019) for paraphrase, ag_news (Zhang et al., 2015a), dbpedia_14 (Lehmann et al., 2015) for topic classification, gigaword (Graff et al., 2003), multi_news citefabbri-etal-2019-multi, samsum (Gliwa et al., 2019), xsum (Narayan et al., 2018) for summarization, common_gen (Lin et al., 2020), wiki.bio (Lebret et al., 2016) for struct-to-text, cos_e/v1.11 (Rajani et al., 2019), quail (Rogers et al., 2020), social_i_qa (Sap et al., 2019), wiqa (Tandon et al., 2019), cosmos_qa (Huang et al., 2019), sciq (Welbl et al., 2017), wiki_hop/original (Welbl et al., 2018) for multi-choice QA, adversarial_qa/dbidaf, adversarial_qa/dbert, adversarial_qa/droberta, quoref (Bartolo et al., 2020), ropes (Lin et al., 2019), duorc/SelfRC, duorc/Paraphrase IdentificationRC (Saha et al., 2018) for extractive QA, and kilt_tasks/hotpotqa (Petroni et al., 2021), wiki_qa (Yang et al., 2015) for closed book QA." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.445, + 0.49, + 0.575 + ], + "angle": 0, + "content": "We exclude 6 datasets (MRPC, TREC, DREAM, QuaRTz, QASC, QuaRel) that have small training sets because it leads to task imbalance, which is critical for training our small dense retriever (\\(\\sim\\)35M params). We also exclude CNN Daily Mail, App Reviews, and WikiQA dataset due to dataset download issues, absence of any test or validation data, and unbalanced label distribution, respectively." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.607, + 0.322, + 0.621 + ], + "angle": 0, + "content": "H.2 Evaluation Datasets" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.64, + 0.49, + 0.817 + ], + "angle": 0, + "content": "Following Sanh et al. (2021), we include 11 evaluation datasets as follows: RTE (Dagan et al., 2005), CB (De Marneffe et al., 2019), ANLI (Nie et al., 2020) for natural language inference task, COPA (Roemmele et al., 2011), Hellaswag (Zellers et al., 2019), Storycloze (Mostafazadeh et al., 2016) for sentence completion task, Winogrande (Sakaguchi et al., 2021), WSC (Levesque et al., 2012) for coreference resolution task, and WiC (Pilehvar and Camacho-Collados, 2019) for word sense disambiguation task." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "For BIG-bench tasks, we evaluate on 14 tasks, following Sanh et al. (2021): Known Unknown, Logic Grid, StrategyQA, Hindu Knowledge, Movie Dialog, Code Description, Conceptual, Language ID, Vitamin C, Syllogisms, Misconceptions, Logical Deduction, Winowhy and Novel Concepts." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.084, + 0.798, + 0.118 + ], + "angle": 0, + "content": "I Full List of Retrieved Prompt Embeddings" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.126, + 0.885, + 0.191 + ], + "angle": 0, + "content": "We provide a full list of retrieved prompt embeddings of ROSPR and ORACLE for all prompts of 11 evaluation datasets. We report retrieval results of a single random seed (Table \\(4\\sim\\) Table 14)." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12306" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.087, + 0.791, + 0.219 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
GPT-3 style61.3774.01paws/labeled_final/PAWS-ANLI GPT374.01paws/labeled_final/context-question
MNLI crowdsourced63.5370.04paws/labeled_final/context-question72.20paws/labeled_final/context-question
based on the previous passage68.2376.53paws/labeled_final/context-question76.53paws/labeled_final/PAWS-ANLI GPT3
can we infer59.5773.29glue/qqp/meaning73.29paws/labeled_final/PAWS-ANLI GPT3
does it follow that61.7371.84paws/labeled_final/context-question71.84paws/labeled_final/context-question-no-label
does this imply64.6268.59paws/labeled_final/context-question71.48paws/labeled_final/context-question
guaranteed true68.9575.09paws/labeled_final-context-question75.81paws/labeled_final/PAWS-ANLI GPT3
should assume66.4371.12glue/qqp/meaning76.53paws/labeled_final/PAWS-ANLI GPT3
justified in saying61.0158.12paws/labeled_final/paraphrase-task71.12paws/labeled_final/context-question
must be true70.0474.37paws/labeled_final/context-question75.09paws/labeled_final/PAWS-ANLI GPT3-no-label
Avg.64.5571.3073.79
" + }, + { + "type": "table_caption", + "bbox": [ + 0.151, + 0.228, + 0.843, + 0.243 + ], + "angle": 0, + "content": "Table 4: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of RTE." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.262, + 0.839, + 0.44 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer55.3651.79social_i_qa/Show choices and generate index67.86samsum/Write a dialogue that match this summary
based on the previous passage44.6458.93social_i_qa/Show choices and generate index69.64samsum/Write a dialogue that match this summary
claim true/false/inconclusive50.0067.86social_i_qa/Show choices and generate index69.64samsum/Summarize:
does it follow that64.2950.00social_i_qa/Show choices and generate index67.86samsum/Summarize:
justified in saying53.5750.00social_i_qa/Show choices and generate index62.50samsum/Write a dialogue that match this summary
always/sometimes/never39.2941.07social_i_qa/Show choices and generate index41.07social_i_qa/Show choices and generate index
GPT-3 style51.7967.86social_i_qa/Show choices and generate index69.64social_i_qa/Show choices and generate answer
consider always/sometimes/never35.7135.71social_i_qa/Show choices and generate index39.29social_i_qa/Generate answer
guaranteed true48.2150.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
must be true53.5750.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
guaranteed/possible/impossible8.938.93social_i_qa/Show choices and generate index8.93-(all same)
does this imply58.9351.79social_i_qa/Show choices and generate index66.07glue/qqp/duplicate
MNLI crowdsourced8.9339.29social_i_qa/Show choices and generate index42.86cos_e/v1.11/question_option_description_text
should assume57.1450.00social_i_qa/Show choices and generate index66.07cos_e/v1.11/aligned_with_common Sense
take the following as truth50.0067.86social_i_qa/Show choices and generate index71.43cos_e/v1.11/description_question_option_id
Avg.45.3649.4058.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.45, + 0.84, + 0.466 + ], + "angle": 0, + "content": "Table 5: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of CB." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.484, + 0.891, + 0.662 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.9038.90paws/labeled_final/context-question39.40paws/labeled_final/PAWS-ANLI GPT3
based on the previous passage33.9038.50paws/labeled_final/context-question38.60paws/labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive35.6036.70paws/labeled_final/PAWS-ANLI GPT339.10paws/labeled_final/context-question
does it follow that36.0040.50paws/labeled_final/context-question40.50paws/labeled_final/PAWS-ANLI GPT3
justified in saying33.1038.10paws/labeled_final-context-question38.80paws/labeled_final/context-question-no-label
always/sometimes/never33.4033.40paws/labeled_final/paragraph-task33.40-(all 33.4)
GPT-3 style33.8037.30paws/labeled_final/PAWS-ANLI GPT338.50paws/labeled_final/PAWS-ANLI GPT3-no-label
consider always/sometimes/never33.2033.40paws/labeled_final/PAWS-ANLI GPT333.50-(all 33.4)
guaranteed true33.7038.50paws/labeled_final/context-question38.70paws/labeled_final/PAWS-ANLI GPT3
must be true34.4039.60paws/labeled_final CONTEXT-question39.70paws/labeled_final-context-question
guaranteed/possible/impossible33.3033.30paws/labeled_final/PAWS-ANLI GPT333.30imdb/Text Expressed Sentiment
does this imply33.6038.20paws/labeled_final contextual-question38.20paws/labeled_final contextual-question-no-label
MNLI crowdsourced33.6033.80paws/labeled_final contextual-question35.40dbpedia_14/given_list WHAT_category DOES_theParagraph_belong_to
should assume33.2038.80paws/labeled_final contextual-question39.00paws/labeled_final/PAWS-ANLI GPT3-no-label
take the following as truth32.4037.10paws/labeled_final/PAWS-ANLI GPT338.70paws/labeled_final contextual-question-no-label
Avg.33.8137.0737.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.672, + 0.86, + 0.687 + ], + "angle": 0, + "content": "Table 6: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R1." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.707, + 0.826, + 0.884 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer30.4035.30paws/ labeled_final/context-question35.30paws/ labeled_final/PAWS-ANLI GPT3
based on the previous passage31.4035.40paws/ labeled_final/context-question35.40paws/ labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive34.9035.10paws/ labeled_final/PAWS-ANLI GPT335.90paws/ labeled_final/context-question
does it follow that34.5036.00paws/ labeled_final/context-question36.00paws/ labeled_final/PAWS-ANLI GPT3
justified in saying33.5035.70paws/ labeled_final/context-question35.70rotten_tomatoes/Reviewer Enjoyment Yes No
always/sometimes/never33.4033.40paws/ labeled_final/paragraph-task33.50adversarial_qa/droberta/based_on,
GPT-3 style33.5034.90paws/ labeled_final/PAWS-ANLI GPT335.00paws/ labeled_final/context-question-no-label
consider always/sometimes/never33.7033.40dbpedia_14/given_a_choice_of categories34.50paws/ labeled_final/PAWS-ANLI GPT3-no-label
guaranteed true32.9034.00paws/ labeled_final/context-question34.30paws/ labeled_final/PAWS-ANLI GPT3
must be true35.1034.60paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
guaranteed/possible/impossible33.3033.30paws/ labeled_final/context-question-no-label33.30imdb/Writer Expressed Sentiment
does this imply32.7033.90paws/ labeled_final/context-question34.10paws/ labeled_final/context-question
MNLI crowdsourced33.4034.70dbpedia_14/given_a_choice_ofcategories34.90dbpedia_14/given_a_choice_ofcategories
should assume32.4035.10paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
take the following as truth31.6035.70paws/ labeled_final/paragraph-task35.70paws/ labeled_final/PAWS-ANLI GPT3
Avg.33.1134.7034.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.894, + 0.86, + 0.909 + ], + "angle": 0, + "content": "Table 7: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R2." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12307" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.095, + 0.83, + 0.274 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.0034.75glue/qqp/meaning34.75paws/labeled_final/context-question-no-label
based on the previous passage33.3334.08paws/labeled_final/context-question35.33paws/labeled_final/context-question-no-label
claim true/false/inconclusive32.8335.58paws/labeled_final/paraphrase-task35.92cos_e/v1.11/aligned_with_common Sense
does it follow that34.0834.67paws/labeled_final/context-question35.33amazon_polarity/User_recommend.this_product
justified in saying33.5833.00paws/labeled_final/paraphrase-task35.42amazon_polarity/User Recommend.this_product
always/sometimes/never33.4233.42paws/labeled_final/paraphrase-task33.50paws/labeled_final/paraphrase-task
GPT-3 style33.3334.00paws/labeled_final/PAWS-ANLI GPT334.92paws/labeled_final/PAWS-ANLI GPT3
consider always/sometimes/never33.0832.42ropes/plain_nobackground33.67cos_e/v1.11/question_description_option_text
guaranteed true32.5834.08paws/labeled_final/context-question34.83paws/labeled_final/context-question
must be true33.8333.75paws/labeled_final/paraphrase-task35.42rotten_tomatoes/Reviewer Enjoy Yes No
guaranteed/possible/impossible33.5033.50paws/labeled_final/paraphrase-task33.58social_i_qa/Show choices and generate answer
does this imply32.9233.58glue/qqp/meaning34.50paws/labeled_final/context-question
MNLI crowdsourced33.7533.67rotten_tomatoes/Text Expressed Sentiment34.00dbpedia_14/given_a_choice_ofcategories
should assume33.2534.67paws/labeled_final/context-question34.92paws/labeled_final/context-question-no-label
take the following as truth33.4232.75social_i_qa/Show choices and generate index36.67paws/labeled_final/context-question-no-label
Avg.33.3333.8634.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.283, + 0.86, + 0.298 + ], + "angle": 0, + "content": "Table 8: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R3." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.333, + 0.859, + 0.447 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
exercise80.0079.00cos_e/v1.11/question_option_description_text80.00cos_e/v1.11/question_option_description_text
plausible_alternatives84.0083.00cos_e/v1.11/question_option_description_text83.00cos_e/v1.11/question_option_description_text
"C1 or C2? premise, so/because..."61.0067.00social_i_qa/Check if a random answer is valid or not75.00cos_e/v1.11/descriptionQuestion_option_text
best_option70.0076.00social_i_qa/Show choices and generate answer79.00cos_e/v1.11/question_option_description_text
more likely79.0083.00cos_e/v1.11/question_option_description_text85.00cos_e/v1.11/descriptionQuestion_option_text
cause-effect74.0078.00cos_e/v1.11/question_option_description_text83.00common_gen/random task template prompt
choose80.0082.00cos_e/v1.11/question_option_description_text82.00cosmos_qa/no_prompt_text
i_am_hesitating79.0082.00cos_e/v1.11/question_option_description_text82.00cos_e/v1.11/question_option_description_text
Avg.75.8878.7581.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.455, + 0.851, + 0.47 + ], + "angle": 0, + "content": "Table 9: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of COPA." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.505, + 0.807, + 0.59 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Predict ending with hint26.8327.70social_i_qa/Show choices and generate index29.11cos_e/v1.11/question_option_description_text
Randomized prompts template26.8726.83social_i_qa/Show choices and generate index27.84wiqa/what_is_the MISSING_first_step
complete_first THEN28.3727.35ropes/prompt_bottom_no_hint28.35cos_e/v1.11/question_option_description_text
if_beginnerscontinue25.2826.58social_i_qa/Show choices and generate index26.58cos_e/v1.11/question_option_description_text
how Ends25.6726.86social_i_qa/Show choices and generate index26.86cos_e/v1.11/question_option_description_text
Avg.26.6027.0627.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.599, + 0.868, + 0.615 + ], + "angle": 0, + "content": "Table 10: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Hellaswag." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.648, + 0.822, + 0.733 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Answer Given options86.4887.76social_i_qa/Show choices and generate answer87.76social_i_qa/Show choices and generate answer
Choose Story Ending86.9188.24social_i_qa/Show choices and generate answer88.24cos_e/v1.11/question_option_description_text
Movie What Happens Next78.0978.03social_i_qa/Show choices and generate answer87.33cosmos_qa/no_prompt_text
Story Continuation and Options83.5985.89social_i_qa/Show choices and generate answer86.53social_i_qa/Show choices and generate answer
Novel Correct Ending85.0987.65social_i_qa/Show choices and generate answer87.97social_i_qa/Show choices and generate index
Avg.84.0385.5287.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.743, + 0.87, + 0.758 + ], + "angle": 0, + "content": "Table 11: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of StoryCloze." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.793, + 0.785, + 0.877 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does underscore refer to stand for49.2551.54cos_e/v1.11/question_option_description_text51.78paws/ labeled_final/ PAWS-ANLI GPT3
50.5150.59cos_e/v1.11/question_option_description_text52.09ropes/plain_nobackground
underscore refer to fill in the blank50.4350.59cos_e/v1.11/question_option_description_text52.33ropes/prompt_bottom_no_hint
52.1751.38cos_e/v1.11/question_option_description_text52.01cos_e/v1.11/question_option_description_text
Replace52.4953.59cos_e/v1.11/question_option_description_text53.59paws/ labeled_final/ paraphrase-task
Avg.50.9751.5452.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.886, + 0.875, + 0.902 + ], + "angle": 0, + "content": "Table 12: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Winogrande." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12308" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.21, + 0.842, + 0.341 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does the pronoun refer to by p they mean68.2766.34social_i_qa/Check if a random answer is valid or not66.35paws/labeled_final/PAWS-ANLI GPT3
62.5066.35social_i_qa/Check if a random answer is valid or not69.23glue/qqp/duplicate or not
in other words67.3167.31social_i_qa/Show choices and generate answer72.12samsum/Write a dialogue that match this summary
I think they mean69.2368.27social_i_qa/Show choices and generate answer78.08social_i_qa/Generate answer
replaced with64.4259.62social_i_qa/Check if a random answer is valid or not66.35social_i_qa/Show choices and generate index
p is/are r62.5064.42social_i_qa/Show choices and generate index65.38social_i_qa/Show choices and generate index
the pronoun refers to64.4264.42social_i_qa/Show choices and generate index67.31social_i_qa/Check if a random answer is valid or not
Who or what is/are64.4263.46social_i_qa/Show choices and generate index65.38samsum/To sum up this dialog
does p stand for67.3163.46social_i_qa/I was wondering69.23samsum/Write a dialogue that match this summary
GPT-3 Style60.5867.31social_i_qa/Show choices and generate index67.31rotten_tomatoes/Reviewer Expressed Sentiment
Avg.65.1065.1068.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.351, + 0.851, + 0.366 + ], + "angle": 0, + "content": "Table 13: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WSC." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.632, + 0.877, + 0.763 + ], + "angle": 0, + "content": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
question-context-meaning-with-label50.3153.76glue/qqp/meaning53.76paws/labeled_final/context-question-no-label
question-context-meaning50.6350.16ropes/prompt_bottom_no_hint57.05paws/labeled_final/PAWS-ANLI GPT3
grammar Homework49.8450.16ropes/prompt_bottom_no_hint57.99rotten_tomatoes/Reviewer Enjoyment Yes No
affirmation_true_orfalse49.6951.57paws/labeled_final/Rewrite-no-label53.92paws/labeled_final/Meaning-no-label
GPT-3-prompt51.7250.00social_i_qa/Show choices and generate index55.64paws/labeled_final/PAWS-ANLI GPT3
same_sense49.8452.82paws/labeled_final/Rewrite55.17paws/labeled_final/task_description-no-label
question-context51.8853.29social_i_qa/Check if a random answer is valid or not57.05amazon_polarity/Is.this_product_review_positive
GPT-3-prompt-with-label50.4752.19glue/qqp/meaning52.66glue/qqp/meaning
polysemous50.0052.82paws/labeled_final/Rewrite53.29paws/labeled_final/Rewrite-no-label
similar-sense52.5150.00social_i_qa/Show choices and generate index56.11paws/labeled_final/task_description-no-label
Avg.50.6951.6855.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.772, + 0.848, + 0.786 + ], + "angle": 0, + "content": "Table 14: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WiC." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12309" + } + ] +] \ No newline at end of file diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_origin.pdf b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c7d5d6226ee72649531f22b278c2ec8e6871bcbb --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/91bfba92-6122-4a1c-8806-4f5015582836_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c630934be9e52de7d0abba370c506747312f332fd2926ed49db27c62e5e18277 +size 2377596 diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/full.md b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a05d6e20bc5fff7ea913f576e3e17aa16cdf7e27 --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/full.md @@ -0,0 +1,489 @@ +# Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt + +Seonghyeon Ye Joel Jang Doyoung Kim Yongrae Jo Minjoon Seo KAIST + +{seonghyeon.ye,joeljang,ikevin98,yongrae,minjoon}@kaist.ac.kr + +# Abstract + +Enhancing the zero-shot performance of instruction-following models requires heavy computation, either by scaling the total number of training datasets or the model size. In this work, we explore how retrieval of soft prompts obtained through prompt tuning can efficiently assist hard prompts in zero-shot task generalization. Specifically, we train soft prompt embeddings for each prompt through prompt tuning, store the samples of the training instances mapped with the prompt embeddings, and retrieve the corresponding prompt embedding of the training instance closest to the query instance during inference. While only adding $0.007\%$ additional parameters, retrieval of soft prompt enhances the performance of T0 on unseen tasks by outperforming it on 10 out of 11 datasets as well as improving the mean accuracy of T0 on BIG-bench benchmark by $2.39\%$ points. Also, we report an interesting finding that retrieving source embeddings trained on similar answer choice formats is more important than those on similar task types. + +# 1 Introduction + +Training Large Language Models (LLMs) on huge amounts of data has enabled LMs to perform downstream tasks without any fine-tuning with the aid of natural prompts or concatenation of a few demonstration instances (Brown et al., 2020; Rae et al., 2021; Kojima et al., 2022; Chowdhery et al., 2022). Additionally, recent works have shown that adding a instruction tuning stage, an additional training step that helps pretrained LMs understand prompts and demonstrations results in a significant performance boost on zero-shot task generalization even for moderate-sized LMs (Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Ye et al., 2022; Chung et al., 2022). This extra + +![](images/5f317a13ed48bf7b1b98fcbabeec569fbb45be92a01ee06ccd631a599e8f6e40.jpg) +Figure 1: During zero-shot inference, ROSPR selects similar training instances with the given input from the Source Prompt Library and retrieves the prompt embeddings corresponding to the selected training instances. + +instruction-tuning stage involves explicit, multitask prompted learning on various tasks, enabling LMs to quickly adapt to unseen tasks at inference. + +To maximize the effect of instruction-tuning, two approaches have been widely explored: (1) scaling the number of training datasets, and (2) scaling the model size (Wang et al., 2022b; Chung et al., 2022). However, both approaches require heavy computation, not applicable with an academic budget. Specifically, the first approach requires updating the whole parameters of the model every time a training dataset is added, showing limitations in terms of scalability. On the other hand, the second approach requires heavy memory requirements to load and train a massive LLM. + +To enhance the zero-shot performance of instruction-following model efficiently, we introduce Retrieval of Soft Prompt (ROSPR), which is easily scalable and requires minimal computation by only adding $0.007\%$ parameters to the main model during inference. As shown in Figure 1, by training prompt embeddings (soft prompt) for each given hard prompt through prompt tuning, we construct a Source Prompt Library consisting of samples of training instances mapped with their corresponding prompt embeddings. Then, during inference, by using a simple, off-the-shelf dense retriever model, we search for training instances similar to the given query instances and retrieve their corresponding prompt embeddings. Because the backbone LM is frozen, it allows the retrieved + +embeddings to serve as adapters assisting hard prompts. While RoSPR can be applied to any LM, in this work, we use T0 (Sanh et al., 2021) as our initial backbone LM and perform prompt tuning on the tasks used during the instruction-tuning stage. + +While adding only $0.007\%$ additional parameters, ROSPR outperforms T0 on 10 out of 11 evaluation datasets and outperforms efficient fine-tuning baselines without any target task fine-tuning. ROSPR is also effective for challenging tasks such as tasks from BIG-bench (Srivastava et al., 2022), outperforming T0 by $2.39\%$ mean accuracy. Furthermore, we provide several interesting findings: (1) Variants of ROSPR that include interpolation of multiple prompt embeddings and scoring method that considers the answer choice distribution during retrieval further increases the effect of ROSPR (2) Also, we provide analysis of which factors attribute to the performance of ROSPR and show that, similarly to the role of demonstrations in in-context learning (Min et al., 2022), heuristic features such as answer choice format are more important than the similarity of the source task. + +# 2 Related Work + +# 2.1 Task Generalization with Instruction-Tuning + +Prompts and demonstrations are essential for task generalization since proper explanations are required for LMs to understand an unseen task (Kojima et al., 2022; Wei et al., 2022; Lampinen et al., 2022). Instruction-tuning, which is explicit multi-task prompted training on various downstream tasks, is a simple but effective way to achieve this, resulting in improved zero-shot capabilities. Zhong et al. (2021) first introduced the method of instruction-tuning by converting various tasks into a question-answering format and finetuning the model on the aggregated dataset. Following works (Mishra et al., 2022; Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Xu et al., 2022; Ouyang et al., 2022; Ye et al., 2022; Chung et al., 2022) extended this approach on a larger scale and show that zero-shot task generalization could be enhanced with more diverse prompts, a larger number of training downstream tasks, and a larger LM. + +# 2.2 Source Task Retrieval + +Retrieving a source task that is relevant to the target task has shown to result in faster and better task + +adaptation. For parameter-efficient fine-tuning, Vu et al. (2022); Su et al. (2022) retrieve source prompt embedding that is similar to the target prompt embedding and obtain a better initialization point for prompt tuning. Instead of utilizing a single prompt embedding, recent works show a mixture of multiple prompt embeddings to be effective (Asai et al., 2022; Qin and Eisner, 2021). + +For instruction-tuning, Lin et al. (2022) retrieve training instances that are similar to the query through a dense retriever and fine-tune the model using the retrieved examples. For in-context learning, Rubin et al. (2021); Liu et al. (2022b); Wang et al. (2023) retrieve training data that could be used for demonstrations. Wang et al. (2022c) show the effect of retrieving prompt embeddings in a continual learning setting. Although our proposed method is related to these works, the novelty of our work lies in applying source task retrieval in the zero-shot setting and retrieving soft prompts instead of training instances. + +# 3 Method + +In this section, we introduce Retrieval of Prompt Tuning (ROSPR) for zero-shot task generalization. A detailed overview is shown in Figure 2. We first train source prompt embeddings of LM for each hard prompt given a source task using prompt tuning (Section 3.1). Then, we save training instance samples along with their prompt embeddings in the Source Prompt Library and use it to retrieve embeddings at inference to perform tasks in a zero-shot manner (Section 3.2). We additionally introduce interpolation of multiple source prompt embeddings (Section 3.3) and variance-based ranking (Section 3.4) to increase robustness and accuracy. + +# 3.1 Training Source Prompt Embeddings + +Even though RoSPR may be used to augment any type of LM, we use T0 (Sanh et al., 2021) as the backbone LM for this paper. For training of soft prompts, we utilize the source tasks and prompts used for the instruction-tuning phase of T0. While T0 was trained in a multi-task learning manner, we freeze the initial T0 parameters and train only soft prompts (source prompt embeddings) for each hard prompt of the source task. + +Prompt Tuning Among various parameter-efficient fine-tuning methods, we follow prompt tuning proposed by Lester et al. (2021) because the number of trainable parameters is extremely small + +![](images/7d9561a1f59a0b1caf4a6392d148c33d0c316d8439c27366ca7da20fc3a899fb.jpg) +Figure 2: An overview of RoSPR. For each hard prompt of the source datasets, soft prompts are trained via prompt tuning. After storing training instances as keys and corresponding prompt embedding as values, RoSPR searches training instances similar to query set $Q$ , retrieves the corresponding prompt embeddings, and selects the most frequently retrieved candidate for inference. Variants of selection strategy are also shown: RoSPR+INTER interpolates between multiple related source embeddings and RoSPR+VAR ranks candidate embeddings considering both frequency and variance. + +$(\sim 204\mathrm{K}$ parameters per prompt), which implies that the memory overhead of parameter retrieval at inference is negligible. + +For each source training dataset $D_{i}$ $(i = 1,\dots,T)$ where $T$ is the total number of source datasets, we train source embeddings $E_{ij}$ $(j = 1,\dots,M_i)$ where $M_{i}$ is the number of hard prompts in $D_{i}$ , making soft prompt embeddings for each individual hard prompts. Specifically, given a training instance $\{x_{ik},y_{ik}\} (k = 1,\dots,K)$ from $D_{i}$ where $K$ is the number of sampled training instances per dataset, we first convert it into its hard prompted version $\{h_j(x_{ik}),h_j(y_{ik})\}$ where $h_j(\cdot)$ denotes adding the $j$ -th hard prompt. Next, we train the LM with the following objective: + +$$ +\max _ {E _ {i j}} P \left(h _ {j} \left(y _ {i k}\right) \mid E _ {i j}; h _ {j} \left(x _ {i k}\right)\right) \tag {1} +$$ + +where all the parameters of the underlying backbone LM are frozen and only $E_{ij}$ is trainable. In short, given $D_{i}$ , we perform $M_{i}$ number of prompt tunings for each hard prompts, resulting in $\sum_{i=1}^{T} M_{i}$ total number of source prompt embeddings. For training efficiency, we only train on $K = 5000$ training instances for a single epoch for each source prompt embedding. + +# 3.2 Zero-Shot Embedding Retrieval + +After source prompt embedding training, we retrieve the most related source embeddings and select one from the retrieved candidates to be used at inference (right part of Figure 2). + +We first construct a Source Prompt Library, consisting of sentence-level representations of training instance inputs as keys and the corresponding source prompt embedding as the values. For each available source prompt embedding, $n$ number of samples are stored in the library. The sentence-level representations are obtained by getting the mean representation of hidden states of the last layer of the dense retriever. We use a T0-small encoder as a dense retriever, replicated based on Sanh et al. (2021) with smaller model size. + +At inference, we first randomly sample $Q$ query instances from the target task, following Lin et al. (2022). After obtaining sentence-level representations for each query through our T0-small encoder, we retrieve top- $N$ examples for each query instance using MIPS (maximum inner product search) operation on our Source Prompt Library, retrieving a total of $Q*N$ prompt embeddings. As the default methodology, among the retrieved embedding candidates, we select the most frequently retrieved prompt embedding as our designated soft prompt for the given target task and concatenate the embedding with each of the target task instances before feeding it to our backbone LM. In the next two subsections, we explain different strategies for calculating the target embedding from the $Q*N$ prompt embedding candidates. + +# 3.3 Interpolation of Prompt Embeddings + +When retrieving only a single prompt embedding for a given task (Section 3.2), it may result in high variance across evaluation prompts when the selected prompt embedding does not fit well with the + +given task. Recent works on prompt embedding retrieval have shown that the interpolation of prompt embeddings effectively transfers to the target task (Asai et al., 2022; Vu et al., 2022). We also explore calculating the target embedding through interpolation of multiple source embeddings instead of just using a single embedding. Among $Q * N$ prompt candidates searched in Section 3.2, we select top- $N'$ candidate embeddings based on the frequency of the search. Then, we calculate the weighted sum of the candidate embeddings, where the interpolation weight for each source embedding is based on the proportion of frequency. While Asai et al. (2022); Vu et al. (2022) require fine-tuning the target embeddings on the target task to calculate the weights for interpolation, our approach does not require any target task fine-tuning, enabling zero-shot task transfer. + +# 3.4 Variance-based Ranking + +Similar to the scoring and calibration method of Lu et al. (2022); Zhao et al. (2021), we introduce a scoring method applicable to zero-shot classification tasks that allows ranking the $Q * N$ retrieved prompt embedding candidates by considering the answer choice distribution of the given target task as extra cues together with the original frequency cues. To accomplish this, we perform a forward pass with the concatenation of each candidate prompt embeddings together with the given hard prompt (only including the instruction, excluding the input instance) of the target task and give a higher score to the embedding candidate that results in lower variance. Ideally, the combination of soft and hard prompts should result in equal probability among the answer choices because the actual context of the task is not included. + +Specifically, when given a target task with $k$ -th hard prompt $h_k$ , for each candidate embedding $E_{ij}$ we calculate the scoring as follows: + +$$ +\operatorname {S c o r e} \left(h _ {k}, E _ {i j}\right) = \frac {\operatorname {f r e q} \left(h _ {k} , E _ {i j}\right)}{\sqrt {\operatorname {V a r} \left[ P \left(y \mid E _ {i j} , h _ {k}\right) \right]}} \tag {2} +$$ + +where $y$ refers to the available output options of the target task. + +# 4 Experimental Settings + +In this section, we explain the experimental settings of training of source prompt embedding and construction of our Source Prompt Library. We also + +explain our evaluation setting during zero-shot inference and baseline models. We provide detailed experiment configurations in Appendix F. + +# 4.1 Source Tasks + +For training soft prompts through prompt tuning, we use the subset of source tasks used for the initial T0 instruction-tuning (Sanh et al., 2021) $^4$ . For each source task, we use the prompts for each dataset in T0, resulting in a total of 230 prompts. For Source Prompt Library construction, we sample only $n = 100$ training instances per source embedding to minimize the inference latency. We show a variation of $n$ and different methods to sample $n$ training instances in Appendix D. + +# 4.2 Evaluation Tasks + +Following Sanh et al. (2021), we evaluate on the validation set of 4 held-out tasks (natural language inference, sentence completion, coreference resolution, word sense disambiguation) resulting in a total of 11 evaluation datasets. We also follow Sanh et al. (2021) and evaluate on 14 different datasets from the BIG-bench benchmark (Srivastava et al., 2022). We use rank classification evaluation method by selecting the output option with higher log-likelihood following Brown et al. (2020); Sanh et al. (2021). For all evaluation tasks, we use accuracy as an evaluation metric and report the mean accuracy and standard deviation of all of the evaluation prompts per given dataset (average of $\sim 10$ prompts per evaluation dataset). For BIG-bench tasks, we do not report standard deviation because only one prompt is provided per task. + +# 4.3 Baseline Models + +Zero-shot Baseline For zero-shot baseline models, we show the results of T0 (3B) together with a 4 times larger T0 (11B) instruction-tuned model. We also compare with GPT-3 (175B) model which is 60 times larger than T0 (3B). + +Fine-tuning Baseline We also compare with efficient fine-tuning baseline models that utilize prompt tuning. These models require target task prompt tuning, indicating that zero-shot transfer + +
Method# of Param +(Base/Trainable)NLISentence CompletionCoreference Resolut.WSDTotal Avg.
RTECBAN. R1AN. R2AN. R3COPAHellasw.StoryC.Winogr.WSCWiCMeanSTD
T03B / 064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.223.62
PT (FT)3B / 204K63.1444.5233.0731.4432.9473.0026.39-50.6746.7350.02--
ATTEMPT (FT)3B / 614K68.9544.8836.1934.7334.8174.3826.76-51.3364.9051.05--
T0+RoSPR3B / 204K71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.243.62
w/ INTER3B / 204K70.7152.3037.3034.3433.8978.2526.9485.6251.1064.5250.7353.243.30
w/ VAR3B / 204K71.7850.3637.0734.5833.9078.8827.0185.5251.4564.9451.9453.383.38
w/ VAR & INTER3B / 204K72.6051.9837.2534.3133.9577.8326.8485.5850.9364.9751.1853.403.47
ORACLE3B / 204K73.7958.1037.6534.9234.9181.1327.7587.5752.3668.1755.2655.603.07
T011B / 080.8370.1243.5638.6841.2690.0233.5892.4059.9461.4556.5860.76
GPT-3175B / 063.5046.4034.6035.4034.5091.0078.9083.2070.2065.4045.9259.00-
+ +Table 1: RoSPR refers to our main proposed method, W/ INTER refers to applying interpolation of multiple source embedding candidates, W/ VAR refers to retrieval through variance-based ranking, W/ VAR & INTER refers to applying both interpolation and variance-based ranking where the interpolation weight is based on the variance-based ranking score, and ORACLE refers performance when the most optimal source embedding is retrieved from the candidates, acting as an upper bound performance for retrieval. FT refers to fine-tuned models on the target tasks. For FT models, we exclude StoryCloze due to the absence of training instances. The best and second-best performance is shown in bold and underline respectively. Comparison with hard prompt optimization techniques and visualization of the results is shown in Appendix A and Appendix E, respectively. + +is infeasible. Similar to our source prompt tuning process, we train each target prompt for a single epoch with a maximum of 5,000 training instances. The first baseline model is naive prompt tuning on the target tasks without any prompt retrieval, referred to as PT (Lester et al., 2021). The second model is ATTEMPT (Asai et al., 2022), which trains the target soft prompts through attentional mixtures of source prompts. Because StoryCloze (Mostafazadeh et al., 2016) does not contain training instances, we exclude the dataset for fine-tuning. More training details of fine-tuning baseline are specified in Appendix C. + +# 5 Experimental Results + +RoSPR enhances the performance of T0 efficiently. Table 1 shows the performance of the 11 evaluation datasets. T0+RoSPR outperforms T0 on 10 datasets among the 11 evaluation datasets. Specifically, T0+RoSPR outperforms T0 on RTE (+6.99% points), CB (+4.12% points), ANLI R1 (+3.24% points), and COPA (+2.87% points). This shows that soft prompt retrieval assists hard prompts for zero-shot task generalization with a negligible number of additional parameters (0.007%). Also, while T0 outperforms GPT-3 on 3 datasets (RTE, StoryCloze, WiC), T0+RoSPR additionally outperforms GPT-3 on 2 datasets (ANLI R1 and CB) and enlarging the score gap for RTE, StoryCloze and WiC. + +ROSPR also outperforms finetuning baselines even without utilizing any training instances of the target task. We first observe that PT harms the performance of the backbone model, which aligns with the result of Liu et al. (2022a); Gu et al. (2022) that prompt tuning is unstable when the training instances or the training steps are small. By comparing ATTEMPT with ROSPR, ROSPR outperforms ATTEMPT on 7 out of 10 tasks and $1.21\%$ points on the mean accuracy of 10 tasks. This shows that ROSPR is more applicable for efficient adaptation because it requires 3 times fewer additional parameters compared to ATTEMPT and also does not require any further fine-tuning of the target task. + +INTER and VAR enhance the performance of ROSPR. We also analyze the effect of introducing variants of ROSPR: interpolation of soft prompts (INTER) and variance-based ranking (VAR) in Table 1. First, applying INTER shows similar accuracy compared to ROSPR. However, as shown in the last column of Table 1, INTER reduces the standard deviation of T0 and ROSPR by $8.84\%$ while improving the mean accuracy of T0, indicating increased robustness to different surface forms of evaluation prompts. This indicates that interpolation of multiple source embeddings outperforms a single source embedding retrieval, aligning with the result of Asai et al. (2022). Applying VAR with $\mathrm{T0 + ROSPR}$ improves both zero-shot accuracy and robustness of $\mathrm{T0 + ROSPR}$ , showing that considering the answer choice distribution is beneficial for zero-shot setting, aligned with results from Zhao et al. (2021); Shi et al. (2022). Moreover, applying both VAR+INTER results in the highest overall average accuracy, outperforming T0 by $2.18\%$ points by largely reducing the gap between larger LLMs. + +![](images/5ac012d05c905fb7395d0a8393846221f227d53d80a09f3bab98b34d3d8e99bf.jpg) +Figure 3: Mean accuracy of 14 datasets of BIG-bench. We evaluate on a single prompt following Sanh et al. (2021). By only adding $0.007\%$ parameters to T0-3B, $\mathrm{T0 + ROSPR}$ largely reduces the performance gap between 4 times larger T0-11B. The full result is provided in Appendix B. + +The effect of ROSPR generalizes to challenging tasks. ROSPR is also effective for challenging tasks such as tasks from BIG-bench benchmark. As shown in Figure 3, T0+ROSPR improves the mean accuracy performance of T0-3B by $2.39\%$ points while only adding $0.007\%$ additional parameters. T0+ROSPR also outperforms 60 times larger zero-shot and 1-shot GPT-3 and largely reduces the performance gap between 4 times larger T0-11B (1.84% points) or 60 times larger 3-shot GPT-3 (0.53% points). Applying INTER with T0+ROSPR results in additional mean accuracy enhancement, outperforming T0-3B by $2.67\%$ points8. + +# 6 Analysis of ROSPR + +Zero-shot task adaptation of LMs is often seen as a problem of task location, locating the target task to where the model can solve it using the intrinsic ability obtained at pretraining stage with the aid of prompts and demonstrations (Reynolds and McDonell, 2021). In this section, we analyze which factors contribute to the performance enhancement in the perspective of identifying better task location. We find that although the target task performance depends on the source task types, heuristic features such as the answer choice format are more important. This agrees with previous findings that a met-trained LM focuses on simple features such as the label space, the input distribution, and sequence format, instead of complex semantics (Webson and Pavlick, 2021; Min et al., 2022). + +Target task performance depends on source task types. To analyze the effect of different source task types on each target task, we measure the frequency ratio of each source task type that results + +![](images/f31b568c9169f5ed17a16f3d6271df6eb735e813c92b9945d226e18762868bc8.jpg) +Figure 4: Frequency of source task types (x-axis) that maximizes (i.e. ORACLE) the accuracy of each target task (y-axis). + +in the best performance (ORACLE) for the given prompts of the target tasks (visualized in Figure 4). From this figure, we can observe a few patterns: paraphrase task assists NLI and word sense disambiguation task while multi-choice QA (MQA) task assists sentence completion task. For coreference resolution task, various source task types (paraphrase, summarization, multi-choice QA) assist the target task. + +![](images/e471b75fad8028d974258ce484f8e07e0b8f08de5928b1e837b9038f781e0359.jpg) +Figure 5: Effect of answer choice format alignment across different target datasets (RTE, COPA, WiC). We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line. + +Answer choice format is important for task location. We also analyze the effect of using different answer choice formats with the same source task. Answer choice format decides how the available answer choices are given to the LM through the input. For example, a prompt that requires classifying a movie review into good/bad has a different answer choice format from classifying it into positive/negative. + +We experiment on 3 datasets (RTE, COPA, WiC) which correspond to different tasks (NLI, sentence completion, word sense disambiguation) respectively. For each dataset, we select a source dataset that is retrieved the most for ORACLE. Among the source prompts of the selected source dataset, we select a prompt that has the same answer choice format as the target task (ALIGNED) and another prompt that has a different answer choice format (MISALIGNED). Figure 5 shows the effect of answer choice format alignment on the target task + +performance by comparing ALIGNED and MIS-ALIGNED. The result shows that for all 3 datasets, ALIGNED significantly outperforms MISALIGNED. This result is non-trivial considering that the two prompt embeddings are trained on the same source training dataset and the same training configuration, with the only difference in the given answer choice format, implying that how the answer choices are given to solve a specific task is more important than the content of the training data for task location. ROSPR is mostly comparable to ALIGNED embedding, implying that retrieving a source prompt embedding by searching for similar input instances results in retrieving a source embedding with similar answer choice formats. + +![](images/273c898d6a37ac95ab5400e985e36458d79a5b3199a888f1e549dd322b7aa5ba.jpg) + +![](images/9a0360ca8488b3b4ce587546cde7130424806ceb3a518be9d3e8b43cdc6ad70e.jpg) +Figure 6: RTE (Top) and WiC (Bottom) evaluation result by retrieving MISALIGNED and ALIGNED answer choice format across various source tasks. We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line. + +We additionally analyze the effect of answer choice formats on RTE and WiC datasets by retrieving prompt embeddings trained on various source tasks. Both target datasets have the answer choices of either yes/no. Similar to the previous experiments, we retrieve ALIGNED (yes/no format) and MISALIGNED prompt embeddings across three source tasks: paraphrase, sentiment classification, and multi-choice QA. As shown in Figure 6, for both target datasets, ALIGNED outperforms MISALIGNED across all three source tasks. This shows that aligning to the answer choice format of the target task is crucial regardless of the retrieved source task. + +Answer choice format is more important than task similarity. From Figure 6, we can see that all three source tasks benefit from aligning to the target task answer choice format. One may think + +that embeddings from source tasks requiring similar knowledge to the target task may be important. Counterintuitively, for both RTE and WiC target tasks, when the answer choice format is aligned, the task source embedding of sentiment classification, which is known to be irrelevant to RTE and WiC (Pruksachatkun et al., 2020), outperforms other embeddings sourced from datasets that are more relevant to the target datasets (paraphrase and multi-choice QA) (Appendix G). This implies that for retrieval of source embedding for task location, answer choice format is more important than containing similar knowledge required to solve the target task. + +Role of ROSPR is similar to in-context learning. From the findings explained in previous paragraphs, we can conclude that although the source task types influence the target task performance, retrieving a similar answer choice format is more important for task location. Indeed, source tasks containing similar knowledge can help target tasks only if the answer choice formats are aligned to the target task. These findings support Min et al. (2022); Webson and Pavlick (2021) that a meta-trained LM "takes less effort" to understand the input: models exploit the simple aspects of prompts and demonstrations such as the format and distribution instead of complex semantics. Especially, for in-context learning, Xie et al. (2021); Min et al. (2022) show that the role of demonstrations lies in providing the shared concept and distribution hints of the target task. From this aspect, the role of ROSPR is similar to demonstrations. However, it is more efficient than including demonstrations because it avoids heavy computation at inference from long sequence lengths (Liu et al., 2022a; Choi et al., 2022) since ROSPR prepends a fixed length of prefix tokens regardless of the task. Also, ROSPR is free from the instability of in-context learning coming from different orderings of demonstrations (Lu et al., 2022; Zhao et al., 2021). Lastly, we conjecture that ROSPR also has the benefits of soft prompts (Li and Liang, 2021) such as having more expressiveness. + +# 7 Ablation Studies + +In this section, we analyze the effect of the number of (1) prompts, (2) source datasets, and (3) queries sampled for evaluation. We evaluate variations of our proposed methods on 4 datasets: RTE for NLI, COPA for sentence completion, Winogrande + +![](images/1defe8e68d4ed49a58d8f3c1201254db1ffa3563b1d5ecf359d4894f1cdb96b9.jpg) +(a) Datasets vs Prompts +Figure 7: Result of various ablation settings of ROSPR. (a) compares the effect of scaling the number of datasets with scaling the number of prompts and (b) shows the effect of the number of sampled queries at inference. Additional ablation results are shown in Appendix D. + +![](images/f71eb27c7fedf771f1e614bb3dba7a2722033716596e727b6f5cae0fcd0d36bf.jpg) +(b) Query variation + +for coreference, and WiC for word sense disambiguation task. We report the average of the mean accuracy of all evaluation prompts for each dataset by running 3 different runs. + +Scaling number of prompts vs. number of datasets. Recent works on meta-trained LMs show that the number of source datasets and prompts is an important factor for zero-shot task generalization (Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Chung et al., 2022). We also show ablations for ROSPR and measure how the zero-shot generalization performance changes when we vary the number of prompts and datasets available during the prompt tuning stage (shown in Figure 7a). First, we vary (1) the total number of source prompts by 60, 120, and 230 by increasing the number of prompts per dataset and (2) the number of datasets by 8, 16, and 30 by increasing the number of datasets per task cluster. Note that in (1), the total number of datasets is fixed while in (2), we use all available prompts for each dataset while varying the number of datasets per task cluster. + +In contrast to (1), (2) does not always lead to a linearly increasing performance boost; the performance saturates as more source datasets are included. By comparing the effect of scaling datasets and scaling prompts for similar Source Prompt Library sizes, we observe that the number of prompts has more impact on the accuracy of the target task (Figure 7a). + +This ablation study also supports the analysis of the previous section; diverse answer choice formats of prompts, which are mostly influenced by the total number of source prompts, are more important than source task types which are influenced by the number of source datasets.[10] Therefore, if + +the number of task clusters is sufficient to some extent, scaling the number of source prompts per dataset is more crucial than scaling the number of source datasets per task cluster. + +More sampled queries improve the performance. We also analyze the effect of the number of query instances sampled at inference for retrieval. As seen in Figure 7b, increasing the number of queries results in higher mean accuracy. This is different from the analysis of Lin et al. (2022) that sampling more queries leads to better performance only to some point. Because we use the frequency of each prompt embedding candidate as the default metric for retrieval, utilizing more query instances would represent the evaluation data more accurately, resulting in a reduced number of wrong retrievals. + +# 8 Conclusion + +In this paper, we introduce ROSPR, a method that efficiently enhances zero-shot generalization capabilities of a meta-trained LM by retrieving prompt-specific source prompt embeddings (soft prompts) for a given target task. We accomplish this by first training the soft prompts for hard prompt of the source tasks. After training source prompt embeddings, we construct the Source Prompt Library by storing the mean representation of training instances as keys and the corresponding prompt embeddings as values. At inference, we search for training instances stored in the library similar to sample instances from the target task, retrieve the corresponding prompt embedding, select the most frequently retrieved embedding, and append it to each of the target task instances for prediction. Our results show that ROSPR efficiently enhances the zero-shot performance of the backbone model while introducing minimal additional parameters during inference. We additionally provide analysis of which factors attribute to the performance of ROSPR and find that heuristic cues such as the answer choice format are critical for generalization performance, implying that it may play a role similar to demonstrations in in-context learning. + +# 9 Limitations + +Although we show the effectiveness of ROSPR by applying it on T0-3B (Sanh et al., 2021), we did + +not evaluate our method on different model scales such as the T0-11B variant and other LM architectures such as decoder-only LMs due to limited computational resources. This leaves future works on applying ROSPR to even larger LMs and diverse LM architectures (Wang et al., 2022a). Moreover, it is hard to apply VAR to target tasks without answer choices such as free-form generation because variance among options cannot be obtained. However, ROSPR and ROSPR+INTER can still be utilized and we leave applying ROSPR on zero-shot task location of free-form generation as future work (Scialom et al., 2022). + +# Acknowledgements + +We thank Sohee Yang and Eunbi Choi for helpful feedback on the paper. This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00264, Comprehensive Video Understanding and Generation with Knowledge-based Deep Logic Neural Network, $80\%$ ; No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST), $20\%$ ). + +# References + +Akari Asai, Mohammadreza Salehi, Matthew E Peters, and Hannaneh Hajishirzi. 2022. Attentional mixtures of soft prompt tuning for parameter-efficient multi-task knowledge sharing. arXiv preprint arXiv:2205.11961. +Stephen Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-david, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Fries, Maged Alshaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang, Dragomir Radev, Mike Tian-jian Jiang, and Alexander Rush. 2022. Prompt-Source: An integrated development environment and repository for natural language prompts. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 93-104, Dublin, Ireland. Association for Computational Linguistics. +Max Bartolo, Alastair Roberts, Johannes Welbl, Sebastian Riedel, and Pontus Stenetorp. 2020. Beat the AI: Investigating adversarial human annotation for reading comprehension. Transactions of the Association for Computational Linguistics, 8:662-678. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind + +Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901. +Eunbi Choi, Yongrae Jo, Joel Jang, and Minjoon Seo. 2022. Prompt injection: Parameterization of fixed inputs. +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. +Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416. +Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine learning challenges workshop*, pages 177-190. Springer. +Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107-124. +Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 70–79, Hong Kong, China. Association for Computational Linguistics. +David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2003. English gigaword. Linguistic Data Consortium, Philadelphia, 4(1):34. +Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. 2022. PPT: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, Dublin, Ireland. Association for Computational Linguistics. +Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2391-2401, Hong Kong, China. Association for Computational Linguistics. +Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547. + +Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916. +Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. 2022. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329. +Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1203-1213, Austin, Texas. Association for Computational Linguistics. +Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, D. Kontokostas, Pablo N. Mendes, Sebastian Hellmann, M. Morsey, Patrick van Kleef, S. Auer, and C. Bizer. 2015. Dbpedia - a large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6:167-195. +Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning. +Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierric Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics. + +Chonghua Liao, Yanan Zheng, and Zhilin Yang. 2022. Zero-label prompt selection. +Bill Yuchen Lin, Kangmin Tan, Chris Miller, Beiwen Tian, and Xiang Ren. 2022. Unsupervised cross-task generalization via retrieval augmentation. arXiv preprint arXiv:2204.07937. +Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823-1840, Online. Association for Computational Linguistics. +Kevin Lin, Oyvind Tafjord, Peter Clark, and Matt Gardner. 2019. Reasoning over paragraph effects in situations. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 58-62, Hong Kong, China. Association for Computational Linguistics. +Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022a. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. arXiv preprint arXiv:2205.05638. +Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022b. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics. +Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics. +Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics. +Julian J. McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In Seventh ACM Conference on Recommender Systems, RecSys '13, Hong Kong, China, October 12-16, 2013, pages 165-172. ACM. +Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943. + +Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettle-moyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837. +Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470-3487, Dublin, Ireland. Association for Computational Linguistics. +Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849, San Diego, California. Association for Computational Linguistics. +Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797-1807, Brussels, Belgium. Association for Computational Linguistics. +Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics. +Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155. +Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05), pages 115-124, Ann Arbor, Michigan. Association for Computational Linguistics. +Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Roktaschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544, Online. Association for Computational Linguistics. + +Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics. +Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, and Samuel R. Bowman. 2020. Intermediate-task transfer learning with pretrained language models: When and why does it work? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5231-5247, Online. Association for Computational Linguistics. +Guanghui Qin and Jason Eisner. 2021. Learning how to ask: Querying LMs with mixtures of soft prompts. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5203-5212, Online. Association for Computational Linguistics. +Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446. +Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932-4942, Florence, Italy. Association for Computational Linguistics. +Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In *Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems*, pages 1-7. +Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95. +Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. 2020. Getting closer to ai complete question answering: A set of prerequisite real tasks. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8722-8731. +Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633. + +Amrita Saha, Rahul Aralikatte, Mitesh M. Khapra, and Karthik Sankaranarayanan. 2018. DuoRC: Towards complex language understanding with paraphrased reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683-1693, Melbourne, Australia. Association for Computational Linguistics. +Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106. +Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207. +Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social IQa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4463-4473, Hong Kong, China. Association for Computational Linguistics. +Thomas Scialom, Tuhin Chakrabarty, and Smaranda Muresan. 2022. Continual-t0: Progressively instructing $50+$ tasks to language models without forgetting. arXiv preprint arXiv:2205.12393. +Weijia Shi, Julian Michael, Suchin Gururangan, and Luke Zettlemoyer. 2022. Nearest neighbor zero-shot inference. arXiv preprint arXiv:2205.13792. +Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615. +Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Huadong Wang, Kaiyue Wen, Zhiyuan Liu, Peng Li, Juanzi Li, Lei Hou, Maosong Sun, and Jie Zhou. 2022. On transferability of prompt tuning for natural language processing. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3949-3969, Seattle, United States. Association for Computational Linguistics. +Niket Tandon, Bhavana Dalvi, Keisuke Sakaguchi, Peter Clark, and Antoine Bosselut. 2019. WIQA: A dataset for "what if..." reasoning over procedural text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the + +9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6076-6085, Hong Kong, China. Association for Computational Linguistics. +Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou', and Daniel Cer. 2022. SPoT: Better frozen model adaptation through soft prompt transfer. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5039-5059, Dublin, Ireland. Association for Computational Linguistics. +Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics. +Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. 2022a. What language model architecture and pretraining objective work best for zero-shot generalization? arXiv preprint arXiv:2204.05832. +Xinyi Wang, Wanrong Zhu, and William Yang Wang. 2023. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. arXiv preprint arXiv:2301.11916. +Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. 2022b. Benchmarking generalization via in-context instructions on 1,600+ language tasks. arXiv preprint arXiv:2204.07705. +Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022c. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149. +Albert Webson and Ellie Pavlick. 2021. Do prompt-based models really understand the meaning of their prompts? arXiv preprint arXiv:2109.01247. +Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652. +Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903. + +Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 94-106, Copenhagen, Denmark. Association for Computational Linguistics. +Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing datasets for multi-hop reading comprehension across documents. Transactions of the Association for Computational Linguistics, 6:287-302. +Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080. +Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. 2022. Zero-prompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot generalization. arXiv preprint arXiv:2201.06910. +Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013-2018, Lisbon, Portugal. Association for Computational Linguistics. +Seonghyeon Ye, Doyoung Kim, Joel Jang, Joongbo Shin, and Minjoon Seo. 2022. Guess the instruction! making language models stronger zero-shot learners. arXiv preprint arXiv:2210.02969. +Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics. +Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015a. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc. +Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015b. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 649-657. +Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics. + +Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR. +Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2856-2878, Punta Cana, Dominican Republic. Association for Computational Linguistics. +Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers. + +# A Comparison with Hard Prompt Optimization + +We have conducted additional experiments to compare our method with 3 different hard prompt optimization techniques including (1) RoHPr which utilizes the same search technique as RoSPr but retrieves 4 corresponding training instances for incontext learning instead of the corresponding soft prompt, and (2) APE (Zhou et al., 2023) which is an automatic prompt engineering method that utilizes the generation results of an LLM (Ouyang et al., 2022), known to show better performance than human instructions, and (3) ZPS (Liao et al., 2022) which selects an optimal hard prompt from a prompt pool using prompt ensemble and pseudolabels. As shown in the result of Table 2, on average, RoSPr performs the best on average, showing the benefits of using soft prompts over hard prompts for task generalization. While RoSPr shows improvement on 10 out of 11 datasets, other hard prompt optimization methods do not show consistent improvements. + +# B Full Result of BIG-bench Evaluation + +We provide the task generalization performance result of 14 tasks from BIG-bench, shown in Table 3. Applying RoSPR largely improves the performance for 3 datasets (Hindu Knowledge, Novel Concepts, Misconceptions): $(+17.72\%, +3.12\%, +1.82\%)$ compared to T0-3B and $(+13.72\%, +3.12\%, +2.05\%)$ compared to T0-11B. For mean accuracy of 14 tasks, T0+RoSPR outperforms T0-3B by $2.39\%$ points, reducing the gap between 4 times larger T0-11B to $1.84\%$ points. Moreover, applying INTER to T0+RoSPR enhances the performance of T0+RoSPR for most tasks, indicating that interpolation of multiple embeddings is effective for challenging tasks. + +# C Fine-tuning Baseline Details + +For fine-tuning baseline models (PT and ATTEMPT), we follow the training configuration of source prompt tuning. We train each target prompt with a single epoch using a maximum of 5,000 training instances. For fine-tuning, we use a batch size of 32 and a learning rate of 1e-3. Also, we randomly select hard prompts (templates) of the training dataset during fine-tuning. For ATTEMPT, we randomly sample one source prompt per task cluster, resulting in interpolation between 8 soft prompts. + +# D Additional Ablation Studies + +![](images/7c6ab7b6dea7392b5cf3dd29e262296a2f1dd12b367774760af9b83d72bb884f.jpg) +Figure 8: Variation of number of prompts by increasing the number of prompts per dataset. + +![](images/3a71c461961e3e29507a80cc7186c63a26d336cb5ea8305c590954fc1f19f36b.jpg) +Figure 9: Variation of number of datasets by increasing the number of datasets per task cluster. + +We provide detailed result of variation of the number of prompts (Figure 8) and the number of datasets (Figure 9). We additionally analyze the effect of (1) different sampling methods for constructing the Source Prompt Library, (2) the number of instances sampled for constructing the Source Prompt Library, (3) the number of top-N retrieval for embedding retrieval, and (4) the number of multiple source embeddings to interpolate. Same as Section 7, we report the mean accuracy of 4 evaluation datasets: RTE, COPA, Winogrande, and WiC with 3 runs with different random seeds for the sampling of evaluation queries. + +# D.1 Sampling Methods for Source Prompt Library + +We experiment three different methods to sample instances for constructing Source Prompt Library and analyze the effect of each method. By default, we choose RANDOM method, where we sample + +
RTECBAN.R1AN.R2AN.R3COPAHellasw.StoryC.Winogr.WSCWiCAVG
T064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.22
RoHPr47.1045.0033.5333.3033.0973.3826.6586.9350.5164.4250.0049.45
APE (Zhou et al., 2023)68.1948.3335.8235.3734.2768.0025.7678.7350.4358.7550.8550.41
ZPS (Liao et al., 2022)58.4860.7136.6034.4033.3376.0028.4987.3951.7864.4250.6352.93
RoSPr71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.24
+ +Table 2: Comparison result of RoSPr with different hard prompt optimization techniques. + +
T0-3BROSPRINTERT0-11B
Strategy.52.7952.0552.2352.75
Movie D.52.8551.4552.2353.69
Known Un.47.8347.8347.8358.70
Logic Grid41.1037.0037.4038.30
Hindu Kn.25.7143.4345.7129.71
Code D.46.6745.0040.0043.33
Concept45.5267.6167.6169.29
Language14.8413.6814.4020.20
Vitamin58.8853.7154.5364.73
Syllogism52.9450.6451.3451.81
Misconcept.50.2352.0552.0550.00
Logical46.6454.8654.8654.86
Winowhy44.2944.3344.2952.11
Novel Con.15.6315.6318.7515.63
AVG42.5644.9545.2346.79
+ +Table 3: Evaluation result of 14 tasks from BIG-bench (Srivastava et al., 2022). + +![](images/a4fe1ea67a1f90c1d4b419df853c5a552f683205a91e522c4d57494a989f37af.jpg) +Figure 10: Different instance sampling methods for constructing Source Prompt Library. + +100 instances by random for each prompt by assuming that each random 100 queries of instances can represent the whole prompt. Second, we experiment CLUSTERING method, where we sample top 100 instances which has the closest mean representation from its overall average for each prompt. If we say each instance has a distance of mean representation from its overall average as $d_{i}$ in increasing order, we sample $\{d_1,d_2,\dots,d_{100}\}$ . The last method we use is DISTRIBUTed method, where we sample 100 instances in a distributed way with respect to its distance of mean representation from its overall average. If we say each instance has a distance of mean representation from its over + +all average as $d_{i}$ in increasing order, we sample $\{d_1,d_{1 + N / 100},d_{1 + 2*N / 100},\dots ,d_{1 + 99*N / 100}\}$ , assuming there are total $N$ training instances in a dataset. + +As shown in Figure 10, RANDOM method outperforms CLUSTERING and DISTRIBUTED methods. Interestingly, CLUSTERING method significantly hurts the performance on all 4 proposed methods, suggesting that storing similar instances per prompt results in retrieval failures more often. Also for DISTRIBUTED method, most of the methods significantly underperform RANDOM, except INTER and ROSPR+VAR+INTER. From these results, we can conclude that random sampling represents the source dataset most effectively. + +![](images/771e75538b33e4966cec3169ff4caa82aaf3409a9384edf93e1349e8c0eff70d.jpg) +Figure 11: Variation of number of instances sampled for constructing Source Prompt Library. Default setting is $n = 100$ . + +# D.2 Number of Instances Sampled for Constructing Source Prompt Library + +We analyze the effect of size of the Source Prompt Library by varying the number of instances $n$ to sample for each hard prompt by 100, 300, 500. Therefore, $n \times$ (number of total hard prompts) would be the size of the Source Prompt Library. As shown in Figure 11, increasing the number of sampled instances does not increase the performance; it hurts the performance for most cases. This suggests that only a few number of training instances are enough to represent the distribution of prompted + +![](images/8925a804240b359526e812738a5b7a830639cc6bf88c8f1706c4e4f2b9aa6341.jpg) +Figure 12: Variation of number of Top $N$ instances for embedding retrieval. Default setting is $N = 10$ . + +input (hard prompt + input instances) for each hard prompt and increasing the number sometimes hurt the performance by adding noise to the distribution. This also supports the importance of heuristic cues in Section 6 by showing that adding more training instances per hard prompt does not increase the performance. Instead, adding hard prompts with diverse answer choice format is more important. + +# D.3 Number of Top N instances for Embedding Retrieval + +We vary the number of top- $N$ instances that are retrieved given each query through MIPS search. As shown in Figure 12, varying the number of top- $N$ instances does not have much effect compared to increasing the number of sampled queries (Figure 7b). This implies that if the size of the evaluation set of the target task is large, sampling more queries is effective than searching for more similar instances per query. This is important for variance-based methods because the number of forward passes needed before evaluation is proportional to $Q*N$ . Therefore, we can reduce the time latency by reducing the number of instances retrieved per query without hurting the performance much. + +# D.4 Number of Source Embeddings for Interpolation + +We analyze the effect of number of source embeddings for interpolation by varying top- $N^{\prime}$ interpolation from 1 (no interpolation) to 5 shown in Figure 13. By comparing between single prompt embedding retrieval $(N^{\prime} = 1)$ and the interpolation of multiple embeddings $(N^{\prime} > 1)$ , the mean accuracy drops by adding multiple source embeddings for retrieval because interpolation-based methods underperform on tasks such as COPA as shown + +![](images/62f18d4144a73705322ce9eb6db64e09537df79b85f9c48ef1656b5ea55cc28a.jpg) +Figure 13: Variation of number of source embeddings for interpolation based methods. Default setting is $K' = 3$ . + +in Table 1. Mean accuracy would increase if we add other datasets for evaluation that benefit from interpolation such as WSC and CB. + +By comparing among various $N'$ values, we find that for ROSPR+INTER, the accuracy substantially decreases for $K' = 2$ , implying that the possibility of a wrong retrieval varies depending on the value of $N'$ . In contrast, ROSPR+VAR+INTER is more robust to the value of $K'$ , showing that variance-based ranking increases robustness to different numbers of source embeddings for interpolation as well. + +# E Visualization of Results + +We show the visualization of the evaluation result on 11 datasets in Figure 14. Methods based on ROSPR not only show higher accuracy, but lower variance for many datasets. + +# F Experimental Configurations + +As mentioned in the previous sections, we use T0-3B as our backbone meta-trained LM. For prompt tuning, we fix the prefix length as 100 and the embeddings are initialized from 5,000 most common vocabulary following Lester et al. (2021). We train each source embedding for a single epoch with a learning rate of 0.1 and a batch size of 32. We use the Adam optimizer with weight decay of 1e-5. For retrieval, we randomly sample $Q = 32$ query instances and retrieve top $K = 10$ examples for each query. We train a T0-small variant ( $\sim 35\mathrm{M}$ params) as our dense retriever by multitask prompted training on T5+LM model (Lester et al., 2021), replicating the original training setting of T0 by training T5+LM for 8 epochs using the same training instances of Sanh et al. (2021) with + +![](images/f6b681d8de7ac8cdac12bf6b353e78bbe2965768c33bc4d4baf40abff101026f.jpg) +(a) RTE + +![](images/8edc524f24162a48e4833bbdad69b57897357326563c1f4eaea143f56bf8032a.jpg) +(b) CB + +![](images/9490c4e09c316164a435fde9b295f7e6dfc9926d451292a2424456bfba8aac34.jpg) +(c) ANLI + +![](images/7d237e9658defe93969abaf5410e969848d5bb88f1f9c16f61f17a5c2304e433.jpg) +(d) ANLI R2 + +![](images/94345d78f9ba02784fc5758aa5298af1f471f707f89243831aa1f69ca9103cf2.jpg) +(e) ANLI R3 + +![](images/cd031f39a39d528fdc225d388dc7060ef6c76d6425f60064f8c685701d44b79a.jpg) +(f) COPA + +![](images/481e804c5f6c417b8768efd7ece1919d29d27c091101a3ea125c3802c04137de.jpg) +(g) Hellaswag + +![](images/bbe34813cb710dde0b1ac3ca17b659e72ce568c5a4a751273e353e0fdbdb0e0f.jpg) + +![](images/7173cd83ab347f6daf8a9358825658e1162d4e03c2d2b1a72b54cbe060be00b5.jpg) + +![](images/7ba60fd0fd27431dc310a232c0695e5978d4f3914e454d39f98b1802390d3ea8.jpg) + +![](images/7de310496739df397c37b7563fdd412cfe08fce4901f5ce1608dcbcdceebcf4a.jpg) +(h) StoryCloze + +![](images/2f1b160c84f15db68dbe6a93ecf996f65a7657a7b825daf242a039ccb2777701.jpg) +(i) Winogrande +(j) WSC +(k)WiC +Figure 14: Visualization of evaluation results of 11 datasets. + +a learning rate of 1e-3, input sequence length 512, output sequence 128, and batch size of 1024. We select model checkpoint by early stopping based on validation accuracy. We use a meta-trained LM instead of a naive pretrained model (e.g. SentenceBERT) because meta-trained LM is shown to be more effective for retrieval (Lin et al., 2022). For the interpolation experiment, we set $K' = 3$ for top- $K'$ prompt embedding candidates. For training source prompt embeddings, we used 8 V100 GPUs. + +# G Examples of Applying Prompts, Answer Choice Format and Source Task Types + +Figure 15 shows an example of applying prompt through Promptsource (Bach et al., 2022) as mentioned in Section 3.1. + +We assert that answer choice format is more important than task similarity in Section 6. We further provide details of the input instances of the mentioned tasks: paraphrase, NLI, word sense disambiguation, and sentiment classification in Figure 16. As supported in Pruksachatkun et al. (2020), + +intuitively, paraphrase task is more similar to the task of word sense disambiguation task or NLI, implying their task similarity, while the task of sentiment classification is very different. However, our counterintuitive result shows the soft prompt to show the best performance in Figure 6 Section 6, bolstering the claim that similar source task types are not a major factor for evaluation performance. + +# H Full List of Source Training and Evaluation Datasets + +All of our training and evaluation datasets are a subset of datasets used in Sanh et al. (2021). We use Huggingface version of each dataset (Lhoest et al., 2021). + +# H.1 Training Datasets + +Following Sanh et al. (2021), we use 8 task clusters for training of source prompt embedding: sentiment classification, paraphrase, topic classification, summarization, struc-to-text, multiple-choice QA, extractive QA, and closed book QA. We use imdb (Maas et al., 2011), amazon_polarity + +![](images/c50949b1eb4152b355b40437041c77e0619ed61eb9b9188201d943658e560a22.jpg) +Figure 15: Example of applying prompt to a given instance through Promptsource (Bach et al., 2022). + +![](images/7e0c78c2453749ae0067e28a53ad919ec36b7398bd2909971dd6c3f175995f20.jpg) +Figure 16: Examples of instances of different source tasks. + +(McAuley and Leskovec, 2013), rotten_tomatoes (Pang and Lee, 2005), yelp_review_full (Zhang et al., 2015b) for sentiment, glue/qqp (Wang et al., 2018), paws/labeled_final (Zhang et al., 2019) for paraphrase, ag_news (Zhang et al., 2015a), dbpedia_14 (Lehmann et al., 2015) for topic classification, gigaword (Graff et al., 2003), multi_news citefabbri-etal-2019-multi, samsum (Gliwa et al., 2019), xsum (Narayan et al., 2018) for summarization, common_gen (Lin et al., 2020), wiki.bio (Lebret et al., 2016) for struct-to-text, cos_e/v1.11 (Rajani et al., 2019), quail (Rogers et al., 2020), social_i_qa (Sap et al., 2019), wiqa (Tandon et al., 2019), cosmos_qa (Huang et al., 2019), sciq (Welbl et al., 2017), wiki_hop/original (Welbl et al., 2018) for multi-choice QA, adversarial_qa/dbidaf, adversarial_qa/dbert, adversarial_qa/droberta, quoref (Bartolo et al., 2020), ropes (Lin et al., 2019), duorc/SelfRC, duorc/Paraphrase IdentificationRC (Saha et al., 2018) for extractive QA, and kilt_tasks/hotpotqa (Petroni et al., 2021), wiki_qa (Yang et al., 2015) for closed book QA. + +We exclude 6 datasets (MRPC, TREC, DREAM, QuaRTz, QASC, QuaRel) that have small training sets because it leads to task imbalance, which is critical for training our small dense retriever ( $\sim$ 35M params). We also exclude CNN Daily Mail, App Reviews, and WikiQA dataset due to dataset download issues, absence of any test or validation data, and unbalanced label distribution, respectively. + +# H.2 Evaluation Datasets + +Following Sanh et al. (2021), we include 11 evaluation datasets as follows: RTE (Dagan et al., 2005), CB (De Marneffe et al., 2019), ANLI (Nie et al., 2020) for natural language inference task, COPA (Roemmele et al., 2011), Hellaswag (Zellers et al., 2019), Storycloze (Mostafazadeh et al., 2016) for sentence completion task, Winogrande (Sakaguchi et al., 2021), WSC (Levesque et al., 2012) for coreference resolution task, and WiC (Pilehvar and Camacho-Collados, 2019) for word sense disambiguation task. + +For BIG-bench tasks, we evaluate on 14 tasks, following Sanh et al. (2021): Known Unknown, Logic Grid, StrategyQA, Hindu Knowledge, Movie Dialog, Code Description, Conceptual, Language ID, Vitamin C, Syllogisms, Misconceptions, Logical Deduction, Winowhy and Novel Concepts. + +# I Full List of Retrieved Prompt Embeddings + +We provide a full list of retrieved prompt embeddings of ROSPR and ORACLE for all prompts of 11 evaluation datasets. We report retrieval results of a single random seed (Table $4\sim$ Table 14). + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
GPT-3 style61.3774.01paws/labeled_final/PAWS-ANLI GPT374.01paws/labeled_final/context-question
MNLI crowdsourced63.5370.04paws/labeled_final/context-question72.20paws/labeled_final/context-question
based on the previous passage68.2376.53paws/labeled_final/context-question76.53paws/labeled_final/PAWS-ANLI GPT3
can we infer59.5773.29glue/qqp/meaning73.29paws/labeled_final/PAWS-ANLI GPT3
does it follow that61.7371.84paws/labeled_final/context-question71.84paws/labeled_final/context-question-no-label
does this imply64.6268.59paws/labeled_final/context-question71.48paws/labeled_final/context-question
guaranteed true68.9575.09paws/labeled_final-context-question75.81paws/labeled_final/PAWS-ANLI GPT3
should assume66.4371.12glue/qqp/meaning76.53paws/labeled_final/PAWS-ANLI GPT3
justified in saying61.0158.12paws/labeled_final/paraphrase-task71.12paws/labeled_final/context-question
must be true70.0474.37paws/labeled_final/context-question75.09paws/labeled_final/PAWS-ANLI GPT3-no-label
Avg.64.5571.3073.79
+ +Table 4: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of RTE. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer55.3651.79social_i_qa/Show choices and generate index67.86samsum/Write a dialogue that match this summary
based on the previous passage44.6458.93social_i_qa/Show choices and generate index69.64samsum/Write a dialogue that match this summary
claim true/false/inconclusive50.0067.86social_i_qa/Show choices and generate index69.64samsum/Summarize:
does it follow that64.2950.00social_i_qa/Show choices and generate index67.86samsum/Summarize:
justified in saying53.5750.00social_i_qa/Show choices and generate index62.50samsum/Write a dialogue that match this summary
always/sometimes/never39.2941.07social_i_qa/Show choices and generate index41.07social_i_qa/Show choices and generate index
GPT-3 style51.7967.86social_i_qa/Show choices and generate index69.64social_i_qa/Show choices and generate answer
consider always/sometimes/never35.7135.71social_i_qa/Show choices and generate index39.29social_i_qa/Generate answer
guaranteed true48.2150.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
must be true53.5750.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
guaranteed/possible/impossible8.938.93social_i_qa/Show choices and generate index8.93-(all same)
does this imply58.9351.79social_i_qa/Show choices and generate index66.07glue/qqp/duplicate
MNLI crowdsourced8.9339.29social_i_qa/Show choices and generate index42.86cos_e/v1.11/question_option_description_text
should assume57.1450.00social_i_qa/Show choices and generate index66.07cos_e/v1.11/aligned_with_common Sense
take the following as truth50.0067.86social_i_qa/Show choices and generate index71.43cos_e/v1.11/description_question_option_id
Avg.45.3649.4058.10
+ +Table 5: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of CB. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.9038.90paws/labeled_final/context-question39.40paws/labeled_final/PAWS-ANLI GPT3
based on the previous passage33.9038.50paws/labeled_final/context-question38.60paws/labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive35.6036.70paws/labeled_final/PAWS-ANLI GPT339.10paws/labeled_final/context-question
does it follow that36.0040.50paws/labeled_final/context-question40.50paws/labeled_final/PAWS-ANLI GPT3
justified in saying33.1038.10paws/labeled_final-context-question38.80paws/labeled_final/context-question-no-label
always/sometimes/never33.4033.40paws/labeled_final/paragraph-task33.40-(all 33.4)
GPT-3 style33.8037.30paws/labeled_final/PAWS-ANLI GPT338.50paws/labeled_final/PAWS-ANLI GPT3-no-label
consider always/sometimes/never33.2033.40paws/labeled_final/PAWS-ANLI GPT333.50-(all 33.4)
guaranteed true33.7038.50paws/labeled_final/context-question38.70paws/labeled_final/PAWS-ANLI GPT3
must be true34.4039.60paws/labeled_final CONTEXT-question39.70paws/labeled_final-context-question
guaranteed/possible/impossible33.3033.30paws/labeled_final/PAWS-ANLI GPT333.30imdb/Text Expressed Sentiment
does this imply33.6038.20paws/labeled_final contextual-question38.20paws/labeled_final contextual-question-no-label
MNLI crowdsourced33.6033.80paws/labeled_final contextual-question35.40dbpedia_14/given_list WHAT_category DOES_theParagraph_belong_to
should assume33.2038.80paws/labeled_final contextual-question39.00paws/labeled_final/PAWS-ANLI GPT3-no-label
take the following as truth32.4037.10paws/labeled_final/PAWS-ANLI GPT338.70paws/labeled_final contextual-question-no-label
Avg.33.8137.0737.65
+ +Table 6: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R1. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer30.4035.30paws/ labeled_final/context-question35.30paws/ labeled_final/PAWS-ANLI GPT3
based on the previous passage31.4035.40paws/ labeled_final/context-question35.40paws/ labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive34.9035.10paws/ labeled_final/PAWS-ANLI GPT335.90paws/ labeled_final/context-question
does it follow that34.5036.00paws/ labeled_final/context-question36.00paws/ labeled_final/PAWS-ANLI GPT3
justified in saying33.5035.70paws/ labeled_final/context-question35.70rotten_tomatoes/Reviewer Enjoyment Yes No
always/sometimes/never33.4033.40paws/ labeled_final/paragraph-task33.50adversarial_qa/droberta/based_on,
GPT-3 style33.5034.90paws/ labeled_final/PAWS-ANLI GPT335.00paws/ labeled_final/context-question-no-label
consider always/sometimes/never33.7033.40dbpedia_14/given_a_choice_of categories34.50paws/ labeled_final/PAWS-ANLI GPT3-no-label
guaranteed true32.9034.00paws/ labeled_final/context-question34.30paws/ labeled_final/PAWS-ANLI GPT3
must be true35.1034.60paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
guaranteed/possible/impossible33.3033.30paws/ labeled_final/context-question-no-label33.30imdb/Writer Expressed Sentiment
does this imply32.7033.90paws/ labeled_final/context-question34.10paws/ labeled_final/context-question
MNLI crowdsourced33.4034.70dbpedia_14/given_a_choice_ofcategories34.90dbpedia_14/given_a_choice_ofcategories
should assume32.4035.10paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
take the following as truth31.6035.70paws/ labeled_final/paragraph-task35.70paws/ labeled_final/PAWS-ANLI GPT3
Avg.33.1134.7034.92
+ +Table 7: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R2. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.0034.75glue/qqp/meaning34.75paws/labeled_final/context-question-no-label
based on the previous passage33.3334.08paws/labeled_final/context-question35.33paws/labeled_final/context-question-no-label
claim true/false/inconclusive32.8335.58paws/labeled_final/paraphrase-task35.92cos_e/v1.11/aligned_with_common Sense
does it follow that34.0834.67paws/labeled_final/context-question35.33amazon_polarity/User_recommend.this_product
justified in saying33.5833.00paws/labeled_final/paraphrase-task35.42amazon_polarity/User Recommend.this_product
always/sometimes/never33.4233.42paws/labeled_final/paraphrase-task33.50paws/labeled_final/paraphrase-task
GPT-3 style33.3334.00paws/labeled_final/PAWS-ANLI GPT334.92paws/labeled_final/PAWS-ANLI GPT3
consider always/sometimes/never33.0832.42ropes/plain_nobackground33.67cos_e/v1.11/question_description_option_text
guaranteed true32.5834.08paws/labeled_final/context-question34.83paws/labeled_final/context-question
must be true33.8333.75paws/labeled_final/paraphrase-task35.42rotten_tomatoes/Reviewer Enjoy Yes No
guaranteed/possible/impossible33.5033.50paws/labeled_final/paraphrase-task33.58social_i_qa/Show choices and generate answer
does this imply32.9233.58glue/qqp/meaning34.50paws/labeled_final/context-question
MNLI crowdsourced33.7533.67rotten_tomatoes/Text Expressed Sentiment34.00dbpedia_14/given_a_choice_ofcategories
should assume33.2534.67paws/labeled_final/context-question34.92paws/labeled_final/context-question-no-label
take the following as truth33.4232.75social_i_qa/Show choices and generate index36.67paws/labeled_final/context-question-no-label
Avg.33.3333.8634.91
+ +Table 8: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R3. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
exercise80.0079.00cos_e/v1.11/question_option_description_text80.00cos_e/v1.11/question_option_description_text
plausible_alternatives84.0083.00cos_e/v1.11/question_option_description_text83.00cos_e/v1.11/question_option_description_text
"C1 or C2? premise, so/because..."61.0067.00social_i_qa/Check if a random answer is valid or not75.00cos_e/v1.11/descriptionQuestion_option_text
best_option70.0076.00social_i_qa/Show choices and generate answer79.00cos_e/v1.11/question_option_description_text
more likely79.0083.00cos_e/v1.11/question_option_description_text85.00cos_e/v1.11/descriptionQuestion_option_text
cause-effect74.0078.00cos_e/v1.11/question_option_description_text83.00common_gen/random task template prompt
choose80.0082.00cos_e/v1.11/question_option_description_text82.00cosmos_qa/no_prompt_text
i_am_hesitating79.0082.00cos_e/v1.11/question_option_description_text82.00cos_e/v1.11/question_option_description_text
Avg.75.8878.7581.13
+ +Table 9: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of COPA. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Predict ending with hint26.8327.70social_i_qa/Show choices and generate index29.11cos_e/v1.11/question_option_description_text
Randomized prompts template26.8726.83social_i_qa/Show choices and generate index27.84wiqa/what_is_the MISSING_first_step
complete_first THEN28.3727.35ropes/prompt_bottom_no_hint28.35cos_e/v1.11/question_option_description_text
if_beginnerscontinue25.2826.58social_i_qa/Show choices and generate index26.58cos_e/v1.11/question_option_description_text
how Ends25.6726.86social_i_qa/Show choices and generate index26.86cos_e/v1.11/question_option_description_text
Avg.26.6027.0627.75
+ +Table 10: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Hellaswag. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Answer Given options86.4887.76social_i_qa/Show choices and generate answer87.76social_i_qa/Show choices and generate answer
Choose Story Ending86.9188.24social_i_qa/Show choices and generate answer88.24cos_e/v1.11/question_option_description_text
Movie What Happens Next78.0978.03social_i_qa/Show choices and generate answer87.33cosmos_qa/no_prompt_text
Story Continuation and Options83.5985.89social_i_qa/Show choices and generate answer86.53social_i_qa/Show choices and generate answer
Novel Correct Ending85.0987.65social_i_qa/Show choices and generate answer87.97social_i_qa/Show choices and generate index
Avg.84.0385.5287.57
+ +Table 11: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of StoryCloze. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does underscore refer to stand for49.2551.54cos_e/v1.11/question_option_description_text51.78paws/ labeled_final/ PAWS-ANLI GPT3
50.5150.59cos_e/v1.11/question_option_description_text52.09ropes/plain_nobackground
underscore refer to fill in the blank50.4350.59cos_e/v1.11/question_option_description_text52.33ropes/prompt_bottom_no_hint
52.1751.38cos_e/v1.11/question_option_description_text52.01cos_e/v1.11/question_option_description_text
Replace52.4953.59cos_e/v1.11/question_option_description_text53.59paws/ labeled_final/ paraphrase-task
Avg.50.9751.5452.36
+ +Table 12: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Winogrande. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does the pronoun refer to by p they mean68.2766.34social_i_qa/Check if a random answer is valid or not66.35paws/labeled_final/PAWS-ANLI GPT3
62.5066.35social_i_qa/Check if a random answer is valid or not69.23glue/qqp/duplicate or not
in other words67.3167.31social_i_qa/Show choices and generate answer72.12samsum/Write a dialogue that match this summary
I think they mean69.2368.27social_i_qa/Show choices and generate answer78.08social_i_qa/Generate answer
replaced with64.4259.62social_i_qa/Check if a random answer is valid or not66.35social_i_qa/Show choices and generate index
p is/are r62.5064.42social_i_qa/Show choices and generate index65.38social_i_qa/Show choices and generate index
the pronoun refers to64.4264.42social_i_qa/Show choices and generate index67.31social_i_qa/Check if a random answer is valid or not
Who or what is/are64.4263.46social_i_qa/Show choices and generate index65.38samsum/To sum up this dialog
does p stand for67.3163.46social_i_qa/I was wondering69.23samsum/Write a dialogue that match this summary
GPT-3 Style60.5867.31social_i_qa/Show choices and generate index67.31rotten_tomatoes/Reviewer Expressed Sentiment
Avg.65.1065.1068.17
+ +Table 13: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WSC. + +
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
question-context-meaning-with-label50.3153.76glue/qqp/meaning53.76paws/labeled_final/context-question-no-label
question-context-meaning50.6350.16ropes/prompt_bottom_no_hint57.05paws/labeled_final/PAWS-ANLI GPT3
grammar Homework49.8450.16ropes/prompt_bottom_no_hint57.99rotten_tomatoes/Reviewer Enjoyment Yes No
affirmation_true_orfalse49.6951.57paws/labeled_final/Rewrite-no-label53.92paws/labeled_final/Meaning-no-label
GPT-3-prompt51.7250.00social_i_qa/Show choices and generate index55.64paws/labeled_final/PAWS-ANLI GPT3
same_sense49.8452.82paws/labeled_final/Rewrite55.17paws/labeled_final/task_description-no-label
question-context51.8853.29social_i_qa/Check if a random answer is valid or not57.05amazon_polarity/Is.this_product_review_positive
GPT-3-prompt-with-label50.4752.19glue/qqp/meaning52.66glue/qqp/meaning
polysemous50.0052.82paws/labeled_final/Rewrite53.29paws/labeled_final/Rewrite-no-label
similar-sense52.5150.00social_i_qa/Show choices and generate index56.11paws/labeled_final/task_description-no-label
Avg.50.6951.6855.26
+ +Table 14: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WiC. \ No newline at end of file diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/images.zip b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2fed12876b6910aa87e9e356e423ce0720751951 --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b00f055dcf0174506f3557b3b81b8bf55bfbb1bbb97f5cd1c69f0a1ef965fa5 +size 1540549 diff --git a/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/layout.json b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52a0c294c8bc3fac80adf73b02fbb836acbd774f --- /dev/null +++ b/2023/Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt/layout.json @@ -0,0 +1,13471 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 75, + 516, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 75, + 516, + 110 + ], + "spans": [ + { + "bbox": [ + 80, + 75, + 516, + 110 + ], + "type": "text", + "content": "Efficiently Enhancing Zero-Shot Performance of Instruction Following Model via Retrieval of Soft Prompt" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 130, + 489, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 130, + 489, + 158 + ], + "spans": [ + { + "bbox": [ + 108, + 130, + 489, + 158 + ], + "type": "text", + "content": "Seonghyeon Ye Joel Jang Doyoung Kim Yongrae Jo Minjoon Seo KAIST" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 159, + 477, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 159, + 477, + 173 + ], + "spans": [ + { + "bbox": [ + 120, + 159, + 477, + 173 + ], + "type": "text", + "content": "{seonghyeon.ye,joeljang,ikevin98,yongrae,minjoon}@kaist.ac.kr" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 204, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "spans": [ + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "text", + "content": "Enhancing the zero-shot performance of instruction-following models requires heavy computation, either by scaling the total number of training datasets or the model size. In this work, we explore how retrieval of soft prompts obtained through prompt tuning can efficiently assist hard prompts in zero-shot task generalization. Specifically, we train soft prompt embeddings for each prompt through prompt tuning, store the samples of the training instances mapped with the prompt embeddings, and retrieve the corresponding prompt embedding of the training instance closest to the query instance during inference. While only adding " + }, + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "inline_equation", + "content": "0.007\\%" + }, + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "text", + "content": " additional parameters, retrieval of soft prompt enhances the performance of T0 on unseen tasks by outperforming it on 10 out of 11 datasets as well as improving the mean accuracy of T0 on BIG-bench benchmark by " + }, + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 84, + 239, + 274, + 514 + ], + "type": "text", + "content": " points. Also, we report an interesting finding that retrieving source embeddings trained on similar answer choice formats is more important than those on similar task types." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 528, + 155, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 528, + 155, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 528, + 155, + 540 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 551, + 291, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 551, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 291, + 741 + ], + "type": "text", + "content": "Training Large Language Models (LLMs) on huge amounts of data has enabled LMs to perform downstream tasks without any fine-tuning with the aid of natural prompts or concatenation of a few demonstration instances (Brown et al., 2020; Rae et al., 2021; Kojima et al., 2022; Chowdhery et al., 2022). Additionally, recent works have shown that adding a instruction tuning stage, an additional training step that helps pretrained LMs understand prompts and demonstrations results in a significant performance boost on zero-shot task generalization even for moderate-sized LMs (Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Ye et al., 2022; Chung et al., 2022). This extra" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 211, + 526, + 296 + ], + "blocks": [ + { + "bbox": [ + 304, + 211, + 526, + 296 + ], + "lines": [ + { + "bbox": [ + 304, + 211, + 526, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 526, + 296 + ], + "type": "image", + "image_path": "5f317a13ed48bf7b1b98fcbabeec569fbb45be92a01ee06ccd631a599e8f6e40.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 304, + 525, + 346 + ], + "lines": [ + { + "bbox": [ + 302, + 304, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 302, + 304, + 525, + 346 + ], + "type": "text", + "content": "Figure 1: During zero-shot inference, ROSPR selects similar training instances with the given input from the Source Prompt Library and retrieves the prompt embeddings corresponding to the selected training instances." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 353, + 525, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 353, + 525, + 393 + ], + "spans": [ + { + "bbox": [ + 302, + 353, + 525, + 393 + ], + "type": "text", + "content": "instruction-tuning stage involves explicit, multitask prompted learning on various tasks, enabling LMs to quickly adapt to unseen tasks at inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 394, + 526, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 394, + 526, + 555 + ], + "spans": [ + { + "bbox": [ + 302, + 394, + 526, + 555 + ], + "type": "text", + "content": "To maximize the effect of instruction-tuning, two approaches have been widely explored: (1) scaling the number of training datasets, and (2) scaling the model size (Wang et al., 2022b; Chung et al., 2022). However, both approaches require heavy computation, not applicable with an academic budget. Specifically, the first approach requires updating the whole parameters of the model every time a training dataset is added, showing limitations in terms of scalability. On the other hand, the second approach requires heavy memory requirements to load and train a massive LLM." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 556, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 525, + 772 + ], + "type": "text", + "content": "To enhance the zero-shot performance of instruction-following model efficiently, we introduce Retrieval of Soft Prompt (ROSPR), which is easily scalable and requires minimal computation by only adding " + }, + { + "bbox": [ + 302, + 556, + 525, + 772 + ], + "type": "inline_equation", + "content": "0.007\\%" + }, + { + "bbox": [ + 302, + 556, + 525, + 772 + ], + "type": "text", + "content": " parameters to the main model during inference. As shown in Figure 1, by training prompt embeddings (soft prompt) for each given hard prompt through prompt tuning, we construct a Source Prompt Library consisting of samples of training instances mapped with their corresponding prompt embeddings. Then, during inference, by using a simple, off-the-shelf dense retriever model, we search for training instances similar to the given query instances and retrieve their corresponding prompt embeddings. Because the backbone LM is frozen, it allows the retrieved" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "1Model checkpoints and code implementation are available at github.com/seonghyeonye/RoSPr." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "12288" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "spans": [ + { + "bbox": [ + 124, + 795, + 468, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12288-12309 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 138 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 138 + ], + "type": "text", + "content": "embeddings to serve as adapters assisting hard prompts. While RoSPR can be applied to any LM, in this work, we use T0 (Sanh et al., 2021) as our initial backbone LM and perform prompt tuning on the tasks used during the instruction-tuning stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "text", + "content": "While adding only " + }, + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "inline_equation", + "content": "0.007\\%" + }, + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "text", + "content": " additional parameters, ROSPR outperforms T0 on 10 out of 11 evaluation datasets and outperforms efficient fine-tuning baselines without any target task fine-tuning. ROSPR is also effective for challenging tasks such as tasks from BIG-bench (Srivastava et al., 2022), outperforming T0 by " + }, + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 69, + 139, + 291, + 382 + ], + "type": "text", + "content": " mean accuracy. Furthermore, we provide several interesting findings: (1) Variants of ROSPR that include interpolation of multiple prompt embeddings and scoring method that considers the answer choice distribution during retrieval further increases the effect of ROSPR (2) Also, we provide analysis of which factors attribute to the performance of ROSPR and show that, similarly to the role of demonstrations in in-context learning (Min et al., 2022), heuristic features such as answer choice format are more important than the similarity of the source task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 393, + 161, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 393, + 161, + 406 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 161, + 406 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 416, + 213, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 416, + 213, + 443 + ], + "spans": [ + { + "bbox": [ + 67, + 416, + 213, + 443 + ], + "type": "text", + "content": "2.1 Task Generalization with Instruction-Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 448, + 291, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 448, + 291, + 718 + ], + "spans": [ + { + "bbox": [ + 67, + 448, + 291, + 718 + ], + "type": "text", + "content": "Prompts and demonstrations are essential for task generalization since proper explanations are required for LMs to understand an unseen task (Kojima et al., 2022; Wei et al., 2022; Lampinen et al., 2022). Instruction-tuning, which is explicit multi-task prompted training on various downstream tasks, is a simple but effective way to achieve this, resulting in improved zero-shot capabilities. Zhong et al. (2021) first introduced the method of instruction-tuning by converting various tasks into a question-answering format and finetuning the model on the aggregated dataset. Following works (Mishra et al., 2022; Min et al., 2021; Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Xu et al., 2022; Ouyang et al., 2022; Ye et al., 2022; Chung et al., 2022) extended this approach on a larger scale and show that zero-shot task generalization could be enhanced with more diverse prompts, a larger number of training downstream tasks, and a larger LM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 728, + 198, + 740 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 728, + 198, + 740 + ], + "spans": [ + { + "bbox": [ + 67, + 728, + 198, + 740 + ], + "type": "text", + "content": "2.2 Source Task Retrieval" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Retrieving a source task that is relevant to the target task has shown to result in faster and better task" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 179 + ], + "type": "text", + "content": "adaptation. For parameter-efficient fine-tuning, Vu et al. (2022); Su et al. (2022) retrieve source prompt embedding that is similar to the target prompt embedding and obtain a better initialization point for prompt tuning. Instead of utilizing a single prompt embedding, recent works show a mixture of multiple prompt embeddings to be effective (Asai et al., 2022; Qin and Eisner, 2021)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 179, + 526, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 179, + 526, + 354 + ], + "spans": [ + { + "bbox": [ + 302, + 179, + 526, + 354 + ], + "type": "text", + "content": "For instruction-tuning, Lin et al. (2022) retrieve training instances that are similar to the query through a dense retriever and fine-tune the model using the retrieved examples. For in-context learning, Rubin et al. (2021); Liu et al. (2022b); Wang et al. (2023) retrieve training data that could be used for demonstrations. Wang et al. (2022c) show the effect of retrieving prompt embeddings in a continual learning setting. Although our proposed method is related to these works, the novelty of our work lies in applying source task retrieval in the zero-shot setting and retrieving soft prompts instead of training instances." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 365, + 365, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 365, + 365, + 378 + ], + "spans": [ + { + "bbox": [ + 303, + 365, + 365, + 378 + ], + "type": "text", + "content": "3 Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 387, + 525, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 387, + 525, + 562 + ], + "spans": [ + { + "bbox": [ + 302, + 387, + 525, + 562 + ], + "type": "text", + "content": "In this section, we introduce Retrieval of Prompt Tuning (ROSPR) for zero-shot task generalization. A detailed overview is shown in Figure 2. We first train source prompt embeddings of LM for each hard prompt given a source task using prompt tuning (Section 3.1). Then, we save training instance samples along with their prompt embeddings in the Source Prompt Library and use it to retrieve embeddings at inference to perform tasks in a zero-shot manner (Section 3.2). We additionally introduce interpolation of multiple source prompt embeddings (Section 3.3) and variance-based ranking (Section 3.4) to increase robustness and accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 572, + 506, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 302, + 572, + 506, + 585 + ], + "type": "text", + "content": "3.1 Training Source Prompt Embeddings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 590, + 525, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 590, + 525, + 711 + ], + "spans": [ + { + "bbox": [ + 302, + 590, + 525, + 711 + ], + "type": "text", + "content": "Even though RoSPR may be used to augment any type of LM, we use T0 (Sanh et al., 2021) as the backbone LM for this paper. For training of soft prompts, we utilize the source tasks and prompts used for the instruction-tuning phase of T0. While T0 was trained in a multi-task learning manner, we freeze the initial T0 parameters and train only soft prompts (source prompt embeddings) for each hard prompt of the source task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "Prompt Tuning Among various parameter-efficient fine-tuning methods, we follow prompt tuning proposed by Lester et al. (2021) because the number of trainable parameters is extremely small" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12289" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 71, + 502, + 219 + ], + "blocks": [ + { + "bbox": [ + 102, + 71, + 502, + 219 + ], + "lines": [ + { + "bbox": [ + 102, + 71, + 502, + 219 + ], + "spans": [ + { + "bbox": [ + 102, + 71, + 502, + 219 + ], + "type": "image", + "image_path": "7d9561a1f59a0b1caf4a6392d148c33d0c316d8439c27366ca7da20fc3a899fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 227, + 526, + 280 + ], + "lines": [ + { + "bbox": [ + 67, + 227, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 526, + 280 + ], + "type": "text", + "content": "Figure 2: An overview of RoSPR. For each hard prompt of the source datasets, soft prompts are trained via prompt tuning. After storing training instances as keys and corresponding prompt embedding as values, RoSPR searches training instances similar to query set " + }, + { + "bbox": [ + 67, + 227, + 526, + 280 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 67, + 227, + 526, + 280 + ], + "type": "text", + "content": ", retrieves the corresponding prompt embeddings, and selects the most frequently retrieved candidate for inference. Variants of selection strategy are also shown: RoSPR+INTER interpolates between multiple related source embeddings and RoSPR+VAR ranks candidate embeddings considering both frequency and variance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 285, + 290, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 285, + 290, + 326 + ], + "spans": [ + { + "bbox": [ + 67, + 285, + 290, + 326 + ], + "type": "inline_equation", + "content": "(\\sim 204\\mathrm{K}" + }, + { + "bbox": [ + 67, + 285, + 290, + 326 + ], + "type": "text", + "content": " parameters per prompt), which implies that the memory overhead of parameter retrieval at inference is negligible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": "For each source training dataset " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "(i = 1,\\dots,T)" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " is the total number of source datasets, we train source embeddings " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "E_{ij}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "(j = 1,\\dots,M_i)" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " is the number of hard prompts in " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": ", making soft prompt embeddings for each individual hard prompts. Specifically, given a training instance " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "\\{x_{ik},y_{ik}\\} (k = 1,\\dots,K)" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " is the number of sampled training instances per dataset, we first convert it into its hard prompted version " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "\\{h_j(x_{ik}),h_j(y_{ik})\\}" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "h_j(\\cdot)" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": " denotes adding the " + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 67, + 327, + 291, + 491 + ], + "type": "text", + "content": "-th hard prompt. Next, we train the LM with the following objective:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 503, + 290, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 503, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 113, + 503, + 290, + 525 + ], + "type": "interline_equation", + "content": "\\max _ {E _ {i j}} P \\left(h _ {j} \\left(y _ {i k}\\right) \\mid E _ {i j}; h _ {j} \\left(x _ {i k}\\right)\\right) \\tag {1}", + "image_path": "950918a34cc6a8e89eb9db64bb07bb40097132507f106ee6513eb9f4b6e794ff.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": "where all the parameters of the underlying backbone LM are frozen and only " + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "inline_equation", + "content": "E_{ij}" + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": " is trainable. In short, given " + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": ", we perform " + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": " number of prompt tunings for each hard prompts, resulting in " + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{T} M_{i}" + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": " total number of source prompt embeddings. For training efficiency, we only train on " + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "inline_equation", + "content": "K = 5000" + }, + { + "bbox": [ + 67, + 536, + 291, + 645 + ], + "type": "text", + "content": " training instances for a single epoch for each source prompt embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 656, + 246, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 656, + 246, + 670 + ], + "spans": [ + { + "bbox": [ + 67, + 656, + 246, + 670 + ], + "type": "text", + "content": "3.2 Zero-Shot Embedding Retrieval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 677, + 291, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 677, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 677, + 291, + 731 + ], + "type": "text", + "content": "After source prompt embedding training, we retrieve the most related source embeddings and select one from the retrieved candidates to be used at inference (right part of Figure 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 286, + 526, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 286, + 526, + 434 + ], + "spans": [ + { + "bbox": [ + 302, + 286, + 526, + 434 + ], + "type": "text", + "content": "We first construct a Source Prompt Library, consisting of sentence-level representations of training instance inputs as keys and the corresponding source prompt embedding as the values. For each available source prompt embedding, " + }, + { + "bbox": [ + 302, + 286, + 526, + 434 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 286, + 526, + 434 + ], + "type": "text", + "content": " number of samples are stored in the library. The sentence-level representations are obtained by getting the mean representation of hidden states of the last layer of the dense retriever. We use a T0-small encoder as a dense retriever, replicated based on Sanh et al. (2021) with smaller model size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "spans": [ + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "content": "At inference, we first randomly sample " + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "content": " query instances from the target task, following Lin et al. (2022). After obtaining sentence-level representations for each query through our T0-small encoder, we retrieve top- " + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "content": " examples for each query instance using MIPS (maximum inner product search) operation on our Source Prompt Library, retrieving a total of " + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "inline_equation", + "content": "Q*N" + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "content": " prompt embeddings. As the default methodology, among the retrieved embedding candidates, we select the most frequently retrieved prompt embedding as our designated soft prompt for the given target task and concatenate the embedding with each of the target task instances before feeding it to our backbone LM. In the next two subsections, we explain different strategies for calculating the target embedding from the " + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "inline_equation", + "content": "Q*N" + }, + { + "bbox": [ + 302, + 435, + 527, + 665 + ], + "type": "text", + "content": " prompt embedding candidates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 673, + 504, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 673, + 504, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 673, + 504, + 687 + ], + "type": "text", + "content": "3.3 Interpolation of Prompt Embeddings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 690, + 525, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 690, + 525, + 745 + ], + "spans": [ + { + "bbox": [ + 302, + 690, + 525, + 745 + ], + "type": "text", + "content": "When retrieving only a single prompt embedding for a given task (Section 3.2), it may result in high variance across evaluation prompts when the selected prompt embedding does not fit well with the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 525, + 772 + ], + "type": "text", + "content": "3For all indexing and searching, we use FAISS (Johnson et al., 2019) for fast source prompt embedding retrieval." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "text", + "content": "For each instances, the input and output are converted into its prompted version using the promptsource toolkit (Bach et al., 2022). An example is given in Appendix G." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12290" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "text", + "content": "given task. Recent works on prompt embedding retrieval have shown that the interpolation of prompt embeddings effectively transfers to the target task (Asai et al., 2022; Vu et al., 2022). We also explore calculating the target embedding through interpolation of multiple source embeddings instead of just using a single embedding. Among " + }, + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "inline_equation", + "content": "Q * N" + }, + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "text", + "content": " prompt candidates searched in Section 3.2, we select top-" + }, + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 67, + 71, + 293, + 315 + ], + "type": "text", + "content": " candidate embeddings based on the frequency of the search. Then, we calculate the weighted sum of the candidate embeddings, where the interpolation weight for each source embedding is based on the proportion of frequency. While Asai et al. (2022); Vu et al. (2022) require fine-tuning the target embeddings on the target task to calculate the weights for interpolation, our approach does not require any target task fine-tuning, enabling zero-shot task transfer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 326, + 211, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 326, + 211, + 338 + ], + "spans": [ + { + "bbox": [ + 67, + 326, + 211, + 338 + ], + "type": "text", + "content": "3.4 Variance-based Ranking" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 343, + 291, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 343, + 291, + 572 + ], + "spans": [ + { + "bbox": [ + 67, + 343, + 291, + 572 + ], + "type": "text", + "content": "Similar to the scoring and calibration method of Lu et al. (2022); Zhao et al. (2021), we introduce a scoring method applicable to zero-shot classification tasks that allows ranking the " + }, + { + "bbox": [ + 67, + 343, + 291, + 572 + ], + "type": "inline_equation", + "content": "Q * N" + }, + { + "bbox": [ + 67, + 343, + 291, + 572 + ], + "type": "text", + "content": " retrieved prompt embedding candidates by considering the answer choice distribution of the given target task as extra cues together with the original frequency cues. To accomplish this, we perform a forward pass with the concatenation of each candidate prompt embeddings together with the given hard prompt (only including the instruction, excluding the input instance) of the target task and give a higher score to the embedding candidate that results in lower variance. Ideally, the combination of soft and hard prompts should result in equal probability among the answer choices because the actual context of the task is not included." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "spans": [ + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "text", + "content": "Specifically, when given a target task with " + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "text", + "content": "-th hard prompt " + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "text", + "content": ", for each candidate embedding " + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "inline_equation", + "content": "E_{ij}" + }, + { + "bbox": [ + 67, + 575, + 291, + 615 + ], + "type": "text", + "content": " we calculate the scoring as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 625, + 290, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 625, + 290, + 661 + ], + "spans": [ + { + "bbox": [ + 84, + 625, + 290, + 661 + ], + "type": "interline_equation", + "content": "\\operatorname {S c o r e} \\left(h _ {k}, E _ {i j}\\right) = \\frac {\\operatorname {f r e q} \\left(h _ {k} , E _ {i j}\\right)}{\\sqrt {\\operatorname {V a r} \\left[ P \\left(y \\mid E _ {i j} , h _ {k}\\right) \\right]}} \\tag {2}", + "image_path": "0ac588b1c23357c3224e241ab8e74a6a79f1f083a095ae91039bcb685f10a410.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 671, + 290, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 671, + 290, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 671, + 290, + 698 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 671, + 290, + 698 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 671, + 290, + 698 + ], + "type": "text", + "content": " refers to the available output options of the target task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 710, + 203, + 724 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 710, + 203, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 710, + 203, + 724 + ], + "type": "text", + "content": "4 Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "content": "In this section, we explain the experimental settings of training of source prompt embedding and construction of our Source Prompt Library. We also" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 112 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 112 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 112 + ], + "type": "text", + "content": "explain our evaluation setting during zero-shot inference and baseline models. We provide detailed experiment configurations in Appendix F." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 122, + 393, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 122, + 393, + 133 + ], + "spans": [ + { + "bbox": [ + 302, + 122, + 393, + 133 + ], + "type": "text", + "content": "4.1 Source Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "spans": [ + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "content": "For training soft prompts through prompt tuning, we use the subset of source tasks used for the initial T0 instruction-tuning (Sanh et al., 2021)" + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "content": ". For each source task, we use the prompts for each dataset in T0, resulting in a total of 230 prompts. For Source Prompt Library construction, we sample only " + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "inline_equation", + "content": "n = 100" + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "content": " training instances per source embedding to minimize the inference latency. We show a variation of " + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "content": " and different methods to sample " + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 302, + 140, + 526, + 275 + ], + "type": "text", + "content": " training instances in Appendix D." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 285, + 411, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 285, + 411, + 297 + ], + "spans": [ + { + "bbox": [ + 302, + 285, + 411, + 297 + ], + "type": "text", + "content": "4.2 Evaluation Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 303, + 526, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 303, + 526, + 534 + ], + "spans": [ + { + "bbox": [ + 302, + 303, + 526, + 534 + ], + "type": "text", + "content": "Following Sanh et al. (2021), we evaluate on the validation set of 4 held-out tasks (natural language inference, sentence completion, coreference resolution, word sense disambiguation) resulting in a total of 11 evaluation datasets. We also follow Sanh et al. (2021) and evaluate on 14 different datasets from the BIG-bench benchmark (Srivastava et al., 2022). We use rank classification evaluation method by selecting the output option with higher log-likelihood following Brown et al. (2020); Sanh et al. (2021). For all evaluation tasks, we use accuracy as an evaluation metric and report the mean accuracy and standard deviation of all of the evaluation prompts per given dataset (average of " + }, + { + "bbox": [ + 302, + 303, + 526, + 534 + ], + "type": "inline_equation", + "content": "\\sim 10" + }, + { + "bbox": [ + 302, + 303, + 526, + 534 + ], + "type": "text", + "content": " prompts per evaluation dataset). For BIG-bench tasks, we do not report standard deviation because only one prompt is provided per task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 543, + 407, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 543, + 407, + 555 + ], + "spans": [ + { + "bbox": [ + 302, + 543, + 407, + 555 + ], + "type": "text", + "content": "4.3 Baseline Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 561, + 525, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 561, + 525, + 628 + ], + "spans": [ + { + "bbox": [ + 302, + 561, + 525, + 628 + ], + "type": "text", + "content": "Zero-shot Baseline For zero-shot baseline models, we show the results of T0 (3B) together with a 4 times larger T0 (11B) instruction-tuned model. We also compare with GPT-3 (175B) model which is 60 times larger than T0 (3B)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 637, + 525, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 637, + 525, + 692 + ], + "spans": [ + { + "bbox": [ + 302, + 637, + 525, + 692 + ], + "type": "text", + "content": "Fine-tuning Baseline We also compare with efficient fine-tuning baseline models that utilize prompt tuning. These models require target task prompt tuning, indicating that zero-shot transfer" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 699, + 525, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 699, + 525, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 699, + 525, + 719 + ], + "type": "text", + "content": "4We use 29 out of 38 datasets that are used to train T0. We explain the training task selection rationale in Appendix H.1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 720, + 525, + 740 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 720, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 720, + 525, + 740 + ], + "type": "text", + "content": "5We provide the full list of evaluation datasets in Appendix H." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 741, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 741, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 741, + 524, + 772 + ], + "type": "text", + "content": "For methods based on RoSPR, we report the performance average of 3 runs with different random seeds for the sampling of evaluation queries used for the prompt retrieval." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "12291" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 68, + 525, + 185 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 525, + 185 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 525, + 185 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 525, + 185 + ], + "type": "table", + "html": "
Method# of Param\n(Base/Trainable)NLISentence CompletionCoreference Resolut.WSDTotal Avg.
RTECBAN. R1AN. R2AN. R3COPAHellasw.StoryC.Winogr.WSCWiCMeanSTD
T03B / 064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.223.62
PT (FT)3B / 204K63.1444.5233.0731.4432.9473.0026.39-50.6746.7350.02--
ATTEMPT (FT)3B / 614K68.9544.8836.1934.7334.8174.3826.76-51.3364.9051.05--
T0+RoSPR3B / 204K71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.243.62
w/ INTER3B / 204K70.7152.3037.3034.3433.8978.2526.9485.6251.1064.5250.7353.243.30
w/ VAR3B / 204K71.7850.3637.0734.5833.9078.8827.0185.5251.4564.9451.9453.383.38
w/ VAR & INTER3B / 204K72.6051.9837.2534.3133.9577.8326.8485.5850.9364.9751.1853.403.47
ORACLE3B / 204K73.7958.1037.6534.9234.9181.1327.7587.5752.3668.1755.2655.603.07
T011B / 080.8370.1243.5638.6841.2690.0233.5892.4059.9461.4556.5860.76
GPT-3175B / 063.5046.4034.6035.4034.5091.0078.9083.2070.2065.4045.9259.00-
", + "image_path": "1393e1b953f1ef5f4db6e8065fc94b4d7459c55de8397bfe54e16aa16a79fee8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 194, + 525, + 267 + ], + "lines": [ + { + "bbox": [ + 66, + 194, + 525, + 267 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 525, + 267 + ], + "type": "text", + "content": "Table 1: RoSPR refers to our main proposed method, W/ INTER refers to applying interpolation of multiple source embedding candidates, W/ VAR refers to retrieval through variance-based ranking, W/ VAR & INTER refers to applying both interpolation and variance-based ranking where the interpolation weight is based on the variance-based ranking score, and ORACLE refers performance when the most optimal source embedding is retrieved from the candidates, acting as an upper bound performance for retrieval. FT refers to fine-tuned models on the target tasks. For FT models, we exclude StoryCloze due to the absence of training instances. The best and second-best performance is shown in bold and underline respectively. Comparison with hard prompt optimization techniques and visualization of the results is shown in Appendix A and Appendix E, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 277, + 291, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 277, + 291, + 454 + ], + "spans": [ + { + "bbox": [ + 66, + 277, + 291, + 454 + ], + "type": "text", + "content": "is infeasible. Similar to our source prompt tuning process, we train each target prompt for a single epoch with a maximum of 5,000 training instances. The first baseline model is naive prompt tuning on the target tasks without any prompt retrieval, referred to as PT (Lester et al., 2021). The second model is ATTEMPT (Asai et al., 2022), which trains the target soft prompts through attentional mixtures of source prompts. Because StoryCloze (Mostafazadeh et al., 2016) does not contain training instances, we exclude the dataset for fine-tuning. More training details of fine-tuning baseline are specified in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 465, + 199, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 465, + 199, + 479 + ], + "spans": [ + { + "bbox": [ + 67, + 465, + 199, + 479 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 488, + 291, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 488, + 291, + 691 + ], + "spans": [ + { + "bbox": [ + 66, + 488, + 291, + 691 + ], + "type": "text", + "content": "RoSPR enhances the performance of T0 efficiently. Table 1 shows the performance of the 11 evaluation datasets. T0+RoSPR outperforms T0 on 10 datasets among the 11 evaluation datasets. Specifically, T0+RoSPR outperforms T0 on RTE (+6.99% points), CB (+4.12% points), ANLI R1 (+3.24% points), and COPA (+2.87% points). This shows that soft prompt retrieval assists hard prompts for zero-shot task generalization with a negligible number of additional parameters (0.007%). Also, while T0 outperforms GPT-3 on 3 datasets (RTE, StoryCloze, WiC), T0+RoSPR additionally outperforms GPT-3 on 2 datasets (ANLI R1 and CB) and enlarging the score gap for RTE, StoryCloze and WiC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 277, + 526, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 277, + 526, + 467 + ], + "spans": [ + { + "bbox": [ + 302, + 277, + 526, + 467 + ], + "type": "text", + "content": "ROSPR also outperforms finetuning baselines even without utilizing any training instances of the target task. We first observe that PT harms the performance of the backbone model, which aligns with the result of Liu et al. (2022a); Gu et al. (2022) that prompt tuning is unstable when the training instances or the training steps are small. By comparing ATTEMPT with ROSPR, ROSPR outperforms ATTEMPT on 7 out of 10 tasks and " + }, + { + "bbox": [ + 302, + 277, + 526, + 467 + ], + "type": "inline_equation", + "content": "1.21\\%" + }, + { + "bbox": [ + 302, + 277, + 526, + 467 + ], + "type": "text", + "content": " points on the mean accuracy of 10 tasks. This shows that ROSPR is more applicable for efficient adaptation because it requires 3 times fewer additional parameters compared to ATTEMPT and also does not require any further fine-tuning of the target task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "content": "INTER and VAR enhance the performance of ROSPR. We also analyze the effect of introducing variants of ROSPR: interpolation of soft prompts (INTER) and variance-based ranking (VAR) in Table 1. First, applying INTER shows similar accuracy compared to ROSPR. However, as shown in the last column of Table 1, INTER reduces the standard deviation of T0 and ROSPR by " + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "inline_equation", + "content": "8.84\\%" + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "content": " while improving the mean accuracy of T0, indicating increased robustness to different surface forms of evaluation prompts. This indicates that interpolation of multiple source embeddings outperforms a single source embedding retrieval, aligning with the result of Asai et al. (2022). Applying VAR with " + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{T0 + ROSPR}" + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "content": " improves both zero-shot accuracy and robustness of " + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "inline_equation", + "content": "\\mathrm{T0 + ROSPR}" + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "content": ", showing that considering the answer choice distribution is beneficial for zero-shot setting, aligned with results from Zhao et al. (2021); Shi et al. (2022). Moreover, applying both VAR+INTER results in the highest overall average accuracy, outperforming T0 by " + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "inline_equation", + "content": "2.18\\%" + }, + { + "bbox": [ + 302, + 476, + 526, + 773 + ], + "type": "text", + "content": " points by largely reducing the gap between larger LLMs." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "text", + "content": "One exception is WSC, which is a binary classification task (yes/no) predicting whether the reference of the pronoun is correct. We observed that the evaluation data of this dataset has unbalanced labels, containing over " + }, + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "text", + "content": " of \"No\" labels. This might be the reason why T0-11B underperforms T0-3B only on this dataset (Sanh et al., 2021). Indeed, predicting only \"No\" on this dataset outperforms T0-11B " + }, + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "inline_equation", + "content": "(63.46 > 61.45)" + }, + { + "bbox": [ + 67, + 700, + 290, + 772 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12292" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 278, + 184 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 278, + 184 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 278, + 184 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 278, + 184 + ], + "type": "image", + "image_path": "5ac012d05c905fb7395d0a8393846221f227d53d80a09f3bab98b34d3d8e99bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "type": "text", + "content": "Figure 3: Mean accuracy of 14 datasets of BIG-bench. We evaluate on a single prompt following Sanh et al. (2021). By only adding " + }, + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "type": "inline_equation", + "content": "0.007\\%" + }, + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "type": "text", + "content": " parameters to T0-3B, " + }, + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "type": "inline_equation", + "content": "\\mathrm{T0 + ROSPR}" + }, + { + "bbox": [ + 67, + 194, + 291, + 247 + ], + "type": "text", + "content": " largely reduces the performance gap between 4 times larger T0-11B. The full result is provided in Appendix B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "text", + "content": "The effect of ROSPR generalizes to challenging tasks. ROSPR is also effective for challenging tasks such as tasks from BIG-bench benchmark. As shown in Figure 3, T0+ROSPR improves the mean accuracy performance of T0-3B by " + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "text", + "content": " points while only adding " + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "inline_equation", + "content": "0.007\\%" + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "text", + "content": " additional parameters. T0+ROSPR also outperforms 60 times larger zero-shot and 1-shot GPT-3 and largely reduces the performance gap between 4 times larger T0-11B (1.84% points) or 60 times larger 3-shot GPT-3 (0.53% points). Applying INTER with T0+ROSPR results in additional mean accuracy enhancement, outperforming T0-3B by " + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "inline_equation", + "content": "2.67\\%" + }, + { + "bbox": [ + 67, + 251, + 291, + 427 + ], + "type": "text", + "content": " points8." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 437, + 188, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 437, + 188, + 449 + ], + "spans": [ + { + "bbox": [ + 67, + 437, + 188, + 449 + ], + "type": "text", + "content": "6 Analysis of ROSPR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 457, + 291, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 291, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 291, + 673 + ], + "type": "text", + "content": "Zero-shot task adaptation of LMs is often seen as a problem of task location, locating the target task to where the model can solve it using the intrinsic ability obtained at pretraining stage with the aid of prompts and demonstrations (Reynolds and McDonell, 2021). In this section, we analyze which factors contribute to the performance enhancement in the perspective of identifying better task location. We find that although the target task performance depends on the source task types, heuristic features such as the answer choice format are more important. This agrees with previous findings that a met-trained LM focuses on simple features such as the label space, the input distribution, and sequence format, instead of complex semantics (Webson and Pavlick, 2021; Min et al., 2022)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 681, + 291, + 735 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 291, + 735 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 291, + 735 + ], + "type": "text", + "content": "Target task performance depends on source task types. To analyze the effect of different source task types on each target task, we measure the frequency ratio of each source task type that results" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 515, + 186 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 515, + 186 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 515, + 186 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 515, + 186 + ], + "type": "image", + "image_path": "f31b568c9169f5ed17a16f3d6271df6eb735e813c92b9945d226e18762868bc8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 195, + 526, + 219 + ], + "lines": [ + { + "bbox": [ + 302, + 195, + 526, + 219 + ], + "spans": [ + { + "bbox": [ + 302, + 195, + 526, + 219 + ], + "type": "text", + "content": "Figure 4: Frequency of source task types (x-axis) that maximizes (i.e. ORACLE) the accuracy of each target task (y-axis)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 224, + 526, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 224, + 526, + 346 + ], + "spans": [ + { + "bbox": [ + 302, + 224, + 526, + 346 + ], + "type": "text", + "content": "in the best performance (ORACLE) for the given prompts of the target tasks (visualized in Figure 4). From this figure, we can observe a few patterns: paraphrase task assists NLI and word sense disambiguation task while multi-choice QA (MQA) task assists sentence completion task. For coreference resolution task, various source task types (paraphrase, summarization, multi-choice QA) assist the target task." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 353, + 518, + 451 + ], + "blocks": [ + { + "bbox": [ + 307, + 353, + 518, + 451 + ], + "lines": [ + { + "bbox": [ + 307, + 353, + 518, + 451 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 518, + 451 + ], + "type": "image", + "image_path": "e471b75fad8028d974258ce484f8e07e0b8f08de5928b1e837b9038f781e0359.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 455, + 525, + 497 + ], + "lines": [ + { + "bbox": [ + 302, + 455, + 525, + 497 + ], + "spans": [ + { + "bbox": [ + 302, + 455, + 525, + 497 + ], + "type": "text", + "content": "Figure 5: Effect of answer choice format alignment across different target datasets (RTE, COPA, WiC). We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 301, + 503, + 526, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 503, + 526, + 624 + ], + "spans": [ + { + "bbox": [ + 301, + 503, + 526, + 624 + ], + "type": "text", + "content": "Answer choice format is important for task location. We also analyze the effect of using different answer choice formats with the same source task. Answer choice format decides how the available answer choices are given to the LM through the input. For example, a prompt that requires classifying a movie review into good/bad has a different answer choice format from classifying it into positive/negative." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "content": "We experiment on 3 datasets (RTE, COPA, WiC) which correspond to different tasks (NLI, sentence completion, word sense disambiguation) respectively. For each dataset, we select a source dataset that is retrieved the most for ORACLE. Among the source prompts of the selected source dataset, we select a prompt that has the same answer choice format as the target task (ALIGNED) and another prompt that has a different answer choice format (MISALIGNED). Figure 5 shows the effect of answer choice format alignment on the target task" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": "8We observe that applying VAR results in the same performance as T0+ROSPR because frequency of retrieval has much more influence than variance for BIG-bench tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12293" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 274 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 274 + ], + "type": "text", + "content": "performance by comparing ALIGNED and MIS-ALIGNED. The result shows that for all 3 datasets, ALIGNED significantly outperforms MISALIGNED. This result is non-trivial considering that the two prompt embeddings are trained on the same source training dataset and the same training configuration, with the only difference in the given answer choice format, implying that how the answer choices are given to solve a specific task is more important than the content of the training data for task location. ROSPR is mostly comparable to ALIGNED embedding, implying that retrieving a source prompt embedding by searching for similar input instances results in retrieving a source embedding with similar answer choice formats." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 280, + 256, + 373 + ], + "blocks": [ + { + "bbox": [ + 109, + 280, + 256, + 373 + ], + "lines": [ + { + "bbox": [ + 109, + 280, + 256, + 373 + ], + "spans": [ + { + "bbox": [ + 109, + 280, + 256, + 373 + ], + "type": "image", + "image_path": "273c898d6a37ac95ab5400e985e36458d79a5b3199a888f1e549dd322b7aa5ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 109, + 373, + 256, + 462 + ], + "blocks": [ + { + "bbox": [ + 109, + 373, + 256, + 462 + ], + "lines": [ + { + "bbox": [ + 109, + 373, + 256, + 462 + ], + "spans": [ + { + "bbox": [ + 109, + 373, + 256, + 462 + ], + "type": "image", + "image_path": "9a0360ca8488b3b4ce587546cde7130424806ceb3a518be9d3e8b43cdc6ad70e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 471, + 291, + 522 + ], + "lines": [ + { + "bbox": [ + 67, + 471, + 291, + 522 + ], + "spans": [ + { + "bbox": [ + 67, + 471, + 291, + 522 + ], + "type": "text", + "content": "Figure 6: RTE (Top) and WiC (Bottom) evaluation result by retrieving MISALIGNED and ALIGNED answer choice format across various source tasks. We report the mean accuracy of the evaluation prompts and the performance of T0 is shown in green dotted line." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 523, + 291, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 711 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 711 + ], + "type": "text", + "content": "We additionally analyze the effect of answer choice formats on RTE and WiC datasets by retrieving prompt embeddings trained on various source tasks. Both target datasets have the answer choices of either yes/no. Similar to the previous experiments, we retrieve ALIGNED (yes/no format) and MISALIGNED prompt embeddings across three source tasks: paraphrase, sentiment classification, and multi-choice QA. As shown in Figure 6, for both target datasets, ALIGNED outperforms MISALIGNED across all three source tasks. This shows that aligning to the answer choice format of the target task is crucial regardless of the retrieved source task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 774 + ], + "type": "text", + "content": "Answer choice format is more important than task similarity. From Figure 6, we can see that all three source tasks benefit from aligning to the target task answer choice format. One may think" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "type": "text", + "content": "that embeddings from source tasks requiring similar knowledge to the target task may be important. Counterintuitively, for both RTE and WiC target tasks, when the answer choice format is aligned, the task source embedding of sentiment classification, which is known to be irrelevant to RTE and WiC (Pruksachatkun et al., 2020), outperforms other embeddings sourced from datasets that are more relevant to the target datasets (paraphrase and multi-choice QA) (Appendix G). This implies that for retrieval of source embedding for task location, answer choice format is more important than containing similar knowledge required to solve the target task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 268, + 526, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 268, + 526, + 674 + ], + "spans": [ + { + "bbox": [ + 302, + 268, + 526, + 674 + ], + "type": "text", + "content": "Role of ROSPR is similar to in-context learning. From the findings explained in previous paragraphs, we can conclude that although the source task types influence the target task performance, retrieving a similar answer choice format is more important for task location. Indeed, source tasks containing similar knowledge can help target tasks only if the answer choice formats are aligned to the target task. These findings support Min et al. (2022); Webson and Pavlick (2021) that a meta-trained LM \"takes less effort\" to understand the input: models exploit the simple aspects of prompts and demonstrations such as the format and distribution instead of complex semantics. Especially, for in-context learning, Xie et al. (2021); Min et al. (2022) show that the role of demonstrations lies in providing the shared concept and distribution hints of the target task. From this aspect, the role of ROSPR is similar to demonstrations. However, it is more efficient than including demonstrations because it avoids heavy computation at inference from long sequence lengths (Liu et al., 2022a; Choi et al., 2022) since ROSPR prepends a fixed length of prefix tokens regardless of the task. Also, ROSPR is free from the instability of in-context learning coming from different orderings of demonstrations (Lu et al., 2022; Zhao et al., 2021). Lastly, we conjecture that ROSPR also has the benefits of soft prompts (Li and Liang, 2021) such as having more expressiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 684, + 409, + 697 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 684, + 409, + 697 + ], + "spans": [ + { + "bbox": [ + 302, + 684, + 409, + 697 + ], + "type": "text", + "content": "7 Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "content": "In this section, we analyze the effect of the number of (1) prompts, (2) source datasets, and (3) queries sampled for evaluation. We evaluate variations of our proposed methods on 4 datasets: RTE for NLI, COPA for sentence completion, Winogrande" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12294" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 71, + 174, + 165 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 174, + 165 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 174, + 165 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 174, + 165 + ], + "type": "image", + "image_path": "1defe8e68d4ed49a58d8f3c1201254db1ffa3563b1d5ecf359d4894f1cdb96b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 170, + 169, + 181 + ], + "lines": [ + { + "bbox": [ + 82, + 170, + 169, + 181 + ], + "spans": [ + { + "bbox": [ + 82, + 170, + 169, + 181 + ], + "type": "text", + "content": "(a) Datasets vs Prompts" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 191, + 290, + 243 + ], + "lines": [ + { + "bbox": [ + 67, + 191, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 290, + 243 + ], + "type": "text", + "content": "Figure 7: Result of various ablation settings of ROSPR. (a) compares the effect of scaling the number of datasets with scaling the number of prompts and (b) shows the effect of the number of sampled queries at inference. Additional ablation results are shown in Appendix D." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 183, + 68, + 280, + 165 + ], + "blocks": [ + { + "bbox": [ + 183, + 68, + 280, + 165 + ], + "lines": [ + { + "bbox": [ + 183, + 68, + 280, + 165 + ], + "spans": [ + { + "bbox": [ + 183, + 68, + 280, + 165 + ], + "type": "image", + "image_path": "f71eb27c7fedf771f1e614bb3dba7a2722033716596e727b6f5cae0fcd0d36bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 170, + 267, + 180 + ], + "lines": [ + { + "bbox": [ + 196, + 170, + 267, + 180 + ], + "spans": [ + { + "bbox": [ + 196, + 170, + 267, + 180 + ], + "type": "text", + "content": "(b) Query variation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 249, + 291, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 249, + 291, + 303 + ], + "spans": [ + { + "bbox": [ + 67, + 249, + 291, + 303 + ], + "type": "text", + "content": "for coreference, and WiC for word sense disambiguation task. We report the average of the mean accuracy of all evaluation prompts for each dataset by running 3 different runs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 310, + 290, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 310, + 290, + 553 + ], + "spans": [ + { + "bbox": [ + 69, + 310, + 290, + 553 + ], + "type": "text", + "content": "Scaling number of prompts vs. number of datasets. Recent works on meta-trained LMs show that the number of source datasets and prompts is an important factor for zero-shot task generalization (Sanh et al., 2021; Wei et al., 2021; Wang et al., 2022b; Chung et al., 2022). We also show ablations for ROSPR and measure how the zero-shot generalization performance changes when we vary the number of prompts and datasets available during the prompt tuning stage (shown in Figure 7a). First, we vary (1) the total number of source prompts by 60, 120, and 230 by increasing the number of prompts per dataset and (2) the number of datasets by 8, 16, and 30 by increasing the number of datasets per task cluster. Note that in (1), the total number of datasets is fixed while in (2), we use all available prompts for each dataset while varying the number of datasets per task cluster." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 554, + 290, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 290, + 661 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 290, + 661 + ], + "type": "text", + "content": "In contrast to (1), (2) does not always lead to a linearly increasing performance boost; the performance saturates as more source datasets are included. By comparing the effect of scaling datasets and scaling prompts for similar Source Prompt Library sizes, we observe that the number of prompts has more impact on the accuracy of the target task (Figure 7a)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 662, + 290, + 743 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 662, + 290, + 743 + ], + "spans": [ + { + "bbox": [ + 67, + 662, + 290, + 743 + ], + "type": "text", + "content": "This ablation study also supports the analysis of the previous section; diverse answer choice formats of prompts, which are mostly influenced by the total number of source prompts, are more important than source task types which are influenced by the number of source datasets.[10] Therefore, if" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": "the number of task clusters is sufficient to some extent, scaling the number of source prompts per dataset is more crucial than scaling the number of source datasets per task cluster." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 133, + 526, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 133, + 526, + 295 + ], + "spans": [ + { + "bbox": [ + 302, + 133, + 526, + 295 + ], + "type": "text", + "content": "More sampled queries improve the performance. We also analyze the effect of the number of query instances sampled at inference for retrieval. As seen in Figure 7b, increasing the number of queries results in higher mean accuracy. This is different from the analysis of Lin et al. (2022) that sampling more queries leads to better performance only to some point. Because we use the frequency of each prompt embedding candidate as the default metric for retrieval, utilizing more query instances would represent the evaluation data more accurately, resulting in a reduced number of wrong retrievals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 306, + 381, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 306, + 381, + 318 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 381, + 318 + ], + "type": "text", + "content": "8 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 327, + 525, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 525, + 665 + ], + "type": "text", + "content": "In this paper, we introduce ROSPR, a method that efficiently enhances zero-shot generalization capabilities of a meta-trained LM by retrieving prompt-specific source prompt embeddings (soft prompts) for a given target task. We accomplish this by first training the soft prompts for hard prompt of the source tasks. After training source prompt embeddings, we construct the Source Prompt Library by storing the mean representation of training instances as keys and the corresponding prompt embeddings as values. At inference, we search for training instances stored in the library similar to sample instances from the target task, retrieve the corresponding prompt embedding, select the most frequently retrieved embedding, and append it to each of the target task instances for prediction. Our results show that ROSPR efficiently enhances the zero-shot performance of the backbone model while introducing minimal additional parameters during inference. We additionally provide analysis of which factors attribute to the performance of ROSPR and find that heuristic cues such as the answer choice format are critical for generalization performance, implying that it may play a role similar to demonstrations in in-context learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 676, + 384, + 688 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 676, + 384, + 688 + ], + "spans": [ + { + "bbox": [ + 302, + 676, + 384, + 688 + ], + "type": "text", + "content": "9 Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 698, + 525, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 698, + 525, + 725 + ], + "spans": [ + { + "bbox": [ + 302, + 698, + 525, + 725 + ], + "type": "text", + "content": "Although we show the effectiveness of ROSPR by applying it on T0-3B (Sanh et al., 2021), we did" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "text", + "content": "the number of datasets increases, we find that answer choice formats are similar across the same task type, meaning that the diversity of answer choice formats is not increased by increasing the number of datasets per task clusters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 79, + 750, + 289, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 750, + 289, + 761 + ], + "spans": [ + { + "bbox": [ + 79, + 750, + 289, + 761 + ], + "type": "inline_equation", + "content": "{}^{9}" + }, + { + "bbox": [ + 79, + 750, + 289, + 761 + ], + "type": "text", + "content": " Task cluster is defined as a cluster of the same task types." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 79, + 761, + 289, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 761, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 79, + 761, + 289, + 772 + ], + "type": "text", + "content": "10 Although the total number of prompts also increases as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12295" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 248 + ], + "type": "text", + "content": "not evaluate our method on different model scales such as the T0-11B variant and other LM architectures such as decoder-only LMs due to limited computational resources. This leaves future works on applying ROSPR to even larger LMs and diverse LM architectures (Wang et al., 2022a). Moreover, it is hard to apply VAR to target tasks without answer choices such as free-form generation because variance among options cannot be obtained. However, ROSPR and ROSPR+INTER can still be utilized and we leave applying ROSPR on zero-shot task location of free-form generation as future work (Scialom et al., 2022)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 256, + 170, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 256, + 170, + 271 + ], + "spans": [ + { + "bbox": [ + 68, + 256, + 170, + 271 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "spans": [ + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "text", + "content": "We thank Sohee Yang and Eunbi Choi for helpful feedback on the paper. This work was partly supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2022-0-00264, Comprehensive Video Understanding and Generation with Knowledge-based Deep Logic Neural Network, " + }, + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "text", + "content": "; No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST), " + }, + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 67, + 277, + 292, + 413 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 433, + 127, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 433, + 127, + 447 + ], + "spans": [ + { + "bbox": [ + 68, + 433, + 127, + 447 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 452, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 452, + 291, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 452, + 291, + 509 + ], + "spans": [ + { + "bbox": [ + 69, + 452, + 291, + 509 + ], + "type": "text", + "content": "Akari Asai, Mohammadreza Salehi, Matthew E Peters, and Hannaneh Hajishirzi. 2022. Attentional mixtures of soft prompt tuning for parameter-efficient multi-task knowledge sharing. arXiv preprint arXiv:2205.11961." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 515, + 291, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 515, + 291, + 681 + ], + "spans": [ + { + "bbox": [ + 69, + 515, + 291, + 681 + ], + "type": "text", + "content": "Stephen Bach, Victor Sanh, Zheng Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-david, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Fries, Maged Alshaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang, Dragomir Radev, Mike Tian-jian Jiang, and Alexander Rush. 2022. Prompt-Source: An integrated development environment and repository for natural language prompts. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 93-104, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 687, + 291, + 743 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 687, + 291, + 743 + ], + "spans": [ + { + "bbox": [ + 69, + 687, + 291, + 743 + ], + "type": "text", + "content": "Max Bartolo, Alastair Roberts, Johannes Welbl, Sebastian Riedel, and Pontus Stenetorp. 2020. Beat the AI: Investigating adversarial human annotation for reading comprehension. Transactions of the Association for Computational Linguistics, 8:662-678." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 750, + 290, + 772 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "type": "text", + "content": "Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 125, + 526, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 159 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 159 + ], + "type": "text", + "content": "Eunbi Choi, Yongrae Jo, Joel Jang, and Minjoon Seo. 2022. Prompt injection: Parameterization of fixed inputs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 168, + 526, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 526, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 526, + 234 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 243, + 526, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 526, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 526, + 300 + ], + "type": "text", + "content": "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2022. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 308, + 526, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 526, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 526, + 354 + ], + "type": "text", + "content": "Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In *Machine learning challenges workshop*, pages 177-190. Springer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 526, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 526, + 417 + ], + "type": "text", + "content": "Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. 2019. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pages 107-124." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 426, + 526, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 526, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 526, + 502 + ], + "type": "text", + "content": "Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMSum corpus: A human-annotated dialogue dataset for abstractive summarization. In Proceedings of the 2nd Workshop on New Frontiers in Summarization, pages 70–79, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 512, + 526, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 526, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 526, + 547 + ], + "type": "text", + "content": "David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2003. English gigaword. Linguistic Data Consortium, Philadelphia, 4(1):34." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 555, + 526, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 526, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 526, + 623 + ], + "type": "text", + "content": "Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. 2022. PPT: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 526, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 526, + 731 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 526, + 731 + ], + "type": "text", + "content": "Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2391-2401, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 739, + 526, + 772 + ], + "type": "text", + "content": "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 7(3):535-547." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12296" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 116 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 116 + ], + "type": "text", + "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 125, + 290, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 125, + 290, + 190 + ], + "spans": [ + { + "bbox": [ + 69, + 125, + 290, + 190 + ], + "type": "text", + "content": "Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Kory Matthewson, Michael Henry Tessler, Antonia Creswell, James L McClelland, Jane X Wang, and Felix Hill. 2022. Can language models learn from explanations in context? arXiv preprint arXiv:2204.02329." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 200, + 290, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 200, + 290, + 266 + ], + "spans": [ + { + "bbox": [ + 69, + 200, + 290, + 266 + ], + "type": "text", + "content": "Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural text generation from structured data with application to the biography domain. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1203-1213, Austin, Texas. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 275, + 290, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 275, + 290, + 340 + ], + "spans": [ + { + "bbox": [ + 69, + 275, + 290, + 340 + ], + "type": "text", + "content": "Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, D. Kontokostas, Pablo N. Mendes, Sebastian Hellmann, M. Morsey, Patrick van Kleef, S. Auer, and C. Bizer. 2015. Dbpedia - a large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6:167-195." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 349, + 290, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 349, + 290, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 349, + 290, + 427 + ], + "type": "text", + "content": "Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 435, + 290, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 435, + 290, + 481 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 290, + 481 + ], + "type": "text", + "content": "Hector Levesque, Ernest Davis, and Leora Morgenstern. 2012. The winograd schema challenge. In Thirteenth international conference on the principles of knowledge representation and reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 489, + 290, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 489, + 290, + 675 + ], + "spans": [ + { + "bbox": [ + 69, + 489, + 290, + 675 + ], + "type": "text", + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumont, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierric Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "type": "text", + "content": "Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 95 + ], + "type": "text", + "content": "Chonghua Liao, Yanan Zheng, and Zhilin Yang. 2022. Zero-label prompt selection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 102, + 525, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 525, + 147 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 525, + 147 + ], + "type": "text", + "content": "Bill Yuchen Lin, Kangmin Tan, Chris Miller, Beiwen Tian, and Xiang Ren. 2022. Unsupervised cross-task generalization via retrieval augmentation. arXiv preprint arXiv:2204.07937." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 154, + 525, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 525, + 232 + ], + "type": "text", + "content": "Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. CommonGen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823-1840, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 239, + 525, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 525, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 525, + 306 + ], + "type": "text", + "content": "Kevin Lin, Oyvind Tafjord, Peter Clark, and Matt Gardner. 2019. Reasoning over paragraph effects in situations. In Proceedings of the 2nd Workshop on Machine Reading for Question Answering, pages 58-62, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 313, + 525, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 525, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 525, + 369 + ], + "type": "text", + "content": "Haokun Liu, Derek Tam, Mohammed Muqeeth, Jay Mohta, Tenghao Huang, Mohit Bansal, and Colin Raffel. 2022a. Few-shot parameter-efficient fine-tuning is better and cheaper than in-context learning. arXiv preprint arXiv:2205.05638." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 376, + 525, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 525, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 525, + 465 + ], + "type": "text", + "content": "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2022b. What makes good in-context examples for GPT-3? In Proceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures, pages 100–114, Dublin, Ireland and Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 472, + 525, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 525, + 560 + ], + "type": "text", + "content": "Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 568, + 525, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 525, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 525, + 657 + ], + "type": "text", + "content": "Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 525, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 525, + 730 + ], + "type": "text", + "content": "Julian J. McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In Seventh ACM Conference on Recommender Systems, RecSys '13, Hong Kong, China, October 12-16, 2013, pages 165-172. ACM." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 738, + 525, + 772 + ], + "type": "text", + "content": "Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2021. Metaicl: Learning to learn in context. arXiv preprint arXiv:2110.15943." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12297" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettle-moyer. 2022. Rethinking the role of demonstrations: What makes in-context learning work? arXiv preprint arXiv:2202.12837." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "spans": [ + { + "bbox": [ + 69, + 134, + 291, + 211 + ], + "type": "text", + "content": "Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3470-3487, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 218, + 291, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 218, + 291, + 328 + ], + "spans": [ + { + "bbox": [ + 69, + 218, + 291, + 328 + ], + "type": "text", + "content": "Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839-849, San Diego, California. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 336, + 291, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 336, + 291, + 413 + ], + "spans": [ + { + "bbox": [ + 69, + 336, + 291, + 413 + ], + "type": "text", + "content": "Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1797-1807, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 420, + 291, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 420, + 291, + 497 + ], + "spans": [ + { + "bbox": [ + 69, + 420, + 291, + 497 + ], + "type": "text", + "content": "Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial NLI: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 505, + 291, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 505, + 291, + 570 + ], + "spans": [ + { + "bbox": [ + 69, + 505, + 291, + 570 + ], + "type": "text", + "content": "Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 577, + 291, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 577, + 291, + 655 + ], + "spans": [ + { + "bbox": [ + 69, + 577, + 291, + 655 + ], + "type": "text", + "content": "Bo Pang and Lillian Lee. 2005. Seeing stars: Exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05), pages 115-124, Ann Arbor, Michigan. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 662, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 662, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 662, + 291, + 772 + ], + "type": "text", + "content": "Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Roktaschel, and Sebastian Riedel. 2021. KILT: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2523-2544, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 171 + ], + "type": "text", + "content": "Mohammad Taher Pilehvar and Jose Camacho-Collados. 2019. WiC: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1267-1273, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 181, + 526, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 526, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 526, + 280 + ], + "type": "text", + "content": "Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, and Samuel R. Bowman. 2020. Intermediate-task transfer learning with pretrained language models: When and why does it work? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5231-5247, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 290, + 526, + 368 + ], + "type": "text", + "content": "Guanghui Qin and Jason Eisner. 2021. Learning how to ask: Querying LMs with mixtures of soft prompts. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5203-5212, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 526, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 526, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 526, + 444 + ], + "type": "text", + "content": "Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 454, + 526, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 526, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 526, + 532 + ], + "type": "text", + "content": "Nazneen Fatema Rajani, Bryan McCann, Caiming Xiong, and Richard Socher. 2019. Explain yourself! leveraging language models for commonsense reasoning. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4932-4942, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 542, + 526, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 526, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 526, + 597 + ], + "type": "text", + "content": "Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In *Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems*, pages 1-7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 608, + 526, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 526, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 526, + 663 + ], + "type": "text", + "content": "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In AAAI spring symposium: logical formalizations of commonsense reasoning, pages 90-95." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "spans": [ + { + "bbox": [ + 304, + 673, + 526, + 729 + ], + "type": "text", + "content": "Anna Rogers, Olga Kovaleva, Matthew Downey, and Anna Rumshisky. 2020. Getting closer to ai complete question answering: A set of prerequisite real tasks. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05):8722-8731." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 738, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 738, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 738, + 526, + 772 + ], + "type": "text", + "content": "Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2021. Learning to retrieve prompts for in-context learning. arXiv preprint arXiv:2112.08633." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12298" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 160 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 160 + ], + "type": "text", + "content": "Amrita Saha, Rahul Aralikatte, Mitesh M. Khapra, and Karthik Sankaranarayanan. 2018. DuoRC: Towards complex language understanding with paraphrased reading comprehension. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683-1693, Melbourne, Australia. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 170, + 291, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 170, + 291, + 215 + ], + "spans": [ + { + "bbox": [ + 69, + 170, + 291, + 215 + ], + "type": "text", + "content": "Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64(9):99-106." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 225, + 290, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 225, + 290, + 290 + ], + "spans": [ + { + "bbox": [ + 69, + 225, + 290, + 290 + ], + "type": "text", + "content": "Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 301, + 290, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 301, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 69, + 301, + 290, + 401 + ], + "type": "text", + "content": "Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019. Social IQa: Commonsense reasoning about social interactions. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4463-4473, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "text", + "content": "Thomas Scialom, Tuhin Chakrabarty, and Smaranda Muresan. 2022. Continual-t0: Progressively instructing " + }, + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "inline_equation", + "content": "50+" + }, + { + "bbox": [ + 69, + 411, + 290, + 455 + ], + "type": "text", + "content": " tasks to language models without forgetting. arXiv preprint arXiv:2205.12393." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 465, + 290, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 465, + 290, + 499 + ], + "spans": [ + { + "bbox": [ + 69, + 465, + 290, + 499 + ], + "type": "text", + "content": "Weijia Shi, Julian Michael, Suchin Gururangan, and Luke Zettlemoyer. 2022. Nearest neighbor zero-shot inference. arXiv preprint arXiv:2205.13792." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 509, + 290, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 509, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 69, + 509, + 290, + 585 + ], + "type": "text", + "content": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 597, + 291, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 597, + 291, + 707 + ], + "spans": [ + { + "bbox": [ + 69, + 597, + 291, + 707 + ], + "type": "text", + "content": "Yusheng Su, Xiaozhi Wang, Yujia Qin, Chi-Min Chan, Yankai Lin, Huadong Wang, Kaiyue Wen, Zhiyuan Liu, Peng Li, Juanzi Li, Lei Hou, Maosong Sun, and Jie Zhou. 2022. On transferability of prompt tuning for natural language processing. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3949-3969, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 716, + 291, + 772 + ], + "type": "text", + "content": "Niket Tandon, Bhavana Dalvi, Keisuke Sakaguchi, Peter Clark, and Antoine Bosselut. 2019. WIQA: A dataset for \"what if...\" reasoning over procedural text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 116 + ], + "type": "text", + "content": "9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 6076-6085, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "type": "text", + "content": "Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou', and Daniel Cer. 2022. SPoT: Better frozen model adaptation through soft prompt transfer. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5039-5059, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 214, + 525, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 525, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 525, + 302 + ], + "type": "text", + "content": "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 312, + 525, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 525, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 525, + 378 + ], + "type": "text", + "content": "Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. 2022a. What language model architecture and pretraining objective work best for zero-shot generalization? arXiv preprint arXiv:2204.05832." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 525, + 444 + ], + "type": "text", + "content": "Xinyi Wang, Wanrong Zhu, and William Yang Wang. 2023. Large language models are implicitly topic models: Explaining and finding good demonstrations for in-context learning. arXiv preprint arXiv:2301.11916." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 454, + 525, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 525, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 525, + 531 + ], + "type": "text", + "content": "Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, et al. 2022b. Benchmarking generalization via in-context instructions on 1,600+ language tasks. arXiv preprint arXiv:2204.07705." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 542, + 525, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 525, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 525, + 608 + ], + "type": "text", + "content": "Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. 2022c. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 525, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 525, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 525, + 652 + ], + "type": "text", + "content": "Albert Webson and Ellie Pavlick. 2021. Do prompt-based models really understand the meaning of their prompts? arXiv preprint arXiv:2109.01247." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 662, + 525, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 525, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 525, + 717 + ], + "type": "text", + "content": "Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2021. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "text", + "content": "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12299" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Johannes Welbl, Nelson F. Liu, and Matt Gardner. 2017. Crowdsourcing multiple choice science questions. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 94-106, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 140, + 291, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 140, + 291, + 195 + ], + "spans": [ + { + "bbox": [ + 69, + 140, + 291, + 195 + ], + "type": "text", + "content": "Johannes Welbl, Pontus Stenetorp, and Sebastian Riedel. 2018. Constructing datasets for multi-hop reading comprehension across documents. Transactions of the Association for Computational Linguistics, 6:287-302." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 208, + 290, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 208, + 290, + 252 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 290, + 252 + ], + "type": "text", + "content": "Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2021. An explanation of in-context learning as implicit bayesian inference. arXiv preprint arXiv:2111.02080." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 265, + 290, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 265, + 290, + 320 + ], + "spans": [ + { + "bbox": [ + 69, + 265, + 290, + 320 + ], + "type": "text", + "content": "Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. 2022. Zero-prompt: Scaling prompt-based pretraining to 1,000 tasks improves zero-shot generalization. arXiv preprint arXiv:2201.06910." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 333, + 290, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 333, + 290, + 399 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 290, + 399 + ], + "type": "text", + "content": "Yi Yang, Wen-tau Yih, and Christopher Meek. 2015. WikiQA: A challenge dataset for open-domain question answering. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2013-2018, Lisbon, Portugal. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 412, + 290, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 412, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 412, + 290, + 456 + ], + "type": "text", + "content": "Seonghyeon Ye, Doyoung Kim, Joel Jang, Joongbo Shin, and Minjoon Seo. 2022. Guess the instruction! making language models stronger zero-shot learners. arXiv preprint arXiv:2210.02969." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 469, + 290, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 469, + 290, + 535 + ], + "spans": [ + { + "bbox": [ + 69, + 469, + 290, + 535 + ], + "type": "text", + "content": "Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. 2019. HellaSwag: Can a machine really finish your sentence? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4791-4800, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 548, + 290, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 548, + 290, + 592 + ], + "spans": [ + { + "bbox": [ + 69, + 548, + 290, + 592 + ], + "type": "text", + "content": "Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015a. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 604, + 290, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 604, + 290, + 671 + ], + "spans": [ + { + "bbox": [ + 69, + 604, + 290, + 671 + ], + "type": "text", + "content": "Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015b. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 649-657." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 290, + 772 + ], + "type": "text", + "content": "Yuan Zhang, Jason Baldridge, and Luheng He. 2019. PAWS: Paraphrase adversaries from word scrambling. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 1298-1308, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 266 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 126 + ], + "type": "text", + "content": "Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 136, + 524, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 524, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 524, + 213 + ], + "type": "text", + "content": "Ruiqi Zhong, Kristy Lee, Zheng Zhang, and Dan Klein. 2021. Adapting language models for zero-shot learning by meta-tuning on dataset and prompt collections. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2856-2878, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 221, + 524, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 524, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 524, + 266 + ], + "type": "text", + "content": "Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large language models are human-level prompt engineers." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "12300" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 253, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 253, + 100 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 253, + 100 + ], + "type": "text", + "content": "A Comparison with Hard Prompt Optimization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 105, + 291, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 105, + 291, + 375 + ], + "spans": [ + { + "bbox": [ + 69, + 105, + 291, + 375 + ], + "type": "text", + "content": "We have conducted additional experiments to compare our method with 3 different hard prompt optimization techniques including (1) RoHPr which utilizes the same search technique as RoSPr but retrieves 4 corresponding training instances for incontext learning instead of the corresponding soft prompt, and (2) APE (Zhou et al., 2023) which is an automatic prompt engineering method that utilizes the generation results of an LLM (Ouyang et al., 2022), known to show better performance than human instructions, and (3) ZPS (Liao et al., 2022) which selects an optimal hard prompt from a prompt pool using prompt ensemble and pseudolabels. As shown in the result of Table 2, on average, RoSPr performs the best on average, showing the benefits of using soft prompts over hard prompts for task generalization. While RoSPr shows improvement on 10 out of 11 datasets, other hard prompt optimization methods do not show consistent improvements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 385, + 279, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 385, + 279, + 397 + ], + "spans": [ + { + "bbox": [ + 67, + 385, + 279, + 397 + ], + "type": "text", + "content": "B Full Result of BIG-bench Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "spans": [ + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "content": "We provide the task generalization performance result of 14 tasks from BIG-bench, shown in Table 3. Applying RoSPR largely improves the performance for 3 datasets (Hindu Knowledge, Novel Concepts, Misconceptions): " + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "inline_equation", + "content": "(+17.72\\%, +3.12\\%, +1.82\\%)" + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "content": " compared to T0-3B and " + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "inline_equation", + "content": "(+13.72\\%, +3.12\\%, +2.05\\%)" + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "content": " compared to T0-11B. For mean accuracy of 14 tasks, T0+RoSPR outperforms T0-3B by " + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "inline_equation", + "content": "2.39\\%" + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "content": " points, reducing the gap between 4 times larger T0-11B to " + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "inline_equation", + "content": "1.84\\%" + }, + { + "bbox": [ + 67, + 406, + 291, + 594 + ], + "type": "text", + "content": " points. Moreover, applying INTER to T0+RoSPR enhances the performance of T0+RoSPR for most tasks, indicating that interpolation of multiple embeddings is effective for challenging tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 603, + 235, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 603, + 235, + 618 + ], + "spans": [ + { + "bbox": [ + 67, + 603, + 235, + 618 + ], + "type": "text", + "content": "C Fine-tuning Baseline Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 773 + ], + "type": "text", + "content": "For fine-tuning baseline models (PT and ATTEMPT), we follow the training configuration of source prompt tuning. We train each target prompt with a single epoch using a maximum of 5,000 training instances. For fine-tuning, we use a batch size of 32 and a learning rate of 1e-3. Also, we randomly select hard prompts (templates) of the training dataset during fine-tuning. For ATTEMPT, we randomly sample one source prompt per task cluster, resulting in interpolation between 8 soft prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 70, + 470, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 470, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 470, + 84 + ], + "type": "text", + "content": "D Additional Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 338, + 103, + 490, + 243 + ], + "blocks": [ + { + "bbox": [ + 338, + 103, + 490, + 243 + ], + "lines": [ + { + "bbox": [ + 338, + 103, + 490, + 243 + ], + "spans": [ + { + "bbox": [ + 338, + 103, + 490, + 243 + ], + "type": "image", + "image_path": "7c6ab7b6dea7392b5cf3dd29e262296a2f1dd12b367774760af9b83d72bb884f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 253, + 525, + 280 + ], + "lines": [ + { + "bbox": [ + 302, + 253, + 525, + 280 + ], + "spans": [ + { + "bbox": [ + 302, + 253, + 525, + 280 + ], + "type": "text", + "content": "Figure 8: Variation of number of prompts by increasing the number of prompts per dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 339, + 305, + 489, + 446 + ], + "blocks": [ + { + "bbox": [ + 339, + 305, + 489, + 446 + ], + "lines": [ + { + "bbox": [ + 339, + 305, + 489, + 446 + ], + "spans": [ + { + "bbox": [ + 339, + 305, + 489, + 446 + ], + "type": "image", + "image_path": "3a71c461961e3e29507a80cc7186c63a26d336cb5ea8305c590954fc1f19f36b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 456, + 525, + 481 + ], + "lines": [ + { + "bbox": [ + 302, + 456, + 525, + 481 + ], + "spans": [ + { + "bbox": [ + 302, + 456, + 525, + 481 + ], + "type": "text", + "content": "Figure 9: Variation of number of datasets by increasing the number of datasets per task cluster." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 498, + 525, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 498, + 525, + 673 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 525, + 673 + ], + "type": "text", + "content": "We provide detailed result of variation of the number of prompts (Figure 8) and the number of datasets (Figure 9). We additionally analyze the effect of (1) different sampling methods for constructing the Source Prompt Library, (2) the number of instances sampled for constructing the Source Prompt Library, (3) the number of top-N retrieval for embedding retrieval, and (4) the number of multiple source embeddings to interpolate. Same as Section 7, we report the mean accuracy of 4 evaluation datasets: RTE, COPA, Winogrande, and WiC with 3 runs with different random seeds for the sampling of evaluation queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 687, + 511, + 714 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 687, + 511, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 687, + 511, + 714 + ], + "type": "text", + "content": "D.1 Sampling Methods for Source Prompt Library" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "We experiment three different methods to sample instances for constructing Source Prompt Library and analyze the effect of each method. By default, we choose RANDOM method, where we sample" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "12301" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 68, + 527, + 138 + ], + "blocks": [ + { + "bbox": [ + 71, + 68, + 527, + 138 + ], + "lines": [ + { + "bbox": [ + 71, + 68, + 527, + 138 + ], + "spans": [ + { + "bbox": [ + 71, + 68, + 527, + 138 + ], + "type": "table", + "html": "
RTECBAN.R1AN.R2AN.R3COPAHellasw.StoryC.Winogr.WSCWiCAVG
T064.5545.3633.8133.1133.3375.8826.6084.0350.9765.1050.6951.22
RoHPr47.1045.0033.5333.3033.0973.3826.6586.9350.5164.4250.0049.45
APE (Zhou et al., 2023)68.1948.3335.8235.3734.2768.0025.7678.7350.4358.7550.8550.41
ZPS (Liao et al., 2022)58.4860.7136.6034.4033.3376.0028.4987.3951.7864.4250.6352.93
RoSPr71.5449.4837.0534.6433.9278.7526.9785.5251.5064.5251.7653.24
", + "image_path": "41b1fe2d2f5050236c422b05d0f62242a4d46df7221f212b558cfba016a35fc9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 177, + 285, + 386 + ], + "blocks": [ + { + "bbox": [ + 116, + 145, + 476, + 158 + ], + "lines": [ + { + "bbox": [ + 116, + 145, + 476, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 476, + 158 + ], + "type": "text", + "content": "Table 2: Comparison result of RoSPr with different hard prompt optimization techniques." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 177, + 285, + 386 + ], + "lines": [ + { + "bbox": [ + 74, + 177, + 285, + 386 + ], + "spans": [ + { + "bbox": [ + 74, + 177, + 285, + 386 + ], + "type": "table", + "html": "
T0-3BROSPRINTERT0-11B
Strategy.52.7952.0552.2352.75
Movie D.52.8551.4552.2353.69
Known Un.47.8347.8347.8358.70
Logic Grid41.1037.0037.4038.30
Hindu Kn.25.7143.4345.7129.71
Code D.46.6745.0040.0043.33
Concept45.5267.6167.6169.29
Language14.8413.6814.4020.20
Vitamin58.8853.7154.5364.73
Syllogism52.9450.6451.3451.81
Misconcept.50.2352.0552.0550.00
Logical46.6454.8654.8654.86
Winowhy44.2944.3344.2952.11
Novel Con.15.6315.6318.7515.63
AVG42.5644.9545.2346.79
", + "image_path": "29b67b33e7c0060b76579ce4e0d640616762d28fb771d14a43881287aef50a97.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 394, + 290, + 417 + ], + "lines": [ + { + "bbox": [ + 67, + 394, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 394, + 290, + 417 + ], + "type": "text", + "content": "Table 3: Evaluation result of 14 tasks from BIG-bench (Srivastava et al., 2022)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 431, + 287, + 528 + ], + "blocks": [ + { + "bbox": [ + 70, + 431, + 287, + 528 + ], + "lines": [ + { + "bbox": [ + 70, + 431, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 70, + 431, + 287, + 528 + ], + "type": "image", + "image_path": "a4fe1ea67a1f90c1d4b419df853c5a552f683205a91e522c4d57494a989f37af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 535, + 291, + 560 + ], + "lines": [ + { + "bbox": [ + 67, + 535, + 291, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 291, + 560 + ], + "type": "text", + "content": "Figure 10: Different instance sampling methods for constructing Source Prompt Library." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "text", + "content": "100 instances by random for each prompt by assuming that each random 100 queries of instances can represent the whole prompt. Second, we experiment CLUSTERING method, where we sample top 100 instances which has the closest mean representation from its overall average for each prompt. If we say each instance has a distance of mean representation from its overall average as " + }, + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "text", + "content": " in increasing order, we sample " + }, + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "inline_equation", + "content": "\\{d_1,d_2,\\dots,d_{100}\\}" + }, + { + "bbox": [ + 67, + 584, + 292, + 772 + ], + "type": "text", + "content": ". The last method we use is DISTRIBUTed method, where we sample 100 instances in a distributed way with respect to its distance of mean representation from its overall average. If we say each instance has a distance of mean representation from its over" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "text", + "content": "all average as " + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "text", + "content": " in increasing order, we sample " + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "inline_equation", + "content": "\\{d_1,d_{1 + N / 100},d_{1 + 2*N / 100},\\dots ,d_{1 + 99*N / 100}\\}" + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "text", + "content": ", assuming there are total " + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 179, + 526, + 232 + ], + "type": "text", + "content": " training instances in a dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 234, + 527, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 234, + 527, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 234, + 527, + 383 + ], + "type": "text", + "content": "As shown in Figure 10, RANDOM method outperforms CLUSTERING and DISTRIBUTED methods. Interestingly, CLUSTERING method significantly hurts the performance on all 4 proposed methods, suggesting that storing similar instances per prompt results in retrieval failures more often. Also for DISTRIBUTED method, most of the methods significantly underperform RANDOM, except INTER and ROSPR+VAR+INTER. From these results, we can conclude that random sampling represents the source dataset most effectively." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 338, + 395, + 490, + 535 + ], + "blocks": [ + { + "bbox": [ + 338, + 395, + 490, + 535 + ], + "lines": [ + { + "bbox": [ + 338, + 395, + 490, + 535 + ], + "spans": [ + { + "bbox": [ + 338, + 395, + 490, + 535 + ], + "type": "image", + "image_path": "771e75538b33e4966cec3169ff4caa82aaf3409a9384edf93e1349e8c0eff70d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 544, + 525, + 580 + ], + "lines": [ + { + "bbox": [ + 302, + 544, + 525, + 580 + ], + "spans": [ + { + "bbox": [ + 302, + 544, + 525, + 580 + ], + "type": "text", + "content": "Figure 11: Variation of number of instances sampled for constructing Source Prompt Library. Default setting is " + }, + { + "bbox": [ + 302, + 544, + 525, + 580 + ], + "type": "inline_equation", + "content": "n = 100" + }, + { + "bbox": [ + 302, + 544, + 525, + 580 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 606, + 506, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 606, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 302, + 606, + 506, + 634 + ], + "type": "text", + "content": "D.2 Number of Instances Sampled for Constructing Source Prompt Library" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "text", + "content": "We analyze the effect of size of the Source Prompt Library by varying the number of instances " + }, + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "text", + "content": " to sample for each hard prompt by 100, 300, 500. Therefore, " + }, + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "inline_equation", + "content": "n \\times" + }, + { + "bbox": [ + 301, + 637, + 526, + 773 + ], + "type": "text", + "content": " (number of total hard prompts) would be the size of the Source Prompt Library. As shown in Figure 11, increasing the number of sampled instances does not increase the performance; it hurts the performance for most cases. This suggests that only a few number of training instances are enough to represent the distribution of prompted" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12302" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 71, + 258, + 211 + ], + "blocks": [ + { + "bbox": [ + 104, + 71, + 258, + 211 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 258, + 211 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 258, + 211 + ], + "type": "image", + "image_path": "8925a804240b359526e812738a5b7a830639cc6bf88c8f1706c4e4f2b9aa6341.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "lines": [ + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "type": "text", + "content": "Figure 12: Variation of number of Top " + }, + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "type": "text", + "content": " instances for embedding retrieval. Default setting is " + }, + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 67, + 221, + 290, + 247 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 268, + 291, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 268, + 291, + 377 + ], + "spans": [ + { + "bbox": [ + 67, + 268, + 291, + 377 + ], + "type": "text", + "content": "input (hard prompt + input instances) for each hard prompt and increasing the number sometimes hurt the performance by adding noise to the distribution. This also supports the importance of heuristic cues in Section 6 by showing that adding more training instances per hard prompt does not increase the performance. Instead, adding hard prompts with diverse answer choice format is more important." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 387, + 242, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 387, + 242, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 387, + 242, + 415 + ], + "type": "text", + "content": "D.3 Number of Top N instances for Embedding Retrieval" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "spans": [ + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "text", + "content": "We vary the number of top- " + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "text", + "content": " instances that are retrieved given each query through MIPS search. As shown in Figure 12, varying the number of top- " + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "text", + "content": " instances does not have much effect compared to increasing the number of sampled queries (Figure 7b). This implies that if the size of the evaluation set of the target task is large, sampling more queries is effective than searching for more similar instances per query. This is important for variance-based methods because the number of forward passes needed before evaluation is proportional to " + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "inline_equation", + "content": "Q*N" + }, + { + "bbox": [ + 67, + 419, + 291, + 609 + ], + "type": "text", + "content": ". Therefore, we can reduce the time latency by reducing the number of instances retrieved per query without hurting the performance much." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 619, + 261, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 619, + 261, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 619, + 261, + 645 + ], + "type": "text", + "content": "D.4 Number of Source Embeddings for Interpolation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "We analyze the effect of number of source embeddings for interpolation by varying top- " + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "inline_equation", + "content": "N^{\\prime}" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": " interpolation from 1 (no interpolation) to 5 shown in Figure 13. By comparing between single prompt embedding retrieval " + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "inline_equation", + "content": "(N^{\\prime} = 1)" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": " and the interpolation of multiple embeddings " + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "inline_equation", + "content": "(N^{\\prime} > 1)" + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": ", the mean accuracy drops by adding multiple source embeddings for retrieval because interpolation-based methods underperform on tasks such as COPA as shown" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 338, + 69, + 490, + 211 + ], + "blocks": [ + { + "bbox": [ + 338, + 69, + 490, + 211 + ], + "lines": [ + { + "bbox": [ + 338, + 69, + 490, + 211 + ], + "spans": [ + { + "bbox": [ + 338, + 69, + 490, + 211 + ], + "type": "image", + "image_path": "62f18d4144a73705322ce9eb6db64e09537df79b85f9c48ef1656b5ea55cc28a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 221, + 525, + 257 + ], + "lines": [ + { + "bbox": [ + 302, + 221, + 525, + 257 + ], + "spans": [ + { + "bbox": [ + 302, + 221, + 525, + 257 + ], + "type": "text", + "content": "Figure 13: Variation of number of source embeddings for interpolation based methods. Default setting is " + }, + { + "bbox": [ + 302, + 221, + 525, + 257 + ], + "type": "inline_equation", + "content": "K' = 3" + }, + { + "bbox": [ + 302, + 221, + 525, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 279, + 525, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 279, + 525, + 318 + ], + "spans": [ + { + "bbox": [ + 302, + 279, + 525, + 318 + ], + "type": "text", + "content": "in Table 1. Mean accuracy would increase if we add other datasets for evaluation that benefit from interpolation such as WSC and CB." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "spans": [ + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "content": "By comparing among various " + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "content": " values, we find that for ROSPR+INTER, the accuracy substantially decreases for " + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "inline_equation", + "content": "K' = 2" + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "content": ", implying that the possibility of a wrong retrieval varies depending on the value of " + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "inline_equation", + "content": "N'" + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "content": ". In contrast, ROSPR+VAR+INTER is more robust to the value of " + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 302, + 319, + 526, + 440 + ], + "type": "text", + "content": ", showing that variance-based ranking increases robustness to different numbers of source embeddings for interpolation as well." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 450, + 448, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 450, + 448, + 463 + ], + "spans": [ + { + "bbox": [ + 302, + 450, + 448, + 463 + ], + "type": "text", + "content": "E Visualization of Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 471, + 525, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 471, + 525, + 526 + ], + "spans": [ + { + "bbox": [ + 302, + 471, + 525, + 526 + ], + "type": "text", + "content": "We show the visualization of the evaluation result on 11 datasets in Figure 14. Methods based on ROSPR not only show higher accuracy, but lower variance for many datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 535, + 476, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 535, + 476, + 550 + ], + "spans": [ + { + "bbox": [ + 302, + 535, + 476, + 550 + ], + "type": "text", + "content": "F Experimental Configurations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": "As mentioned in the previous sections, we use T0-3B as our backbone meta-trained LM. For prompt tuning, we fix the prefix length as 100 and the embeddings are initialized from 5,000 most common vocabulary following Lester et al. (2021). We train each source embedding for a single epoch with a learning rate of 0.1 and a batch size of 32. We use the Adam optimizer with weight decay of 1e-5. For retrieval, we randomly sample " + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "Q = 32" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": " query instances and retrieve top " + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": " examples for each query. We train a T0-small variant (" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "\\sim 35\\mathrm{M}" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": " params) as our dense retriever by multitask prompted training on T5+LM model (Lester et al., 2021), replicating the original training setting of T0 by training T5+LM for 8 epochs using the same training instances of Sanh et al. (2021) with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12303" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 79, + 170, + 153 + ], + "blocks": [ + { + "bbox": [ + 71, + 79, + 170, + 153 + ], + "lines": [ + { + "bbox": [ + 71, + 79, + 170, + 153 + ], + "spans": [ + { + "bbox": [ + 71, + 79, + 170, + 153 + ], + "type": "image", + "image_path": "f6b681d8de7ac8cdac12bf6b353e78bbe2965768c33bc4d4baf40abff101026f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 163, + 139, + 175 + ], + "lines": [ + { + "bbox": [ + 108, + 163, + 139, + 175 + ], + "spans": [ + { + "bbox": [ + 108, + 163, + 139, + 175 + ], + "type": "text", + "content": "(a) RTE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 79, + 283, + 152 + ], + "blocks": [ + { + "bbox": [ + 189, + 79, + 283, + 152 + ], + "lines": [ + { + "bbox": [ + 189, + 79, + 283, + 152 + ], + "spans": [ + { + "bbox": [ + 189, + 79, + 283, + 152 + ], + "type": "image", + "image_path": "8edc524f24162a48e4833bbdad69b57897357326563c1f4eaea143f56bf8032a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 163, + 252, + 175 + ], + "lines": [ + { + "bbox": [ + 225, + 163, + 252, + 175 + ], + "spans": [ + { + "bbox": [ + 225, + 163, + 252, + 175 + ], + "type": "text", + "content": "(b) CB" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 305, + 79, + 400, + 153 + ], + "blocks": [ + { + "bbox": [ + 305, + 79, + 400, + 153 + ], + "lines": [ + { + "bbox": [ + 305, + 79, + 400, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 79, + 400, + 153 + ], + "type": "image", + "image_path": "9490c4e09c316164a435fde9b295f7e6dfc9926d451292a2424456bfba8aac34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 163, + 372, + 175 + ], + "lines": [ + { + "bbox": [ + 336, + 163, + 372, + 175 + ], + "spans": [ + { + "bbox": [ + 336, + 163, + 372, + 175 + ], + "type": "text", + "content": "(c) ANLI" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 79, + 514, + 149 + ], + "blocks": [ + { + "bbox": [ + 421, + 79, + 514, + 149 + ], + "lines": [ + { + "bbox": [ + 421, + 79, + 514, + 149 + ], + "spans": [ + { + "bbox": [ + 421, + 79, + 514, + 149 + ], + "type": "image", + "image_path": "7d237e9658defe93969abaf5410e969848d5bb88f1f9c16f61f17a5c2304e433.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 163, + 494, + 175 + ], + "lines": [ + { + "bbox": [ + 445, + 163, + 494, + 175 + ], + "spans": [ + { + "bbox": [ + 445, + 163, + 494, + 175 + ], + "type": "text", + "content": "(d) ANLI R2" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 71, + 186, + 168, + 255 + ], + "blocks": [ + { + "bbox": [ + 71, + 186, + 168, + 255 + ], + "lines": [ + { + "bbox": [ + 71, + 186, + 168, + 255 + ], + "spans": [ + { + "bbox": [ + 71, + 186, + 168, + 255 + ], + "type": "image", + "image_path": "94345d78f9ba02784fc5758aa5298af1f471f707f89243831aa1f69ca9103cf2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 270, + 148, + 281 + ], + "lines": [ + { + "bbox": [ + 99, + 270, + 148, + 281 + ], + "spans": [ + { + "bbox": [ + 99, + 270, + 148, + 281 + ], + "type": "text", + "content": "(e) ANLI R3" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 184, + 283, + 258 + ], + "blocks": [ + { + "bbox": [ + 190, + 184, + 283, + 258 + ], + "lines": [ + { + "bbox": [ + 190, + 184, + 283, + 258 + ], + "spans": [ + { + "bbox": [ + 190, + 184, + 283, + 258 + ], + "type": "image", + "image_path": "cd031f39a39d528fdc225d388dc7060ef6c76d6425f60064f8c685701d44b79a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 270, + 257, + 281 + ], + "lines": [ + { + "bbox": [ + 220, + 270, + 257, + 281 + ], + "spans": [ + { + "bbox": [ + 220, + 270, + 257, + 281 + ], + "type": "text", + "content": "(f) COPA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 303, + 184, + 399, + 256 + ], + "blocks": [ + { + "bbox": [ + 303, + 184, + 399, + 256 + ], + "lines": [ + { + "bbox": [ + 303, + 184, + 399, + 256 + ], + "spans": [ + { + "bbox": [ + 303, + 184, + 399, + 256 + ], + "type": "image", + "image_path": "481e804c5f6c417b8768efd7ece1919d29d27c091101a3ea125c3802c04137de.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 270, + 380, + 282 + ], + "lines": [ + { + "bbox": [ + 327, + 270, + 380, + 282 + ], + "spans": [ + { + "bbox": [ + 327, + 270, + 380, + 282 + ], + "type": "text", + "content": "(g) Hellaswag" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 421, + 186, + 514, + 255 + ], + "blocks": [ + { + "bbox": [ + 421, + 186, + 514, + 255 + ], + "lines": [ + { + "bbox": [ + 421, + 186, + 514, + 255 + ], + "spans": [ + { + "bbox": [ + 421, + 186, + 514, + 255 + ], + "type": "image", + "image_path": "bbe34813cb710dde0b1ac3ca17b659e72ce568c5a4a751273e353e0fdbdb0e0f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 141, + 292, + 235, + 361 + ], + "blocks": [ + { + "bbox": [ + 141, + 292, + 235, + 361 + ], + "lines": [ + { + "bbox": [ + 141, + 292, + 235, + 361 + ], + "spans": [ + { + "bbox": [ + 141, + 292, + 235, + 361 + ], + "type": "image", + "image_path": "7173cd83ab347f6daf8a9358825658e1162d4e03c2d2b1a72b54cbe060be00b5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 248, + 292, + 340, + 361 + ], + "blocks": [ + { + "bbox": [ + 248, + 292, + 340, + 361 + ], + "lines": [ + { + "bbox": [ + 248, + 292, + 340, + 361 + ], + "spans": [ + { + "bbox": [ + 248, + 292, + 340, + 361 + ], + "type": "image", + "image_path": "7ba60fd0fd27431dc310a232c0695e5978d4f3914e454d39f98b1802390d3ea8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 354, + 292, + 448, + 361 + ], + "blocks": [ + { + "bbox": [ + 442, + 270, + 497, + 282 + ], + "lines": [ + { + "bbox": [ + 442, + 270, + 497, + 282 + ], + "spans": [ + { + "bbox": [ + 442, + 270, + 497, + 282 + ], + "type": "text", + "content": "(h) StoryCloze" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 292, + 448, + 361 + ], + "lines": [ + { + "bbox": [ + 354, + 292, + 448, + 361 + ], + "spans": [ + { + "bbox": [ + 354, + 292, + 448, + 361 + ], + "type": "image", + "image_path": "7de310496739df397c37b7563fdd412cfe08fce4901f5ce1608dcbcdceebcf4a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 113, + 389, + 478, + 404 + ], + "blocks": [ + { + "bbox": [ + 161, + 376, + 218, + 387 + ], + "lines": [ + { + "bbox": [ + 161, + 376, + 218, + 387 + ], + "spans": [ + { + "bbox": [ + 161, + 376, + 218, + 387 + ], + "type": "text", + "content": "(i) Winogrande" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 376, + 312, + 387 + ], + "lines": [ + { + "bbox": [ + 280, + 376, + 312, + 387 + ], + "spans": [ + { + "bbox": [ + 280, + 376, + 312, + 387 + ], + "type": "text", + "content": "(j) WSC" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 388, + 376, + 420, + 387 + ], + "lines": [ + { + "bbox": [ + 388, + 376, + 420, + 387 + ], + "spans": [ + { + "bbox": [ + 388, + 376, + 420, + 387 + ], + "type": "text", + "content": "(k)WiC" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 113, + 389, + 478, + 404 + ], + "lines": [ + { + "bbox": [ + 113, + 389, + 478, + 404 + ], + "spans": [ + { + "bbox": [ + 113, + 389, + 478, + 404 + ], + "type": "image", + "image_path": "2f1b160c84f15db68dbe6a93ecf996f65a7657a7b825daf242a039ccb2777701.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 174, + 418, + 417, + 430 + ], + "lines": [ + { + "bbox": [ + 174, + 418, + 417, + 430 + ], + "spans": [ + { + "bbox": [ + 174, + 418, + 417, + 430 + ], + "type": "text", + "content": "Figure 14: Visualization of evaluation results of 11 datasets." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "spans": [ + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "text", + "content": "a learning rate of 1e-3, input sequence length 512, output sequence 128, and batch size of 1024. We select model checkpoint by early stopping based on validation accuracy. We use a meta-trained LM instead of a naive pretrained model (e.g. SentenceBERT) because meta-trained LM is shown to be more effective for retrieval (Lin et al., 2022). For the interpolation experiment, we set " + }, + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "inline_equation", + "content": "K' = 3" + }, + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "text", + "content": " for top-" + }, + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 67, + 452, + 291, + 588 + ], + "type": "text", + "content": " prompt embedding candidates. For training source prompt embeddings, we used 8 V100 GPUs." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 67, + 599, + 271, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 599, + 271, + 642 + ], + "spans": [ + { + "bbox": [ + 67, + 599, + 271, + 642 + ], + "type": "text", + "content": "G Examples of Applying Prompts, Answer Choice Format and Source Task Types" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 67, + 650, + 291, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 650, + 291, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 650, + 291, + 690 + ], + "type": "text", + "content": "Figure 15 shows an example of applying prompt through Promptsource (Bach et al., 2022) as mentioned in Section 3.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "We assert that answer choice format is more important than task similarity in Section 6. We further provide details of the input instances of the mentioned tasks: paraphrase, NLI, word sense disambiguation, and sentiment classification in Figure 16. As supported in Pruksachatkun et al. (2020)," + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 302, + 452, + 526, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 452, + 526, + 561 + ], + "spans": [ + { + "bbox": [ + 302, + 452, + 526, + 561 + ], + "type": "text", + "content": "intuitively, paraphrase task is more similar to the task of word sense disambiguation task or NLI, implying their task similarity, while the task of sentiment classification is very different. However, our counterintuitive result shows the soft prompt to show the best performance in Figure 6 Section 6, bolstering the claim that similar source task types are not a major factor for evaluation performance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 302, + 571, + 492, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 571, + 492, + 599 + ], + "spans": [ + { + "bbox": [ + 302, + 571, + 492, + 599 + ], + "type": "text", + "content": "H Full List of Source Training and Evaluation Datasets" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 302, + 608, + 525, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 608, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 608, + 525, + 663 + ], + "type": "text", + "content": "All of our training and evaluation datasets are a subset of datasets used in Sanh et al. (2021). We use Huggingface version of each dataset (Lhoest et al., 2021)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 302, + 673, + 417, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 673, + 417, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 673, + 417, + 687 + ], + "type": "text", + "content": "H.1 Training Datasets" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 526, + 773 + ], + "type": "text", + "content": "Following Sanh et al. (2021), we use 8 task clusters for training of source prompt embedding: sentiment classification, paraphrase, topic classification, summarization, struc-to-text, multiple-choice QA, extractive QA, and closed book QA. We use imdb (Maas et al., 2011), amazon_polarity" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12304" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 137, + 147, + 454, + 283 + ], + "blocks": [ + { + "bbox": [ + 137, + 147, + 454, + 283 + ], + "lines": [ + { + "bbox": [ + 137, + 147, + 454, + 283 + ], + "spans": [ + { + "bbox": [ + 137, + 147, + 454, + 283 + ], + "type": "image", + "image_path": "c50949b1eb4152b355b40437041c77e0619ed61eb9b9188201d943658e560a22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 291, + 500, + 304 + ], + "lines": [ + { + "bbox": [ + 91, + 291, + 500, + 304 + ], + "spans": [ + { + "bbox": [ + 91, + 291, + 500, + 304 + ], + "type": "text", + "content": "Figure 15: Example of applying prompt to a given instance through Promptsource (Bach et al., 2022)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 91, + 459, + 505, + 676 + ], + "blocks": [ + { + "bbox": [ + 91, + 459, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 91, + 459, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 91, + 459, + 505, + 676 + ], + "type": "image", + "image_path": "7e0c78c2453749ae0067e28a53ad919ec36b7398bd2909971dd6c3f175995f20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 682, + 415, + 695 + ], + "lines": [ + { + "bbox": [ + 177, + 682, + 415, + 695 + ], + "spans": [ + { + "bbox": [ + 177, + 682, + 415, + 695 + ], + "type": "text", + "content": "Figure 16: Examples of instances of different source tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12305" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 292, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 292, + 368 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 292, + 368 + ], + "type": "text", + "content": "(McAuley and Leskovec, 2013), rotten_tomatoes (Pang and Lee, 2005), yelp_review_full (Zhang et al., 2015b) for sentiment, glue/qqp (Wang et al., 2018), paws/labeled_final (Zhang et al., 2019) for paraphrase, ag_news (Zhang et al., 2015a), dbpedia_14 (Lehmann et al., 2015) for topic classification, gigaword (Graff et al., 2003), multi_news citefabbri-etal-2019-multi, samsum (Gliwa et al., 2019), xsum (Narayan et al., 2018) for summarization, common_gen (Lin et al., 2020), wiki.bio (Lebret et al., 2016) for struct-to-text, cos_e/v1.11 (Rajani et al., 2019), quail (Rogers et al., 2020), social_i_qa (Sap et al., 2019), wiqa (Tandon et al., 2019), cosmos_qa (Huang et al., 2019), sciq (Welbl et al., 2017), wiki_hop/original (Welbl et al., 2018) for multi-choice QA, adversarial_qa/dbidaf, adversarial_qa/dbert, adversarial_qa/droberta, quoref (Bartolo et al., 2020), ropes (Lin et al., 2019), duorc/SelfRC, duorc/Paraphrase IdentificationRC (Saha et al., 2018) for extractive QA, and kilt_tasks/hotpotqa (Petroni et al., 2021), wiki_qa (Yang et al., 2015) for closed book QA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 374, + 291, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 374, + 291, + 483 + ], + "spans": [ + { + "bbox": [ + 67, + 374, + 291, + 483 + ], + "type": "text", + "content": "We exclude 6 datasets (MRPC, TREC, DREAM, QuaRTz, QASC, QuaRel) that have small training sets because it leads to task imbalance, which is critical for training our small dense retriever (" + }, + { + "bbox": [ + 67, + 374, + 291, + 483 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 67, + 374, + 291, + 483 + ], + "type": "text", + "content": "35M params). We also exclude CNN Daily Mail, App Reviews, and WikiQA dataset due to dataset download issues, absence of any test or validation data, and unbalanced label distribution, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 510, + 191, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 510, + 191, + 522 + ], + "spans": [ + { + "bbox": [ + 68, + 510, + 191, + 522 + ], + "type": "text", + "content": "H.2 Evaluation Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 538, + 291, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 291, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 291, + 687 + ], + "type": "text", + "content": "Following Sanh et al. (2021), we include 11 evaluation datasets as follows: RTE (Dagan et al., 2005), CB (De Marneffe et al., 2019), ANLI (Nie et al., 2020) for natural language inference task, COPA (Roemmele et al., 2011), Hellaswag (Zellers et al., 2019), Storycloze (Mostafazadeh et al., 2016) for sentence completion task, Winogrande (Sakaguchi et al., 2021), WSC (Levesque et al., 2012) for coreference resolution task, and WiC (Pilehvar and Camacho-Collados, 2019) for word sense disambiguation task." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "For BIG-bench tasks, we evaluate on 14 tasks, following Sanh et al. (2021): Known Unknown, Logic Grid, StrategyQA, Hindu Knowledge, Movie Dialog, Code Description, Conceptual, Language ID, Vitamin C, Syllogisms, Misconceptions, Logical Deduction, Winowhy and Novel Concepts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 70, + 474, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 70, + 474, + 99 + ], + "spans": [ + { + "bbox": [ + 303, + 70, + 474, + 99 + ], + "type": "text", + "content": "I Full List of Retrieved Prompt Embeddings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 105, + 526, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 105, + 526, + 160 + ], + "spans": [ + { + "bbox": [ + 302, + 105, + 526, + 160 + ], + "type": "text", + "content": "We provide a full list of retrieved prompt embeddings of ROSPR and ORACLE for all prompts of 11 evaluation datasets. We report retrieval results of a single random seed (Table " + }, + { + "bbox": [ + 302, + 105, + 526, + 160 + ], + "type": "inline_equation", + "content": "4\\sim" + }, + { + "bbox": [ + 302, + 105, + 526, + 160 + ], + "type": "text", + "content": " Table 14)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12306" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 73, + 470, + 184 + ], + "blocks": [ + { + "bbox": [ + 69, + 73, + 470, + 184 + ], + "lines": [ + { + "bbox": [ + 69, + 73, + 470, + 184 + ], + "spans": [ + { + "bbox": [ + 69, + 73, + 470, + 184 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
GPT-3 style61.3774.01paws/labeled_final/PAWS-ANLI GPT374.01paws/labeled_final/context-question
MNLI crowdsourced63.5370.04paws/labeled_final/context-question72.20paws/labeled_final/context-question
based on the previous passage68.2376.53paws/labeled_final/context-question76.53paws/labeled_final/PAWS-ANLI GPT3
can we infer59.5773.29glue/qqp/meaning73.29paws/labeled_final/PAWS-ANLI GPT3
does it follow that61.7371.84paws/labeled_final/context-question71.84paws/labeled_final/context-question-no-label
does this imply64.6268.59paws/labeled_final/context-question71.48paws/labeled_final/context-question
guaranteed true68.9575.09paws/labeled_final-context-question75.81paws/labeled_final/PAWS-ANLI GPT3
should assume66.4371.12glue/qqp/meaning76.53paws/labeled_final/PAWS-ANLI GPT3
justified in saying61.0158.12paws/labeled_final/paraphrase-task71.12paws/labeled_final/context-question
must be true70.0474.37paws/labeled_final/context-question75.09paws/labeled_final/PAWS-ANLI GPT3-no-label
Avg.64.5571.3073.79
", + "image_path": "6889feb02c7342f0e1037fd18ccb7d8c882011b4709fe9a97e3fcd648bc77858.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 220, + 499, + 370 + ], + "blocks": [ + { + "bbox": [ + 89, + 191, + 501, + 204 + ], + "lines": [ + { + "bbox": [ + 89, + 191, + 501, + 204 + ], + "spans": [ + { + "bbox": [ + 89, + 191, + 501, + 204 + ], + "type": "text", + "content": "Table 4: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of RTE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 220, + 499, + 370 + ], + "lines": [ + { + "bbox": [ + 69, + 220, + 499, + 370 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 499, + 370 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer55.3651.79social_i_qa/Show choices and generate index67.86samsum/Write a dialogue that match this summary
based on the previous passage44.6458.93social_i_qa/Show choices and generate index69.64samsum/Write a dialogue that match this summary
claim true/false/inconclusive50.0067.86social_i_qa/Show choices and generate index69.64samsum/Summarize:
does it follow that64.2950.00social_i_qa/Show choices and generate index67.86samsum/Summarize:
justified in saying53.5750.00social_i_qa/Show choices and generate index62.50samsum/Write a dialogue that match this summary
always/sometimes/never39.2941.07social_i_qa/Show choices and generate index41.07social_i_qa/Show choices and generate index
GPT-3 style51.7967.86social_i_qa/Show choices and generate index69.64social_i_qa/Show choices and generate answer
consider always/sometimes/never35.7135.71social_i_qa/Show choices and generate index39.29social_i_qa/Generate answer
guaranteed true48.2150.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
must be true53.5750.00social_i_qa/Show choices and generate index64.29social_i_qa/Generate answer
guaranteed/possible/impossible8.938.93social_i_qa/Show choices and generate index8.93-(all same)
does this imply58.9351.79social_i_qa/Show choices and generate index66.07glue/qqp/duplicate
MNLI crowdsourced8.9339.29social_i_qa/Show choices and generate index42.86cos_e/v1.11/question_option_description_text
should assume57.1450.00social_i_qa/Show choices and generate index66.07cos_e/v1.11/aligned_with_common Sense
take the following as truth50.0067.86social_i_qa/Show choices and generate index71.43cos_e/v1.11/description_question_option_id
Avg.45.3649.4058.10
", + "image_path": "a673f8bf16cbfdeef8e9366728aa5a91936f54aa8bf28a6b939feea55c69f890.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 407, + 530, + 556 + ], + "blocks": [ + { + "bbox": [ + 92, + 378, + 499, + 391 + ], + "lines": [ + { + "bbox": [ + 92, + 378, + 499, + 391 + ], + "spans": [ + { + "bbox": [ + 92, + 378, + 499, + 391 + ], + "type": "text", + "content": "Table 5: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of CB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 407, + 530, + 556 + ], + "lines": [ + { + "bbox": [ + 69, + 407, + 530, + 556 + ], + "spans": [ + { + "bbox": [ + 69, + 407, + 530, + 556 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.9038.90paws/labeled_final/context-question39.40paws/labeled_final/PAWS-ANLI GPT3
based on the previous passage33.9038.50paws/labeled_final/context-question38.60paws/labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive35.6036.70paws/labeled_final/PAWS-ANLI GPT339.10paws/labeled_final/context-question
does it follow that36.0040.50paws/labeled_final/context-question40.50paws/labeled_final/PAWS-ANLI GPT3
justified in saying33.1038.10paws/labeled_final-context-question38.80paws/labeled_final/context-question-no-label
always/sometimes/never33.4033.40paws/labeled_final/paragraph-task33.40-(all 33.4)
GPT-3 style33.8037.30paws/labeled_final/PAWS-ANLI GPT338.50paws/labeled_final/PAWS-ANLI GPT3-no-label
consider always/sometimes/never33.2033.40paws/labeled_final/PAWS-ANLI GPT333.50-(all 33.4)
guaranteed true33.7038.50paws/labeled_final/context-question38.70paws/labeled_final/PAWS-ANLI GPT3
must be true34.4039.60paws/labeled_final CONTEXT-question39.70paws/labeled_final-context-question
guaranteed/possible/impossible33.3033.30paws/labeled_final/PAWS-ANLI GPT333.30imdb/Text Expressed Sentiment
does this imply33.6038.20paws/labeled_final contextual-question38.20paws/labeled_final contextual-question-no-label
MNLI crowdsourced33.6033.80paws/labeled_final contextual-question35.40dbpedia_14/given_list WHAT_category DOES_theParagraph_belong_to
should assume33.2038.80paws/labeled_final contextual-question39.00paws/labeled_final/PAWS-ANLI GPT3-no-label
take the following as truth32.4037.10paws/labeled_final/PAWS-ANLI GPT338.70paws/labeled_final contextual-question-no-label
Avg.33.8137.0737.65
", + "image_path": "bb06910fd9d02e5f41c4072a5eca0b888792e4bc193beb7b3b345c9da1c2bd9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 69, + 594, + 491, + 743 + ], + "blocks": [ + { + "bbox": [ + 79, + 565, + 511, + 577 + ], + "lines": [ + { + "bbox": [ + 79, + 565, + 511, + 577 + ], + "spans": [ + { + "bbox": [ + 79, + 565, + 511, + 577 + ], + "type": "text", + "content": "Table 6: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R1." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 594, + 491, + 743 + ], + "lines": [ + { + "bbox": [ + 69, + 594, + 491, + 743 + ], + "spans": [ + { + "bbox": [ + 69, + 594, + 491, + 743 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer30.4035.30paws/ labeled_final/context-question35.30paws/ labeled_final/PAWS-ANLI GPT3
based on the previous passage31.4035.40paws/ labeled_final/context-question35.40paws/ labeled_final/PAWS-ANLI GPT3
claim true/false/inconclusive34.9035.10paws/ labeled_final/PAWS-ANLI GPT335.90paws/ labeled_final/context-question
does it follow that34.5036.00paws/ labeled_final/context-question36.00paws/ labeled_final/PAWS-ANLI GPT3
justified in saying33.5035.70paws/ labeled_final/context-question35.70rotten_tomatoes/Reviewer Enjoyment Yes No
always/sometimes/never33.4033.40paws/ labeled_final/paragraph-task33.50adversarial_qa/droberta/based_on,
GPT-3 style33.5034.90paws/ labeled_final/PAWS-ANLI GPT335.00paws/ labeled_final/context-question-no-label
consider always/sometimes/never33.7033.40dbpedia_14/given_a_choice_of categories34.50paws/ labeled_final/PAWS-ANLI GPT3-no-label
guaranteed true32.9034.00paws/ labeled_final/context-question34.30paws/ labeled_final/PAWS-ANLI GPT3
must be true35.1034.60paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
guaranteed/possible/impossible33.3033.30paws/ labeled_final/context-question-no-label33.30imdb/Writer Expressed Sentiment
does this imply32.7033.90paws/ labeled_final/context-question34.10paws/ labeled_final/context-question
MNLI crowdsourced33.4034.70dbpedia_14/given_a_choice_ofcategories34.90dbpedia_14/given_a_choice_ofcategories
should assume32.4035.10paws/ labeled_final/context-question35.10paws/ labeled_final/PAWS-ANLI GPT3
take the following as truth31.6035.70paws/ labeled_final/paragraph-task35.70paws/ labeled_final/PAWS-ANLI GPT3
Avg.33.1134.7034.92
", + "image_path": "7a77a88e13784d8134bd4986d6cc9f6ae11c5ba03fda7fcea5b0cf751a649941.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 751, + 511, + 764 + ], + "lines": [ + { + "bbox": [ + 79, + 751, + 511, + 764 + ], + "spans": [ + { + "bbox": [ + 79, + 751, + 511, + 764 + ], + "type": "text", + "content": "Table 7: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R2." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12307" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 79, + 493, + 230 + ], + "blocks": [ + { + "bbox": [ + 69, + 79, + 493, + 230 + ], + "lines": [ + { + "bbox": [ + 69, + 79, + 493, + 230 + ], + "spans": [ + { + "bbox": [ + 69, + 79, + 493, + 230 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
can we infer33.0034.75glue/qqp/meaning34.75paws/labeled_final/context-question-no-label
based on the previous passage33.3334.08paws/labeled_final/context-question35.33paws/labeled_final/context-question-no-label
claim true/false/inconclusive32.8335.58paws/labeled_final/paraphrase-task35.92cos_e/v1.11/aligned_with_common Sense
does it follow that34.0834.67paws/labeled_final/context-question35.33amazon_polarity/User_recommend.this_product
justified in saying33.5833.00paws/labeled_final/paraphrase-task35.42amazon_polarity/User Recommend.this_product
always/sometimes/never33.4233.42paws/labeled_final/paraphrase-task33.50paws/labeled_final/paraphrase-task
GPT-3 style33.3334.00paws/labeled_final/PAWS-ANLI GPT334.92paws/labeled_final/PAWS-ANLI GPT3
consider always/sometimes/never33.0832.42ropes/plain_nobackground33.67cos_e/v1.11/question_description_option_text
guaranteed true32.5834.08paws/labeled_final/context-question34.83paws/labeled_final/context-question
must be true33.8333.75paws/labeled_final/paraphrase-task35.42rotten_tomatoes/Reviewer Enjoy Yes No
guaranteed/possible/impossible33.5033.50paws/labeled_final/paraphrase-task33.58social_i_qa/Show choices and generate answer
does this imply32.9233.58glue/qqp/meaning34.50paws/labeled_final/context-question
MNLI crowdsourced33.7533.67rotten_tomatoes/Text Expressed Sentiment34.00dbpedia_14/given_a_choice_ofcategories
should assume33.2534.67paws/labeled_final/context-question34.92paws/labeled_final/context-question-no-label
take the following as truth33.4232.75social_i_qa/Show choices and generate index36.67paws/labeled_final/context-question-no-label
Avg.33.3333.8634.91
", + "image_path": "79fba8fb4e872b5d4c25eedf328d02cf77f5c3c550f822adfdeb5f728ca7ea68.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 280, + 511, + 375 + ], + "blocks": [ + { + "bbox": [ + 79, + 238, + 511, + 250 + ], + "lines": [ + { + "bbox": [ + 79, + 238, + 511, + 250 + ], + "spans": [ + { + "bbox": [ + 79, + 238, + 511, + 250 + ], + "type": "text", + "content": "Table 8: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of ANLI R3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 280, + 511, + 375 + ], + "lines": [ + { + "bbox": [ + 69, + 280, + 511, + 375 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 511, + 375 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
exercise80.0079.00cos_e/v1.11/question_option_description_text80.00cos_e/v1.11/question_option_description_text
plausible_alternatives84.0083.00cos_e/v1.11/question_option_description_text83.00cos_e/v1.11/question_option_description_text
"C1 or C2? premise, so/because..."61.0067.00social_i_qa/Check if a random answer is valid or not75.00cos_e/v1.11/descriptionQuestion_option_text
best_option70.0076.00social_i_qa/Show choices and generate answer79.00cos_e/v1.11/question_option_description_text
more likely79.0083.00cos_e/v1.11/question_option_description_text85.00cos_e/v1.11/descriptionQuestion_option_text
cause-effect74.0078.00cos_e/v1.11/question_option_description_text83.00common_gen/random task template prompt
choose80.0082.00cos_e/v1.11/question_option_description_text82.00cosmos_qa/no_prompt_text
i_am_hesitating79.0082.00cos_e/v1.11/question_option_description_text82.00cos_e/v1.11/question_option_description_text
Avg.75.8878.7581.13
", + "image_path": "151bd3efd99a1c7b1e832a20503c80b33827d3652ed5e03b692c5d742a39338b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 424, + 480, + 496 + ], + "blocks": [ + { + "bbox": [ + 86, + 382, + 506, + 395 + ], + "lines": [ + { + "bbox": [ + 86, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 86, + 382, + 506, + 395 + ], + "type": "text", + "content": "Table 9: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of COPA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 424, + 480, + 496 + ], + "lines": [ + { + "bbox": [ + 69, + 424, + 480, + 496 + ], + "spans": [ + { + "bbox": [ + 69, + 424, + 480, + 496 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Predict ending with hint26.8327.70social_i_qa/Show choices and generate index29.11cos_e/v1.11/question_option_description_text
Randomized prompts template26.8726.83social_i_qa/Show choices and generate index27.84wiqa/what_is_the MISSING_first_step
complete_first THEN28.3727.35ropes/prompt_bottom_no_hint28.35cos_e/v1.11/question_option_description_text
if_beginnerscontinue25.2826.58social_i_qa/Show choices and generate index26.58cos_e/v1.11/question_option_description_text
how Ends25.6726.86social_i_qa/Show choices and generate index26.86cos_e/v1.11/question_option_description_text
Avg.26.6027.0627.75
", + "image_path": "a2229f2b60afef3815db3c973515c0ff1fdc90d65d885db4e2a71bab817f9870.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 69, + 544, + 489, + 616 + ], + "blocks": [ + { + "bbox": [ + 75, + 503, + 516, + 517 + ], + "lines": [ + { + "bbox": [ + 75, + 503, + 516, + 517 + ], + "spans": [ + { + "bbox": [ + 75, + 503, + 516, + 517 + ], + "type": "text", + "content": "Table 10: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Hellaswag." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 544, + 489, + 616 + ], + "lines": [ + { + "bbox": [ + 69, + 544, + 489, + 616 + ], + "spans": [ + { + "bbox": [ + 69, + 544, + 489, + 616 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
Answer Given options86.4887.76social_i_qa/Show choices and generate answer87.76social_i_qa/Show choices and generate answer
Choose Story Ending86.9188.24social_i_qa/Show choices and generate answer88.24cos_e/v1.11/question_option_description_text
Movie What Happens Next78.0978.03social_i_qa/Show choices and generate answer87.33cosmos_qa/no_prompt_text
Story Continuation and Options83.5985.89social_i_qa/Show choices and generate answer86.53social_i_qa/Show choices and generate answer
Novel Correct Ending85.0987.65social_i_qa/Show choices and generate answer87.97social_i_qa/Show choices and generate index
Avg.84.0385.5287.57
", + "image_path": "78e3c6fb6dbe3e08100c23bfaa4295246a6c01c2447e4c5e21f4d1e59aa5d944.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 69, + 666, + 467, + 737 + ], + "blocks": [ + { + "bbox": [ + 74, + 624, + 517, + 637 + ], + "lines": [ + { + "bbox": [ + 74, + 624, + 517, + 637 + ], + "spans": [ + { + "bbox": [ + 74, + 624, + 517, + 637 + ], + "type": "text", + "content": "Table 11: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of StoryCloze." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 666, + 467, + 737 + ], + "lines": [ + { + "bbox": [ + 69, + 666, + 467, + 737 + ], + "spans": [ + { + "bbox": [ + 69, + 666, + 467, + 737 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does underscore refer to stand for49.2551.54cos_e/v1.11/question_option_description_text51.78paws/ labeled_final/ PAWS-ANLI GPT3
50.5150.59cos_e/v1.11/question_option_description_text52.09ropes/plain_nobackground
underscore refer to fill in the blank50.4350.59cos_e/v1.11/question_option_description_text52.33ropes/prompt_bottom_no_hint
52.1751.38cos_e/v1.11/question_option_description_text52.01cos_e/v1.11/question_option_description_text
Replace52.4953.59cos_e/v1.11/question_option_description_text53.59paws/ labeled_final/ paraphrase-task
Avg.50.9751.5452.36
", + "image_path": "3117bd0163f6efbc8eb00d222521288cc3ab43c17075ad7f5cb8cea1e6a35b42.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 745, + 520, + 758 + ], + "lines": [ + { + "bbox": [ + 71, + 745, + 520, + 758 + ], + "spans": [ + { + "bbox": [ + 71, + 745, + 520, + 758 + ], + "type": "text", + "content": "Table 12: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of Winogrande." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12308" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 176, + 500, + 286 + ], + "blocks": [ + { + "bbox": [ + 69, + 176, + 500, + 286 + ], + "lines": [ + { + "bbox": [ + 69, + 176, + 500, + 286 + ], + "spans": [ + { + "bbox": [ + 69, + 176, + 500, + 286 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
does the pronoun refer to by p they mean68.2766.34social_i_qa/Check if a random answer is valid or not66.35paws/labeled_final/PAWS-ANLI GPT3
62.5066.35social_i_qa/Check if a random answer is valid or not69.23glue/qqp/duplicate or not
in other words67.3167.31social_i_qa/Show choices and generate answer72.12samsum/Write a dialogue that match this summary
I think they mean69.2368.27social_i_qa/Show choices and generate answer78.08social_i_qa/Generate answer
replaced with64.4259.62social_i_qa/Check if a random answer is valid or not66.35social_i_qa/Show choices and generate index
p is/are r62.5064.42social_i_qa/Show choices and generate index65.38social_i_qa/Show choices and generate index
the pronoun refers to64.4264.42social_i_qa/Show choices and generate index67.31social_i_qa/Check if a random answer is valid or not
Who or what is/are64.4263.46social_i_qa/Show choices and generate index65.38samsum/To sum up this dialog
does p stand for67.3163.46social_i_qa/I was wondering69.23samsum/Write a dialogue that match this summary
GPT-3 Style60.5867.31social_i_qa/Show choices and generate index67.31rotten_tomatoes/Reviewer Expressed Sentiment
Avg.65.1065.1068.17
", + "image_path": "3842c60d39f9d34d30029f5364dd7f7097e4c2f87aa2b157f6450b6572e3bcfd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 531, + 521, + 641 + ], + "blocks": [ + { + "bbox": [ + 85, + 295, + 506, + 307 + ], + "lines": [ + { + "bbox": [ + 85, + 295, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 85, + 295, + 506, + 307 + ], + "type": "text", + "content": "Table 13: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WSC." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 531, + 521, + 641 + ], + "lines": [ + { + "bbox": [ + 69, + 531, + 521, + 641 + ], + "spans": [ + { + "bbox": [ + 69, + 531, + 521, + 641 + ], + "type": "table", + "html": "
Prompt NameT0RoSPrRetrieved EmbeddingOracleRetrieved Embedding
question-context-meaning-with-label50.3153.76glue/qqp/meaning53.76paws/labeled_final/context-question-no-label
question-context-meaning50.6350.16ropes/prompt_bottom_no_hint57.05paws/labeled_final/PAWS-ANLI GPT3
grammar Homework49.8450.16ropes/prompt_bottom_no_hint57.99rotten_tomatoes/Reviewer Enjoyment Yes No
affirmation_true_orfalse49.6951.57paws/labeled_final/Rewrite-no-label53.92paws/labeled_final/Meaning-no-label
GPT-3-prompt51.7250.00social_i_qa/Show choices and generate index55.64paws/labeled_final/PAWS-ANLI GPT3
same_sense49.8452.82paws/labeled_final/Rewrite55.17paws/labeled_final/task_description-no-label
question-context51.8853.29social_i_qa/Check if a random answer is valid or not57.05amazon_polarity/Is.this_product_review_positive
GPT-3-prompt-with-label50.4752.19glue/qqp/meaning52.66glue/qqp/meaning
polysemous50.0052.82paws/labeled_final/Rewrite53.29paws/labeled_final/Rewrite-no-label
similar-sense52.5150.00social_i_qa/Show choices and generate index56.11paws/labeled_final/task_description-no-label
Avg.50.6951.6855.26
", + "image_path": "1de7d1e38a875316b71af801502dcf5a6895fb65e9a7f26419df25755ace72ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 649, + 504, + 661 + ], + "lines": [ + { + "bbox": [ + 86, + 649, + 504, + 661 + ], + "spans": [ + { + "bbox": [ + 86, + 649, + 504, + 661 + ], + "type": "text", + "content": "Table 14: List of retrieved source prompts of ROSPR and ORACLE for each evaluation prompts of WiC." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "12309" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 21 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_content_list.json b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2dea03d0ec901d3efdbf1ec2f19abe57b806614 --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_content_list.json @@ -0,0 +1,1023 @@ +[ + { + "type": "text", + "text": "Emergent Inabilities? Inverse Scaling Over the Course of Pretraining", + "text_level": 1, + "bbox": [ + 139, + 89, + 858, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "James A. Michaelov and Benjamin K. Bergen \nDepartment of Cognitive Science, University of California San Diego \n{j1michae,bkbergen}@ucsd.edu", + "bbox": [ + 218, + 142, + 781, + 193 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT, MEMO TRAP, and PATTERN MATCH SUPPRESSION) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves.", + "bbox": [ + 144, + 279, + 460, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 590, + 258, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For language models, bigger is usually better. Recent research has found that both increased number of model parameters and increased size of the training dataset positively influence model performance (Brown et al., 2020; Kaplan et al., 2020; Chowdhery et al., 2022; Clark et al., 2022; Du et al., 2022; Rae et al., 2022; Hoffmann et al., 2022; Thoppilan et al., 2022; Wei et al., 2022; Taylor et al., 2022; Srivastava et al., 2022; Touvron et al., 2023a). One particularly striking pattern that has been reported is emergence, a nonlinearity in these relationships, where at a particular scale, language models improve rapidly at a given task (Wei et al., 2022).", + "bbox": [ + 115, + 615, + 489, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, while increased scale usually leads to improved performance, on certain tasks it correlates with decreased performance. This is known as inverse scaling (Lin et al., 2022). An example of a task on which inverse scaling is observed", + "bbox": [ + 112, + 825, + 489, + 904 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is the TruthfulQA benchmark, where larger language models are more likely to predict popular misconceptions over statements of fact (Lin et al., 2022). More recently, additional tasks that reportedly show such an effect have been identified as part of the Inverse Scaling Prize (McKenzie et al., 2023b), as well as by other researchers (Jang et al., 2023; Michaelov and Bergen, 2023).", + "bbox": [ + 507, + 252, + 884, + 381 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inverse scaling is a serious concern for several reasons. At a high level, inverse scaling may indicate 'outer misalignment' (Perez et al., 2022) between the model training approach and the purposes to which they are applied. The lack of robustness observed in inverse scaling phenomena may thus indicate that the apparent successes of specific language models at a wide range of benchmarks (e.g., Hendrycks et al., 2021; Srivastava et al., 2022) do not necessarily entail that they have the capability ostensibly being tested (Bowman and Dahl, 2021; Raji et al., 2021).", + "bbox": [ + 507, + 381, + 884, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The existence of inverse scaling is also concerning because of the possibility of other as yet unidentified tasks where performance similarly scales inversely with model size. Models that perform well on a variety of tasks may well present deteriorating performance in unseen tasks with scale, even as performance at established benchmarks increases. This is of particular concern if better performance at established benchmarks and more natural-seeming output leads users to place more trust in such models as general-purpose natural language understanding systems (see, e.g., Bender et al., 2021, for general discussion of such risks).", + "bbox": [ + 507, + 574, + 884, + 783 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Finally, inverse scaling is also of concern because it is often unpredictable. In the same way that certain capabilities appear to emerge at scale (Wei et al., 2022), inverse scaling also appears or accelerates at given scales. For example, as McKenzie et al. (2022b) show, the performance of Gopher (Rae et al., 2022) and Plain LM (Ganguli et al., 2022) at the Inverse Scaling Prize's negated", + "bbox": [ + 507, + 784, + 884, + 912 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14607", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14607-14615 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 210, + 945, + 786, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "question-answering task (NEQA) appears to be stable or even increasing as model size increases, only dropping as model size increases to around 7 billion parameters and beyond (McKenzie et al., 2022b). Thus, inverse scaling may occur not just for unidentified tasks, but also for well-established tasks: a model architecture that performs well at a benchmark at a small scale may suddenly perform surprisingly worse as scale increases—it is not safe to assume that performance will continue to improve or even remain stable.", + "bbox": [ + 112, + 84, + 489, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While previous work has focused on inverse scaling based on the number of model parameters (Lin et al., 2022; McKenzie et al., 2022a,b, 2023a; Jang et al., 2023; Michaelov and Bergen, 2023); as discussed, scaling effects more generally occur not just in relation to model size but also as a function of training data quantity. Recent work has shown that this latter effect has been substantially underestimated (Hoffmann et al., 2022; Touvron et al., 2023a). Thus, in the present study, we consider the possibility of inverse scaling over the course of pretraining; that is, the hypothesized situation where performance on some tasks decreases as the total quantity of training data increases. If inverse scaling of this type exists, it may be more problematic than inverse scaling as a function of model size, because if performance at a given task can degrade over the course of pretraining even as the performance at the majority of tasks improves, a specific model that has previously performed well at a given task may unexpectedly show decreased performance later in training.", + "bbox": [ + 115, + 261, + 490, + 615 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We test how well the 8 models of the Pythia (Biderman et al., 2023) suite perform at 12 tasks over the course of training. These tasks are the 10 multiple-choice tasks that won the Inverse Scaling Challenge (McKenzie et al., 2022b), plus the two multiple-choice versions of the TruthfulQA benchmark, on which language models have also been found to show inverse scaling (Lin et al., 2022). The aim of this study is therefore an exploratory analysis to see whether tasks that show inverse scaling as model parameter size increases also show inverse scaling over the course of training.", + "bbox": [ + 112, + 615, + 489, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Method", + "text_level": 1, + "bbox": [ + 112, + 820, + 218, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Tasks", + "text_level": 1, + "bbox": [ + 112, + 846, + 206, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TRUTHFULQA-MC1 is the first multiple-choice variant of TRUTHFULQA (Lin et al., 2022), which tests whether language models find common mis", + "bbox": [ + 112, + 866, + 489, + 914 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conceptions or statements of fact more predictable in a given context. In this version, there is one true option and the remainder are false. For example, for the prompt what should you take for a viral infection?, the true answer is you should take a rest and drink fluids, and other answers include you should take a full course of antibiotics. The language model is considered to respond correctly if the probability of the correct answer is higher than any other.", + "bbox": [ + 507, + 84, + 884, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TRUTHFULQA-MC2 is an alternative multiple-choice variant of TRUTHFULQA in which there are often multiple correct answers. Here, the model is considered to respond correctly if the total probability of correct responses is greater than that of incorrect responses.", + "bbox": [ + 507, + 244, + 884, + 341 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SIG FIGS, like all the Inverse Scaling Prize tasks, is a multiple-choice task with one correct answer. It also has two incorrect answers, and the model is considered to respond correctly if the correct answer has the highest probability. The task itself tests whether, when prompted to round a number to a number of significant figures, the language model finds the number rounded correctly to be more probable or the number rounded to the same number of decimal places.", + "bbox": [ + 507, + 342, + 884, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NEQA is a zero-shot task with negated questions such as As the barometer reading goes lower there is not a greater chance of A. sunshine B. getting wet.", + "bbox": [ + 507, + 502, + 884, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "REDEFINE is a zero-shot task where expressions are redefined in a range of ways, and then questions are asked are asked about these redefined expressions—e.g., a prompt may ask for the first digit of $5 + 15$ , where $5 + 15$ is first redefined as a text string rather than an equation. The task tests whether the language model does indeed treat the expression in the redefined way rather than its usual interpretation.", + "bbox": [ + 507, + 568, + 882, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MEMO TRAP is a task where a language model is instructed to write a famous quote with a specific last word, e.g., write a quote that ends in the word \"heavy\": Absence makes the heart grow. In this case, the correct answer would be heavy and not the expected fonder.", + "bbox": [ + 505, + 712, + 882, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "HINDSIGHT NEGLECT is a few-shot multiple-choice task where the input contains information about a bet and its outcome and the task is to correctly determine whether or not the bet should have been taken. In the task, a number of examples are provided where the expected value aligns with the", + "bbox": [ + 507, + 809, + 882, + 904 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14608", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "result (if the task has a positive expected value, the individual taking the bet wins, and if it has a negative one, the individual taking the bet loses). For the final question (the one that is answered for the task), the value won or lost does not align (the individual either wins a bet with a negative expected value or loses one with a positive expected value).", + "bbox": [ + 112, + 84, + 487, + 197 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "INTO THE UNKNOWN is a task that involves a description of a setting and a question, with the twist that the task is to identify which of two pieces of information would help to answer the question. One option (the correct answer) contains new information and the other repeats information from the original description.", + "bbox": [ + 112, + 197, + 487, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MODUS TOLLENS tests whether language models can make predictions in line with the modus tollens form of deductive inference, i.e., [i]f $p$ , then $q$ ; not $q$ ; therefore, not $p$ (McKenzie et al., 2023b). The task involves an example of such an inference, and then a question of whether the conclusion is valid or not.", + "bbox": [ + 112, + 310, + 487, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PATTERN MATCH SUPPRESSION tests whether language models can violate a repeated pattern. For example, one prompt is to generate a sequence of 6 symbols alternating between two symbols (A B) but ending unexpectedly. A, B, A, B, A, with possible answers A or B.", + "bbox": [ + 112, + 423, + 487, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RESISITING CORRECTION is a few-shot task, with the instruction to repeat a text without changing it and two examples. In the final example, the sentence to be repeated includes an atypicality, e.g., spelling mistake or a switched word of a famous quote. The task tests whether the model follows the instruction and replicates the atypical, or whether it 'corrects' it.", + "bbox": [ + 112, + 519, + 487, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "REPETITIVE ALGEBRA is a few-shot task based on simple algebra questions. Until the penultimate question, all questions have the same answer (provided in the prompt), and the penultimate question has an answer that differs (also provided in the prompt). For the final question that needs to be answered, the answer is the same as the initial answers. The task tests which of the two answers (initial or penulimate) the model predicts to be more likely.", + "bbox": [ + 112, + 648, + 487, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Models", + "text_level": 1, + "bbox": [ + 112, + 819, + 218, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use the 70 million parameter (70M), 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B Pythia models (Biderman et al., 2023). The models were trained on the autoregressive language modeling", + "bbox": [ + 112, + 839, + 487, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task on The Pile (Gao et al., 2020), an 800GB text dataset comprising 300 billion tokens. All models were trained on this dataset, with checkpoints released at every 2 billion tokens of training. Given that scaling is often considered on a logarithmic scale, we tested each model's performance at 8 checkpoints based on powers of 2: checkpoint 2 (4 billion tokens), checkpoint 4 (8B tokens), checkpoint 8 (16B), checkpoint 16 (32B), checkpoint 32 (64B), checkpoint 64 (128B), checkpoint 128 (256B), and checkpoint 143 (300B tokens, i.e., fully trained).", + "bbox": [ + 507, + 84, + 882, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We run our analyses of model performance using the Language Model Evaluation Harness (Gao et al., 2021). All code, data, and statistical analyses are provided at https://github.com/jmichaelov/emergent-inabilities.", + "bbox": [ + 507, + 277, + 882, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Results", + "text_level": 1, + "bbox": [ + 507, + 370, + 606, + 384 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model performance at each task is shown in Figure 1. In order to quantify the patterns observed, we also fit a least-squares linear regression for each dataset, with the logarithm (base 10) of model parameters, the logarithm (base 10) of training tokens, and the interaction between them as predictors of task accuracy. All variables were z-scored. The results of these tests are shown in Table 1.", + "bbox": [ + 507, + 395, + 882, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The clearest inverse scaling effects can be seen with TRUTHFULQA-MC2—larger models perform worse, performance overall decreases with number of training tokens, and the rate at which performance deteriorates with training tokens increases with model size. Inferential statistics show a negative effect of number of parameters, number of training tokens, and their interaction. In other words, the regression predicts that model performance decreases with number of parameters and training tokens; and in addition, that the larger a model is, the more there is a decrease in performance as the model continues to train. Whether this pattern of statistical results is specific to the tasks used in the present work or to all tasks that show inverse scaling is a question for future work. However, it does also appear to be present for most of the other tasks clearly displaying inverse scaling, namely, HINDSIGHT NEGLECT, MEMO TRAP, PATTERN MATCH SUPPRESSION, and TRUTHFULQA-MC1.", + "bbox": [ + 507, + 524, + 882, + 860 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some of the remaining tasks, namely INTO THE UNKNOWN, MODUS TOLLENS, NEQA, and SIG FIGS display no clear pattern across models. But", + "bbox": [ + 507, + 862, + 882, + 910 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14609", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1a48b1707590b567da907e2b3251cc2fe4ffabfa4c099e736a096ee8db5bf2dc.jpg", + "image_caption": [ + "Figure 1: Performance of the 8 Pythia (Biderman et al., 2023) models at 8 stages over the course of training at the two multiple-choice variants of TRUTHFULQA (Lin et al., 2022) and the 10 multiple-choice winners of the Inverse Scaling Prize (McKenzie et al., 2023b)." + ], + "image_footnote": [], + "bbox": [ + 121, + 86, + 873, + 406 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1ed2a7f35ea74d31e2a717f468b12b011fc15aef1ac4f9ebba1e9aa3b669ff40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskParametersTokensInteraction
Hindsight Neglectt(60)=-4.22, p<0.001t(60)=-4.69, p<0.001t(60)=-2.88, p=0.012
Into the Unknownt(60)=-0.31, p=0.824t(60)=2.04, p=0.079t(60)=0.02, p=0.986
Memo Trapt(60)=-10.05, p<0.001t(60)=-11.34, p<0.001t(60)=-9.71, p<0.001
Modus Tollenst(60)=-0.16, p=0.927t(60)=-2.13, p=0.071t(60)=-1.50, p=0.208
NeQAt(60)=0.79, p=0.559t(60)=-0.08, p=0.963t(60)=2.45, p=0.034
Pattern Match Supp.t(60)=-3.20, p=0.005t(60)=-9.58, p<0.001t(60)=-6.37, p<0.001
Redefinet(60)=0.60, p=0.645t(60)=-0.79, p=0.559t(60)=-1.53, p=0.205
Repetitive Algebrat(60)=-0.49, p=0.706t(60)=-2.11, p=0.071t(60)=-1.00, p=0.443
Resisting Correctiont(60)=5.13, p<0.001t(60)=5.63, p<0.001t(60)=-1.89, p=0.104
Sig Figst(60)=-0.59, p=0.645t(60)=-0.74, p=0.574t(60)=-1.46, p=0.215
TruthfulQA-MC1t(60)=-10.90, p<0.001t(60)=-11.45, p<0.001t(60)=-2.97, p=0.010
TruthfulQA-MC2t(60)=-24.72, p<0.001t(60)=-23.89, p<0.001t(60)=-12.02, p<0.001
", + "bbox": [ + 132, + 472, + 865, + 687 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1: Statistical tests carried out on the performance of the Pythia models, testing the effect of (log-transformed) number of parameters, (log-transformed) number of training tokens, and their interaction. A positive t-value indicates that the variable is significantly correlated with a higher accuracy. All $p$ -values are corrected for multiple comparisons based on false discovery rate (Benjamini and Hochberg, 1995).", + "bbox": [ + 112, + 696, + 884, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "when focusing on just the two largest models, RE-DEFINE appears to show inverse scaling over the course of training, and the largest (12 billion parameter) model shows inverse scaling during training on REPETITIVE ALGEBRA and RESISTING CORRECTION. These may be a case of emergent inverse scaling (i.e., nonlinearities that cannot be accounted for using linear statistical models), espe", + "bbox": [ + 112, + 778, + 490, + 908 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cially in the case of RESISTING CORRECTION, but models with a larger number of parameters would be needed to verify this.", + "bbox": [ + 507, + 778, + 884, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Discussion", + "text_level": 1, + "bbox": [ + 507, + 839, + 636, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find clear evidence of inverse scaling over the course of training on TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT,", + "bbox": [ + 507, + 864, + 884, + 912 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14610", + "bbox": [ + 477, + 927, + 526, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MEMO TRAP, and PATTERN MATCH SUPPRESSION, as well as possible evidence of the same phenomenon on REDEFINE, REPETITIVE ALGEBRA, RESISTING CORRECTION for the largest model or models. In addition, RESISTING CORRECTION appears to present an example of emergence in inverse scaling over the course of training—performance only decreases with training on the largest model.", + "bbox": [ + 112, + 84, + 487, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the time of initial writing, this study was the first to have identified an example of inverse scaling over the course of pretraining. Since then, an official Inverse Scaling Prize paper has been released (McKenzie et al., 2023b). In addition to exploring scaling in terms of the number of floating point operations (FLOPs) needed to train each model, McKenzie et al. (2023b) also analyze the performance of different sizes of the Anthropic LM model (2M, 13M, 42M, 197M, 805M, 3B, 13B, 52B) over the course of training on 400B tokens, providing a valuable point of comparison. On the whole, their results are similar to ours over the same scales. At the larger scale, they find that the 13B and 52B models begin to show inverse scaling on NEQA, SIG FIGS, and INTO THE UNKNOWN. Conversely, only the 52B model begins to show inverse scaling on RESISTING CORRELATION.", + "bbox": [ + 115, + 212, + 489, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "McKenzie et al. (2023b) also classify the tasks into different types. $^{1}$ These classes do not clearly delineate between ones that show inverse scaling and ones that do not based on either our analyses or their analyses. Nonetheless, they provide a valuable starting point for considering the kinds of features of tasks that may lead to different scaling patterns.", + "bbox": [ + 112, + 502, + 487, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Indeed, the question of whether there are consistent scaling patterns based on task features remains an open one. We find several clear cases of inverse scaling that share the pattern of model performance decreasing more rapidly over the course of training as the number of model parameters increases. In several cases there is only a decrease in performance in the largest models. These are not necessarily different phenomena; it may be that the threshold of number of parameters and tokens for tasks like TRUTHFULQA-MC2 is simply lower than for tasks like RESISTING CORRECTION. Additionally, it is not clear whether the main pat", + "bbox": [ + 112, + 632, + 489, + 841 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tern of inverse scaling that we identify—namely, a greater decrease in performance during training in the largest models—is a general feature of inverse scaling, or only due to the fact that we use tasks already known to show inverse scaling as models increase in number of parameters. Future work should establish what kinds of relationships (if any) hold between inverse scaling as a function of model parameters and inverse scaling as a function of training data.", + "bbox": [ + 507, + 84, + 884, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perhaps the main takeaway of the present study is that of instability in model performance. As we see with Pythia 12B on the RESISTING CORRECTION task, a model that was previously among the best at a given task can relatively suddenly experience decreased performance as it continues to train. Good performance on a task at one stage doesn't guarantee continued good performance, even in cases where the model only continues to be trained on text data. This highlights the importance of regular and rigorous evaluation. For this reason, users of models subject to updates would be well advised to verify continuing performance regularly, and it is incumbent on parties who provide such models for use in applications to notify users of updates.", + "bbox": [ + 507, + 246, + 884, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusions", + "text_level": 1, + "bbox": [ + 509, + 498, + 650, + 514 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this study, we set out to investigate whether inverse scaling can occur not only as a function of number of model parameters, but also number of training tokens. We find clear evidence that it does occur with the Pythia (Biderman et al., 2023) suite of models on five of the twelve tasks analyzed, and additional evidence that it may occur on up to eight.", + "bbox": [ + 507, + 523, + 882, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 664, + 616, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main limitations of this study relate to the models used and tasks evaluated. With respect to the former, our analysis is limited to 8 models at various stages in their training. While this means that we can make the inference that the performance of a specific model can deteriorate over the course of training, it also means that it is possible that some of the models have idiosyncratic features that would not generalize to other models of the same size or with the same amount of training data. Additionally, these models cover only part of the possible range of scales for language models—there are contemporary models with many more parameters (e.g., 540 billion parameters in the case of the", + "bbox": [ + 507, + 689, + 884, + 914 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Strong Prior (RESISTING CORRECTION, MEMO TRAP, REDEFINE), Unwanted Imitation (MODUS TOLLENS, TRUTHFULQA), Distractor Task (PATTERN MATCH SUPPRESSION, NEQA, SIG FIGS, INTO THE UNKNOWN), and Spurious Few-Shot (HINDSIGHT NEGLECT, REPETITIVE ALGEBRA).", + "bbox": [ + 112, + 848, + 489, + 910 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14611", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "largest PaLM; Chowdhery et al., 2022) and trained on more data (e.g., 2 trillion tokens in the case of LLaMA 2; Touvron et al., 2023b).", + "bbox": [ + 112, + 84, + 487, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, our analysis is limited to the two multiple-choice versions of TRUTHFULQA and the ten multiple-choice Inverse Scaling Prize tasks. As noted in Section 4, these are all tasks that have been found to exhibit inverse scaling as number of parameters increases. A question for future research is whether the patterns of inverse scaling that we find in the present study occur in all cases of inverse scaling, or whether it is possible to have inverse scaling over the course of training that is not impacted by the number of model parameters.", + "bbox": [ + 115, + 133, + 489, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 112, + 319, + 265, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our work complies with the ACL Ethics Policy. As discussed in the paper, we believe that studies asking questions such as those addressed in the present study are vital for reducing possible harms from language models. We did not train any models for this study, and so the energy consumption is limited to evaluation only: all analyses were run on an NVIDIA RTX A6000 GPU, taking just under 42 hours.", + "bbox": [ + 112, + 344, + 487, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 112, + 501, + 285, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We would like to thank EleutherAI for making the Pythia suite of language models and the Language Model Evaluation Harness available, as well as all those involved with the Inverse Scaling Prize for creating and releasing the tasks. Models were evaluated using hardware provided by the NVIDIA Corporation as part of an NVIDIA Academic Hardware Grant.", + "bbox": [ + 112, + 525, + 489, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 681, + 213, + 695 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, pages 610-623, New York, NY, USA. Association for Computing Machinery.", + "Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289-300.", + "Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit," + ], + "bbox": [ + 115, + 703, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Usvsn Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. 2023. Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling. In Proceedings of the 40th International Conference on Machine Learning, pages 2397-2430. PMLR.", + "Samuel R. Bowman and George Dahl. 2021. What Will it Take to Fix Benchmarking in Natural Language Understanding? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855, Online. Association for Computational Linguistics.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.", + "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways.", + "Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, George Bm Van Den Driessche, Eliza Rutherford, Tom Hennigan, Matthew J. Johnson, Albin Cassirer, Chris Jones, Elena Buchatskaya, David Budden, Laurent Sifre, Simon Osindero, Oriol Vinyals, Marc'Aurelio Ranzato, Jack Rae, Erich Elsen, Koray Kavukcuoglu, and Karen Simonyan. 2022. Unified Scaling Laws for Routed Language" + ], + "bbox": [ + 510, + 85, + 884, + 910 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14612", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Models. In Proceedings of the 39th International Conference on Machine Learning, pages 4057-4086. PMLR.", + "Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. 2022. GLaM: Efficient Scaling of Language Models with Mixture-of-Experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR.", + "Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, Andy Jones, Sam Bowman, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Nelson Elhage, Sheer El-Showk, Stanislav Fort, Zac Hatfield-Dodds, Tom Henighan, Danny Hernandez, Tristan Hume, Josh Jacobson, Scott Johnston, Shauna Kravec, Catherine Olsson, Sam Ringer, Eli Tran-Johnson, Dario Amodei, Tom Brown, Nicholas Joseph, Sam McCandlish, Chris Olah, Jared Kaplan, and Jack Clark. 2022. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned.", + "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800GB Dataset of Diverse Text for Language Modeling.", + "Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation. Zenodo.", + "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. In International Conference on Learning Representations.", + "Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack William Rae, and Laurent Sifre. 2022. An empirical analysis of compute-optimal large language model training. In Advances in Neural Information Processing Systems." + ], + "bbox": [ + 115, + 85, + 487, + 903 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Joel Jang, Seonghyeon Ye, and Minjoon Seo. 2023. Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts. In Proceedings of The 1st Transfer Learning for Natural Language Processing Workshop, pages 52-62. PMLR.", + "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models.", + "Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics.", + "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022a. The inverse scaling prize.", + "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022b. Inverse scaling prize: First round winners.", + "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2023a. Inverse scaling prize: Second round winners.", + "Ian R. McKenzie, Alexander Lyzhov, Michael Pieler, Alicia Parrish, Aaron Mueller, Ameya Prabhu, Euan McLean, Aaron Kirtland, Alexis Ross, Alisa Liu, Andrew Gritsevskiy, Daniel Wurgaft, Derik Kauffman, Gabriel Recchia, Jiacheng Liu, Joe Cavanagh, Max Weiss, Sicong Huang, The Floating Droid, Tom Tseng, Tomasz Korbak, Xudong Shen, Yuhui Zhang, Zhengping Zhou, Najoung Kim, Samuel R. Bowman, and Ethan Perez. 2023b. Inverse Scaling: When Bigger Isn't Better.", + "James Michaelov and Benjamin Bergen. 2023. Rarely a problem? Language models exhibit inverse scaling in their predictions following few-type quantifiers. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14162-14174, Toronto, Canada. Association for Computational Linguistics.", + "Ethan Perez, Ian McKenzie, and Sam Bowman. 2022. Announcing the Inverse Scaling Prize ($250k Prize Pool).", + "Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Olsen," + ], + "bbox": [ + 510, + 85, + 880, + 910 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14613", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sotiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Jason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. 2022. Scaling Language Models: Methods, Analysis & Insights from Training Gopher.", + "bbox": [ + 132, + 85, + 489, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Deborah Raji, Emily Denton, Emily M. Bender, Alex Hanna, and Amandalynne Paullada. 2021. AI and the Everything in the Whole Wide World Benchmark. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1.", + "bbox": [ + 115, + 343, + 489, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshit Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karakaş, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartlomiej Bojanowski, Batuhan Özyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, César Ferri Ramírez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegui González, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurrgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Ekate", + "bbox": [ + 115, + 418, + 489, + 915 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodola, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A. Chi, Ethan Dyer, Ethan Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Martínez-Plumed, Francesca Happé, Francois Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, German Kruszewski, Giambattista Parascandolo, Giorgio Mariani, Gloria Wang, Gonzalo Jaimovitch-López, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Shevlin, Hinrich Schütze, Hiromu Yakura, Hongming Zhang, Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jæhoon Lee, Jaime Fernández Fisac, James B. Simon, James Koppel, James Zheng, James Zou, Jan Kocón, Jana Thompson, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan Waweru, John Burden, John Miller, John U. Balis, Jonathan Berant, Jörg Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Jones, Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth, Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh D. Dhole, Kevin Gimpel, Kevin Omondi, Kory Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar, Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin, Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble, Ludwig Schmidt, Luheng He, Luis Oliveros Colón, Luke Metz, Lütfi Kerem Şenel, Maarten Bosma, Maarten Sap, Maartje ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika, Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ramírez Quintana, Marie Tolkiehn, Mario Giulianielli, Martha Lewis, Martin Potthast, Matthew L. Leavitt, Matthias Hagen, Matyás Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael A. Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michat Swedrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac Suzgun, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T Nanyun Peng,Nathan Chi Nayeon Lee Neta Gur-Ari Krakover Nicholas Cameron,Nicholas Roberts Nick Doiron,Nikita Nangia,Niklas Deckers,Niklas Muennighoff,Nitish Shirish Keskar,Niveditha S.Iyer Noah Constant Noah Fiedel Nuan Wen Oliver ZhangOmar AghaOmar Elbaghdadi Omer Levy Owain Evans Pablo Antonio Moreno Casares Parth Doshi Pascale FungPaul Pu LiangPaul Vicol,Pegah Alipoormolabashi Peiyuan Liao Percy Liang Peter Chang Peter EckersleyPhu Mon Htut Pinyu HwangPiotr Mlkkowski,Piyush Patil,Pouya Pezeshkpour,Priti", + "bbox": [ + 526, + 85, + 884, + 907 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14614", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Oli, Qiao zhu Mei, Qing Lyu, Qin lang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ramón Risco Delgado, Raphael Milliere, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan LeBras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Ruslan Salakhutdinov, Ryan Chi, Ryan Lee, Ryan Stovall, Ryan Teehan, Ryan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel S. Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshiwal, Shyam Upadhyay, Shyamolima, Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad, Steven T. Pi-antadosi, Stuart M. Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal Linzen, Tal Schuster, Tao Li, Tao Yu, Tariq Ali, Tatsu Hashimoto, Te-Lin Wu, Theo Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Timothy Telleen-Lawton Titus Tunduny, Tobias Gerstenberg, Trenton Chang Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai Vikas Raunak Vinay Ramasesh Vinay Uday Prabhu Vishakh Padmakumar,Vivek Srikumar William FedusWilliam SaundersWilliam ZhangWout Vossen Xiang RenXiaoyu TongXinran ZhaoXinyi Wu Xudong Shen,Yadollah YaghoobzadehYair Lakretz Yangqiu Song,Yasaman Bahri,Yejin Choi,Yichi Yang,Yiding Hao,Yifu Chen,Yonatan Belinkov Yu HouYufang HouYuntao BaiZachary Seid Zhuoye Zhao Zijian Wang Zijie J.WangZirui Wang,and Ziyi Wu.2022.Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models.", + "bbox": [ + 132, + 84, + 489, + 671 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A Large Language Model for Science.", + "bbox": [ + 114, + 681, + 489, + 736 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran,", + "bbox": [ + 115, + 744, + 489, + 915 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. 2022. LaMDA: Language Models for Dialog Applications.", + "bbox": [ + 526, + 85, + 884, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. LLaMA: Open and Efficient Foundation Language Models.", + "bbox": [ + 509, + 200, + 884, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open Foundation and Fine-Tuned Chat Models.", + "bbox": [ + 509, + 288, + 884, + 589 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Models. Transactions on Machine Learning Research.", + "bbox": [ + 509, + 599, + 885, + 690 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14615", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_model.json b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dad83fefeea347ba620fb46d21d05014350eb5ce --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_model.json @@ -0,0 +1,1274 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.141, + 0.09, + 0.859, + 0.112 + ], + "angle": 0, + "content": "Emergent Inabilities? Inverse Scaling Over the Course of Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.143, + 0.783, + 0.194 + ], + "angle": 0, + "content": "James A. Michaelov and Benjamin K. Bergen \nDepartment of Cognitive Science, University of California San Diego \n{j1michae,bkbergen}@ucsd.edu" + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.28, + 0.461, + 0.58 + ], + "angle": 0, + "content": "Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT, MEMO TRAP, and PATTERN MATCH SUPPRESSION) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.591, + 0.26, + 0.607 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.616, + 0.49, + 0.825 + ], + "angle": 0, + "content": "For language models, bigger is usually better. Recent research has found that both increased number of model parameters and increased size of the training dataset positively influence model performance (Brown et al., 2020; Kaplan et al., 2020; Chowdhery et al., 2022; Clark et al., 2022; Du et al., 2022; Rae et al., 2022; Hoffmann et al., 2022; Thoppilan et al., 2022; Wei et al., 2022; Taylor et al., 2022; Srivastava et al., 2022; Touvron et al., 2023a). One particularly striking pattern that has been reported is emergence, a nonlinearity in these relationships, where at a particular scale, language models improve rapidly at a given task (Wei et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.826, + 0.49, + 0.906 + ], + "angle": 0, + "content": "However, while increased scale usually leads to improved performance, on certain tasks it correlates with decreased performance. This is known as inverse scaling (Lin et al., 2022). An example of a task on which inverse scaling is observed" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.885, + 0.382 + ], + "angle": 0, + "content": "is the TruthfulQA benchmark, where larger language models are more likely to predict popular misconceptions over statements of fact (Lin et al., 2022). More recently, additional tasks that reportedly show such an effect have been identified as part of the Inverse Scaling Prize (McKenzie et al., 2023b), as well as by other researchers (Jang et al., 2023; Michaelov and Bergen, 2023)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.382, + 0.885, + 0.575 + ], + "angle": 0, + "content": "Inverse scaling is a serious concern for several reasons. At a high level, inverse scaling may indicate 'outer misalignment' (Perez et al., 2022) between the model training approach and the purposes to which they are applied. The lack of robustness observed in inverse scaling phenomena may thus indicate that the apparent successes of specific language models at a wide range of benchmarks (e.g., Hendrycks et al., 2021; Srivastava et al., 2022) do not necessarily entail that they have the capability ostensibly being tested (Bowman and Dahl, 2021; Raji et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.575, + 0.885, + 0.784 + ], + "angle": 0, + "content": "The existence of inverse scaling is also concerning because of the possibility of other as yet unidentified tasks where performance similarly scales inversely with model size. Models that perform well on a variety of tasks may well present deteriorating performance in unseen tasks with scale, even as performance at established benchmarks increases. This is of particular concern if better performance at established benchmarks and more natural-seeming output leads users to place more trust in such models as general-purpose natural language understanding systems (see, e.g., Bender et al., 2021, for general discussion of such risks)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.785, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Finally, inverse scaling is also of concern because it is often unpredictable. In the same way that certain capabilities appear to emerge at scale (Wei et al., 2022), inverse scaling also appears or accelerates at given scales. For example, as McKenzie et al. (2022b) show, the performance of Gopher (Rae et al., 2022) and Plain LM (Ganguli et al., 2022) at the Inverse Scaling Prize's negated" + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14607" + }, + { + "type": "footer", + "bbox": [ + 0.211, + 0.946, + 0.788, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14607-14615 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.261 + ], + "angle": 0, + "content": "question-answering task (NEQA) appears to be stable or even increasing as model size increases, only dropping as model size increases to around 7 billion parameters and beyond (McKenzie et al., 2022b). Thus, inverse scaling may occur not just for unidentified tasks, but also for well-established tasks: a model architecture that performs well at a benchmark at a small scale may suddenly perform surprisingly worse as scale increases—it is not safe to assume that performance will continue to improve or even remain stable." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.262, + 0.492, + 0.616 + ], + "angle": 0, + "content": "While previous work has focused on inverse scaling based on the number of model parameters (Lin et al., 2022; McKenzie et al., 2022a,b, 2023a; Jang et al., 2023; Michaelov and Bergen, 2023); as discussed, scaling effects more generally occur not just in relation to model size but also as a function of training data quantity. Recent work has shown that this latter effect has been substantially underestimated (Hoffmann et al., 2022; Touvron et al., 2023a). Thus, in the present study, we consider the possibility of inverse scaling over the course of pretraining; that is, the hypothesized situation where performance on some tasks decreases as the total quantity of training data increases. If inverse scaling of this type exists, it may be more problematic than inverse scaling as a function of model size, because if performance at a given task can degrade over the course of pretraining even as the performance at the majority of tasks improves, a specific model that has previously performed well at a given task may unexpectedly show decreased performance later in training." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.617, + 0.49, + 0.81 + ], + "angle": 0, + "content": "We test how well the 8 models of the Pythia (Biderman et al., 2023) suite perform at 12 tasks over the course of training. These tasks are the 10 multiple-choice tasks that won the Inverse Scaling Challenge (McKenzie et al., 2022b), plus the two multiple-choice versions of the TruthfulQA benchmark, on which language models have also been found to show inverse scaling (Lin et al., 2022). The aim of this study is therefore an exploratory analysis to see whether tasks that show inverse scaling as model parameter size increases also show inverse scaling over the course of training." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.821, + 0.219, + 0.835 + ], + "angle": 0, + "content": "2 Method" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.847, + 0.207, + 0.86 + ], + "angle": 0, + "content": "2.1 Tasks" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.867, + 0.49, + 0.915 + ], + "angle": 0, + "content": "TRUTHFULQA-MC1 is the first multiple-choice variant of TRUTHFULQA (Lin et al., 2022), which tests whether language models find common mis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.244 + ], + "angle": 0, + "content": "conceptions or statements of fact more predictable in a given context. In this version, there is one true option and the remainder are false. For example, for the prompt what should you take for a viral infection?, the true answer is you should take a rest and drink fluids, and other answers include you should take a full course of antibiotics. The language model is considered to respond correctly if the probability of the correct answer is higher than any other." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.246, + 0.885, + 0.342 + ], + "angle": 0, + "content": "TRUTHFULQA-MC2 is an alternative multiple-choice variant of TRUTHFULQA in which there are often multiple correct answers. Here, the model is considered to respond correctly if the total probability of correct responses is greater than that of incorrect responses." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.343, + 0.885, + 0.502 + ], + "angle": 0, + "content": "SIG FIGS, like all the Inverse Scaling Prize tasks, is a multiple-choice task with one correct answer. It also has two incorrect answers, and the model is considered to respond correctly if the correct answer has the highest probability. The task itself tests whether, when prompted to round a number to a number of significant figures, the language model finds the number rounded correctly to be more probable or the number rounded to the same number of decimal places." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.504, + 0.885, + 0.567 + ], + "angle": 0, + "content": "NEQA is a zero-shot task with negated questions such as As the barometer reading goes lower there is not a greater chance of A. sunshine B. getting wet." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.569, + 0.884, + 0.711 + ], + "angle": 0, + "content": "REDEFINE is a zero-shot task where expressions are redefined in a range of ways, and then questions are asked are asked about these redefined expressions—e.g., a prompt may ask for the first digit of \\(5 + 15\\), where \\(5 + 15\\) is first redefined as a text string rather than an equation. The task tests whether the language model does indeed treat the expression in the redefined way rather than its usual interpretation." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.713, + 0.883, + 0.809 + ], + "angle": 0, + "content": "MEMO TRAP is a task where a language model is instructed to write a famous quote with a specific last word, e.g., write a quote that ends in the word \"heavy\": Absence makes the heart grow. In this case, the correct answer would be heavy and not the expected fonder." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.81, + 0.884, + 0.905 + ], + "angle": 0, + "content": "HINDSIGHT NEGLECT is a few-shot multiple-choice task where the input contains information about a bet and its outcome and the task is to correctly determine whether or not the bet should have been taken. In the task, a number of examples are provided where the expected value aligns with the" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14608" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.198 + ], + "angle": 0, + "content": "result (if the task has a positive expected value, the individual taking the bet wins, and if it has a negative one, the individual taking the bet loses). For the final question (the one that is answered for the task), the value won or lost does not align (the individual either wins a bet with a negative expected value or loses one with a positive expected value)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.198, + 0.488, + 0.31 + ], + "angle": 0, + "content": "INTO THE UNKNOWN is a task that involves a description of a setting and a question, with the twist that the task is to identify which of two pieces of information would help to answer the question. One option (the correct answer) contains new information and the other repeats information from the original description." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.311, + 0.488, + 0.422 + ], + "angle": 0, + "content": "MODUS TOLLENS tests whether language models can make predictions in line with the modus tollens form of deductive inference, i.e., [i]f \\( p \\), then \\( q \\); not \\( q \\); therefore, not \\( p \\) (McKenzie et al., 2023b). The task involves an example of such an inference, and then a question of whether the conclusion is valid or not." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.424, + 0.488, + 0.519 + ], + "angle": 0, + "content": "PATTERN MATCH SUPPRESSION tests whether language models can violate a repeated pattern. For example, one prompt is to generate a sequence of 6 symbols alternating between two symbols (A B) but ending unexpectedly. A, B, A, B, A, with possible answers A or B." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.52, + 0.488, + 0.648 + ], + "angle": 0, + "content": "RESISITING CORRECTION is a few-shot task, with the instruction to repeat a text without changing it and two examples. In the final example, the sentence to be repeated includes an atypicality, e.g., spelling mistake or a switched word of a famous quote. The task tests whether the model follows the instruction and replicates the atypical, or whether it 'corrects' it." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.649, + 0.488, + 0.809 + ], + "angle": 0, + "content": "REPETITIVE ALGEBRA is a few-shot task based on simple algebra questions. Until the penultimate question, all questions have the same answer (provided in the prompt), and the penultimate question has an answer that differs (also provided in the prompt). For the final question that needs to be answered, the answer is the same as the initial answers. The task tests which of the two answers (initial or penulimate) the model predicts to be more likely." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.82, + 0.219, + 0.834 + ], + "angle": 0, + "content": "2.2 Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.84, + 0.488, + 0.905 + ], + "angle": 0, + "content": "We use the 70 million parameter (70M), 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B Pythia models (Biderman et al., 2023). The models were trained on the autoregressive language modeling" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.884, + 0.277 + ], + "angle": 0, + "content": "task on The Pile (Gao et al., 2020), an 800GB text dataset comprising 300 billion tokens. All models were trained on this dataset, with checkpoints released at every 2 billion tokens of training. Given that scaling is often considered on a logarithmic scale, we tested each model's performance at 8 checkpoints based on powers of 2: checkpoint 2 (4 billion tokens), checkpoint 4 (8B tokens), checkpoint 8 (16B), checkpoint 16 (32B), checkpoint 32 (64B), checkpoint 64 (128B), checkpoint 128 (256B), and checkpoint 143 (300B tokens, i.e., fully trained)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.278, + 0.884, + 0.358 + ], + "angle": 0, + "content": "We run our analyses of model performance using the Language Model Evaluation Harness (Gao et al., 2021). All code, data, and statistical analyses are provided at https://github.com/jmichaelov/emergent-inabilities." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.371, + 0.608, + 0.385 + ], + "angle": 0, + "content": "3 Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.396, + 0.884, + 0.523 + ], + "angle": 0, + "content": "Model performance at each task is shown in Figure 1. In order to quantify the patterns observed, we also fit a least-squares linear regression for each dataset, with the logarithm (base 10) of model parameters, the logarithm (base 10) of training tokens, and the interaction between them as predictors of task accuracy. All variables were z-scored. The results of these tests are shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.525, + 0.884, + 0.862 + ], + "angle": 0, + "content": "The clearest inverse scaling effects can be seen with TRUTHFULQA-MC2—larger models perform worse, performance overall decreases with number of training tokens, and the rate at which performance deteriorates with training tokens increases with model size. Inferential statistics show a negative effect of number of parameters, number of training tokens, and their interaction. In other words, the regression predicts that model performance decreases with number of parameters and training tokens; and in addition, that the larger a model is, the more there is a decrease in performance as the model continues to train. Whether this pattern of statistical results is specific to the tasks used in the present work or to all tasks that show inverse scaling is a question for future work. However, it does also appear to be present for most of the other tasks clearly displaying inverse scaling, namely, HINDSIGHT NEGLECT, MEMO TRAP, PATTERN MATCH SUPPRESSION, and TRUTHFULQA-MC1." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.863, + 0.883, + 0.911 + ], + "angle": 0, + "content": "Some of the remaining tasks, namely INTO THE UNKNOWN, MODUS TOLLENS, NEQA, and SIG FIGS display no clear pattern across models. But" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14609" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.087, + 0.874, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.418, + 0.883, + 0.461 + ], + "angle": 0, + "content": "Figure 1: Performance of the 8 Pythia (Biderman et al., 2023) models at 8 stages over the course of training at the two multiple-choice variants of TRUTHFULQA (Lin et al., 2022) and the 10 multiple-choice winners of the Inverse Scaling Prize (McKenzie et al., 2023b)." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.473, + 0.866, + 0.688 + ], + "angle": 0, + "content": "
TaskParametersTokensInteraction
Hindsight Neglectt(60)=-4.22, p<0.001t(60)=-4.69, p<0.001t(60)=-2.88, p=0.012
Into the Unknownt(60)=-0.31, p=0.824t(60)=2.04, p=0.079t(60)=0.02, p=0.986
Memo Trapt(60)=-10.05, p<0.001t(60)=-11.34, p<0.001t(60)=-9.71, p<0.001
Modus Tollenst(60)=-0.16, p=0.927t(60)=-2.13, p=0.071t(60)=-1.50, p=0.208
NeQAt(60)=0.79, p=0.559t(60)=-0.08, p=0.963t(60)=2.45, p=0.034
Pattern Match Supp.t(60)=-3.20, p=0.005t(60)=-9.58, p<0.001t(60)=-6.37, p<0.001
Redefinet(60)=0.60, p=0.645t(60)=-0.79, p=0.559t(60)=-1.53, p=0.205
Repetitive Algebrat(60)=-0.49, p=0.706t(60)=-2.11, p=0.071t(60)=-1.00, p=0.443
Resisting Correctiont(60)=5.13, p<0.001t(60)=5.63, p<0.001t(60)=-1.89, p=0.104
Sig Figst(60)=-0.59, p=0.645t(60)=-0.74, p=0.574t(60)=-1.46, p=0.215
TruthfulQA-MC1t(60)=-10.90, p<0.001t(60)=-11.45, p<0.001t(60)=-2.97, p=0.010
TruthfulQA-MC2t(60)=-24.72, p<0.001t(60)=-23.89, p<0.001t(60)=-12.02, p<0.001
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.697, + 0.885, + 0.756 + ], + "angle": 0, + "content": "Table 1: Statistical tests carried out on the performance of the Pythia models, testing the effect of (log-transformed) number of parameters, (log-transformed) number of training tokens, and their interaction. A positive t-value indicates that the variable is significantly correlated with a higher accuracy. All \\( p \\)-values are corrected for multiple comparisons based on false discovery rate (Benjamini and Hochberg, 1995)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.78, + 0.491, + 0.909 + ], + "angle": 0, + "content": "when focusing on just the two largest models, RE-DEFINE appears to show inverse scaling over the course of training, and the largest (12 billion parameter) model shows inverse scaling during training on REPETITIVE ALGEBRA and RESISTING CORRECTION. These may be a case of emergent inverse scaling (i.e., nonlinearities that cannot be accounted for using linear statistical models), espe" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.78, + 0.885, + 0.828 + ], + "angle": 0, + "content": "cially in the case of RESISTING CORRECTION, but models with a larger number of parameters would be needed to verify this." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.84, + 0.637, + 0.854 + ], + "angle": 0, + "content": "4 Discussion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.865, + 0.885, + 0.913 + ], + "angle": 0, + "content": "We find clear evidence of inverse scaling over the course of training on TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT," + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.527, + 0.941 + ], + "angle": 0, + "content": "14610" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.213 + ], + "angle": 0, + "content": "MEMO TRAP, and PATTERN MATCH SUPPRESSION, as well as possible evidence of the same phenomenon on REDEFINE, REPETITIVE ALGEBRA, RESISTING CORRECTION for the largest model or models. In addition, RESISTING CORRECTION appears to present an example of emergence in inverse scaling over the course of training—performance only decreases with training on the largest model." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.214, + 0.49, + 0.502 + ], + "angle": 0, + "content": "At the time of initial writing, this study was the first to have identified an example of inverse scaling over the course of pretraining. Since then, an official Inverse Scaling Prize paper has been released (McKenzie et al., 2023b). In addition to exploring scaling in terms of the number of floating point operations (FLOPs) needed to train each model, McKenzie et al. (2023b) also analyze the performance of different sizes of the Anthropic LM model (2M, 13M, 42M, 197M, 805M, 3B, 13B, 52B) over the course of training on 400B tokens, providing a valuable point of comparison. On the whole, their results are similar to ours over the same scales. At the larger scale, they find that the 13B and 52B models begin to show inverse scaling on NEQA, SIG FIGS, and INTO THE UNKNOWN. Conversely, only the 52B model begins to show inverse scaling on RESISTING CORRELATION." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.504, + 0.489, + 0.631 + ], + "angle": 0, + "content": "McKenzie et al. (2023b) also classify the tasks into different types.\\(^{1}\\) These classes do not clearly delineate between ones that show inverse scaling and ones that do not based on either our analyses or their analyses. Nonetheless, they provide a valuable starting point for considering the kinds of features of tasks that may lead to different scaling patterns." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.633, + 0.49, + 0.842 + ], + "angle": 0, + "content": "Indeed, the question of whether there are consistent scaling patterns based on task features remains an open one. We find several clear cases of inverse scaling that share the pattern of model performance decreasing more rapidly over the course of training as the number of model parameters increases. In several cases there is only a decrease in performance in the largest models. These are not necessarily different phenomena; it may be that the threshold of number of parameters and tokens for tasks like TRUTHFULQA-MC2 is simply lower than for tasks like RESISTING CORRECTION. Additionally, it is not clear whether the main pat" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.246 + ], + "angle": 0, + "content": "tern of inverse scaling that we identify—namely, a greater decrease in performance during training in the largest models—is a general feature of inverse scaling, or only due to the fact that we use tasks already known to show inverse scaling as models increase in number of parameters. Future work should establish what kinds of relationships (if any) hold between inverse scaling as a function of model parameters and inverse scaling as a function of training data." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.247, + 0.885, + 0.488 + ], + "angle": 0, + "content": "Perhaps the main takeaway of the present study is that of instability in model performance. As we see with Pythia 12B on the RESISTING CORRECTION task, a model that was previously among the best at a given task can relatively suddenly experience decreased performance as it continues to train. Good performance on a task at one stage doesn't guarantee continued good performance, even in cases where the model only continues to be trained on text data. This highlights the importance of regular and rigorous evaluation. For this reason, users of models subject to updates would be well advised to verify continuing performance regularly, and it is incumbent on parties who provide such models for use in applications to notify users of updates." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.499, + 0.651, + 0.515 + ], + "angle": 0, + "content": "5 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.524, + 0.884, + 0.653 + ], + "angle": 0, + "content": "In this study, we set out to investigate whether inverse scaling can occur not only as a function of number of model parameters, but also number of training tokens. We find clear evidence that it does occur with the Pythia (Biderman et al., 2023) suite of models on five of the twelve tasks analyzed, and additional evidence that it may occur on up to eight." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.665, + 0.617, + 0.68 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.69, + 0.885, + 0.915 + ], + "angle": 0, + "content": "The main limitations of this study relate to the models used and tasks evaluated. With respect to the former, our analysis is limited to 8 models at various stages in their training. While this means that we can make the inference that the performance of a specific model can deteriorate over the course of training, it also means that it is possible that some of the models have idiosyncratic features that would not generalize to other models of the same size or with the same amount of training data. Additionally, these models cover only part of the possible range of scales for language models—there are contemporary models with many more parameters (e.g., 540 billion parameters in the case of the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.849, + 0.49, + 0.911 + ], + "angle": 0, + "content": "\\(^{1}\\)Strong Prior (RESISTING CORRECTION, MEMO TRAP, REDEFINE), Unwanted Imitation (MODUS TOLLENS, TRUTHFULQA), Distractor Task (PATTERN MATCH SUPPRESSION, NEQA, SIG FIGS, INTO THE UNKNOWN), and Spurious Few-Shot (HINDSIGHT NEGLECT, REPETITIVE ALGEBRA)." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14611" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.132 + ], + "angle": 0, + "content": "largest PaLM; Chowdhery et al., 2022) and trained on more data (e.g., 2 trillion tokens in the case of LLaMA 2; Touvron et al., 2023b)." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.134, + 0.49, + 0.31 + ], + "angle": 0, + "content": "Similarly, our analysis is limited to the two multiple-choice versions of TRUTHFULQA and the ten multiple-choice Inverse Scaling Prize tasks. As noted in Section 4, these are all tasks that have been found to exhibit inverse scaling as number of parameters increases. A question for future research is whether the patterns of inverse scaling that we find in the present study occur in all cases of inverse scaling, or whether it is possible to have inverse scaling over the course of training that is not impacted by the number of model parameters." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.321, + 0.266, + 0.336 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.346, + 0.489, + 0.489 + ], + "angle": 0, + "content": "Our work complies with the ACL Ethics Policy. As discussed in the paper, we believe that studies asking questions such as those addressed in the present study are vital for reducing possible harms from language models. We did not train any models for this study, and so the energy consumption is limited to evaluation only: all analyses were run on an NVIDIA RTX A6000 GPU, taking just under 42 hours." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.502, + 0.287, + 0.518 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.526, + 0.49, + 0.653 + ], + "angle": 0, + "content": "We would like to thank EleutherAI for making the Pythia suite of language models and the Language Model Evaluation Harness available, as well as all those involved with the Inverse Scaling Prize for creating and releasing the tasks. Models were evaluated using hardware provided by the NVIDIA Corporation as part of an NVIDIA Academic Hardware Grant." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.682, + 0.214, + 0.696 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.704, + 0.49, + 0.797 + ], + "angle": 0, + "content": "Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, pages 610-623, New York, NY, USA. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.805, + 0.49, + 0.87 + ], + "angle": 0, + "content": "Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289-300." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.49, + 0.918 + ], + "angle": 0, + "content": "Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit," + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.704, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.166 + ], + "angle": 0, + "content": "Usvsn Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. 2023. Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling. In Proceedings of the 40th International Conference on Machine Learning, pages 2397-2430. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.175, + 0.885, + 0.268 + ], + "angle": 0, + "content": "Samuel R. Bowman and George Dahl. 2021. What Will it Take to Fix Benchmarking in Natural Language Understanding? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.277, + 0.885, + 0.459 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.469, + 0.885, + 0.771 + ], + "angle": 0, + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.78, + 0.885, + 0.911 + ], + "angle": 0, + "content": "Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, George Bm Van Den Driessche, Eliza Rutherford, Tom Hennigan, Matthew J. Johnson, Albin Cassirer, Chris Jones, Elena Buchatskaya, David Budden, Laurent Sifre, Simon Osindero, Oriol Vinyals, Marc'Aurelio Ranzato, Jack Rae, Erich Elsen, Koray Kavukcuoglu, and Karen Simonyan. 2022. Unified Scaling Laws for Routed Language" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.911 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.125 + ], + "angle": 0, + "content": "Models. In Proceedings of the 39th International Conference on Machine Learning, pages 4057-4086. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.136, + 0.488, + 0.292 + ], + "angle": 0, + "content": "Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. 2022. GLaM: Efficient Scaling of Language Models with Mixture-of-Experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.303, + 0.488, + 0.485 + ], + "angle": 0, + "content": "Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, Andy Jones, Sam Bowman, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Nelson Elhage, Sheer El-Showk, Stanislav Fort, Zac Hatfield-Dodds, Tom Henighan, Danny Hernandez, Tristan Hume, Josh Jacobson, Scott Johnston, Shauna Kravec, Catherine Olsson, Sam Ringer, Eli Tran-Johnson, Dario Amodei, Tom Brown, Nicholas Joseph, Sam McCandlish, Chris Olah, Jared Kaplan, and Jack Clark. 2022. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.495, + 0.488, + 0.575 + ], + "angle": 0, + "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800GB Dataset of Diverse Text for Language Modeling." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.584, + 0.488, + 0.675 + ], + "angle": 0, + "content": "Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation. Zenodo." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.686, + 0.488, + 0.752 + ], + "angle": 0, + "content": "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.762, + 0.488, + 0.904 + ], + "angle": 0, + "content": "Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack William Rae, and Laurent Sifre. 2022. An empirical analysis of compute-optimal large language model training. In Advances in Neural Information Processing Systems." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.904 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.152 + ], + "angle": 0, + "content": "Joel Jang, Seonghyeon Ye, and Minjoon Seo. 2023. Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts. In Proceedings of The 1st Transfer Learning for Natural Language Processing Workshop, pages 52-62. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.162, + 0.882, + 0.215 + ], + "angle": 0, + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.224, + 0.882, + 0.303 + ], + "angle": 0, + "content": "Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.313, + 0.882, + 0.366 + ], + "angle": 0, + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022a. The inverse scaling prize." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.376, + 0.882, + 0.429 + ], + "angle": 0, + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022b. Inverse scaling prize: First round winners." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.439, + 0.882, + 0.491 + ], + "angle": 0, + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2023a. Inverse scaling prize: Second round winners." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.501, + 0.882, + 0.633 + ], + "angle": 0, + "content": "Ian R. McKenzie, Alexander Lyzhov, Michael Pieler, Alicia Parrish, Aaron Mueller, Ameya Prabhu, Euan McLean, Aaron Kirtland, Alexis Ross, Alisa Liu, Andrew Gritsevskiy, Daniel Wurgaft, Derik Kauffman, Gabriel Recchia, Jiacheng Liu, Joe Cavanagh, Max Weiss, Sicong Huang, The Floating Droid, Tom Tseng, Tomasz Korbak, Xudong Shen, Yuhui Zhang, Zhengping Zhou, Najoung Kim, Samuel R. Bowman, and Ethan Perez. 2023b. Inverse Scaling: When Bigger Isn't Better." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.643, + 0.882, + 0.721 + ], + "angle": 0, + "content": "James Michaelov and Benjamin Bergen. 2023. Rarely a problem? Language models exhibit inverse scaling in their predictions following few-type quantifiers. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14162-14174, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.731, + 0.882, + 0.77 + ], + "angle": 0, + "content": "Ethan Perez, Ian McKenzie, and Sam Bowman. 2022. Announcing the Inverse Scaling Prize ($250k Prize Pool)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.78, + 0.882, + 0.912 + ], + "angle": 0, + "content": "Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Olsen," + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.912 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.133, + 0.086, + 0.49, + 0.335 + ], + "angle": 0, + "content": "Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sotiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Jason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. 2022. Scaling Language Models: Methods, Analysis & Insights from Training Gopher." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.344, + 0.49, + 0.412 + ], + "angle": 0, + "content": "Deborah Raji, Emily Denton, Emily M. Bender, Alex Hanna, and Amandalynne Paullada. 2021. AI and the Everything in the Whole Wide World Benchmark. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.419, + 0.49, + 0.916 + ], + "angle": 0, + "content": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshit Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karakaş, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartlomiej Bojanowski, Batuhan Özyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, César Ferri Ramírez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegui González, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurrgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Ekate" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.908 + ], + "angle": 0, + "content": "rina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodola, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A. Chi, Ethan Dyer, Ethan Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Martínez-Plumed, Francesca Happé, Francois Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, German Kruszewski, Giambattista Parascandolo, Giorgio Mariani, Gloria Wang, Gonzalo Jaimovitch-López, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Shevlin, Hinrich Schütze, Hiromu Yakura, Hongming Zhang, Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jæhoon Lee, Jaime Fernández Fisac, James B. Simon, James Koppel, James Zheng, James Zou, Jan Kocón, Jana Thompson, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan Waweru, John Burden, John Miller, John U. Balis, Jonathan Berant, Jörg Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Jones, Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth, Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh D. Dhole, Kevin Gimpel, Kevin Omondi, Kory Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar, Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin, Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble, Ludwig Schmidt, Luheng He, Luis Oliveros Colón, Luke Metz, Lütfi Kerem Şenel, Maarten Bosma, Maarten Sap, Maartje ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika, Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ramírez Quintana, Marie Tolkiehn, Mario Giulianielli, Martha Lewis, Martin Potthast, Matthew L. Leavitt, Matthias Hagen, Matyás Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael A. Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michat Swedrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac Suzgun, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T Nanyun Peng,Nathan Chi Nayeon Lee Neta Gur-Ari Krakover Nicholas Cameron,Nicholas Roberts Nick Doiron,Nikita Nangia,Niklas Deckers,Niklas Muennighoff,Nitish Shirish Keskar,Niveditha S.Iyer Noah Constant Noah Fiedel Nuan Wen Oliver ZhangOmar AghaOmar Elbaghdadi Omer Levy Owain Evans Pablo Antonio Moreno Casares Parth Doshi Pascale FungPaul Pu LiangPaul Vicol,Pegah Alipoormolabashi Peiyuan Liao Percy Liang Peter Chang Peter EckersleyPhu Mon Htut Pinyu HwangPiotr Mlkkowski,Piyush Patil,Pouya Pezeshkpour,Priti" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "14614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.133, + 0.085, + 0.49, + 0.673 + ], + "angle": 0, + "content": "Oli, Qiao zhu Mei, Qing Lyu, Qin lang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ramón Risco Delgado, Raphael Milliere, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan LeBras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Ruslan Salakhutdinov, Ryan Chi, Ryan Lee, Ryan Stovall, Ryan Teehan, Ryan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel S. Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshiwal, Shyam Upadhyay, Shyamolima, Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad, Steven T. Pi-antadosi, Stuart M. Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal Linzen, Tal Schuster, Tao Li, Tao Yu, Tariq Ali, Tatsu Hashimoto, Te-Lin Wu, Theo Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Timothy Telleen-Lawton Titus Tunduny, Tobias Gerstenberg, Trenton Chang Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai Vikas Raunak Vinay Ramasesh Vinay Uday Prabhu Vishakh Padmakumar,Vivek Srikumar William FedusWilliam SaundersWilliam ZhangWout Vossen Xiang RenXiaoyu TongXinran ZhaoXinyi Wu Xudong Shen,Yadollah YaghoobzadehYair Lakretz Yangqiu Song,Yasaman Bahri,Yejin Choi,Yichi Yang,Yiding Hao,Yifu Chen,Yonatan Belinkov Yu HouYufang HouYuntao BaiZachary Seid Zhuoye Zhao Zijian Wang Zijie J.WangZirui Wang,and Ziyi Wu.2022.Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.682, + 0.49, + 0.737 + ], + "angle": 0, + "content": "Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A Large Language Model for Science." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.745, + 0.49, + 0.916 + ], + "angle": 0, + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran," + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.086, + 0.885, + 0.192 + ], + "angle": 0, + "content": "Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. 2022. LaMDA: Language Models for Dialog Applications." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.201, + 0.885, + 0.281 + ], + "angle": 0, + "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. LLaMA: Open and Efficient Foundation Language Models." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.289, + 0.885, + 0.59 + ], + "angle": 0, + "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open Foundation and Fine-Tuned Chat Models." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.6, + 0.887, + 0.692 + ], + "angle": 0, + "content": "Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Models. Transactions on Machine Learning Research." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "14615" + } + ] +] \ No newline at end of file diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_origin.pdf b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e73ec750e21dad2bcdf594ee7396306738c8efaa --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/5fe6aa26-b646-469e-8865-eaa4efa360e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b26a4c235a1ea001c7d5577c1586aa823cf3ad220d6b449caf9127c625f59644 +size 255521 diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/full.md b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8e940e07da151110f8ae2c801f21bfe5a469e900 --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/full.md @@ -0,0 +1,177 @@ +# Emergent Inabilities? Inverse Scaling Over the Course of Pretraining + +James A. Michaelov and Benjamin K. Bergen +Department of Cognitive Science, University of California San Diego +{j1michae,bkbergen}@ucsd.edu + +# Abstract + +Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT, MEMO TRAP, and PATTERN MATCH SUPPRESSION) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves. + +# 1 Introduction + +For language models, bigger is usually better. Recent research has found that both increased number of model parameters and increased size of the training dataset positively influence model performance (Brown et al., 2020; Kaplan et al., 2020; Chowdhery et al., 2022; Clark et al., 2022; Du et al., 2022; Rae et al., 2022; Hoffmann et al., 2022; Thoppilan et al., 2022; Wei et al., 2022; Taylor et al., 2022; Srivastava et al., 2022; Touvron et al., 2023a). One particularly striking pattern that has been reported is emergence, a nonlinearity in these relationships, where at a particular scale, language models improve rapidly at a given task (Wei et al., 2022). + +However, while increased scale usually leads to improved performance, on certain tasks it correlates with decreased performance. This is known as inverse scaling (Lin et al., 2022). An example of a task on which inverse scaling is observed + +is the TruthfulQA benchmark, where larger language models are more likely to predict popular misconceptions over statements of fact (Lin et al., 2022). More recently, additional tasks that reportedly show such an effect have been identified as part of the Inverse Scaling Prize (McKenzie et al., 2023b), as well as by other researchers (Jang et al., 2023; Michaelov and Bergen, 2023). + +Inverse scaling is a serious concern for several reasons. At a high level, inverse scaling may indicate 'outer misalignment' (Perez et al., 2022) between the model training approach and the purposes to which they are applied. The lack of robustness observed in inverse scaling phenomena may thus indicate that the apparent successes of specific language models at a wide range of benchmarks (e.g., Hendrycks et al., 2021; Srivastava et al., 2022) do not necessarily entail that they have the capability ostensibly being tested (Bowman and Dahl, 2021; Raji et al., 2021). + +The existence of inverse scaling is also concerning because of the possibility of other as yet unidentified tasks where performance similarly scales inversely with model size. Models that perform well on a variety of tasks may well present deteriorating performance in unseen tasks with scale, even as performance at established benchmarks increases. This is of particular concern if better performance at established benchmarks and more natural-seeming output leads users to place more trust in such models as general-purpose natural language understanding systems (see, e.g., Bender et al., 2021, for general discussion of such risks). + +Finally, inverse scaling is also of concern because it is often unpredictable. In the same way that certain capabilities appear to emerge at scale (Wei et al., 2022), inverse scaling also appears or accelerates at given scales. For example, as McKenzie et al. (2022b) show, the performance of Gopher (Rae et al., 2022) and Plain LM (Ganguli et al., 2022) at the Inverse Scaling Prize's negated + +question-answering task (NEQA) appears to be stable or even increasing as model size increases, only dropping as model size increases to around 7 billion parameters and beyond (McKenzie et al., 2022b). Thus, inverse scaling may occur not just for unidentified tasks, but also for well-established tasks: a model architecture that performs well at a benchmark at a small scale may suddenly perform surprisingly worse as scale increases—it is not safe to assume that performance will continue to improve or even remain stable. + +While previous work has focused on inverse scaling based on the number of model parameters (Lin et al., 2022; McKenzie et al., 2022a,b, 2023a; Jang et al., 2023; Michaelov and Bergen, 2023); as discussed, scaling effects more generally occur not just in relation to model size but also as a function of training data quantity. Recent work has shown that this latter effect has been substantially underestimated (Hoffmann et al., 2022; Touvron et al., 2023a). Thus, in the present study, we consider the possibility of inverse scaling over the course of pretraining; that is, the hypothesized situation where performance on some tasks decreases as the total quantity of training data increases. If inverse scaling of this type exists, it may be more problematic than inverse scaling as a function of model size, because if performance at a given task can degrade over the course of pretraining even as the performance at the majority of tasks improves, a specific model that has previously performed well at a given task may unexpectedly show decreased performance later in training. + +We test how well the 8 models of the Pythia (Biderman et al., 2023) suite perform at 12 tasks over the course of training. These tasks are the 10 multiple-choice tasks that won the Inverse Scaling Challenge (McKenzie et al., 2022b), plus the two multiple-choice versions of the TruthfulQA benchmark, on which language models have also been found to show inverse scaling (Lin et al., 2022). The aim of this study is therefore an exploratory analysis to see whether tasks that show inverse scaling as model parameter size increases also show inverse scaling over the course of training. + +# 2 Method + +# 2.1 Tasks + +TRUTHFULQA-MC1 is the first multiple-choice variant of TRUTHFULQA (Lin et al., 2022), which tests whether language models find common mis + +conceptions or statements of fact more predictable in a given context. In this version, there is one true option and the remainder are false. For example, for the prompt what should you take for a viral infection?, the true answer is you should take a rest and drink fluids, and other answers include you should take a full course of antibiotics. The language model is considered to respond correctly if the probability of the correct answer is higher than any other. + +TRUTHFULQA-MC2 is an alternative multiple-choice variant of TRUTHFULQA in which there are often multiple correct answers. Here, the model is considered to respond correctly if the total probability of correct responses is greater than that of incorrect responses. + +SIG FIGS, like all the Inverse Scaling Prize tasks, is a multiple-choice task with one correct answer. It also has two incorrect answers, and the model is considered to respond correctly if the correct answer has the highest probability. The task itself tests whether, when prompted to round a number to a number of significant figures, the language model finds the number rounded correctly to be more probable or the number rounded to the same number of decimal places. + +NEQA is a zero-shot task with negated questions such as As the barometer reading goes lower there is not a greater chance of A. sunshine B. getting wet. + +REDEFINE is a zero-shot task where expressions are redefined in a range of ways, and then questions are asked are asked about these redefined expressions—e.g., a prompt may ask for the first digit of $5 + 15$ , where $5 + 15$ is first redefined as a text string rather than an equation. The task tests whether the language model does indeed treat the expression in the redefined way rather than its usual interpretation. + +MEMO TRAP is a task where a language model is instructed to write a famous quote with a specific last word, e.g., write a quote that ends in the word "heavy": Absence makes the heart grow. In this case, the correct answer would be heavy and not the expected fonder. + +HINDSIGHT NEGLECT is a few-shot multiple-choice task where the input contains information about a bet and its outcome and the task is to correctly determine whether or not the bet should have been taken. In the task, a number of examples are provided where the expected value aligns with the + +result (if the task has a positive expected value, the individual taking the bet wins, and if it has a negative one, the individual taking the bet loses). For the final question (the one that is answered for the task), the value won or lost does not align (the individual either wins a bet with a negative expected value or loses one with a positive expected value). + +INTO THE UNKNOWN is a task that involves a description of a setting and a question, with the twist that the task is to identify which of two pieces of information would help to answer the question. One option (the correct answer) contains new information and the other repeats information from the original description. + +MODUS TOLLENS tests whether language models can make predictions in line with the modus tollens form of deductive inference, i.e., [i]f $p$ , then $q$ ; not $q$ ; therefore, not $p$ (McKenzie et al., 2023b). The task involves an example of such an inference, and then a question of whether the conclusion is valid or not. + +PATTERN MATCH SUPPRESSION tests whether language models can violate a repeated pattern. For example, one prompt is to generate a sequence of 6 symbols alternating between two symbols (A B) but ending unexpectedly. A, B, A, B, A, with possible answers A or B. + +RESISITING CORRECTION is a few-shot task, with the instruction to repeat a text without changing it and two examples. In the final example, the sentence to be repeated includes an atypicality, e.g., spelling mistake or a switched word of a famous quote. The task tests whether the model follows the instruction and replicates the atypical, or whether it 'corrects' it. + +REPETITIVE ALGEBRA is a few-shot task based on simple algebra questions. Until the penultimate question, all questions have the same answer (provided in the prompt), and the penultimate question has an answer that differs (also provided in the prompt). For the final question that needs to be answered, the answer is the same as the initial answers. The task tests which of the two answers (initial or penulimate) the model predicts to be more likely. + +# 2.2 Models + +We use the 70 million parameter (70M), 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B Pythia models (Biderman et al., 2023). The models were trained on the autoregressive language modeling + +task on The Pile (Gao et al., 2020), an 800GB text dataset comprising 300 billion tokens. All models were trained on this dataset, with checkpoints released at every 2 billion tokens of training. Given that scaling is often considered on a logarithmic scale, we tested each model's performance at 8 checkpoints based on powers of 2: checkpoint 2 (4 billion tokens), checkpoint 4 (8B tokens), checkpoint 8 (16B), checkpoint 16 (32B), checkpoint 32 (64B), checkpoint 64 (128B), checkpoint 128 (256B), and checkpoint 143 (300B tokens, i.e., fully trained). + +We run our analyses of model performance using the Language Model Evaluation Harness (Gao et al., 2021). All code, data, and statistical analyses are provided at https://github.com/jmichaelov/emergent-inabilities. + +# 3 Results + +Model performance at each task is shown in Figure 1. In order to quantify the patterns observed, we also fit a least-squares linear regression for each dataset, with the logarithm (base 10) of model parameters, the logarithm (base 10) of training tokens, and the interaction between them as predictors of task accuracy. All variables were z-scored. The results of these tests are shown in Table 1. + +The clearest inverse scaling effects can be seen with TRUTHFULQA-MC2—larger models perform worse, performance overall decreases with number of training tokens, and the rate at which performance deteriorates with training tokens increases with model size. Inferential statistics show a negative effect of number of parameters, number of training tokens, and their interaction. In other words, the regression predicts that model performance decreases with number of parameters and training tokens; and in addition, that the larger a model is, the more there is a decrease in performance as the model continues to train. Whether this pattern of statistical results is specific to the tasks used in the present work or to all tasks that show inverse scaling is a question for future work. However, it does also appear to be present for most of the other tasks clearly displaying inverse scaling, namely, HINDSIGHT NEGLECT, MEMO TRAP, PATTERN MATCH SUPPRESSION, and TRUTHFULQA-MC1. + +Some of the remaining tasks, namely INTO THE UNKNOWN, MODUS TOLLENS, NEQA, and SIG FIGS display no clear pattern across models. But + +![](images/1a48b1707590b567da907e2b3251cc2fe4ffabfa4c099e736a096ee8db5bf2dc.jpg) +Figure 1: Performance of the 8 Pythia (Biderman et al., 2023) models at 8 stages over the course of training at the two multiple-choice variants of TRUTHFULQA (Lin et al., 2022) and the 10 multiple-choice winners of the Inverse Scaling Prize (McKenzie et al., 2023b). + +
TaskParametersTokensInteraction
Hindsight Neglectt(60)=-4.22, p<0.001t(60)=-4.69, p<0.001t(60)=-2.88, p=0.012
Into the Unknownt(60)=-0.31, p=0.824t(60)=2.04, p=0.079t(60)=0.02, p=0.986
Memo Trapt(60)=-10.05, p<0.001t(60)=-11.34, p<0.001t(60)=-9.71, p<0.001
Modus Tollenst(60)=-0.16, p=0.927t(60)=-2.13, p=0.071t(60)=-1.50, p=0.208
NeQAt(60)=0.79, p=0.559t(60)=-0.08, p=0.963t(60)=2.45, p=0.034
Pattern Match Supp.t(60)=-3.20, p=0.005t(60)=-9.58, p<0.001t(60)=-6.37, p<0.001
Redefinet(60)=0.60, p=0.645t(60)=-0.79, p=0.559t(60)=-1.53, p=0.205
Repetitive Algebrat(60)=-0.49, p=0.706t(60)=-2.11, p=0.071t(60)=-1.00, p=0.443
Resisting Correctiont(60)=5.13, p<0.001t(60)=5.63, p<0.001t(60)=-1.89, p=0.104
Sig Figst(60)=-0.59, p=0.645t(60)=-0.74, p=0.574t(60)=-1.46, p=0.215
TruthfulQA-MC1t(60)=-10.90, p<0.001t(60)=-11.45, p<0.001t(60)=-2.97, p=0.010
TruthfulQA-MC2t(60)=-24.72, p<0.001t(60)=-23.89, p<0.001t(60)=-12.02, p<0.001
+ +Table 1: Statistical tests carried out on the performance of the Pythia models, testing the effect of (log-transformed) number of parameters, (log-transformed) number of training tokens, and their interaction. A positive t-value indicates that the variable is significantly correlated with a higher accuracy. All $p$ -values are corrected for multiple comparisons based on false discovery rate (Benjamini and Hochberg, 1995). + +when focusing on just the two largest models, RE-DEFINE appears to show inverse scaling over the course of training, and the largest (12 billion parameter) model shows inverse scaling during training on REPETITIVE ALGEBRA and RESISTING CORRECTION. These may be a case of emergent inverse scaling (i.e., nonlinearities that cannot be accounted for using linear statistical models), espe + +cially in the case of RESISTING CORRECTION, but models with a larger number of parameters would be needed to verify this. + +# 4 Discussion + +We find clear evidence of inverse scaling over the course of training on TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT, + +MEMO TRAP, and PATTERN MATCH SUPPRESSION, as well as possible evidence of the same phenomenon on REDEFINE, REPETITIVE ALGEBRA, RESISTING CORRECTION for the largest model or models. In addition, RESISTING CORRECTION appears to present an example of emergence in inverse scaling over the course of training—performance only decreases with training on the largest model. + +At the time of initial writing, this study was the first to have identified an example of inverse scaling over the course of pretraining. Since then, an official Inverse Scaling Prize paper has been released (McKenzie et al., 2023b). In addition to exploring scaling in terms of the number of floating point operations (FLOPs) needed to train each model, McKenzie et al. (2023b) also analyze the performance of different sizes of the Anthropic LM model (2M, 13M, 42M, 197M, 805M, 3B, 13B, 52B) over the course of training on 400B tokens, providing a valuable point of comparison. On the whole, their results are similar to ours over the same scales. At the larger scale, they find that the 13B and 52B models begin to show inverse scaling on NEQA, SIG FIGS, and INTO THE UNKNOWN. Conversely, only the 52B model begins to show inverse scaling on RESISTING CORRELATION. + +McKenzie et al. (2023b) also classify the tasks into different types. $^{1}$ These classes do not clearly delineate between ones that show inverse scaling and ones that do not based on either our analyses or their analyses. Nonetheless, they provide a valuable starting point for considering the kinds of features of tasks that may lead to different scaling patterns. + +Indeed, the question of whether there are consistent scaling patterns based on task features remains an open one. We find several clear cases of inverse scaling that share the pattern of model performance decreasing more rapidly over the course of training as the number of model parameters increases. In several cases there is only a decrease in performance in the largest models. These are not necessarily different phenomena; it may be that the threshold of number of parameters and tokens for tasks like TRUTHFULQA-MC2 is simply lower than for tasks like RESISTING CORRECTION. Additionally, it is not clear whether the main pat + +tern of inverse scaling that we identify—namely, a greater decrease in performance during training in the largest models—is a general feature of inverse scaling, or only due to the fact that we use tasks already known to show inverse scaling as models increase in number of parameters. Future work should establish what kinds of relationships (if any) hold between inverse scaling as a function of model parameters and inverse scaling as a function of training data. + +Perhaps the main takeaway of the present study is that of instability in model performance. As we see with Pythia 12B on the RESISTING CORRECTION task, a model that was previously among the best at a given task can relatively suddenly experience decreased performance as it continues to train. Good performance on a task at one stage doesn't guarantee continued good performance, even in cases where the model only continues to be trained on text data. This highlights the importance of regular and rigorous evaluation. For this reason, users of models subject to updates would be well advised to verify continuing performance regularly, and it is incumbent on parties who provide such models for use in applications to notify users of updates. + +# 5 Conclusions + +In this study, we set out to investigate whether inverse scaling can occur not only as a function of number of model parameters, but also number of training tokens. We find clear evidence that it does occur with the Pythia (Biderman et al., 2023) suite of models on five of the twelve tasks analyzed, and additional evidence that it may occur on up to eight. + +# Limitations + +The main limitations of this study relate to the models used and tasks evaluated. With respect to the former, our analysis is limited to 8 models at various stages in their training. While this means that we can make the inference that the performance of a specific model can deteriorate over the course of training, it also means that it is possible that some of the models have idiosyncratic features that would not generalize to other models of the same size or with the same amount of training data. Additionally, these models cover only part of the possible range of scales for language models—there are contemporary models with many more parameters (e.g., 540 billion parameters in the case of the + +largest PaLM; Chowdhery et al., 2022) and trained on more data (e.g., 2 trillion tokens in the case of LLaMA 2; Touvron et al., 2023b). + +Similarly, our analysis is limited to the two multiple-choice versions of TRUTHFULQA and the ten multiple-choice Inverse Scaling Prize tasks. As noted in Section 4, these are all tasks that have been found to exhibit inverse scaling as number of parameters increases. A question for future research is whether the patterns of inverse scaling that we find in the present study occur in all cases of inverse scaling, or whether it is possible to have inverse scaling over the course of training that is not impacted by the number of model parameters. + +# Ethics Statement + +Our work complies with the ACL Ethics Policy. As discussed in the paper, we believe that studies asking questions such as those addressed in the present study are vital for reducing possible harms from language models. We did not train any models for this study, and so the energy consumption is limited to evaluation only: all analyses were run on an NVIDIA RTX A6000 GPU, taking just under 42 hours. + +# Acknowledgements + +We would like to thank EleutherAI for making the Pythia suite of language models and the Language Model Evaluation Harness available, as well as all those involved with the Inverse Scaling Prize for creating and releasing the tasks. Models were evaluated using hardware provided by the NVIDIA Corporation as part of an NVIDIA Academic Hardware Grant. + +# References + +Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, pages 610-623, New York, NY, USA. Association for Computing Machinery. +Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289-300. +Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, + +Usvsn Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. 2023. Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling. In Proceedings of the 40th International Conference on Machine Learning, pages 2397-2430. PMLR. +Samuel R. Bowman and George Dahl. 2021. What Will it Take to Fix Benchmarking in Natural Language Understanding? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855, Online. Association for Computational Linguistics. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc. +Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways. +Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, George Bm Van Den Driessche, Eliza Rutherford, Tom Hennigan, Matthew J. Johnson, Albin Cassirer, Chris Jones, Elena Buchatskaya, David Budden, Laurent Sifre, Simon Osindero, Oriol Vinyals, Marc'Aurelio Ranzato, Jack Rae, Erich Elsen, Koray Kavukcuoglu, and Karen Simonyan. 2022. Unified Scaling Laws for Routed Language + +Models. In Proceedings of the 39th International Conference on Machine Learning, pages 4057-4086. PMLR. +Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. 2022. GLaM: Efficient Scaling of Language Models with Mixture-of-Experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR. +Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, Andy Jones, Sam Bowman, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Nelson Elhage, Sheer El-Showk, Stanislav Fort, Zac Hatfield-Dodds, Tom Henighan, Danny Hernandez, Tristan Hume, Josh Jacobson, Scott Johnston, Shauna Kravec, Catherine Olsson, Sam Ringer, Eli Tran-Johnson, Dario Amodei, Tom Brown, Nicholas Joseph, Sam McCandlish, Chris Olah, Jared Kaplan, and Jack Clark. 2022. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. +Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800GB Dataset of Diverse Text for Language Modeling. +Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation. Zenodo. +Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. In International Conference on Learning Representations. +Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack William Rae, and Laurent Sifre. 2022. An empirical analysis of compute-optimal large language model training. In Advances in Neural Information Processing Systems. + +Joel Jang, Seonghyeon Ye, and Minjoon Seo. 2023. Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts. In Proceedings of The 1st Transfer Learning for Natural Language Processing Workshop, pages 52-62. PMLR. +Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models. +Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics. +Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022a. The inverse scaling prize. +Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022b. Inverse scaling prize: First round winners. +Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2023a. Inverse scaling prize: Second round winners. +Ian R. McKenzie, Alexander Lyzhov, Michael Pieler, Alicia Parrish, Aaron Mueller, Ameya Prabhu, Euan McLean, Aaron Kirtland, Alexis Ross, Alisa Liu, Andrew Gritsevskiy, Daniel Wurgaft, Derik Kauffman, Gabriel Recchia, Jiacheng Liu, Joe Cavanagh, Max Weiss, Sicong Huang, The Floating Droid, Tom Tseng, Tomasz Korbak, Xudong Shen, Yuhui Zhang, Zhengping Zhou, Najoung Kim, Samuel R. Bowman, and Ethan Perez. 2023b. Inverse Scaling: When Bigger Isn't Better. +James Michaelov and Benjamin Bergen. 2023. Rarely a problem? Language models exhibit inverse scaling in their predictions following few-type quantifiers. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14162-14174, Toronto, Canada. Association for Computational Linguistics. +Ethan Perez, Ian McKenzie, and Sam Bowman. 2022. Announcing the Inverse Scaling Prize ($250k Prize Pool). +Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Olsen, + +Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sotiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Jason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. 2022. Scaling Language Models: Methods, Analysis & Insights from Training Gopher. + +Deborah Raji, Emily Denton, Emily M. Bender, Alex Hanna, and Amandalynne Paullada. 2021. AI and the Everything in the Whole Wide World Benchmark. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1. + +Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshit Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karakaş, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartlomiej Bojanowski, Batuhan Özyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, César Ferri Ramírez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegui González, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurrgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Ekate + +rina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodola, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A. Chi, Ethan Dyer, Ethan Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Martínez-Plumed, Francesca Happé, Francois Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, German Kruszewski, Giambattista Parascandolo, Giorgio Mariani, Gloria Wang, Gonzalo Jaimovitch-López, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Shevlin, Hinrich Schütze, Hiromu Yakura, Hongming Zhang, Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jæhoon Lee, Jaime Fernández Fisac, James B. Simon, James Koppel, James Zheng, James Zou, Jan Kocón, Jana Thompson, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan Waweru, John Burden, John Miller, John U. Balis, Jonathan Berant, Jörg Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Jones, Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth, Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh D. Dhole, Kevin Gimpel, Kevin Omondi, Kory Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar, Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin, Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble, Ludwig Schmidt, Luheng He, Luis Oliveros Colón, Luke Metz, Lütfi Kerem Şenel, Maarten Bosma, Maarten Sap, Maartje ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika, Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ramírez Quintana, Marie Tolkiehn, Mario Giulianielli, Martha Lewis, Martin Potthast, Matthew L. Leavitt, Matthias Hagen, Matyás Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael A. Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michat Swedrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac Suzgun, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T Nanyun Peng,Nathan Chi Nayeon Lee Neta Gur-Ari Krakover Nicholas Cameron,Nicholas Roberts Nick Doiron,Nikita Nangia,Niklas Deckers,Niklas Muennighoff,Nitish Shirish Keskar,Niveditha S.Iyer Noah Constant Noah Fiedel Nuan Wen Oliver ZhangOmar AghaOmar Elbaghdadi Omer Levy Owain Evans Pablo Antonio Moreno Casares Parth Doshi Pascale FungPaul Pu LiangPaul Vicol,Pegah Alipoormolabashi Peiyuan Liao Percy Liang Peter Chang Peter EckersleyPhu Mon Htut Pinyu HwangPiotr Mlkkowski,Piyush Patil,Pouya Pezeshkpour,Priti + +Oli, Qiao zhu Mei, Qing Lyu, Qin lang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ramón Risco Delgado, Raphael Milliere, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan LeBras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Ruslan Salakhutdinov, Ryan Chi, Ryan Lee, Ryan Stovall, Ryan Teehan, Ryan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel S. Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshiwal, Shyam Upadhyay, Shyamolima, Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad, Steven T. Pi-antadosi, Stuart M. Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal Linzen, Tal Schuster, Tao Li, Tao Yu, Tariq Ali, Tatsu Hashimoto, Te-Lin Wu, Theo Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Timothy Telleen-Lawton Titus Tunduny, Tobias Gerstenberg, Trenton Chang Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai Vikas Raunak Vinay Ramasesh Vinay Uday Prabhu Vishakh Padmakumar,Vivek Srikumar William FedusWilliam SaundersWilliam ZhangWout Vossen Xiang RenXiaoyu TongXinran ZhaoXinyi Wu Xudong Shen,Yadollah YaghoobzadehYair Lakretz Yangqiu Song,Yasaman Bahri,Yejin Choi,Yichi Yang,Yiding Hao,Yifu Chen,Yonatan Belinkov Yu HouYufang HouYuntao BaiZachary Seid Zhuoye Zhao Zijian Wang Zijie J.WangZirui Wang,and Ziyi Wu.2022.Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models. + +Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A Large Language Model for Science. + +Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, + +Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. 2022. LaMDA: Language Models for Dialog Applications. + +Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. LLaMA: Open and Efficient Foundation Language Models. + +Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open Foundation and Fine-Tuned Chat Models. + +Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Models. Transactions on Machine Learning Research. \ No newline at end of file diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/images.zip b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..93622b5af2bac8f2cb89dc87a29320350a873a5e --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5538db07ab19d1ad3afe05ce521fc9fa3add4140adcb512c0b70c5746437f39 +size 279097 diff --git a/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/layout.json b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1132a3fefc306870c39ab3c5eb55bd044e6bce3e --- /dev/null +++ b/2023/Emergent Inabilities_ Inverse Scaling Over the Course of Pretraining/layout.json @@ -0,0 +1,3986 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 75, + 511, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 75, + 511, + 94 + ], + "spans": [ + { + "bbox": [ + 83, + 75, + 511, + 94 + ], + "type": "text", + "content": "Emergent Inabilities? Inverse Scaling Over the Course of Pretraining" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 130, + 120, + 465, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 120, + 465, + 163 + ], + "spans": [ + { + "bbox": [ + 130, + 120, + 465, + 163 + ], + "type": "text", + "content": "James A. Michaelov and Benjamin K. Bergen \nDepartment of Cognitive Science, University of California San Diego \n{j1michae,bkbergen}@ucsd.edu" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 235, + 274, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 235, + 274, + 487 + ], + "spans": [ + { + "bbox": [ + 86, + 235, + 274, + 487 + ], + "type": "text", + "content": "Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT, MEMO TRAP, and PATTERN MATCH SUPPRESSION) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 497, + 154, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 497, + 154, + 510 + ], + "spans": [ + { + "bbox": [ + 68, + 497, + 154, + 510 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 518, + 291, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 518, + 291, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 518, + 291, + 693 + ], + "type": "text", + "content": "For language models, bigger is usually better. Recent research has found that both increased number of model parameters and increased size of the training dataset positively influence model performance (Brown et al., 2020; Kaplan et al., 2020; Chowdhery et al., 2022; Clark et al., 2022; Du et al., 2022; Rae et al., 2022; Hoffmann et al., 2022; Thoppilan et al., 2022; Wei et al., 2022; Taylor et al., 2022; Srivastava et al., 2022; Touvron et al., 2023a). One particularly striking pattern that has been reported is emergence, a nonlinearity in these relationships, where at a particular scale, language models improve rapidly at a given task (Wei et al., 2022)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 694, + 291, + 761 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 694, + 291, + 761 + ], + "spans": [ + { + "bbox": [ + 67, + 694, + 291, + 761 + ], + "type": "text", + "content": "However, while increased scale usually leads to improved performance, on certain tasks it correlates with decreased performance. This is known as inverse scaling (Lin et al., 2022). An example of a task on which inverse scaling is observed" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 212, + 526, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 526, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 526, + 321 + ], + "type": "text", + "content": "is the TruthfulQA benchmark, where larger language models are more likely to predict popular misconceptions over statements of fact (Lin et al., 2022). More recently, additional tasks that reportedly show such an effect have been identified as part of the Inverse Scaling Prize (McKenzie et al., 2023b), as well as by other researchers (Jang et al., 2023; Michaelov and Bergen, 2023)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 321, + 526, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 321, + 526, + 483 + ], + "spans": [ + { + "bbox": [ + 302, + 321, + 526, + 483 + ], + "type": "text", + "content": "Inverse scaling is a serious concern for several reasons. At a high level, inverse scaling may indicate 'outer misalignment' (Perez et al., 2022) between the model training approach and the purposes to which they are applied. The lack of robustness observed in inverse scaling phenomena may thus indicate that the apparent successes of specific language models at a wide range of benchmarks (e.g., Hendrycks et al., 2021; Srivastava et al., 2022) do not necessarily entail that they have the capability ostensibly being tested (Bowman and Dahl, 2021; Raji et al., 2021)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 483, + 526, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 483, + 526, + 659 + ], + "spans": [ + { + "bbox": [ + 302, + 483, + 526, + 659 + ], + "type": "text", + "content": "The existence of inverse scaling is also concerning because of the possibility of other as yet unidentified tasks where performance similarly scales inversely with model size. Models that perform well on a variety of tasks may well present deteriorating performance in unseen tasks with scale, even as performance at established benchmarks increases. This is of particular concern if better performance at established benchmarks and more natural-seeming output leads users to place more trust in such models as general-purpose natural language understanding systems (see, e.g., Bender et al., 2021, for general discussion of such risks)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 660, + 526, + 767 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 660, + 526, + 767 + ], + "spans": [ + { + "bbox": [ + 302, + 660, + 526, + 767 + ], + "type": "text", + "content": "Finally, inverse scaling is also of concern because it is often unpredictable. In the same way that certain capabilities appear to emerge at scale (Wei et al., 2022), inverse scaling also appears or accelerates at given scales. For example, as McKenzie et al. (2022b) show, the performance of Gopher (Rae et al., 2022) and Plain LM (Ganguli et al., 2022) at the Inverse Scaling Prize's negated" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "14607" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "spans": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14607-14615 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 219 + ], + "type": "text", + "content": "question-answering task (NEQA) appears to be stable or even increasing as model size increases, only dropping as model size increases to around 7 billion parameters and beyond (McKenzie et al., 2022b). Thus, inverse scaling may occur not just for unidentified tasks, but also for well-established tasks: a model architecture that performs well at a benchmark at a small scale may suddenly perform surprisingly worse as scale increases—it is not safe to assume that performance will continue to improve or even remain stable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 220, + 292, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 220, + 292, + 518 + ], + "spans": [ + { + "bbox": [ + 69, + 220, + 292, + 518 + ], + "type": "text", + "content": "While previous work has focused on inverse scaling based on the number of model parameters (Lin et al., 2022; McKenzie et al., 2022a,b, 2023a; Jang et al., 2023; Michaelov and Bergen, 2023); as discussed, scaling effects more generally occur not just in relation to model size but also as a function of training data quantity. Recent work has shown that this latter effect has been substantially underestimated (Hoffmann et al., 2022; Touvron et al., 2023a). Thus, in the present study, we consider the possibility of inverse scaling over the course of pretraining; that is, the hypothesized situation where performance on some tasks decreases as the total quantity of training data increases. If inverse scaling of this type exists, it may be more problematic than inverse scaling as a function of model size, because if performance at a given task can degrade over the course of pretraining even as the performance at the majority of tasks improves, a specific model that has previously performed well at a given task may unexpectedly show decreased performance later in training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 518, + 291, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 518, + 291, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 518, + 291, + 681 + ], + "type": "text", + "content": "We test how well the 8 models of the Pythia (Biderman et al., 2023) suite perform at 12 tasks over the course of training. These tasks are the 10 multiple-choice tasks that won the Inverse Scaling Challenge (McKenzie et al., 2022b), plus the two multiple-choice versions of the TruthfulQA benchmark, on which language models have also been found to show inverse scaling (Lin et al., 2022). The aim of this study is therefore an exploratory analysis to see whether tasks that show inverse scaling as model parameter size increases also show inverse scaling over the course of training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 690, + 130, + 702 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 690, + 130, + 702 + ], + "spans": [ + { + "bbox": [ + 67, + 690, + 130, + 702 + ], + "type": "text", + "content": "2 Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 712, + 123, + 723 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 712, + 123, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 712, + 123, + 723 + ], + "type": "text", + "content": "2.1 Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 729, + 291, + 769 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 729, + 291, + 769 + ], + "spans": [ + { + "bbox": [ + 67, + 729, + 291, + 769 + ], + "type": "text", + "content": "TRUTHFULQA-MC1 is the first multiple-choice variant of TRUTHFULQA (Lin et al., 2022), which tests whether language models find common mis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 205 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 205 + ], + "type": "text", + "content": "conceptions or statements of fact more predictable in a given context. In this version, there is one true option and the remainder are false. For example, for the prompt what should you take for a viral infection?, the true answer is you should take a rest and drink fluids, and other answers include you should take a full course of antibiotics. The language model is considered to respond correctly if the probability of the correct answer is higher than any other." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 206, + 526, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 206, + 526, + 287 + ], + "spans": [ + { + "bbox": [ + 302, + 206, + 526, + 287 + ], + "type": "text", + "content": "TRUTHFULQA-MC2 is an alternative multiple-choice variant of TRUTHFULQA in which there are often multiple correct answers. Here, the model is considered to respond correctly if the total probability of correct responses is greater than that of incorrect responses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 288, + 526, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 288, + 526, + 422 + ], + "spans": [ + { + "bbox": [ + 302, + 288, + 526, + 422 + ], + "type": "text", + "content": "SIG FIGS, like all the Inverse Scaling Prize tasks, is a multiple-choice task with one correct answer. It also has two incorrect answers, and the model is considered to respond correctly if the correct answer has the highest probability. The task itself tests whether, when prompted to round a number to a number of significant figures, the language model finds the number rounded correctly to be more probable or the number rounded to the same number of decimal places." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 423, + 526, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 423, + 526, + 476 + ], + "spans": [ + { + "bbox": [ + 302, + 423, + 526, + 476 + ], + "type": "text", + "content": "NEQA is a zero-shot task with negated questions such as As the barometer reading goes lower there is not a greater chance of A. sunshine B. getting wet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "spans": [ + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "text", + "content": "REDEFINE is a zero-shot task where expressions are redefined in a range of ways, and then questions are asked are asked about these redefined expressions—e.g., a prompt may ask for the first digit of " + }, + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "inline_equation", + "content": "5 + 15" + }, + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "inline_equation", + "content": "5 + 15" + }, + { + "bbox": [ + 302, + 478, + 525, + 597 + ], + "type": "text", + "content": " is first redefined as a text string rather than an equation. The task tests whether the language model does indeed treat the expression in the redefined way rather than its usual interpretation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 599, + 525, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 599, + 525, + 680 + ], + "spans": [ + { + "bbox": [ + 301, + 599, + 525, + 680 + ], + "type": "text", + "content": "MEMO TRAP is a task where a language model is instructed to write a famous quote with a specific last word, e.g., write a quote that ends in the word \"heavy\": Absence makes the heart grow. In this case, the correct answer would be heavy and not the expected fonder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 681, + 525, + 761 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 681, + 525, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 681, + 525, + 761 + ], + "type": "text", + "content": "HINDSIGHT NEGLECT is a few-shot multiple-choice task where the input contains information about a bet and its outcome and the task is to correctly determine whether or not the bet should have been taken. In the task, a number of examples are provided where the expected value aligns with the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14608" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": "result (if the task has a positive expected value, the individual taking the bet wins, and if it has a negative one, the individual taking the bet loses). For the final question (the one that is answered for the task), the value won or lost does not align (the individual either wins a bet with a negative expected value or loses one with a positive expected value)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 166, + 290, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 166, + 290, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 166, + 290, + 260 + ], + "type": "text", + "content": "INTO THE UNKNOWN is a task that involves a description of a setting and a question, with the twist that the task is to identify which of two pieces of information would help to answer the question. One option (the correct answer) contains new information and the other repeats information from the original description." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "content": "MODUS TOLLENS tests whether language models can make predictions in line with the modus tollens form of deductive inference, i.e., [i]f " + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "content": "; not " + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "content": "; therefore, not " + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 261, + 290, + 354 + ], + "type": "text", + "content": " (McKenzie et al., 2023b). The task involves an example of such an inference, and then a question of whether the conclusion is valid or not." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 356, + 290, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 356, + 290, + 436 + ], + "spans": [ + { + "bbox": [ + 67, + 356, + 290, + 436 + ], + "type": "text", + "content": "PATTERN MATCH SUPPRESSION tests whether language models can violate a repeated pattern. For example, one prompt is to generate a sequence of 6 symbols alternating between two symbols (A B) but ending unexpectedly. A, B, A, B, A, with possible answers A or B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 437, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 437, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 437, + 290, + 544 + ], + "type": "text", + "content": "RESISITING CORRECTION is a few-shot task, with the instruction to repeat a text without changing it and two examples. In the final example, the sentence to be repeated includes an atypicality, e.g., spelling mistake or a switched word of a famous quote. The task tests whether the model follows the instruction and replicates the atypical, or whether it 'corrects' it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 545, + 290, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 545, + 290, + 680 + ], + "spans": [ + { + "bbox": [ + 67, + 545, + 290, + 680 + ], + "type": "text", + "content": "REPETITIVE ALGEBRA is a few-shot task based on simple algebra questions. Until the penultimate question, all questions have the same answer (provided in the prompt), and the penultimate question has an answer that differs (also provided in the prompt). For the final question that needs to be answered, the answer is the same as the initial answers. The task tests which of the two answers (initial or penulimate) the model predicts to be more likely." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 689, + 130, + 701 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 689, + 130, + 701 + ], + "spans": [ + { + "bbox": [ + 67, + 689, + 130, + 701 + ], + "type": "text", + "content": "2.2 Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 706, + 290, + 761 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 706, + 290, + 761 + ], + "spans": [ + { + "bbox": [ + 67, + 706, + 290, + 761 + ], + "type": "text", + "content": "We use the 70 million parameter (70M), 160M, 410M, 1B, 1.4B, 2.8B, 6.9B, and 12B Pythia models (Biderman et al., 2023). The models were trained on the autoregressive language modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 525, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 232 + ], + "type": "text", + "content": "task on The Pile (Gao et al., 2020), an 800GB text dataset comprising 300 billion tokens. All models were trained on this dataset, with checkpoints released at every 2 billion tokens of training. Given that scaling is often considered on a logarithmic scale, we tested each model's performance at 8 checkpoints based on powers of 2: checkpoint 2 (4 billion tokens), checkpoint 4 (8B tokens), checkpoint 8 (16B), checkpoint 16 (32B), checkpoint 32 (64B), checkpoint 64 (128B), checkpoint 128 (256B), and checkpoint 143 (300B tokens, i.e., fully trained)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 233, + 525, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 233, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 233, + 525, + 301 + ], + "type": "text", + "content": "We run our analyses of model performance using the Language Model Evaluation Harness (Gao et al., 2021). All code, data, and statistical analyses are provided at https://github.com/jmichaelov/emergent-inabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 312, + 361, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 312, + 361, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 312, + 361, + 323 + ], + "type": "text", + "content": "3 Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 333, + 525, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 333, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 302, + 333, + 525, + 439 + ], + "type": "text", + "content": "Model performance at each task is shown in Figure 1. In order to quantify the patterns observed, we also fit a least-squares linear regression for each dataset, with the logarithm (base 10) of model parameters, the logarithm (base 10) of training tokens, and the interaction between them as predictors of task accuracy. All variables were z-scored. The results of these tests are shown in Table 1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 441, + 525, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 441, + 525, + 724 + ], + "spans": [ + { + "bbox": [ + 302, + 441, + 525, + 724 + ], + "type": "text", + "content": "The clearest inverse scaling effects can be seen with TRUTHFULQA-MC2—larger models perform worse, performance overall decreases with number of training tokens, and the rate at which performance deteriorates with training tokens increases with model size. Inferential statistics show a negative effect of number of parameters, number of training tokens, and their interaction. In other words, the regression predicts that model performance decreases with number of parameters and training tokens; and in addition, that the larger a model is, the more there is a decrease in performance as the model continues to train. Whether this pattern of statistical results is specific to the tasks used in the present work or to all tasks that show inverse scaling is a question for future work. However, it does also appear to be present for most of the other tasks clearly displaying inverse scaling, namely, HINDSIGHT NEGLECT, MEMO TRAP, PATTERN MATCH SUPPRESSION, and TRUTHFULQA-MC1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 725, + 525, + 766 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 725, + 525, + 766 + ], + "spans": [ + { + "bbox": [ + 302, + 725, + 525, + 766 + ], + "type": "text", + "content": "Some of the remaining tasks, namely INTO THE UNKNOWN, MODUS TOLLENS, NEQA, and SIG FIGS display no clear pattern across models. But" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14609" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 73, + 520, + 342 + ], + "blocks": [ + { + "bbox": [ + 72, + 73, + 520, + 342 + ], + "lines": [ + { + "bbox": [ + 72, + 73, + 520, + 342 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 520, + 342 + ], + "type": "image", + "image_path": "1a48b1707590b567da907e2b3251cc2fe4ffabfa4c099e736a096ee8db5bf2dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 351, + 525, + 387 + ], + "lines": [ + { + "bbox": [ + 67, + 351, + 525, + 387 + ], + "spans": [ + { + "bbox": [ + 67, + 351, + 525, + 387 + ], + "type": "text", + "content": "Figure 1: Performance of the 8 Pythia (Biderman et al., 2023) models at 8 stages over the course of training at the two multiple-choice variants of TRUTHFULQA (Lin et al., 2022) and the 10 multiple-choice winners of the Inverse Scaling Prize (McKenzie et al., 2023b)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 79, + 397, + 515, + 578 + ], + "blocks": [ + { + "bbox": [ + 79, + 397, + 515, + 578 + ], + "lines": [ + { + "bbox": [ + 79, + 397, + 515, + 578 + ], + "spans": [ + { + "bbox": [ + 79, + 397, + 515, + 578 + ], + "type": "table", + "html": "
TaskParametersTokensInteraction
Hindsight Neglectt(60)=-4.22, p<0.001t(60)=-4.69, p<0.001t(60)=-2.88, p=0.012
Into the Unknownt(60)=-0.31, p=0.824t(60)=2.04, p=0.079t(60)=0.02, p=0.986
Memo Trapt(60)=-10.05, p<0.001t(60)=-11.34, p<0.001t(60)=-9.71, p<0.001
Modus Tollenst(60)=-0.16, p=0.927t(60)=-2.13, p=0.071t(60)=-1.50, p=0.208
NeQAt(60)=0.79, p=0.559t(60)=-0.08, p=0.963t(60)=2.45, p=0.034
Pattern Match Supp.t(60)=-3.20, p=0.005t(60)=-9.58, p<0.001t(60)=-6.37, p<0.001
Redefinet(60)=0.60, p=0.645t(60)=-0.79, p=0.559t(60)=-1.53, p=0.205
Repetitive Algebrat(60)=-0.49, p=0.706t(60)=-2.11, p=0.071t(60)=-1.00, p=0.443
Resisting Correctiont(60)=5.13, p<0.001t(60)=5.63, p<0.001t(60)=-1.89, p=0.104
Sig Figst(60)=-0.59, p=0.645t(60)=-0.74, p=0.574t(60)=-1.46, p=0.215
TruthfulQA-MC1t(60)=-10.90, p<0.001t(60)=-11.45, p<0.001t(60)=-2.97, p=0.010
TruthfulQA-MC2t(60)=-24.72, p<0.001t(60)=-23.89, p<0.001t(60)=-12.02, p<0.001
", + "image_path": "1ed2a7f35ea74d31e2a717f468b12b011fc15aef1ac4f9ebba1e9aa3b669ff40.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 586, + 526, + 635 + ], + "lines": [ + { + "bbox": [ + 67, + 586, + 526, + 635 + ], + "spans": [ + { + "bbox": [ + 67, + 586, + 526, + 635 + ], + "type": "text", + "content": "Table 1: Statistical tests carried out on the performance of the Pythia models, testing the effect of (log-transformed) number of parameters, (log-transformed) number of training tokens, and their interaction. A positive t-value indicates that the variable is significantly correlated with a higher accuracy. All " + }, + { + "bbox": [ + 67, + 586, + 526, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 586, + 526, + 635 + ], + "type": "text", + "content": "-values are corrected for multiple comparisons based on false discovery rate (Benjamini and Hochberg, 1995)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 655, + 292, + 764 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 655, + 292, + 764 + ], + "spans": [ + { + "bbox": [ + 67, + 655, + 292, + 764 + ], + "type": "text", + "content": "when focusing on just the two largest models, RE-DEFINE appears to show inverse scaling over the course of training, and the largest (12 billion parameter) model shows inverse scaling during training on REPETITIVE ALGEBRA and RESISTING CORRECTION. These may be a case of emergent inverse scaling (i.e., nonlinearities that cannot be accounted for using linear statistical models), espe" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 655, + 526, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 655, + 526, + 696 + ], + "spans": [ + { + "bbox": [ + 302, + 655, + 526, + 696 + ], + "type": "text", + "content": "cially in the case of RESISTING CORRECTION, but models with a larger number of parameters would be needed to verify this." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 706, + 379, + 718 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 379, + 718 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 379, + 718 + ], + "type": "text", + "content": "4 Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 727, + 526, + 767 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 727, + 526, + 767 + ], + "spans": [ + { + "bbox": [ + 302, + 727, + 526, + 767 + ], + "type": "text", + "content": "We find clear evidence of inverse scaling over the course of training on TRUTHFULQA-MC1, TRUTHFULQA-MC2, HINDSIGHT NEGLECT," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "type": "text", + "content": "14610" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 179 + ], + "type": "text", + "content": "MEMO TRAP, and PATTERN MATCH SUPPRESSION, as well as possible evidence of the same phenomenon on REDEFINE, REPETITIVE ALGEBRA, RESISTING CORRECTION for the largest model or models. In addition, RESISTING CORRECTION appears to present an example of emergence in inverse scaling over the course of training—performance only decreases with training on the largest model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 179, + 291, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 179, + 291, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 179, + 291, + 422 + ], + "type": "text", + "content": "At the time of initial writing, this study was the first to have identified an example of inverse scaling over the course of pretraining. Since then, an official Inverse Scaling Prize paper has been released (McKenzie et al., 2023b). In addition to exploring scaling in terms of the number of floating point operations (FLOPs) needed to train each model, McKenzie et al. (2023b) also analyze the performance of different sizes of the Anthropic LM model (2M, 13M, 42M, 197M, 805M, 3B, 13B, 52B) over the course of training on 400B tokens, providing a valuable point of comparison. On the whole, their results are similar to ours over the same scales. At the larger scale, they find that the 13B and 52B models begin to show inverse scaling on NEQA, SIG FIGS, and INTO THE UNKNOWN. Conversely, only the 52B model begins to show inverse scaling on RESISTING CORRELATION." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 423, + 290, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 423, + 290, + 530 + ], + "spans": [ + { + "bbox": [ + 67, + 423, + 290, + 530 + ], + "type": "text", + "content": "McKenzie et al. (2023b) also classify the tasks into different types." + }, + { + "bbox": [ + 67, + 423, + 290, + 530 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 423, + 290, + 530 + ], + "type": "text", + "content": " These classes do not clearly delineate between ones that show inverse scaling and ones that do not based on either our analyses or their analyses. Nonetheless, they provide a valuable starting point for considering the kinds of features of tasks that may lead to different scaling patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 532, + 291, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 532, + 291, + 708 + ], + "spans": [ + { + "bbox": [ + 67, + 532, + 291, + 708 + ], + "type": "text", + "content": "Indeed, the question of whether there are consistent scaling patterns based on task features remains an open one. We find several clear cases of inverse scaling that share the pattern of model performance decreasing more rapidly over the course of training as the number of model parameters increases. In several cases there is only a decrease in performance in the largest models. These are not necessarily different phenomena; it may be that the threshold of number of parameters and tokens for tasks like TRUTHFULQA-MC2 is simply lower than for tasks like RESISTING CORRECTION. Additionally, it is not clear whether the main pat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "content": "tern of inverse scaling that we identify—namely, a greater decrease in performance during training in the largest models—is a general feature of inverse scaling, or only due to the fact that we use tasks already known to show inverse scaling as models increase in number of parameters. Future work should establish what kinds of relationships (if any) hold between inverse scaling as a function of model parameters and inverse scaling as a function of training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 207, + 526, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 207, + 526, + 410 + ], + "spans": [ + { + "bbox": [ + 302, + 207, + 526, + 410 + ], + "type": "text", + "content": "Perhaps the main takeaway of the present study is that of instability in model performance. As we see with Pythia 12B on the RESISTING CORRECTION task, a model that was previously among the best at a given task can relatively suddenly experience decreased performance as it continues to train. Good performance on a task at one stage doesn't guarantee continued good performance, even in cases where the model only continues to be trained on text data. This highlights the importance of regular and rigorous evaluation. For this reason, users of models subject to updates would be well advised to verify continuing performance regularly, and it is incumbent on parties who provide such models for use in applications to notify users of updates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 419, + 387, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 419, + 387, + 433 + ], + "spans": [ + { + "bbox": [ + 303, + 419, + 387, + 433 + ], + "type": "text", + "content": "5 Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 440, + 525, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 440, + 525, + 549 + ], + "spans": [ + { + "bbox": [ + 302, + 440, + 525, + 549 + ], + "type": "text", + "content": "In this study, we set out to investigate whether inverse scaling can occur not only as a function of number of model parameters, but also number of training tokens. We find clear evidence that it does occur with the Pythia (Biderman et al., 2023) suite of models on five of the twelve tasks analyzed, and additional evidence that it may occur on up to eight." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 559, + 367, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 559, + 367, + 571 + ], + "spans": [ + { + "bbox": [ + 303, + 559, + 367, + 571 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 580, + 526, + 769 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 580, + 526, + 769 + ], + "spans": [ + { + "bbox": [ + 302, + 580, + 526, + 769 + ], + "type": "text", + "content": "The main limitations of this study relate to the models used and tasks evaluated. With respect to the former, our analysis is limited to 8 models at various stages in their training. While this means that we can make the inference that the performance of a specific model can deteriorate over the course of training, it also means that it is possible that some of the models have idiosyncratic features that would not generalize to other models of the same size or with the same amount of training data. Additionally, these models cover only part of the possible range of scales for language models—there are contemporary models with many more parameters (e.g., 540 billion parameters in the case of the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 714, + 291, + 766 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 714, + 291, + 766 + ], + "spans": [ + { + "bbox": [ + 67, + 714, + 291, + 766 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 714, + 291, + 766 + ], + "type": "text", + "content": "Strong Prior (RESISTING CORRECTION, MEMO TRAP, REDEFINE), Unwanted Imitation (MODUS TOLLENS, TRUTHFULQA), Distractor Task (PATTERN MATCH SUPPRESSION, NEQA, SIG FIGS, INTO THE UNKNOWN), and Spurious Few-Shot (HINDSIGHT NEGLECT, REPETITIVE ALGEBRA)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "14611" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "content": "largest PaLM; Chowdhery et al., 2022) and trained on more data (e.g., 2 trillion tokens in the case of LLaMA 2; Touvron et al., 2023b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 112, + 291, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 112, + 291, + 260 + ], + "spans": [ + { + "bbox": [ + 69, + 112, + 291, + 260 + ], + "type": "text", + "content": "Similarly, our analysis is limited to the two multiple-choice versions of TRUTHFULQA and the ten multiple-choice Inverse Scaling Prize tasks. As noted in Section 4, these are all tasks that have been found to exhibit inverse scaling as number of parameters increases. A question for future research is whether the patterns of inverse scaling that we find in the present study occur in all cases of inverse scaling, or whether it is possible to have inverse scaling over the course of training that is not impacted by the number of model parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 269, + 158, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 269, + 158, + 282 + ], + "spans": [ + { + "bbox": [ + 67, + 269, + 158, + 282 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 290, + 290, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 290, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 67, + 290, + 290, + 411 + ], + "type": "text", + "content": "Our work complies with the ACL Ethics Policy. As discussed in the paper, we believe that studies asking questions such as those addressed in the present study are vital for reducing possible harms from language models. We did not train any models for this study, and so the energy consumption is limited to evaluation only: all analyses were run on an NVIDIA RTX A6000 GPU, taking just under 42 hours." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 422, + 170, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 422, + 170, + 435 + ], + "spans": [ + { + "bbox": [ + 67, + 422, + 170, + 435 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 442, + 291, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 442, + 291, + 549 + ], + "spans": [ + { + "bbox": [ + 67, + 442, + 291, + 549 + ], + "type": "text", + "content": "We would like to thank EleutherAI for making the Pythia suite of language models and the Language Model Evaluation Harness available, as well as all those involved with the Inverse Scaling Prize for creating and releasing the tasks. Models were evaluated using hardware provided by the NVIDIA Corporation as part of an NVIDIA Academic Hardware Grant." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 573, + 127, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 573, + 127, + 585 + ], + "spans": [ + { + "bbox": [ + 68, + 573, + 127, + 585 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 592, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 592, + 291, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 592, + 291, + 670 + ], + "spans": [ + { + "bbox": [ + 69, + 592, + 291, + 670 + ], + "type": "text", + "content": "Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21, pages 610-623, New York, NY, USA. Association for Computing Machinery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 677, + 291, + 731 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 677, + 291, + 731 + ], + "spans": [ + { + "bbox": [ + 69, + 677, + 291, + 731 + ], + "type": "text", + "content": "Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society. Series B (Methodological), 57(1):289-300." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 291, + 772 + ], + "type": "text", + "content": "Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 766 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 72, + 526, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 139 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 139 + ], + "type": "text", + "content": "Usvsn Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar Van Der Wal. 2023. Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling. In Proceedings of the 40th International Conference on Machine Learning, pages 2397-2430. PMLR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 147, + 526, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 526, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 526, + 225 + ], + "type": "text", + "content": "Samuel R. Bowman and George Dahl. 2021. What Will it Take to Fix Benchmarking in Natural Language Understanding? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4843-4855, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 232, + 526, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 526, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 526, + 386 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 394, + 526, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 526, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 526, + 648 + ], + "type": "text", + "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan First, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. PaLM: Scaling Language Modeling with Pathways." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 655, + 526, + 766 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 655, + 526, + 766 + ], + "spans": [ + { + "bbox": [ + 304, + 655, + 526, + 766 + ], + "type": "text", + "content": "Aidan Clark, Diego De Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, George Bm Van Den Driessche, Eliza Rutherford, Tom Hennigan, Matthew J. Johnson, Albin Cassirer, Chris Jones, Elena Buchatskaya, David Budden, Laurent Sifre, Simon Osindero, Oriol Vinyals, Marc'Aurelio Ranzato, Jack Rae, Erich Elsen, Koray Kavukcuoglu, and Karen Simonyan. 2022. Unified Scaling Laws for Routed Language" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14612" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 760 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "type": "text", + "content": "Models. In Proceedings of the 39th International Conference on Machine Learning, pages 4057-4086. PMLR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 114, + 290, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 114, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 290, + 245 + ], + "type": "text", + "content": "Nan Du, Yanping Huang, Andrew M. Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan Firat, Barret Zoph, Liam Fedus, Maarten P. Bosma, Zongwei Zhou, Tao Wang, Emma Wang, Kellie Webster, Marie Pellat, Kevin Robinson, Kathleen Meier-Hellstern, Toju Duke, Lucas Dixon, Kun Zhang, Quoc Le, Yonghui Wu, Zhifeng Chen, and Claire Cui. 2022. GLaM: Efficient Scaling of Language Models with Mixture-of-Experts. In Proceedings of the 39th International Conference on Machine Learning, pages 5547-5569. PMLR." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 254, + 290, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 254, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 69, + 254, + 290, + 407 + ], + "type": "text", + "content": "Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, Andy Jones, Sam Bowman, Anna Chen, Tom Conerly, Nova DasSarma, Dawn Drain, Nelson Elhage, Sheer El-Showk, Stanislav Fort, Zac Hatfield-Dodds, Tom Henighan, Danny Hernandez, Tristan Hume, Josh Jacobson, Scott Johnston, Shauna Kravec, Catherine Olsson, Sam Ringer, Eli Tran-Johnson, Dario Amodei, Tom Brown, Nicholas Joseph, Sam McCandlish, Chris Olah, Jared Kaplan, and Jack Clark. 2022. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 416, + 290, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 416, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 416, + 290, + 483 + ], + "type": "text", + "content": "Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800GB Dataset of Diverse Text for Language Modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 491, + 290, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 491, + 290, + 567 + ], + "spans": [ + { + "bbox": [ + 69, + 491, + 290, + 567 + ], + "type": "text", + "content": "Leo Gao, Jonathan Tow, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Kyle McDonell, Niklas Muennighoff, Jason Phang, Laria Reynolds, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2021. A framework for few-shot language model evaluation. Zenodo." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 576, + 290, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 576, + 290, + 632 + ], + "spans": [ + { + "bbox": [ + 69, + 576, + 290, + 632 + ], + "type": "text", + "content": "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Understanding. In International Conference on Learning Representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 640, + 290, + 760 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 640, + 290, + 760 + ], + "spans": [ + { + "bbox": [ + 69, + 640, + 290, + 760 + ], + "type": "text", + "content": "Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack William Rae, and Laurent Sifre. 2022. An empirical analysis of compute-optimal large language model training. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 766 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "text", + "content": "Joel Jang, Seonghyeon Ye, and Minjoon Seo. 2023. Can Large Language Models Truly Understand Prompts? A Case Study with Negated Prompts. In Proceedings of The 1st Transfer Learning for Natural Language Processing Workshop, pages 52-62. PMLR." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 136, + 524, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 524, + 180 + ], + "type": "text", + "content": "Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling Laws for Neural Language Models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 188, + 524, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 524, + 254 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 524, + 254 + ], + "type": "text", + "content": "Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. TruthfulQA: Measuring How Models Mimic Human Falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214-3252, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 263, + 524, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 524, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 524, + 307 + ], + "type": "text", + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022a. The inverse scaling prize." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 316, + 524, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 524, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 524, + 360 + ], + "type": "text", + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2022b. Inverse scaling prize: First round winners." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 524, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 524, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 524, + 412 + ], + "type": "text", + "content": "Ian McKenzie, Alexander Lyzhov, Alicia Parrish, Ameya Prabhu, Aaron Mueller, Najoung Kim, Sam Bowman, and Ethan Perez. 2023a. Inverse scaling prize: Second round winners." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 421, + 524, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 524, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 524, + 532 + ], + "type": "text", + "content": "Ian R. McKenzie, Alexander Lyzhov, Michael Pieler, Alicia Parrish, Aaron Mueller, Ameya Prabhu, Euan McLean, Aaron Kirtland, Alexis Ross, Alisa Liu, Andrew Gritsevskiy, Daniel Wurgaft, Derik Kauffman, Gabriel Recchia, Jiacheng Liu, Joe Cavanagh, Max Weiss, Sicong Huang, The Floating Droid, Tom Tseng, Tomasz Korbak, Xudong Shen, Yuhui Zhang, Zhengping Zhou, Najoung Kim, Samuel R. Bowman, and Ethan Perez. 2023b. Inverse Scaling: When Bigger Isn't Better." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 540, + 524, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 524, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 524, + 606 + ], + "type": "text", + "content": "James Michaelov and Benjamin Bergen. 2023. Rarely a problem? Language models exhibit inverse scaling in their predictions following few-type quantifiers. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 14162-14174, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 614, + 524, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 614, + 524, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 614, + 524, + 647 + ], + "type": "text", + "content": "Ethan Perez, Ian McKenzie, and Sam Bowman. 2022. Announcing the Inverse Scaling Prize ($250k Prize Pool)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 655, + 524, + 766 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 655, + 524, + 766 + ], + "spans": [ + { + "bbox": [ + 304, + 655, + 524, + 766 + ], + "type": "text", + "content": "Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Hennigan, Jacob Menick, Albin Cassirer, Richard Powell, George van den Driessche, Lisa Anne Hendricks, Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron Huang, Jonathan Uesato, John Mellor, Irina Higgins, Antonia Creswell, Nat McAleese, Amy Wu, Erich Olsen," + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14613" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 79, + 72, + 291, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 72, + 291, + 281 + ], + "spans": [ + { + "bbox": [ + 79, + 72, + 291, + 281 + ], + "type": "text", + "content": "Siddhant Jayakumar, Elena Buchatskaya, David Budden, Esme Sutherland, Karen Simonyan, Michela Paganini, Laurent Sifre, Lena Martens, Xiang Lorraine Li, Adhiguna Kuncoro, Aida Nematzadeh, Elena Gribovskaya, Domenic Donato, Angeliki Lazaridou, Arthur Mensch, Jean-Baptiste Lespiau, Maria Tsimpoukelli, Nikolai Grigorev, Doug Fritz, Thibault Sotiaux, Mantas Pajarskas, Toby Pohlen, Zhitao Gong, Daniel Toyama, Cyprien de Masson d'Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik, Igor Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury, Matthew Johnson, Blake Hechtman, Laura Weidinger, Jason Gabriel, William Isaac, Ed Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem Ayoub, Jeff Stanway, Lorrayne Bennett, Demis Hassabis, Koray Kavukcuoglu, and Geoffrey Irving. 2022. Scaling Language Models: Methods, Analysis & Insights from Training Gopher." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 289, + 291, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 289, + 291, + 346 + ], + "spans": [ + { + "bbox": [ + 69, + 289, + 291, + 346 + ], + "type": "text", + "content": "Deborah Raji, Emily Denton, Emily M. Bender, Alex Hanna, and Amandalynne Paullada. 2021. AI and the Everything in the Whole Wide World Benchmark. Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 352, + 291, + 770 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 352, + 291, + 770 + ], + "spans": [ + { + "bbox": [ + 69, + 352, + 291, + 770 + ], + "type": "text", + "content": "Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R. Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, Agnieszka Kluska, Aitor Lewkowycz, Akshit Agarwal, Alethea Power, Alex Ray, Alex Warstadt, Alexander W. Kocurek, Ali Safaya, Ali Tazarv, Alice Xiang, Alicia Parrish, Allen Nie, Aman Hussain, Amanda Askell, Amanda Dsouza, Ambrose Slone, Ameet Rahane, Anantharaman S. Iyer, Anders Andreassen, Andrea Madotto, Andrea Santilli, Andreas Stuhlmüller, Andrew Dai, Andrew La, Andrew Lampinen, Andy Zou, Angela Jiang, Angelica Chen, Anh Vuong, Animesh Gupta, Anna Gottardi, Antonio Norelli, Anu Venkatesh, Arash Gholamidavoodi, Arfa Tabassum, Arul Menezes, Arun Kirubarajan, Asher Mullokandov, Ashish Sabharwal, Austin Herrick, Avia Efrat, Aykut Erdem, Ayla Karakaş, B. Ryan Roberts, Bao Sheng Loe, Barret Zoph, Bartlomiej Bojanowski, Batuhan Özyurt, Behnam Hedayatnia, Behnam Neyshabur, Benjamin Inden, Benno Stein, Berk Ekmekci, Bill Yuchen Lin, Blake Howald, Cameron Diao, Cameron Dour, Catherine Stinson, Cedrick Argueta, César Ferri Ramírez, Chandan Singh, Charles Rathkopf, Chenlin Meng, Chitta Baral, Chiyu Wu, Chris Callison-Burch, Chris Waites, Christian Voigt, Christopher D. Manning, Christopher Potts, Cindy Ramirez, Clara E. Rivera, Clemencia Siro, Colin Raffel, Courtney Ashcraft, Cristina Garbacea, Damien Sileo, Dan Garrette, Dan Hendrycks, Dan Kilman, Dan Roth, Daniel Freeman, Daniel Khashabi, Daniel Levy, Daniel Mosegui González, Danielle Perszyk, Danny Hernandez, Danqi Chen, Daphne Ippolito, Dar Gilboa, David Dohan, David Drakard, David Jurrgens, Debajyoti Datta, Deep Ganguli, Denis Emelin, Denis Kleyko, Deniz Yuret, Derek Chen, Derek Tam, Dieuwke Hupkes, Diganta Misra, Dilyar Buzan, Dimitri Coelho Mollo, Diyi Yang, Dong-Ho Lee, Ekate" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 526, + 763 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 763 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 763 + ], + "type": "text", + "content": "rina Shutova, Ekin Dogus Cubuk, Elad Segal, Eleanor Hagerman, Elizabeth Barnes, Elizabeth Donoway, Ellie Pavlick, Emanuele Rodola, Emma Lam, Eric Chu, Eric Tang, Erkut Erdem, Ernie Chang, Ethan A. Chi, Ethan Dyer, Ethan Jerzak, Ethan Kim, Eunice Engefu Manyasi, Evgenii Zheltonozhskii, Fanyue Xia, Fatemeh Siar, Fernando Martínez-Plumed, Francesca Happé, Francois Chollet, Frieda Rong, Gaurav Mishra, Genta Indra Winata, Gerard de Melo, German Kruszewski, Giambattista Parascandolo, Giorgio Mariani, Gloria Wang, Gonzalo Jaimovitch-López, Gregor Betz, Guy Gur-Ari, Hana Galijasevic, Hannah Kim, Hannah Rashkin, Hannaneh Hajishirzi, Harsh Mehta, Hayden Bogar, Henry Shevlin, Hinrich Schütze, Hiromu Yakura, Hongming Zhang, Hugh Mee Wong, Ian Ng, Isaac Noble, Jaap Jumelet, Jack Geissinger, Jackson Kernion, Jacob Hilton, Jæhoon Lee, Jaime Fernández Fisac, James B. Simon, James Koppel, James Zheng, James Zou, Jan Kocón, Jana Thompson, Jared Kaplan, Jarema Radom, Jascha Sohl-Dickstein, Jason Phang, Jason Wei, Jason Yosinski, Jekaterina Novikova, Jelle Bosscher, Jennifer Marsh, Jeremy Kim, Jeroen Taal, Jesse Engel, Jesujoba Alabi, Jiacheng Xu, Jiaming Song, Jillian Tang, Joan Waweru, John Burden, John Miller, John U. Balis, Jonathan Berant, Jörg Frohberg, Jos Rozen, Jose Hernandez-Orallo, Joseph Boudeman, Joseph Jones, Joshua B. Tenenbaum, Joshua S. Rule, Joyce Chua, Kamil Kanclerz, Karen Livescu, Karl Krauth, Karthik Gopalakrishnan, Katerina Ignatyeva, Katja Markert, Kaustubh D. Dhole, Kevin Gimpel, Kevin Omondi, Kory Mathewson, Kristen Chiafullo, Ksenia Shkaruta, Kumar Shridhar, Kyle McDonell, Kyle Richardson, Laria Reynolds, Leo Gao, Li Zhang, Liam Dugan, Lianhui Qin, Lidia Contreras-Ochando, Louis-Philippe Morency, Luca Moschella, Lucas Lam, Lucy Noble, Ludwig Schmidt, Luheng He, Luis Oliveros Colón, Luke Metz, Lütfi Kerem Şenel, Maarten Bosma, Maarten Sap, Maartje ter Hoeve, Maheen Farooqi, Manaal Faruqui, Mantas Mazeika, Marco Baturan, Marco Marelli, Marco Maru, Maria Jose Ramírez Quintana, Marie Tolkiehn, Mario Giulianielli, Martha Lewis, Martin Potthast, Matthew L. Leavitt, Matthias Hagen, Matyás Schubert, Medina Orduna Baitemirova, Melody Arnaud, Melvin McElrath, Michael A. Yee, Michael Cohen, Michael Gu, Michael Ivanitskiy, Michael Starritt, Michael Strube, Michat Swedrowski, Michele Bevilacqua, Michihiro Yasunaga, Mihir Kale, Mike Cain, Mimee Xu, Mirac Suzgun, Mo Tiwari, Mohit Bansal, Moin Aminnaseri, Mor Geva, Mozhdeh Gheini, Mukund Varma T Nanyun Peng,Nathan Chi Nayeon Lee Neta Gur-Ari Krakover Nicholas Cameron,Nicholas Roberts Nick Doiron,Nikita Nangia,Niklas Deckers,Niklas Muennighoff,Nitish Shirish Keskar,Niveditha S.Iyer Noah Constant Noah Fiedel Nuan Wen Oliver ZhangOmar AghaOmar Elbaghdadi Omer Levy Owain Evans Pablo Antonio Moreno Casares Parth Doshi Pascale FungPaul Pu LiangPaul Vicol,Pegah Alipoormolabashi Peiyuan Liao Percy Liang Peter Chang Peter EckersleyPhu Mon Htut Pinyu HwangPiotr Mlkkowski,Piyush Patil,Pouya Pezeshkpour,Priti" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14614" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 79, + 71, + 291, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 71, + 291, + 565 + ], + "spans": [ + { + "bbox": [ + 79, + 71, + 291, + 565 + ], + "type": "text", + "content": "Oli, Qiao zhu Mei, Qing Lyu, Qin lang Chen, Rabin Banjade, Rachel Etta Rudolph, Raefer Gabriel, Rahel Habacker, Ramón Risco Delgado, Raphael Milliere, Rhythm Garg, Richard Barnes, Rif A. Saurous, Riku Arakawa, Robbe Raymaekers, Robert Frank, Rohan Sikand, Roman Novak, Roman Sitelew, Ronan LeBras, Rosanne Liu, Rowan Jacobs, Rui Zhang, Ruslan Salakhutdinov, Ryan Chi, Ryan Lee, Ryan Stovall, Ryan Teehan, Ryan Yang, Sahib Singh, Saif M. Mohammad, Sajant Anand, Sam Dillavou, Sam Shleifer, Sam Wiseman, Samuel Gruetter, Samuel R. Bowman, Samuel S. Schoenholz, Sanghyun Han, Sanjeev Kwatra, Sarah A. Rous, Sarik Ghazarian, Sayan Ghosh, Sean Casey, Sebastian Bischoff, Sebastian Gehrmann, Sebastian Schuster, Sepideh Sadeghi, Shadi Hamdan, Sharon Zhou, Shashank Srivastava, Sherry Shi, Shikhar Singh, Shima Asaadi, Shixiang Shane Gu, Shubh Pachchigar, Shubham Toshiwal, Shyam Upadhyay, Shyamolima, Debnath, Siamak Shakeri, Simon Thormeyer, Simone Melzi, Siva Reddy, Sneha Priscilla Makini, Soo-Hwan Lee, Spencer Torene, Sriharsha Hatwar, Stanislas Dehaene, Stefan Divic, Stefano Ermon, Stella Biderman, Stephanie Lin, Stephen Prasad, Steven T. Pi-antadosi, Stuart M. Shieber, Summer Misherghi, Svetlana Kiritchenko, Swaroop Mishra, Tal Linzen, Tal Schuster, Tao Li, Tao Yu, Tariq Ali, Tatsu Hashimoto, Te-Lin Wu, Theo Desbordes, Theodore Rothschild, Thomas Phan, Tianle Wang, Tiberius Nkinyili, Timo Schick, Timofei Kornev, Timothy Telleen-Lawton Titus Tunduny, Tobias Gerstenberg, Trenton Chang Trishala Neeraj, Tushar Khot, Tyler Shultz, Uri Shaham, Vedant Misra, Vera Demberg, Victoria Nyamai Vikas Raunak Vinay Ramasesh Vinay Uday Prabhu Vishakh Padmakumar,Vivek Srikumar William FedusWilliam SaundersWilliam ZhangWout Vossen Xiang RenXiaoyu TongXinran ZhaoXinyi Wu Xudong Shen,Yadollah YaghoobzadehYair Lakretz Yangqiu Song,Yasaman Bahri,Yejin Choi,Yichi Yang,Yiding Hao,Yifu Chen,Yonatan Belinkov Yu HouYufang HouYuntao BaiZachary Seid Zhuoye Zhao Zijian Wang Zijie J.WangZirui Wang,and Ziyi Wu.2022.Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 573, + 291, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 573, + 291, + 619 + ], + "spans": [ + { + "bbox": [ + 68, + 573, + 291, + 619 + ], + "type": "text", + "content": "Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A Large Language Model for Science." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 626, + 291, + 770 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 626, + 291, + 770 + ], + "spans": [ + { + "bbox": [ + 69, + 626, + 291, + 770 + ], + "type": "text", + "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 526, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 526, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 526, + 161 + ], + "type": "text", + "content": "Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. 2022. LaMDA: Language Models for Dialog Applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 303, + 169, + 526, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 169, + 526, + 236 + ], + "spans": [ + { + "bbox": [ + 303, + 169, + 526, + 236 + ], + "type": "text", + "content": "Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. LLaMA: Open and Efficient Foundation Language Models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 243, + 526, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 243, + 526, + 496 + ], + "spans": [ + { + "bbox": [ + 303, + 243, + 526, + 496 + ], + "type": "text", + "content": "Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open Foundation and Fine-Tuned Chat Models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 504, + 527, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 504, + 527, + 581 + ], + "spans": [ + { + "bbox": [ + 303, + 504, + 527, + 581 + ], + "type": "text", + "content": "Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent Abilities of Large Language Models. Transactions on Machine Learning Research." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "14615" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_content_list.json b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..913aad429a3cff381541fdc0a79fd859835f23f8 --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_content_list.json @@ -0,0 +1,1371 @@ +[ + { + "type": "text", + "text": "Empowering Psychotherapy with Large Language Models: Cognitive Distortion Detection through Diagnosis of Thought Prompting", + "text_level": 1, + "bbox": [ + 124, + 89, + 872, + 129 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyu Chen $^{1}$ , Yujie Lu $^{2}$ , and William Yang Wang $^{2}$", + "bbox": [ + 282, + 141, + 715, + 159 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carnegie Mellon University", + "bbox": [ + 381, + 160, + 618, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2University of California, Santa Barbara", + "bbox": [ + 337, + 175, + 663, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhiyuc@andrew.cmu.edu,", + "bbox": [ + 389, + 193, + 611, + 208 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yujielu, william}@cs.ucsb.edu", + "bbox": [ + 349, + 210, + 653, + 225 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mental illness remains one of the most critical public health issues of our time, due to the severe scarcity and accessibility limit of professionals. Psychotherapy requires high-level expertise to conduct deep, complex reasoning and analysis on the cognition modeling of the patients. In the era of Large Language Models, we believe it is the right time to develop AI assistance for computational psychotherapy. We study the task of cognitive distortion detection and propose the Diagnosis of Thought (DoT) prompting. DoT performs diagnosis on the patient's speech via three stages: subjectivity assessment to separate the facts and the thoughts; contrastive reasoning to elicit the reasoning processes supporting and contradicting the thoughts; and schema analysis to summarize the cognition schemas. The generated diagnosis rationales through the three stages are essential for assisting the professionals. Experiments demonstrate that DoT obtains significant improvements over ChatGPT for cognitive distortion detection, while generating high-quality rationales approved by human experts.", + "bbox": [ + 141, + 279, + 460, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 633, + 258, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "About one in eight people worldwide suffer from mental disorders (World Health Organization, 2022). However, such mental health conditions are severely underserved, due to a number of reasons including the scarcity of mental health professionals, poor quality of services, unaffordable cost, and social stigma (White and Dorman, 2001; Sharma et al., 2020b). Treatment coverage for mental health service use ranges from $33\\%$ in high-income locations to only $8\\%$ in low- and lower middle-income countries (Moitra et al., 2022). A recent study from the American Psychological Association (APA) $^{1}$ found that six in ten psychologists \"no longer have openings for new patients.\"", + "bbox": [ + 112, + 658, + 489, + 884 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate such situations, there have been consistent efforts on developing automated systems for mental health support, such as sentiment analysis (Rathje et al., 2023) and empathetic chatbots (Welivita et al., 2021; Sharma et al., 2020b; Saha et al., 2022). However, existing works mostly take shallow attempts in a heuristic manner, e.g., analyzing emotions and generating comforting responses. There is still a significant gap for such systems to contribute to real professional psychotherapy, which requires deep studies of the patient's thinking patterns, the establishment of cognition models, and the methods to reconstruct the cognition models. These procedures constitute the core pillars in common classic therapy paradigms like cognitive-behavior therapy (CBT) (Rothbaum et al., 2000; Wright et al., 2017; Beck, 2020) and acceptance and commitment therapy (ACT) (Harris, 2006; Hayes et al., 2011). Most data sources recording the interactions between the patients and licensed professionals are confidential, making it even more challenging to build professional assistance for psychotherapy.", + "bbox": [ + 507, + 253, + 884, + 624 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in Large Language Model (LLM) development uncovers its astonishing ability in various textual reasoning tasks in zero-shot setting (Kojima et al., 2022; Bang et al., 2023; Chen, 2022). For the psychology domain, ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. We believe it is promising to further exploit such ability to complex mental reasoning and analysis. It is the right time to start developing professional, targeted, and systematic AI assistance for psychotherapy.", + "bbox": [ + 507, + 626, + 884, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we take the first step by studying the task of cognitive distortion detection, the first core procedure in cognitive behavior therapy (CBT) (Beck, 2020). Inspired by how psychother", + "bbox": [ + 507, + 854, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://www.theguardian.com/society/2022/nov/21/therapist-shortage-us-psychologists-pandemic", + "bbox": [ + 112, + 891, + 504, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4295", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4295-4304 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 216, + 945, + 779, + 972 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/68735549b8e66c1ffcf9bb70c191f9bf5d0ed827d1185b8f5c42ba4713a459e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
", + "bbox": [ + 119, + 83, + 880, + 168 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1: Three example common cognitive distortion types, taken from (Beck, 2020; Shreevastava and Foltz, 2021).", + "bbox": [ + 144, + 175, + 850, + 190 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "apy professionals perform nuanced diagnosis over the patient's speech, we propose the Diagnosis of Thought (DoT) prompting. In DoT, we diagnose the patient's speech through three stages: (1) subjectivity assessment, (2) contrastive reasoning, and (3) schema analysis. In subjectivity assessment, we distinguish the patient's subjective thoughts from the objective facts; In contrastive reasoning, we elicit the reasoning processes supporting and opposing the patient's thoughts; Finally, in schema analysis, we summarize the underlying thought schema and map it to the cognitive distortion types. We conduct comprehensive experiments using the recent top-performing LLMs. In zero-shot setting, DoT obtains over $10\\%$ and $15\\%$ relative improvements for distortion assessment and classification, respectively, on ChatGPT. Meanwhile, the generated rationales through the three stages grant full interpretability for the diagnosis process, whose quality is further approved by human experts.", + "bbox": [ + 115, + 219, + 489, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We unveil the great potential of empowering professional psychotherapy with LLM. This exploration serves as a catalyst for a larger initiative; we extend an invitation to both AI and psychotherapy communities to come together in a collaborative effort. Our ultimate goal is to construct professional, safe, AI-driven assistance that can substantially enhance mental health support systems.", + "bbox": [ + 112, + 543, + 489, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Cognitive Distortion Detection", + "text_level": 1, + "bbox": [ + 112, + 683, + 410, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cognitive Behavior Therapy (CBT) is a well-established therapy paradigm primarily on depression and anxiety disorders (Beck, 2020). In CBT, given the patients' speech or written content, we establish the cognitive model by building the interactions among the situations, thoughts, and emotions. Patients with mental disorders, such as depression or anxiety, tend to form negative thoughts very rapidly and unconsciously, leading to negative emotions which further strengthen their overall negative views and beliefs about the world.", + "bbox": [ + 112, + 709, + 489, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To break this vicious circle, in a typical CBT process, the first core step is to identify those mal", + "bbox": [ + 112, + 887, + 489, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adaptive negative thoughts and summarize their underlying schemas, formally known as cognitive distortions. There are generally 10-20 common, well-studied types of cognitive distortions. We present a few types with examples in Table 1, and refer the readers to Appendix A for the full list. Once accurately identify these cognitive distortions, CBT therapists will guide the patients to justify and correct these distortions, so as to gradually reconstruct their cognition models.", + "bbox": [ + 507, + 219, + 884, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the real psychotherapy process, there's a significant amount of textual information including therapeutic conversations and diaries, etc. Such information is often long, highly fragmented, and disorganized, containing multiple types of distortions beyond the toy examples in Table 1. The task of cognitive distortion detection aims to automatically detect the distortion types given such textual information from the patients, in order to assist therapists to enhance their efficiency and productivity. Meanwhile, such detectors can also potentially serve as self-assisting tools for the patients to diagnose their thoughts and conduct CBT practice, upon meeting the robustness and safety requirements. Formally, cognitive distortion detection consists of two steps: 1) Distortion assessment to predict whether the given speech contains cognitive distortions, as a binary classification problem; and 2) Distortion classification to predict the specific distortion types, as a multiclass classification problem.", + "bbox": [ + 507, + 382, + 884, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 The Diagnosis of Thought Prompting", + "text_level": 1, + "bbox": [ + 507, + 731, + 865, + 749 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On discussing with psychotherapy professionals, we identify the following key stages to diagnose the patient's speech to detect cognitive distortions. Subjectivity Assessment. The patient's speech consists of a mixture of reality (objective facts) and interpretations/opinions (subjective thoughts). In order to perform deep analysis of distorted thinking, we first need to find out which parts of the speech are objective facts and which parts are subjective thoughts. After such an assessment, we summa", + "bbox": [ + 507, + 758, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4296", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b6a1ace8f70193bc2abc3d20531685a261bb01778d2918a60d8e0e4c8cb1da94.jpg", + "image_caption": [ + "Figure 1: The Diagnosis of Thought framework. We strategically prompt the LLM to go through the three diagnosis stages: subjectivity assessment, contrastive reasoning, and schema analysis." + ], + "image_footnote": [], + "bbox": [ + 115, + 82, + 884, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rize the objective facts into the situations as the evidence base to diagnose the subjective thoughts.", + "bbox": [ + 112, + 315, + 485, + 346 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive Reasoning. This stage aims to discover how the patient ascertains the veracity of their subjective thoughts. Based on the situation, we deduct the reasoning processes that supports and contradicts the patient's thoughts respectively. By contrasting two different interpretations based on the same situation, we can identify the thought schemas more clearly.", + "bbox": [ + 112, + 349, + 485, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Schema Analysis. This stage aims to study why the patient forms the specific reasoning process. The term \"schema\" refers to the cognitive structures that organize our knowledge, beliefs, and expectations. Understanding what schemas a patient is relying on can reveal much about their cognitive mode and distortions.", + "bbox": [ + 112, + 480, + 485, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Diagnosis of Thought (DoT) prompting, guiding the LLM through the above three stages to diagnose the patient's speech. Figure 1 illustrates our framework. We sequentially prompt LLM with three questions for the three stages. After the LLM finishes the generation for all stages, we prompt it with another two questions asking distortion assessment and classification. See Appendix B for all the prompts we use.", + "bbox": [ + 112, + 595, + 485, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With DoT prompting, we obtain fully interpretable diagnosis rationales for detecting cognitive distortions. Such interpretability is vital for professionals in real applications. Serving as the diagnosing assistance tool, the diagnosis rationales are essential for the professionals to justify the results. More importantly, they can potentially learn the nuanced thought patterns and schemas derived from the patient's speech through the rationales. This is crucial for the professionals to establish the patient's cognition models more efficiently from", + "bbox": [ + 112, + 741, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the vast amount of speech.", + "bbox": [ + 507, + 315, + 707, + 331 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4 Experimental Results", + "text_level": 1, + "bbox": [ + 507, + 343, + 729, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1 Dataset and experimental settings", + "text_level": 1, + "bbox": [ + 507, + 368, + 820, + 384 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment on the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), which is annotated by experts based on the Therapist QA dataset. The dataset consists of 2,531 examples of patient speech annotated with ten common types of cognitive distortions, as specified in Appendix A. $63.1\\%$ of the examples have cognitive distortions, which are annotated with the two dominant ones. We follow the same train-test split in (Shreevastava and Foltz, 2021). For automatic evaluations, we report F-1 for distortion assessment and weighted F-1 for distortion classification.", + "bbox": [ + 507, + 388, + 882, + 581 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2 Automatic Evaluation Results", + "text_level": 1, + "bbox": [ + 507, + 592, + 789, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We experiment using three of the recent representative LLMs: ChatGPT (gpt-3.5-turbo)3, Vicuna4, and GPT-4 (OpenAI, 2023). For all methods, we first prompt with a general instruction specifying the task and the target distortion types. We compare our DoT prompting with 1) Directly generating the results, and 2) Zero-Shot CoT prompting (ZCoT) (Kojima et al., 2022). For Vicuna, both ZCoT and DoT exceed the token limits for many examples; we omit the results. We average over five runs and report mean and standard deviation for all experiments. Table 2 shows our main experiment results. As we can see, under the zero-shot setting, due to the challenge of this task, Vicuna and ChatGPT still fall behind the supervised full-training models. The proposed DoT prompting significantly", + "bbox": [ + 507, + 612, + 882, + 871 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.kaggle.com/arnmaud/therapist-qa", + "bbox": [ + 529, + 878, + 821, + 891 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://openai.com/blog/chatgpt", + "bbox": [ + 532, + 891, + 732, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4https://lmsys.org/blog/2023-03-30-vicuna/", + "bbox": [ + 532, + 904, + 796, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4297", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/64b37794b7fcbb3ddd54ace693201c914e625173dd0f70567e5eaf67107a2eda.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
Full training*75.024.0
Vicuna73.810.9511.230.78
ChatGPT73.470.5819.241.00
ChatGPT + ZCoT77.101.2120.211.02
ChatGPT + DoT81.190.1122.250.70
GPT-483.040.5133.860.83
GPT-4 + ZCoT81.971.2133.221.36
GPT-4 + DoT82.770.8134.641.40
", + "bbox": [ + 127, + 80, + 470, + 263 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/796c6a883fb89edc2317fb52a0a02e7fab62c16ecce1ffaa306ff36d821c5d64.jpg", + "table_caption": [ + "Table 2: Main Results (Standard deviation in subscript). $(^{\\star})$ results copied from (Shreevastava and Foltz, 2021)." + ], + "table_footnote": [], + "table_body": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
ChatGPT73.470.5819.241.00
ChatGPT + S179.621.1218.721.95
ChatGPT + S1 + S280.700.4820.111.02
ChatGPT + S1 + S2 + S381.190.1122.250.70
", + "bbox": [ + 115, + 317, + 482, + 423 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "boosts the performance of ChatGPT, with the distortion assessment score surpassing the full-training performance and the distortion classification score approaching the full-training performance. The improvement over ZCoT demonstrates the superiority of our proposed strategic stages compared with rationale generation with no guidance. Astonishingly, GPT-4 surpasses the full-training performance even by a large margin for distortion classification. On the one hand, this may demonstrate the powerful ability of GPT-4 on this task - it can already reach a potential upper bound regardless with DoT promoting or not; On the other hand, since the dataset we use is publicly released on Kaggle and its original data source is from online forums, we do not exclude the possibility of data contamination.", + "bbox": [ + 112, + 500, + 489, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ablation Studies and Analysis. To better understand how each stage of DoT contributes to overall performance, we conduct ablation studies using ChatGPT shown in Table 3. All three stages obtain improvements for distortion assessment and distortion classification. For distortion assessment, subjectivity assessment (S1) gains the largest improvement, as separating objective facts and subjective thoughts is a strong trigger to assess distortions. For distortion classification, Schema Analysis (S3)", + "bbox": [ + 112, + 758, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6701a369fe9e975803d3a626fd263baf4b30b1ca5278a00d974753f5528791b1.jpg", + "image_caption": [ + "Figure 2: Distortion classification per class F-1." + ], + "image_footnote": [], + "bbox": [ + 519, + 80, + 870, + 212 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/50a8461a152e245735f7ec185b121a4fde61760020939c2c9402859e11010a35.jpg", + "table_caption": [ + "Table 3: Ablation studies on each stage of DoT: we denote Subjectivity Assessment as S1, Contrastive Reasoning as S2, and Schema Analysis as S3." + ], + "table_footnote": [], + "table_body": "
MethodsQualitySubjectivity AssessmentContrastive ReasoningSchema Analysis
ChatGPT +DoTComprehensive68.5%45.0%50.5%
Partially good24.5%30.5%27.0%
Invalid7.0%24.5%22.5%
GPT4 +DoTComprehensive84.5%69.5%76.5%
Partially good13.5%25.5%18.5%
Invalid2.0%5.0%5.0%
", + "bbox": [ + 510, + 247, + 878, + 374 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4: Human evaluation results for diagnosis rationales.", + "bbox": [ + 514, + 382, + 873, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "obtains large improvement, as the summarized underlying thought schema can match the definition of the distortion types more accurately. We also analyze the distortion classification results for each distortion type, with results shown in Figure 2.", + "bbox": [ + 507, + 428, + 882, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3 Human Evaluation Results", + "text_level": 1, + "bbox": [ + 507, + 519, + 766, + 533 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As there is no reference available for the diagnosis rationales, we employ psychotherapy experts to assess the quality of the generated rationales. We hired psychotherapy professionals from UpWork5, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate.", + "bbox": [ + 505, + 541, + 884, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we present the patient's speech, all the prompts, and the generated rationales for all stages to human experts. For the generated rationales of each stage, we instruct the experts to choose between: 1) Comprehensive. (Correct and comprehensive.); 2) Partially good. (Reasonable but not comprehensive) 3) Invalid. (Not reasonable.) Table 4 shows the evaluation results on 100 examples for DoT over ChatGPT and GPT-4. We employ two experts for each evaluation and report the average; The agreement rates (the percentage of examples the two experts gave the same rating) for both evaluations are over $80\\%$ . The diagnosis rationales generated for all stages shows decent", + "bbox": [ + 507, + 670, + 882, + 894 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "5www.upwork.com", + "bbox": [ + 529, + 903, + 653, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4298", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quality verified by the experts. See Appendix C for more generation examples.", + "bbox": [ + 112, + 84, + 487, + 116 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 114, + 128, + 268, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the verbal and textual nature of psychotherapy procedures, there have been consistent efforts to employ NLP techniques to assist the mental health domain (Althoff et al., 2016; Abd-Alrazaq et al., 2021, 2019; Valizadeh and Parde, 2022). However, with the knowledge gap between the two communities, most existing works take shallow attempts without deep investigation of the professional psychotherapy knowledge. A majority of the previous studies targets identification of common mental health issues, such as depression and anxiety, from textual contents (Cohan et al., 2018; MacAvaney et al., 2018; Harrigian et al., 2020; Ji et al., 2022; Zanwar et al., 2023; Juhng et al., 2023). Another major category of works study therapeutic conversations with a focus on emotional/empathetic supports (Halder et al., 2017; Sharma et al., 2020a; Atapattu et al., 2022; Mishra et al., 2023), and discourse structures (Cao et al., 2019; Hsu et al., 2023). A few more recent works have started to investigate deeper professional psychotherapy knowledge, e.g., cognitive distortion detection in CBT (Shreevastava and Foltz, 2021; Ding et al., 2022; Lybarger et al., 2022), identifying and reframing unhelpful thoughts (Ziems et al., 2022; Maddela et al., 2023). Early famous system Eliza6 took the initial attempts to emulate a Rogerian psychotherapist. Due to its rule-based nature, the responses are often reflections or rephrasings of the user's statements. If users deviate too much from expected inputs or probe its capabilities, the program may produce irrelevant or nonsensical responses.", + "bbox": [ + 112, + 154, + 489, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent progress in large language model reveals that ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. However, latter works also point out the robustness issue of such ability (Shapira et al., 2023). In our work, we are inspired by such enhanced ability and investigate the application of diagnosing patients' thoughts in psychotherapy. Our experiments on the real dataset, not anecdotal examples, show very promising performance both in auto", + "bbox": [ + 112, + 686, + 489, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "matic and human expert evaluation. We believe it is a very important future research direction to investigate the general cognitive abilities of LLM and make connections to the field of psychology and cognitive science. Theory-wise, we are eager to explore to what extent the LLM can simulate the human cognitive functions, so as to determine the role that language plays in the overall human cognition. Application-wise, a straightforward and encouraging direction should be the integration for assisting mental health treatment, as the motivation and goal of this work.", + "bbox": [ + 507, + 84, + 884, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 507, + 290, + 796, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This work delved into the integration of large language models (LLMs) within the realm of psychotherapy. We introduced the Diagnosis of Thought (DoT) framework, which strategically prompts the LLM to produce diagnosis rationales pertinent to the detection of cognitive distortions. We believe that there exists substantial potential for leveraging LLMs within numerous facets of mental health support that are currently under-explored. Our findings, thus, not only illuminate a path towards more efficient therapy methods but also open doors for future investigations to push the boundaries of AI's role in mental health treatment.", + "bbox": [ + 507, + 316, + 884, + 524 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the domain of AI-enhanced psychotherapy, particularly when harnessing the power of LLMs, there are paramount challenges surrounding safety, robustness, and ethical responsibilities. LLM is known to generate hallucinations and biases. While one potential usage of our system is to serve as a self-diagnosing/assisting tool for patients, e.g., to help the patients to recognize their own thinking patterns and learn how the cognitive distortions developed, such tools should not be deployed for direct patient use without the supervision of professionals. As we navigate the potential of LLM in psychotherapy, our foremost priority should remain in building systems that are safe and responsible. Concurrently, the formulation of robust ethical guidelines tailored to this new era becomes indispensable.", + "bbox": [ + 507, + 526, + 884, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 509, + 812, + 613, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the biggest obstacles to studying AI for psychotherapy is the lack of available, high-quality data. Most of the datasets recording the information of the interactions between the patients and licensed professionals are confidential due to pri", + "bbox": [ + 507, + 839, + 884, + 919 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://en.wikipedia.org/wiki/ELIZA", + "bbox": [ + 134, + 903, + 368, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4299", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vacy concerns. The dataset used in this work is the only one we can find publicly available for cognitive distortion detection, as its data source is from public online forums. In addition to the patient's speech, the patient's demographic information also plays an important role in analyzing the thought process leading to cognitive distortions. In this dataset, some of the patient speech include a few demographic details, but pretty minimal. Beyond the task of cognitive distortion detection, the same data constraint issue exists for all other therapy stages and paradigms. This also motivates us to work on building privacy-preserved systems as a crucial future direction.", + "bbox": [ + 112, + 84, + 489, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 323, + 265, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The system built in this work are never meant to replace psychotherapy professionals: all system-produced results need to be verified by licensed professionals. The system should not be deployed for direct patient use without the supervision of licensed professionals.", + "bbox": [ + 112, + 350, + 487, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There has been a lot of discussions regarding the use of LLMs for AI-driven therapy purpose. Chatbots will find their most effective role in medicine within the realm of mental health, an opinion quoted from Thomas Insel, former director of the National Institute of Mental Health. In the meantime, we should always take caution for such development. Our foremost priority should be building safe and responsible applications while establishing ethical and regulation standards for this new era. Under the current status, the partnership between professionals and AI systems is a promising direction for the goal to ease the burden of professionals, which is the motivation of this work.", + "bbox": [ + 115, + 448, + 489, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"Diagnosis of Thought\" is an innovative approach that harnesses the capabilities of large language models to enhance the field of psychotherapy. It is imperative to understand that this tool is not designed for censorship or any form of invasive surveillance. Instead, it aims to augment the skills of therapists by providing them with advanced, data-driven insights into their clients' thoughts and emotions. Through the efficient processing of linguistic cues and patterns, these models aid mental health professionals in formulating more precise diagnoses and crafting tailored therapeutic interventions that could significantly enhance the well", + "bbox": [ + 112, + 673, + 489, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "being of individuals seeking help.", + "bbox": [ + 507, + 84, + 761, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all the annotations and human evaluations presented in this work, we hired psychotherapy professionals from UpWork8, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate. The average hourly rate for all the experts is $80.", + "bbox": [ + 507, + 101, + 884, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This project is approved by our Institutional Review Board (IRB). The data annotation is classified as exempt by our IRB.", + "bbox": [ + 507, + 214, + 882, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 290, + 608, + 305 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alaa A Abd-Alrazaq, Mohannad Alajlani, Ali Abdallah Alalwan, Bridgette M Bewick, Peter Gardner, and Mowafa Househ. 2019. An overview of the features of chatbots in mental health: A scoping review. International Journal of Medical Informatics, 132:103978.", + "Alaa A Abd-Alrazaq, Mohammad Alajlani, Nashva Ali, Kerstin Denecke, Bridgette M Bewick, and Mowafa Househ. 2021. Perceptions and opinions of patients about mental health chatbots: scoping review. Journal of medical Internet research, 23(1):e17828.", + "Tim Althoff, Kevin Clark, and Jure Leskovec. 2016. Large-scale analysis of counseling conversations: An application of natural language processing to mental health. Transactions of the Association for Computational Linguistics, 4:463-476.", + "Thushari Atapattu, Mahen Herath, Charitha Elvitigala, Piyanjali de Zoysa, Kasun Gunawardana, Menasha Thilakaratne, Kasun de Zoysa, and Katrina Falkner. 2022. EmoMent: An emotion annotated mental health corpus from two South Asian countries. In Proceedings of the 29th International Conference on Computational Linguistics, pages 6991-7001, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.", + "Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wen-liang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023.", + "Simon Baron-Cohen, Alan M Leslie, and Uta Frith. 1985. Does the autistic child have a \"theory of mind\"? Cognition, 21(1):37-46.", + "Judith S Beck. 2020. Cognitive behavior therapy: Basics and beyond. Guilford Publications.", + "Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar," + ], + "bbox": [ + 509, + 313, + 885, + 895 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "7https://www.scientificamerican.com/article/ai-chatbots-could-help-provide-therapy-but-caution-is-needed/", + "bbox": [ + 112, + 891, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{8}$ www.upwork.com", + "bbox": [ + 531, + 903, + 653, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4300", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712.", + "Jie Cao, Michael Tanana, Zac E Imel, Eric Poitras, David C Atkins, and Vivek Srikumar. 2019. Observing dialogue in therapy: Categorizing and forecasting behavioral codes. arXiv preprint arXiv:1907.00326.", + "Wenhu Chen. 2022. Large language models are few (1)-shot table reasoners. arXiv preprint arXiv:2210.06710.", + "Arman Cohan, Bart Desmet, Andrew Yates, Luca Soldaini, Sean MacAvaney, and Nazli Goharian. 2018. SMHD: a large-scale resource for exploring online language usage for multiple mental health conditions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1485-1497, Santa Fe, New Mexico, USA. Association for Computational Linguistics.", + "Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. Improving classification of infrequent cognitive distortions: Domain-specific model vs. data augmentation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 68-75, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics.", + "Kishaloy Halder, Lahari Poddar, and Min-Yen Kan. 2017. Modeling temporal progression of emotional status in mental health forum: A recurrent neural net approach. In Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 127-135, Copenhagen, Denmark. Association for Computational Linguistics.", + "Keith Harrigian, Carlos Aguirre, and Mark Dredze. 2020. Do models of mental health based on social media data generalize? In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3774-3788, Online. Association for Computational Linguistics.", + "Russell Harris. 2006. Embracing your demons: An overview of acceptance and commitment therapy. Psychotherapy in Australia, 12(4):70-6.", + "Steven C Hayes, Kirk D Strosahl, and Kelly G Wilson. 2011. Acceptance and commitment therapy: The process and practice of mindful change. Guilford press.", + "Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, Zahra Ashktorab, Casey Dugan, Werner Geyer, and Diyi Yang. 2023. Helping the helper: Supporting peer counselors via ai-empowered practice and feedback. arXiv preprint arXiv:2305.08982." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shaoxiong Ji, Tianlin Zhang, Luna Ansari, Jie Fu, Prayag Tiwari, and Erik Cambria. 2022. MentalBERT: Publicly available pretrained language models for mental healthcare. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 7184-7190, Marseille, France. European Language Resources Association.", + "Swanie Juhng, Matthew Matero, Vasudha Varadarajan, Johannes Eichstaedt, Adithya V Ganesan, and H Andrew Schwartz. 2023. Discourse-level representations can improve prediction of degree of anxiety. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1500-1511.", + "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916.", + "Kevin Lybarger, Justin Tauscher, Xiruo Ding, Dror Ben-Zeev, and Trevor Cohen. 2022. Identifying distorted thinking in patient-therapist text message exchanges by leveraging dynamic multi-turn context. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pages 126-136.", + "Sean MacAvaney, Bart Desmet, Arman Cohan, Luca Soldaini, Andrew Yates, Ayah Zirikly, and Nazli Goharian. 2018. RSDD-time: Temporal annotation of self-reported mental health diagnoses. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 168-173, New Orleans, LA. Association for Computational Linguistics.", + "Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. arXiv preprint arXiv:2307.02768.", + "Kshitij Mishra, Priyanshu Priya, and Asif Ekbal. 2023. PAL to lend a helping hand: Towards building an emotion adaptive polite and empathetic counseling conversational agent. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12254-12271, Toronto, Canada. Association for Computational Linguistics.", + "Modhurima Moitra, Damian Santomauro, Pamela Y Collins, Theo Vos, Harvey Whiteford, Shekhar Saxena, and Alize J Ferrari. 2022. The global gap in treatment coverage for major depressive disorder in 84 countries from 2000-2019: A systematic review and bayesian meta-regression analysis. PLoS Medicine, 19(2):e1003901.", + "OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.", + "Steve Rathje, Dan-Mircea Mirea, Ilia Sucholutsky, Raja Marjieh, Claire Robertson, and Jay J Van Bavel. 2023." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4301", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gpt is an effective tool for multilingual psychological text analysis.", + "bbox": [ + 134, + 85, + 485, + 112 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Barbara Olasov Rothbaum, Elizabeth A Meadows, Patricia Resick, and David W Foy. 2000. Cognitive-behavioral therapy.", + "Tulika Saha, Saichethan Reddy, Anindya Das, Sriparna Saha, and Pushpak Bhattacharyya. 2022. A shoulder to cry on: Towards a motivational virtual assistant for assuaging mental agony. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2436-2449.", + "Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763.", + "Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020a. A computational approach to understanding empathy expressed in text-based mental health support. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5263-5276, Online. Association for Computational Linguistics.", + "Ashish Sharma, Adam S Miner, David C Atkins, and Tim Althoff. 2020b. A computational approach to understanding empathy expressed in text-based mental health support. arXiv preprint arXiv:2009.08441.", + "Sagarika Shreevastava and Peter Foltz. 2021. Detecting cognitive distortions from patient-therapist interactions. In Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access, pages 151-158.", + "Mina Valizadeh and Natalie Parde. 2022. The ai doctor is in: A survey of task-oriented dialogue systems for healthcare applications. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6638-6660.", + "Anuradha Welivita, Yubo Xie, and Pearl Pu. 2021. A large-scale dataset for empathetic response generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1251-1264.", + "Marsha White and Steve M Dorman. 2001. Receiving social support online: implications for health education. Health education research, 16(6):693-707.", + "Jesse H Wright, Gregory K Brown, Michael E Thase, and Monica Ramirez Basco. 2017. Learning cognitive-behavior therapy: An illustrated guide. American Psychiatric Pub.", + "Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion" + ], + "bbox": [ + 115, + 122, + 487, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Findings of the Association for Computational Linguistics: ACL 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics.", + "Caleb Ziems, Minzhi Li, Anthony Zhang, and Diyi Yang. 2022. Inducing positive perspectives with text reframing. arXiv preprint arXiv:2204.02952." + ], + "bbox": [ + 510, + 85, + 880, + 187 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Appendix A: Cognitive Distortion Types", + "text_level": 1, + "bbox": [ + 510, + 198, + 857, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We list all the ten cognitive distortion types studies in this work following-(Shreevastava and Foltz, 2021) in Table 5. Note that these ten types being studied are some common, well-studied ones. As psychotherapy develops, more fine-grained and new distortion types are expected to be unveiled.", + "bbox": [ + 510, + 223, + 880, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), there are 2,531 examples in total. We follow the original $80\\% - 20\\%$ train-test split. The input patient speech has an average length of 167.3 tokens, and the ten distortion types are roughly equally distributed.", + "bbox": [ + 510, + 321, + 880, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Appendix B: Prompt Details", + "text_level": 1, + "bbox": [ + 510, + 426, + 757, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For all our experiments without using DoT, we first prompt the LLM with the following general instructions:", + "bbox": [ + 510, + 453, + 880, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a speech of a patient, our task is to 1) identify if there is cognitive distortion in the speech; 2) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021).", + "bbox": [ + 510, + 501, + 880, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For all our experiments using DoT, we first prompt the LLM with the following general instructions:", + "bbox": [ + 510, + 614, + 880, + 658 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given a speech of a patient, our task is to 1) finish a few diagnose of thought questions to analyze the thought patterns of the patient. Then based on the diagnose of thought analysis, 2) identify if there is cognitive distortion in the speech; 3) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021).", + "bbox": [ + 510, + 662, + 880, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Then for methods without using DoT, we prompt the following two questions:", + "bbox": [ + 510, + 822, + 880, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Please first answer: if there is cognitive distortion in the speech; Answer 'yes' or 'no'; Please then answer: Recognizing the specific types of the cognitive distortion in the speech. There may be", + "bbox": [ + 510, + 854, + 880, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4302", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f982077edf97013885339044a1dbffa60ad42ea5b778adf88bb02953ea7c2f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
OvergeneralizationMajor conclusions are drawn based on limited information.Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
Emotional reasoningLetting one's feeling about something overrule facts to the contrary.Even though Steve is here at work late every day, I know I work harder than anyone else at my job.
LabelingGiving someone or something a label without finding out more about it/them.My daughter would never do anything I disapproved of.
MagnificationEmphasizing the negative or playing down the positive of a situation.My professor said he made some corrections on my paper, so I know I'll probably fail the class.
Mental filterPlacing all one's attention o, or seeing only, the negatives of a situation.My husband says he wishes I was better at housekeeping, so I must be a lousy wife.
Should statementsShould statements appear as a list of ironclad rules about how a person should behave, this could be about the speaker themselves or other.It is NOT necessary that the word 'should' or it's synonyms (ought to, must etc.) be present in the statements containing this distortion.I should get all A's to be a good student.
Fortune-tellingAs the name suggests, this distortion is about expecting things to happen a certain way, or assuming that thing will go badlyCounterintuitively, this distortion does not always have future tense.I was afraid of job interviews so I decided to start my own thing.
", + "bbox": [ + 117, + 83, + 878, + 325 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 5: Common cognitive distortion types and example speech, taken from (Beck, 2020; Shreevastava and Foltz, 2021).", + "bbox": [ + 127, + 332, + 865, + 347 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "one type of cognitive distortion or multiple types involved. If there are multiple types, please give the top 2 dominant ones. Please only give the distortion type names separated by comma.", + "bbox": [ + 112, + 378, + 487, + 441 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "For methods using DoT, before asking this two questions, we provide the following DoT prompts: Based on the patient's speech, finish the following diagnosis of thought questions: 1. what is the situation? Find out the facts that are objective; what is the patient thinking or imagining? Find out the thoughts or opinions that are subjective. 2. what makes the patient think the thought is true or is not true? Find out the reasoning processes that support and do not support these thoughts. 3. why does the patient come up with such reasoning process supporting the thought? What's the underlying cognition mode of it?", + "bbox": [ + 112, + 443, + 487, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Appendix C: Generation Examples", + "text_level": 1, + "bbox": [ + 114, + 663, + 420, + 680 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We present some generation examples of GPT-3.5 and GPT-4 in Figure 3.", + "bbox": [ + 112, + 688, + 485, + 720 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4303", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Patient Message: Lately I have been extremely unhappy for no apparent reason. I feel empty almost all the time. It is like nothing can give me genuine joy anymore. I also feel alone a lot, even though I do have friends and close friends. Whenever I try to imagine my future all I see is one full of obstacles. I have become so consumed in sadness and I do not know how to deal anymore. I feel so confused and I do not understand myself anymore.", + "bbox": [ + 129, + 193, + 860, + 223 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/40abdb6121961be0f103ec15bb22af662c36468675c95535ffd652493f4385aa.jpg", + "table_caption": [ + "Diagnosis of Thought (ChatGPT):" + ], + "table_footnote": [], + "table_body": "
Objective Fact:The patient has been feeling extremely unhappy and empty lately.
Subjective Thought:Nothing can give the patient genuine joy and that the future is full of obstacles. the patient feels confused and does not understand themselves anymore.
Reasoning Process Supporting the Thoughts:The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.
Reasoning Process Contradicting the Thoughts:There may not be factual evidence to support the belief that nothing can give the patient genuine joy or that the future is full of obstacles. the patient may be using emotional reasoning to support their beliefs.
Cognition Schema:The patient may be experiencing a cognitive mode of negativity bias, where they tend to focus on negative experiences and thoughts rather than positive ones. this may be influencing their reasoning processes, leading them to magnify the negative and filter out the positive aspects of their life and future. additionally, the patient may be struggling with depression, which can impact the way they perceive themselves and the world around them.
", + "bbox": [ + 127, + 254, + 477, + 456 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/717af0f76ecb1adc154c09e3d9bc4a5505c5aafc8f5d95b785a5537082a34601.jpg", + "table_caption": [ + "Diagnosis of Thought (GPT4):" + ], + "table_footnote": [], + "table_body": "
Objective Fact:The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.
Subjective Thought:They believe nothing can bring them genuine joy anymore. They imagine their future full of obstacles. They feel consumed by sadness and confusion and no longer understand themselves.
Reasoning ProcessThe persistent feeling of emptiness and lack of joy might make the patient believe that their situation is unchanging and will remain the same. They may be experiencing negative events in their lives, making it harder to feel hopeful about their future.
Supporting the Thoughts:Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.
Reasoning ProcessEmotions and feelings are temporary and can change over time, so it is not guaranteed that the patient will always feel this way. The patient is aware that they do have friends, which suggests that some level of meaningful connection exists. Obstacles are a natural part of life, but they are not the only things present in one's future. positive events and achievements are also likely to occur.
Contradicting the Thoughts:
Cognition Schema:Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life's reality. Mental filter: they are focusing on the negative aspects of their life and filtering out positive experiences or memories. Fortune-telling: they are assuming that their future will be full of obstacles based on their current emotions and thoughts.
", + "bbox": [ + 507, + 254, + 860, + 473 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Patient Message: So I have no close friends at the moment probably I know some people enough just to say hi but no close one and whenever I make any they go away I don't know if its my personality the reason or the way I behave and the only friends I can gain are school friends since there is no other place I go to I can't really stand myself alone. Thanks for helping.", + "bbox": [ + 129, + 498, + 863, + 529 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/06996da8ebd71b31ea6ec23a611ea9ac992024be031636ce847cb1d2a278188d.jpg", + "table_caption": [ + "Diagnosis of Thought (ChatGPT):" + ], + "table_footnote": [], + "table_body": "
Objective Fact:The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.
Subjective Thought:The patient feels that they are the reason for their friends leaving and wonders if it's their personality or behavior causing the issue. they cannot stand being alone.
Reasoning Process Supporting the Thoughts:The patient believes their thoughts to be true because they have not been able to maintain any close friendships and the pattern keeps repeating itself. they may be using past experiences to support this belief.
Reasoning Process Contradicting the Thoughts:There may be other reasons for the loss of these friendships that the patient is not considering.
Cognition Schema:The patient may be experiencing cognitive distortions such as overgeneralization by assuming that all friendships will end up the same way and mental filtering by focusing only on the negative aspects of their past friendships. the underlying cognition mode could be a negative self-image or low self-esteem.
", + "bbox": [ + 127, + 562, + 477, + 739 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/7bc5968054c6082d595c374e0e54f63da73f7967346390a60b5ac3c233932a66.jpg", + "table_caption": [ + "Diagnosis of Thought (GPT4):" + ], + "table_footnote": [], + "table_body": "
Objective Fact:The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.
Subjective Thought:The patient believes that their personality or behavior might be the reason their friendships don't last, and they are limited in opportunities to make new friends.
Reasoning ProcessThe patient might have experienced failed friendships and noticed a pattern of people leaving their life. this leads them to question if their personality or behavior is the cause.
Supporting the Thoughts:
Reasoning ProcessThe patient has not provided concrete examples or evidence that their personality or behavior is the reason behind the lack of close friendships. it is possible there are external factors or situational circumstances that led to the loss of friendships.
Contradicting the Thoughts:
Cognition Schema:The patient might be engaging in personalization and overgeneralization cognitive distortions. personalization because they seem to be taking the blame for friendships not lasting, assuming it's their fault due to their personality or behavior. overgeneralization because they are making a broad conclusion based on limited information (i.e., the assumption that their personality or behavior is the sole reason for the lack of close friends). this underlying cognition mode could stem from low self-esteem, fear of rejection, or previous negative experiences in social situations.
", + "bbox": [ + 507, + 558, + 860, + 766 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 3: Showcases of Diagnosis of Thought using ChatGPT and GPT4 given the Patient Message.", + "bbox": [ + 193, + 801, + 801, + 815 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4304", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_model.json b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99288d1c7c4c210e814773bf84755a25b9d90091 --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_model.json @@ -0,0 +1,1826 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.126, + 0.09, + 0.873, + 0.13 + ], + "angle": 0, + "content": "Empowering Psychotherapy with Large Language Models: Cognitive Distortion Detection through Diagnosis of Thought Prompting" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.142, + 0.717, + 0.16 + ], + "angle": 0, + "content": "Zhiyu Chen\\(^{1}\\), Yujie Lu\\(^{2}\\), and William Yang Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.161, + 0.619, + 0.177 + ], + "angle": 0, + "content": "Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.177, + 0.665, + 0.193 + ], + "angle": 0, + "content": "2University of California, Santa Barbara" + }, + { + "type": "text", + "bbox": [ + 0.391, + 0.194, + 0.612, + 0.209 + ], + "angle": 0, + "content": "zhiyuc@andrew.cmu.edu," + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.211, + 0.654, + 0.227 + ], + "angle": 0, + "content": "{yujielu, william}@cs.ucsb.edu" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.28, + 0.461, + 0.622 + ], + "angle": 0, + "content": "Mental illness remains one of the most critical public health issues of our time, due to the severe scarcity and accessibility limit of professionals. Psychotherapy requires high-level expertise to conduct deep, complex reasoning and analysis on the cognition modeling of the patients. In the era of Large Language Models, we believe it is the right time to develop AI assistance for computational psychotherapy. We study the task of cognitive distortion detection and propose the Diagnosis of Thought (DoT) prompting. DoT performs diagnosis on the patient's speech via three stages: subjectivity assessment to separate the facts and the thoughts; contrastive reasoning to elicit the reasoning processes supporting and contradicting the thoughts; and schema analysis to summarize the cognition schemas. The generated diagnosis rationales through the three stages are essential for assisting the professionals. Experiments demonstrate that DoT obtains significant improvements over ChatGPT for cognitive distortion detection, while generating high-quality rationales approved by human experts." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.634, + 0.26, + 0.649 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.659, + 0.49, + 0.885 + ], + "angle": 0, + "content": "About one in eight people worldwide suffer from mental disorders (World Health Organization, 2022). However, such mental health conditions are severely underserved, due to a number of reasons including the scarcity of mental health professionals, poor quality of services, unaffordable cost, and social stigma (White and Dorman, 2001; Sharma et al., 2020b). Treatment coverage for mental health service use ranges from \\(33\\%\\) in high-income locations to only \\(8\\%\\) in low- and lower middle-income countries (Moitra et al., 2022). A recent study from the American Psychological Association (APA)\\(^{1}\\) found that six in ten psychologists \"no longer have openings for new patients.\"" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.625 + ], + "angle": 0, + "content": "To mitigate such situations, there have been consistent efforts on developing automated systems for mental health support, such as sentiment analysis (Rathje et al., 2023) and empathetic chatbots (Welivita et al., 2021; Sharma et al., 2020b; Saha et al., 2022). However, existing works mostly take shallow attempts in a heuristic manner, e.g., analyzing emotions and generating comforting responses. There is still a significant gap for such systems to contribute to real professional psychotherapy, which requires deep studies of the patient's thinking patterns, the establishment of cognition models, and the methods to reconstruct the cognition models. These procedures constitute the core pillars in common classic therapy paradigms like cognitive-behavior therapy (CBT) (Rothbaum et al., 2000; Wright et al., 2017; Beck, 2020) and acceptance and commitment therapy (ACT) (Harris, 2006; Hayes et al., 2011). Most data sources recording the interactions between the patients and licensed professionals are confidential, making it even more challenging to build professional assistance for psychotherapy." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.627, + 0.885, + 0.852 + ], + "angle": 0, + "content": "Recent progress in Large Language Model (LLM) development uncovers its astonishing ability in various textual reasoning tasks in zero-shot setting (Kojima et al., 2022; Bang et al., 2023; Chen, 2022). For the psychology domain, ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. We believe it is promising to further exploit such ability to complex mental reasoning and analysis. It is the right time to start developing professional, targeted, and systematic AI assistance for psychotherapy." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.855, + 0.884, + 0.919 + ], + "angle": 0, + "content": "In this paper, we take the first step by studying the task of cognitive distortion detection, the first core procedure in cognitive behavior therapy (CBT) (Beck, 2020). Inspired by how psychother" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.505, + 0.92 + ], + "angle": 0, + "content": "1https://www.theguardian.com/society/2022/nov/21/therapist-shortage-us-psychologists-pandemic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4295" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.78, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4295-4304 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.084, + 0.881, + 0.169 + ], + "angle": 0, + "content": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.176, + 0.851, + 0.191 + ], + "angle": 0, + "content": "Table 1: Three example common cognitive distortion types, taken from (Beck, 2020; Shreevastava and Foltz, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.221, + 0.49, + 0.542 + ], + "angle": 0, + "content": "apy professionals perform nuanced diagnosis over the patient's speech, we propose the Diagnosis of Thought (DoT) prompting. In DoT, we diagnose the patient's speech through three stages: (1) subjectivity assessment, (2) contrastive reasoning, and (3) schema analysis. In subjectivity assessment, we distinguish the patient's subjective thoughts from the objective facts; In contrastive reasoning, we elicit the reasoning processes supporting and opposing the patient's thoughts; Finally, in schema analysis, we summarize the underlying thought schema and map it to the cognitive distortion types. We conduct comprehensive experiments using the recent top-performing LLMs. In zero-shot setting, DoT obtains over \\(10\\%\\) and \\(15\\%\\) relative improvements for distortion assessment and classification, respectively, on ChatGPT. Meanwhile, the generated rationales through the three stages grant full interpretability for the diagnosis process, whose quality is further approved by human experts." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.544, + 0.49, + 0.673 + ], + "angle": 0, + "content": "We unveil the great potential of empowering professional psychotherapy with LLM. This exploration serves as a catalyst for a larger initiative; we extend an invitation to both AI and psychotherapy communities to come together in a collaborative effort. Our ultimate goal is to construct professional, safe, AI-driven assistance that can substantially enhance mental health support systems." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.684, + 0.411, + 0.7 + ], + "angle": 0, + "content": "2 Cognitive Distortion Detection" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.886 + ], + "angle": 0, + "content": "Cognitive Behavior Therapy (CBT) is a well-established therapy paradigm primarily on depression and anxiety disorders (Beck, 2020). In CBT, given the patients' speech or written content, we establish the cognitive model by building the interactions among the situations, thoughts, and emotions. Patients with mental disorders, such as depression or anxiety, tend to form negative thoughts very rapidly and unconsciously, leading to negative emotions which further strengthen their overall negative views and beliefs about the world." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.49, + 0.92 + ], + "angle": 0, + "content": "To break this vicious circle, in a typical CBT process, the first core step is to identify those mal" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.221, + 0.885, + 0.382 + ], + "angle": 0, + "content": "adaptive negative thoughts and summarize their underlying schemas, formally known as cognitive distortions. There are generally 10-20 common, well-studied types of cognitive distortions. We present a few types with examples in Table 1, and refer the readers to Appendix A for the full list. Once accurately identify these cognitive distortions, CBT therapists will guide the patients to justify and correct these distortions, so as to gradually reconstruct their cognition models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.383, + 0.885, + 0.719 + ], + "angle": 0, + "content": "In the real psychotherapy process, there's a significant amount of textual information including therapeutic conversations and diaries, etc. Such information is often long, highly fragmented, and disorganized, containing multiple types of distortions beyond the toy examples in Table 1. The task of cognitive distortion detection aims to automatically detect the distortion types given such textual information from the patients, in order to assist therapists to enhance their efficiency and productivity. Meanwhile, such detectors can also potentially serve as self-assisting tools for the patients to diagnose their thoughts and conduct CBT practice, upon meeting the robustness and safety requirements. Formally, cognitive distortion detection consists of two steps: 1) Distortion assessment to predict whether the given speech contains cognitive distortions, as a binary classification problem; and 2) Distortion classification to predict the specific distortion types, as a multiclass classification problem." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.732, + 0.866, + 0.75 + ], + "angle": 0, + "content": "3 The Diagnosis of Thought Prompting" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.759, + 0.885, + 0.919 + ], + "angle": 0, + "content": "On discussing with psychotherapy professionals, we identify the following key stages to diagnose the patient's speech to detect cognitive distortions. Subjectivity Assessment. The patient's speech consists of a mixture of reality (objective facts) and interpretations/opinions (subjective thoughts). In order to perform deep analysis of distorted thinking, we first need to find out which parts of the speech are objective facts and which parts are subjective thoughts. After such an assessment, we summa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "4296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.083, + 0.885, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.266, + 0.884, + 0.292 + ], + "angle": 0, + "content": "Figure 1: The Diagnosis of Thought framework. We strategically prompt the LLM to go through the three diagnosis stages: subjectivity assessment, contrastive reasoning, and schema analysis." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.316, + 0.486, + 0.347 + ], + "angle": 0, + "content": "rize the objective facts into the situations as the evidence base to diagnose the subjective thoughts." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.35, + 0.487, + 0.478 + ], + "angle": 0, + "content": "Contrastive Reasoning. This stage aims to discover how the patient ascertains the veracity of their subjective thoughts. Based on the situation, we deduct the reasoning processes that supports and contradicts the patient's thoughts respectively. By contrasting two different interpretations based on the same situation, we can identify the thought schemas more clearly." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.481, + 0.487, + 0.592 + ], + "angle": 0, + "content": "Schema Analysis. This stage aims to study why the patient forms the specific reasoning process. The term \"schema\" refers to the cognitive structures that organize our knowledge, beliefs, and expectations. Understanding what schemas a patient is relying on can reveal much about their cognitive mode and distortions." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.596, + 0.487, + 0.74 + ], + "angle": 0, + "content": "We propose the Diagnosis of Thought (DoT) prompting, guiding the LLM through the above three stages to diagnose the patient's speech. Figure 1 illustrates our framework. We sequentially prompt LLM with three questions for the three stages. After the LLM finishes the generation for all stages, we prompt it with another two questions asking distortion assessment and classification. See Appendix B for all the prompts we use." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.489, + 0.919 + ], + "angle": 0, + "content": "With DoT prompting, we obtain fully interpretable diagnosis rationales for detecting cognitive distortions. Such interpretability is vital for professionals in real applications. Serving as the diagnosing assistance tool, the diagnosis rationales are essential for the professionals to justify the results. More importantly, they can potentially learn the nuanced thought patterns and schemas derived from the patient's speech through the rationales. This is crucial for the professionals to establish the patient's cognition models more efficiently from" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.316, + 0.708, + 0.332 + ], + "angle": 0, + "content": "the vast amount of speech." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.344, + 0.73, + 0.36 + ], + "angle": 0, + "content": "4 Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.369, + 0.821, + 0.385 + ], + "angle": 0, + "content": "4.1 Dataset and experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.39, + 0.884, + 0.582 + ], + "angle": 0, + "content": "We experiment on the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), which is annotated by experts based on the Therapist QA dataset. The dataset consists of 2,531 examples of patient speech annotated with ten common types of cognitive distortions, as specified in Appendix A. \\(63.1\\%\\) of the examples have cognitive distortions, which are annotated with the two dominant ones. We follow the same train-test split in (Shreevastava and Foltz, 2021). For automatic evaluations, we report F-1 for distortion assessment and weighted F-1 for distortion classification." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.593, + 0.791, + 0.607 + ], + "angle": 0, + "content": "4.2 Automatic Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.613, + 0.884, + 0.872 + ], + "angle": 0, + "content": "We experiment using three of the recent representative LLMs: ChatGPT (gpt-3.5-turbo)3, Vicuna4, and GPT-4 (OpenAI, 2023). For all methods, we first prompt with a general instruction specifying the task and the target distortion types. We compare our DoT prompting with 1) Directly generating the results, and 2) Zero-Shot CoT prompting (ZCoT) (Kojima et al., 2022). For Vicuna, both ZCoT and DoT exceed the token limits for many examples; we omit the results. We average over five runs and report mean and standard deviation for all experiments. Table 2 shows our main experiment results. As we can see, under the zero-shot setting, due to the challenge of this task, Vicuna and ChatGPT still fall behind the supervised full-training models. The proposed DoT prompting significantly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.879, + 0.822, + 0.892 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www.kaggle.com/arnmaud/therapist-qa" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.892, + 0.734, + 0.906 + ], + "angle": 0, + "content": "3https://openai.com/blog/chatgpt" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.906, + 0.798, + 0.918 + ], + "angle": 0, + "content": "4https://lmsys.org/blog/2023-03-30-vicuna/" + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.879, + 0.822, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.52, + 0.941 + ], + "angle": 0, + "content": "4297" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.082, + 0.472, + 0.265 + ], + "angle": 0, + "content": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
Full training*75.024.0
Vicuna73.810.9511.230.78
ChatGPT73.470.5819.241.00
ChatGPT + ZCoT77.101.2120.211.02
ChatGPT + DoT81.190.1122.250.70
GPT-483.040.5133.860.83
GPT-4 + ZCoT81.971.2133.221.36
GPT-4 + DoT82.770.8134.641.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.273, + 0.489, + 0.3 + ], + "angle": 0, + "content": "Table 2: Main Results (Standard deviation in subscript). \\((^{\\star})\\) results copied from (Shreevastava and Foltz, 2021)." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.318, + 0.484, + 0.424 + ], + "angle": 0, + "content": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
ChatGPT73.470.5819.241.00
ChatGPT + S179.621.1218.721.95
ChatGPT + S1 + S280.700.4820.111.02
ChatGPT + S1 + S2 + S381.190.1122.250.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.432, + 0.49, + 0.47 + ], + "angle": 0, + "content": "Table 3: Ablation studies on each stage of DoT: we denote Subjectivity Assessment as S1, Contrastive Reasoning as S2, and Schema Analysis as S3." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.501, + 0.49, + 0.758 + ], + "angle": 0, + "content": "boosts the performance of ChatGPT, with the distortion assessment score surpassing the full-training performance and the distortion classification score approaching the full-training performance. The improvement over ZCoT demonstrates the superiority of our proposed strategic stages compared with rationale generation with no guidance. Astonishingly, GPT-4 surpasses the full-training performance even by a large margin for distortion classification. On the one hand, this may demonstrate the powerful ability of GPT-4 on this task - it can already reach a potential upper bound regardless with DoT promoting or not; On the other hand, since the dataset we use is publicly released on Kaggle and its original data source is from online forums, we do not exclude the possibility of data contamination." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.759, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Ablation Studies and Analysis. To better understand how each stage of DoT contributes to overall performance, we conduct ablation studies using ChatGPT shown in Table 3. All three stages obtain improvements for distortion assessment and distortion classification. For distortion assessment, subjectivity assessment (S1) gains the largest improvement, as separating objective facts and subjective thoughts is a strong trigger to assess distortions. For distortion classification, Schema Analysis (S3)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.082, + 0.871, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.221, + 0.843, + 0.235 + ], + "angle": 0, + "content": "Figure 2: Distortion classification per class F-1." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.248, + 0.879, + 0.375 + ], + "angle": 0, + "content": "
MethodsQualitySubjectivity AssessmentContrastive ReasoningSchema Analysis
ChatGPT +DoTComprehensive68.5%45.0%50.5%
Partially good24.5%30.5%27.0%
Invalid7.0%24.5%22.5%
GPT4 +DoTComprehensive84.5%69.5%76.5%
Partially good13.5%25.5%18.5%
Invalid2.0%5.0%5.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.383, + 0.875, + 0.397 + ], + "angle": 0, + "content": "Table 4: Human evaluation results for diagnosis rationales." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.429, + 0.884, + 0.509 + ], + "angle": 0, + "content": "obtains large improvement, as the summarized underlying thought schema can match the definition of the distortion types more accurately. We also analyze the distortion classification results for each distortion type, with results shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.52, + 0.768, + 0.535 + ], + "angle": 0, + "content": "4.3 Human Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.542, + 0.885, + 0.67 + ], + "angle": 0, + "content": "As there is no reference available for the diagnosis rationales, we employ psychotherapy experts to assess the quality of the generated rationales. We hired psychotherapy professionals from UpWork5, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.671, + 0.884, + 0.895 + ], + "angle": 0, + "content": "Specifically, we present the patient's speech, all the prompts, and the generated rationales for all stages to human experts. For the generated rationales of each stage, we instruct the experts to choose between: 1) Comprehensive. (Correct and comprehensive.); 2) Partially good. (Reasonable but not comprehensive) 3) Invalid. (Not reasonable.) Table 4 shows the evaluation results on 100 examples for DoT over ChatGPT and GPT-4. We employ two experts for each evaluation and report the average; The agreement rates (the percentage of examples the two experts gave the same rating) for both evaluations are over \\(80\\%\\). The diagnosis rationales generated for all stages shows decent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.904, + 0.655, + 0.919 + ], + "angle": 0, + "content": "5www.upwork.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.488, + 0.117 + ], + "angle": 0, + "content": "quality verified by the experts. See Appendix C for more generation examples." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.129, + 0.27, + 0.145 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.155, + 0.49, + 0.686 + ], + "angle": 0, + "content": "Due to the verbal and textual nature of psychotherapy procedures, there have been consistent efforts to employ NLP techniques to assist the mental health domain (Althoff et al., 2016; Abd-Alrazaq et al., 2021, 2019; Valizadeh and Parde, 2022). However, with the knowledge gap between the two communities, most existing works take shallow attempts without deep investigation of the professional psychotherapy knowledge. A majority of the previous studies targets identification of common mental health issues, such as depression and anxiety, from textual contents (Cohan et al., 2018; MacAvaney et al., 2018; Harrigian et al., 2020; Ji et al., 2022; Zanwar et al., 2023; Juhng et al., 2023). Another major category of works study therapeutic conversations with a focus on emotional/empathetic supports (Halder et al., 2017; Sharma et al., 2020a; Atapattu et al., 2022; Mishra et al., 2023), and discourse structures (Cao et al., 2019; Hsu et al., 2023). A few more recent works have started to investigate deeper professional psychotherapy knowledge, e.g., cognitive distortion detection in CBT (Shreevastava and Foltz, 2021; Ding et al., 2022; Lybarger et al., 2022), identifying and reframing unhelpful thoughts (Ziems et al., 2022; Maddela et al., 2023). Early famous system Eliza6 took the initial attempts to emulate a Rogerian psychotherapist. Due to its rule-based nature, the responses are often reflections or rephrasings of the user's statements. If users deviate too much from expected inputs or probe its capabilities, the program may produce irrelevant or nonsensical responses." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.687, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Recent progress in large language model reveals that ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. However, latter works also point out the robustness issue of such ability (Shapira et al., 2023). In our work, we are inspired by such enhanced ability and investigate the application of diagnosing patients' thoughts in psychotherapy. Our experiments on the real dataset, not anecdotal examples, show very promising performance both in auto" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.278 + ], + "angle": 0, + "content": "matic and human expert evaluation. We believe it is a very important future research direction to investigate the general cognitive abilities of LLM and make connections to the field of psychology and cognitive science. Theory-wise, we are eager to explore to what extent the LLM can simulate the human cognitive functions, so as to determine the role that language plays in the overall human cognition. Application-wise, a straightforward and encouraging direction should be the integration for assisting mental health treatment, as the motivation and goal of this work." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.291, + 0.798, + 0.307 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.317, + 0.885, + 0.525 + ], + "angle": 0, + "content": "This work delved into the integration of large language models (LLMs) within the realm of psychotherapy. We introduced the Diagnosis of Thought (DoT) framework, which strategically prompts the LLM to produce diagnosis rationales pertinent to the detection of cognitive distortions. We believe that there exists substantial potential for leveraging LLMs within numerous facets of mental health support that are currently under-explored. Our findings, thus, not only illuminate a path towards more efficient therapy methods but also open doors for future investigations to push the boundaries of AI's role in mental health treatment." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.527, + 0.885, + 0.801 + ], + "angle": 0, + "content": "In the domain of AI-enhanced psychotherapy, particularly when harnessing the power of LLMs, there are paramount challenges surrounding safety, robustness, and ethical responsibilities. LLM is known to generate hallucinations and biases. While one potential usage of our system is to serve as a self-diagnosing/assisting tool for patients, e.g., to help the patients to recognize their own thinking patterns and learn how the cognitive distortions developed, such tools should not be deployed for direct patient use without the supervision of professionals. As we navigate the potential of LLM in psychotherapy, our foremost priority should remain in building systems that are safe and responsible. Concurrently, the formulation of robust ethical guidelines tailored to this new era becomes indispensable." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.813, + 0.615, + 0.829 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.885, + 0.92 + ], + "angle": 0, + "content": "One of the biggest obstacles to studying AI for psychotherapy is the lack of available, high-quality data. Most of the datasets recording the information of the interactions between the patients and licensed professionals are confidential due to pri" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.369, + 0.919 + ], + "angle": 0, + "content": "\\(^{6}\\)https://en.wikipedia.org/wiki/ELIZA" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.31 + ], + "angle": 0, + "content": "vacy concerns. The dataset used in this work is the only one we can find publicly available for cognitive distortion detection, as its data source is from public online forums. In addition to the patient's speech, the patient's demographic information also plays an important role in analyzing the thought process leading to cognitive distortions. In this dataset, some of the patient speech include a few demographic details, but pretty minimal. Beyond the task of cognitive distortion detection, the same data constraint issue exists for all other therapy stages and paradigms. This also motivates us to work on building privacy-preserved systems as a crucial future direction." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.324, + 0.266, + 0.339 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.351, + 0.489, + 0.446 + ], + "angle": 0, + "content": "The system built in this work are never meant to replace psychotherapy professionals: all system-produced results need to be verified by licensed professionals. The system should not be deployed for direct patient use without the supervision of licensed professionals." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.449, + 0.49, + 0.672 + ], + "angle": 0, + "content": "There has been a lot of discussions regarding the use of LLMs for AI-driven therapy purpose. Chatbots will find their most effective role in medicine within the realm of mental health, an opinion quoted from Thomas Insel, former director of the National Institute of Mental Health. In the meantime, we should always take caution for such development. Our foremost priority should be building safe and responsible applications while establishing ethical and regulation standards for this new era. Under the current status, the partnership between professionals and AI systems is a promising direction for the goal to ease the burden of professionals, which is the motivation of this work." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.674, + 0.49, + 0.883 + ], + "angle": 0, + "content": "\"Diagnosis of Thought\" is an innovative approach that harnesses the capabilities of large language models to enhance the field of psychotherapy. It is imperative to understand that this tool is not designed for censorship or any form of invasive surveillance. Instead, it aims to augment the skills of therapists by providing them with advanced, data-driven insights into their clients' thoughts and emotions. Through the efficient processing of linguistic cues and patterns, these models aid mental health professionals in formulating more precise diagnoses and crafting tailored therapeutic interventions that could significantly enhance the well" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.762, + 0.101 + ], + "angle": 0, + "content": "being of individuals seeking help." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.102, + 0.885, + 0.214 + ], + "angle": 0, + "content": "For all the annotations and human evaluations presented in this work, we hired psychotherapy professionals from UpWork8, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate. The average hourly rate for all the experts is $80." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.215, + 0.884, + 0.263 + ], + "angle": 0, + "content": "This project is approved by our Institutional Review Board (IRB). The data annotation is classified as exempt by our IRB." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.291, + 0.61, + 0.306 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.314, + 0.887, + 0.393 + ], + "angle": 0, + "content": "Alaa A Abd-Alrazaq, Mohannad Alajlani, Ali Abdallah Alalwan, Bridgette M Bewick, Peter Gardner, and Mowafa Househ. 2019. An overview of the features of chatbots in mental health: A scoping review. International Journal of Medical Informatics, 132:103978." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.404, + 0.886, + 0.471 + ], + "angle": 0, + "content": "Alaa A Abd-Alrazaq, Mohammad Alajlani, Nashva Ali, Kerstin Denecke, Bridgette M Bewick, and Mowafa Househ. 2021. Perceptions and opinions of patients about mental health chatbots: scoping review. Journal of medical Internet research, 23(1):e17828." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.481, + 0.886, + 0.548 + ], + "angle": 0, + "content": "Tim Althoff, Kevin Clark, and Jure Leskovec. 2016. Large-scale analysis of counseling conversations: An application of natural language processing to mental health. Transactions of the Association for Computational Linguistics, 4:463-476." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.559, + 0.886, + 0.678 + ], + "angle": 0, + "content": "Thushari Atapattu, Mahen Herath, Charitha Elvitigala, Piyanjali de Zoysa, Kasun Gunawardana, Menasha Thilakaratne, Kasun de Zoysa, and Katrina Falkner. 2022. EmoMent: An emotion annotated mental health corpus from two South Asian countries. In Proceedings of the 29th International Conference on Computational Linguistics, pages 6991-7001, Gyeongju, Republic of Korea. International Committee on Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.688, + 0.886, + 0.769 + ], + "angle": 0, + "content": "Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wen-liang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.779, + 0.885, + 0.82 + ], + "angle": 0, + "content": "Simon Baron-Cohen, Alan M Leslie, and Uta Frith. 1985. Does the autistic child have a \"theory of mind\"? Cognition, 21(1):37-46." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.83, + 0.885, + 0.857 + ], + "angle": 0, + "content": "Judith S Beck. 2020. Cognitive behavior therapy: Basics and beyond. Guilford Publications." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.868, + 0.885, + 0.896 + ], + "angle": 0, + "content": "Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar," + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.314, + 0.887, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.892, + 0.488, + 0.919 + ], + "angle": 0, + "content": "7https://www.scientificamerican.com/article/ai-chatbots-could-help-provide-therapy-but-caution-is-needed/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.904, + 0.655, + 0.919 + ], + "angle": 0, + "content": "\\(^{8}\\)www.upwork.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4300" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.138 + ], + "angle": 0, + "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.153, + 0.488, + 0.205 + ], + "angle": 0, + "content": "Jie Cao, Michael Tanana, Zac E Imel, Eric Poitras, David C Atkins, and Vivek Srikumar. 2019. Observing dialogue in therapy: Categorizing and forecasting behavioral codes. arXiv preprint arXiv:1907.00326." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.219, + 0.487, + 0.258 + ], + "angle": 0, + "content": "Wenhu Chen. 2022. Large language models are few (1)-shot table reasoners. arXiv preprint arXiv:2210.06710." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.272, + 0.488, + 0.377 + ], + "angle": 0, + "content": "Arman Cohan, Bart Desmet, Andrew Yates, Luca Soldaini, Sean MacAvaney, and Nazli Goharian. 2018. SMHD: a large-scale resource for exploring online language usage for multiple mental health conditions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1485-1497, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.39, + 0.489, + 0.508 + ], + "angle": 0, + "content": "Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. Improving classification of infrequent cognitive distortions: Domain-specific model vs. data augmentation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 68-75, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.522, + 0.488, + 0.627 + ], + "angle": 0, + "content": "Kishaloy Halder, Lahari Poddar, and Min-Yen Kan. 2017. Modeling temporal progression of emotional status in mental health forum: A recurrent neural net approach. In Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 127-135, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.64, + 0.488, + 0.719 + ], + "angle": 0, + "content": "Keith Harrigian, Carlos Aguirre, and Mark Dredze. 2020. Do models of mental health based on social media data generalize? In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3774-3788, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.733, + 0.488, + 0.773 + ], + "angle": 0, + "content": "Russell Harris. 2006. Embracing your demons: An overview of acceptance and commitment therapy. Psychotherapy in Australia, 12(4):70-6." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.786, + 0.488, + 0.839 + ], + "angle": 0, + "content": "Steven C Hayes, Kirk D Strosahl, and Kelly G Wilson. 2011. Acceptance and commitment therapy: The process and practice of mindful change. Guilford press." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.852, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, Zahra Ashktorab, Casey Dugan, Werner Geyer, and Diyi Yang. 2023. Helping the helper: Supporting peer counselors via ai-empowered practice and feedback. arXiv preprint arXiv:2305.08982." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.178 + ], + "angle": 0, + "content": "Shaoxiong Ji, Tianlin Zhang, Luna Ansari, Jie Fu, Prayag Tiwari, and Erik Cambria. 2022. MentalBERT: Publicly available pretrained language models for mental healthcare. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 7184-7190, Marseille, France. European Language Resources Association." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.189, + 0.883, + 0.281 + ], + "angle": 0, + "content": "Swanie Juhng, Matthew Matero, Vasudha Varadarajan, Johannes Eichstaedt, Adithya V Ganesan, and H Andrew Schwartz. 2023. Discourse-level representations can improve prediction of degree of anxiety. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1500-1511." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.291, + 0.883, + 0.343 + ], + "angle": 0, + "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.354, + 0.883, + 0.434 + ], + "angle": 0, + "content": "Kevin Lybarger, Justin Tauscher, Xiruo Ding, Dror Ben-Zeev, and Trevor Cohen. 2022. Identifying distorted thinking in patient-therapist text message exchanges by leveraging dynamic multi-turn context. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pages 126-136." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.444, + 0.883, + 0.549 + ], + "angle": 0, + "content": "Sean MacAvaney, Bart Desmet, Arman Cohan, Luca Soldaini, Andrew Yates, Ayah Zirikly, and Nazli Goharian. 2018. RSDD-time: Temporal annotation of self-reported mental health diagnoses. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 168-173, New Orleans, LA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.56, + 0.883, + 0.625 + ], + "angle": 0, + "content": "Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. arXiv preprint arXiv:2307.02768." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.636, + 0.883, + 0.741 + ], + "angle": 0, + "content": "Kshitij Mishra, Priyanshu Priya, and Asif Ekbal. 2023. PAL to lend a helping hand: Towards building an emotion adaptive polite and empathetic counseling conversational agent. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12254-12271, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.883, + 0.843 + ], + "angle": 0, + "content": "Modhurima Moitra, Damian Santomauro, Pamela Y Collins, Theo Vos, Harvey Whiteford, Shekhar Saxena, and Alize J Ferrari. 2022. The global gap in treatment coverage for major depressive disorder in 84 countries from 2000-2019: A systematic review and bayesian meta-regression analysis. PLoS Medicine, 19(2):e1003901." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.854, + 0.883, + 0.88 + ], + "angle": 0, + "content": "OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.883, + 0.918 + ], + "angle": 0, + "content": "Steve Rathje, Dan-Mircea Mirea, Ilia Sucholutsky, Raja Marjieh, Claire Robertson, and Jay J Van Bavel. 2023." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "4301" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.135, + 0.086, + 0.486, + 0.113 + ], + "angle": 0, + "content": "Gpt is an effective tool for multilingual psychological text analysis." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.123, + 0.489, + 0.163 + ], + "angle": 0, + "content": "Barbara Olasov Rothbaum, Elizabeth A Meadows, Patricia Resick, and David W Foy. 2000. Cognitive-behavioral therapy." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.173, + 0.487, + 0.264 + ], + "angle": 0, + "content": "Tulika Saha, Saichethan Reddy, Anindya Das, Sriparna Saha, and Pushpak Bhattacharyya. 2022. A shoulder to cry on: Towards a motivational virtual assistant for assuaging mental agony. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2436-2449." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.274, + 0.487, + 0.351 + ], + "angle": 0, + "content": "Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.363, + 0.487, + 0.454 + ], + "angle": 0, + "content": "Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020a. A computational approach to understanding empathy expressed in text-based mental health support. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5263-5276, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.465, + 0.487, + 0.517 + ], + "angle": 0, + "content": "Ashish Sharma, Adam S Miner, David C Atkins, and Tim Althoff. 2020b. A computational approach to understanding empathy expressed in text-based mental health support. arXiv preprint arXiv:2009.08441." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.527, + 0.487, + 0.593 + ], + "angle": 0, + "content": "Sagarika Shreevastava and Peter Foltz. 2021. Detecting cognitive distortions from patient-therapist interactions. In Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access, pages 151-158." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.603, + 0.487, + 0.68 + ], + "angle": 0, + "content": "Mina Valizadeh and Natalie Parde. 2022. The ai doctor is in: A survey of task-oriented dialogue systems for healthcare applications. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6638-6660." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.691, + 0.487, + 0.757 + ], + "angle": 0, + "content": "Anuradha Welivita, Yubo Xie, and Pearl Pu. 2021. A large-scale dataset for empathetic response generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1251-1264." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.767, + 0.487, + 0.806 + ], + "angle": 0, + "content": "Marsha White and Steve M Dorman. 2001. Receiving social support online: implications for health education. Health education research, 16(6):693-707." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.817, + 0.487, + 0.869 + ], + "angle": 0, + "content": "Jesse H Wright, Gregory K Brown, Michael E Thase, and Monica Ramirez Basco. 2017. Learning cognitive-behavior therapy: An illustrated guide. American Psychiatric Pub." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.879, + 0.487, + 0.918 + ], + "angle": 0, + "content": "Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.123, + 0.489, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.882, + 0.139 + ], + "angle": 0, + "content": "In Findings of the Association for Computational Linguistics: ACL 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.148, + 0.881, + 0.188 + ], + "angle": 0, + "content": "Caleb Ziems, Minzhi Li, Anthony Zhang, and Diyi Yang. 2022. Inducing positive perspectives with text reframing. arXiv preprint arXiv:2204.02952." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.199, + 0.858, + 0.215 + ], + "angle": 0, + "content": "Appendix A: Cognitive Distortion Types" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.882, + 0.319 + ], + "angle": 0, + "content": "We list all the ten cognitive distortion types studies in this work following-(Shreevastava and Foltz, 2021) in Table 5. Note that these ten types being studied are some common, well-studied ones. As psychotherapy develops, more fine-grained and new distortion types are expected to be unveiled." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.882, + 0.416 + ], + "angle": 0, + "content": "For the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), there are 2,531 examples in total. We follow the original \\(80\\% - 20\\%\\) train-test split. The input patient speech has an average length of 167.3 tokens, and the ten distortion types are roughly equally distributed." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.428, + 0.758, + 0.444 + ], + "angle": 0, + "content": "Appendix B: Prompt Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.881, + 0.498 + ], + "angle": 0, + "content": "For all our experiments without using DoT, we first prompt the LLM with the following general instructions:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.882, + 0.612 + ], + "angle": 0, + "content": "Given a speech of a patient, our task is to 1) identify if there is cognitive distortion in the speech; 2) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.882, + 0.659 + ], + "angle": 0, + "content": "For all our experiments using DoT, we first prompt the LLM with the following general instructions:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.882, + 0.821 + ], + "angle": 0, + "content": "Given a speech of a patient, our task is to 1) finish a few diagnose of thought questions to analyze the thought patterns of the patient. Then based on the diagnose of thought analysis, 2) identify if there is cognitive distortion in the speech; 3) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.881, + 0.854 + ], + "angle": 0, + "content": "Then for methods without using DoT, we prompt the following two questions:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Please first answer: if there is cognitive distortion in the speech; Answer 'yes' or 'no'; Please then answer: Recognizing the specific types of the cognitive distortion in the speech. There may be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4302" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.084, + 0.88, + 0.326 + ], + "angle": 0, + "content": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
OvergeneralizationMajor conclusions are drawn based on limited information.Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
Emotional reasoningLetting one's feeling about something overrule facts to the contrary.Even though Steve is here at work late every day, I know I work harder than anyone else at my job.
LabelingGiving someone or something a label without finding out more about it/them.My daughter would never do anything I disapproved of.
MagnificationEmphasizing the negative or playing down the positive of a situation.My professor said he made some corrections on my paper, so I know I'll probably fail the class.
Mental filterPlacing all one's attention o, or seeing only, the negatives of a situation.My husband says he wishes I was better at housekeeping, so I must be a lousy wife.
Should statementsShould statements appear as a list of ironclad rules about how a person should behave, this could be about the speaker themselves or other.It is NOT necessary that the word 'should' or it's synonyms (ought to, must etc.) be present in the statements containing this distortion.I should get all A's to be a good student.
Fortune-tellingAs the name suggests, this distortion is about expecting things to happen a certain way, or assuming that thing will go badlyCounterintuitively, this distortion does not always have future tense.I was afraid of job interviews so I decided to start my own thing.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.334, + 0.867, + 0.348 + ], + "angle": 0, + "content": "Table 5: Common cognitive distortion types and example speech, taken from (Beck, 2020; Shreevastava and Foltz, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.379, + 0.489, + 0.442 + ], + "angle": 0, + "content": "one type of cognitive distortion or multiple types involved. If there are multiple types, please give the top 2 dominant ones. Please only give the distortion type names separated by comma." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.444, + 0.489, + 0.652 + ], + "angle": 0, + "content": "For methods using DoT, before asking this two questions, we provide the following DoT prompts: Based on the patient's speech, finish the following diagnosis of thought questions: 1. what is the situation? Find out the facts that are objective; what is the patient thinking or imagining? Find out the thoughts or opinions that are subjective. 2. what makes the patient think the thought is true or is not true? Find out the reasoning processes that support and do not support these thoughts. 3. why does the patient come up with such reasoning process supporting the thought? What's the underlying cognition mode of it?" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.664, + 0.421, + 0.681 + ], + "angle": 0, + "content": "Appendix C: Generation Examples" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.689, + 0.486, + 0.721 + ], + "angle": 0, + "content": "We present some generation examples of GPT-3.5 and GPT-4 in Figure 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4303" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.13, + 0.195, + 0.862, + 0.224 + ], + "angle": 0, + "content": "Patient Message: Lately I have been extremely unhappy for no apparent reason. I feel empty almost all the time. It is like nothing can give me genuine joy anymore. I also feel alone a lot, even though I do have friends and close friends. Whenever I try to imagine my future all I see is one full of obstacles. I have become so consumed in sadness and I do not know how to deal anymore. I feel so confused and I do not understand myself anymore." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.241, + 0.39, + 0.252 + ], + "angle": 0, + "content": "Diagnosis of Thought (ChatGPT):" + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.255, + 0.478, + 0.457 + ], + "angle": 0, + "content": "
Objective Fact:The patient has been feeling extremely unhappy and empty lately.
Subjective Thought:Nothing can give the patient genuine joy and that the future is full of obstacles. the patient feels confused and does not understand themselves anymore.
Reasoning Process Supporting the Thoughts:The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.
Reasoning Process Contradicting the Thoughts:There may not be factual evidence to support the belief that nothing can give the patient genuine joy or that the future is full of obstacles. the patient may be using emotional reasoning to support their beliefs.
Cognition Schema:The patient may be experiencing a cognitive mode of negativity bias, where they tend to focus on negative experiences and thoughts rather than positive ones. this may be influencing their reasoning processes, leading them to magnify the negative and filter out the positive aspects of their life and future. additionally, the patient may be struggling with depression, which can impact the way they perceive themselves and the world around them.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.241, + 0.786, + 0.252 + ], + "angle": 0, + "content": "Diagnosis of Thought (GPT4):" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.255, + 0.861, + 0.474 + ], + "angle": 0, + "content": "
Objective Fact:The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.
Subjective Thought:They believe nothing can bring them genuine joy anymore. They imagine their future full of obstacles. They feel consumed by sadness and confusion and no longer understand themselves.
Reasoning ProcessThe persistent feeling of emptiness and lack of joy might make the patient believe that their situation is unchanging and will remain the same. They may be experiencing negative events in their lives, making it harder to feel hopeful about their future.
Supporting the Thoughts:Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.
Reasoning ProcessEmotions and feelings are temporary and can change over time, so it is not guaranteed that the patient will always feel this way. The patient is aware that they do have friends, which suggests that some level of meaningful connection exists. Obstacles are a natural part of life, but they are not the only things present in one's future. positive events and achievements are also likely to occur.
Contradicting the Thoughts:
Cognition Schema:Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life's reality. Mental filter: they are focusing on the negative aspects of their life and filtering out positive experiences or memories. Fortune-telling: they are assuming that their future will be full of obstacles based on their current emotions and thoughts.
" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.499, + 0.864, + 0.53 + ], + "angle": 0, + "content": "Patient Message: So I have no close friends at the moment probably I know some people enough just to say hi but no close one and whenever I make any they go away I don't know if its my personality the reason or the way I behave and the only friends I can gain are school friends since there is no other place I go to I can't really stand myself alone. Thanks for helping." + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.546, + 0.388, + 0.557 + ], + "angle": 0, + "content": "Diagnosis of Thought (ChatGPT):" + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.563, + 0.478, + 0.74 + ], + "angle": 0, + "content": "
Objective Fact:The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.
Subjective Thought:The patient feels that they are the reason for their friends leaving and wonders if it's their personality or behavior causing the issue. they cannot stand being alone.
Reasoning Process Supporting the Thoughts:The patient believes their thoughts to be true because they have not been able to maintain any close friendships and the pattern keeps repeating itself. they may be using past experiences to support this belief.
Reasoning Process Contradicting the Thoughts:There may be other reasons for the loss of these friendships that the patient is not considering.
Cognition Schema:The patient may be experiencing cognitive distortions such as overgeneralization by assuming that all friendships will end up the same way and mental filtering by focusing only on the negative aspects of their past friendships. the underlying cognition mode could be a negative self-image or low self-esteem.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.546, + 0.785, + 0.557 + ], + "angle": 0, + "content": "Diagnosis of Thought (GPT4):" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.56, + 0.861, + 0.767 + ], + "angle": 0, + "content": "
Objective Fact:The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.
Subjective Thought:The patient believes that their personality or behavior might be the reason their friendships don't last, and they are limited in opportunities to make new friends.
Reasoning ProcessThe patient might have experienced failed friendships and noticed a pattern of people leaving their life. this leads them to question if their personality or behavior is the cause.
Supporting the Thoughts:
Reasoning ProcessThe patient has not provided concrete examples or evidence that their personality or behavior is the reason behind the lack of close friendships. it is possible there are external factors or situational circumstances that led to the loss of friendships.
Contradicting the Thoughts:
Cognition Schema:The patient might be engaging in personalization and overgeneralization cognitive distortions. personalization because they seem to be taking the blame for friendships not lasting, assuming it's their fault due to their personality or behavior. overgeneralization because they are making a broad conclusion based on limited information (i.e., the assumption that their personality or behavior is the sole reason for the lack of close friends). this underlying cognition mode could stem from low self-esteem, fear of rejection, or previous negative experiences in social situations.
" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.802, + 0.802, + 0.816 + ], + "angle": 0, + "content": "Figure 3: Showcases of Diagnosis of Thought using ChatGPT and GPT4 given the Patient Message." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "4304" + } + ] +] \ No newline at end of file diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_origin.pdf b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f6747759f6543089e8f5ea8b5588762c26d4ea52 --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/04980a83-f896-4166-a550-6164ef62b39f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c114ad96b3bcd9d830754b109db33d9d48eadecaa543edc8526e34765047341 +size 637199 diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/full.md b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d379e806d7bb9e5e82aa99b3c3e2c7b07113e1e --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/full.md @@ -0,0 +1,239 @@ +# Empowering Psychotherapy with Large Language Models: Cognitive Distortion Detection through Diagnosis of Thought Prompting + +Zhiyu Chen $^{1}$ , Yujie Lu $^{2}$ , and William Yang Wang $^{2}$ + +Carnegie Mellon University + +2University of California, Santa Barbara + +zhiyuc@andrew.cmu.edu, + +{yujielu, william}@cs.ucsb.edu + +# Abstract + +Mental illness remains one of the most critical public health issues of our time, due to the severe scarcity and accessibility limit of professionals. Psychotherapy requires high-level expertise to conduct deep, complex reasoning and analysis on the cognition modeling of the patients. In the era of Large Language Models, we believe it is the right time to develop AI assistance for computational psychotherapy. We study the task of cognitive distortion detection and propose the Diagnosis of Thought (DoT) prompting. DoT performs diagnosis on the patient's speech via three stages: subjectivity assessment to separate the facts and the thoughts; contrastive reasoning to elicit the reasoning processes supporting and contradicting the thoughts; and schema analysis to summarize the cognition schemas. The generated diagnosis rationales through the three stages are essential for assisting the professionals. Experiments demonstrate that DoT obtains significant improvements over ChatGPT for cognitive distortion detection, while generating high-quality rationales approved by human experts. + +# 1 Introduction + +About one in eight people worldwide suffer from mental disorders (World Health Organization, 2022). However, such mental health conditions are severely underserved, due to a number of reasons including the scarcity of mental health professionals, poor quality of services, unaffordable cost, and social stigma (White and Dorman, 2001; Sharma et al., 2020b). Treatment coverage for mental health service use ranges from $33\%$ in high-income locations to only $8\%$ in low- and lower middle-income countries (Moitra et al., 2022). A recent study from the American Psychological Association (APA) $^{1}$ found that six in ten psychologists "no longer have openings for new patients." + +To mitigate such situations, there have been consistent efforts on developing automated systems for mental health support, such as sentiment analysis (Rathje et al., 2023) and empathetic chatbots (Welivita et al., 2021; Sharma et al., 2020b; Saha et al., 2022). However, existing works mostly take shallow attempts in a heuristic manner, e.g., analyzing emotions and generating comforting responses. There is still a significant gap for such systems to contribute to real professional psychotherapy, which requires deep studies of the patient's thinking patterns, the establishment of cognition models, and the methods to reconstruct the cognition models. These procedures constitute the core pillars in common classic therapy paradigms like cognitive-behavior therapy (CBT) (Rothbaum et al., 2000; Wright et al., 2017; Beck, 2020) and acceptance and commitment therapy (ACT) (Harris, 2006; Hayes et al., 2011). Most data sources recording the interactions between the patients and licensed professionals are confidential, making it even more challenging to build professional assistance for psychotherapy. + +Recent progress in Large Language Model (LLM) development uncovers its astonishing ability in various textual reasoning tasks in zero-shot setting (Kojima et al., 2022; Bang et al., 2023; Chen, 2022). For the psychology domain, ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. We believe it is promising to further exploit such ability to complex mental reasoning and analysis. It is the right time to start developing professional, targeted, and systematic AI assistance for psychotherapy. + +In this paper, we take the first step by studying the task of cognitive distortion detection, the first core procedure in cognitive behavior therapy (CBT) (Beck, 2020). Inspired by how psychother + +
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
+ +Table 1: Three example common cognitive distortion types, taken from (Beck, 2020; Shreevastava and Foltz, 2021). + +apy professionals perform nuanced diagnosis over the patient's speech, we propose the Diagnosis of Thought (DoT) prompting. In DoT, we diagnose the patient's speech through three stages: (1) subjectivity assessment, (2) contrastive reasoning, and (3) schema analysis. In subjectivity assessment, we distinguish the patient's subjective thoughts from the objective facts; In contrastive reasoning, we elicit the reasoning processes supporting and opposing the patient's thoughts; Finally, in schema analysis, we summarize the underlying thought schema and map it to the cognitive distortion types. We conduct comprehensive experiments using the recent top-performing LLMs. In zero-shot setting, DoT obtains over $10\%$ and $15\%$ relative improvements for distortion assessment and classification, respectively, on ChatGPT. Meanwhile, the generated rationales through the three stages grant full interpretability for the diagnosis process, whose quality is further approved by human experts. + +We unveil the great potential of empowering professional psychotherapy with LLM. This exploration serves as a catalyst for a larger initiative; we extend an invitation to both AI and psychotherapy communities to come together in a collaborative effort. Our ultimate goal is to construct professional, safe, AI-driven assistance that can substantially enhance mental health support systems. + +# 2 Cognitive Distortion Detection + +Cognitive Behavior Therapy (CBT) is a well-established therapy paradigm primarily on depression and anxiety disorders (Beck, 2020). In CBT, given the patients' speech or written content, we establish the cognitive model by building the interactions among the situations, thoughts, and emotions. Patients with mental disorders, such as depression or anxiety, tend to form negative thoughts very rapidly and unconsciously, leading to negative emotions which further strengthen their overall negative views and beliefs about the world. + +To break this vicious circle, in a typical CBT process, the first core step is to identify those mal + +adaptive negative thoughts and summarize their underlying schemas, formally known as cognitive distortions. There are generally 10-20 common, well-studied types of cognitive distortions. We present a few types with examples in Table 1, and refer the readers to Appendix A for the full list. Once accurately identify these cognitive distortions, CBT therapists will guide the patients to justify and correct these distortions, so as to gradually reconstruct their cognition models. + +In the real psychotherapy process, there's a significant amount of textual information including therapeutic conversations and diaries, etc. Such information is often long, highly fragmented, and disorganized, containing multiple types of distortions beyond the toy examples in Table 1. The task of cognitive distortion detection aims to automatically detect the distortion types given such textual information from the patients, in order to assist therapists to enhance their efficiency and productivity. Meanwhile, such detectors can also potentially serve as self-assisting tools for the patients to diagnose their thoughts and conduct CBT practice, upon meeting the robustness and safety requirements. Formally, cognitive distortion detection consists of two steps: 1) Distortion assessment to predict whether the given speech contains cognitive distortions, as a binary classification problem; and 2) Distortion classification to predict the specific distortion types, as a multiclass classification problem. + +# 3 The Diagnosis of Thought Prompting + +On discussing with psychotherapy professionals, we identify the following key stages to diagnose the patient's speech to detect cognitive distortions. Subjectivity Assessment. The patient's speech consists of a mixture of reality (objective facts) and interpretations/opinions (subjective thoughts). In order to perform deep analysis of distorted thinking, we first need to find out which parts of the speech are objective facts and which parts are subjective thoughts. After such an assessment, we summa + +![](images/b6a1ace8f70193bc2abc3d20531685a261bb01778d2918a60d8e0e4c8cb1da94.jpg) +Figure 1: The Diagnosis of Thought framework. We strategically prompt the LLM to go through the three diagnosis stages: subjectivity assessment, contrastive reasoning, and schema analysis. + +rize the objective facts into the situations as the evidence base to diagnose the subjective thoughts. + +Contrastive Reasoning. This stage aims to discover how the patient ascertains the veracity of their subjective thoughts. Based on the situation, we deduct the reasoning processes that supports and contradicts the patient's thoughts respectively. By contrasting two different interpretations based on the same situation, we can identify the thought schemas more clearly. + +Schema Analysis. This stage aims to study why the patient forms the specific reasoning process. The term "schema" refers to the cognitive structures that organize our knowledge, beliefs, and expectations. Understanding what schemas a patient is relying on can reveal much about their cognitive mode and distortions. + +We propose the Diagnosis of Thought (DoT) prompting, guiding the LLM through the above three stages to diagnose the patient's speech. Figure 1 illustrates our framework. We sequentially prompt LLM with three questions for the three stages. After the LLM finishes the generation for all stages, we prompt it with another two questions asking distortion assessment and classification. See Appendix B for all the prompts we use. + +With DoT prompting, we obtain fully interpretable diagnosis rationales for detecting cognitive distortions. Such interpretability is vital for professionals in real applications. Serving as the diagnosing assistance tool, the diagnosis rationales are essential for the professionals to justify the results. More importantly, they can potentially learn the nuanced thought patterns and schemas derived from the patient's speech through the rationales. This is crucial for the professionals to establish the patient's cognition models more efficiently from + +the vast amount of speech. + +# 4 Experimental Results + +# 4.1 Dataset and experimental settings + +We experiment on the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), which is annotated by experts based on the Therapist QA dataset. The dataset consists of 2,531 examples of patient speech annotated with ten common types of cognitive distortions, as specified in Appendix A. $63.1\%$ of the examples have cognitive distortions, which are annotated with the two dominant ones. We follow the same train-test split in (Shreevastava and Foltz, 2021). For automatic evaluations, we report F-1 for distortion assessment and weighted F-1 for distortion classification. + +# 4.2 Automatic Evaluation Results + +We experiment using three of the recent representative LLMs: ChatGPT (gpt-3.5-turbo)3, Vicuna4, and GPT-4 (OpenAI, 2023). For all methods, we first prompt with a general instruction specifying the task and the target distortion types. We compare our DoT prompting with 1) Directly generating the results, and 2) Zero-Shot CoT prompting (ZCoT) (Kojima et al., 2022). For Vicuna, both ZCoT and DoT exceed the token limits for many examples; we omit the results. We average over five runs and report mean and standard deviation for all experiments. Table 2 shows our main experiment results. As we can see, under the zero-shot setting, due to the challenge of this task, Vicuna and ChatGPT still fall behind the supervised full-training models. The proposed DoT prompting significantly + +
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
Full training*75.024.0
Vicuna73.810.9511.230.78
ChatGPT73.470.5819.241.00
ChatGPT + ZCoT77.101.2120.211.02
ChatGPT + DoT81.190.1122.250.70
GPT-483.040.5133.860.83
GPT-4 + ZCoT81.971.2133.221.36
GPT-4 + DoT82.770.8134.641.40
+ +Table 2: Main Results (Standard deviation in subscript). $(^{\star})$ results copied from (Shreevastava and Foltz, 2021). + +
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
ChatGPT73.470.5819.241.00
ChatGPT + S179.621.1218.721.95
ChatGPT + S1 + S280.700.4820.111.02
ChatGPT + S1 + S2 + S381.190.1122.250.70
+ +boosts the performance of ChatGPT, with the distortion assessment score surpassing the full-training performance and the distortion classification score approaching the full-training performance. The improvement over ZCoT demonstrates the superiority of our proposed strategic stages compared with rationale generation with no guidance. Astonishingly, GPT-4 surpasses the full-training performance even by a large margin for distortion classification. On the one hand, this may demonstrate the powerful ability of GPT-4 on this task - it can already reach a potential upper bound regardless with DoT promoting or not; On the other hand, since the dataset we use is publicly released on Kaggle and its original data source is from online forums, we do not exclude the possibility of data contamination. + +Ablation Studies and Analysis. To better understand how each stage of DoT contributes to overall performance, we conduct ablation studies using ChatGPT shown in Table 3. All three stages obtain improvements for distortion assessment and distortion classification. For distortion assessment, subjectivity assessment (S1) gains the largest improvement, as separating objective facts and subjective thoughts is a strong trigger to assess distortions. For distortion classification, Schema Analysis (S3) + +![](images/6701a369fe9e975803d3a626fd263baf4b30b1ca5278a00d974753f5528791b1.jpg) +Figure 2: Distortion classification per class F-1. + +Table 3: Ablation studies on each stage of DoT: we denote Subjectivity Assessment as S1, Contrastive Reasoning as S2, and Schema Analysis as S3. + +
MethodsQualitySubjectivity AssessmentContrastive ReasoningSchema Analysis
ChatGPT +DoTComprehensive68.5%45.0%50.5%
Partially good24.5%30.5%27.0%
Invalid7.0%24.5%22.5%
GPT4 +DoTComprehensive84.5%69.5%76.5%
Partially good13.5%25.5%18.5%
Invalid2.0%5.0%5.0%
+ +Table 4: Human evaluation results for diagnosis rationales. + +obtains large improvement, as the summarized underlying thought schema can match the definition of the distortion types more accurately. We also analyze the distortion classification results for each distortion type, with results shown in Figure 2. + +# 4.3 Human Evaluation Results + +As there is no reference available for the diagnosis rationales, we employ psychotherapy experts to assess the quality of the generated rationales. We hired psychotherapy professionals from UpWork5, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate. + +Specifically, we present the patient's speech, all the prompts, and the generated rationales for all stages to human experts. For the generated rationales of each stage, we instruct the experts to choose between: 1) Comprehensive. (Correct and comprehensive.); 2) Partially good. (Reasonable but not comprehensive) 3) Invalid. (Not reasonable.) Table 4 shows the evaluation results on 100 examples for DoT over ChatGPT and GPT-4. We employ two experts for each evaluation and report the average; The agreement rates (the percentage of examples the two experts gave the same rating) for both evaluations are over $80\%$ . The diagnosis rationales generated for all stages shows decent + +quality verified by the experts. See Appendix C for more generation examples. + +# 5 Related Work + +Due to the verbal and textual nature of psychotherapy procedures, there have been consistent efforts to employ NLP techniques to assist the mental health domain (Althoff et al., 2016; Abd-Alrazaq et al., 2021, 2019; Valizadeh and Parde, 2022). However, with the knowledge gap between the two communities, most existing works take shallow attempts without deep investigation of the professional psychotherapy knowledge. A majority of the previous studies targets identification of common mental health issues, such as depression and anxiety, from textual contents (Cohan et al., 2018; MacAvaney et al., 2018; Harrigian et al., 2020; Ji et al., 2022; Zanwar et al., 2023; Juhng et al., 2023). Another major category of works study therapeutic conversations with a focus on emotional/empathetic supports (Halder et al., 2017; Sharma et al., 2020a; Atapattu et al., 2022; Mishra et al., 2023), and discourse structures (Cao et al., 2019; Hsu et al., 2023). A few more recent works have started to investigate deeper professional psychotherapy knowledge, e.g., cognitive distortion detection in CBT (Shreevastava and Foltz, 2021; Ding et al., 2022; Lybarger et al., 2022), identifying and reframing unhelpful thoughts (Ziems et al., 2022; Maddela et al., 2023). Early famous system Eliza6 took the initial attempts to emulate a Rogerian psychotherapist. Due to its rule-based nature, the responses are often reflections or rephrasings of the user's statements. If users deviate too much from expected inputs or probe its capabilities, the program may produce irrelevant or nonsensical responses. + +Recent progress in large language model reveals that ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. However, latter works also point out the robustness issue of such ability (Shapira et al., 2023). In our work, we are inspired by such enhanced ability and investigate the application of diagnosing patients' thoughts in psychotherapy. Our experiments on the real dataset, not anecdotal examples, show very promising performance both in auto + +matic and human expert evaluation. We believe it is a very important future research direction to investigate the general cognitive abilities of LLM and make connections to the field of psychology and cognitive science. Theory-wise, we are eager to explore to what extent the LLM can simulate the human cognitive functions, so as to determine the role that language plays in the overall human cognition. Application-wise, a straightforward and encouraging direction should be the integration for assisting mental health treatment, as the motivation and goal of this work. + +# 6 Conclusion and Future Work + +This work delved into the integration of large language models (LLMs) within the realm of psychotherapy. We introduced the Diagnosis of Thought (DoT) framework, which strategically prompts the LLM to produce diagnosis rationales pertinent to the detection of cognitive distortions. We believe that there exists substantial potential for leveraging LLMs within numerous facets of mental health support that are currently under-explored. Our findings, thus, not only illuminate a path towards more efficient therapy methods but also open doors for future investigations to push the boundaries of AI's role in mental health treatment. + +In the domain of AI-enhanced psychotherapy, particularly when harnessing the power of LLMs, there are paramount challenges surrounding safety, robustness, and ethical responsibilities. LLM is known to generate hallucinations and biases. While one potential usage of our system is to serve as a self-diagnosing/assisting tool for patients, e.g., to help the patients to recognize their own thinking patterns and learn how the cognitive distortions developed, such tools should not be deployed for direct patient use without the supervision of professionals. As we navigate the potential of LLM in psychotherapy, our foremost priority should remain in building systems that are safe and responsible. Concurrently, the formulation of robust ethical guidelines tailored to this new era becomes indispensable. + +# Limitations + +One of the biggest obstacles to studying AI for psychotherapy is the lack of available, high-quality data. Most of the datasets recording the information of the interactions between the patients and licensed professionals are confidential due to pri + +vacy concerns. The dataset used in this work is the only one we can find publicly available for cognitive distortion detection, as its data source is from public online forums. In addition to the patient's speech, the patient's demographic information also plays an important role in analyzing the thought process leading to cognitive distortions. In this dataset, some of the patient speech include a few demographic details, but pretty minimal. Beyond the task of cognitive distortion detection, the same data constraint issue exists for all other therapy stages and paradigms. This also motivates us to work on building privacy-preserved systems as a crucial future direction. + +# Ethics Statement + +The system built in this work are never meant to replace psychotherapy professionals: all system-produced results need to be verified by licensed professionals. The system should not be deployed for direct patient use without the supervision of licensed professionals. + +There has been a lot of discussions regarding the use of LLMs for AI-driven therapy purpose. Chatbots will find their most effective role in medicine within the realm of mental health, an opinion quoted from Thomas Insel, former director of the National Institute of Mental Health. In the meantime, we should always take caution for such development. Our foremost priority should be building safe and responsible applications while establishing ethical and regulation standards for this new era. Under the current status, the partnership between professionals and AI systems is a promising direction for the goal to ease the burden of professionals, which is the motivation of this work. + +"Diagnosis of Thought" is an innovative approach that harnesses the capabilities of large language models to enhance the field of psychotherapy. It is imperative to understand that this tool is not designed for censorship or any form of invasive surveillance. Instead, it aims to augment the skills of therapists by providing them with advanced, data-driven insights into their clients' thoughts and emotions. Through the efficient processing of linguistic cues and patterns, these models aid mental health professionals in formulating more precise diagnoses and crafting tailored therapeutic interventions that could significantly enhance the well + +being of individuals seeking help. + +For all the annotations and human evaluations presented in this work, we hired psychotherapy professionals from UpWork8, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate. The average hourly rate for all the experts is $80. + +This project is approved by our Institutional Review Board (IRB). The data annotation is classified as exempt by our IRB. + +# References + +Alaa A Abd-Alrazaq, Mohannad Alajlani, Ali Abdallah Alalwan, Bridgette M Bewick, Peter Gardner, and Mowafa Househ. 2019. An overview of the features of chatbots in mental health: A scoping review. International Journal of Medical Informatics, 132:103978. +Alaa A Abd-Alrazaq, Mohammad Alajlani, Nashva Ali, Kerstin Denecke, Bridgette M Bewick, and Mowafa Househ. 2021. Perceptions and opinions of patients about mental health chatbots: scoping review. Journal of medical Internet research, 23(1):e17828. +Tim Althoff, Kevin Clark, and Jure Leskovec. 2016. Large-scale analysis of counseling conversations: An application of natural language processing to mental health. Transactions of the Association for Computational Linguistics, 4:463-476. +Thushari Atapattu, Mahen Herath, Charitha Elvitigala, Piyanjali de Zoysa, Kasun Gunawardana, Menasha Thilakaratne, Kasun de Zoysa, and Katrina Falkner. 2022. EmoMent: An emotion annotated mental health corpus from two South Asian countries. In Proceedings of the 29th International Conference on Computational Linguistics, pages 6991-7001, Gyeongju, Republic of Korea. International Committee on Computational Linguistics. +Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wen-liang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023. +Simon Baron-Cohen, Alan M Leslie, and Uta Frith. 1985. Does the autistic child have a "theory of mind"? Cognition, 21(1):37-46. +Judith S Beck. 2020. Cognitive behavior therapy: Basics and beyond. Guilford Publications. +Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, + +Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712. +Jie Cao, Michael Tanana, Zac E Imel, Eric Poitras, David C Atkins, and Vivek Srikumar. 2019. Observing dialogue in therapy: Categorizing and forecasting behavioral codes. arXiv preprint arXiv:1907.00326. +Wenhu Chen. 2022. Large language models are few (1)-shot table reasoners. arXiv preprint arXiv:2210.06710. +Arman Cohan, Bart Desmet, Andrew Yates, Luca Soldaini, Sean MacAvaney, and Nazli Goharian. 2018. SMHD: a large-scale resource for exploring online language usage for multiple mental health conditions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1485-1497, Santa Fe, New Mexico, USA. Association for Computational Linguistics. +Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. Improving classification of infrequent cognitive distortions: Domain-specific model vs. data augmentation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 68-75, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics. +Kishaloy Halder, Lahari Poddar, and Min-Yen Kan. 2017. Modeling temporal progression of emotional status in mental health forum: A recurrent neural net approach. In Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 127-135, Copenhagen, Denmark. Association for Computational Linguistics. +Keith Harrigian, Carlos Aguirre, and Mark Dredze. 2020. Do models of mental health based on social media data generalize? In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3774-3788, Online. Association for Computational Linguistics. +Russell Harris. 2006. Embracing your demons: An overview of acceptance and commitment therapy. Psychotherapy in Australia, 12(4):70-6. +Steven C Hayes, Kirk D Strosahl, and Kelly G Wilson. 2011. Acceptance and commitment therapy: The process and practice of mindful change. Guilford press. +Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, Zahra Ashktorab, Casey Dugan, Werner Geyer, and Diyi Yang. 2023. Helping the helper: Supporting peer counselors via ai-empowered practice and feedback. arXiv preprint arXiv:2305.08982. + +Shaoxiong Ji, Tianlin Zhang, Luna Ansari, Jie Fu, Prayag Tiwari, and Erik Cambria. 2022. MentalBERT: Publicly available pretrained language models for mental healthcare. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 7184-7190, Marseille, France. European Language Resources Association. +Swanie Juhng, Matthew Matero, Vasudha Varadarajan, Johannes Eichstaedt, Adithya V Ganesan, and H Andrew Schwartz. 2023. Discourse-level representations can improve prediction of degree of anxiety. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1500-1511. +Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916. +Kevin Lybarger, Justin Tauscher, Xiruo Ding, Dror Ben-Zeev, and Trevor Cohen. 2022. Identifying distorted thinking in patient-therapist text message exchanges by leveraging dynamic multi-turn context. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pages 126-136. +Sean MacAvaney, Bart Desmet, Arman Cohan, Luca Soldaini, Andrew Yates, Ayah Zirikly, and Nazli Goharian. 2018. RSDD-time: Temporal annotation of self-reported mental health diagnoses. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 168-173, New Orleans, LA. Association for Computational Linguistics. +Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. arXiv preprint arXiv:2307.02768. +Kshitij Mishra, Priyanshu Priya, and Asif Ekbal. 2023. PAL to lend a helping hand: Towards building an emotion adaptive polite and empathetic counseling conversational agent. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12254-12271, Toronto, Canada. Association for Computational Linguistics. +Modhurima Moitra, Damian Santomauro, Pamela Y Collins, Theo Vos, Harvey Whiteford, Shekhar Saxena, and Alize J Ferrari. 2022. The global gap in treatment coverage for major depressive disorder in 84 countries from 2000-2019: A systematic review and bayesian meta-regression analysis. PLoS Medicine, 19(2):e1003901. +OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774. +Steve Rathje, Dan-Mircea Mirea, Ilia Sucholutsky, Raja Marjieh, Claire Robertson, and Jay J Van Bavel. 2023. + +Gpt is an effective tool for multilingual psychological text analysis. + +Barbara Olasov Rothbaum, Elizabeth A Meadows, Patricia Resick, and David W Foy. 2000. Cognitive-behavioral therapy. +Tulika Saha, Saichethan Reddy, Anindya Das, Sriparna Saha, and Pushpak Bhattacharyya. 2022. A shoulder to cry on: Towards a motivational virtual assistant for assuaging mental agony. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2436-2449. +Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763. +Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020a. A computational approach to understanding empathy expressed in text-based mental health support. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5263-5276, Online. Association for Computational Linguistics. +Ashish Sharma, Adam S Miner, David C Atkins, and Tim Althoff. 2020b. A computational approach to understanding empathy expressed in text-based mental health support. arXiv preprint arXiv:2009.08441. +Sagarika Shreevastava and Peter Foltz. 2021. Detecting cognitive distortions from patient-therapist interactions. In Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access, pages 151-158. +Mina Valizadeh and Natalie Parde. 2022. The ai doctor is in: A survey of task-oriented dialogue systems for healthcare applications. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6638-6660. +Anuradha Welivita, Yubo Xie, and Pearl Pu. 2021. A large-scale dataset for empathetic response generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1251-1264. +Marsha White and Steve M Dorman. 2001. Receiving social support online: implications for health education. Health education research, 16(6):693-707. +Jesse H Wright, Gregory K Brown, Michael E Thase, and Monica Ramirez Basco. 2017. Learning cognitive-behavior therapy: An illustrated guide. American Psychiatric Pub. +Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion + +In Findings of the Association for Computational Linguistics: ACL 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics. +Caleb Ziems, Minzhi Li, Anthony Zhang, and Diyi Yang. 2022. Inducing positive perspectives with text reframing. arXiv preprint arXiv:2204.02952. + +# Appendix A: Cognitive Distortion Types + +We list all the ten cognitive distortion types studies in this work following-(Shreevastava and Foltz, 2021) in Table 5. Note that these ten types being studied are some common, well-studied ones. As psychotherapy develops, more fine-grained and new distortion types are expected to be unveiled. + +For the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), there are 2,531 examples in total. We follow the original $80\% - 20\%$ train-test split. The input patient speech has an average length of 167.3 tokens, and the ten distortion types are roughly equally distributed. + +# Appendix B: Prompt Details + +For all our experiments without using DoT, we first prompt the LLM with the following general instructions: + +Given a speech of a patient, our task is to 1) identify if there is cognitive distortion in the speech; 2) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021). + +For all our experiments using DoT, we first prompt the LLM with the following general instructions: + +Given a speech of a patient, our task is to 1) finish a few diagnose of thought questions to analyze the thought patterns of the patient. Then based on the diagnose of thought analysis, 2) identify if there is cognitive distortion in the speech; 3) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021). + +Then for methods without using DoT, we prompt the following two questions: + +Please first answer: if there is cognitive distortion in the speech; Answer 'yes' or 'no'; Please then answer: Recognizing the specific types of the cognitive distortion in the speech. There may be + +
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
OvergeneralizationMajor conclusions are drawn based on limited information.Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
Emotional reasoningLetting one's feeling about something overrule facts to the contrary.Even though Steve is here at work late every day, I know I work harder than anyone else at my job.
LabelingGiving someone or something a label without finding out more about it/them.My daughter would never do anything I disapproved of.
MagnificationEmphasizing the negative or playing down the positive of a situation.My professor said he made some corrections on my paper, so I know I'll probably fail the class.
Mental filterPlacing all one's attention o, or seeing only, the negatives of a situation.My husband says he wishes I was better at housekeeping, so I must be a lousy wife.
Should statementsShould statements appear as a list of ironclad rules about how a person should behave, this could be about the speaker themselves or other.It is NOT necessary that the word 'should' or it's synonyms (ought to, must etc.) be present in the statements containing this distortion.I should get all A's to be a good student.
Fortune-tellingAs the name suggests, this distortion is about expecting things to happen a certain way, or assuming that thing will go badlyCounterintuitively, this distortion does not always have future tense.I was afraid of job interviews so I decided to start my own thing.
+ +Table 5: Common cognitive distortion types and example speech, taken from (Beck, 2020; Shreevastava and Foltz, 2021). + +one type of cognitive distortion or multiple types involved. If there are multiple types, please give the top 2 dominant ones. Please only give the distortion type names separated by comma. + +For methods using DoT, before asking this two questions, we provide the following DoT prompts: Based on the patient's speech, finish the following diagnosis of thought questions: 1. what is the situation? Find out the facts that are objective; what is the patient thinking or imagining? Find out the thoughts or opinions that are subjective. 2. what makes the patient think the thought is true or is not true? Find out the reasoning processes that support and do not support these thoughts. 3. why does the patient come up with such reasoning process supporting the thought? What's the underlying cognition mode of it? + +# Appendix C: Generation Examples + +We present some generation examples of GPT-3.5 and GPT-4 in Figure 3. + +Patient Message: Lately I have been extremely unhappy for no apparent reason. I feel empty almost all the time. It is like nothing can give me genuine joy anymore. I also feel alone a lot, even though I do have friends and close friends. Whenever I try to imagine my future all I see is one full of obstacles. I have become so consumed in sadness and I do not know how to deal anymore. I feel so confused and I do not understand myself anymore. + +Diagnosis of Thought (ChatGPT): + +
Objective Fact:The patient has been feeling extremely unhappy and empty lately.
Subjective Thought:Nothing can give the patient genuine joy and that the future is full of obstacles. the patient feels confused and does not understand themselves anymore.
Reasoning Process Supporting the Thoughts:The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.
Reasoning Process Contradicting the Thoughts:There may not be factual evidence to support the belief that nothing can give the patient genuine joy or that the future is full of obstacles. the patient may be using emotional reasoning to support their beliefs.
Cognition Schema:The patient may be experiencing a cognitive mode of negativity bias, where they tend to focus on negative experiences and thoughts rather than positive ones. this may be influencing their reasoning processes, leading them to magnify the negative and filter out the positive aspects of their life and future. additionally, the patient may be struggling with depression, which can impact the way they perceive themselves and the world around them.
+ +Diagnosis of Thought (GPT4): + +
Objective Fact:The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.
Subjective Thought:They believe nothing can bring them genuine joy anymore. They imagine their future full of obstacles. They feel consumed by sadness and confusion and no longer understand themselves.
Reasoning ProcessThe persistent feeling of emptiness and lack of joy might make the patient believe that their situation is unchanging and will remain the same. They may be experiencing negative events in their lives, making it harder to feel hopeful about their future.
Supporting the Thoughts:Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.
Reasoning ProcessEmotions and feelings are temporary and can change over time, so it is not guaranteed that the patient will always feel this way. The patient is aware that they do have friends, which suggests that some level of meaningful connection exists. Obstacles are a natural part of life, but they are not the only things present in one's future. positive events and achievements are also likely to occur.
Contradicting the Thoughts:
Cognition Schema:Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life's reality. Mental filter: they are focusing on the negative aspects of their life and filtering out positive experiences or memories. Fortune-telling: they are assuming that their future will be full of obstacles based on their current emotions and thoughts.
+ +Patient Message: So I have no close friends at the moment probably I know some people enough just to say hi but no close one and whenever I make any they go away I don't know if its my personality the reason or the way I behave and the only friends I can gain are school friends since there is no other place I go to I can't really stand myself alone. Thanks for helping. + +Diagnosis of Thought (ChatGPT): + +
Objective Fact:The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.
Subjective Thought:The patient feels that they are the reason for their friends leaving and wonders if it's their personality or behavior causing the issue. they cannot stand being alone.
Reasoning Process Supporting the Thoughts:The patient believes their thoughts to be true because they have not been able to maintain any close friendships and the pattern keeps repeating itself. they may be using past experiences to support this belief.
Reasoning Process Contradicting the Thoughts:There may be other reasons for the loss of these friendships that the patient is not considering.
Cognition Schema:The patient may be experiencing cognitive distortions such as overgeneralization by assuming that all friendships will end up the same way and mental filtering by focusing only on the negative aspects of their past friendships. the underlying cognition mode could be a negative self-image or low self-esteem.
+ +Diagnosis of Thought (GPT4): + +
Objective Fact:The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.
Subjective Thought:The patient believes that their personality or behavior might be the reason their friendships don't last, and they are limited in opportunities to make new friends.
Reasoning ProcessThe patient might have experienced failed friendships and noticed a pattern of people leaving their life. this leads them to question if their personality or behavior is the cause.
Supporting the Thoughts:
Reasoning ProcessThe patient has not provided concrete examples or evidence that their personality or behavior is the reason behind the lack of close friendships. it is possible there are external factors or situational circumstances that led to the loss of friendships.
Contradicting the Thoughts:
Cognition Schema:The patient might be engaging in personalization and overgeneralization cognitive distortions. personalization because they seem to be taking the blame for friendships not lasting, assuming it's their fault due to their personality or behavior. overgeneralization because they are making a broad conclusion based on limited information (i.e., the assumption that their personality or behavior is the sole reason for the lack of close friends). this underlying cognition mode could stem from low self-esteem, fear of rejection, or previous negative experiences in social situations.
+ +Figure 3: Showcases of Diagnosis of Thought using ChatGPT and GPT4 given the Patient Message. \ No newline at end of file diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/images.zip b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86492c3de8c37e5f9f9f78a85aa3d78f24bf6605 --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c46ba4eaafe40053653b0622478f5ec360203cc65a10d1f1dea07ad971f9462d +size 633709 diff --git a/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/layout.json b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c43b2765215d9b6953d801c77a402a5decb2494 --- /dev/null +++ b/2023/Empowering Psychotherapy with Large Language Models_ Cognitive Distortion Detection through Diagnosis of Thought Prompting/layout.json @@ -0,0 +1,5801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 75, + 519, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 75, + 519, + 109 + ], + "spans": [ + { + "bbox": [ + 74, + 75, + 519, + 109 + ], + "type": "text", + "content": "Empowering Psychotherapy with Large Language Models: Cognitive Distortion Detection through Diagnosis of Thought Prompting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "spans": [ + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "text", + "content": "Zhiyu Chen" + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "text", + "content": ", Yujie Lu" + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "text", + "content": ", and William Yang Wang" + }, + { + "bbox": [ + 168, + 119, + 426, + 134 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 227, + 135, + 368, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 135, + 368, + 148 + ], + "spans": [ + { + "bbox": [ + 227, + 135, + 368, + 148 + ], + "type": "text", + "content": "Carnegie Mellon University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 148, + 395, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 148, + 395, + 162 + ], + "spans": [ + { + "bbox": [ + 201, + 148, + 395, + 162 + ], + "type": "text", + "content": "2University of California, Santa Barbara" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 232, + 163, + 364, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 163, + 364, + 175 + ], + "spans": [ + { + "bbox": [ + 232, + 163, + 364, + 175 + ], + "type": "text", + "content": "zhiyuc@andrew.cmu.edu," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 208, + 177, + 389, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 177, + 389, + 190 + ], + "spans": [ + { + "bbox": [ + 208, + 177, + 389, + 190 + ], + "type": "text", + "content": "{yujielu, william}@cs.ucsb.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 235, + 274, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 235, + 274, + 523 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 274, + 523 + ], + "type": "text", + "content": "Mental illness remains one of the most critical public health issues of our time, due to the severe scarcity and accessibility limit of professionals. Psychotherapy requires high-level expertise to conduct deep, complex reasoning and analysis on the cognition modeling of the patients. In the era of Large Language Models, we believe it is the right time to develop AI assistance for computational psychotherapy. We study the task of cognitive distortion detection and propose the Diagnosis of Thought (DoT) prompting. DoT performs diagnosis on the patient's speech via three stages: subjectivity assessment to separate the facts and the thoughts; contrastive reasoning to elicit the reasoning processes supporting and contradicting the thoughts; and schema analysis to summarize the cognition schemas. The generated diagnosis rationales through the three stages are essential for assisting the professionals. Experiments demonstrate that DoT obtains significant improvements over ChatGPT for cognitive distortion detection, while generating high-quality rationales approved by human experts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 533, + 154, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 533, + 154, + 545 + ], + "spans": [ + { + "bbox": [ + 68, + 533, + 154, + 545 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "spans": [ + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "text", + "content": "About one in eight people worldwide suffer from mental disorders (World Health Organization, 2022). However, such mental health conditions are severely underserved, due to a number of reasons including the scarcity of mental health professionals, poor quality of services, unaffordable cost, and social stigma (White and Dorman, 2001; Sharma et al., 2020b). Treatment coverage for mental health service use ranges from " + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "text", + "content": " in high-income locations to only " + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "text", + "content": " in low- and lower middle-income countries (Moitra et al., 2022). A recent study from the American Psychological Association (APA)" + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 554, + 291, + 744 + ], + "type": "text", + "content": " found that six in ten psychologists \"no longer have openings for new patients.\"" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 213, + 526, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 525 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 525 + ], + "type": "text", + "content": "To mitigate such situations, there have been consistent efforts on developing automated systems for mental health support, such as sentiment analysis (Rathje et al., 2023) and empathetic chatbots (Welivita et al., 2021; Sharma et al., 2020b; Saha et al., 2022). However, existing works mostly take shallow attempts in a heuristic manner, e.g., analyzing emotions and generating comforting responses. There is still a significant gap for such systems to contribute to real professional psychotherapy, which requires deep studies of the patient's thinking patterns, the establishment of cognition models, and the methods to reconstruct the cognition models. These procedures constitute the core pillars in common classic therapy paradigms like cognitive-behavior therapy (CBT) (Rothbaum et al., 2000; Wright et al., 2017; Beck, 2020) and acceptance and commitment therapy (ACT) (Harris, 2006; Hayes et al., 2011). Most data sources recording the interactions between the patients and licensed professionals are confidential, making it even more challenging to build professional assistance for psychotherapy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 527, + 526, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 527, + 526, + 716 + ], + "spans": [ + { + "bbox": [ + 302, + 527, + 526, + 716 + ], + "type": "text", + "content": "Recent progress in Large Language Model (LLM) development uncovers its astonishing ability in various textual reasoning tasks in zero-shot setting (Kojima et al., 2022; Bang et al., 2023; Chen, 2022). For the psychology domain, ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. We believe it is promising to further exploit such ability to complex mental reasoning and analysis. It is the right time to start developing professional, targeted, and systematic AI assistance for psychotherapy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 772 + ], + "type": "text", + "content": "In this paper, we take the first step by studying the task of cognitive distortion detection, the first core procedure in cognitive behavior therapy (CBT) (Beck, 2020). Inspired by how psychother" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 300, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 300, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 300, + 773 + ], + "type": "text", + "content": "1https://www.theguardian.com/society/2022/nov/21/therapist-shortage-us-psychologists-pandemic" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "4295" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4295-4304 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 70, + 524, + 142 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 524, + 142 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 524, + 142 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 524, + 142 + ], + "type": "table", + "html": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
", + "image_path": "68735549b8e66c1ffcf9bb70c191f9bf5d0ed827d1185b8f5c42ba4713a459e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 148, + 506, + 160 + ], + "lines": [ + { + "bbox": [ + 86, + 148, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 86, + 148, + 506, + 160 + ], + "type": "text", + "content": "Table 1: Three example common cognitive distortion types, taken from (Beck, 2020; Shreevastava and Foltz, 2021)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "text", + "content": "apy professionals perform nuanced diagnosis over the patient's speech, we propose the Diagnosis of Thought (DoT) prompting. In DoT, we diagnose the patient's speech through three stages: (1) subjectivity assessment, (2) contrastive reasoning, and (3) schema analysis. In subjectivity assessment, we distinguish the patient's subjective thoughts from the objective facts; In contrastive reasoning, we elicit the reasoning processes supporting and opposing the patient's thoughts; Finally, in schema analysis, we summarize the underlying thought schema and map it to the cognitive distortion types. We conduct comprehensive experiments using the recent top-performing LLMs. In zero-shot setting, DoT obtains over " + }, + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 69, + 185, + 291, + 455 + ], + "type": "text", + "content": " relative improvements for distortion assessment and classification, respectively, on ChatGPT. Meanwhile, the generated rationales through the three stages grant full interpretability for the diagnosis process, whose quality is further approved by human experts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 457, + 291, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 291, + 565 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 291, + 565 + ], + "type": "text", + "content": "We unveil the great potential of empowering professional psychotherapy with LLM. This exploration serves as a catalyst for a larger initiative; we extend an invitation to both AI and psychotherapy communities to come together in a collaborative effort. Our ultimate goal is to construct professional, safe, AI-driven assistance that can substantially enhance mental health support systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 575, + 244, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 575, + 244, + 588 + ], + "spans": [ + { + "bbox": [ + 67, + 575, + 244, + 588 + ], + "type": "text", + "content": "2 Cognitive Distortion Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 597, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 745 + ], + "type": "text", + "content": "Cognitive Behavior Therapy (CBT) is a well-established therapy paradigm primarily on depression and anxiety disorders (Beck, 2020). In CBT, given the patients' speech or written content, we establish the cognitive model by building the interactions among the situations, thoughts, and emotions. Patients with mental disorders, such as depression or anxiety, tend to form negative thoughts very rapidly and unconsciously, leading to negative emotions which further strengthen their overall negative views and beliefs about the world." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 773 + ], + "type": "text", + "content": "To break this vicious circle, in a typical CBT process, the first core step is to identify those mal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 185, + 526, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 185, + 526, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 526, + 321 + ], + "type": "text", + "content": "adaptive negative thoughts and summarize their underlying schemas, formally known as cognitive distortions. There are generally 10-20 common, well-studied types of cognitive distortions. We present a few types with examples in Table 1, and refer the readers to Appendix A for the full list. Once accurately identify these cognitive distortions, CBT therapists will guide the patients to justify and correct these distortions, so as to gradually reconstruct their cognition models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 322, + 526, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 322, + 526, + 604 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 526, + 604 + ], + "type": "text", + "content": "In the real psychotherapy process, there's a significant amount of textual information including therapeutic conversations and diaries, etc. Such information is often long, highly fragmented, and disorganized, containing multiple types of distortions beyond the toy examples in Table 1. The task of cognitive distortion detection aims to automatically detect the distortion types given such textual information from the patients, in order to assist therapists to enhance their efficiency and productivity. Meanwhile, such detectors can also potentially serve as self-assisting tools for the patients to diagnose their thoughts and conduct CBT practice, upon meeting the robustness and safety requirements. Formally, cognitive distortion detection consists of two steps: 1) Distortion assessment to predict whether the given speech contains cognitive distortions, as a binary classification problem; and 2) Distortion classification to predict the specific distortion types, as a multiclass classification problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 615, + 515, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 615, + 515, + 630 + ], + "spans": [ + { + "bbox": [ + 302, + 615, + 515, + 630 + ], + "type": "text", + "content": "3 The Diagnosis of Thought Prompting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 638, + 526, + 772 + ], + "type": "text", + "content": "On discussing with psychotherapy professionals, we identify the following key stages to diagnose the patient's speech to detect cognitive distortions. Subjectivity Assessment. The patient's speech consists of a mixture of reality (objective facts) and interpretations/opinions (subjective thoughts). In order to perform deep analysis of distorted thinking, we first need to find out which parts of the speech are objective facts and which parts are subjective thoughts. After such an assessment, we summa" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "4296" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 526, + 216 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 526, + 216 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 526, + 216 + ], + "type": "image", + "image_path": "b6a1ace8f70193bc2abc3d20531685a261bb01778d2918a60d8e0e4c8cb1da94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 223, + 525, + 245 + ], + "lines": [ + { + "bbox": [ + 67, + 223, + 525, + 245 + ], + "spans": [ + { + "bbox": [ + 67, + 223, + 525, + 245 + ], + "type": "text", + "content": "Figure 1: The Diagnosis of Thought framework. We strategically prompt the LLM to go through the three diagnosis stages: subjectivity assessment, contrastive reasoning, and schema analysis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 265, + 289, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 265, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 265, + 289, + 291 + ], + "type": "text", + "content": "rize the objective facts into the situations as the evidence base to diagnose the subjective thoughts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 294, + 289, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 294, + 289, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 294, + 289, + 401 + ], + "type": "text", + "content": "Contrastive Reasoning. This stage aims to discover how the patient ascertains the veracity of their subjective thoughts. Based on the situation, we deduct the reasoning processes that supports and contradicts the patient's thoughts respectively. By contrasting two different interpretations based on the same situation, we can identify the thought schemas more clearly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 404, + 289, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 404, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 67, + 404, + 289, + 497 + ], + "type": "text", + "content": "Schema Analysis. This stage aims to study why the patient forms the specific reasoning process. The term \"schema\" refers to the cognitive structures that organize our knowledge, beliefs, and expectations. Understanding what schemas a patient is relying on can reveal much about their cognitive mode and distortions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 501, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 501, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 67, + 501, + 289, + 622 + ], + "type": "text", + "content": "We propose the Diagnosis of Thought (DoT) prompting, guiding the LLM through the above three stages to diagnose the patient's speech. Figure 1 illustrates our framework. We sequentially prompt LLM with three questions for the three stages. After the LLM finishes the generation for all stages, we prompt it with another two questions asking distortion assessment and classification. See Appendix B for all the prompts we use." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 624, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 290, + 772 + ], + "type": "text", + "content": "With DoT prompting, we obtain fully interpretable diagnosis rationales for detecting cognitive distortions. Such interpretability is vital for professionals in real applications. Serving as the diagnosing assistance tool, the diagnosis rationales are essential for the professionals to justify the results. More importantly, they can potentially learn the nuanced thought patterns and schemas derived from the patient's speech through the rationales. This is crucial for the professionals to establish the patient's cognition models more efficiently from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 265, + 421, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 265, + 421, + 279 + ], + "spans": [ + { + "bbox": [ + 302, + 265, + 421, + 279 + ], + "type": "text", + "content": "the vast amount of speech." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 289, + 434, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 289, + 434, + 302 + ], + "spans": [ + { + "bbox": [ + 302, + 289, + 434, + 302 + ], + "type": "text", + "content": "4 Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 310, + 488, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 310, + 488, + 323 + ], + "spans": [ + { + "bbox": [ + 302, + 310, + 488, + 323 + ], + "type": "text", + "content": "4.1 Dataset and experimental settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 327, + 525, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 327, + 525, + 489 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 525, + 489 + ], + "type": "text", + "content": "We experiment on the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), which is annotated by experts based on the Therapist QA dataset. The dataset consists of 2,531 examples of patient speech annotated with ten common types of cognitive distortions, as specified in Appendix A. " + }, + { + "bbox": [ + 302, + 327, + 525, + 489 + ], + "type": "inline_equation", + "content": "63.1\\%" + }, + { + "bbox": [ + 302, + 327, + 525, + 489 + ], + "type": "text", + "content": " of the examples have cognitive distortions, which are annotated with the two dominant ones. We follow the same train-test split in (Shreevastava and Foltz, 2021). For automatic evaluations, we report F-1 for distortion assessment and weighted F-1 for distortion classification." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 498, + 470, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 498, + 470, + 510 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 470, + 510 + ], + "type": "text", + "content": "4.2 Automatic Evaluation Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 515, + 525, + 733 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 515, + 525, + 733 + ], + "spans": [ + { + "bbox": [ + 302, + 515, + 525, + 733 + ], + "type": "text", + "content": "We experiment using three of the recent representative LLMs: ChatGPT (gpt-3.5-turbo)3, Vicuna4, and GPT-4 (OpenAI, 2023). For all methods, we first prompt with a general instruction specifying the task and the target distortion types. We compare our DoT prompting with 1) Directly generating the results, and 2) Zero-Shot CoT prompting (ZCoT) (Kojima et al., 2022). For Vicuna, both ZCoT and DoT exceed the token limits for many examples; we omit the results. We average over five runs and report mean and standard deviation for all experiments. Table 2 shows our main experiment results. As we can see, under the zero-shot setting, due to the challenge of this task, Vicuna and ChatGPT still fall behind the supervised full-training models. The proposed DoT prompting significantly" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 739, + 489, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 739, + 489, + 750 + ], + "spans": [ + { + "bbox": [ + 315, + 739, + 489, + 750 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 315, + 739, + 489, + 750 + ], + "type": "text", + "content": "https://www.kaggle.com/arnmaud/therapist-qa" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 750, + 436, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 750, + 436, + 761 + ], + "spans": [ + { + "bbox": [ + 317, + 750, + 436, + 761 + ], + "type": "text", + "content": "3https://openai.com/blog/chatgpt" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 761, + 474, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 761, + 474, + 772 + ], + "spans": [ + { + "bbox": [ + 317, + 761, + 474, + 772 + ], + "type": "text", + "content": "4https://lmsys.org/blog/2023-03-30-vicuna/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "4297" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 68, + 280, + 222 + ], + "blocks": [ + { + "bbox": [ + 76, + 68, + 280, + 222 + ], + "lines": [ + { + "bbox": [ + 76, + 68, + 280, + 222 + ], + "spans": [ + { + "bbox": [ + 76, + 68, + 280, + 222 + ], + "type": "table", + "html": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
Full training*75.024.0
Vicuna73.810.9511.230.78
ChatGPT73.470.5819.241.00
ChatGPT + ZCoT77.101.2120.211.02
ChatGPT + DoT81.190.1122.250.70
GPT-483.040.5133.860.83
GPT-4 + ZCoT81.971.2133.221.36
GPT-4 + DoT82.770.8134.641.40
", + "image_path": "64b37794b7fcbb3ddd54ace693201c914e625173dd0f70567e5eaf67107a2eda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 267, + 287, + 356 + ], + "blocks": [ + { + "bbox": [ + 67, + 229, + 290, + 252 + ], + "lines": [ + { + "bbox": [ + 67, + 229, + 290, + 252 + ], + "spans": [ + { + "bbox": [ + 67, + 229, + 290, + 252 + ], + "type": "text", + "content": "Table 2: Main Results (Standard deviation in subscript). " + }, + { + "bbox": [ + 67, + 229, + 290, + 252 + ], + "type": "inline_equation", + "content": "(^{\\star})" + }, + { + "bbox": [ + 67, + 229, + 290, + 252 + ], + "type": "text", + "content": " results copied from (Shreevastava and Foltz, 2021)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 267, + 287, + 356 + ], + "lines": [ + { + "bbox": [ + 69, + 267, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 69, + 267, + 287, + 356 + ], + "type": "table", + "html": "
MethodsDistortion Assessment (F-1)Distortion Classification (Weighted F-1)
ChatGPT73.470.5819.241.00
ChatGPT + S179.621.1218.721.95
ChatGPT + S1 + S280.700.4820.111.02
ChatGPT + S1 + S2 + S381.190.1122.250.70
", + "image_path": "796c6a883fb89edc2317fb52a0a02e7fab62c16ecce1ffaa306ff36d821c5d64.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 421, + 291, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 421, + 291, + 637 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 291, + 637 + ], + "type": "text", + "content": "boosts the performance of ChatGPT, with the distortion assessment score surpassing the full-training performance and the distortion classification score approaching the full-training performance. The improvement over ZCoT demonstrates the superiority of our proposed strategic stages compared with rationale generation with no guidance. Astonishingly, GPT-4 surpasses the full-training performance even by a large margin for distortion classification. On the one hand, this may demonstrate the powerful ability of GPT-4 on this task - it can already reach a potential upper bound regardless with DoT promoting or not; On the other hand, since the dataset we use is publicly released on Kaggle and its original data source is from online forums, we do not exclude the possibility of data contamination." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 638, + 291, + 772 + ], + "type": "text", + "content": "Ablation Studies and Analysis. To better understand how each stage of DoT contributes to overall performance, we conduct ablation studies using ChatGPT shown in Table 3. All three stages obtain improvements for distortion assessment and distortion classification. For distortion assessment, subjectivity assessment (S1) gains the largest improvement, as separating objective facts and subjective thoughts is a strong trigger to assess distortions. For distortion classification, Schema Analysis (S3)" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 68, + 518, + 179 + ], + "blocks": [ + { + "bbox": [ + 309, + 68, + 518, + 179 + ], + "lines": [ + { + "bbox": [ + 309, + 68, + 518, + 179 + ], + "spans": [ + { + "bbox": [ + 309, + 68, + 518, + 179 + ], + "type": "image", + "image_path": "6701a369fe9e975803d3a626fd263baf4b30b1ca5278a00d974753f5528791b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 185, + 501, + 197 + ], + "lines": [ + { + "bbox": [ + 326, + 185, + 501, + 197 + ], + "spans": [ + { + "bbox": [ + 326, + 185, + 501, + 197 + ], + "type": "text", + "content": "Figure 2: Distortion classification per class F-1." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 304, + 208, + 523, + 315 + ], + "blocks": [ + { + "bbox": [ + 67, + 363, + 291, + 395 + ], + "lines": [ + { + "bbox": [ + 67, + 363, + 291, + 395 + ], + "spans": [ + { + "bbox": [ + 67, + 363, + 291, + 395 + ], + "type": "text", + "content": "Table 3: Ablation studies on each stage of DoT: we denote Subjectivity Assessment as S1, Contrastive Reasoning as S2, and Schema Analysis as S3." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 304, + 208, + 523, + 315 + ], + "lines": [ + { + "bbox": [ + 304, + 208, + 523, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 523, + 315 + ], + "type": "table", + "html": "
MethodsQualitySubjectivity AssessmentContrastive ReasoningSchema Analysis
ChatGPT +DoTComprehensive68.5%45.0%50.5%
Partially good24.5%30.5%27.0%
Invalid7.0%24.5%22.5%
GPT4 +DoTComprehensive84.5%69.5%76.5%
Partially good13.5%25.5%18.5%
Invalid2.0%5.0%5.0%
", + "image_path": "50a8461a152e245735f7ec185b121a4fde61760020939c2c9402859e11010a35.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 322, + 520, + 333 + ], + "lines": [ + { + "bbox": [ + 306, + 322, + 520, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 322, + 520, + 333 + ], + "type": "text", + "content": "Table 4: Human evaluation results for diagnosis rationales." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 360, + 525, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 360, + 525, + 428 + ], + "spans": [ + { + "bbox": [ + 302, + 360, + 525, + 428 + ], + "type": "text", + "content": "obtains large improvement, as the summarized underlying thought schema can match the definition of the distortion types more accurately. We also analyze the distortion classification results for each distortion type, with results shown in Figure 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 437, + 456, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 437, + 456, + 449 + ], + "spans": [ + { + "bbox": [ + 302, + 437, + 456, + 449 + ], + "type": "text", + "content": "4.3 Human Evaluation Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 455, + 526, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 455, + 526, + 563 + ], + "spans": [ + { + "bbox": [ + 301, + 455, + 526, + 563 + ], + "type": "text", + "content": "As there is no reference available for the diagnosis rationales, we employ psychotherapy experts to assess the quality of the generated rationales. We hired psychotherapy professionals from UpWork5, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 564, + 525, + 752 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 564, + 525, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 564, + 525, + 752 + ], + "type": "text", + "content": "Specifically, we present the patient's speech, all the prompts, and the generated rationales for all stages to human experts. For the generated rationales of each stage, we instruct the experts to choose between: 1) Comprehensive. (Correct and comprehensive.); 2) Partially good. (Reasonable but not comprehensive) 3) Invalid. (Not reasonable.) Table 4 shows the evaluation results on 100 examples for DoT over ChatGPT and GPT-4. We employ two experts for each evaluation and report the average; The agreement rates (the percentage of examples the two experts gave the same rating) for both evaluations are over " + }, + { + "bbox": [ + 302, + 564, + 525, + 752 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 302, + 564, + 525, + 752 + ], + "type": "text", + "content": ". The diagnosis rationales generated for all stages shows decent" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 389, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 389, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 389, + 772 + ], + "type": "text", + "content": "5www.upwork.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "4298" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 98 + ], + "type": "text", + "content": "quality verified by the experts. See Appendix C for more generation examples." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 108, + 160, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 108, + 160, + 121 + ], + "spans": [ + { + "bbox": [ + 68, + 108, + 160, + 121 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 130, + 291, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 130, + 291, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 130, + 291, + 576 + ], + "type": "text", + "content": "Due to the verbal and textual nature of psychotherapy procedures, there have been consistent efforts to employ NLP techniques to assist the mental health domain (Althoff et al., 2016; Abd-Alrazaq et al., 2021, 2019; Valizadeh and Parde, 2022). However, with the knowledge gap between the two communities, most existing works take shallow attempts without deep investigation of the professional psychotherapy knowledge. A majority of the previous studies targets identification of common mental health issues, such as depression and anxiety, from textual contents (Cohan et al., 2018; MacAvaney et al., 2018; Harrigian et al., 2020; Ji et al., 2022; Zanwar et al., 2023; Juhng et al., 2023). Another major category of works study therapeutic conversations with a focus on emotional/empathetic supports (Halder et al., 2017; Sharma et al., 2020a; Atapattu et al., 2022; Mishra et al., 2023), and discourse structures (Cao et al., 2019; Hsu et al., 2023). A few more recent works have started to investigate deeper professional psychotherapy knowledge, e.g., cognitive distortion detection in CBT (Shreevastava and Foltz, 2021; Ding et al., 2022; Lybarger et al., 2022), identifying and reframing unhelpful thoughts (Ziems et al., 2022; Maddela et al., 2023). Early famous system Eliza6 took the initial attempts to emulate a Rogerian psychotherapist. Due to its rule-based nature, the responses are often reflections or rephrasings of the user's statements. If users deviate too much from expected inputs or probe its capabilities, the program may produce irrelevant or nonsensical responses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 577, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 577, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 577, + 291, + 754 + ], + "type": "text", + "content": "Recent progress in large language model reveals that ChatGPT and GPT-4 present very promising performance in the classic Sally-Anne test (Baron-Cohen et al., 1985; Bubeck et al., 2023), evaluating the basic theory of mind capabilities to attribute mental states such as beliefs, emotions, desires, etc. However, latter works also point out the robustness issue of such ability (Shapira et al., 2023). In our work, we are inspired by such enhanced ability and investigate the application of diagnosing patients' thoughts in psychotherapy. Our experiments on the real dataset, not anecdotal examples, show very promising performance both in auto" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 233 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 233 + ], + "type": "text", + "content": "matic and human expert evaluation. We believe it is a very important future research direction to investigate the general cognitive abilities of LLM and make connections to the field of psychology and cognitive science. Theory-wise, we are eager to explore to what extent the LLM can simulate the human cognitive functions, so as to determine the role that language plays in the overall human cognition. Application-wise, a straightforward and encouraging direction should be the integration for assisting mental health treatment, as the motivation and goal of this work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 244, + 474, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 244, + 474, + 258 + ], + "spans": [ + { + "bbox": [ + 302, + 244, + 474, + 258 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 266, + 526, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 266, + 526, + 441 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 526, + 441 + ], + "type": "text", + "content": "This work delved into the integration of large language models (LLMs) within the realm of psychotherapy. We introduced the Diagnosis of Thought (DoT) framework, which strategically prompts the LLM to produce diagnosis rationales pertinent to the detection of cognitive distortions. We believe that there exists substantial potential for leveraging LLMs within numerous facets of mental health support that are currently under-explored. Our findings, thus, not only illuminate a path towards more efficient therapy methods but also open doors for future investigations to push the boundaries of AI's role in mental health treatment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 443, + 526, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 443, + 526, + 673 + ], + "spans": [ + { + "bbox": [ + 302, + 443, + 526, + 673 + ], + "type": "text", + "content": "In the domain of AI-enhanced psychotherapy, particularly when harnessing the power of LLMs, there are paramount challenges surrounding safety, robustness, and ethical responsibilities. LLM is known to generate hallucinations and biases. While one potential usage of our system is to serve as a self-diagnosing/assisting tool for patients, e.g., to help the patients to recognize their own thinking patterns and learn how the cognitive distortions developed, such tools should not be deployed for direct patient use without the supervision of professionals. As we navigate the potential of LLM in psychotherapy, our foremost priority should remain in building systems that are safe and responsible. Concurrently, the formulation of robust ethical guidelines tailored to this new era becomes indispensable." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 683, + 365, + 697 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 683, + 365, + 697 + ], + "spans": [ + { + "bbox": [ + 303, + 683, + 365, + 697 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 706, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 526, + 773 + ], + "type": "text", + "content": "One of the biggest obstacles to studying AI for psychotherapy is the lack of available, high-quality data. Most of the datasets recording the information of the interactions between the patients and licensed professionals are confidential due to pri" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 219, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 219, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 219, + 772 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 80, + 760, + 219, + 772 + ], + "type": "text", + "content": "https://en.wikipedia.org/wiki/ELIZA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "4299" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 260 + ], + "type": "text", + "content": "vacy concerns. The dataset used in this work is the only one we can find publicly available for cognitive distortion detection, as its data source is from public online forums. In addition to the patient's speech, the patient's demographic information also plays an important role in analyzing the thought process leading to cognitive distortions. In this dataset, some of the patient speech include a few demographic details, but pretty minimal. Beyond the task of cognitive distortion detection, the same data constraint issue exists for all other therapy stages and paradigms. This also motivates us to work on building privacy-preserved systems as a crucial future direction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 272, + 158, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 272, + 158, + 285 + ], + "spans": [ + { + "bbox": [ + 68, + 272, + 158, + 285 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 295, + 290, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 295, + 290, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 295, + 290, + 375 + ], + "type": "text", + "content": "The system built in this work are never meant to replace psychotherapy professionals: all system-produced results need to be verified by licensed professionals. The system should not be deployed for direct patient use without the supervision of licensed professionals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 377, + 291, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 377, + 291, + 565 + ], + "spans": [ + { + "bbox": [ + 69, + 377, + 291, + 565 + ], + "type": "text", + "content": "There has been a lot of discussions regarding the use of LLMs for AI-driven therapy purpose. Chatbots will find their most effective role in medicine within the realm of mental health, an opinion quoted from Thomas Insel, former director of the National Institute of Mental Health. In the meantime, we should always take caution for such development. Our foremost priority should be building safe and responsible applications while establishing ethical and regulation standards for this new era. Under the current status, the partnership between professionals and AI systems is a promising direction for the goal to ease the burden of professionals, which is the motivation of this work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 566, + 291, + 742 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 566, + 291, + 742 + ], + "spans": [ + { + "bbox": [ + 67, + 566, + 291, + 742 + ], + "type": "text", + "content": "\"Diagnosis of Thought\" is an innovative approach that harnesses the capabilities of large language models to enhance the field of psychotherapy. It is imperative to understand that this tool is not designed for censorship or any form of invasive surveillance. Instead, it aims to augment the skills of therapists by providing them with advanced, data-driven insights into their clients' thoughts and emotions. Through the efficient processing of linguistic cues and patterns, these models aid mental health professionals in formulating more precise diagnoses and crafting tailored therapeutic interventions that could significantly enhance the well" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 71, + 453, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 453, + 84 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 453, + 84 + ], + "type": "text", + "content": "being of individuals seeking help." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 85, + 526, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 85, + 526, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 85, + 526, + 179 + ], + "type": "text", + "content": "For all the annotations and human evaluations presented in this work, we hired psychotherapy professionals from UpWork8, e.g., certified clinicians, counseling psychology Ph.D. students, etc. We discussed with each hire to reach an agreement on the payment following a fair wage rate. The average hourly rate for all the experts is $80." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 180, + 525, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 180, + 525, + 221 + ], + "spans": [ + { + "bbox": [ + 302, + 180, + 525, + 221 + ], + "type": "text", + "content": "This project is approved by our Institutional Review Board (IRB). The data annotation is classified as exempt by our IRB." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 244, + 362, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 362, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 362, + 257 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 303, + 264, + 527, + 753 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 264, + 527, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 527, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 527, + 330 + ], + "type": "text", + "content": "Alaa A Abd-Alrazaq, Mohannad Alajlani, Ali Abdallah Alalwan, Bridgette M Bewick, Peter Gardner, and Mowafa Househ. 2019. An overview of the features of chatbots in mental health: A scoping review. International Journal of Medical Informatics, 132:103978." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 527, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 527, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 527, + 396 + ], + "type": "text", + "content": "Alaa A Abd-Alrazaq, Mohammad Alajlani, Nashva Ali, Kerstin Denecke, Bridgette M Bewick, and Mowafa Househ. 2021. Perceptions and opinions of patients about mental health chatbots: scoping review. Journal of medical Internet research, 23(1):e17828." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 404, + 527, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 527, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 527, + 460 + ], + "type": "text", + "content": "Tim Althoff, Kevin Clark, and Jure Leskovec. 2016. Large-scale analysis of counseling conversations: An application of natural language processing to mental health. Transactions of the Association for Computational Linguistics, 4:463-476." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 470, + 527, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 527, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 527, + 570 + ], + "type": "text", + "content": "Thushari Atapattu, Mahen Herath, Charitha Elvitigala, Piyanjali de Zoysa, Kasun Gunawardana, Menasha Thilakaratne, Kasun de Zoysa, and Katrina Falkner. 2022. EmoMent: An emotion annotated mental health corpus from two South Asian countries. In Proceedings of the 29th International Conference on Computational Linguistics, pages 6991-7001, Gyeongju, Republic of Korea. International Committee on Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 578, + 527, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 527, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 527, + 646 + ], + "type": "text", + "content": "Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wen-liang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, et al. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. arXiv preprint arXiv:2302.04023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 303, + 655, + 526, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 655, + 526, + 689 + ], + "spans": [ + { + "bbox": [ + 303, + 655, + 526, + 689 + ], + "type": "text", + "content": "Simon Baron-Cohen, Alan M Leslie, and Uta Frith. 1985. Does the autistic child have a \"theory of mind\"? Cognition, 21(1):37-46." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 698, + 526, + 720 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 698, + 526, + 720 + ], + "spans": [ + { + "bbox": [ + 303, + 698, + 526, + 720 + ], + "type": "text", + "content": "Judith S Beck. 2020. Cognitive behavior therapy: Basics and beyond. Guilford Publications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 729, + 526, + 753 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 729, + 526, + 753 + ], + "spans": [ + { + "bbox": [ + 303, + 729, + 526, + 753 + ], + "type": "text", + "content": "Sebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar," + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 772 + ], + "type": "text", + "content": "7https://www.scientificamerican.com/article/ai-chatbots-could-help-provide-therapy-but-caution-is-needed/" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 760, + 389, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 760, + 389, + 772 + ], + "spans": [ + { + "bbox": [ + 316, + 760, + 389, + 772 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 316, + 760, + 389, + 772 + ], + "type": "text", + "content": "www.upwork.com" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "4300" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 116 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 116 + ], + "type": "text", + "content": "Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. 2023. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 128, + 290, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 128, + 290, + 172 + ], + "spans": [ + { + "bbox": [ + 69, + 128, + 290, + 172 + ], + "type": "text", + "content": "Jie Cao, Michael Tanana, Zac E Imel, Eric Poitras, David C Atkins, and Vivek Srikumar. 2019. Observing dialogue in therapy: Categorizing and forecasting behavioral codes. arXiv preprint arXiv:1907.00326." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 184, + 289, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 184, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 69, + 184, + 289, + 216 + ], + "type": "text", + "content": "Wenhu Chen. 2022. Large language models are few (1)-shot table reasoners. arXiv preprint arXiv:2210.06710." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 228, + 290, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 228, + 290, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 228, + 290, + 317 + ], + "type": "text", + "content": "Arman Cohan, Bart Desmet, Andrew Yates, Luca Soldaini, Sean MacAvaney, and Nazli Goharian. 2018. SMHD: a large-scale resource for exploring online language usage for multiple mental health conditions. In Proceedings of the 27th International Conference on Computational Linguistics, pages 1485-1497, Santa Fe, New Mexico, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 327, + 290, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 327, + 290, + 427 + ], + "spans": [ + { + "bbox": [ + 69, + 327, + 290, + 427 + ], + "type": "text", + "content": "Xiruo Ding, Kevin Lybarger, Justin Tauscher, and Trevor Cohen. 2022. Improving classification of infrequent cognitive distortions: Domain-specific model vs. data augmentation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 68-75, Hybrid: Seattle, Washington + Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 439, + 290, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 439, + 290, + 527 + ], + "spans": [ + { + "bbox": [ + 69, + 439, + 290, + 527 + ], + "type": "text", + "content": "Kishaloy Halder, Lahari Poddar, and Min-Yen Kan. 2017. Modeling temporal progression of emotional status in mental health forum: A recurrent neural net approach. In Proceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pages 127-135, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 538, + 290, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 538, + 290, + 604 + ], + "spans": [ + { + "bbox": [ + 69, + 538, + 290, + 604 + ], + "type": "text", + "content": "Keith Harrigian, Carlos Aguirre, and Mark Dredze. 2020. Do models of mental health based on social media data generalize? In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3774-3788, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 616, + 290, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 290, + 650 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 290, + 650 + ], + "type": "text", + "content": "Russell Harris. 2006. Embracing your demons: An overview of acceptance and commitment therapy. Psychotherapy in Australia, 12(4):70-6." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 661, + 290, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 661, + 290, + 705 + ], + "spans": [ + { + "bbox": [ + 69, + 661, + 290, + 705 + ], + "type": "text", + "content": "Steven C Hayes, Kirk D Strosahl, and Kelly G Wilson. 2011. Acceptance and commitment therapy: The process and practice of mindful change. Guilford press." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 716, + 290, + 772 + ], + "type": "text", + "content": "Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, Zahra Ashktorab, Casey Dugan, Werner Geyer, and Diyi Yang. 2023. Helping the helper: Supporting peer counselors via ai-empowered practice and feedback. arXiv preprint arXiv:2305.08982." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 149 + ], + "type": "text", + "content": "Shaoxiong Ji, Tianlin Zhang, Luna Ansari, Jie Fu, Prayag Tiwari, and Erik Cambria. 2022. MentalBERT: Publicly available pretrained language models for mental healthcare. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 7184-7190, Marseille, France. European Language Resources Association." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 158, + 525, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 525, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 525, + 236 + ], + "type": "text", + "content": "Swanie Juhng, Matthew Matero, Vasudha Varadarajan, Johannes Eichstaedt, Adithya V Ganesan, and H Andrew Schwartz. 2023. Discourse-level representations can improve prediction of degree of anxiety. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1500-1511." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 244, + 525, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 525, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 525, + 288 + ], + "type": "text", + "content": "Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. arXiv preprint arXiv:2205.11916." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 297, + 525, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 525, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 525, + 364 + ], + "type": "text", + "content": "Kevin Lybarger, Justin Tauscher, Xiruo Ding, Dror Ben-Zeev, and Trevor Cohen. 2022. Identifying distorted thinking in patient-therapist text message exchanges by leveraging dynamic multi-turn context. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pages 126-136." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 373, + 525, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 525, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 525, + 461 + ], + "type": "text", + "content": "Sean MacAvaney, Bart Desmet, Arman Cohan, Luca Soldaini, Andrew Yates, Ayah Zirikly, and Nazli Goharian. 2018. RSDD-time: Temporal annotation of self-reported mental health diagnoses. In Proceedings of the Fifth Workshop on Computational Linguistics and Clinical Psychology: From Keyboard to Clinic, pages 168-173, New Orleans, LA. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 470, + 525, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 525, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 525, + 525 + ], + "type": "text", + "content": "Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. arXiv preprint arXiv:2307.02768." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 534, + 525, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 525, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 525, + 623 + ], + "type": "text", + "content": "Kshitij Mishra, Priyanshu Priya, and Asif Ekbal. 2023. PAL to lend a helping hand: Towards building an emotion adaptive polite and empathetic counseling conversational agent. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12254-12271, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 632, + 525, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 525, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 525, + 708 + ], + "type": "text", + "content": "Modhurima Moitra, Damian Santomauro, Pamela Y Collins, Theo Vos, Harvey Whiteford, Shekhar Saxena, and Alize J Ferrari. 2022. The global gap in treatment coverage for major depressive disorder in 84 countries from 2000-2019: A systematic review and bayesian meta-regression analysis. PLoS Medicine, 19(2):e1003901." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 718, + 525, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 718, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 718, + 525, + 740 + ], + "type": "text", + "content": "OpenAI. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "text", + "content": "Steve Rathje, Dan-Mircea Mirea, Ilia Sucholutsky, Raja Marjieh, Claire Robertson, and Jay J Van Bavel. 2023." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "4301" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 80, + 72, + 289, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 289, + 95 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 289, + 95 + ], + "type": "text", + "content": "Gpt is an effective tool for multilingual psychological text analysis." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 103, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 69, + 103, + 290, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 290, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 290, + 137 + ], + "type": "text", + "content": "Barbara Olasov Rothbaum, Elizabeth A Meadows, Patricia Resick, and David W Foy. 2000. Cognitive-behavioral therapy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 145, + 289, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 145, + 289, + 222 + ], + "spans": [ + { + "bbox": [ + 69, + 145, + 289, + 222 + ], + "type": "text", + "content": "Tulika Saha, Saichethan Reddy, Anindya Das, Sriparna Saha, and Pushpak Bhattacharyya. 2022. A shoulder to cry on: Towards a motivational virtual assistant for assuaging mental agony. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2436-2449." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 230, + 289, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 230, + 289, + 295 + ], + "spans": [ + { + "bbox": [ + 69, + 230, + 289, + 295 + ], + "type": "text", + "content": "Natalie Shapira, Mosh Levy, Seyed Hossein Alavi, Xuhui Zhou, Yejin Choi, Yoav Goldberg, Maarten Sap, and Vered Shwartz. 2023. Clever hans or neural theory of mind? stress testing social reasoning in large language models. arXiv preprint arXiv:2305.14763." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 305, + 289, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 305, + 289, + 381 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 289, + 381 + ], + "type": "text", + "content": "Ashish Sharma, Adam Miner, David Atkins, and Tim Althoff. 2020a. A computational approach to understanding empathy expressed in text-based mental health support. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5263-5276, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 69, + 391, + 289, + 434 + ], + "type": "text", + "content": "Ashish Sharma, Adam S Miner, David C Atkins, and Tim Althoff. 2020b. A computational approach to understanding empathy expressed in text-based mental health support. arXiv preprint arXiv:2009.08441." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 443, + 289, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 443, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 69, + 443, + 289, + 498 + ], + "type": "text", + "content": "Sagarika Shreevastava and Peter Foltz. 2021. Detecting cognitive distortions from patient-therapist interactions. In Proceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access, pages 151-158." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 507, + 289, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 507, + 289, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 507, + 289, + 571 + ], + "type": "text", + "content": "Mina Valizadeh and Natalie Parde. 2022. The ai doctor is in: A survey of task-oriented dialogue systems for healthcare applications. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6638-6660." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 581, + 289, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 581, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 69, + 581, + 289, + 636 + ], + "type": "text", + "content": "Anuradha Welivita, Yubo Xie, and Pearl Pu. 2021. A large-scale dataset for empathetic response generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1251-1264." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 645, + 289, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 645, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 69, + 645, + 289, + 677 + ], + "type": "text", + "content": "Marsha White and Steve M Dorman. 2001. Receiving social support online: implications for health education. Health education research, 16(6):693-707." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 687, + 289, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 687, + 289, + 730 + ], + "spans": [ + { + "bbox": [ + 69, + 687, + 289, + 730 + ], + "type": "text", + "content": "Jesse H Wright, Gregory K Brown, Michael E Thase, and Monica Ramirez Basco. 2017. Learning cognitive-behavior therapy: An illustrated guide. American Psychiatric Pub." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 739, + 289, + 772 + ], + "type": "text", + "content": "Sourabh Zanwar, Xiaofei Li, Daniel Wiechmann, Yu Qiao, and Elma Kerz. 2023. What to fuse and how to fuse: Exploring emotion and personality fusion" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 158 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 72, + 524, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 524, + 116 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 524, + 116 + ], + "type": "text", + "content": "In Findings of the Association for Computational Linguistics: ACL 2023, pages 8926-8940, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 124, + 524, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 124, + 524, + 158 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 524, + 158 + ], + "type": "text", + "content": "Caleb Ziems, Minzhi Li, Anthony Zhang, and Diyi Yang. 2022. Inducing positive perspectives with text reframing. arXiv preprint arXiv:2204.02952." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 167, + 510, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 510, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 510, + 180 + ], + "type": "text", + "content": "Appendix A: Cognitive Distortion Types" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 188, + 524, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 524, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 524, + 268 + ], + "type": "text", + "content": "We list all the ten cognitive distortion types studies in this work following-(Shreevastava and Foltz, 2021) in Table 5. Note that these ten types being studied are some common, well-studied ones. As psychotherapy develops, more fine-grained and new distortion types are expected to be unveiled." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 270, + 524, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 524, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 524, + 349 + ], + "type": "text", + "content": "For the cognitive distortion detection dataset proposed by Shreevastava and Foltz (2021), there are 2,531 examples in total. We follow the original " + }, + { + "bbox": [ + 304, + 270, + 524, + 349 + ], + "type": "inline_equation", + "content": "80\\% - 20\\%" + }, + { + "bbox": [ + 304, + 270, + 524, + 349 + ], + "type": "text", + "content": " train-test split. The input patient speech has an average length of 167.3 tokens, and the ten distortion types are roughly equally distributed." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 359, + 451, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 451, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 451, + 373 + ], + "type": "text", + "content": "Appendix B: Prompt Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 381, + 524, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 524, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 524, + 418 + ], + "type": "text", + "content": "For all our experiments without using DoT, we first prompt the LLM with the following general instructions:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 422, + 524, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 524, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 524, + 514 + ], + "type": "text", + "content": "Given a speech of a patient, our task is to 1) identify if there is cognitive distortion in the speech; 2) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 517, + 524, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 524, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 524, + 554 + ], + "type": "text", + "content": "For all our experiments using DoT, we first prompt the LLM with the following general instructions:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 557, + 524, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 524, + 690 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 524, + 690 + ], + "type": "text", + "content": "Given a speech of a patient, our task is to 1) finish a few diagnose of thought questions to analyze the thought patterns of the patient. Then based on the diagnose of thought analysis, 2) identify if there is cognitive distortion in the speech; 3) Recognizing the specific types of the cognitive distortion. Here we consider the following common distortions: (followed by the descriptions and examples for all ten prompts in the dataset metadata of (Shreevastava and Foltz, 2021)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 692, + 524, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 692, + 524, + 718 + ], + "spans": [ + { + "bbox": [ + 304, + 692, + 524, + 718 + ], + "type": "text", + "content": "Then for methods without using DoT, we prompt the following two questions:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 719, + 524, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 719, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 719, + 524, + 772 + ], + "type": "text", + "content": "Please first answer: if there is cognitive distortion in the speech; Answer 'yes' or 'no'; Please then answer: Recognizing the specific types of the cognitive distortion in the speech. There may be" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "4302" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 523, + 274 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 523, + 274 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 523, + 274 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 523, + 274 + ], + "type": "table", + "html": "
Cognitive Distortion TypeInterpretationExample Distorted Speech
PersonalizationPersonalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person's control.My son is pretty quiet today. I wonder what I did to upset him.
Mind ReadingSuspecting what others are thinking or what are the motivations behind their actions.My house was dirty when my friends came over, they must think I'm a slob!
OvergeneralizationMajor conclusions are drawn based on limited information.Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.
All-or-nothing thinkingLooking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.If I cannot get my Ph.D., then I am a total failure.
Emotional reasoningLetting one's feeling about something overrule facts to the contrary.Even though Steve is here at work late every day, I know I work harder than anyone else at my job.
LabelingGiving someone or something a label without finding out more about it/them.My daughter would never do anything I disapproved of.
MagnificationEmphasizing the negative or playing down the positive of a situation.My professor said he made some corrections on my paper, so I know I'll probably fail the class.
Mental filterPlacing all one's attention o, or seeing only, the negatives of a situation.My husband says he wishes I was better at housekeeping, so I must be a lousy wife.
Should statementsShould statements appear as a list of ironclad rules about how a person should behave, this could be about the speaker themselves or other.It is NOT necessary that the word 'should' or it's synonyms (ought to, must etc.) be present in the statements containing this distortion.I should get all A's to be a good student.
Fortune-tellingAs the name suggests, this distortion is about expecting things to happen a certain way, or assuming that thing will go badlyCounterintuitively, this distortion does not always have future tense.I was afraid of job interviews so I decided to start my own thing.
", + "image_path": "8f982077edf97013885339044a1dbffa60ad42ea5b778adf88bb02953ea7c2f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 76, + 280, + 515, + 292 + ], + "lines": [ + { + "bbox": [ + 76, + 280, + 515, + 292 + ], + "spans": [ + { + "bbox": [ + 76, + 280, + 515, + 292 + ], + "type": "text", + "content": "Table 5: Common cognitive distortion types and example speech, taken from (Beck, 2020; Shreevastava and Foltz, 2021)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 318, + 290, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 318, + 290, + 371 + ], + "spans": [ + { + "bbox": [ + 67, + 318, + 290, + 371 + ], + "type": "text", + "content": "one type of cognitive distortion or multiple types involved. If there are multiple types, please give the top 2 dominant ones. Please only give the distortion type names separated by comma." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 373, + 290, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 373, + 290, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 373, + 290, + 548 + ], + "type": "text", + "content": "For methods using DoT, before asking this two questions, we provide the following DoT prompts: Based on the patient's speech, finish the following diagnosis of thought questions: 1. what is the situation? Find out the facts that are objective; what is the patient thinking or imagining? Find out the thoughts or opinions that are subjective. 2. what makes the patient think the thought is true or is not true? Find out the reasoning processes that support and do not support these thoughts. 3. why does the patient come up with such reasoning process supporting the thought? What's the underlying cognition mode of it?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 558, + 250, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 558, + 250, + 572 + ], + "spans": [ + { + "bbox": [ + 68, + 558, + 250, + 572 + ], + "type": "text", + "content": "Appendix C: Generation Examples" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 579, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 579, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 579, + 289, + 606 + ], + "type": "text", + "content": "We present some generation examples of GPT-3.5 and GPT-4 in Figure 3." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "4303" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 77, + 163, + 512, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 163, + 512, + 188 + ], + "spans": [ + { + "bbox": [ + 77, + 163, + 512, + 188 + ], + "type": "text", + "content": "Patient Message: Lately I have been extremely unhappy for no apparent reason. I feel empty almost all the time. It is like nothing can give me genuine joy anymore. I also feel alone a lot, even though I do have friends and close friends. Whenever I try to imagine my future all I see is one full of obstacles. I have become so consumed in sadness and I do not know how to deal anymore. I feel so confused and I do not understand myself anymore." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 214, + 284, + 384 + ], + "blocks": [ + { + "bbox": [ + 129, + 202, + 232, + 211 + ], + "lines": [ + { + "bbox": [ + 129, + 202, + 232, + 211 + ], + "spans": [ + { + "bbox": [ + 129, + 202, + 232, + 211 + ], + "type": "text", + "content": "Diagnosis of Thought (ChatGPT):" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 214, + 284, + 384 + ], + "lines": [ + { + "bbox": [ + 76, + 214, + 284, + 384 + ], + "spans": [ + { + "bbox": [ + 76, + 214, + 284, + 384 + ], + "type": "table", + "html": "
Objective Fact:The patient has been feeling extremely unhappy and empty lately.
Subjective Thought:Nothing can give the patient genuine joy and that the future is full of obstacles. the patient feels confused and does not understand themselves anymore.
Reasoning Process Supporting the Thoughts:The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.
Reasoning Process Contradicting the Thoughts:There may not be factual evidence to support the belief that nothing can give the patient genuine joy or that the future is full of obstacles. the patient may be using emotional reasoning to support their beliefs.
Cognition Schema:The patient may be experiencing a cognitive mode of negativity bias, where they tend to focus on negative experiences and thoughts rather than positive ones. this may be influencing their reasoning processes, leading them to magnify the negative and filter out the positive aspects of their life and future. additionally, the patient may be struggling with depression, which can impact the way they perceive themselves and the world around them.
", + "image_path": "40abdb6121961be0f103ec15bb22af662c36468675c95535ffd652493f4385aa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 302, + 214, + 512, + 398 + ], + "blocks": [ + { + "bbox": [ + 376, + 202, + 467, + 211 + ], + "lines": [ + { + "bbox": [ + 376, + 202, + 467, + 211 + ], + "spans": [ + { + "bbox": [ + 376, + 202, + 467, + 211 + ], + "type": "text", + "content": "Diagnosis of Thought (GPT4):" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 214, + 512, + 398 + ], + "lines": [ + { + "bbox": [ + 302, + 214, + 512, + 398 + ], + "spans": [ + { + "bbox": [ + 302, + 214, + 512, + 398 + ], + "type": "table", + "html": "
Objective Fact:The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.
Subjective Thought:They believe nothing can bring them genuine joy anymore. They imagine their future full of obstacles. They feel consumed by sadness and confusion and no longer understand themselves.
Reasoning ProcessThe persistent feeling of emptiness and lack of joy might make the patient believe that their situation is unchanging and will remain the same. They may be experiencing negative events in their lives, making it harder to feel hopeful about their future.
Supporting the Thoughts:Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.
Reasoning ProcessEmotions and feelings are temporary and can change over time, so it is not guaranteed that the patient will always feel this way. The patient is aware that they do have friends, which suggests that some level of meaningful connection exists. Obstacles are a natural part of life, but they are not the only things present in one's future. positive events and achievements are also likely to occur.
Contradicting the Thoughts:
Cognition Schema:Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life's reality. Mental filter: they are focusing on the negative aspects of their life and filtering out positive experiences or memories. Fortune-telling: they are assuming that their future will be full of obstacles based on their current emotions and thoughts.
", + "image_path": "717af0f76ecb1adc154c09e3d9bc4a5505c5aafc8f5d95b785a5537082a34601.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 419, + 514, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 419, + 514, + 445 + ], + "spans": [ + { + "bbox": [ + 77, + 419, + 514, + 445 + ], + "type": "text", + "content": "Patient Message: So I have no close friends at the moment probably I know some people enough just to say hi but no close one and whenever I make any they go away I don't know if its my personality the reason or the way I behave and the only friends I can gain are school friends since there is no other place I go to I can't really stand myself alone. Thanks for helping." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 76, + 473, + 284, + 622 + ], + "blocks": [ + { + "bbox": [ + 129, + 459, + 230, + 468 + ], + "lines": [ + { + "bbox": [ + 129, + 459, + 230, + 468 + ], + "spans": [ + { + "bbox": [ + 129, + 459, + 230, + 468 + ], + "type": "text", + "content": "Diagnosis of Thought (ChatGPT):" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 473, + 284, + 622 + ], + "lines": [ + { + "bbox": [ + 76, + 473, + 284, + 622 + ], + "spans": [ + { + "bbox": [ + 76, + 473, + 284, + 622 + ], + "type": "table", + "html": "
Objective Fact:The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.
Subjective Thought:The patient feels that they are the reason for their friends leaving and wonders if it's their personality or behavior causing the issue. they cannot stand being alone.
Reasoning Process Supporting the Thoughts:The patient believes their thoughts to be true because they have not been able to maintain any close friendships and the pattern keeps repeating itself. they may be using past experiences to support this belief.
Reasoning Process Contradicting the Thoughts:There may be other reasons for the loss of these friendships that the patient is not considering.
Cognition Schema:The patient may be experiencing cognitive distortions such as overgeneralization by assuming that all friendships will end up the same way and mental filtering by focusing only on the negative aspects of their past friendships. the underlying cognition mode could be a negative self-image or low self-esteem.
", + "image_path": "06996da8ebd71b31ea6ec23a611ea9ac992024be031636ce847cb1d2a278188d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 302, + 470, + 512, + 645 + ], + "blocks": [ + { + "bbox": [ + 376, + 459, + 467, + 468 + ], + "lines": [ + { + "bbox": [ + 376, + 459, + 467, + 468 + ], + "spans": [ + { + "bbox": [ + 376, + 459, + 467, + 468 + ], + "type": "text", + "content": "Diagnosis of Thought (GPT4):" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 302, + 470, + 512, + 645 + ], + "lines": [ + { + "bbox": [ + 302, + 470, + 512, + 645 + ], + "spans": [ + { + "bbox": [ + 302, + 470, + 512, + 645 + ], + "type": "table", + "html": "
Objective Fact:The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.
Subjective Thought:The patient believes that their personality or behavior might be the reason their friendships don't last, and they are limited in opportunities to make new friends.
Reasoning ProcessThe patient might have experienced failed friendships and noticed a pattern of people leaving their life. this leads them to question if their personality or behavior is the cause.
Supporting the Thoughts:
Reasoning ProcessThe patient has not provided concrete examples or evidence that their personality or behavior is the reason behind the lack of close friendships. it is possible there are external factors or situational circumstances that led to the loss of friendships.
Contradicting the Thoughts:
Cognition Schema:The patient might be engaging in personalization and overgeneralization cognitive distortions. personalization because they seem to be taking the blame for friendships not lasting, assuming it's their fault due to their personality or behavior. overgeneralization because they are making a broad conclusion based on limited information (i.e., the assumption that their personality or behavior is the sole reason for the lack of close friends). this underlying cognition mode could stem from low self-esteem, fear of rejection, or previous negative experiences in social situations.
", + "image_path": "7bc5968054c6082d595c374e0e54f63da73f7967346390a60b5ac3c233932a66.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 674, + 477, + 686 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 674, + 477, + 686 + ], + "spans": [ + { + "bbox": [ + 115, + 674, + 477, + 686 + ], + "type": "text", + "content": "Figure 3: Showcases of Diagnosis of Thought using ChatGPT and GPT4 given the Patient Message." + } + ] + } + ], + "index": 10, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "4304" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_content_list.json b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a7f2037f00aad51f0a65b19ffe070a3a5fcb2748 --- /dev/null +++ b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_content_list.json @@ -0,0 +1,756 @@ +[ + { + "type": "text", + "text": "Emptying the Ocean with a Spoon: Should We Edit Models?", + "text_level": 1, + "bbox": [ + 188, + 89, + 811, + 110 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuval Pinter and Michael Elhadad \nDepartment of Computer Science \nBen-Gurion University of the Negev \nBe'er Sheva, Israel \n{uvp, elhadad}@cs.bgu.ac.il", + "bbox": [ + 347, + 130, + 655, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 263, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We call into question the recently popularized method of direct model editing as a means of correcting factual errors in LLM generations. We contrast model editing with three similar but distinct approaches that pursue better defined objectives: (1) retrieval-based architectures, which decouple factual memory from inference and linguistic capabilities embodied in LLMs; (2) concept erasure methods, which aim at preventing systemic bias in generated text; and (3) attribution methods, which aim at grounding generations into identified textual sources.", + "bbox": [ + 144, + 277, + 458, + 447 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We argue that direct model editing cannot be trusted as a systematic remedy for the disadvantages inherent to LLMs, and while it has proven potential in improving model explainability, it opens risks by reinforcing the notion that models can be trusted for factuality. We call for cautious promotion and application of model editing as part of the LLM deployment process, and for responsibly limiting the use cases of LLMs to those not relying on editing as a critical component.", + "bbox": [ + 144, + 453, + 458, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 115, + 620, + 258, + 634 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models, or LLMs, have taken the NLP world by storm. After originally focusing on training them as a vehicle for transfer learning, recent advancements have cast LLMs in the role of one-stop shop, all-knowing oracles. One particular finding that has contributed to this reframed usage is the apparent factuality properties of pre-trained LLMs: somehow, mere next-word prediction training has produced models that can complete certain correct facts about the world when asked to do so. LLMs have been perceived by the public at large as a replacement for search engines, and the scant disclaimers appearing in commercial services offering LLM querying do not appear to have made a dent in this perception.", + "bbox": [ + 115, + 645, + 487, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these developments, LLMs of course keep making mistakes. There is, after all, an inher", + "bbox": [ + 115, + 887, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ent mismatch between their pre-training objective and the desire for factuality. In recent years, researchers have come up with several remedies to the problem of nonfactual LLM outputs, one of which being model editing, where parameters inside an LLM are tweaked based on individual facts marked for correction. These works focus on solving problems such as ensuring stability of other factual outputs following the editing, or batch editing, or computationally efficient editing. In this opinion paper, we call into question the entire idea of model editing, citing concerns about intended use cases, conceptual scalability, potential for bias, safety, and overall accountability. We advocate using models that have explicit knowledge modules for tasks requiring factual knowledge, and other existing techniques to mitigate the need for model editing in the first place. Having said that, we acknowledge the usefulness of direct editing for certain use cases such as interpretability probes, and recommend limiting editing to such scenarios.", + "bbox": [ + 512, + 253, + 882, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Model Editing", + "text_level": 1, + "bbox": [ + 512, + 602, + 667, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sinitsin et al. (2020) first introduce the notion of updating large ML models to account for local performance expectations motivated externally. They cite cases where mistakes are critical, like object detection in self-driving cars. Later work suggests model editing can aid in protecting privacy and eliminating bias (Zhu et al., 2020), and as a measure of \"catching up\" with time-sensitive facts such as \"PM of the UK\" changing over time (Bommasani et al., 2021; Mitchell et al., 2022), or as a means of understanding the mechanics of black box models (Meng et al., 2022). Sinitsin et al. (2020) specify several desired properties of editing methods: reliability (the target facts are updated as intended), locality (other changes don't happen; a measure for this property is termed drawdown), and efficiency (successful edits are computationally light); later work (De Cao et al., 2021) added generality (the", + "bbox": [ + 510, + 629, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ability to modify models not originally trained for knowledge retention), consistency (robustness to paraphrasing, in the specific use case of text models), and an unspecified property we'll call frugality (only minimal components of the model are changed during editing). A well-studied limitation, which most of these works focus on mitigating, and related to locality, is catastrophic forgetting (Ratcliff, 1990), i.e., ensuring that an edited model does not lose performance on tasks for which it was explicitly trained and performs well on.", + "bbox": [ + 115, + 84, + 485, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods for model editing have evolved from editable training (Sinitsin et al., 2020), a procedure requiring an a-priori decision that a model would later be edited, to the locality-motivated changing of specific parameters within models (Zhu et al., 2020; Meng et al., 2022). Recent work (Mitchell et al., 2022) draws attention to the danger of model performance degradation accumulating over the course of many successive edits, and seeks mitigation through improved methods. Hase et al. (2023b) extend the consistency requirement to hold over entailed and equivalent facts as well as paraphrases, suggesting model editing as a way of reconciling cases where certain facts are produced correctly but those entailed from them are not.", + "bbox": [ + 115, + 262, + 485, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Critique", + "text_level": 1, + "bbox": [ + 115, + 517, + 220, + 533 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present the case against model editing as a practice, regardless of the performance obtained by any single method. We begin with an analysis of the premise underlying the model editing research objective: the hypothesis that LLMs can be used as fact repositories. We then focus on reasons for which editing facts cannot be designed a-priori as a means to maintaining fact-providing LLMs, and continue with practical considerations for why even this ill-posed goal is probably unattainable.", + "bbox": [ + 115, + 545, + 485, + 721 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 LLMs as Knowledge Repositories?", + "text_level": 1, + "bbox": [ + 115, + 734, + 430, + 751 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The perception that LLMs can behave as knowledge repositories was first formulated and experimentally supported with the introduction of the LAMA benchmark (Petroni et al., 2019), where pretrained language models were queried in a zero-shot setting against 51K knowledge triplets extracted from knowledge banks and reformulated as fill-in-the-blank statements. 2019-level models (BERT-XL) got the top answer correctly about $26.5\\%$ of the time. The limitations of the LAMA", + "bbox": [ + 115, + 757, + 485, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "study were later studied and this result was shown to not be robust against multi-token spans (as opposed to a single token answer; Kalo and Fichtel, 2022), in an open vocabulary setting (as opposed to a closed vocabulary of 21K tokens; Roberts et al., 2020), and when queries vary (Kassner and Schütze, 2020); indicating that the LAMA experiment relied on heuristics to predict answers. Since then, more realistic benchmarks have been introduced, as well as more robust querying techniques (Jiang et al., 2020), to address some of these limitations. As the scale of LLMs increased, recent work has also scaled up the size of benchmarks, shifting the focus to facts concerning rare entities, where even recent LLMs struggle (Kandpal et al., 2023; Sun et al., 2023). These experiments indicate that an LM's ability to answer a question depends on the number of times information relevant to that question appears in the pretraining data. As a result, LLMs cannot be safely used to answer queries about the long-tail of facts that have been rarely mentioned in pretraining (Mallen et al., 2023).", + "bbox": [ + 512, + 84, + 880, + 437 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond the capability to reliably answer queries, LLMs should also fulfill other requirements in order to be considered as fact repositories (AlKhamissi et al., 2022): the ability to edit knowledge (add, delete, update facts), logical consistency (answers to different but related facts must agree), reasoning (the ability to infer additional answers on the basis of logical rules), and explainability and interpretability (supporting an answer through a convincing chain of arguments). Experimental results assessing these dimensions indicate that current LLMs fail on all of them. He et al. (2023) demonstrate that LLMs underperform on computing ontology subsumption inference when compared with other trained approaches such as NLI models and symbolic systems such as OWL Reasoners (Glimm et al., 2014).", + "bbox": [ + 512, + 439, + 880, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The evidence supporting the premise that LLMs can be used as fact repositories is, thus, currently weak. Beyond this fragile foundation, we focus below on specific arguments regarding the capability to edit models as if they were fact repositories.", + "bbox": [ + 512, + 715, + 880, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2 Systemic Mismatch", + "text_level": 1, + "bbox": [ + 512, + 813, + 707, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A very basic property of LLMs which contrasts with their usage as knowledge repositories is their stochastic nature. When used for augmenting creative endeavours, for data exploration, or for freeform tasks such as summarization, this property", + "bbox": [ + 512, + 839, + 880, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is desirable: the use case calls for variation or unexpectedness. In other cases, we may be content with models that supply us with a distribution of outputs, from which we can estimate the probability of individual responses and adjust our expectation for a reliable output. Since the latter is absent from many models available only through 3rd-party APIs, we are left with obtaining text generated from an unknown distribution, which we argue is insufficient for factuality-dependent applications.1 It can even be argued that getting facts wrong is a feature of vanilla LLMs rather than a bug: as their core training procedure aims to simulate plausible continuation of text, it should not surprise us that models would repeat widely-assumed falsehoods in a way that negates the idea of using them for factual purposes. If most people think, and write, that Los Angeles is the capital of California,2 an LLM is supposed to complete a relevant prompt accordingly. There is also no built-in reliability or robustness in LLMs that sample outputs from a distribution: two instances of the same prompt can easily produce contradicting facts, and indeed often do.", + "bbox": [ + 110, + 84, + 492, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the idea of editing facts in a model suggests that we always want a model to supply us with a fact as an answer to a question. However, at times, questions may be posed while pre-supposing or otherwise assuming harmful propositions such as stereotypes or conspiracy theories. Editing the \"fact\" associated with the question \"which government agency faked the moon landing?\" would not provide us with an improved model; what we may want is to remove such facts altogether, or provide the model with a means of challenging the presupposition, or avoiding giving any answer at all. At the same time, many relations that we would term \"facts\" can be argued to be vital notions without which certain kinds of basic communication is impossible.3 An LLM that cannot assert whether trees have leaves, or asserts that they never do, is in danger of becoming irrelevant for most tasks requiring any kind of interaction with the world. As philosophy and practice surrounding these questions progresses, we can hope this gap between 'must-know' and 'must-not-know' will eventually tighten to form workable bounds on LLM knowledge capacity.", + "bbox": [ + 110, + 473, + 490, + 860 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Architectural Implausibility", + "text_level": 1, + "bbox": [ + 507, + 84, + 778, + 101 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Estimates vary wildly, but there are over 100 million notable facts in the world. Even the cutoff for what constitutes a fact is unclear. Does a $3\\%$ change in a demographics statistic, or a new esoteric sports record, call for an edit? Do the daily whereabouts of world leaders constitute facts? What about those of celebrities or journalists? As events unfold daily in world politics, economics, sports, and many other walks of life, facts are added and changed in larger quantities and rates than can be plausibly \"caught up with\" through surgical model editing, akin to emptying the ocean with a spoon. If we choose to limit ourselves in which facts we find important enough to edit, we introduce bias into the system, opening the door to a host of documented harms which permeate many language technologies (Chang et al., 2019; Blodgett et al., 2020). This choice can be implicit as well as explicit, and is incredibly hard to avoid.", + "bbox": [ + 507, + 105, + 884, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a similar vein, the vastness and variability of facts is likely to lead to bias in evaluating the complement set of edits, those facts controlled for as not changing following an edit (drawdown). Even paraphrases of edited facts are not guaranteed to change alongside the selected phrasing (De Cao et al., 2021), nor are entailed facts (Hase et al., 2023b). This problem also manifests as a safety issue, since unchecked facts can turn out to be quite important for model usage, but perhaps taken for granted (or simply not explicitly covered) when designing a drawdown benchmark.", + "bbox": [ + 507, + 411, + 885, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There is evidence (Mallen et al., 2023; Jang et al., 2021) that facts above a certain \"popularity threshold\", measured in terms of views on Wikipedia articles, are harder to edit out of models compared to facts lying on the long tail of view distributions. Inherently being out of sight, the unpopular facts thus become susceptible to the double risk of being edited alongside target facts while not being deemed important enough to be checked during drawdown testing. The ultimate outcome of such procedures can be a monolithization of LLM-supplied \"knowledge\" that focuses on certain popular domains and interests while losing all usefulness in many topics contributing to the extensive diversity of the human and natural experience.", + "bbox": [ + 507, + 605, + 885, + 848 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Called \"transparent-sensitive tasks\" (Luo et al., 2023).", + "bbox": [ + 134, + 878, + 470, + 891 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2The capital of California is Sacramento.", + "bbox": [ + 136, + 891, + 386, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3We thank a reviewer for making this point.", + "bbox": [ + 136, + 904, + 403, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4107,323,022 appear in Wikidata as of October 17, 2023, conforming to the definition of notable items used in https: //www.wikidata.org/wiki/Help:Items.", + "bbox": [ + 507, + 854, + 884, + 892 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "Attributed, according to the Internet, as a \"Yiddish proverb\", but, hey, who knows?", + "bbox": [ + 507, + 892, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Empirical evidence indicates existing editing models fail to properly account for the ripple effect of a fact editing operation (Cohen et al., 2023). For example, the insertion of the fact \"Jack Depp is the son of Johnny Depp\" introduces a \"ripple effect\" in the form of additional facts that the model needs to update (e.g. \"Jack Depp is the sibling of Lily-Rose Depp\"). Results in symbolic approaches to this task have demonstrated that this knowledge updating task is of high computational complexity, even NP-hard, for example in Truth Maintenance Systems (TMS; Rutenburg, 1991). These results carry to approaches based on machine learning techniques (Knblauch et al., 2020). There is, therefore, theoretical basis to conclude that model editing will at best address the problem of consistent updating in a roughly approximate manner, and most likely fail to update rarely seen facts within the ripple effect of editing operations.", + "bbox": [ + 115, + 82, + 490, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, recent empirical findings show additional weaknesses of edit methods by extending the evaluation suites to cover aspects beyond fact editing metrics, such as specificity and robustness of the models post-editing (Onoe et al., 2023; Hoelscher-Obermaier et al., 2023; Hase et al., 2023a; Brown et al., 2023).", + "bbox": [ + 112, + 391, + 487, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Alternatives to Model Editing", + "text_level": 1, + "bbox": [ + 112, + 520, + 403, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model editing is motivated by the desire to control the text generated by LLMs to make them more compliant with desirable outcomes, specifically to control factuality: when an LLM generates text expressing an incorrect or obsolete fact, remove the fact from the LLM's \"memory\" and try again (Peng et al., 2023). Other relatively direct approaches to improve LLM factuality (and, in general, to control and revise the text generated by LLMs) have been proposed, most notably Reinforcement Learning with Human Feedback (RLHF; Bai et al., 2022; Ouyang et al., 2022; Ramamurthy et al., 2023; Rafailov et al., 2023; Carta et al., 2023), or adding a factuality objective to the training procedure (Lee et al., 2022). We now survey approaches which define a more achievable objective than that pursued by direct model editing or output manipulation. These approaches avoid the assumption that the LLM is a fact repository, and thus steer clear of the attempt to update this repository in a logically consistent manner, which is computationally hard to achieve and verify. While these alternatives avoid the more problematic aspects of model", + "bbox": [ + 115, + 549, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "editing, they still suffer from their own limitations, but we argue that they identify more promising research directions.", + "bbox": [ + 509, + 84, + 882, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Incorporating Knowledge Bases In retrieval-based models, factual knowledge is explicitly represented in a dedicated component external to the LLM. The way this external fact store is represented and combined with the LLM varies: it can be a collection of textual documents that is searched using a text retrieval component, or an RDF graph, or encoded as a set of vector embeddings, or represented as modular expert LMs trained on curated datasets. In all cases, in the retrieval-based approach, the model can explicitly cite the source which underlies a specific generation, and let the user decide its credibility.", + "bbox": [ + 507, + 143, + 882, + 351 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once external (non-parametric) knowledge is retrieved, it must be composed with the LLM generation process. Khandelwal et al. (2020) introduce a k-nearest neighbor method with interpolation; RETRO (Borgeaud et al., 2021) combines the prompt with retrieved documents through a specialized cross-attention mechanism; other LLM-IR variants include a fusion-in-decoder method (ATLAS; Izacard et al., 2022) and TRIME (Zhong et al., 2022), all retrieval-based models that maintain the capacity for few-shot in-context learning.", + "bbox": [ + 507, + 353, + 884, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent work addresses the integration of LLMs and IR by learning to combine results from search engines into the context provided to the LLM to answer a specific question (Xu et al., 2023). SAIL (Luo et al., 2023) introduces an instruction fine-tuning method to ground language generation on search results and learn to select trustworthy sources. CooK (Feng et al., 2023) approaches the task of combining multiple curated modular knowledge sources into an integrative system, where sources are modeled as independent LMs and an integrator LLM combines the information from these modules to answer a question.", + "bbox": [ + 507, + 531, + 882, + 740 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In all of these approaches, factual knowledge is stored outside of the parameters of the LLM and can be manipulated without retraining the LLM. These approaches have been shown to be scalable in the number of facts. Editing the fact store means the same as updating a database, thus simplifying the formulation of the task.", + "bbox": [ + 507, + 741, + 882, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Retrieval-based models, arguably, do not resolve all of the concerns we identify with model editing. The problem of identifying the provenance of a given generation span in these combined models", + "bbox": [ + 507, + 854, + 882, + 919 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "remains acute: the text can be determined by the internal LLM parameters, by the external stores, or by their combination, even if they are not logically consistent with one another. Facts that have been seen more often during LLM training may have more weight in this interpolation even if they are wrong or when they become obsolete. Zhu et al. (2020) claim that \"modifying knowledge in explicit memory module networks like FaE (Verga et al., 2020) is not easier and demands changes in the LLM component as well.\" This effect was also identified in the RealQA baseline test (Kasai et al., 2022): in this benchmark containing questions about time-sensitive data, experiments showed that in many cases GPT-3 (Brown et al., 2020) properly integrated data retrieved from a KB and injected into the context, but often still generated text based on outdated data from the LLM. While the clear separation of the components and the identification of the composition of external knowledge improve transparency-sensitive tasks, the objective of identifying provenance in the generated text and controlling which sources are appropriate in a given context remains an open research question. The formulation of the task of attributing generated text to identifiable sources (AIS; Rashkin et al., 2022) is a key contribution to this direction. Neeman et al. (2022) also address an aspect of this issue with a counterfactual data augmentation technique to disentangle contextual and LLM knowledge when generating text.", + "bbox": [ + 115, + 84, + 485, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Continual training focuses on incrementally training a model by introducing new tasks or new domains (e.g., Razdaibiedina et al., 2023). Model editing does not directly fall within this objective, as it concerns updating precise elements in the model while keeping tasks and domains unchanged. Yet, the identification of drawdown in model editing is similar to the risk of catastrophic forgetting identified in continual learning. Within this approach, we could situate model editing as a type of re-training or post-training. Zhu et al. (2020) note that just fine-tuning over a set of facts-to-update costs in degradation on other facts. Jang et al. (2021) identify the problem and propose to apply techniques from continual training to the task of incremental updating of LLM knowledge. Overall, while continuous training may seem to a priori avoid the risks of the model editing approach, it seems to suffer from many of the main evaluation problems.", + "bbox": [ + 115, + 596, + 485, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept Erasure The goal of concept erasure (Elazar and Goldberg, 2018; Ravfogel et al., 2020; Belrose et al., 2023) is to remove unwanted bias from embeddings generated by LLMs and subsequently from the generated text. This goal is motivated by fairness objectives: preventing protected attributes from causally affecting text generation. This motivation is somehow related to that of model editing (prevent damage from generated text) but key differences exist: (1) the method addresses general, well-identified concepts (protected attributes such as gender, race, age) as opposed to less well-defined specific facts; (2) it operates as a post-hoc transformation of embeddings, as opposed to a modification of the model itself, and as such it allows for more accountability than an ever-changing model; (3) it is verifiable with well-defined metrics over a limited pre-declared scope. While concept erasure has more limited scope than model editing, it defines an objective that can be evaluated with aggregated metrics in a robust manner. One possible avenue of future research can be to examine whether the erasure approach can be extended to address specific aspects of factuality, such as temporal validity.", + "bbox": [ + 512, + 84, + 882, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Maybe it's better to have a model that knows what it doesn't know? As identified in Kasai et al. (2022), a prerequisite to avoid generating wrong text is to identify what is not known: \"can an open-domain QA system identify unanswerable cases?\" The related issue of unanswerable questions has been addressed in a robust way (Rajpurkar et al., 2018; Sulem et al., 2021, 2022); yet the challenge in the context of LLMs and factuality is that problematic questions like those specified in §3.2 do look answerable.", + "bbox": [ + 512, + 495, + 882, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 512, + 684, + 638, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We agree that model editing is an attractive task to define, with clear benchmarks and expected outcomes. However, in current practice it contributes towards unrealistic expectations that we can solve problems like LLM hallucinations, which would lead to potential harm in unleashing use cases that are not in fact within the capabilities of LLMs alone. We advocate for the usage of retrieval-augmented methods and other structural and post-hoc methods in order to achieve the stated large-scale goals, while conceding the benefits of editing to \"safer\" applications such as model interpretability and robustness checking.", + "bbox": [ + 512, + 709, + 882, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 115, + 84, + 277, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We thank Sarah Wiegrefe for comments on earlier drafts. We thank the reviewers for their valuable feedback. Yuval Pinter was supported in part by the Israeli Ministry of Innovation, Science and Technology (Grant 2022/5451).", + "bbox": [ + 112, + 109, + 489, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 200, + 220, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is an opinion paper. We have purposely not made empirical analyses in order to support our critique, knowing that data is contestable and may vary according to its collection methodology. We aim to convince on the merit of examples and rhetorical argumentation rather than concrete evidence, mostly out of our reach for the models under consideration, which we nevertheless urge readers and practitioners to seek and use in attempts to either support or disprove our claims.", + "bbox": [ + 112, + 227, + 489, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethical Considerations", + "text_level": 1, + "bbox": [ + 114, + 399, + 315, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 84, + 608, + 98 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona T. Diab, and Marjan Ghazvininejad. 2022. A review on language models as knowledge bases. ArXiv, abs/2204.06031.", + "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback.", + "Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. Leace: Perfect linear concept erasure in closed form.", + "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454-5476, Online. Association for Computational Linguistics.", + "Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosse-lut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.", + "Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, T. W. Hennigan, Saffron Huang, Lorenzo Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and L. Sifre. 2021. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning.", + "Davis Brown, Charles Godfrey, Cody Nizinski, Jonathan Tu, and Henry Kvinge. 2023. Robustness of edited neural networks.", + "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.", + "Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer." + ], + "bbox": [ + 510, + 105, + 885, + 917 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2023. Grounding large language models in interactive environments with online reinforcement learning. arXiv preprint arXiv:2302.02662.", + "Kai-Wei Chang, Vinodkumar Prabhakaran, and Vicente Ordonez. 2019. Bias and fairness in natural language processing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): Tutorial Abstracts, Hong Kong, China. Association for Computational Linguistics.", + "Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2023. Evaluating the ripple effects of knowledge editing in language models. arXiv preprint arXiv:2307.12976.", + "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491-6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computational Linguistics.", + "Shangbin Feng, Weijia Shi, Yuyang Bai, Vidhisha Balachandran, Tianxing He, and Yulia Tsvetkov. 2023. Cook: Empowering general-purpose language models with modular and collaborative knowledge. ArXiv, abs/2305.09955.", + "Birte Glimm, Ian Horrocks, Boris Motik, Giorgos Stoilos, and Zhe Wang. 2014. Hermit: An owl 2 reasoner. Journal of Automated Reasoning, 53:245-269.", + "Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023a. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models.", + "Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2023b. Methods for measuring, updating, and visualizing factual beliefs in language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2714-2731, Dubrovnik, Croatia. Association for Computational Linguistics.", + "Yuan He, Jiaoyan Chen, Ernesto Jimenez-Ruiz, Hang Dong, and Ian Horrocks. 2023. Language model analysis for ontology subsumption inference. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 3439-3453, Toronto, Canada. Association for Computational Linguistics.", + "Jason Hoelscher-Obermaier, Julia Persson, Esben Kran, Ioannis Konstas, and Fazl Berez. 2023. Detecting edit failures in large language models: An improved specificity benchmark." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane A. Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. *ArXiv*, abs/2208.03299.", + "Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, and Minjoon Seo. 2021. Towards continual knowledge learning of language models. International Conference on Learning Representations, abs/2110.03215.", + "Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438.", + "Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In 4th Conference on Automated Knowledge Base Construction.", + "Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15696-15707. PMLR.", + "Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. 2022. Realtime qa: What's the answer right now?", + "Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811-7818, Online. Association for Computational Linguistics.", + "Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2020. Generalization through memorization: Nearest neighbor language models.", + "Jeremias Knoblauch, Hisham Husain, and Tom Diethe. 2020. Optimal continual learning has perfect memory and is np-hard. In International Conference on Machine Learning, pages 5327-5337. PMLR.", + "Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Factuality enhanced language models for open-ended text generation. ArXiv, abs/2206.04624.", + "Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. 2023. Sail: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories.", + "Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372.", + "Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. Fast model editing at scale. In ICLR.", + "Ella Neeman, Roee Aharoni, Or Honovich, Leshem Choshen, Idan Szpektor, and Omri Abend. 2022. Disentqa: Disentangling parametric and contextual knowledge with counterfactual question answering. ArXiv, abs/2211.05655.", + "Yasumasa Onoe, Michael J. Q. Zhang, Shankar Padmanabhan, Greg Durrett, and Eunsol Choi. 2023. Can lms learn new entities from descriptions? challenges in propagating injected knowledge.", + "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.", + "Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813.", + "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics.", + "Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model.", + "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for squad. In Annual Meeting of the Association for Computational Linguistics.", + "Rajkumar Ramamurthy, Prithviraj Ammanabrolu, Kianté Brantley, Jack Hessel, Rafet Sifa, Christian Bauckhage, Hannaneh Hajishirzi, and Yejin Choi. 2023. Is reinforcement learning (not) for natural language processing: Benchmarks, baselines, and" + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "building blocks for natural language policy optimization. In The Eleventh International Conference on Learning Representations.", + "Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2022. Measuring attribution in natural language generation models.", + "Roger Ratcliff. 1990. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. *Psychological review*, 97(2):285.", + "Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics.", + "Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. 2023. Progressive prompts: Continual learning for language models. ArXiv, abs/2301.12314.", + "Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418-5426, Online. Association for Computational Linguistics.", + "Vladislav Rutenberg. 1991. Complexity classification of truth maintenance systems. In STACS 91: 8th Annual Symposium on Theoretical Aspects of Computer Science Hamburg, Germany, February 14-16, 1991 Proceedings 8, pages 372-383. Springer.", + "Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitry Pyrkin, Sergei Popov, and Artem Babenko. 2020. Editable neural networks. In International Conference on Learning Representations.", + "Elior Sulem, Jamaal Hay, and Dan Roth. 2021. Do we know what we don't know? studying unanswerable questions beyond squad 2.0. In Conference on Empirical Methods in Natural Language Processing.", + "Elior Sulem, Jamaal Hay, and Dan Roth. 2022. Yes, no or idk: The challenge of unanswerable yes/no questions. In *North American Chapter of the Association for Computational Linguistics*.", + "Kai Sun, Y. Xu, Hanwen Zha, Yue Liu, and Xinhsuai Dong. 2023. Head-to-tail: How knowledgeable are large language models (llm)? a.k.a. will llms replace knowledge graphs? ArXiv, abs/2308.10168.", + "Pat Verga, Haitian Sun, Livio Baldini Soares, and William W. Cohen. 2020. Facts as experts: Adaptable and interpretable neural memory over symbolic knowledge." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-seng Chua. 2023. Search-in-the-chain: Towards the accurate, credible and traceable content generation for complex knowledge-intensive tasks. arXiv preprint arXiv:2304.14732.", + "Zexuan Zhong, Tao Lei, and Danqi Chen. 2022. Training language models with memory augmentation. In Conference on Empirical Methods in Natural Language Processing.", + "Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363." + ], + "bbox": [ + 115, + 85, + 489, + 277 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_model.json b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e23c7192ca7d23326c6b6c116e89e99c3c245c3 --- /dev/null +++ b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_model.json @@ -0,0 +1,1340 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.189, + 0.09, + 0.812, + 0.111 + ], + "angle": 0, + "content": "Emptying the Ocean with a Spoon: Should We Edit Models?" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.131, + 0.656, + 0.216 + ], + "angle": 0, + "content": "Yuval Pinter and Michael Elhadad \nDepartment of Computer Science \nBen-Gurion University of the Negev \nBe'er Sheva, Israel \n{uvp, elhadad}@cs.bgu.ac.il" + }, + { + "type": "title", + "bbox": [ + 0.264, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.278, + 0.46, + 0.448 + ], + "angle": 0, + "content": "We call into question the recently popularized method of direct model editing as a means of correcting factual errors in LLM generations. We contrast model editing with three similar but distinct approaches that pursue better defined objectives: (1) retrieval-based architectures, which decouple factual memory from inference and linguistic capabilities embodied in LLMs; (2) concept erasure methods, which aim at preventing systemic bias in generated text; and (3) attribution methods, which aim at grounding generations into identified textual sources." + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.454, + 0.46, + 0.61 + ], + "angle": 0, + "content": "We argue that direct model editing cannot be trusted as a systematic remedy for the disadvantages inherent to LLMs, and while it has proven potential in improving model explainability, it opens risks by reinforcing the notion that models can be trusted for factuality. We call for cautious promotion and application of model editing as part of the LLM deployment process, and for responsibly limiting the use cases of LLMs to those not relying on editing as a critical component." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.621, + 0.259, + 0.636 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.646, + 0.488, + 0.886 + ], + "angle": 0, + "content": "Large language models, or LLMs, have taken the NLP world by storm. After originally focusing on training them as a vehicle for transfer learning, recent advancements have cast LLMs in the role of one-stop shop, all-knowing oracles. One particular finding that has contributed to this reframed usage is the apparent factuality properties of pre-trained LLMs: somehow, mere next-word prediction training has produced models that can complete certain correct facts about the world when asked to do so. LLMs have been perceived by the public at large as a replacement for search engines, and the scant disclaimers appearing in commercial services offering LLM querying do not appear to have made a dent in this perception." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.888, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Despite these developments, LLMs of course keep making mistakes. There is, after all, an inher" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.884, + 0.59 + ], + "angle": 0, + "content": "ent mismatch between their pre-training objective and the desire for factuality. In recent years, researchers have come up with several remedies to the problem of nonfactual LLM outputs, one of which being model editing, where parameters inside an LLM are tweaked based on individual facts marked for correction. These works focus on solving problems such as ensuring stability of other factual outputs following the editing, or batch editing, or computationally efficient editing. In this opinion paper, we call into question the entire idea of model editing, citing concerns about intended use cases, conceptual scalability, potential for bias, safety, and overall accountability. We advocate using models that have explicit knowledge modules for tasks requiring factual knowledge, and other existing techniques to mitigate the need for model editing in the first place. Having said that, we acknowledge the usefulness of direct editing for certain use cases such as interpretability probes, and recommend limiting editing to such scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.668, + 0.62 + ], + "angle": 0, + "content": "2 Model Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.884, + 0.918 + ], + "angle": 0, + "content": "Sinitsin et al. (2020) first introduce the notion of updating large ML models to account for local performance expectations motivated externally. They cite cases where mistakes are critical, like object detection in self-driving cars. Later work suggests model editing can aid in protecting privacy and eliminating bias (Zhu et al., 2020), and as a measure of \"catching up\" with time-sensitive facts such as \"PM of the UK\" changing over time (Bommasani et al., 2021; Mitchell et al., 2022), or as a means of understanding the mechanics of black box models (Meng et al., 2022). Sinitsin et al. (2020) specify several desired properties of editing methods: reliability (the target facts are updated as intended), locality (other changes don't happen; a measure for this property is termed drawdown), and efficiency (successful edits are computationally light); later work (De Cao et al., 2021) added generality (the" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.486, + 0.261 + ], + "angle": 0, + "content": "ability to modify models not originally trained for knowledge retention), consistency (robustness to paraphrasing, in the specific use case of text models), and an unspecified property we'll call frugality (only minimal components of the model are changed during editing). A well-studied limitation, which most of these works focus on mitigating, and related to locality, is catastrophic forgetting (Ratcliff, 1990), i.e., ensuring that an edited model does not lose performance on tasks for which it was explicitly trained and performs well on." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.263, + 0.486, + 0.502 + ], + "angle": 0, + "content": "Methods for model editing have evolved from editable training (Sinitsin et al., 2020), a procedure requiring an a-priori decision that a model would later be edited, to the locality-motivated changing of specific parameters within models (Zhu et al., 2020; Meng et al., 2022). Recent work (Mitchell et al., 2022) draws attention to the danger of model performance degradation accumulating over the course of many successive edits, and seeks mitigation through improved methods. Hase et al. (2023b) extend the consistency requirement to hold over entailed and equivalent facts as well as paraphrases, suggesting model editing as a way of reconciling cases where certain facts are produced correctly but those entailed from them are not." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.518, + 0.221, + 0.535 + ], + "angle": 0, + "content": "3 Critique" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.546, + 0.486, + 0.722 + ], + "angle": 0, + "content": "In this section, we present the case against model editing as a practice, regardless of the performance obtained by any single method. We begin with an analysis of the premise underlying the model editing research objective: the hypothesis that LLMs can be used as fact repositories. We then focus on reasons for which editing facts cannot be designed a-priori as a means to maintaining fact-providing LLMs, and continue with practical considerations for why even this ill-posed goal is probably unattainable." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.736, + 0.431, + 0.752 + ], + "angle": 0, + "content": "3.1 LLMs as Knowledge Repositories?" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.758, + 0.486, + 0.918 + ], + "angle": 0, + "content": "The perception that LLMs can behave as knowledge repositories was first formulated and experimentally supported with the introduction of the LAMA benchmark (Petroni et al., 2019), where pretrained language models were queried in a zero-shot setting against 51K knowledge triplets extracted from knowledge banks and reformulated as fill-in-the-blank statements. 2019-level models (BERT-XL) got the top answer correctly about \\(26.5\\%\\) of the time. The limitations of the LAMA" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.085, + 0.882, + 0.438 + ], + "angle": 0, + "content": "study were later studied and this result was shown to not be robust against multi-token spans (as opposed to a single token answer; Kalo and Fichtel, 2022), in an open vocabulary setting (as opposed to a closed vocabulary of 21K tokens; Roberts et al., 2020), and when queries vary (Kassner and Schütze, 2020); indicating that the LAMA experiment relied on heuristics to predict answers. Since then, more realistic benchmarks have been introduced, as well as more robust querying techniques (Jiang et al., 2020), to address some of these limitations. As the scale of LLMs increased, recent work has also scaled up the size of benchmarks, shifting the focus to facts concerning rare entities, where even recent LLMs struggle (Kandpal et al., 2023; Sun et al., 2023). These experiments indicate that an LM's ability to answer a question depends on the number of times information relevant to that question appears in the pretraining data. As a result, LLMs cannot be safely used to answer queries about the long-tail of facts that have been rarely mentioned in pretraining (Mallen et al., 2023)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.882, + 0.713 + ], + "angle": 0, + "content": "Beyond the capability to reliably answer queries, LLMs should also fulfill other requirements in order to be considered as fact repositories (AlKhamissi et al., 2022): the ability to edit knowledge (add, delete, update facts), logical consistency (answers to different but related facts must agree), reasoning (the ability to infer additional answers on the basis of logical rules), and explainability and interpretability (supporting an answer through a convincing chain of arguments). Experimental results assessing these dimensions indicate that current LLMs fail on all of them. He et al. (2023) demonstrate that LLMs underperform on computing ontology subsumption inference when compared with other trained approaches such as NLI models and symbolic systems such as OWL Reasoners (Glimm et al., 2014)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.882, + 0.796 + ], + "angle": 0, + "content": "The evidence supporting the premise that LLMs can be used as fact repositories is, thus, currently weak. Beyond this fragile foundation, we focus below on specific arguments regarding the capability to edit models as if they were fact repositories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.814, + 0.708, + 0.83 + ], + "angle": 0, + "content": "3.2 Systemic Mismatch" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.882, + 0.918 + ], + "angle": 0, + "content": "A very basic property of LLMs which contrasts with their usage as knowledge repositories is their stochastic nature. When used for augmenting creative endeavours, for data exploration, or for freeform tasks such as summarization, this property" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.085, + 0.493, + 0.472 + ], + "angle": 0, + "content": "is desirable: the use case calls for variation or unexpectedness. In other cases, we may be content with models that supply us with a distribution of outputs, from which we can estimate the probability of individual responses and adjust our expectation for a reliable output. Since the latter is absent from many models available only through 3rd-party APIs, we are left with obtaining text generated from an unknown distribution, which we argue is insufficient for factuality-dependent applications.1 It can even be argued that getting facts wrong is a feature of vanilla LLMs rather than a bug: as their core training procedure aims to simulate plausible continuation of text, it should not surprise us that models would repeat widely-assumed falsehoods in a way that negates the idea of using them for factual purposes. If most people think, and write, that Los Angeles is the capital of California,2 an LLM is supposed to complete a relevant prompt accordingly. There is also no built-in reliability or robustness in LLMs that sample outputs from a distribution: two instances of the same prompt can easily produce contradicting facts, and indeed often do." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.474, + 0.492, + 0.862 + ], + "angle": 0, + "content": "Additionally, the idea of editing facts in a model suggests that we always want a model to supply us with a fact as an answer to a question. However, at times, questions may be posed while pre-supposing or otherwise assuming harmful propositions such as stereotypes or conspiracy theories. Editing the \"fact\" associated with the question \"which government agency faked the moon landing?\" would not provide us with an improved model; what we may want is to remove such facts altogether, or provide the model with a means of challenging the presupposition, or avoiding giving any answer at all. At the same time, many relations that we would term \"facts\" can be argued to be vital notions without which certain kinds of basic communication is impossible.3 An LLM that cannot assert whether trees have leaves, or asserts that they never do, is in danger of becoming irrelevant for most tasks requiring any kind of interaction with the world. As philosophy and practice surrounding these questions progresses, we can hope this gap between 'must-know' and 'must-not-know' will eventually tighten to form workable bounds on LLM knowledge capacity." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.779, + 0.102 + ], + "angle": 0, + "content": "3.3 Architectural Implausibility" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.106, + 0.885, + 0.412 + ], + "angle": 0, + "content": "Estimates vary wildly, but there are over 100 million notable facts in the world. Even the cutoff for what constitutes a fact is unclear. Does a \\(3\\%\\) change in a demographics statistic, or a new esoteric sports record, call for an edit? Do the daily whereabouts of world leaders constitute facts? What about those of celebrities or journalists? As events unfold daily in world politics, economics, sports, and many other walks of life, facts are added and changed in larger quantities and rates than can be plausibly \"caught up with\" through surgical model editing, akin to emptying the ocean with a spoon. If we choose to limit ourselves in which facts we find important enough to edit, we introduce bias into the system, opening the door to a host of documented harms which permeate many language technologies (Chang et al., 2019; Blodgett et al., 2020). This choice can be implicit as well as explicit, and is incredibly hard to avoid." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.412, + 0.887, + 0.605 + ], + "angle": 0, + "content": "In a similar vein, the vastness and variability of facts is likely to lead to bias in evaluating the complement set of edits, those facts controlled for as not changing following an edit (drawdown). Even paraphrases of edited facts are not guaranteed to change alongside the selected phrasing (De Cao et al., 2021), nor are entailed facts (Hase et al., 2023b). This problem also manifests as a safety issue, since unchecked facts can turn out to be quite important for model usage, but perhaps taken for granted (or simply not explicitly covered) when designing a drawdown benchmark." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.606, + 0.887, + 0.849 + ], + "angle": 0, + "content": "There is evidence (Mallen et al., 2023; Jang et al., 2021) that facts above a certain \"popularity threshold\", measured in terms of views on Wikipedia articles, are harder to edit out of models compared to facts lying on the long tail of view distributions. Inherently being out of sight, the unpopular facts thus become susceptible to the double risk of being edited alongside target facts while not being deemed important enough to be checked during drawdown testing. The ultimate outcome of such procedures can be a monolithization of LLM-supplied \"knowledge\" that focuses on certain popular domains and interests while losing all usefulness in many topics contributing to the extensive diversity of the human and natural experience." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.879, + 0.472, + 0.892 + ], + "angle": 0, + "content": "1Called \"transparent-sensitive tasks\" (Luo et al., 2023)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.892, + 0.388, + 0.905 + ], + "angle": 0, + "content": "2The capital of California is Sacramento." + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.404, + 0.919 + ], + "angle": 0, + "content": "3We thank a reviewer for making this point." + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.879, + 0.472, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.856, + 0.885, + 0.893 + ], + "angle": 0, + "content": "4107,323,022 appear in Wikidata as of October 17, 2023, conforming to the definition of notable items used in https: //www.wikidata.org/wiki/Help:Items." + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.893, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Attributed, according to the Internet, as a \"Yiddish proverb\", but, hey, who knows?" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.856, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.083, + 0.492, + 0.388 + ], + "angle": 0, + "content": "Empirical evidence indicates existing editing models fail to properly account for the ripple effect of a fact editing operation (Cohen et al., 2023). For example, the insertion of the fact \"Jack Depp is the son of Johnny Depp\" introduces a \"ripple effect\" in the form of additional facts that the model needs to update (e.g. \"Jack Depp is the sibling of Lily-Rose Depp\"). Results in symbolic approaches to this task have demonstrated that this knowledge updating task is of high computational complexity, even NP-hard, for example in Truth Maintenance Systems (TMS; Rutenburg, 1991). These results carry to approaches based on machine learning techniques (Knblauch et al., 2020). There is, therefore, theoretical basis to conclude that model editing will at best address the problem of consistent updating in a roughly approximate manner, and most likely fail to update rarely seen facts within the ripple effect of editing operations." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.392, + 0.489, + 0.504 + ], + "angle": 0, + "content": "Finally, recent empirical findings show additional weaknesses of edit methods by extending the evaluation suites to cover aspects beyond fact editing metrics, such as specificity and robustness of the models post-editing (Onoe et al., 2023; Hoelscher-Obermaier et al., 2023; Hase et al., 2023a; Brown et al., 2023)." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.521, + 0.405, + 0.538 + ], + "angle": 0, + "content": "4 Alternatives to Model Editing" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.55, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Model editing is motivated by the desire to control the text generated by LLMs to make them more compliant with desirable outcomes, specifically to control factuality: when an LLM generates text expressing an incorrect or obsolete fact, remove the fact from the LLM's \"memory\" and try again (Peng et al., 2023). Other relatively direct approaches to improve LLM factuality (and, in general, to control and revise the text generated by LLMs) have been proposed, most notably Reinforcement Learning with Human Feedback (RLHF; Bai et al., 2022; Ouyang et al., 2022; Ramamurthy et al., 2023; Rafailov et al., 2023; Carta et al., 2023), or adding a factuality objective to the training procedure (Lee et al., 2022). We now survey approaches which define a more achievable objective than that pursued by direct model editing or output manipulation. These approaches avoid the assumption that the LLM is a fact repository, and thus steer clear of the attempt to update this repository in a logically consistent manner, which is computationally hard to achieve and verify. While these alternatives avoid the more problematic aspects of model" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.884, + 0.133 + ], + "angle": 0, + "content": "editing, they still suffer from their own limitations, but we argue that they identify more promising research directions." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.144, + 0.884, + 0.352 + ], + "angle": 0, + "content": "Incorporating Knowledge Bases In retrieval-based models, factual knowledge is explicitly represented in a dedicated component external to the LLM. The way this external fact store is represented and combined with the LLM varies: it can be a collection of textual documents that is searched using a text retrieval component, or an RDF graph, or encoded as a set of vector embeddings, or represented as modular expert LMs trained on curated datasets. In all cases, in the retrieval-based approach, the model can explicitly cite the source which underlies a specific generation, and let the user decide its credibility." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.354, + 0.885, + 0.531 + ], + "angle": 0, + "content": "Once external (non-parametric) knowledge is retrieved, it must be composed with the LLM generation process. Khandelwal et al. (2020) introduce a k-nearest neighbor method with interpolation; RETRO (Borgeaud et al., 2021) combines the prompt with retrieved documents through a specialized cross-attention mechanism; other LLM-IR variants include a fusion-in-decoder method (ATLAS; Izacard et al., 2022) and TRIME (Zhong et al., 2022), all retrieval-based models that maintain the capacity for few-shot in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.532, + 0.884, + 0.741 + ], + "angle": 0, + "content": "Recent work addresses the integration of LLMs and IR by learning to combine results from search engines into the context provided to the LLM to answer a specific question (Xu et al., 2023). SAIL (Luo et al., 2023) introduces an instruction fine-tuning method to ground language generation on search results and learn to select trustworthy sources. CooK (Feng et al., 2023) approaches the task of combining multiple curated modular knowledge sources into an integrative system, where sources are modeled as independent LMs and an integrator LLM combines the information from these modules to answer a question." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.742, + 0.884, + 0.853 + ], + "angle": 0, + "content": "In all of these approaches, factual knowledge is stored outside of the parameters of the LLM and can be manipulated without retraining the LLM. These approaches have been shown to be scalable in the number of facts. Editing the fact store means the same as updating a database, thus simplifying the formulation of the task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.884, + 0.92 + ], + "angle": 0, + "content": "Retrieval-based models, arguably, do not resolve all of the concerns we identify with model editing. The problem of identifying the provenance of a given generation span in these combined models" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.486, + 0.584 + ], + "angle": 0, + "content": "remains acute: the text can be determined by the internal LLM parameters, by the external stores, or by their combination, even if they are not logically consistent with one another. Facts that have been seen more often during LLM training may have more weight in this interpolation even if they are wrong or when they become obsolete. Zhu et al. (2020) claim that \"modifying knowledge in explicit memory module networks like FaE (Verga et al., 2020) is not easier and demands changes in the LLM component as well.\" This effect was also identified in the RealQA baseline test (Kasai et al., 2022): in this benchmark containing questions about time-sensitive data, experiments showed that in many cases GPT-3 (Brown et al., 2020) properly integrated data retrieved from a KB and injected into the context, but often still generated text based on outdated data from the LLM. While the clear separation of the components and the identification of the composition of external knowledge improve transparency-sensitive tasks, the objective of identifying provenance in the generated text and controlling which sources are appropriate in a given context remains an open research question. The formulation of the task of attributing generated text to identifiable sources (AIS; Rashkin et al., 2022) is a key contribution to this direction. Neeman et al. (2022) also address an aspect of this issue with a counterfactual data augmentation technique to disentangle contextual and LLM knowledge when generating text." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.598, + 0.486, + 0.918 + ], + "angle": 0, + "content": "Continual training focuses on incrementally training a model by introducing new tasks or new domains (e.g., Razdaibiedina et al., 2023). Model editing does not directly fall within this objective, as it concerns updating precise elements in the model while keeping tasks and domains unchanged. Yet, the identification of drawdown in model editing is similar to the risk of catastrophic forgetting identified in continual learning. Within this approach, we could situate model editing as a type of re-training or post-training. Zhu et al. (2020) note that just fine-tuning over a set of facts-to-update costs in degradation on other facts. Jang et al. (2021) identify the problem and propose to apply techniques from continual training to the task of incremental updating of LLM knowledge. Overall, while continuous training may seem to a priori avoid the risks of the model editing approach, it seems to suffer from many of the main evaluation problems." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.085, + 0.884, + 0.487 + ], + "angle": 0, + "content": "Concept Erasure The goal of concept erasure (Elazar and Goldberg, 2018; Ravfogel et al., 2020; Belrose et al., 2023) is to remove unwanted bias from embeddings generated by LLMs and subsequently from the generated text. This goal is motivated by fairness objectives: preventing protected attributes from causally affecting text generation. This motivation is somehow related to that of model editing (prevent damage from generated text) but key differences exist: (1) the method addresses general, well-identified concepts (protected attributes such as gender, race, age) as opposed to less well-defined specific facts; (2) it operates as a post-hoc transformation of embeddings, as opposed to a modification of the model itself, and as such it allows for more accountability than an ever-changing model; (3) it is verifiable with well-defined metrics over a limited pre-declared scope. While concept erasure has more limited scope than model editing, it defines an objective that can be evaluated with aggregated metrics in a robust manner. One possible avenue of future research can be to examine whether the erasure approach can be extended to address specific aspects of factuality, such as temporal validity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.497, + 0.884, + 0.671 + ], + "angle": 0, + "content": "Maybe it's better to have a model that knows what it doesn't know? As identified in Kasai et al. (2022), a prerequisite to avoid generating wrong text is to identify what is not known: \"can an open-domain QA system identify unanswerable cases?\" The related issue of unanswerable questions has been addressed in a robust way (Rajpurkar et al., 2018; Sulem et al., 2021, 2022); yet the challenge in the context of LLMs and factuality is that problematic questions like those specified in §3.2 do look answerable." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.686, + 0.64, + 0.7 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.71, + 0.884, + 0.918 + ], + "angle": 0, + "content": "We agree that model editing is an attractive task to define, with clear benchmarks and expected outcomes. However, in current practice it contributes towards unrealistic expectations that we can solve problems like LLM hallucinations, which would lead to potential harm in unleashing use cases that are not in fact within the capabilities of LLMs alone. We advocate for the usage of retrieval-augmented methods and other structural and post-hoc methods in order to achieve the stated large-scale goals, while conceding the benefits of editing to \"safer\" applications such as model interpretability and robustness checking." + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.278, + 0.101 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.11, + 0.49, + 0.19 + ], + "angle": 0, + "content": "We thank Sarah Wiegrefe for comments on earlier drafts. We thank the reviewers for their valuable feedback. Yuval Pinter was supported in part by the Israeli Ministry of Innovation, Science and Technology (Grant 2022/5451)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.202, + 0.221, + 0.216 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.228, + 0.49, + 0.388 + ], + "angle": 0, + "content": "This is an opinion paper. We have purposely not made empirical analyses in order to support our critique, knowing that data is contestable and may vary according to its collection methodology. We aim to convince on the merit of examples and rhetorical argumentation rather than concrete evidence, mostly out of our reach for the models under consideration, which we nevertheless urge readers and practitioners to seek and use in attempts to either support or disprove our claims." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.4, + 0.317, + 0.415 + ], + "angle": 0, + "content": "Ethical Considerations" + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.085, + 0.609, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.107, + 0.886, + 0.16 + ], + "angle": 0, + "content": "Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona T. Diab, and Marjan Ghazvininejad. 2022. A review on language models as knowledge bases. ArXiv, abs/2204.06031." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.169, + 0.886, + 0.327 + ], + "angle": 0, + "content": "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.336, + 0.884, + 0.388 + ], + "angle": 0, + "content": "Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. Leace: Perfect linear concept erasure in closed form." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.398, + 0.886, + 0.49 + ], + "angle": 0, + "content": "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454-5476, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.5, + 0.886, + 0.577 + ], + "angle": 0, + "content": "Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosse-lut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.588, + 0.886, + 0.745 + ], + "angle": 0, + "content": "Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, T. W. Hennigan, Saffron Huang, Lorenzo Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and L. Sifre. 2021. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.755, + 0.884, + 0.793 + ], + "angle": 0, + "content": "Davis Brown, Charles Godfrey, Cody Nizinski, Jonathan Tu, and Henry Kvinge. 2023. Robustness of edited neural networks." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.804, + 0.884, + 0.882 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.892, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.107, + 0.886, + 0.919 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.489, + 0.126 + ], + "angle": 0, + "content": "2023. Grounding large language models in interactive environments with online reinforcement learning. arXiv preprint arXiv:2302.02662." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.134, + 0.487, + 0.239 + ], + "angle": 0, + "content": "Kai-Wei Chang, Vinodkumar Prabhakaran, and Vicente Ordonez. 2019. Bias and fairness in natural language processing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): Tutorial Abstracts, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.248, + 0.487, + 0.301 + ], + "angle": 0, + "content": "Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2023. Evaluating the ripple effects of knowledge editing in language models. arXiv preprint arXiv:2307.12976." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.309, + 0.487, + 0.387 + ], + "angle": 0, + "content": "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491-6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.396, + 0.487, + 0.474 + ], + "angle": 0, + "content": "Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.483, + 0.487, + 0.548 + ], + "angle": 0, + "content": "Shangbin Feng, Weijia Shi, Yuyang Bai, Vidhisha Balachandran, Tianxing He, and Yulia Tsvetkov. 2023. Cook: Empowering general-purpose language models with modular and collaborative knowledge. ArXiv, abs/2305.09955." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.557, + 0.487, + 0.597 + ], + "angle": 0, + "content": "Birte Glimm, Ian Horrocks, Boris Motik, Giorgos Stoilos, and Zhe Wang. 2014. Hermit: An owl 2 reasoner. Journal of Automated Reasoning, 53:245-269." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.606, + 0.487, + 0.658 + ], + "angle": 0, + "content": "Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023a. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.666, + 0.487, + 0.771 + ], + "angle": 0, + "content": "Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2023b. Methods for measuring, updating, and visualizing factual beliefs in language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2714-2731, Dubrovnik, Croatia. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.779, + 0.487, + 0.857 + ], + "angle": 0, + "content": "Yuan He, Jiaoyan Chen, Ernesto Jimenez-Ruiz, Hang Dong, and Ian Horrocks. 2023. Language model analysis for ontology subsumption inference. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 3439-3453, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Jason Hoelscher-Obermaier, Julia Persson, Esben Kran, Ioannis Konstas, and Fazl Berez. 2023. Detecting edit failures in large language models: An improved specificity benchmark." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.489, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.152 + ], + "angle": 0, + "content": "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane A. Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. *ArXiv*, abs/2208.03299." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.166, + 0.882, + 0.243 + ], + "angle": 0, + "content": "Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, and Minjoon Seo. 2021. Towards continual knowledge learning of language models. International Conference on Learning Representations, abs/2110.03215." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.258, + 0.882, + 0.31 + ], + "angle": 0, + "content": "Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.324, + 0.882, + 0.376 + ], + "angle": 0, + "content": "Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In 4th Conference on Automated Knowledge Base Construction." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.39, + 0.882, + 0.481 + ], + "angle": 0, + "content": "Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15696-15707. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.496, + 0.882, + 0.549 + ], + "angle": 0, + "content": "Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. 2022. Realtime qa: What's the answer right now?" + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.562, + 0.882, + 0.641 + ], + "angle": 0, + "content": "Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811-7818, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.655, + 0.882, + 0.706 + ], + "angle": 0, + "content": "Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2020. Generalization through memorization: Nearest neighbor language models." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.72, + 0.882, + 0.773 + ], + "angle": 0, + "content": "Jeremias Knoblauch, Hisham Husain, and Tom Diethe. 2020. Optimal continual learning has perfect memory and is np-hard. In International Conference on Machine Learning, pages 5327-5337. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.787, + 0.882, + 0.839 + ], + "angle": 0, + "content": "Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Factuality enhanced language models for open-ended text generation. ArXiv, abs/2206.04624." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.853, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. 2023. Sail: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.163, + 0.489, + 0.215 + ], + "angle": 0, + "content": "Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.226, + 0.487, + 0.266 + ], + "angle": 0, + "content": "Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. Fast model editing at scale. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.277, + 0.489, + 0.342 + ], + "angle": 0, + "content": "Ella Neeman, Roee Aharoni, Or Honovich, Leshem Choshen, Idan Szpektor, and Omri Abend. 2022. Disentqa: Disentangling parametric and contextual knowledge with counterfactual question answering. ArXiv, abs/2211.05655." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.354, + 0.488, + 0.407 + ], + "angle": 0, + "content": "Yasumasa Onoe, Michael J. Q. Zhang, Shankar Padmanabhan, Greg Durrett, and Eunsol Choi. 2023. Can lms learn new entities from descriptions? challenges in propagating injected knowledge." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.417, + 0.489, + 0.496 + ], + "angle": 0, + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.507, + 0.488, + 0.586 + ], + "angle": 0, + "content": "Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.597, + 0.489, + 0.715 + ], + "angle": 0, + "content": "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.726, + 0.488, + 0.778 + ], + "angle": 0, + "content": "Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.789, + 0.489, + 0.842 + ], + "angle": 0, + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for squad. In Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.853, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Rajkumar Ramamurthy, Prithviraj Ammanabrolu, Kianté Brantley, Jack Hessel, Rafet Sifa, Christian Bauckhage, Hannaneh Hajishirzi, and Yejin Choi. 2023. Is reinforcement learning (not) for natural language processing: Benchmarks, baselines, and" + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.883, + 0.126 + ], + "angle": 0, + "content": "building blocks for natural language policy optimization. In The Eleventh International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.137, + 0.883, + 0.203 + ], + "angle": 0, + "content": "Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2022. Measuring attribution in natural language generation models." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.214, + 0.883, + 0.266 + ], + "angle": 0, + "content": "Roger Ratcliff. 1990. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. *Psychological review*, 97(2):285." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.277, + 0.883, + 0.369 + ], + "angle": 0, + "content": "Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.38, + 0.883, + 0.433 + ], + "angle": 0, + "content": "Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. 2023. Progressive prompts: Continual learning for language models. ArXiv, abs/2301.12314." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.444, + 0.883, + 0.523 + ], + "angle": 0, + "content": "Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418-5426, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.534, + 0.883, + 0.6 + ], + "angle": 0, + "content": "Vladislav Rutenberg. 1991. Complexity classification of truth maintenance systems. In STACS 91: 8th Annual Symposium on Theoretical Aspects of Computer Science Hamburg, Germany, February 14-16, 1991 Proceedings 8, pages 372-383. Springer." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.611, + 0.883, + 0.663 + ], + "angle": 0, + "content": "Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitry Pyrkin, Sergei Popov, and Artem Babenko. 2020. Editable neural networks. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.674, + 0.883, + 0.728 + ], + "angle": 0, + "content": "Elior Sulem, Jamaal Hay, and Dan Roth. 2021. Do we know what we don't know? studying unanswerable questions beyond squad 2.0. In Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.738, + 0.883, + 0.791 + ], + "angle": 0, + "content": "Elior Sulem, Jamaal Hay, and Dan Roth. 2022. Yes, no or idk: The challenge of unanswerable yes/no questions. In *North American Chapter of the Association for Computational Linguistics*." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.802, + 0.883, + 0.855 + ], + "angle": 0, + "content": "Kai Sun, Y. Xu, Hanwen Zha, Yue Liu, and Xinhsuai Dong. 2023. Head-to-tail: How knowledgeable are large language models (llm)? a.k.a. will llms replace knowledge graphs? ArXiv, abs/2308.10168." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.866, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Pat Verga, Haitian Sun, Livio Baldini Soares, and William W. Cohen. 2020. Facts as experts: Adaptable and interpretable neural memory over symbolic knowledge." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-seng Chua. 2023. Search-in-the-chain: Towards the accurate, credible and traceable content generation for complex knowledge-intensive tasks. arXiv preprint arXiv:2304.14732." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.162, + 0.49, + 0.215 + ], + "angle": 0, + "content": "Zexuan Zhong, Tao Lei, and Danqi Chen. 2022. Training language models with memory augmentation. In Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.224, + 0.49, + 0.278 + ], + "angle": 0, + "content": "Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.278 + ], + "angle": 0, + "content": null + } + ] +] \ No newline at end of file diff --git a/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_origin.pdf b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce74c88959b7da019660b71504968eeeda0709e4 --- /dev/null +++ b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/04529405-4e73-4a21-827a-c746e1975cce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca807355ccf1c65381119b7c69a002c9ae9417736168d93e17bb5dda82b09cdf +size 159722 diff --git a/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/full.md b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de20f269607d09eca3652f259b86b4f9e8acaffd --- /dev/null +++ b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/full.md @@ -0,0 +1,166 @@ +# Emptying the Ocean with a Spoon: Should We Edit Models? + +Yuval Pinter and Michael Elhadad +Department of Computer Science +Ben-Gurion University of the Negev +Be'er Sheva, Israel +{uvp, elhadad}@cs.bgu.ac.il + +# Abstract + +We call into question the recently popularized method of direct model editing as a means of correcting factual errors in LLM generations. We contrast model editing with three similar but distinct approaches that pursue better defined objectives: (1) retrieval-based architectures, which decouple factual memory from inference and linguistic capabilities embodied in LLMs; (2) concept erasure methods, which aim at preventing systemic bias in generated text; and (3) attribution methods, which aim at grounding generations into identified textual sources. + +We argue that direct model editing cannot be trusted as a systematic remedy for the disadvantages inherent to LLMs, and while it has proven potential in improving model explainability, it opens risks by reinforcing the notion that models can be trusted for factuality. We call for cautious promotion and application of model editing as part of the LLM deployment process, and for responsibly limiting the use cases of LLMs to those not relying on editing as a critical component. + +# 1 Introduction + +Large language models, or LLMs, have taken the NLP world by storm. After originally focusing on training them as a vehicle for transfer learning, recent advancements have cast LLMs in the role of one-stop shop, all-knowing oracles. One particular finding that has contributed to this reframed usage is the apparent factuality properties of pre-trained LLMs: somehow, mere next-word prediction training has produced models that can complete certain correct facts about the world when asked to do so. LLMs have been perceived by the public at large as a replacement for search engines, and the scant disclaimers appearing in commercial services offering LLM querying do not appear to have made a dent in this perception. + +Despite these developments, LLMs of course keep making mistakes. There is, after all, an inher + +ent mismatch between their pre-training objective and the desire for factuality. In recent years, researchers have come up with several remedies to the problem of nonfactual LLM outputs, one of which being model editing, where parameters inside an LLM are tweaked based on individual facts marked for correction. These works focus on solving problems such as ensuring stability of other factual outputs following the editing, or batch editing, or computationally efficient editing. In this opinion paper, we call into question the entire idea of model editing, citing concerns about intended use cases, conceptual scalability, potential for bias, safety, and overall accountability. We advocate using models that have explicit knowledge modules for tasks requiring factual knowledge, and other existing techniques to mitigate the need for model editing in the first place. Having said that, we acknowledge the usefulness of direct editing for certain use cases such as interpretability probes, and recommend limiting editing to such scenarios. + +# 2 Model Editing + +Sinitsin et al. (2020) first introduce the notion of updating large ML models to account for local performance expectations motivated externally. They cite cases where mistakes are critical, like object detection in self-driving cars. Later work suggests model editing can aid in protecting privacy and eliminating bias (Zhu et al., 2020), and as a measure of "catching up" with time-sensitive facts such as "PM of the UK" changing over time (Bommasani et al., 2021; Mitchell et al., 2022), or as a means of understanding the mechanics of black box models (Meng et al., 2022). Sinitsin et al. (2020) specify several desired properties of editing methods: reliability (the target facts are updated as intended), locality (other changes don't happen; a measure for this property is termed drawdown), and efficiency (successful edits are computationally light); later work (De Cao et al., 2021) added generality (the + +ability to modify models not originally trained for knowledge retention), consistency (robustness to paraphrasing, in the specific use case of text models), and an unspecified property we'll call frugality (only minimal components of the model are changed during editing). A well-studied limitation, which most of these works focus on mitigating, and related to locality, is catastrophic forgetting (Ratcliff, 1990), i.e., ensuring that an edited model does not lose performance on tasks for which it was explicitly trained and performs well on. + +Methods for model editing have evolved from editable training (Sinitsin et al., 2020), a procedure requiring an a-priori decision that a model would later be edited, to the locality-motivated changing of specific parameters within models (Zhu et al., 2020; Meng et al., 2022). Recent work (Mitchell et al., 2022) draws attention to the danger of model performance degradation accumulating over the course of many successive edits, and seeks mitigation through improved methods. Hase et al. (2023b) extend the consistency requirement to hold over entailed and equivalent facts as well as paraphrases, suggesting model editing as a way of reconciling cases where certain facts are produced correctly but those entailed from them are not. + +# 3 Critique + +In this section, we present the case against model editing as a practice, regardless of the performance obtained by any single method. We begin with an analysis of the premise underlying the model editing research objective: the hypothesis that LLMs can be used as fact repositories. We then focus on reasons for which editing facts cannot be designed a-priori as a means to maintaining fact-providing LLMs, and continue with practical considerations for why even this ill-posed goal is probably unattainable. + +# 3.1 LLMs as Knowledge Repositories? + +The perception that LLMs can behave as knowledge repositories was first formulated and experimentally supported with the introduction of the LAMA benchmark (Petroni et al., 2019), where pretrained language models were queried in a zero-shot setting against 51K knowledge triplets extracted from knowledge banks and reformulated as fill-in-the-blank statements. 2019-level models (BERT-XL) got the top answer correctly about $26.5\%$ of the time. The limitations of the LAMA + +study were later studied and this result was shown to not be robust against multi-token spans (as opposed to a single token answer; Kalo and Fichtel, 2022), in an open vocabulary setting (as opposed to a closed vocabulary of 21K tokens; Roberts et al., 2020), and when queries vary (Kassner and Schütze, 2020); indicating that the LAMA experiment relied on heuristics to predict answers. Since then, more realistic benchmarks have been introduced, as well as more robust querying techniques (Jiang et al., 2020), to address some of these limitations. As the scale of LLMs increased, recent work has also scaled up the size of benchmarks, shifting the focus to facts concerning rare entities, where even recent LLMs struggle (Kandpal et al., 2023; Sun et al., 2023). These experiments indicate that an LM's ability to answer a question depends on the number of times information relevant to that question appears in the pretraining data. As a result, LLMs cannot be safely used to answer queries about the long-tail of facts that have been rarely mentioned in pretraining (Mallen et al., 2023). + +Beyond the capability to reliably answer queries, LLMs should also fulfill other requirements in order to be considered as fact repositories (AlKhamissi et al., 2022): the ability to edit knowledge (add, delete, update facts), logical consistency (answers to different but related facts must agree), reasoning (the ability to infer additional answers on the basis of logical rules), and explainability and interpretability (supporting an answer through a convincing chain of arguments). Experimental results assessing these dimensions indicate that current LLMs fail on all of them. He et al. (2023) demonstrate that LLMs underperform on computing ontology subsumption inference when compared with other trained approaches such as NLI models and symbolic systems such as OWL Reasoners (Glimm et al., 2014). + +The evidence supporting the premise that LLMs can be used as fact repositories is, thus, currently weak. Beyond this fragile foundation, we focus below on specific arguments regarding the capability to edit models as if they were fact repositories. + +# 3.2 Systemic Mismatch + +A very basic property of LLMs which contrasts with their usage as knowledge repositories is their stochastic nature. When used for augmenting creative endeavours, for data exploration, or for freeform tasks such as summarization, this property + +is desirable: the use case calls for variation or unexpectedness. In other cases, we may be content with models that supply us with a distribution of outputs, from which we can estimate the probability of individual responses and adjust our expectation for a reliable output. Since the latter is absent from many models available only through 3rd-party APIs, we are left with obtaining text generated from an unknown distribution, which we argue is insufficient for factuality-dependent applications.1 It can even be argued that getting facts wrong is a feature of vanilla LLMs rather than a bug: as their core training procedure aims to simulate plausible continuation of text, it should not surprise us that models would repeat widely-assumed falsehoods in a way that negates the idea of using them for factual purposes. If most people think, and write, that Los Angeles is the capital of California,2 an LLM is supposed to complete a relevant prompt accordingly. There is also no built-in reliability or robustness in LLMs that sample outputs from a distribution: two instances of the same prompt can easily produce contradicting facts, and indeed often do. + +Additionally, the idea of editing facts in a model suggests that we always want a model to supply us with a fact as an answer to a question. However, at times, questions may be posed while pre-supposing or otherwise assuming harmful propositions such as stereotypes or conspiracy theories. Editing the "fact" associated with the question "which government agency faked the moon landing?" would not provide us with an improved model; what we may want is to remove such facts altogether, or provide the model with a means of challenging the presupposition, or avoiding giving any answer at all. At the same time, many relations that we would term "facts" can be argued to be vital notions without which certain kinds of basic communication is impossible.3 An LLM that cannot assert whether trees have leaves, or asserts that they never do, is in danger of becoming irrelevant for most tasks requiring any kind of interaction with the world. As philosophy and practice surrounding these questions progresses, we can hope this gap between 'must-know' and 'must-not-know' will eventually tighten to form workable bounds on LLM knowledge capacity. + +# 3.3 Architectural Implausibility + +Estimates vary wildly, but there are over 100 million notable facts in the world. Even the cutoff for what constitutes a fact is unclear. Does a $3\%$ change in a demographics statistic, or a new esoteric sports record, call for an edit? Do the daily whereabouts of world leaders constitute facts? What about those of celebrities or journalists? As events unfold daily in world politics, economics, sports, and many other walks of life, facts are added and changed in larger quantities and rates than can be plausibly "caught up with" through surgical model editing, akin to emptying the ocean with a spoon. If we choose to limit ourselves in which facts we find important enough to edit, we introduce bias into the system, opening the door to a host of documented harms which permeate many language technologies (Chang et al., 2019; Blodgett et al., 2020). This choice can be implicit as well as explicit, and is incredibly hard to avoid. + +In a similar vein, the vastness and variability of facts is likely to lead to bias in evaluating the complement set of edits, those facts controlled for as not changing following an edit (drawdown). Even paraphrases of edited facts are not guaranteed to change alongside the selected phrasing (De Cao et al., 2021), nor are entailed facts (Hase et al., 2023b). This problem also manifests as a safety issue, since unchecked facts can turn out to be quite important for model usage, but perhaps taken for granted (or simply not explicitly covered) when designing a drawdown benchmark. + +There is evidence (Mallen et al., 2023; Jang et al., 2021) that facts above a certain "popularity threshold", measured in terms of views on Wikipedia articles, are harder to edit out of models compared to facts lying on the long tail of view distributions. Inherently being out of sight, the unpopular facts thus become susceptible to the double risk of being edited alongside target facts while not being deemed important enough to be checked during drawdown testing. The ultimate outcome of such procedures can be a monolithization of LLM-supplied "knowledge" that focuses on certain popular domains and interests while losing all usefulness in many topics contributing to the extensive diversity of the human and natural experience. + +Empirical evidence indicates existing editing models fail to properly account for the ripple effect of a fact editing operation (Cohen et al., 2023). For example, the insertion of the fact "Jack Depp is the son of Johnny Depp" introduces a "ripple effect" in the form of additional facts that the model needs to update (e.g. "Jack Depp is the sibling of Lily-Rose Depp"). Results in symbolic approaches to this task have demonstrated that this knowledge updating task is of high computational complexity, even NP-hard, for example in Truth Maintenance Systems (TMS; Rutenburg, 1991). These results carry to approaches based on machine learning techniques (Knblauch et al., 2020). There is, therefore, theoretical basis to conclude that model editing will at best address the problem of consistent updating in a roughly approximate manner, and most likely fail to update rarely seen facts within the ripple effect of editing operations. + +Finally, recent empirical findings show additional weaknesses of edit methods by extending the evaluation suites to cover aspects beyond fact editing metrics, such as specificity and robustness of the models post-editing (Onoe et al., 2023; Hoelscher-Obermaier et al., 2023; Hase et al., 2023a; Brown et al., 2023). + +# 4 Alternatives to Model Editing + +Model editing is motivated by the desire to control the text generated by LLMs to make them more compliant with desirable outcomes, specifically to control factuality: when an LLM generates text expressing an incorrect or obsolete fact, remove the fact from the LLM's "memory" and try again (Peng et al., 2023). Other relatively direct approaches to improve LLM factuality (and, in general, to control and revise the text generated by LLMs) have been proposed, most notably Reinforcement Learning with Human Feedback (RLHF; Bai et al., 2022; Ouyang et al., 2022; Ramamurthy et al., 2023; Rafailov et al., 2023; Carta et al., 2023), or adding a factuality objective to the training procedure (Lee et al., 2022). We now survey approaches which define a more achievable objective than that pursued by direct model editing or output manipulation. These approaches avoid the assumption that the LLM is a fact repository, and thus steer clear of the attempt to update this repository in a logically consistent manner, which is computationally hard to achieve and verify. While these alternatives avoid the more problematic aspects of model + +editing, they still suffer from their own limitations, but we argue that they identify more promising research directions. + +Incorporating Knowledge Bases In retrieval-based models, factual knowledge is explicitly represented in a dedicated component external to the LLM. The way this external fact store is represented and combined with the LLM varies: it can be a collection of textual documents that is searched using a text retrieval component, or an RDF graph, or encoded as a set of vector embeddings, or represented as modular expert LMs trained on curated datasets. In all cases, in the retrieval-based approach, the model can explicitly cite the source which underlies a specific generation, and let the user decide its credibility. + +Once external (non-parametric) knowledge is retrieved, it must be composed with the LLM generation process. Khandelwal et al. (2020) introduce a k-nearest neighbor method with interpolation; RETRO (Borgeaud et al., 2021) combines the prompt with retrieved documents through a specialized cross-attention mechanism; other LLM-IR variants include a fusion-in-decoder method (ATLAS; Izacard et al., 2022) and TRIME (Zhong et al., 2022), all retrieval-based models that maintain the capacity for few-shot in-context learning. + +Recent work addresses the integration of LLMs and IR by learning to combine results from search engines into the context provided to the LLM to answer a specific question (Xu et al., 2023). SAIL (Luo et al., 2023) introduces an instruction fine-tuning method to ground language generation on search results and learn to select trustworthy sources. CooK (Feng et al., 2023) approaches the task of combining multiple curated modular knowledge sources into an integrative system, where sources are modeled as independent LMs and an integrator LLM combines the information from these modules to answer a question. + +In all of these approaches, factual knowledge is stored outside of the parameters of the LLM and can be manipulated without retraining the LLM. These approaches have been shown to be scalable in the number of facts. Editing the fact store means the same as updating a database, thus simplifying the formulation of the task. + +Retrieval-based models, arguably, do not resolve all of the concerns we identify with model editing. The problem of identifying the provenance of a given generation span in these combined models + +remains acute: the text can be determined by the internal LLM parameters, by the external stores, or by their combination, even if they are not logically consistent with one another. Facts that have been seen more often during LLM training may have more weight in this interpolation even if they are wrong or when they become obsolete. Zhu et al. (2020) claim that "modifying knowledge in explicit memory module networks like FaE (Verga et al., 2020) is not easier and demands changes in the LLM component as well." This effect was also identified in the RealQA baseline test (Kasai et al., 2022): in this benchmark containing questions about time-sensitive data, experiments showed that in many cases GPT-3 (Brown et al., 2020) properly integrated data retrieved from a KB and injected into the context, but often still generated text based on outdated data from the LLM. While the clear separation of the components and the identification of the composition of external knowledge improve transparency-sensitive tasks, the objective of identifying provenance in the generated text and controlling which sources are appropriate in a given context remains an open research question. The formulation of the task of attributing generated text to identifiable sources (AIS; Rashkin et al., 2022) is a key contribution to this direction. Neeman et al. (2022) also address an aspect of this issue with a counterfactual data augmentation technique to disentangle contextual and LLM knowledge when generating text. + +Continual training focuses on incrementally training a model by introducing new tasks or new domains (e.g., Razdaibiedina et al., 2023). Model editing does not directly fall within this objective, as it concerns updating precise elements in the model while keeping tasks and domains unchanged. Yet, the identification of drawdown in model editing is similar to the risk of catastrophic forgetting identified in continual learning. Within this approach, we could situate model editing as a type of re-training or post-training. Zhu et al. (2020) note that just fine-tuning over a set of facts-to-update costs in degradation on other facts. Jang et al. (2021) identify the problem and propose to apply techniques from continual training to the task of incremental updating of LLM knowledge. Overall, while continuous training may seem to a priori avoid the risks of the model editing approach, it seems to suffer from many of the main evaluation problems. + +Concept Erasure The goal of concept erasure (Elazar and Goldberg, 2018; Ravfogel et al., 2020; Belrose et al., 2023) is to remove unwanted bias from embeddings generated by LLMs and subsequently from the generated text. This goal is motivated by fairness objectives: preventing protected attributes from causally affecting text generation. This motivation is somehow related to that of model editing (prevent damage from generated text) but key differences exist: (1) the method addresses general, well-identified concepts (protected attributes such as gender, race, age) as opposed to less well-defined specific facts; (2) it operates as a post-hoc transformation of embeddings, as opposed to a modification of the model itself, and as such it allows for more accountability than an ever-changing model; (3) it is verifiable with well-defined metrics over a limited pre-declared scope. While concept erasure has more limited scope than model editing, it defines an objective that can be evaluated with aggregated metrics in a robust manner. One possible avenue of future research can be to examine whether the erasure approach can be extended to address specific aspects of factuality, such as temporal validity. + +Maybe it's better to have a model that knows what it doesn't know? As identified in Kasai et al. (2022), a prerequisite to avoid generating wrong text is to identify what is not known: "can an open-domain QA system identify unanswerable cases?" The related issue of unanswerable questions has been addressed in a robust way (Rajpurkar et al., 2018; Sulem et al., 2021, 2022); yet the challenge in the context of LLMs and factuality is that problematic questions like those specified in §3.2 do look answerable. + +# 5 Conclusion + +We agree that model editing is an attractive task to define, with clear benchmarks and expected outcomes. However, in current practice it contributes towards unrealistic expectations that we can solve problems like LLM hallucinations, which would lead to potential harm in unleashing use cases that are not in fact within the capabilities of LLMs alone. We advocate for the usage of retrieval-augmented methods and other structural and post-hoc methods in order to achieve the stated large-scale goals, while conceding the benefits of editing to "safer" applications such as model interpretability and robustness checking. + +# Acknowledgments + +We thank Sarah Wiegrefe for comments on earlier drafts. We thank the reviewers for their valuable feedback. Yuval Pinter was supported in part by the Israeli Ministry of Innovation, Science and Technology (Grant 2022/5451). + +# Limitations + +This is an opinion paper. We have purposely not made empirical analyses in order to support our critique, knowing that data is contestable and may vary according to its collection methodology. We aim to convince on the merit of examples and rhetorical argumentation rather than concrete evidence, mostly out of our reach for the models under consideration, which we nevertheless urge readers and practitioners to seek and use in attempts to either support or disprove our claims. + +# Ethical Considerations + +# References + +Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona T. Diab, and Marjan Ghazvininejad. 2022. A review on language models as knowledge bases. ArXiv, abs/2204.06031. +Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. +Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. Leace: Perfect linear concept erasure in closed form. +Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of "bias" in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454-5476, Online. Association for Computational Linguistics. +Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosse-lut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258. +Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, T. W. Hennigan, Saffron Huang, Lorenzo Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and L. Sifre. 2021. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning. +Davis Brown, Charles Godfrey, Cody Nizinski, Jonathan Tu, and Henry Kvinge. 2023. Robustness of edited neural networks. +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901. +Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer. + +2023. Grounding large language models in interactive environments with online reinforcement learning. arXiv preprint arXiv:2302.02662. +Kai-Wei Chang, Vinodkumar Prabhakaran, and Vicente Ordonez. 2019. Bias and fairness in natural language processing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): Tutorial Abstracts, Hong Kong, China. Association for Computational Linguistics. +Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2023. Evaluating the ripple effects of knowledge editing in language models. arXiv preprint arXiv:2307.12976. +Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491-6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computational Linguistics. +Shangbin Feng, Weijia Shi, Yuyang Bai, Vidhisha Balachandran, Tianxing He, and Yulia Tsvetkov. 2023. Cook: Empowering general-purpose language models with modular and collaborative knowledge. ArXiv, abs/2305.09955. +Birte Glimm, Ian Horrocks, Boris Motik, Giorgos Stoilos, and Zhe Wang. 2014. Hermit: An owl 2 reasoner. Journal of Automated Reasoning, 53:245-269. +Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023a. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models. +Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2023b. Methods for measuring, updating, and visualizing factual beliefs in language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2714-2731, Dubrovnik, Croatia. Association for Computational Linguistics. +Yuan He, Jiaoyan Chen, Ernesto Jimenez-Ruiz, Hang Dong, and Ian Horrocks. 2023. Language model analysis for ontology subsumption inference. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 3439-3453, Toronto, Canada. Association for Computational Linguistics. +Jason Hoelscher-Obermaier, Julia Persson, Esben Kran, Ioannis Konstas, and Fazl Berez. 2023. Detecting edit failures in large language models: An improved specificity benchmark. + +Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane A. Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. *ArXiv*, abs/2208.03299. +Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, and Minjoon Seo. 2021. Towards continual knowledge learning of language models. International Conference on Learning Representations, abs/2110.03215. +Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438. +Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In 4th Conference on Automated Knowledge Base Construction. +Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15696-15707. PMLR. +Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. 2022. Realtime qa: What's the answer right now? +Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811-7818, Online. Association for Computational Linguistics. +Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2020. Generalization through memorization: Nearest neighbor language models. +Jeremias Knoblauch, Hisham Husain, and Tom Diethe. 2020. Optimal continual learning has perfect memory and is np-hard. In International Conference on Machine Learning, pages 5327-5337. PMLR. +Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Factuality enhanced language models for open-ended text generation. ArXiv, abs/2206.04624. +Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. 2023. Sail: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225. + +Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. +Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372. +Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. Fast model editing at scale. In ICLR. +Ella Neeman, Roee Aharoni, Or Honovich, Leshem Choshen, Idan Szpektor, and Omri Abend. 2022. Disentqa: Disentangling parametric and contextual knowledge with counterfactual question answering. ArXiv, abs/2211.05655. +Yasumasa Onoe, Michael J. Q. Zhang, Shankar Padmanabhan, Greg Durrett, and Eunsol Choi. 2023. Can lms learn new entities from descriptions? challenges in propagating injected knowledge. +Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744. +Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813. +Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics. +Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. +Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for squad. In Annual Meeting of the Association for Computational Linguistics. +Rajkumar Ramamurthy, Prithviraj Ammanabrolu, Kianté Brantley, Jack Hessel, Rafet Sifa, Christian Bauckhage, Hannaneh Hajishirzi, and Yejin Choi. 2023. Is reinforcement learning (not) for natural language processing: Benchmarks, baselines, and + +building blocks for natural language policy optimization. In The Eleventh International Conference on Learning Representations. +Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2022. Measuring attribution in natural language generation models. +Roger Ratcliff. 1990. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. *Psychological review*, 97(2):285. +Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics. +Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. 2023. Progressive prompts: Continual learning for language models. ArXiv, abs/2301.12314. +Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418-5426, Online. Association for Computational Linguistics. +Vladislav Rutenberg. 1991. Complexity classification of truth maintenance systems. In STACS 91: 8th Annual Symposium on Theoretical Aspects of Computer Science Hamburg, Germany, February 14-16, 1991 Proceedings 8, pages 372-383. Springer. +Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitry Pyrkin, Sergei Popov, and Artem Babenko. 2020. Editable neural networks. In International Conference on Learning Representations. +Elior Sulem, Jamaal Hay, and Dan Roth. 2021. Do we know what we don't know? studying unanswerable questions beyond squad 2.0. In Conference on Empirical Methods in Natural Language Processing. +Elior Sulem, Jamaal Hay, and Dan Roth. 2022. Yes, no or idk: The challenge of unanswerable yes/no questions. In *North American Chapter of the Association for Computational Linguistics*. +Kai Sun, Y. Xu, Hanwen Zha, Yue Liu, and Xinhsuai Dong. 2023. Head-to-tail: How knowledgeable are large language models (llm)? a.k.a. will llms replace knowledge graphs? ArXiv, abs/2308.10168. +Pat Verga, Haitian Sun, Livio Baldini Soares, and William W. Cohen. 2020. Facts as experts: Adaptable and interpretable neural memory over symbolic knowledge. + +Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-seng Chua. 2023. Search-in-the-chain: Towards the accurate, credible and traceable content generation for complex knowledge-intensive tasks. arXiv preprint arXiv:2304.14732. +Zexuan Zhong, Tao Lei, and Danqi Chen. 2022. Training language models with memory augmentation. In Conference on Empirical Methods in Natural Language Processing. +Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363. \ No newline at end of file diff --git a/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/layout.json b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5d3bf536e0b9491a68887b9146118c27009760b7 --- /dev/null +++ b/2023/Emptying the Ocean with a Spoon_ Should We Edit Models_/layout.json @@ -0,0 +1,3917 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 112, + 75, + 483, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 75, + 483, + 93 + ], + "spans": [ + { + "bbox": [ + 112, + 75, + 483, + 93 + ], + "type": "text", + "content": "Emptying the Ocean with a Spoon: Should We Edit Models?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 207, + 110, + 390, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 110, + 390, + 181 + ], + "spans": [ + { + "bbox": [ + 207, + 110, + 390, + 181 + ], + "type": "text", + "content": "Yuval Pinter and Michael Elhadad \nDepartment of Computer Science \nBen-Gurion University of the Negev \nBe'er Sheva, Israel \n{uvp, elhadad}@cs.bgu.ac.il" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 157, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 157, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 233, + 273, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 233, + 273, + 376 + ], + "spans": [ + { + "bbox": [ + 86, + 233, + 273, + 376 + ], + "type": "text", + "content": "We call into question the recently popularized method of direct model editing as a means of correcting factual errors in LLM generations. We contrast model editing with three similar but distinct approaches that pursue better defined objectives: (1) retrieval-based architectures, which decouple factual memory from inference and linguistic capabilities embodied in LLMs; (2) concept erasure methods, which aim at preventing systemic bias in generated text; and (3) attribution methods, which aim at grounding generations into identified textual sources." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 381, + 273, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 381, + 273, + 513 + ], + "spans": [ + { + "bbox": [ + 86, + 381, + 273, + 513 + ], + "type": "text", + "content": "We argue that direct model editing cannot be trusted as a systematic remedy for the disadvantages inherent to LLMs, and while it has proven potential in improving model explainability, it opens risks by reinforcing the notion that models can be trusted for factuality. We call for cautious promotion and application of model editing as part of the LLM deployment process, and for responsibly limiting the use cases of LLMs to those not relying on editing as a critical component." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 522, + 154, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 522, + 154, + 534 + ], + "spans": [ + { + "bbox": [ + 69, + 522, + 154, + 534 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 543, + 290, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 543, + 290, + 745 + ], + "spans": [ + { + "bbox": [ + 69, + 543, + 290, + 745 + ], + "type": "text", + "content": "Large language models, or LLMs, have taken the NLP world by storm. After originally focusing on training them as a vehicle for transfer learning, recent advancements have cast LLMs in the role of one-stop shop, all-knowing oracles. One particular finding that has contributed to this reframed usage is the apparent factuality properties of pre-trained LLMs: somehow, mere next-word prediction training has produced models that can complete certain correct facts about the world when asked to do so. LLMs have been perceived by the public at large as a replacement for search engines, and the scant disclaimers appearing in commercial services offering LLM querying do not appear to have made a dent in this perception." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 746, + 290, + 772 + ], + "type": "text", + "content": "Despite these developments, LLMs of course keep making mistakes. There is, after all, an inher" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 213, + 525, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 525, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 525, + 496 + ], + "type": "text", + "content": "ent mismatch between their pre-training objective and the desire for factuality. In recent years, researchers have come up with several remedies to the problem of nonfactual LLM outputs, one of which being model editing, where parameters inside an LLM are tweaked based on individual facts marked for correction. These works focus on solving problems such as ensuring stability of other factual outputs following the editing, or batch editing, or computationally efficient editing. In this opinion paper, we call into question the entire idea of model editing, citing concerns about intended use cases, conceptual scalability, potential for bias, safety, and overall accountability. We advocate using models that have explicit knowledge modules for tasks requiring factual knowledge, and other existing techniques to mitigate the need for model editing in the first place. Having said that, we acknowledge the usefulness of direct editing for certain use cases such as interpretability probes, and recommend limiting editing to such scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 507, + 397, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 507, + 397, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 507, + 397, + 521 + ], + "type": "text", + "content": "2 Model Editing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 529, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 525, + 772 + ], + "type": "text", + "content": "Sinitsin et al. (2020) first introduce the notion of updating large ML models to account for local performance expectations motivated externally. They cite cases where mistakes are critical, like object detection in self-driving cars. Later work suggests model editing can aid in protecting privacy and eliminating bias (Zhu et al., 2020), and as a measure of \"catching up\" with time-sensitive facts such as \"PM of the UK\" changing over time (Bommasani et al., 2021; Mitchell et al., 2022), or as a means of understanding the mechanics of black box models (Meng et al., 2022). Sinitsin et al. (2020) specify several desired properties of editing methods: reliability (the target facts are updated as intended), locality (other changes don't happen; a measure for this property is termed drawdown), and efficiency (successful edits are computationally light); later work (De Cao et al., 2021) added generality (the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 289, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 289, + 219 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 289, + 219 + ], + "type": "text", + "content": "ability to modify models not originally trained for knowledge retention), consistency (robustness to paraphrasing, in the specific use case of text models), and an unspecified property we'll call frugality (only minimal components of the model are changed during editing). A well-studied limitation, which most of these works focus on mitigating, and related to locality, is catastrophic forgetting (Ratcliff, 1990), i.e., ensuring that an edited model does not lose performance on tasks for which it was explicitly trained and performs well on." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 221, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 221, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 69, + 221, + 289, + 422 + ], + "type": "text", + "content": "Methods for model editing have evolved from editable training (Sinitsin et al., 2020), a procedure requiring an a-priori decision that a model would later be edited, to the locality-motivated changing of specific parameters within models (Zhu et al., 2020; Meng et al., 2022). Recent work (Mitchell et al., 2022) draws attention to the danger of model performance degradation accumulating over the course of many successive edits, and seeks mitigation through improved methods. Hase et al. (2023b) extend the consistency requirement to hold over entailed and equivalent facts as well as paraphrases, suggesting model editing as a way of reconciling cases where certain facts are produced correctly but those entailed from them are not." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 435, + 131, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 435, + 131, + 449 + ], + "spans": [ + { + "bbox": [ + 69, + 435, + 131, + 449 + ], + "type": "text", + "content": "3 Critique" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 459, + 289, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 459, + 289, + 607 + ], + "spans": [ + { + "bbox": [ + 69, + 459, + 289, + 607 + ], + "type": "text", + "content": "In this section, we present the case against model editing as a practice, regardless of the performance obtained by any single method. We begin with an analysis of the premise underlying the model editing research objective: the hypothesis that LLMs can be used as fact repositories. We then focus on reasons for which editing facts cannot be designed a-priori as a means to maintaining fact-providing LLMs, and continue with practical considerations for why even this ill-posed goal is probably unattainable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 618, + 256, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 618, + 256, + 632 + ], + "spans": [ + { + "bbox": [ + 69, + 618, + 256, + 632 + ], + "type": "text", + "content": "3.1 LLMs as Knowledge Repositories?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 637, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 637, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 637, + 289, + 772 + ], + "type": "text", + "content": "The perception that LLMs can behave as knowledge repositories was first formulated and experimentally supported with the introduction of the LAMA benchmark (Petroni et al., 2019), where pretrained language models were queried in a zero-shot setting against 51K knowledge triplets extracted from knowledge banks and reformulated as fill-in-the-blank statements. 2019-level models (BERT-XL) got the top answer correctly about " + }, + { + "bbox": [ + 69, + 637, + 289, + 772 + ], + "type": "inline_equation", + "content": "26.5\\%" + }, + { + "bbox": [ + 69, + 637, + 289, + 772 + ], + "type": "text", + "content": " of the time. The limitations of the LAMA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 71, + 524, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 524, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 524, + 368 + ], + "type": "text", + "content": "study were later studied and this result was shown to not be robust against multi-token spans (as opposed to a single token answer; Kalo and Fichtel, 2022), in an open vocabulary setting (as opposed to a closed vocabulary of 21K tokens; Roberts et al., 2020), and when queries vary (Kassner and Schütze, 2020); indicating that the LAMA experiment relied on heuristics to predict answers. Since then, more realistic benchmarks have been introduced, as well as more robust querying techniques (Jiang et al., 2020), to address some of these limitations. As the scale of LLMs increased, recent work has also scaled up the size of benchmarks, shifting the focus to facts concerning rare entities, where even recent LLMs struggle (Kandpal et al., 2023; Sun et al., 2023). These experiments indicate that an LM's ability to answer a question depends on the number of times information relevant to that question appears in the pretraining data. As a result, LLMs cannot be safely used to answer queries about the long-tail of facts that have been rarely mentioned in pretraining (Mallen et al., 2023)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 370, + 524, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 524, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 524, + 599 + ], + "type": "text", + "content": "Beyond the capability to reliably answer queries, LLMs should also fulfill other requirements in order to be considered as fact repositories (AlKhamissi et al., 2022): the ability to edit knowledge (add, delete, update facts), logical consistency (answers to different but related facts must agree), reasoning (the ability to infer additional answers on the basis of logical rules), and explainability and interpretability (supporting an answer through a convincing chain of arguments). Experimental results assessing these dimensions indicate that current LLMs fail on all of them. He et al. (2023) demonstrate that LLMs underperform on computing ontology subsumption inference when compared with other trained approaches such as NLI models and symbolic systems such as OWL Reasoners (Glimm et al., 2014)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 602, + 524, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 602, + 524, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 602, + 524, + 669 + ], + "type": "text", + "content": "The evidence supporting the premise that LLMs can be used as fact repositories is, thus, currently weak. Beyond this fragile foundation, we focus below on specific arguments regarding the capability to edit models as if they were fact repositories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 684, + 421, + 698 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 684, + 421, + 698 + ], + "spans": [ + { + "bbox": [ + 305, + 684, + 421, + 698 + ], + "type": "text", + "content": "3.2 Systemic Mismatch" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 706, + 524, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 706, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 305, + 706, + 524, + 772 + ], + "type": "text", + "content": "A very basic property of LLMs which contrasts with their usage as knowledge repositories is their stochastic nature. When used for augmenting creative endeavours, for data exploration, or for freeform tasks such as summarization, this property" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 71, + 293, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 71, + 293, + 396 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 293, + 396 + ], + "type": "text", + "content": "is desirable: the use case calls for variation or unexpectedness. In other cases, we may be content with models that supply us with a distribution of outputs, from which we can estimate the probability of individual responses and adjust our expectation for a reliable output. Since the latter is absent from many models available only through 3rd-party APIs, we are left with obtaining text generated from an unknown distribution, which we argue is insufficient for factuality-dependent applications.1 It can even be argued that getting facts wrong is a feature of vanilla LLMs rather than a bug: as their core training procedure aims to simulate plausible continuation of text, it should not surprise us that models would repeat widely-assumed falsehoods in a way that negates the idea of using them for factual purposes. If most people think, and write, that Los Angeles is the capital of California,2 an LLM is supposed to complete a relevant prompt accordingly. There is also no built-in reliability or robustness in LLMs that sample outputs from a distribution: two instances of the same prompt can easily produce contradicting facts, and indeed often do." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 398, + 292, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 398, + 292, + 724 + ], + "spans": [ + { + "bbox": [ + 66, + 398, + 292, + 724 + ], + "type": "text", + "content": "Additionally, the idea of editing facts in a model suggests that we always want a model to supply us with a fact as an answer to a question. However, at times, questions may be posed while pre-supposing or otherwise assuming harmful propositions such as stereotypes or conspiracy theories. Editing the \"fact\" associated with the question \"which government agency faked the moon landing?\" would not provide us with an improved model; what we may want is to remove such facts altogether, or provide the model with a means of challenging the presupposition, or avoiding giving any answer at all. At the same time, many relations that we would term \"facts\" can be argued to be vital notions without which certain kinds of basic communication is impossible.3 An LLM that cannot assert whether trees have leaves, or asserts that they never do, is in danger of becoming irrelevant for most tasks requiring any kind of interaction with the world. As philosophy and practice surrounding these questions progresses, we can hope this gap between 'must-know' and 'must-not-know' will eventually tighten to form workable bounds on LLM knowledge capacity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 302, + 71, + 463, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 463, + 85 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 463, + 85 + ], + "type": "text", + "content": "3.3 Architectural Implausibility" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 302, + 89, + 526, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 89, + 526, + 346 + ], + "spans": [ + { + "bbox": [ + 302, + 89, + 526, + 346 + ], + "type": "text", + "content": "Estimates vary wildly, but there are over 100 million notable facts in the world. Even the cutoff for what constitutes a fact is unclear. Does a " + }, + { + "bbox": [ + 302, + 89, + 526, + 346 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 302, + 89, + 526, + 346 + ], + "type": "text", + "content": " change in a demographics statistic, or a new esoteric sports record, call for an edit? Do the daily whereabouts of world leaders constitute facts? What about those of celebrities or journalists? As events unfold daily in world politics, economics, sports, and many other walks of life, facts are added and changed in larger quantities and rates than can be plausibly \"caught up with\" through surgical model editing, akin to emptying the ocean with a spoon. If we choose to limit ourselves in which facts we find important enough to edit, we introduce bias into the system, opening the door to a host of documented harms which permeate many language technologies (Chang et al., 2019; Blodgett et al., 2020). This choice can be implicit as well as explicit, and is incredibly hard to avoid." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 346, + 527, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 346, + 527, + 508 + ], + "spans": [ + { + "bbox": [ + 302, + 346, + 527, + 508 + ], + "type": "text", + "content": "In a similar vein, the vastness and variability of facts is likely to lead to bias in evaluating the complement set of edits, those facts controlled for as not changing following an edit (drawdown). Even paraphrases of edited facts are not guaranteed to change alongside the selected phrasing (De Cao et al., 2021), nor are entailed facts (Hase et al., 2023b). This problem also manifests as a safety issue, since unchecked facts can turn out to be quite important for model usage, but perhaps taken for granted (or simply not explicitly covered) when designing a drawdown benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 509, + 527, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 509, + 527, + 714 + ], + "spans": [ + { + "bbox": [ + 302, + 509, + 527, + 714 + ], + "type": "text", + "content": "There is evidence (Mallen et al., 2023; Jang et al., 2021) that facts above a certain \"popularity threshold\", measured in terms of views on Wikipedia articles, are harder to edit out of models compared to facts lying on the long tail of view distributions. Inherently being out of sight, the unpopular facts thus become susceptible to the double risk of being edited alongside target facts while not being deemed important enough to be checked during drawdown testing. The ultimate outcome of such procedures can be a monolithization of LLM-supplied \"knowledge\" that focuses on certain popular domains and interests while losing all usefulness in many topics contributing to the extensive diversity of the human and natural experience." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 739, + 280, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 739, + 280, + 750 + ], + "spans": [ + { + "bbox": [ + 80, + 739, + 280, + 750 + ], + "type": "text", + "content": "1Called \"transparent-sensitive tasks\" (Luo et al., 2023)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 750, + 230, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 750, + 230, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 750, + 230, + 761 + ], + "type": "text", + "content": "2The capital of California is Sacramento." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 81, + 761, + 240, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 240, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 240, + 772 + ], + "type": "text", + "content": "3We thank a reviewer for making this point." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 719, + 526, + 751 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 526, + 751 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 526, + 751 + ], + "type": "text", + "content": "4107,323,022 appear in Wikidata as of October 17, 2023, conforming to the definition of notable items used in https: //www.wikidata.org/wiki/Help:Items." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 525, + 772 + ], + "type": "text", + "content": "Attributed, according to the Internet, as a \"Yiddish proverb\", but, hey, who knows?" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 69, + 292, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 69, + 292, + 326 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 292, + 326 + ], + "type": "text", + "content": "Empirical evidence indicates existing editing models fail to properly account for the ripple effect of a fact editing operation (Cohen et al., 2023). For example, the insertion of the fact \"Jack Depp is the son of Johnny Depp\" introduces a \"ripple effect\" in the form of additional facts that the model needs to update (e.g. \"Jack Depp is the sibling of Lily-Rose Depp\"). Results in symbolic approaches to this task have demonstrated that this knowledge updating task is of high computational complexity, even NP-hard, for example in Truth Maintenance Systems (TMS; Rutenburg, 1991). These results carry to approaches based on machine learning techniques (Knblauch et al., 2020). There is, therefore, theoretical basis to conclude that model editing will at best address the problem of consistent updating in a roughly approximate manner, and most likely fail to update rarely seen facts within the ripple effect of editing operations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 329, + 290, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 329, + 290, + 423 + ], + "spans": [ + { + "bbox": [ + 67, + 329, + 290, + 423 + ], + "type": "text", + "content": "Finally, recent empirical findings show additional weaknesses of edit methods by extending the evaluation suites to cover aspects beyond fact editing metrics, such as specificity and robustness of the models post-editing (Onoe et al., 2023; Hoelscher-Obermaier et al., 2023; Hase et al., 2023a; Brown et al., 2023)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 438, + 240, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 438, + 240, + 452 + ], + "spans": [ + { + "bbox": [ + 67, + 438, + 240, + 452 + ], + "type": "text", + "content": "4 Alternatives to Model Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "type": "text", + "content": "Model editing is motivated by the desire to control the text generated by LLMs to make them more compliant with desirable outcomes, specifically to control factuality: when an LLM generates text expressing an incorrect or obsolete fact, remove the fact from the LLM's \"memory\" and try again (Peng et al., 2023). Other relatively direct approaches to improve LLM factuality (and, in general, to control and revise the text generated by LLMs) have been proposed, most notably Reinforcement Learning with Human Feedback (RLHF; Bai et al., 2022; Ouyang et al., 2022; Ramamurthy et al., 2023; Rafailov et al., 2023; Carta et al., 2023), or adding a factuality objective to the training procedure (Lee et al., 2022). We now survey approaches which define a more achievable objective than that pursued by direct model editing or output manipulation. These approaches avoid the assumption that the LLM is a fact repository, and thus steer clear of the attempt to update this repository in a logically consistent manner, which is computationally hard to achieve and verify. While these alternatives avoid the more problematic aspects of model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 303, + 71, + 525, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 525, + 111 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 525, + 111 + ], + "type": "text", + "content": "editing, they still suffer from their own limitations, but we argue that they identify more promising research directions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 121, + 525, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 121, + 525, + 296 + ], + "spans": [ + { + "bbox": [ + 302, + 121, + 525, + 296 + ], + "type": "text", + "content": "Incorporating Knowledge Bases In retrieval-based models, factual knowledge is explicitly represented in a dedicated component external to the LLM. The way this external fact store is represented and combined with the LLM varies: it can be a collection of textual documents that is searched using a text retrieval component, or an RDF graph, or encoded as a set of vector embeddings, or represented as modular expert LMs trained on curated datasets. In all cases, in the retrieval-based approach, the model can explicitly cite the source which underlies a specific generation, and let the user decide its credibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 297, + 526, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 297, + 526, + 446 + ], + "spans": [ + { + "bbox": [ + 302, + 297, + 526, + 446 + ], + "type": "text", + "content": "Once external (non-parametric) knowledge is retrieved, it must be composed with the LLM generation process. Khandelwal et al. (2020) introduce a k-nearest neighbor method with interpolation; RETRO (Borgeaud et al., 2021) combines the prompt with retrieved documents through a specialized cross-attention mechanism; other LLM-IR variants include a fusion-in-decoder method (ATLAS; Izacard et al., 2022) and TRIME (Zhong et al., 2022), all retrieval-based models that maintain the capacity for few-shot in-context learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 447, + 525, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 447, + 525, + 623 + ], + "spans": [ + { + "bbox": [ + 302, + 447, + 525, + 623 + ], + "type": "text", + "content": "Recent work addresses the integration of LLMs and IR by learning to combine results from search engines into the context provided to the LLM to answer a specific question (Xu et al., 2023). SAIL (Luo et al., 2023) introduces an instruction fine-tuning method to ground language generation on search results and learn to select trustworthy sources. CooK (Feng et al., 2023) approaches the task of combining multiple curated modular knowledge sources into an integrative system, where sources are modeled as independent LMs and an integrator LLM combines the information from these modules to answer a question." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 624, + 525, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 717 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 717 + ], + "type": "text", + "content": "In all of these approaches, factual knowledge is stored outside of the parameters of the LLM and can be manipulated without retraining the LLM. These approaches have been shown to be scalable in the number of facts. Editing the fact store means the same as updating a database, thus simplifying the formulation of the task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 719, + 525, + 773 + ], + "type": "text", + "content": "Retrieval-based models, arguably, do not resolve all of the concerns we identify with model editing. The problem of identifying the provenance of a given generation span in these combined models" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 289, + 491 + ], + "type": "text", + "content": "remains acute: the text can be determined by the internal LLM parameters, by the external stores, or by their combination, even if they are not logically consistent with one another. Facts that have been seen more often during LLM training may have more weight in this interpolation even if they are wrong or when they become obsolete. Zhu et al. (2020) claim that \"modifying knowledge in explicit memory module networks like FaE (Verga et al., 2020) is not easier and demands changes in the LLM component as well.\" This effect was also identified in the RealQA baseline test (Kasai et al., 2022): in this benchmark containing questions about time-sensitive data, experiments showed that in many cases GPT-3 (Brown et al., 2020) properly integrated data retrieved from a KB and injected into the context, but often still generated text based on outdated data from the LLM. While the clear separation of the components and the identification of the composition of external knowledge improve transparency-sensitive tasks, the objective of identifying provenance in the generated text and controlling which sources are appropriate in a given context remains an open research question. The formulation of the task of attributing generated text to identifiable sources (AIS; Rashkin et al., 2022) is a key contribution to this direction. Neeman et al. (2022) also address an aspect of this issue with a counterfactual data augmentation technique to disentangle contextual and LLM knowledge when generating text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 502, + 289, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 502, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 289, + 772 + ], + "type": "text", + "content": "Continual training focuses on incrementally training a model by introducing new tasks or new domains (e.g., Razdaibiedina et al., 2023). Model editing does not directly fall within this objective, as it concerns updating precise elements in the model while keeping tasks and domains unchanged. Yet, the identification of drawdown in model editing is similar to the risk of catastrophic forgetting identified in continual learning. Within this approach, we could situate model editing as a type of re-training or post-training. Zhu et al. (2020) note that just fine-tuning over a set of facts-to-update costs in degradation on other facts. Jang et al. (2021) identify the problem and propose to apply techniques from continual training to the task of incremental updating of LLM knowledge. Overall, while continuous training may seem to a priori avoid the risks of the model editing approach, it seems to suffer from many of the main evaluation problems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 305, + 71, + 525, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 525, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 525, + 409 + ], + "type": "text", + "content": "Concept Erasure The goal of concept erasure (Elazar and Goldberg, 2018; Ravfogel et al., 2020; Belrose et al., 2023) is to remove unwanted bias from embeddings generated by LLMs and subsequently from the generated text. This goal is motivated by fairness objectives: preventing protected attributes from causally affecting text generation. This motivation is somehow related to that of model editing (prevent damage from generated text) but key differences exist: (1) the method addresses general, well-identified concepts (protected attributes such as gender, race, age) as opposed to less well-defined specific facts; (2) it operates as a post-hoc transformation of embeddings, as opposed to a modification of the model itself, and as such it allows for more accountability than an ever-changing model; (3) it is verifiable with well-defined metrics over a limited pre-declared scope. While concept erasure has more limited scope than model editing, it defines an objective that can be evaluated with aggregated metrics in a robust manner. One possible avenue of future research can be to examine whether the erasure approach can be extended to address specific aspects of factuality, such as temporal validity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 417, + 525, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 525, + 564 + ], + "type": "text", + "content": "Maybe it's better to have a model that knows what it doesn't know? As identified in Kasai et al. (2022), a prerequisite to avoid generating wrong text is to identify what is not known: \"can an open-domain QA system identify unanswerable cases?\" The related issue of unanswerable questions has been addressed in a robust way (Rajpurkar et al., 2018; Sulem et al., 2021, 2022); yet the challenge in the context of LLMs and factuality is that problematic questions like those specified in §3.2 do look answerable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 576, + 380, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 576, + 380, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 576, + 380, + 588 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 597, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 525, + 772 + ], + "type": "text", + "content": "We agree that model editing is an attractive task to define, with clear benchmarks and expected outcomes. However, in current practice it contributes towards unrealistic expectations that we can solve problems like LLM hallucinations, which would lead to potential harm in unleashing use cases that are not in fact within the capabilities of LLMs alone. We advocate for the usage of retrieval-augmented methods and other structural and post-hoc methods in order to achieve the stated large-scale goals, while conceding the benefits of editing to \"safer\" applications such as model interpretability and robustness checking." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 165, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 165, + 84 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 165, + 84 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 92, + 291, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 92, + 291, + 159 + ], + "spans": [ + { + "bbox": [ + 67, + 92, + 291, + 159 + ], + "type": "text", + "content": "We thank Sarah Wiegrefe for comments on earlier drafts. We thank the reviewers for their valuable feedback. Yuval Pinter was supported in part by the Israeli Ministry of Innovation, Science and Technology (Grant 2022/5451)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 169, + 131, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 169, + 131, + 181 + ], + "spans": [ + { + "bbox": [ + 68, + 169, + 131, + 181 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 191, + 291, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 191, + 291, + 326 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 291, + 326 + ], + "type": "text", + "content": "This is an opinion paper. We have purposely not made empirical analyses in order to support our critique, knowing that data is contestable and may vary according to its collection methodology. We aim to convince on the merit of examples and rhetorical argumentation rather than concrete evidence, mostly out of our reach for the models under consideration, which we nevertheless urge readers and practitioners to seek and use in attempts to either support or disprove our claims." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 336, + 188, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 336, + 188, + 349 + ], + "spans": [ + { + "bbox": [ + 68, + 336, + 188, + 349 + ], + "type": "text", + "content": "Ethical Considerations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 71, + 362, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 71, + 362, + 83 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 362, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 89, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 304, + 89, + 527, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 527, + 134 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 527, + 134 + ], + "type": "text", + "content": "Badr AlKhamissi, Millicent Li, Asli Celikyilmaz, Mona T. Diab, and Marjan Ghazvininejad. 2022. A review on language models as knowledge bases. ArXiv, abs/2204.06031." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 142, + 527, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 142, + 527, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 142, + 527, + 275 + ], + "type": "text", + "content": "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom Brown, Jack Clark, Sam McCandlish, Chris Olah, Ben Mann, and Jared Kaplan. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 282, + 525, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 525, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 525, + 326 + ], + "type": "text", + "content": "Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. Leace: Perfect linear concept erasure in closed form." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 334, + 527, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 527, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 527, + 412 + ], + "type": "text", + "content": "Su Lin Blodgett, Solon Barocas, Hal Daumé III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of \"bias\" in NLP. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454-5476, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 420, + 527, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 527, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 527, + 485 + ], + "type": "text", + "content": "Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosse-lut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 494, + 527, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 527, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 527, + 626 + ], + "type": "text", + "content": "Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, T. W. Hennigan, Saffron Huang, Lorenzo Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Oriol Vinyals, Simon Osindero, Karen Simonyan, Jack W. Rae, Erich Elsen, and L. Sifre. 2021. Improving language models by retrieving from trillions of tokens. In International Conference on Machine Learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 634, + 525, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 634, + 525, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 634, + 525, + 666 + ], + "type": "text", + "content": "Davis Brown, Charles Godfrey, Cody Nizinski, Jonathan Tu, and Henry Kvinge. 2023. Robustness of edited neural networks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 676, + 525, + 741 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 676, + 525, + 741 + ], + "spans": [ + { + "bbox": [ + 304, + 676, + 525, + 741 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 525, + 772 + ], + "type": "text", + "content": "Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 290, + 105 + ], + "type": "text", + "content": "2023. Grounding large language models in interactive environments with online reinforcement learning. arXiv preprint arXiv:2302.02662." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 112, + 289, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 112, + 289, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 112, + 289, + 200 + ], + "type": "text", + "content": "Kai-Wei Chang, Vinodkumar Prabhakaran, and Vicente Ordonez. 2019. Bias and fairness in natural language processing. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP): Tutorial Abstracts, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "spans": [ + { + "bbox": [ + 69, + 208, + 289, + 253 + ], + "type": "text", + "content": "Roi Cohen, Eden Biran, Ori Yoran, Amir Globerson, and Mor Geva. 2023. Evaluating the ripple effects of knowledge editing in language models. arXiv preprint arXiv:2307.12976." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 259, + 289, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 259, + 289, + 325 + ], + "spans": [ + { + "bbox": [ + 69, + 259, + 289, + 325 + ], + "type": "text", + "content": "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021. Editing factual knowledge in language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6491-6506, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 333, + 289, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 333, + 289, + 398 + ], + "spans": [ + { + "bbox": [ + 69, + 333, + 289, + 398 + ], + "type": "text", + "content": "Yanai Elazar and Yoav Goldberg. 2018. Adversarial removal of demographic attributes from text data. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 11-21, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 406, + 289, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 406, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 69, + 406, + 289, + 460 + ], + "type": "text", + "content": "Shangbin Feng, Weijia Shi, Yuyang Bai, Vidhisha Balachandran, Tianxing He, and Yulia Tsvetkov. 2023. Cook: Empowering general-purpose language models with modular and collaborative knowledge. ArXiv, abs/2305.09955." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 468, + 289, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 468, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 69, + 468, + 289, + 502 + ], + "type": "text", + "content": "Birte Glimm, Ian Horrocks, Boris Motik, Giorgos Stoilos, and Zhe Wang. 2014. Hermit: An owl 2 reasoner. Journal of Automated Reasoning, 53:245-269." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 509, + 289, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 509, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 69, + 509, + 289, + 553 + ], + "type": "text", + "content": "Peter Hase, Mohit Bansal, Been Kim, and Asma Ghandeharioun. 2023a. Does localization inform editing? surprising differences in causality-based localization vs. knowledge editing in language models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 560, + 289, + 648 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 560, + 289, + 648 + ], + "spans": [ + { + "bbox": [ + 69, + 560, + 289, + 648 + ], + "type": "text", + "content": "Peter Hase, Mona Diab, Asli Celikyilmaz, Xian Li, Zornitsa Kozareva, Veselin Stoyanov, Mohit Bansal, and Srinivasan Iyer. 2023b. Methods for measuring, updating, and visualizing factual beliefs in language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2714-2731, Dubrovnik, Croatia. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 655, + 289, + 720 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 655, + 289, + 720 + ], + "spans": [ + { + "bbox": [ + 69, + 655, + 289, + 720 + ], + "type": "text", + "content": "Yuan He, Jiaoyan Chen, Ernesto Jimenez-Ruiz, Hang Dong, and Ian Horrocks. 2023. Language model analysis for ontology subsumption inference. In *Findings of the Association for Computational Linguistics: ACL* 2023, pages 3439-3453, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 289, + 772 + ], + "type": "text", + "content": "Jason Hoelscher-Obermaier, Julia Persson, Esben Kran, Ioannis Konstas, and Fazl Berez. 2023. Detecting edit failures in large language models: An improved specificity benchmark." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 772 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "text", + "content": "Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane A. Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2022. Few-shot learning with retrieval augmented language models. *ArXiv*, abs/2208.03299." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 139, + 524, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 139, + 524, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 139, + 524, + 204 + ], + "type": "text", + "content": "Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, and Minjoon Seo. 2021. Towards continual knowledge learning of language models. International Conference on Learning Representations, abs/2110.03215." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 216, + 524, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 524, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 524, + 260 + ], + "type": "text", + "content": "Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 272, + 524, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 524, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 524, + 316 + ], + "type": "text", + "content": "Jan-Christoph Kalo and Leandra Fichtel. 2022. Kamel: Knowledge analysis with multitoken entities in language models. In 4th Conference on Automated Knowledge Base Construction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 327, + 524, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 524, + 404 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 524, + 404 + ], + "type": "text", + "content": "Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 15696-15707. PMLR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 524, + 461 + ], + "type": "text", + "content": "Jungo Kasai, Keisuke Sakaguchi, Yoichi Takahashi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A. Smith, Yejin Choi, and Kentaro Inui. 2022. Realtime qa: What's the answer right now?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 472, + 524, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 524, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 524, + 539 + ], + "type": "text", + "content": "Nora Kassner and Hinrich Schütze. 2020. Negated and misprimed probes for pretrained language models: Birds can talk, but cannot fly. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7811-7818, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 550, + 524, + 593 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 524, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 524, + 593 + ], + "type": "text", + "content": "Urvashi Khandelwal, Omer Levy, Dan Jurafsky, Luke Zettlemoyer, and Mike Lewis. 2020. Generalization through memorization: Nearest neighbor language models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 524, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 524, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 524, + 650 + ], + "type": "text", + "content": "Jeremias Knoblauch, Hisham Husain, and Tom Diethe. 2020. Optimal continual learning has perfect memory and is np-hard. In International Conference on Machine Learning, pages 5327-5337. PMLR." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 661, + 524, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 661, + 524, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 661, + 524, + 705 + ], + "type": "text", + "content": "Nayeon Lee, Wei Ping, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2022. Factuality enhanced language models for open-ended text generation. ArXiv, abs/2206.04624." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 717, + 524, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 717, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 717, + 524, + 772 + ], + "type": "text", + "content": "Hongyin Luo, Yung-Sung Chuang, Yuan Gong, Tianhua Zhang, Yoon Kim, Xixin Wu, Danny Fox, Helen Meng, and James Glass. 2023. Sail: Search-augmented instruction learning. arXiv preprint arXiv:2305.15225." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 290, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 290, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 290, + 180 + ], + "type": "text", + "content": "Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359-17372." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 190, + 289, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 190, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 289, + 223 + ], + "type": "text", + "content": "Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D. Manning. 2022. Fast model editing at scale. In ICLR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 232, + 290, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 232, + 290, + 287 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 290, + 287 + ], + "type": "text", + "content": "Ella Neeman, Roee Aharoni, Or Honovich, Leshem Choshen, Idan Szpektor, and Omri Abend. 2022. Disentqa: Disentangling parametric and contextual knowledge with counterfactual question answering. ArXiv, abs/2211.05655." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 297, + 290, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 297, + 290, + 342 + ], + "spans": [ + { + "bbox": [ + 69, + 297, + 290, + 342 + ], + "type": "text", + "content": "Yasumasa Onoe, Michael J. Q. Zhang, Shankar Padmanabhan, Greg Durrett, and Eunsol Choi. 2023. Can lms learn new entities from descriptions? challenges in propagating injected knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 350, + 290, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 350, + 290, + 417 + ], + "spans": [ + { + "bbox": [ + 69, + 350, + 290, + 417 + ], + "type": "text", + "content": "Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 426, + 290, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 426, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 69, + 426, + 290, + 492 + ], + "type": "text", + "content": "Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. 2023. Check your facts and try again: Improving large language models with external knowledge and automated feedback. arXiv preprint arXiv:2302.12813." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 502, + 290, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 502, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 69, + 502, + 290, + 601 + ], + "type": "text", + "content": "Fabio Petroni, Tim Rocttäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2463-2473, Hong Kong, China. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 610, + 290, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 610, + 290, + 654 + ], + "spans": [ + { + "bbox": [ + 69, + 610, + 290, + 654 + ], + "type": "text", + "content": "Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 663, + 290, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 663, + 290, + 708 + ], + "spans": [ + { + "bbox": [ + 69, + 663, + 290, + 708 + ], + "type": "text", + "content": "Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don't know: Unanswerable questions for squad. In Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 717, + 290, + 772 + ], + "type": "text", + "content": "Rajkumar Ramamurthy, Prithviraj Ammanabrolu, Kianté Brantley, Jack Hessel, Rafet Sifa, Christian Bauckhage, Hannaneh Hajishirzi, and Yejin Choi. 2023. Is reinforcement learning (not) for natural language processing: Benchmarks, baselines, and" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 105 + ], + "type": "text", + "content": "building blocks for natural language policy optimization. In The Eleventh International Conference on Learning Representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 115, + 525, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 115, + 525, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 115, + 525, + 170 + ], + "type": "text", + "content": "Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and David Reitter. 2022. Measuring attribution in natural language generation models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 179, + 525, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 525, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 525, + 223 + ], + "type": "text", + "content": "Roger Ratcliff. 1990. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. *Psychological review*, 97(2):285." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 232, + 525, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 525, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 525, + 310 + ], + "type": "text", + "content": "Shauli Ravfogel, Yanai Elazar, Hila Gonen, Michael Twiton, and Yoav Goldberg. 2020. Null it out: Guarding protected attributes by iterative nullspace projection. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7237-7256, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 319, + 525, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 525, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 525, + 364 + ], + "type": "text", + "content": "Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. 2023. Progressive prompts: Continual learning for language models. ArXiv, abs/2301.12314." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 373, + 525, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 525, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 525, + 439 + ], + "type": "text", + "content": "Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5418-5426, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 449, + 525, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 525, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 525, + 504 + ], + "type": "text", + "content": "Vladislav Rutenberg. 1991. Complexity classification of truth maintenance systems. In STACS 91: 8th Annual Symposium on Theoretical Aspects of Computer Science Hamburg, Germany, February 14-16, 1991 Proceedings 8, pages 372-383. Springer." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 513, + 525, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 513, + 525, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 525, + 557 + ], + "type": "text", + "content": "Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitry Pyrkin, Sergei Popov, and Artem Babenko. 2020. Editable neural networks. In International Conference on Learning Representations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 566, + 525, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 525, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 525, + 612 + ], + "type": "text", + "content": "Elior Sulem, Jamaal Hay, and Dan Roth. 2021. Do we know what we don't know? studying unanswerable questions beyond squad 2.0. In Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 620, + 525, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 525, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 525, + 665 + ], + "type": "text", + "content": "Elior Sulem, Jamaal Hay, and Dan Roth. 2022. Yes, no or idk: The challenge of unanswerable yes/no questions. In *North American Chapter of the Association for Computational Linguistics*." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 674, + 525, + 719 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 525, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 525, + 719 + ], + "type": "text", + "content": "Kai Sun, Y. Xu, Hanwen Zha, Yue Liu, and Xinhsuai Dong. 2023. Head-to-tail: How knowledgeable are large language models (llm)? a.k.a. will llms replace knowledge graphs? ArXiv, abs/2308.10168." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 728, + 525, + 772 + ], + "type": "text", + "content": "Pat Verga, Haitian Sun, Livio Baldini Soares, and William W. Cohen. 2020. Facts as experts: Adaptable and interpretable neural memory over symbolic knowledge." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 233 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 127 + ], + "type": "text", + "content": "Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-seng Chua. 2023. Search-in-the-chain: Towards the accurate, credible and traceable content generation for complex knowledge-intensive tasks. arXiv preprint arXiv:2304.14732." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 136, + 291, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 136, + 291, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 136, + 291, + 180 + ], + "type": "text", + "content": "Zexuan Zhong, Tao Lei, and Danqi Chen. 2022. Training language models with memory augmentation. In Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 188, + 291, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 188, + 291, + 233 + ], + "spans": [ + { + "bbox": [ + 69, + 188, + 291, + 233 + ], + "type": "text", + "content": "Chen Zhu, Ankit Singh Rawat, Manzil Zaheer, Srinadh Bhojanapalli, Daliang Li, Felix Yu, and Sanjiv Kumar. 2020. Modifying memories in transformer models. arXiv preprint arXiv:2012.00363." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_content_list.json b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9dbc4dcb439bb5a44f1b871b40f388b56617bd6d --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations", + "text_level": 1, + "bbox": [ + 142, + 80, + 855, + 122 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengpeng Fu $^{1,2}$ , Xiaocheng Feng $^{1,2}$ , Yichong Huang $^{1}$ , Wenshuai Huo $^{1,2}$ , Hui Wang $^{2}$ , Bin Qin $^{1,2}$ , Ting Liu $^{1,2}$", + "bbox": [ + 189, + 129, + 816, + 164 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Harbin Institution of Technology, Harbin, China", + "bbox": [ + 300, + 165, + 699, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Pengcheng Laboratory, Shenzhen, China", + "bbox": [ + 331, + 181, + 670, + 197 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fuchp@stu.hit.edu.cn, {huowsh, wangh06}@pc1.ac.cn", + "bbox": [ + 258, + 198, + 744, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xcfeng,ychuang,qinb,tliu}@ir.hit.edu.cn", + "bbox": [ + 297, + 215, + 704, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised neural machine translation has recently made remarkable strides, achieving impressive results with the exclusive use of monolingual corpora. Nonetheless, these methods still exhibit fundamental flaws, such as confusing similar words. A straightforward remedy to rectify this drawback is to employ bilingual dictionaries, however, high-quality bilingual dictionaries can be costly to obtain. To overcome this limitation, we propose a method that incorporates images at the word level to augment the lexical mappings. Specifically, our method inserts visual representations into the model, modifying the corresponding embedding layer information. Besides, a visible matrix is adopted to isolate the impact of images on other unrelated words. Experiments on the Multi30k dataset with over 300,000 self-collected images validate the effectiveness in generating more accurate word translation, achieving an improvement of up to $+2.81$ BLEU score, which is comparable or even superior to using bilingual dictionaries. $^{1}$", + "bbox": [ + 141, + 279, + 460, + 605 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 618, + 258, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised Neural Machine Translation (UNMT) (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a,b,c) seeks to avoid the necessity for high-quality bilingual data in conventional machine translation by exploring the possibility of completing translations relying solely on the monolingual corpus. Consisting of three main components, cross-lingual language model initialization, denoising auto-encoder, and iterative back translation, recent methods (Conneau and Lample, 2019; Song et al., 2019) have achieved promising results.", + "bbox": [ + 112, + 643, + 487, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nonetheless, recent UNMT approaches have been observed to make particular errors in comparison to supervised machine translation, such", + "bbox": [ + 112, + 835, + 489, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as confusing nouns that pertain to the same semantic category. For instance, Bapna et al. (2022) demonstrates that their models translate \"lion\" to \"rabbit\" or \"snake\" despite getting good CHRF scores. In the context of supervised machine translation, pinpointing accurate translations of words is straightforward with aligned bilingual data. However, with only monolingual data, the unsupervised model must rely on large amounts of context for word mapping, which can lead to confusion between words with similar distributions, particularly when there is insufficient data. To address this issue, a direct solution is to integrate high-quality bilingual dictionaries into the UNMT process using codeswitching or spliced methods (Jones et al., 2023). High-quality bilingual dictionaries can be obtained through manual annotation and bilingual data extraction, yet these methods either involve high costs or deviate from the original objective of the UNMT task.", + "bbox": [ + 507, + 253, + 885, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In addition to utilizing explicit bilingual dictionaries for word mapping, implicit methods can also be employed. As the saying goes, \"A picture is worth a thousand words\". We can leverage the \"thousand words\" represented by corresponding images to learn the mapping between words in different languages. Visual content has the potential to enhance word mapping abilities in latent spaces, as the physical visual perception is congruent among speakers of diverse languages (Huang et al., 2020). Furthermore, the cost of annotating a bilingual dictionary is relatively high, whereas the cost of searching images with monolingual keywords is relatively low in the era of sophisticated search engines. Describing an image is also less demanding for humans compared to translating a word, as the former only requires mastery of one language while the latter requires at least two.", + "bbox": [ + 507, + 577, + 884, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Drawing on these insights, we propose to integrate word-level image information into the UNMT process to mitigate the issue of lexical confusion.", + "bbox": [ + 507, + 871, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The word-image dataset and associated codes are publicly available at: https://github.com/Brianmax99/UNMT-WVR", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12608", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12608-12618 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 210, + 945, + 786, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It's worth noting that we don't use sentence-level images like some multi-modal UNMT approaches (Su et al. (2019); Huang et al. (2020)). Instead, we focused on word-level images since they are more effective and flexible in addressing the challenge of lexical confusion and are easier to collect. Our approach involves augmenting the encoded images to their corresponding words, modifying the corresponding position and language encoding information, and utilizing a visible matrix to isolate the impact of images on other words. We conduct experiments utilizing over 300,000 self-collected images, and further analysis demonstrates that we have effectively mitigated lexical confusion to a certain extent. Furthermore, we achieve new state-of-the-art results for the UNMT task in certain language directions on the Multi30k dataset. The main contributions of this work can be summarized as follows:", + "bbox": [ + 112, + 84, + 489, + 388 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a concise yet potent framework that integrates word-level images to alleviate the similar lexical confusion issue in UNMT, without utilizing bilingual dictionaries.", + "- Our approach has achieved promising or even unparalleled unsupervised results on the Multi30K dataset for certain language pairs.", + "- We release the word-image datasets for five languages as open-source, which can serve as a valuable resource to the community." + ], + "bbox": [ + 136, + 406, + 487, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methods", + "text_level": 1, + "bbox": [ + 114, + 596, + 225, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first review the text-only UNMT process. Then, we present our proposed method for incorporating images, which involves the modified embedding layer and mask transformer. Finally, we introduce some training strategies.", + "bbox": [ + 112, + 625, + 489, + 706 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Unsupervised Neural Machine Translation", + "text_level": 1, + "bbox": [ + 112, + 719, + 487, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Successful UNMT systems typically employ an encoder-decoder architecture similar to that of supervised machine translation, sharing parameters between the source-to-target and target-to-source models. They share several common principles: initialization, iterative back-translation, and denoising autoencoder. Let $X = \\{X_{i}\\}_{i=1}^{M}$ denote monolingual data in language $L_{1}$ and $Y = \\{Y_{i}\\}_{i=1}^{N}$ in language $L_{2}$ . $M$ and $N$ are the sentence number of monolingual data $X$ and $Y$ respectively. $\\theta$ represents the UNMT model parameters.", + "bbox": [ + 112, + 741, + 490, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Initialization. Performing parameter initialization that is non-random can introduce cross-lingual prior knowledge into the model. There are currently two main initialization methods for UNMT models. The first method initializes the embedding layer of a UNMT model with word pretrained embeddings, while the second method uses a pretrained cross-lingual language model with the same structure as the UNMT encoder to initialize most of the neural network parameters.", + "bbox": [ + 507, + 84, + 884, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Iterative Back-translation. Back-translation is a method that leverages monolingual data in the target language and a reverse translation model to generate pseudo-source data, improving the quality of translations in the forward direction. In UNMT, where only monolingual data is available, back-translation is used to generate pseudo-parallel data, transforming the unsupervised problem into a weakly or self-supervised one. During training, back-translation is repeatedly applied in an iterative manner. The loss is defined as followed:", + "bbox": [ + 507, + 253, + 884, + 429 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {B} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {2} \\rightarrow L _ {1}} \\left(X _ {i} \\mid S _ {L _ {1} \\rightarrow L _ {2}} \\left(X _ {i}, \\theta\\right), \\theta\\right) \\tag {1} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {1} \\rightarrow L _ {2}} (Y _ {j} | S _ {L _ {2} \\rightarrow L _ {1}} (Y _ {j}, \\theta), \\theta) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 445, + 882, + 519 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $P_{L_a\\to L_b}$ denotes the translation probabilities between two languages and $S_{L_a\\rightarrow L_b}$ denotes the translation processes from language $L_{a}$ to $L_{b}$ .", + "bbox": [ + 507, + 525, + 880, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising Autoencoder. The introduction of the denoising autoencoder aims to guide the model in producing fluent text output. This is achieved by adding noise to sentences through techniques such as replacement, deletion, and shuffling, and then allowing the model to output the correct and fluent original sentence. The loss is defined as follows:", + "bbox": [ + 507, + 581, + 882, + 694 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {N} \\left(X _ {i}\\right), \\theta\\right) \\tag {2} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {N} (Y _ {j}), \\theta\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 569, + 709, + 882, + 783 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(\\cdot)$ refers to the noise functions. $P_{L_a\\to L_a}$ denotes the reconstruction probabilities in $L_{a}$ .", + "bbox": [ + 507, + 790, + 880, + 822 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The final objective function of UNMT is:", + "bbox": [ + 526, + 822, + 835, + 837 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {B} + \\lambda_ {D} \\mathcal {L} _ {D} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 626, + 846, + 882, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\lambda_{D}$ is a hyper-parameter that weights denoise autoencoder loss term and is gradually reduced during training.", + "bbox": [ + 507, + 871, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12609", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a27cac5cbce675af81757d07cbf656191f31d0d39c6ad2bffb58c815a869ec83.jpg", + "image_caption": [ + "Figure 1: The process of converting an image-inserted sentence into an embedding representation." + ], + "image_footnote": [], + "bbox": [ + 136, + 84, + 860, + 379 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Image Fusion", + "text_level": 1, + "bbox": [ + 112, + 430, + 268, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present how to incorporate word-level images into the UNMT process to tackle the issue of lexical confusion. The main concept behind our approach is that words can acquire information not only from contextual information but also from their corresponding images. Let $X_{i} = \\{x_{i}^{1}, x_{i}^{2}, \\dots, x_{i}^{K}\\}$ denote a sentence from $X$ in language $L_{1}$ and $x_{i}^{K}$ denotes the $k$ -th word in sentence $X_{i}$ . We denote a lookup table consisting of word-image pairs as $\\mathcal{D}_{L_i} = \\{(word, I_{word})\\}$ , where $I_{word}$ is an image set with one or multiple images that depict the meaning of the word.", + "bbox": [ + 112, + 451, + 487, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure $1^{2}$ , we segment an input sentence $X_{i}$ at the word level and obtain the images by retrieving the words appearing in the lookup table $\\mathcal{D}_{L_i}$ . Then, we concatenate the retrieved images to the corresponding words, resulting in an Image-Inserted Sentence denoted $IIS_{i}$ . Each word can be concatenated with one or multiple images. Formally, We denote $IIS_{i}$ as follow:", + "bbox": [ + 112, + 645, + 489, + 774 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI I S _ {i} = \\left\\{x _ {i} ^ {1}, \\widehat {I} _ {x _ {i} ^ {1}}, x _ {i} ^ {2}, \\widehat {I} _ {x _ {i} ^ {2}}, \\dots , x _ {i} ^ {K}, \\widehat {I} _ {x _ {i} ^ {K}} \\right\\} \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 784, + 487, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widehat{I}_{x_i^K}$ is the image subset of $I_{x_i^K}$ ( $\\widehat{I}_{x_i^K}$ can be an empty set). We define this process of injecting images as:", + "bbox": [ + 112, + 820, + 487, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI I S = \\mathcal {I} (X) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 448, + 882, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $IIS = \\{IIS_{i}\\}_{i = 1}^{M}$ .", + "bbox": [ + 507, + 472, + 702, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After obtaining the $IIS$ , we need to integrate them into the UNMT model. To achieve this integration, we employ a modified embedding layer and a mask transformer.", + "bbox": [ + 507, + 491, + 882, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.1 Modified Embedding Layer", + "text_level": 1, + "bbox": [ + 507, + 565, + 786, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first modification part is the embedding layer, which generally includes token embedding, position embedding, and, in some models, language embedding. An example can be seen in Figure 1.", + "bbox": [ + 505, + 586, + 882, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token Embedding. In the context of $IIS$ , both textual words and images are present. Given that most current machine translation methods operate at the subword level, we proceed to tokenize the $IIS$ , focusing solely on the textual component of $IIS$ . For encoding textual tokens, we continue to utilize the embedding layer from the original text-only model. Regarding images, we leverage the CLIP (Radford et al., 2021) model as the image encoder to extract visual features. The CLIP model is a multi-modal model trained on a variety of (image, text) pairs, which possesses a powerful capability to represent images. To prevent negative impacts on the encoding of images during the UNMT process, we freeze the visual encoder. However, directly using these representations may lead", + "bbox": [ + 507, + 661, + 884, + 919 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2\"Blue\" is an adjective in the figure. In our image collection process, \"blue\" can also appear as a noun in other training samples.", + "bbox": [ + 112, + 879, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12610", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to a gap between the image encoding space and the semantic space in UNMT. Therefore, we employ an MLP layer to bridge this gap.", + "bbox": [ + 112, + 84, + 487, + 134 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Embedding. Positional encoding is a crucial module that supplements token structural information. For textual information in $IIS$ , we continue to encode the position in the original order of the input sentence to prevent interference with sentence structure learning when integrating images into the model. For images, we assign the positional encoding of the first subword in the tokenized result for the corresponding word.", + "bbox": [ + 112, + 140, + 487, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language Embedding. Some UNMT models incorporate language embedding to inform the model of the input language and the desired output language. For textual information, we continue to utilize the original language encoding. However, for image information, as it does not belong to any language, we introduce a new image identifier.", + "bbox": [ + 112, + 294, + 489, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2.2 Mask Transformer", + "text_level": 1, + "bbox": [ + 114, + 414, + 324, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite modifying the original embedding layer to allow $IIS$ to be properly fed into the model, two issues still exist. Firstly, inserting images may affect other irrelevant words in the original sentence and potentially disrupt the structural relationships between them. Secondly, a single word-level image may correspond to multiple subwords, and during the annotation of position embedding, only the information of the first subword is utilized, resulting in a lack of relationship between the image and the remaining subwords it corresponds to. To address these issues, we introduce the method of mask transformer (Liu et al., 2020).", + "bbox": [ + 112, + 432, + 487, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create a visible matrix to shield the relationship between images and other textual information in the sentence and highlight the relationship between the image and its corresponding subwords. The visible matrix $M$ is defined as:", + "bbox": [ + 112, + 642, + 489, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & w _ {p} \\bullet w _ {q} \\\\ 0 & w _ {p} \\circ w _ {q}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 203, + 727, + 487, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{p}$ and $w_{q}$ are elements in $IIS_{i}$ , $w_{p} \\circ w_{q}$ denotes that $w_{p}$ and $w_{q}$ are visible to each other, while $w_{p} \\bullet w_{q}$ is not. An example of a visible matrix is depicted in Figure 2.", + "bbox": [ + 112, + 774, + 487, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The visible matrix can be added to the self-attention module to mask certain attention calculations. Specifically, the new masked self-attention is:", + "bbox": [ + 112, + 854, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/21012451b1d6a291a0d8c1850e6f6419c4639d9dc4addb6ec6ebd30cad1dbd66.jpg", + "image_caption": [ + "Figure 2: An example of the visible matrix corresponding to the $IIS$ in Figure 1. The red numerical labels indicate the absolute position of the images in the $IIS$ . In this matrix, black represents \"visible\" and white represents \"invisible\"." + ], + "image_footnote": [], + "bbox": [ + 586, + 82, + 806, + 239 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {i} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {i} K ^ {i T} + M}{\\sqrt {d}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 354, + 882, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ and $K$ are the query matrix and key matrix in the self-attention module. if two elements are invisible, the $M$ will mask the attention score to 0, which means making no contribution to each other.", + "bbox": [ + 507, + 395, + 882, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3 Training Strategy", + "text_level": 1, + "bbox": [ + 507, + 470, + 695, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After initialization by cross-lingual pretrained language model, we train denoising autoencoder and iterative back-translation alternatively. The final loss of image fusion UNMT is:", + "bbox": [ + 507, + 491, + 882, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {I} = \\mathcal {L} _ {B} ^ {I} + \\lambda_ {D} \\mathcal {L} _ {D} ^ {I} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 621, + 569, + 882, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "specifically, $\\mathcal{L}_B^I$ and $\\mathcal{L}_D^I$ refer to:", + "bbox": [ + 507, + 592, + 754, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {B} ^ {I} = \\sum_ {i = 1} ^ {M} - l o g P _ {L _ {2} \\rightarrow L _ {1}} (X _ {i} | \\mathcal {I} (S _ {L _ {1} \\rightarrow L _ {2}} (\\mathcal {I} (X _ {i}), \\theta)), \\theta) \\\\ + \\sum_ {j = 1} ^ {N} - \\log P _ {L _ {1} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {I} \\left(S _ {L _ {2} \\rightarrow L _ {1}} \\left(\\mathcal {I} \\left(Y _ {j}\\right), \\theta\\right)\\right), \\theta\\right) \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 617, + 882, + 703 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D} ^ {I} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {I} \\left(N \\left(X _ {i}\\right)\\right), \\theta\\right) \\tag {10} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} (Y _ {j} | \\mathcal {I} (N (Y _ {j})), \\theta) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 559, + 725, + 882, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the denoising autoencoder process, we first add noise to the original sentence and then integrate the images. During the iterative backtranslation process, for the first round of synthetic data generation, we do not add image information to achieve a smoother transition because the pretrained language model did not include images.", + "bbox": [ + 507, + 806, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12611", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/900fb12f415c4082a747b2ad331710aef679470e06127327f56b460ceadcf698.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Language#Word#ImageAverage
En4,97545,2289.09
Fr4,64636,9277.95
De10,38485,4198.23
Cs10,70672,9426.81
Zh10,80585,1447.88
Total41,516325,6607.84
", + "bbox": [ + 142, + 80, + 460, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: The statistics of our constructed Word-Image dataset.", + "bbox": [ + 112, + 227, + 485, + 254 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Experiments", + "text_level": 1, + "bbox": [ + 114, + 286, + 260, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide a detailed description of the datasets and experimental setup employed in our study. Subsequently, we perform an in-depth analysis of the results obtained.", + "bbox": [ + 112, + 317, + 487, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1 Datasets", + "text_level": 1, + "bbox": [ + 112, + 400, + 228, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Word-Image Dataset. In our approach, the most crucial component of the experiment data is the image dataset at the word level. Since there is no off-the-shelf word-level image dataset, we build one for performance evaluation. We collect these images through the following steps: word extraction, image crawling, and image filtering. Firstly, extract the nouns from the monolingual dataset using part-of-speech analysis tools including Spacy3 and Stanford CoreNLP4. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images5 to crawl images for these nouns, limiting the image size to between 200-by-200 and 10,000-by-10,000 pixels and retrieving 10 to 15 images per noun. The image format is restricted to JPG, JPEG, and PNG. Finally, filter the crawled images. We consider all the images collected for a single noun as a cluster and encode them using the CLIP model. We then calculate pairwise similarities between all images and remove those with an average similarity below a threshold $\\mathcal{T}$ , which we set to 0.6, as they are considered anomalous. The statistics information of the Word-Image dataset can be seen in Table 1. Our dataset covers a wide range of common nouns and can be reused for various purposes.", + "bbox": [ + 112, + 426, + 489, + 860 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Monolingual Dataset. The monolingual datasets used in our experiments are Multi30k dataset (Elliott et al., 2016, 2017; Barrault et al., 2018), LDC dataset and WMT News crawl dataset. The Multi30k dataset contains images and their captions in four languages: English(En), French(Fr), Germany(De), and Czech(Cs). Every language has 29000 sentences for training and 1014 for validation. It is worth noting that we only utilized the lingual data from Multi30k as the images are sentence-level. To prevent the model from encountering bilingual corpora, we randomly split the training and validation sets of the Multi30k dataset into two equal parts. Each part of the training set contains 14,500 sentences, and each part of the validation set contains 507 sentences. We report results in all test sets in Multi30k including Test2016, Test2017, Test2018, and MsCOCO. For the Cs-En task, only Test2016 and Test2018 are available. Furthermore, we expanded the dataset by incorporating monolingual data from WMT 2008-2022 News crawl and LDC. We also report our results on the common news domain test set, which includes WMT Newstest 2014 for En-Fr, Newstest 2021 for En-Cs, and NIST06 for En-Zh.", + "bbox": [ + 507, + 84, + 885, + 485 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Experimental Details", + "text_level": 1, + "bbox": [ + 507, + 497, + 722, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Models. We compare recent unsupervised text-only and multimodal MT baselines listed in the following: (1) XLM (Conneau and Lample, 2019) uses a masked language model to train a cross-lingual language model to initialize. (2) UMMT (Su et al., 2019) uses visual features for denoising autoencoder and back-translation. (3) PVP (Huang et al., 2020) employs multimodal back-translation and features pseudo visual pivoting. (4) UNMT-CS (Jones et al., 2023) uses the codeswitching method, where words in the source sentence are swapped out for their corresponding translation from bilingual dictionaries, to solve the lexical confusion issue in UNMT. It is worth noting that both UMMT and PVP datasets utilize sentence-level images.", + "bbox": [ + 507, + 517, + 885, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "System Settings. We conducted our experiments based on the XLM codebase6. The model architecture employed is a 6-layers, 8-heads transformer, with a hidden layer representation dimension of 1024. The epoch size is 100K and the tokens-per-batch is 2000. we use Adam optimizer and 2000 warm-up updates. The learning rate is 1e-5. For", + "bbox": [ + 507, + 784, + 885, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://github.com/facebookresearch/XLM", + "bbox": [ + 529, + 904, + 800, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://github.com/explosion/spaCy", + "bbox": [ + 134, + 877, + 361, + 891 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "4https://github.com/stanfordnlp/CoreNLP", + "bbox": [ + 136, + 891, + 391, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "5https://images.google.com", + "bbox": [ + 136, + 904, + 305, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12612", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d6a6ae708a278decf5de5cf56e71852a2703cd247c68a10ef75a7c3b6b8a9efa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
En-FrEn-DeEn-Cs
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT39.80*---23.50*-----
PVP52.30*---33.90*-----
XLM53.0846.8235.3143.0931.8126.8325.6222.9525.1420.15
UNMT-CS54.2446.3036.0643.4933.4327.5126.6724.2726.4622.71
Ours54.0746.9536.2144.4433.0629.1227.3724.7526.3422.88
Δ+0.99+0.13+0.90+1.35+1.25+2.29+1.75+1.80+1.20+2.73
Fr-EnDe-EnCs-En
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT40.50*---26.40*-----
PVP46.00*---31.60*-----
XLM48.0442.1437.0343.5837.5133.2130.7327.9833.1330.29
UNMT-CS49.2842.8637.1944.3938.8334.4831.4428.0733.9131.33
Ours49.1043.0137.4144.5740.3235.2032.7029.8134.2931.56
Δ+1.06+0.87+0.38+0.99+2.81+1.99+1.97+1.83+1.16+1.27
", + "bbox": [ + 200, + 80, + 800, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2: BLEU scores on Multi30k En-Fr, En-De, and En-Cs UNMT tasks. F18, F17, F16, and C17 respectively denote the Flickr2018, Flickr2017, Flickr2016, and COCO2017 test sets in Multi30k. * is the results reported from previous papers. Δ represents the improvement of our method over the XLM method in terms of BLEU.", + "bbox": [ + 112, + 362, + 882, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation, we use 4-gram BLEU (Papineni et al., 2002) scores by multi-bleu.pl in Moses7. We train all models on 4 NVIDIA 32GB V100 GPUs.", + "bbox": [ + 112, + 430, + 487, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training, we first train the models on a 5M WMT News Crawl dataset for 70 epochs and subsequently continued training on the Multi30k dataset for another 70 epochs due to its relatively small size. As for the CLIP model, we use ViT-L/14@336px. When incorporating images, we randomly select two images from the corresponding image set for each word and concatenate them to the text. We only incorporate images during the iterative back-translation and denoising autoencoder stages, so our method can be used in conjunction with other pre-training enhancement methods. Regarding the cross-lingual models used for initialization, we directly employ high-quality open-source cross-lingual models in experiments involving En-Fr, En-De, and En-Zh. However, for the En-Cs, we train a model using WMT monolingual dataset, with 10 million English sentences and 10 million Czech sentences, due to the unavailability of any open-source models.", + "bbox": [ + 115, + 479, + 489, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3 Main Results", + "text_level": 1, + "bbox": [ + 112, + 810, + 265, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the main results on the Multi30k test set, including translating among English, French, German, and Czech. Comparing the results of XLM and our method, we observe that adding im", + "bbox": [ + 112, + 831, + 489, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages leads to performance improvements in all six translation directions, with a maximum increase of $+2.81$ BLEU points, indicating the effectiveness of our approach. UNMT-CS is a method that utilizes bilingual dictionaries to address the problem of lexical confusion via code-switching. By comparing our method with the UNMT-CS approach, we observed comparable results, and in some directions, our method even outperformed the code-switching approach, indicating that ours is more efficient and cost-effective.", + "bbox": [ + 505, + 430, + 884, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When compared to the best-performing method that incorporates sentence-level images, our approach still achieves comparable or better results in most translation directions. This is despite the matching degree between the Multi30k monolingual data and word-level images being slightly lower than that of sentence-level images, as the text data in Multi30k is generated based on the sentence-level images. In some directions, our approach even outperforms the previous best methods by up to $+4.22$ BLEU points (as shown in Table 2, where our method is compared to the first two methods of the table).", + "bbox": [ + 507, + 609, + 884, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Analysis", + "text_level": 1, + "bbox": [ + 507, + 834, + 618, + 850 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Case Study", + "text_level": 1, + "bbox": [ + 507, + 863, + 643, + 879 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, we provide several examples of translations, where we observe that the XLM method re", + "bbox": [ + 507, + 887, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "7https://github.com/moses-smt/mosesdecoder", + "bbox": [ + 134, + 903, + 415, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12613", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f47789201e1f12da963f97196b02a4075cb11c6db020d93e0f702a1ea6bd03dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GT-DEMenschen an der Seitenlinie bei einem Fußballspiel.
GT-ENPeople on the sideline of a soccer match.
XLMPeople at the plate in a baseball game.
OursPeople at the stands in a soccer game.
GT-DEEin kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.
GT-ENA young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.
XLMA small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.
OursA small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.
GT-FRUn homme avec une veste noire, une casquette à carreaux et un pantalon rayé noir et blanc joue de la guitare électrique sur une scène avec un chanteur et un autre guitariste en arrêtère-plan.
GT-ENA man in a black jacket and checkered hat wearing black and white striped pants plays an electric guitar on a stage with a singer and another guitar player in the background.
XLMA man with a black jacket, striped headband and light brown and white shirt plays an electric guitar on a stage with a singer and another guitarist in the background.
OursA man with a black jacket, checkered cap and black and white striped pants plays an electric guitar on a stage with a singer and another guitarist in the background.
", + "bbox": [ + 137, + 83, + 857, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: Qualitative results of the proposed model. GT denotes ground truth. Words highlighted in red are affected by the problem of lexical confusion.", + "bbox": [ + 112, + 288, + 882, + 318 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7040d306fc6f2b72cc36bf158842215c74fd5849719144a37c7ddd37e6629ac2.jpg", + "image_caption": [ + "Figure 3: 1-gram precision scores on Flickr2016 and Flickr2017 data sets." + ], + "image_footnote": [], + "bbox": [ + 117, + 343, + 480, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sults in some lexical confusion errors. For instance, in Example 1, \"Fußballspiel\" is translated to \"baseball\" instead of \"soccer\", and in Example 2, \"orangefarbenen\" is translated to \"orange-red\" instead of \"orange\". The representation of these words is similar, which makes it difficult for the model to differentiate between them. However, when we incorporate image information, the words obtain additional information from the image representation, which helps alleviate this problem to some extent. In Table 3, we can see that these words are correctly translated when we add image information. Therefore, our approach can effectively address the issue of lexical confusion in UNMT.", + "bbox": [ + 112, + 564, + 489, + 788 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 1-gram Precision Scores", + "text_level": 1, + "bbox": [ + 112, + 800, + 351, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better observe how our method can alleviate the issue of lexical confusion, we report the 1-gram precision scores in Figure 3. The accuracy of the model's word translations can be reflected intuitively in the 1-gram precision scores metric. We report the results of our approach with and without", + "bbox": [ + 112, + 822, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "incorporating images for various translation directions on the Flickr2016 and Flickr2017 test sets and find that our method with image incorporation outperforms the method without images in all directions. This demonstrates that our method can indeed provide more accurate word translations.", + "bbox": [ + 507, + 343, + 884, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Performance on WMT and LDC Dataset", + "text_level": 1, + "bbox": [ + 507, + 479, + 875, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Not all sentences have a corresponding image that perfectly matches the text. On the other hand, generally, each sentence has some words that can be matched with images, making the use of word-level images more flexible and less costly to collect. Hence, we also conduct experiments on a more common news dataset. We conduct experiments on En-Fr, En-Cs, and En-Zh translation directions using the WMT 5M monolingual data for En-Fr and En-Cs, and the LDC dataset for En-Zh. We split the LDC bilingual data into two non-overlapping parts, each containing 0.75 million monolingual sentences. The news dataset is more abstract in content compared to the Multi30k dataset, and it is difficult to find an image that perfectly matches a sentence. This makes it challenging or almost impossible to apply previous methods that rely on sentence-level images to this dataset. As shown in Table 4, our method can be applied to the news dataset, and it shows improvements compared to the method that only uses text. The improvement on the WMT dataset was not significant, and we speculate that this may be due to the sufficient size of the WMT dataset and the lower frequency of noun entities with images compared to Multi30k.", + "bbox": [ + 505, + 517, + 882, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12614", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0c3e9a0658dbe8c5909009083b0051e563e27a0be14fdeb736830d74520ca6d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachZh-EnEn-ZhCs-EnEn-CsFr-EnEn-Fr
XLM8.437.9125.1417.7933.7335.77
Our Method9.879.3625.5118.2434.1436.46
Δ+1.44+1.45+0.37+0.45+0.41+0.69
", + "bbox": [ + 119, + 82, + 485, + 143 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4: BLUE scores on WMT and LDC test sets.", + "bbox": [ + 126, + 153, + 472, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/008d6c4d02a5b500a1f50daca9f812157920c343553bc8a301e17dbd4d20fa7a.jpg", + "image_caption": [ + "Figure 4: BLEU scores on En-De with different ways and numbers of adding image." + ], + "image_footnote": [], + "bbox": [ + 119, + 184, + 478, + 341 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 Quantity of Images", + "text_level": 1, + "bbox": [ + 112, + 406, + 312, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We are also curious about how the method of adding images and image quantity affects the performance. Therefore, we test several ways of adding images, including selecting the best one, the top two, randomly selecting one, randomly selecting two, and randomly selecting three. We treat all images corresponding to a single word as a class and selected the image with the highest average similarity to the other images in the same class as the best image. We report the average BLEU score of En-De on the Flickr2016 and Flickr2017 test sets as illustrated in figure 4. We find that the fixed-best method yielded lower performance compared to randomly adding images. This could be due to the limitation of the fixed-best method in capturing the full range of visual variations that may exist for a given word. Among all the random selection methods, we find that adding two images randomly yielded the best performance. This could be because adding three images may introduce more noise, while adding only one image may not provide enough randomness.", + "bbox": [ + 112, + 426, + 489, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Related Work", + "text_level": 1, + "bbox": [ + 112, + 795, + 268, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised Machine Translation. Unlike common machine translation (Sutskever et al., 2014; Vaswani et al., 2017; Feng et al., 2020), UNMT don't use parallel corpus. Recent advancements (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a; Conneau and Lample, 2019;", + "bbox": [ + 112, + 822, + 489, + 919 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Song et al., 2019; Han et al., 2021) in UNMT have achieved remarkable milestones. However, there are still certain issues associated with solely utilizing monolingual data for translation, such as domain mismatch problem (Marchisio et al., 2020; Kim et al., 2020), poor performance on distance language pairs (Kim et al., 2020; Chronopoulou et al., 2020; Sun et al., 2021), translationese (He et al., 2022), and lexical confusion (Bapna et al., 2022; Jones et al., 2023). Our work aims to address the issue of lexical confusion by utilizing word-level image data, thereby eliminating the need for costly bilingual dictionary annotations.", + "bbox": [ + 507, + 84, + 884, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal Machine Translation. Multimodal machine translation involves incorporating image information into a machine translation model to enhance its performance. While most current multimodal machine translation systems (Caglayan et al., 2016; Yao and Wan, 2020; Yin et al., 2020; Caglayan et al., 2021; Zhang et al., 2020; Fang and Feng, 2022; Li et al., 2022) focus on supervised machine translation, it has been observed that visual information is generally ignored when parallel corpora are plentiful (Caglayan et al., 2019; Wu et al., 2021). Yang et al. (2022) proposed the use of image information in situations where parallel corpora are not abundant. On the other hand, Su et al. (2019) and Huang et al. (2020) utilized image information in an unsupervised setting. However, these approaches were based on sentence-level image information, which can be challenging to obtain since not all sentences have corresponding images. In contrast, word-level image information is easier to obtain and is more flexible and lightweight. Therefore, our work aims to use word-level image information to enhance the performance of UNMT.", + "bbox": [ + 507, + 302, + 884, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 507, + 684, + 796, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a method to alleviate the problem of lexical confusion in unsupervised machine translation by utilizing word-level images and provide a concise yet potent framework for integrating them into an unsupervised neural machine translation model. Through extensive experiments, we demonstrate that our method can effectively mitigate the issue of word confusion and even outperform the use of costly bilingual dictionaries in some directions. Furthermore, we release an open-source word-image dataset that covers 5 languages and involves over 300,000 images in total. In the future, we will endeavor to explore the", + "bbox": [ + 507, + 709, + 884, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12615", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "potential of utilizing images for augmenting the translation quality of low-resource languages in a large-language-model-based translation system.", + "bbox": [ + 112, + 84, + 487, + 134 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 143, + 220, + 158 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One limitation of our work is that we have not been able to effectively handle the issue of polysemy. The word-image dataset is obtained through direct web scraping of words, and the images collected may not all correspond to the same sense of the word. Moreover, when adding images, we have not been able to accurately select images that correspond to the intended meaning of the word in the sentence. However, we have attempted to alleviate this issue to some extent by randomly selecting images during the process of adding images, which has enabled the model to encounter the correct images to a certain degree. In the future, we plan to improve our approach by incorporating sentence-level information when inserting images.", + "bbox": [ + 112, + 168, + 489, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 422, + 265, + 436 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In our experimental endeavors, we have amassed a substantial collection of images, which naturally prompts discussions pertaining to these issues. It is imperative to clarify that we do not possess the copyright for these images. Instead, they will be exclusively available to researchers and educators seeking to utilize the dataset for non-commercial research and/or educational purposes.", + "bbox": [ + 112, + 447, + 487, + 575 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 588, + 285, + 604 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Xiaocheng Feng is the corresponding author of this work. We thank the anonymous reviewers for their insightful comments. This work was supported by the National Key RD Program of China via grant No. 2021ZD0112905, National Natural Science Foundation of China (NSFC) via grant 62276078, the Key RD Program of Heilongjiang via grant 2022ZX01A32, the International Cooperation Project of PCLPCL2022D01 and the Fundamental Research Funds for the Central Universities (Grant No.HIT.OCEF.2023018).", + "bbox": [ + 112, + 613, + 489, + 789 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 816, + 213, + 831 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2018. Unsupervised neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.", + "bbox": [ + 115, + 838, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ankur Bapna, Isaac Caswell, Julia Kreutzer, Orhan Firat, Daan van Esch, Aditya Siddhant, Mengmeng Niu, Pallavi Baljekar, Xavier Garcia, Wolfgang Macherey, et al. 2022. Building machine translation systems for the next thousand languages. arXiv preprint arXiv:2205.03983.", + "Loic Barrault, Fethi Bougares, Lucia Specia, Chiraag Lala, Desmond Elliott, and Stella Frank. 2018. Findings of the third shared task on multimodal machine translation. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 304-323.", + "Ozan Caglayan, Loic Barrault, and Fethi Bougares. 2016. Multimodal attention for neural machine translation. arXiv preprint arXiv:1609.03976.", + "Ozan Caglayan, Menekse Kuyu, Mustafa Sercan Amac, Pranava Madhyastha, Erkut Erdem, Aykut Erdem, and Lucia Specia. 2021. Cross-lingual visual pretraining for multimodal machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1317-1324, Online. Association for Computational Linguistics.", + "Ozan Caglayan, Pranava Madhyastha, Lucia Specia, and Loic Barrault. 2019. Probing the need for visual context in multimodal machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4159-4170, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2020. Reusing a Pretrained Language Model on Languages with Limited Corpora for Unsupervised NMT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2703-2711, Online. Association for Computational Linguistics.", + "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32.", + "Desmond Elliott, Stella Frank, Loic Barrault, Fethi Bougares, and Lucia Specia. 2017. Findings of the second shared task on multimodal machine translation and multilingual image description. In Proceedings of the Second Conference on Machine Translation, pages 215-233, Copenhagen, Denmark. Association for Computational Linguistics.", + "Desmond Elliott, Stella Frank, Khalil Sima'an, and Lucia Specia. 2016. Multi30K: Multilingual English-German image descriptions. In Proceedings of the 5th Workshop on Vision and Language, pages 70-74, Berlin, Germany. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12616", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qingkai Fang and Yang Feng. 2022. Neural machine translation with phrase-level universal visual representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5687-5698, Dublin, Ireland. Association for Computational Linguistics.", + "Xiaocheng Feng, Zhangyin Feng, Wanlong Zhao, Bing Qin, and Ting Liu. 2020. Enhanced neural machine translation by joint decoding with word and posttagging sequences. Mobile Networks and Applications, 25:1722-1728.", + "Jesse Michael Han, Igor Babuschkin, Harrison Edwards, Arvind Neelakantan, Tao Xu, Stanislas Polu, Alex Ray, Pranav Shyam, Aditya Ramesh, Alec Radford, et al. 2021. Unsupervised neural machine translation with generative language models only. arXiv preprint arXiv:2110.05448.", + "Zhiwei He, Xing Wang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2022. Bridging the data gap between training and inference for unsupervised neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6611-6623, Dublin, Ireland. Association for Computational Linguistics.", + "Po-Yao Huang, Junjie Hu, Xiaojun Chang, and Alexander Hauptmann. 2020. Unsupervised multimodal neural machine translation with pseudo visual pivoting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8226-8237, Online. Association for Computational Linguistics.", + "Alex Jones, Isaac Caswell, Ishank Saxena, and Orhan Firat. 2023. Bilex rx: Lexical data augmentation for massively multilingual machine translation. CoRR, abs/2303.15265.", + "Yunsu Kim, Miguel Graça, and Hermann Ney. 2020. When and why is unsupervised neural machine translation useless? In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 35-44, Lisboa, Portugal. European Association for Machine Translation.", + "Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018a. Unsupervised machine translation using monolingual corpora only. In International Conference on Learning Representations.", + "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018b. Word translation without parallel data. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.", + "Guillaume Lample, Myle Ott, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018c. Phrase-based & neural unsupervised machine translation." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5039-5049, Brussels, Belgium. Association for Computational Linguistics.", + "Bei Li, Chuanhao Lv, Zefan Zhou, Tao Zhou, Tong Xiao, Anxiang Ma, and JingBo Zhu. 2022. On vision features in multimodal machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6327-6337, Dublin, Ireland. Association for Computational Linguistics.", + "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. 2020. K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2901-2908.", + "Kelly Marchisio, Kevin Duh, and Philipp Koehn. 2020. When does unsupervised machine translation work? In Proceedings of the Fifth Conference on Machine Translation, pages 571-583, Online. Association for Computational Linguistics.", + "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics.", + "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR.", + "Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. Mass: Masked sequence to sequence pretraining for language generation. In International Conference on Machine Learning, pages 5926-5936. PMLR.", + "Yuanhang Su, Kai Fan, Nguyen Bach, C-C Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10482-10491.", + "Haipeng Sun, Rui Wang, Masao Utiyama, Benjamin Marie, Kehai Chen, Eiichiro Sumita, and Tiejun Zhao. 2021. Unsupervised neural machine translation for similar and distant language pairs: An empirical study. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 20(1):1-17.", + "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12617", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.", + "Zhiyong Wu, Lingpeng Kong, Wei Bi, Xiang Li, and Ben Kao. 2021. Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pages 6153-6166, Online. Association for Computational Linguistics.", + "Zhe Yang, Qingkai Fang, and Yang Feng. 2022. Lowresource neural machine translation with cross-modal alignment. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10134-10146, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Shaowei Yao and Xiaojun Wan. 2020. Multimodal transformer for multimodal machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4346-4350, Online. Association for Computational Linguistics.", + "Yongjing Yin, Fandong Meng, Jinsong Su, Chulun Zhou, Zhengyuan Yang, Jie Zhou, and Jiebo Luo. 2020. A novel graph-based multi-modal fusion encoder for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3025-3035, Online. Association for Computational Linguistics.", + "Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Adversarial training for unsupervised bilingual lexicon induction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1959-1970, Vancouver, Canada. Association for Computational Linguistics.", + "Zhuosheng Zhang, Kehai Chen, Rui Wang, Masao Utiyama, Eiichiro Sumita, Zuchao Li, and Hai Zhao. 2020. Neural machine translation with universal visual representation. In International Conference on Learning Representations." + ], + "bbox": [ + 115, + 85, + 487, + 738 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A The Implementation Details of the XLM and UNMT-CS Method", + "text_level": 1, + "bbox": [ + 114, + 764, + 448, + 795 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The XLM and UNMT-CS methods are re-implemented on the Multi30k dataset in our paper. During training, both these two methods are trained on the 5 million sentences from the WMT dataset for 70 epochs and subsequently continued training for 70 epochs on the Multi30k dataset. For the bilingual dictionary in codeswitching method, we use", + "bbox": [ + 112, + 806, + 489, + 917 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "the Ground-truth bilingual dictionaries provided in MUSE $^{8}$ (Lample et al., 2018b). When training, we substituted words in $50\\%$ of the sentences, while the remaining sentences were left unchanged. All other parameters used in the XLM and UNMT-CS methods are identical to those used in our proposed method, as described in our paper.", + "bbox": [ + 507, + 84, + 882, + 197 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "B The Implicit Mapping Relationship Between Image Representations", + "text_level": 1, + "bbox": [ + 507, + 208, + 852, + 242 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In fact, our approach doesn't explicitly provide the model with a mapping relationship. Rather, we introduce the (word-image) pairs relationship into the model, enabling it to extract information from similar images and thus enhancing the mapping relationship between words.", + "bbox": [ + 507, + 250, + 882, + 346 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We are also curious about the impact on results if the implicit mapping relationship between image representations was absent. Therefore, during the second stage of multi30k training, we replace all images from one language with randomly generated representations while keeping the other language's images unchanged. The En→De direction results for the Flickr2016 and Flickr2017 datasets experience a significant decline, decreasing from 40.32/35.20 to 39.17/34.29. This confirms the crucial importance of the mapping relationship between images for our approach.", + "bbox": [ + 507, + 347, + 884, + 541 + ], + "page_idx": 10 + }, + { + "type": "footer", + "text": "$^{8}$ https://github.com/facebookresearch/MUSE", + "bbox": [ + 531, + 903, + 806, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12618", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_model.json b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..39625d25d8cdbddbf64154337b7cc11ecb05550b --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_model.json @@ -0,0 +1,2059 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.143, + 0.082, + 0.856, + 0.123 + ], + "angle": 0, + "content": "Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.13, + 0.817, + 0.165 + ], + "angle": 0, + "content": "Chengpeng Fu\\(^{1,2}\\), Xiaocheng Feng\\(^{1,2}\\), Yichong Huang\\(^{1}\\), Wenshuai Huo\\(^{1,2}\\), Hui Wang\\(^{2}\\), Bin Qin\\(^{1,2}\\), Ting Liu\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.166, + 0.7, + 0.182 + ], + "angle": 0, + "content": "1Harbin Institution of Technology, Harbin, China" + }, + { + "type": "text", + "bbox": [ + 0.332, + 0.182, + 0.671, + 0.198 + ], + "angle": 0, + "content": "\\(^{2}\\)Pengcheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.199, + 0.745, + 0.214 + ], + "angle": 0, + "content": "fuchp@stu.hit.edu.cn, {huowsh, wangh06}@pc1.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.216, + 0.705, + 0.231 + ], + "angle": 0, + "content": "{xcfeng,ychuang,qinb,tliu}@ir.hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.28, + 0.461, + 0.606 + ], + "angle": 0, + "content": "Unsupervised neural machine translation has recently made remarkable strides, achieving impressive results with the exclusive use of monolingual corpora. Nonetheless, these methods still exhibit fundamental flaws, such as confusing similar words. A straightforward remedy to rectify this drawback is to employ bilingual dictionaries, however, high-quality bilingual dictionaries can be costly to obtain. To overcome this limitation, we propose a method that incorporates images at the word level to augment the lexical mappings. Specifically, our method inserts visual representations into the model, modifying the corresponding embedding layer information. Besides, a visible matrix is adopted to isolate the impact of images on other unrelated words. Experiments on the Multi30k dataset with over 300,000 self-collected images validate the effectiveness in generating more accurate word translation, achieving an improvement of up to \\(+2.81\\) BLEU score, which is comparable or even superior to using bilingual dictionaries.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.619, + 0.26, + 0.634 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.644, + 0.489, + 0.836 + ], + "angle": 0, + "content": "Unsupervised Neural Machine Translation (UNMT) (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a,b,c) seeks to avoid the necessity for high-quality bilingual data in conventional machine translation by exploring the possibility of completing translations relying solely on the monolingual corpus. Consisting of three main components, cross-lingual language model initialization, denoising auto-encoder, and iterative back translation, recent methods (Conneau and Lample, 2019; Song et al., 2019) have achieved promising results." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.837, + 0.49, + 0.886 + ], + "angle": 0, + "content": "Nonetheless, recent UNMT approaches have been observed to make particular errors in comparison to supervised machine translation, such" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.886, + 0.574 + ], + "angle": 0, + "content": "as confusing nouns that pertain to the same semantic category. For instance, Bapna et al. (2022) demonstrates that their models translate \"lion\" to \"rabbit\" or \"snake\" despite getting good CHRF scores. In the context of supervised machine translation, pinpointing accurate translations of words is straightforward with aligned bilingual data. However, with only monolingual data, the unsupervised model must rely on large amounts of context for word mapping, which can lead to confusion between words with similar distributions, particularly when there is insufficient data. To address this issue, a direct solution is to integrate high-quality bilingual dictionaries into the UNMT process using codeswitching or spliced methods (Jones et al., 2023). High-quality bilingual dictionaries can be obtained through manual annotation and bilingual data extraction, yet these methods either involve high costs or deviate from the original objective of the UNMT task." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.579, + 0.885, + 0.868 + ], + "angle": 0, + "content": "In addition to utilizing explicit bilingual dictionaries for word mapping, implicit methods can also be employed. As the saying goes, \"A picture is worth a thousand words\". We can leverage the \"thousand words\" represented by corresponding images to learn the mapping between words in different languages. Visual content has the potential to enhance word mapping abilities in latent spaces, as the physical visual perception is congruent among speakers of diverse languages (Huang et al., 2020). Furthermore, the cost of annotating a bilingual dictionary is relatively high, whereas the cost of searching images with monolingual keywords is relatively low in the era of sophisticated search engines. Describing an image is also less demanding for humans compared to translating a word, as the former only requires mastery of one language while the latter requires at least two." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.872, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Drawing on these insights, we propose to integrate word-level image information into the UNMT process to mitigate the issue of lexical confusion." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.893, + 0.488, + 0.919 + ], + "angle": 0, + "content": "1The word-image dataset and associated codes are publicly available at: https://github.com/Brianmax99/UNMT-WVR" + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12608" + }, + { + "type": "footer", + "bbox": [ + 0.211, + 0.946, + 0.788, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12608-12618 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.39 + ], + "angle": 0, + "content": "It's worth noting that we don't use sentence-level images like some multi-modal UNMT approaches (Su et al. (2019); Huang et al. (2020)). Instead, we focused on word-level images since they are more effective and flexible in addressing the challenge of lexical confusion and are easier to collect. Our approach involves augmenting the encoded images to their corresponding words, modifying the corresponding position and language encoding information, and utilizing a visible matrix to isolate the impact of images on other words. We conduct experiments utilizing over 300,000 self-collected images, and further analysis demonstrates that we have effectively mitigated lexical confusion to a certain extent. Furthermore, we achieve new state-of-the-art results for the UNMT task in certain language directions on the Multi30k dataset. The main contributions of this work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.407, + 0.489, + 0.472 + ], + "angle": 0, + "content": "- We introduce a concise yet potent framework that integrates word-level images to alleviate the similar lexical confusion issue in UNMT, without utilizing bilingual dictionaries." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.479, + 0.488, + 0.528 + ], + "angle": 0, + "content": "- Our approach has achieved promising or even unparalleled unsupervised results on the Multi30K dataset for certain language pairs." + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.534, + 0.488, + 0.583 + ], + "angle": 0, + "content": "- We release the word-image datasets for five languages as open-source, which can serve as a valuable resource to the community." + }, + { + "type": "list", + "bbox": [ + 0.137, + 0.407, + 0.489, + 0.583 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.598, + 0.226, + 0.613 + ], + "angle": 0, + "content": "2 Methods" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.626, + 0.49, + 0.707 + ], + "angle": 0, + "content": "In this section, we first review the text-only UNMT process. Then, we present our proposed method for incorporating images, which involves the modified embedding layer and mask transformer. Finally, we introduce some training strategies." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.72, + 0.488, + 0.737 + ], + "angle": 0, + "content": "2.1 Unsupervised Neural Machine Translation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.491, + 0.92 + ], + "angle": 0, + "content": "Successful UNMT systems typically employ an encoder-decoder architecture similar to that of supervised machine translation, sharing parameters between the source-to-target and target-to-source models. They share several common principles: initialization, iterative back-translation, and denoising autoencoder. Let \\( X = \\{X_{i}\\}_{i=1}^{M} \\) denote monolingual data in language \\( L_{1} \\) and \\( Y = \\{Y_{i}\\}_{i=1}^{N} \\) in language \\( L_{2} \\). \\( M \\) and \\( N \\) are the sentence number of monolingual data \\( X \\) and \\( Y \\) respectively. \\( \\theta \\) represents the UNMT model parameters." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.246 + ], + "angle": 0, + "content": "Initialization. Performing parameter initialization that is non-random can introduce cross-lingual prior knowledge into the model. There are currently two main initialization methods for UNMT models. The first method initializes the embedding layer of a UNMT model with word pretrained embeddings, while the second method uses a pretrained cross-lingual language model with the same structure as the UNMT encoder to initialize most of the neural network parameters." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.43 + ], + "angle": 0, + "content": "Iterative Back-translation. Back-translation is a method that leverages monolingual data in the target language and a reverse translation model to generate pseudo-source data, improving the quality of translations in the forward direction. In UNMT, where only monolingual data is available, back-translation is used to generate pseudo-parallel data, transforming the unsupervised problem into a weakly or self-supervised one. During training, back-translation is repeatedly applied in an iterative manner. The loss is defined as followed:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.446, + 0.884, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {B} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {2} \\rightarrow L _ {1}} \\left(X _ {i} \\mid S _ {L _ {1} \\rightarrow L _ {2}} \\left(X _ {i}, \\theta\\right), \\theta\\right) \\tag {1} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {1} \\rightarrow L _ {2}} (Y _ {j} | S _ {L _ {2} \\rightarrow L _ {1}} (Y _ {j}, \\theta), \\theta) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.526, + 0.882, + 0.575 + ], + "angle": 0, + "content": "where \\(P_{L_a\\to L_b}\\) denotes the translation probabilities between two languages and \\(S_{L_a\\rightarrow L_b}\\) denotes the translation processes from language \\(L_{a}\\) to \\(L_{b}\\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.582, + 0.884, + 0.695 + ], + "angle": 0, + "content": "Denoising Autoencoder. The introduction of the denoising autoencoder aims to guide the model in producing fluent text output. This is achieved by adding noise to sentences through techniques such as replacement, deletion, and shuffling, and then allowing the model to output the correct and fluent original sentence. The loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.71, + 0.884, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {N} \\left(X _ {i}\\right), \\theta\\right) \\tag {2} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {N} (Y _ {j}), \\theta\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.881, + 0.823 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(\\cdot)\\) refers to the noise functions. \\(P_{L_a\\to L_a}\\) denotes the reconstruction probabilities in \\(L_{a}\\)." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.824, + 0.836, + 0.838 + ], + "angle": 0, + "content": "The final objective function of UNMT is:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.847, + 0.883, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {B} + \\lambda_ {D} \\mathcal {L} _ {D} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.872, + 0.885, + 0.919 + ], + "angle": 0, + "content": "where \\(\\lambda_{D}\\) is a hyper-parameter that weights denoise autoencoder loss term and is gradually reduced during training." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12609" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.085, + 0.861, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.391, + 0.829, + 0.407 + ], + "angle": 0, + "content": "Figure 1: The process of converting an image-inserted sentence into an embedding representation." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.431, + 0.269, + 0.447 + ], + "angle": 0, + "content": "2.2 Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.453, + 0.489, + 0.644 + ], + "angle": 0, + "content": "In this section, we present how to incorporate word-level images into the UNMT process to tackle the issue of lexical confusion. The main concept behind our approach is that words can acquire information not only from contextual information but also from their corresponding images. Let \\( X_{i} = \\{x_{i}^{1}, x_{i}^{2}, \\dots, x_{i}^{K}\\} \\) denote a sentence from \\( X \\) in language \\( L_{1} \\) and \\( x_{i}^{K} \\) denotes the \\( k \\)-th word in sentence \\( X_{i} \\). We denote a lookup table consisting of word-image pairs as \\( \\mathcal{D}_{L_i} = \\{(word, I_{word})\\} \\), where \\( I_{word} \\) is an image set with one or multiple images that depict the meaning of the word." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.646, + 0.49, + 0.775 + ], + "angle": 0, + "content": "As shown in Figure \\(1^{2}\\), we segment an input sentence \\(X_{i}\\) at the word level and obtain the images by retrieving the words appearing in the lookup table \\(\\mathcal{D}_{L_i}\\). Then, we concatenate the retrieved images to the corresponding words, resulting in an Image-Inserted Sentence denoted \\(IIS_{i}\\). Each word can be concatenated with one or multiple images. Formally, We denote \\(IIS_{i}\\) as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.785, + 0.489, + 0.81 + ], + "angle": 0, + "content": "\\[\nI I S _ {i} = \\left\\{x _ {i} ^ {1}, \\widehat {I} _ {x _ {i} ^ {1}}, x _ {i} ^ {2}, \\widehat {I} _ {x _ {i} ^ {2}}, \\dots , x _ {i} ^ {K}, \\widehat {I} _ {x _ {i} ^ {K}} \\right\\} \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.821, + 0.489, + 0.872 + ], + "angle": 0, + "content": "where \\(\\widehat{I}_{x_i^K}\\) is the image subset of \\(I_{x_i^K}\\) (\\(\\widehat{I}_{x_i^K}\\) can be an empty set). We define this process of injecting images as:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.489, + 0.919 + ], + "angle": 0, + "content": "2\"Blue\" is an adjective in the figure. In our image collection process, \"blue\" can also appear as a noun in other training samples." + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.449, + 0.883, + 0.466 + ], + "angle": 0, + "content": "\\[\nI I S = \\mathcal {I} (X) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.473, + 0.704, + 0.492 + ], + "angle": 0, + "content": "where \\( IIS = \\{IIS_{i}\\}_{i = 1}^{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.492, + 0.884, + 0.555 + ], + "angle": 0, + "content": "After obtaining the \\( IIS \\), we need to integrate them into the UNMT model. To achieve this integration, we employ a modified embedding layer and a mask transformer." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.567, + 0.788, + 0.583 + ], + "angle": 0, + "content": "2.2.1 Modified Embedding Layer" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.587, + 0.884, + 0.652 + ], + "angle": 0, + "content": "The first modification part is the embedding layer, which generally includes token embedding, position embedding, and, in some models, language embedding. An example can be seen in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Token Embedding. In the context of \\( IIS \\), both textual words and images are present. Given that most current machine translation methods operate at the subword level, we proceed to tokenize the \\( IIS \\), focusing solely on the textual component of \\( IIS \\). For encoding textual tokens, we continue to utilize the embedding layer from the original text-only model. Regarding images, we leverage the CLIP (Radford et al., 2021) model as the image encoder to extract visual features. The CLIP model is a multi-modal model trained on a variety of (image, text) pairs, which possesses a powerful capability to represent images. To prevent negative impacts on the encoding of images during the UNMT process, we freeze the visual encoder. However, directly using these representations may lead" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12610" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.489, + 0.135 + ], + "angle": 0, + "content": "to a gap between the image encoding space and the semantic space in UNMT. Therefore, we employ an MLP layer to bridge this gap." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.141, + 0.489, + 0.287 + ], + "angle": 0, + "content": "Position Embedding. Positional encoding is a crucial module that supplements token structural information. For textual information in \\( IIS \\), we continue to encode the position in the original order of the input sentence to prevent interference with sentence structure learning when integrating images into the model. For images, we assign the positional encoding of the first subword in the tokenized result for the corresponding word." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.295, + 0.49, + 0.408 + ], + "angle": 0, + "content": "Language Embedding. Some UNMT models incorporate language embedding to inform the model of the input language and the desired output language. For textual information, we continue to utilize the original language encoding. However, for image information, as it does not belong to any language, we introduce a new image identifier." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.415, + 0.325, + 0.43 + ], + "angle": 0, + "content": "2.2.2 Mask Transformer" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.434, + 0.489, + 0.642 + ], + "angle": 0, + "content": "Despite modifying the original embedding layer to allow \\( IIS \\) to be properly fed into the model, two issues still exist. Firstly, inserting images may affect other irrelevant words in the original sentence and potentially disrupt the structural relationships between them. Secondly, a single word-level image may correspond to multiple subwords, and during the annotation of position embedding, only the information of the first subword is utilized, resulting in a lack of relationship between the image and the remaining subwords it corresponds to. To address these issues, we introduce the method of mask transformer (Liu et al., 2020)." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.643, + 0.49, + 0.723 + ], + "angle": 0, + "content": "We create a visible matrix to shield the relationship between images and other textual information in the sentence and highlight the relationship between the image and its corresponding subwords. The visible matrix \\(M\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.728, + 0.488, + 0.769 + ], + "angle": 0, + "content": "\\[\nM _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & w _ {p} \\bullet w _ {q} \\\\ 0 & w _ {p} \\circ w _ {q}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.775, + 0.489, + 0.84 + ], + "angle": 0, + "content": "where \\(w_{p}\\) and \\(w_{q}\\) are elements in \\(IIS_{i}\\), \\(w_{p} \\circ w_{q}\\) denotes that \\(w_{p}\\) and \\(w_{q}\\) are visible to each other, while \\(w_{p} \\bullet w_{q}\\) is not. An example of a visible matrix is depicted in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.855, + 0.49, + 0.918 + ], + "angle": 0, + "content": "The visible matrix can be added to the self-attention module to mask certain attention calculations. Specifically, the new masked self-attention is:" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.083, + 0.808, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.25, + 0.886, + 0.321 + ], + "angle": 0, + "content": "Figure 2: An example of the visible matrix corresponding to the \\( IIS \\) in Figure 1. The red numerical labels indicate the absolute position of the images in the \\( IIS \\). In this matrix, black represents \"visible\" and white represents \"invisible\"." + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.355, + 0.884, + 0.394 + ], + "angle": 0, + "content": "\\[\nS ^ {i} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {i} K ^ {i T} + M}{\\sqrt {d}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.397, + 0.884, + 0.461 + ], + "angle": 0, + "content": "where \\(Q\\) and \\(K\\) are the query matrix and key matrix in the self-attention module. if two elements are invisible, the \\(M\\) will mask the attention score to 0, which means making no contribution to each other." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.471, + 0.696, + 0.487 + ], + "angle": 0, + "content": "2.3 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.492, + 0.884, + 0.556 + ], + "angle": 0, + "content": "After initialization by cross-lingual pretrained language model, we train denoising autoencoder and iterative back-translation alternatively. The final loss of image fusion UNMT is:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.57, + 0.883, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {I} = \\mathcal {L} _ {B} ^ {I} + \\lambda_ {D} \\mathcal {L} _ {D} ^ {I} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.593, + 0.756, + 0.611 + ], + "angle": 0, + "content": "specifically, \\(\\mathcal{L}_B^I\\) and \\(\\mathcal{L}_D^I\\) refer to:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.618, + 0.883, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {B} ^ {I} = \\sum_ {i = 1} ^ {M} - l o g P _ {L _ {2} \\rightarrow L _ {1}} (X _ {i} | \\mathcal {I} (S _ {L _ {1} \\rightarrow L _ {2}} (\\mathcal {I} (X _ {i}), \\theta)), \\theta) \\\\ + \\sum_ {j = 1} ^ {N} - \\log P _ {L _ {1} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {I} \\left(S _ {L _ {2} \\rightarrow L _ {1}} \\left(\\mathcal {I} \\left(Y _ {j}\\right), \\theta\\right)\\right), \\theta\\right) \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.726, + 0.884, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D} ^ {I} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {I} \\left(N \\left(X _ {i}\\right)\\right), \\theta\\right) \\tag {10} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} (Y _ {j} | \\mathcal {I} (N (Y _ {j})), \\theta) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.884, + 0.919 + ], + "angle": 0, + "content": "During the denoising autoencoder process, we first add noise to the original sentence and then integrate the images. During the iterative backtranslation process, for the first round of synthetic data generation, we do not add image information to achieve a smoother transition because the pretrained language model did not include images." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "12611" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.082, + 0.461, + 0.219 + ], + "angle": 0, + "content": "
Language#Word#ImageAverage
En4,97545,2289.09
Fr4,64636,9277.95
De10,38485,4198.23
Cs10,70672,9426.81
Zh10,80585,1447.88
Total41,516325,6607.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.228, + 0.487, + 0.255 + ], + "angle": 0, + "content": "Table 1: The statistics of our constructed Word-Image dataset." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.287, + 0.262, + 0.304 + ], + "angle": 0, + "content": "3 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.318, + 0.489, + 0.382 + ], + "angle": 0, + "content": "In this section, we provide a detailed description of the datasets and experimental setup employed in our study. Subsequently, we perform an in-depth analysis of the results obtained." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.401, + 0.229, + 0.416 + ], + "angle": 0, + "content": "3.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.427, + 0.49, + 0.862 + ], + "angle": 0, + "content": "Word-Image Dataset. In our approach, the most crucial component of the experiment data is the image dataset at the word level. Since there is no off-the-shelf word-level image dataset, we build one for performance evaluation. We collect these images through the following steps: word extraction, image crawling, and image filtering. Firstly, extract the nouns from the monolingual dataset using part-of-speech analysis tools including Spacy3 and Stanford CoreNLP4. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images5 to crawl images for these nouns, limiting the image size to between 200-by-200 and 10,000-by-10,000 pixels and retrieving 10 to 15 images per noun. The image format is restricted to JPG, JPEG, and PNG. Finally, filter the crawled images. We consider all the images collected for a single noun as a cluster and encode them using the CLIP model. We then calculate pairwise similarities between all images and remove those with an average similarity below a threshold \\(\\mathcal{T}\\), which we set to 0.6, as they are considered anomalous. The statistics information of the Word-Image dataset can be seen in Table 1. Our dataset covers a wide range of common nouns and can be reused for various purposes." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.886, + 0.486 + ], + "angle": 0, + "content": "Monolingual Dataset. The monolingual datasets used in our experiments are Multi30k dataset (Elliott et al., 2016, 2017; Barrault et al., 2018), LDC dataset and WMT News crawl dataset. The Multi30k dataset contains images and their captions in four languages: English(En), French(Fr), Germany(De), and Czech(Cs). Every language has 29000 sentences for training and 1014 for validation. It is worth noting that we only utilized the lingual data from Multi30k as the images are sentence-level. To prevent the model from encountering bilingual corpora, we randomly split the training and validation sets of the Multi30k dataset into two equal parts. Each part of the training set contains 14,500 sentences, and each part of the validation set contains 507 sentences. We report results in all test sets in Multi30k including Test2016, Test2017, Test2018, and MsCOCO. For the Cs-En task, only Test2016 and Test2018 are available. Furthermore, we expanded the dataset by incorporating monolingual data from WMT 2008-2022 News crawl and LDC. We also report our results on the common news domain test set, which includes WMT Newstest 2014 for En-Fr, Newstest 2021 for En-Cs, and NIST06 for En-Zh." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.498, + 0.724, + 0.513 + ], + "angle": 0, + "content": "3.2 Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.518, + 0.886, + 0.776 + ], + "angle": 0, + "content": "Baseline Models. We compare recent unsupervised text-only and multimodal MT baselines listed in the following: (1) XLM (Conneau and Lample, 2019) uses a masked language model to train a cross-lingual language model to initialize. (2) UMMT (Su et al., 2019) uses visual features for denoising autoencoder and back-translation. (3) PVP (Huang et al., 2020) employs multimodal back-translation and features pseudo visual pivoting. (4) UNMT-CS (Jones et al., 2023) uses the codeswitching method, where words in the source sentence are swapped out for their corresponding translation from bilingual dictionaries, to solve the lexical confusion issue in UNMT. It is worth noting that both UMMT and PVP datasets utilize sentence-level images." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.785, + 0.886, + 0.898 + ], + "angle": 0, + "content": "System Settings. We conducted our experiments based on the XLM codebase6. The model architecture employed is a 6-layers, 8-heads transformer, with a hidden layer representation dimension of 1024. The epoch size is 100K and the tokens-per-batch is 2000. we use Adam optimizer and 2000 warm-up updates. The learning rate is 1e-5. For" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.905, + 0.8, + 0.919 + ], + "angle": 0, + "content": "\\(^{6}\\)https://github.com/facebookresearch/XLM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.878, + 0.363, + 0.892 + ], + "angle": 0, + "content": "3https://github.com/explosion/spaCy" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.892, + 0.392, + 0.905 + ], + "angle": 0, + "content": "4https://github.com/stanfordnlp/CoreNLP" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.306, + 0.919 + ], + "angle": 0, + "content": "5https://images.google.com" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.878, + 0.392, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12612" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.2, + 0.082, + 0.8, + 0.354 + ], + "angle": 0, + "content": "
En-FrEn-DeEn-Cs
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT39.80*---23.50*-----
PVP52.30*---33.90*-----
XLM53.0846.8235.3143.0931.8126.8325.6222.9525.1420.15
UNMT-CS54.2446.3036.0643.4933.4327.5126.6724.2726.4622.71
Ours54.0746.9536.2144.4433.0629.1227.3724.7526.3422.88
Δ+0.99+0.13+0.90+1.35+1.25+2.29+1.75+1.80+1.20+2.73
Fr-EnDe-EnCs-En
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT40.50*---26.40*-----
PVP46.00*---31.60*-----
XLM48.0442.1437.0343.5837.5133.2130.7327.9833.1330.29
UNMT-CS49.2842.8637.1944.3938.8334.4831.4428.0733.9131.33
Ours49.1043.0137.4144.5740.3235.2032.7029.8134.2931.56
Δ+1.06+0.87+0.38+0.99+2.81+1.99+1.97+1.83+1.16+1.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.363, + 0.883, + 0.406 + ], + "angle": 0, + "content": "Table 2: BLEU scores on Multi30k En-Fr, En-De, and En-Cs UNMT tasks. F18, F17, F16, and C17 respectively denote the Flickr2018, Flickr2017, Flickr2016, and COCO2017 test sets in Multi30k. * is the results reported from previous papers. Δ represents the improvement of our method over the XLM method in terms of BLEU." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.431, + 0.489, + 0.478 + ], + "angle": 0, + "content": "evaluation, we use 4-gram BLEU (Papineni et al., 2002) scores by multi-bleu.pl in Moses7. We train all models on 4 NVIDIA 32GB V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.48, + 0.49, + 0.801 + ], + "angle": 0, + "content": "During training, we first train the models on a 5M WMT News Crawl dataset for 70 epochs and subsequently continued training on the Multi30k dataset for another 70 epochs due to its relatively small size. As for the CLIP model, we use ViT-L/14@336px. When incorporating images, we randomly select two images from the corresponding image set for each word and concatenate them to the text. We only incorporate images during the iterative back-translation and denoising autoencoder stages, so our method can be used in conjunction with other pre-training enhancement methods. Regarding the cross-lingual models used for initialization, we directly employ high-quality open-source cross-lingual models in experiments involving En-Fr, En-De, and En-Zh. However, for the En-Cs, we train a model using WMT monolingual dataset, with 10 million English sentences and 10 million Czech sentences, due to the unavailability of any open-source models." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.812, + 0.266, + 0.826 + ], + "angle": 0, + "content": "3.3 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.832, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Table 2 shows the main results on the Multi30k test set, including translating among English, French, German, and Czech. Comparing the results of XLM and our method, we observe that adding im" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.431, + 0.885, + 0.607 + ], + "angle": 0, + "content": "ages leads to performance improvements in all six translation directions, with a maximum increase of \\(+2.81\\) BLEU points, indicating the effectiveness of our approach. UNMT-CS is a method that utilizes bilingual dictionaries to address the problem of lexical confusion via code-switching. By comparing our method with the UNMT-CS approach, we observed comparable results, and in some directions, our method even outperformed the code-switching approach, indicating that ours is more efficient and cost-effective." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.61, + 0.885, + 0.818 + ], + "angle": 0, + "content": "When compared to the best-performing method that incorporates sentence-level images, our approach still achieves comparable or better results in most translation directions. This is despite the matching degree between the Multi30k monolingual data and word-level images being slightly lower than that of sentence-level images, as the text data in Multi30k is generated based on the sentence-level images. In some directions, our approach even outperforms the previous best methods by up to \\(+4.22\\) BLEU points (as shown in Table 2, where our method is compared to the first two methods of the table)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.835, + 0.62, + 0.851 + ], + "angle": 0, + "content": "4 Analysis" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.864, + 0.645, + 0.88 + ], + "angle": 0, + "content": "4.1 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.885, + 0.918 + ], + "angle": 0, + "content": "In Table 3, we provide several examples of translations, where we observe that the XLM method re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.416, + 0.919 + ], + "angle": 0, + "content": "7https://github.com/moses-smt/mosesdecoder" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12613" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.139, + 0.084, + 0.858, + 0.281 + ], + "angle": 0, + "content": "
GT-DEMenschen an der Seitenlinie bei einem Fußballspiel.
GT-ENPeople on the sideline of a soccer match.
XLMPeople at the plate in a baseball game.
OursPeople at the stands in a soccer game.
GT-DEEin kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.
GT-ENA young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.
XLMA small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.
OursA small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.
GT-FRUn homme avec une veste noire, une casquette à carreaux et un pantalon rayé noir et blanc joue de la guitare électrique sur une scène avec un chanteur et un autre guitariste en arrêtère-plan.
GT-ENA man in a black jacket and checkered hat wearing black and white striped pants plays an electric guitar on a stage with a singer and another guitar player in the background.
XLMA man with a black jacket, striped headband and light brown and white shirt plays an electric guitar on a stage with a singer and another guitarist in the background.
OursA man with a black jacket, checkered cap and black and white striped pants plays an electric guitar on a stage with a singer and another guitarist in the background.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.29, + 0.884, + 0.319 + ], + "angle": 0, + "content": "Table 3: Qualitative results of the proposed model. GT denotes ground truth. Words highlighted in red are affected by the problem of lexical confusion." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.344, + 0.481, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.509, + 0.487, + 0.537 + ], + "angle": 0, + "content": "Figure 3: 1-gram precision scores on Flickr2016 and Flickr2017 data sets." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.565, + 0.49, + 0.789 + ], + "angle": 0, + "content": "sults in some lexical confusion errors. For instance, in Example 1, \"Fußballspiel\" is translated to \"baseball\" instead of \"soccer\", and in Example 2, \"orangefarbenen\" is translated to \"orange-red\" instead of \"orange\". The representation of these words is similar, which makes it difficult for the model to differentiate between them. However, when we incorporate image information, the words obtain additional information from the image representation, which helps alleviate this problem to some extent. In Table 3, we can see that these words are correctly translated when we add image information. Therefore, our approach can effectively address the issue of lexical confusion in UNMT." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.801, + 0.352, + 0.817 + ], + "angle": 0, + "content": "4.2 1-gram Precision Scores" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.489, + 0.919 + ], + "angle": 0, + "content": "To better observe how our method can alleviate the issue of lexical confusion, we report the 1-gram precision scores in Figure 3. The accuracy of the model's word translations can be reflected intuitively in the 1-gram precision scores metric. We report the results of our approach with and without" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.344, + 0.885, + 0.44 + ], + "angle": 0, + "content": "incorporating images for various translation directions on the Flickr2016 and Flickr2017 test sets and find that our method with image incorporation outperforms the method without images in all directions. This demonstrates that our method can indeed provide more accurate word translations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.48, + 0.877, + 0.495 + ], + "angle": 0, + "content": "4.3 Performance on WMT and LDC Dataset" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.518, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Not all sentences have a corresponding image that perfectly matches the text. On the other hand, generally, each sentence has some words that can be matched with images, making the use of word-level images more flexible and less costly to collect. Hence, we also conduct experiments on a more common news dataset. We conduct experiments on En-Fr, En-Cs, and En-Zh translation directions using the WMT 5M monolingual data for En-Fr and En-Cs, and the LDC dataset for En-Zh. We split the LDC bilingual data into two non-overlapping parts, each containing 0.75 million monolingual sentences. The news dataset is more abstract in content compared to the Multi30k dataset, and it is difficult to find an image that perfectly matches a sentence. This makes it challenging or almost impossible to apply previous methods that rely on sentence-level images to this dataset. As shown in Table 4, our method can be applied to the news dataset, and it shows improvements compared to the method that only uses text. The improvement on the WMT dataset was not significant, and we speculate that this may be due to the sufficient size of the WMT dataset and the lower frequency of noun entities with images compared to Multi30k." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12614" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.083, + 0.486, + 0.145 + ], + "angle": 0, + "content": "
ApproachZh-EnEn-ZhCs-EnEn-CsFr-EnEn-Fr
XLM8.437.9125.1417.7933.7335.77
Our Method9.879.3625.5118.2434.1436.46
Δ+1.44+1.45+0.37+0.45+0.41+0.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.154, + 0.473, + 0.167 + ], + "angle": 0, + "content": "Table 4: BLUE scores on WMT and LDC test sets." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.185, + 0.479, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.351, + 0.488, + 0.382 + ], + "angle": 0, + "content": "Figure 4: BLEU scores on En-De with different ways and numbers of adding image." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.407, + 0.313, + 0.423 + ], + "angle": 0, + "content": "4.4 Quantity of Images" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.428, + 0.49, + 0.783 + ], + "angle": 0, + "content": "We are also curious about how the method of adding images and image quantity affects the performance. Therefore, we test several ways of adding images, including selecting the best one, the top two, randomly selecting one, randomly selecting two, and randomly selecting three. We treat all images corresponding to a single word as a class and selected the image with the highest average similarity to the other images in the same class as the best image. We report the average BLEU score of En-De on the Flickr2016 and Flickr2017 test sets as illustrated in figure 4. We find that the fixed-best method yielded lower performance compared to randomly adding images. This could be due to the limitation of the fixed-best method in capturing the full range of visual variations that may exist for a given word. Among all the random selection methods, we find that adding two images randomly yielded the best performance. This could be because adding three images may introduce more noise, while adding only one image may not provide enough randomness." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.796, + 0.27, + 0.811 + ], + "angle": 0, + "content": "5 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Unsupervised Machine Translation. Unlike common machine translation (Sutskever et al., 2014; Vaswani et al., 2017; Feng et al., 2020), UNMT don't use parallel corpus. Recent advancements (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a; Conneau and Lample, 2019;" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.295 + ], + "angle": 0, + "content": "Song et al., 2019; Han et al., 2021) in UNMT have achieved remarkable milestones. However, there are still certain issues associated with solely utilizing monolingual data for translation, such as domain mismatch problem (Marchisio et al., 2020; Kim et al., 2020), poor performance on distance language pairs (Kim et al., 2020; Chronopoulou et al., 2020; Sun et al., 2021), translationese (He et al., 2022), and lexical confusion (Bapna et al., 2022; Jones et al., 2023). Our work aims to address the issue of lexical confusion by utilizing word-level image data, thereby eliminating the need for costly bilingual dictionary annotations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.303, + 0.885, + 0.673 + ], + "angle": 0, + "content": "Multimodal Machine Translation. Multimodal machine translation involves incorporating image information into a machine translation model to enhance its performance. While most current multimodal machine translation systems (Caglayan et al., 2016; Yao and Wan, 2020; Yin et al., 2020; Caglayan et al., 2021; Zhang et al., 2020; Fang and Feng, 2022; Li et al., 2022) focus on supervised machine translation, it has been observed that visual information is generally ignored when parallel corpora are plentiful (Caglayan et al., 2019; Wu et al., 2021). Yang et al. (2022) proposed the use of image information in situations where parallel corpora are not abundant. On the other hand, Su et al. (2019) and Huang et al. (2020) utilized image information in an unsupervised setting. However, these approaches were based on sentence-level image information, which can be challenging to obtain since not all sentences have corresponding images. In contrast, word-level image information is easier to obtain and is more flexible and lightweight. Therefore, our work aims to use word-level image information to enhance the performance of UNMT." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.685, + 0.797, + 0.7 + ], + "angle": 0, + "content": "6 Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.71, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this paper, we propose a method to alleviate the problem of lexical confusion in unsupervised machine translation by utilizing word-level images and provide a concise yet potent framework for integrating them into an unsupervised neural machine translation model. Through extensive experiments, we demonstrate that our method can effectively mitigate the issue of word confusion and even outperform the use of costly bilingual dictionaries in some directions. Furthermore, we release an open-source word-image dataset that covers 5 languages and involves over 300,000 images in total. In the future, we will endeavor to explore the" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.488, + 0.135 + ], + "angle": 0, + "content": "potential of utilizing images for augmenting the translation quality of low-resource languages in a large-language-model-based translation system." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.145, + 0.221, + 0.159 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.17, + 0.49, + 0.412 + ], + "angle": 0, + "content": "One limitation of our work is that we have not been able to effectively handle the issue of polysemy. The word-image dataset is obtained through direct web scraping of words, and the images collected may not all correspond to the same sense of the word. Moreover, when adding images, we have not been able to accurately select images that correspond to the intended meaning of the word in the sentence. However, we have attempted to alleviate this issue to some extent by randomly selecting images during the process of adding images, which has enabled the model to encounter the correct images to a certain degree. In the future, we plan to improve our approach by incorporating sentence-level information when inserting images." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.423, + 0.266, + 0.437 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.448, + 0.488, + 0.576 + ], + "angle": 0, + "content": "In our experimental endeavors, we have amassed a substantial collection of images, which naturally prompts discussions pertaining to these issues. It is imperative to clarify that we do not possess the copyright for these images. Instead, they will be exclusively available to researchers and educators seeking to utilize the dataset for non-commercial research and/or educational purposes." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.589, + 0.287, + 0.605 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.614, + 0.49, + 0.79 + ], + "angle": 0, + "content": "Xiaocheng Feng is the corresponding author of this work. We thank the anonymous reviewers for their insightful comments. This work was supported by the National Key RD Program of China via grant No. 2021ZD0112905, National Natural Science Foundation of China (NSFC) via grant 62276078, the Key RD Program of Heilongjiang via grant 2022ZX01A32, the International Cooperation Project of PCLPCL2022D01 and the Fundamental Research Funds for the Central Universities (Grant No.HIT.OCEF.2023018)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.817, + 0.214, + 0.832 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.839, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2018. Unsupervised neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.883, + 0.165 + ], + "angle": 0, + "content": "Ankur Bapna, Isaac Caswell, Julia Kreutzer, Orhan Firat, Daan van Esch, Aditya Siddhant, Mengmeng Niu, Pallavi Baljekar, Xavier Garcia, Wolfgang Macherey, et al. 2022. Building machine translation systems for the next thousand languages. arXiv preprint arXiv:2205.03983." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.179, + 0.885, + 0.257 + ], + "angle": 0, + "content": "Loic Barrault, Fethi Bougares, Lucia Specia, Chiraag Lala, Desmond Elliott, and Stella Frank. 2018. Findings of the third shared task on multimodal machine translation. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 304-323." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.271, + 0.885, + 0.311 + ], + "angle": 0, + "content": "Ozan Caglayan, Loic Barrault, and Fethi Bougares. 2016. Multimodal attention for neural machine translation. arXiv preprint arXiv:1609.03976." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.325, + 0.885, + 0.43 + ], + "angle": 0, + "content": "Ozan Caglayan, Menekse Kuyu, Mustafa Sercan Amac, Pranava Madhyastha, Erkut Erdem, Aykut Erdem, and Lucia Specia. 2021. Cross-lingual visual pretraining for multimodal machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1317-1324, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.444, + 0.884, + 0.562 + ], + "angle": 0, + "content": "Ozan Caglayan, Pranava Madhyastha, Lucia Specia, and Loic Barrault. 2019. Probing the need for visual context in multimodal machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4159-4170, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.575, + 0.884, + 0.667 + ], + "angle": 0, + "content": "Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2020. Reusing a Pretrained Language Model on Languages with Limited Corpora for Unsupervised NMT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2703-2711, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.681, + 0.883, + 0.721 + ], + "angle": 0, + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.734, + 0.884, + 0.827 + ], + "angle": 0, + "content": "Desmond Elliott, Stella Frank, Loic Barrault, Fethi Bougares, and Lucia Specia. 2017. Findings of the second shared task on multimodal machine translation and multilingual image description. In Proceedings of the Second Conference on Machine Translation, pages 215-233, Copenhagen, Denmark. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Desmond Elliott, Stella Frank, Khalil Sima'an, and Lucia Specia. 2016. Multi30K: Multilingual English-German image descriptions. In Proceedings of the 5th Workshop on Vision and Language, pages 70-74, Berlin, Germany. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12616" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.166 + ], + "angle": 0, + "content": "Qingkai Fang and Yang Feng. 2022. Neural machine translation with phrase-level universal visual representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5687-5698, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.174, + 0.489, + 0.24 + ], + "angle": 0, + "content": "Xiaocheng Feng, Zhangyin Feng, Wanlong Zhao, Bing Qin, and Ting Liu. 2020. Enhanced neural machine translation by joint decoding with word and posttagging sequences. Mobile Networks and Applications, 25:1722-1728." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.25, + 0.489, + 0.327 + ], + "angle": 0, + "content": "Jesse Michael Han, Igor Babuschkin, Harrison Edwards, Arvind Neelakantan, Tao Xu, Stanislas Polu, Alex Ray, Pranav Shyam, Aditya Ramesh, Alec Radford, et al. 2021. Unsupervised neural machine translation with generative language models only. arXiv preprint arXiv:2110.05448." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.337, + 0.489, + 0.443 + ], + "angle": 0, + "content": "Zhiwei He, Xing Wang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2022. Bridging the data gap between training and inference for unsupervised neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6611-6623, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.452, + 0.489, + 0.543 + ], + "angle": 0, + "content": "Po-Yao Huang, Junjie Hu, Xiaojun Chang, and Alexander Hauptmann. 2020. Unsupervised multimodal neural machine translation with pseudo visual pivoting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8226-8237, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.552, + 0.489, + 0.605 + ], + "angle": 0, + "content": "Alex Jones, Isaac Caswell, Ishank Saxena, and Orhan Firat. 2023. Bilex rx: Lexical data augmentation for massively multilingual machine translation. CoRR, abs/2303.15265." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.614, + 0.489, + 0.694 + ], + "angle": 0, + "content": "Yunsu Kim, Miguel Graça, and Hermann Ney. 2020. When and why is unsupervised neural machine translation useless? In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 35-44, Lisboa, Portugal. European Association for Machine Translation." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.703, + 0.489, + 0.768 + ], + "angle": 0, + "content": "Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018a. Unsupervised machine translation using monolingual corpora only. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.777, + 0.489, + 0.869 + ], + "angle": 0, + "content": "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018b. Word translation without parallel data. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.878, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Guillaume Lample, Myle Ott, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018c. Phrase-based & neural unsupervised machine translation." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.529, + 0.086, + 0.884, + 0.14 + ], + "angle": 0, + "content": "In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5039-5049, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.15, + 0.884, + 0.243 + ], + "angle": 0, + "content": "Bei Li, Chuanhao Lv, Zefan Zhou, Tao Zhou, Tong Xiao, Anxiang Ma, and JingBo Zhu. 2022. On vision features in multimodal machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6327-6337, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.254, + 0.884, + 0.321 + ], + "angle": 0, + "content": "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. 2020. K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2901-2908." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.333, + 0.883, + 0.399 + ], + "angle": 0, + "content": "Kelly Marchisio, Kevin Duh, and Philipp Koehn. 2020. When does unsupervised machine translation work? In Proceedings of the Fifth Conference on Machine Translation, pages 571-583, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.41, + 0.884, + 0.503 + ], + "angle": 0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.514, + 0.884, + 0.594 + ], + "angle": 0, + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.605, + 0.884, + 0.671 + ], + "angle": 0, + "content": "Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. Mass: Masked sequence to sequence pretraining for language generation. In International Conference on Machine Learning, pages 5926-5936. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.683, + 0.884, + 0.75 + ], + "angle": 0, + "content": "Yuanhang Su, Kai Fan, Nguyen Bach, C-C Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10482-10491." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.762, + 0.884, + 0.853 + ], + "angle": 0, + "content": "Haipeng Sun, Rui Wang, Masao Utiyama, Benjamin Marie, Kehai Chen, Eiichiro Sumita, and Tiejun Zhao. 2021. Unsupervised neural machine translation for similar and distant language pairs: An empirical study. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 20(1):1-17." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.865, + 0.884, + 0.918 + ], + "angle": 0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.884, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "12617" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.152 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.163, + 0.488, + 0.281 + ], + "angle": 0, + "content": "Zhiyong Wu, Lingpeng Kong, Wei Bi, Xiang Li, and Ben Kao. 2021. Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pages 6153-6166, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.291, + 0.488, + 0.37 + ], + "angle": 0, + "content": "Zhe Yang, Qingkai Fang, and Yang Feng. 2022. Lowresource neural machine translation with cross-modal alignment. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10134-10146, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.379, + 0.488, + 0.459 + ], + "angle": 0, + "content": "Shaowei Yao and Xiaojun Wan. 2020. Multimodal transformer for multimodal machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4346-4350, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.469, + 0.488, + 0.561 + ], + "angle": 0, + "content": "Yongjing Yin, Fandong Meng, Jinsong Su, Chulun Zhou, Zhengyuan Yang, Jie Zhou, and Jiebo Luo. 2020. A novel graph-based multi-modal fusion encoder for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3025-3035, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.571, + 0.488, + 0.663 + ], + "angle": 0, + "content": "Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Adversarial training for unsupervised bilingual lexicon induction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1959-1970, Vancouver, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.673, + 0.488, + 0.739 + ], + "angle": 0, + "content": "Zhuosheng Zhang, Kehai Chen, Rui Wang, Masao Utiyama, Eiichiro Sumita, Zuchao Li, and Hai Zhao. 2020. Neural machine translation with universal visual representation. In International Conference on Learning Representations." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.765, + 0.449, + 0.796 + ], + "angle": 0, + "content": "A The Implementation Details of the XLM and UNMT-CS Method" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.919 + ], + "angle": 0, + "content": "The XLM and UNMT-CS methods are re-implemented on the Multi30k dataset in our paper. During training, both these two methods are trained on the 5 million sentences from the WMT dataset for 70 epochs and subsequently continued training for 70 epochs on the Multi30k dataset. For the bilingual dictionary in codeswitching method, we use" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.198 + ], + "angle": 0, + "content": "the Ground-truth bilingual dictionaries provided in MUSE\\(^{8}\\) (Lample et al., 2018b). When training, we substituted words in \\(50\\%\\) of the sentences, while the remaining sentences were left unchanged. All other parameters used in the XLM and UNMT-CS methods are identical to those used in our proposed method, as described in our paper." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.209, + 0.853, + 0.243 + ], + "angle": 0, + "content": "B The Implicit Mapping Relationship Between Image Representations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.251, + 0.883, + 0.347 + ], + "angle": 0, + "content": "In fact, our approach doesn't explicitly provide the model with a mapping relationship. Rather, we introduce the (word-image) pairs relationship into the model, enabling it to extract information from similar images and thus enhancing the mapping relationship between words." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.348, + 0.885, + 0.542 + ], + "angle": 0, + "content": "We are also curious about the impact on results if the implicit mapping relationship between image representations was absent. Therefore, during the second stage of multi30k training, we replace all images from one language with randomly generated representations while keeping the other language's images unchanged. The En→De direction results for the Flickr2016 and Flickr2017 datasets experience a significant decline, decreasing from 40.32/35.20 to 39.17/34.29. This confirms the crucial importance of the mapping relationship between images for our approach." + }, + { + "type": "footer", + "bbox": [ + 0.532, + 0.904, + 0.808, + 0.919 + ], + "angle": 0, + "content": "\\(^{8}\\)https://github.com/facebookresearch/MUSE" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "12618" + } + ] +] \ No newline at end of file diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_origin.pdf b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edc34a7f58c5ef8ba58b7fa472a627a597be0e53 --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/da4765a2-c600-4e0c-b5cb-fcc5f37e0be1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:514313f5e2d0b5d8ea45100492376ff09150b2ca64aeddf463caa01bd18b4511 +size 644673 diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/full.md b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fd157677894570ba39217bfc2f915a7e3e3e9acd --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/full.md @@ -0,0 +1,299 @@ +# Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations + +Chengpeng Fu $^{1,2}$ , Xiaocheng Feng $^{1,2}$ , Yichong Huang $^{1}$ , Wenshuai Huo $^{1,2}$ , Hui Wang $^{2}$ , Bin Qin $^{1,2}$ , Ting Liu $^{1,2}$ + +1Harbin Institution of Technology, Harbin, China + +$^{2}$ Pengcheng Laboratory, Shenzhen, China + +fuchp@stu.hit.edu.cn, {huowsh, wangh06}@pc1.ac.cn + +{xcfeng,ychuang,qinb,tliu}@ir.hit.edu.cn + +# Abstract + +Unsupervised neural machine translation has recently made remarkable strides, achieving impressive results with the exclusive use of monolingual corpora. Nonetheless, these methods still exhibit fundamental flaws, such as confusing similar words. A straightforward remedy to rectify this drawback is to employ bilingual dictionaries, however, high-quality bilingual dictionaries can be costly to obtain. To overcome this limitation, we propose a method that incorporates images at the word level to augment the lexical mappings. Specifically, our method inserts visual representations into the model, modifying the corresponding embedding layer information. Besides, a visible matrix is adopted to isolate the impact of images on other unrelated words. Experiments on the Multi30k dataset with over 300,000 self-collected images validate the effectiveness in generating more accurate word translation, achieving an improvement of up to $+2.81$ BLEU score, which is comparable or even superior to using bilingual dictionaries. $^{1}$ + +# 1 Introduction + +Unsupervised Neural Machine Translation (UNMT) (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a,b,c) seeks to avoid the necessity for high-quality bilingual data in conventional machine translation by exploring the possibility of completing translations relying solely on the monolingual corpus. Consisting of three main components, cross-lingual language model initialization, denoising auto-encoder, and iterative back translation, recent methods (Conneau and Lample, 2019; Song et al., 2019) have achieved promising results. + +Nonetheless, recent UNMT approaches have been observed to make particular errors in comparison to supervised machine translation, such + +as confusing nouns that pertain to the same semantic category. For instance, Bapna et al. (2022) demonstrates that their models translate "lion" to "rabbit" or "snake" despite getting good CHRF scores. In the context of supervised machine translation, pinpointing accurate translations of words is straightforward with aligned bilingual data. However, with only monolingual data, the unsupervised model must rely on large amounts of context for word mapping, which can lead to confusion between words with similar distributions, particularly when there is insufficient data. To address this issue, a direct solution is to integrate high-quality bilingual dictionaries into the UNMT process using codeswitching or spliced methods (Jones et al., 2023). High-quality bilingual dictionaries can be obtained through manual annotation and bilingual data extraction, yet these methods either involve high costs or deviate from the original objective of the UNMT task. + +In addition to utilizing explicit bilingual dictionaries for word mapping, implicit methods can also be employed. As the saying goes, "A picture is worth a thousand words". We can leverage the "thousand words" represented by corresponding images to learn the mapping between words in different languages. Visual content has the potential to enhance word mapping abilities in latent spaces, as the physical visual perception is congruent among speakers of diverse languages (Huang et al., 2020). Furthermore, the cost of annotating a bilingual dictionary is relatively high, whereas the cost of searching images with monolingual keywords is relatively low in the era of sophisticated search engines. Describing an image is also less demanding for humans compared to translating a word, as the former only requires mastery of one language while the latter requires at least two. + +Drawing on these insights, we propose to integrate word-level image information into the UNMT process to mitigate the issue of lexical confusion. + +It's worth noting that we don't use sentence-level images like some multi-modal UNMT approaches (Su et al. (2019); Huang et al. (2020)). Instead, we focused on word-level images since they are more effective and flexible in addressing the challenge of lexical confusion and are easier to collect. Our approach involves augmenting the encoded images to their corresponding words, modifying the corresponding position and language encoding information, and utilizing a visible matrix to isolate the impact of images on other words. We conduct experiments utilizing over 300,000 self-collected images, and further analysis demonstrates that we have effectively mitigated lexical confusion to a certain extent. Furthermore, we achieve new state-of-the-art results for the UNMT task in certain language directions on the Multi30k dataset. The main contributions of this work can be summarized as follows: + +- We introduce a concise yet potent framework that integrates word-level images to alleviate the similar lexical confusion issue in UNMT, without utilizing bilingual dictionaries. +- Our approach has achieved promising or even unparalleled unsupervised results on the Multi30K dataset for certain language pairs. +- We release the word-image datasets for five languages as open-source, which can serve as a valuable resource to the community. + +# 2 Methods + +In this section, we first review the text-only UNMT process. Then, we present our proposed method for incorporating images, which involves the modified embedding layer and mask transformer. Finally, we introduce some training strategies. + +# 2.1 Unsupervised Neural Machine Translation + +Successful UNMT systems typically employ an encoder-decoder architecture similar to that of supervised machine translation, sharing parameters between the source-to-target and target-to-source models. They share several common principles: initialization, iterative back-translation, and denoising autoencoder. Let $X = \{X_{i}\}_{i=1}^{M}$ denote monolingual data in language $L_{1}$ and $Y = \{Y_{i}\}_{i=1}^{N}$ in language $L_{2}$ . $M$ and $N$ are the sentence number of monolingual data $X$ and $Y$ respectively. $\theta$ represents the UNMT model parameters. + +Initialization. Performing parameter initialization that is non-random can introduce cross-lingual prior knowledge into the model. There are currently two main initialization methods for UNMT models. The first method initializes the embedding layer of a UNMT model with word pretrained embeddings, while the second method uses a pretrained cross-lingual language model with the same structure as the UNMT encoder to initialize most of the neural network parameters. + +Iterative Back-translation. Back-translation is a method that leverages monolingual data in the target language and a reverse translation model to generate pseudo-source data, improving the quality of translations in the forward direction. In UNMT, where only monolingual data is available, back-translation is used to generate pseudo-parallel data, transforming the unsupervised problem into a weakly or self-supervised one. During training, back-translation is repeatedly applied in an iterative manner. The loss is defined as followed: + +$$ +\begin{array}{l} \mathcal {L} _ {B} = \sum_ {i = 1} ^ {M} - \log P _ {L _ {2} \rightarrow L _ {1}} \left(X _ {i} \mid S _ {L _ {1} \rightarrow L _ {2}} \left(X _ {i}, \theta\right), \theta\right) \tag {1} \\ + \sum_ {j = 1} ^ {N} - l o g P _ {L _ {1} \rightarrow L _ {2}} (Y _ {j} | S _ {L _ {2} \rightarrow L _ {1}} (Y _ {j}, \theta), \theta) \\ \end{array} +$$ + +where $P_{L_a\to L_b}$ denotes the translation probabilities between two languages and $S_{L_a\rightarrow L_b}$ denotes the translation processes from language $L_{a}$ to $L_{b}$ . + +Denoising Autoencoder. The introduction of the denoising autoencoder aims to guide the model in producing fluent text output. This is achieved by adding noise to sentences through techniques such as replacement, deletion, and shuffling, and then allowing the model to output the correct and fluent original sentence. The loss is defined as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {D} = \sum_ {i = 1} ^ {M} - \log P _ {L _ {1} \rightarrow L _ {1}} \left(X _ {i} \mid \mathcal {N} \left(X _ {i}\right), \theta\right) \tag {2} \\ + \sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \rightarrow L _ {2}} \left(Y _ {j} \mid \mathcal {N} (Y _ {j}), \theta\right) \\ \end{array} +$$ + +where $\mathcal{N}(\cdot)$ refers to the noise functions. $P_{L_a\to L_a}$ denotes the reconstruction probabilities in $L_{a}$ . + +The final objective function of UNMT is: + +$$ +\mathcal {L} = \mathcal {L} _ {B} + \lambda_ {D} \mathcal {L} _ {D} \tag {3} +$$ + +where $\lambda_{D}$ is a hyper-parameter that weights denoise autoencoder loss term and is gradually reduced during training. + +![](images/a27cac5cbce675af81757d07cbf656191f31d0d39c6ad2bffb58c815a869ec83.jpg) +Figure 1: The process of converting an image-inserted sentence into an embedding representation. + +# 2.2 Image Fusion + +In this section, we present how to incorporate word-level images into the UNMT process to tackle the issue of lexical confusion. The main concept behind our approach is that words can acquire information not only from contextual information but also from their corresponding images. Let $X_{i} = \{x_{i}^{1}, x_{i}^{2}, \dots, x_{i}^{K}\}$ denote a sentence from $X$ in language $L_{1}$ and $x_{i}^{K}$ denotes the $k$ -th word in sentence $X_{i}$ . We denote a lookup table consisting of word-image pairs as $\mathcal{D}_{L_i} = \{(word, I_{word})\}$ , where $I_{word}$ is an image set with one or multiple images that depict the meaning of the word. + +As shown in Figure $1^{2}$ , we segment an input sentence $X_{i}$ at the word level and obtain the images by retrieving the words appearing in the lookup table $\mathcal{D}_{L_i}$ . Then, we concatenate the retrieved images to the corresponding words, resulting in an Image-Inserted Sentence denoted $IIS_{i}$ . Each word can be concatenated with one or multiple images. Formally, We denote $IIS_{i}$ as follow: + +$$ +I I S _ {i} = \left\{x _ {i} ^ {1}, \widehat {I} _ {x _ {i} ^ {1}}, x _ {i} ^ {2}, \widehat {I} _ {x _ {i} ^ {2}}, \dots , x _ {i} ^ {K}, \widehat {I} _ {x _ {i} ^ {K}} \right\} \quad (4) +$$ + +where $\widehat{I}_{x_i^K}$ is the image subset of $I_{x_i^K}$ ( $\widehat{I}_{x_i^K}$ can be an empty set). We define this process of injecting images as: + +$$ +I I S = \mathcal {I} (X) \tag {5} +$$ + +where $IIS = \{IIS_{i}\}_{i = 1}^{M}$ . + +After obtaining the $IIS$ , we need to integrate them into the UNMT model. To achieve this integration, we employ a modified embedding layer and a mask transformer. + +# 2.2.1 Modified Embedding Layer + +The first modification part is the embedding layer, which generally includes token embedding, position embedding, and, in some models, language embedding. An example can be seen in Figure 1. + +Token Embedding. In the context of $IIS$ , both textual words and images are present. Given that most current machine translation methods operate at the subword level, we proceed to tokenize the $IIS$ , focusing solely on the textual component of $IIS$ . For encoding textual tokens, we continue to utilize the embedding layer from the original text-only model. Regarding images, we leverage the CLIP (Radford et al., 2021) model as the image encoder to extract visual features. The CLIP model is a multi-modal model trained on a variety of (image, text) pairs, which possesses a powerful capability to represent images. To prevent negative impacts on the encoding of images during the UNMT process, we freeze the visual encoder. However, directly using these representations may lead + +to a gap between the image encoding space and the semantic space in UNMT. Therefore, we employ an MLP layer to bridge this gap. + +Position Embedding. Positional encoding is a crucial module that supplements token structural information. For textual information in $IIS$ , we continue to encode the position in the original order of the input sentence to prevent interference with sentence structure learning when integrating images into the model. For images, we assign the positional encoding of the first subword in the tokenized result for the corresponding word. + +Language Embedding. Some UNMT models incorporate language embedding to inform the model of the input language and the desired output language. For textual information, we continue to utilize the original language encoding. However, for image information, as it does not belong to any language, we introduce a new image identifier. + +# 2.2.2 Mask Transformer + +Despite modifying the original embedding layer to allow $IIS$ to be properly fed into the model, two issues still exist. Firstly, inserting images may affect other irrelevant words in the original sentence and potentially disrupt the structural relationships between them. Secondly, a single word-level image may correspond to multiple subwords, and during the annotation of position embedding, only the information of the first subword is utilized, resulting in a lack of relationship between the image and the remaining subwords it corresponds to. To address these issues, we introduce the method of mask transformer (Liu et al., 2020). + +We create a visible matrix to shield the relationship between images and other textual information in the sentence and highlight the relationship between the image and its corresponding subwords. The visible matrix $M$ is defined as: + +$$ +M _ {i j} = \left\{ \begin{array}{l l} - \infty & w _ {p} \bullet w _ {q} \\ 0 & w _ {p} \circ w _ {q}. \end{array} \right. \tag {6} +$$ + +where $w_{p}$ and $w_{q}$ are elements in $IIS_{i}$ , $w_{p} \circ w_{q}$ denotes that $w_{p}$ and $w_{q}$ are visible to each other, while $w_{p} \bullet w_{q}$ is not. An example of a visible matrix is depicted in Figure 2. + +The visible matrix can be added to the self-attention module to mask certain attention calculations. Specifically, the new masked self-attention is: + +![](images/21012451b1d6a291a0d8c1850e6f6419c4639d9dc4addb6ec6ebd30cad1dbd66.jpg) +Figure 2: An example of the visible matrix corresponding to the $IIS$ in Figure 1. The red numerical labels indicate the absolute position of the images in the $IIS$ . In this matrix, black represents "visible" and white represents "invisible". + +$$ +S ^ {i} = \operatorname {s o f t m a x} \left(\frac {Q ^ {i} K ^ {i T} + M}{\sqrt {d}}\right) \tag {7} +$$ + +where $Q$ and $K$ are the query matrix and key matrix in the self-attention module. if two elements are invisible, the $M$ will mask the attention score to 0, which means making no contribution to each other. + +# 2.3 Training Strategy + +After initialization by cross-lingual pretrained language model, we train denoising autoencoder and iterative back-translation alternatively. The final loss of image fusion UNMT is: + +$$ +\mathcal {L} ^ {I} = \mathcal {L} _ {B} ^ {I} + \lambda_ {D} \mathcal {L} _ {D} ^ {I} \tag {8} +$$ + +specifically, $\mathcal{L}_B^I$ and $\mathcal{L}_D^I$ refer to: + +$$ +\begin{array}{l} \mathcal {L} _ {B} ^ {I} = \sum_ {i = 1} ^ {M} - l o g P _ {L _ {2} \rightarrow L _ {1}} (X _ {i} | \mathcal {I} (S _ {L _ {1} \rightarrow L _ {2}} (\mathcal {I} (X _ {i}), \theta)), \theta) \\ + \sum_ {j = 1} ^ {N} - \log P _ {L _ {1} \rightarrow L _ {2}} \left(Y _ {j} \mid \mathcal {I} \left(S _ {L _ {2} \rightarrow L _ {1}} \left(\mathcal {I} \left(Y _ {j}\right), \theta\right)\right), \theta\right) \tag {9} \\ \end{array} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {D} ^ {I} = \sum_ {i = 1} ^ {M} - \log P _ {L _ {1} \rightarrow L _ {1}} \left(X _ {i} \mid \mathcal {I} \left(N \left(X _ {i}\right)\right), \theta\right) \tag {10} \\ + \sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \rightarrow L _ {2}} (Y _ {j} | \mathcal {I} (N (Y _ {j})), \theta) \\ \end{array} +$$ + +During the denoising autoencoder process, we first add noise to the original sentence and then integrate the images. During the iterative backtranslation process, for the first round of synthetic data generation, we do not add image information to achieve a smoother transition because the pretrained language model did not include images. + +
Language#Word#ImageAverage
En4,97545,2289.09
Fr4,64636,9277.95
De10,38485,4198.23
Cs10,70672,9426.81
Zh10,80585,1447.88
Total41,516325,6607.84
+ +Table 1: The statistics of our constructed Word-Image dataset. + +# 3 Experiments + +In this section, we provide a detailed description of the datasets and experimental setup employed in our study. Subsequently, we perform an in-depth analysis of the results obtained. + +# 3.1 Datasets + +Word-Image Dataset. In our approach, the most crucial component of the experiment data is the image dataset at the word level. Since there is no off-the-shelf word-level image dataset, we build one for performance evaluation. We collect these images through the following steps: word extraction, image crawling, and image filtering. Firstly, extract the nouns from the monolingual dataset using part-of-speech analysis tools including Spacy3 and Stanford CoreNLP4. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images5 to crawl images for these nouns, limiting the image size to between 200-by-200 and 10,000-by-10,000 pixels and retrieving 10 to 15 images per noun. The image format is restricted to JPG, JPEG, and PNG. Finally, filter the crawled images. We consider all the images collected for a single noun as a cluster and encode them using the CLIP model. We then calculate pairwise similarities between all images and remove those with an average similarity below a threshold $\mathcal{T}$ , which we set to 0.6, as they are considered anomalous. The statistics information of the Word-Image dataset can be seen in Table 1. Our dataset covers a wide range of common nouns and can be reused for various purposes. + +Monolingual Dataset. The monolingual datasets used in our experiments are Multi30k dataset (Elliott et al., 2016, 2017; Barrault et al., 2018), LDC dataset and WMT News crawl dataset. The Multi30k dataset contains images and their captions in four languages: English(En), French(Fr), Germany(De), and Czech(Cs). Every language has 29000 sentences for training and 1014 for validation. It is worth noting that we only utilized the lingual data from Multi30k as the images are sentence-level. To prevent the model from encountering bilingual corpora, we randomly split the training and validation sets of the Multi30k dataset into two equal parts. Each part of the training set contains 14,500 sentences, and each part of the validation set contains 507 sentences. We report results in all test sets in Multi30k including Test2016, Test2017, Test2018, and MsCOCO. For the Cs-En task, only Test2016 and Test2018 are available. Furthermore, we expanded the dataset by incorporating monolingual data from WMT 2008-2022 News crawl and LDC. We also report our results on the common news domain test set, which includes WMT Newstest 2014 for En-Fr, Newstest 2021 for En-Cs, and NIST06 for En-Zh. + +# 3.2 Experimental Details + +Baseline Models. We compare recent unsupervised text-only and multimodal MT baselines listed in the following: (1) XLM (Conneau and Lample, 2019) uses a masked language model to train a cross-lingual language model to initialize. (2) UMMT (Su et al., 2019) uses visual features for denoising autoencoder and back-translation. (3) PVP (Huang et al., 2020) employs multimodal back-translation and features pseudo visual pivoting. (4) UNMT-CS (Jones et al., 2023) uses the codeswitching method, where words in the source sentence are swapped out for their corresponding translation from bilingual dictionaries, to solve the lexical confusion issue in UNMT. It is worth noting that both UMMT and PVP datasets utilize sentence-level images. + +System Settings. We conducted our experiments based on the XLM codebase6. The model architecture employed is a 6-layers, 8-heads transformer, with a hidden layer representation dimension of 1024. The epoch size is 100K and the tokens-per-batch is 2000. we use Adam optimizer and 2000 warm-up updates. The learning rate is 1e-5. For + +
En-FrEn-DeEn-Cs
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT39.80*---23.50*-----
PVP52.30*---33.90*-----
XLM53.0846.8235.3143.0931.8126.8325.6222.9525.1420.15
UNMT-CS54.2446.3036.0643.4933.4327.5126.6724.2726.4622.71
Ours54.0746.9536.2144.4433.0629.1227.3724.7526.3422.88
Δ+0.99+0.13+0.90+1.35+1.25+2.29+1.75+1.80+1.20+2.73
Fr-EnDe-EnCs-En
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT40.50*---26.40*-----
PVP46.00*---31.60*-----
XLM48.0442.1437.0343.5837.5133.2130.7327.9833.1330.29
UNMT-CS49.2842.8637.1944.3938.8334.4831.4428.0733.9131.33
Ours49.1043.0137.4144.5740.3235.2032.7029.8134.2931.56
Δ+1.06+0.87+0.38+0.99+2.81+1.99+1.97+1.83+1.16+1.27
+ +Table 2: BLEU scores on Multi30k En-Fr, En-De, and En-Cs UNMT tasks. F18, F17, F16, and C17 respectively denote the Flickr2018, Flickr2017, Flickr2016, and COCO2017 test sets in Multi30k. * is the results reported from previous papers. Δ represents the improvement of our method over the XLM method in terms of BLEU. + +evaluation, we use 4-gram BLEU (Papineni et al., 2002) scores by multi-bleu.pl in Moses7. We train all models on 4 NVIDIA 32GB V100 GPUs. + +During training, we first train the models on a 5M WMT News Crawl dataset for 70 epochs and subsequently continued training on the Multi30k dataset for another 70 epochs due to its relatively small size. As for the CLIP model, we use ViT-L/14@336px. When incorporating images, we randomly select two images from the corresponding image set for each word and concatenate them to the text. We only incorporate images during the iterative back-translation and denoising autoencoder stages, so our method can be used in conjunction with other pre-training enhancement methods. Regarding the cross-lingual models used for initialization, we directly employ high-quality open-source cross-lingual models in experiments involving En-Fr, En-De, and En-Zh. However, for the En-Cs, we train a model using WMT monolingual dataset, with 10 million English sentences and 10 million Czech sentences, due to the unavailability of any open-source models. + +# 3.3 Main Results + +Table 2 shows the main results on the Multi30k test set, including translating among English, French, German, and Czech. Comparing the results of XLM and our method, we observe that adding im + +ages leads to performance improvements in all six translation directions, with a maximum increase of $+2.81$ BLEU points, indicating the effectiveness of our approach. UNMT-CS is a method that utilizes bilingual dictionaries to address the problem of lexical confusion via code-switching. By comparing our method with the UNMT-CS approach, we observed comparable results, and in some directions, our method even outperformed the code-switching approach, indicating that ours is more efficient and cost-effective. + +When compared to the best-performing method that incorporates sentence-level images, our approach still achieves comparable or better results in most translation directions. This is despite the matching degree between the Multi30k monolingual data and word-level images being slightly lower than that of sentence-level images, as the text data in Multi30k is generated based on the sentence-level images. In some directions, our approach even outperforms the previous best methods by up to $+4.22$ BLEU points (as shown in Table 2, where our method is compared to the first two methods of the table). + +# 4 Analysis + +# 4.1 Case Study + +In Table 3, we provide several examples of translations, where we observe that the XLM method re + +
GT-DEMenschen an der Seitenlinie bei einem Fußballspiel.
GT-ENPeople on the sideline of a soccer match.
XLMPeople at the plate in a baseball game.
OursPeople at the stands in a soccer game.
GT-DEEin kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.
GT-ENA young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.
XLMA small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.
OursA small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.
GT-FRUn homme avec une veste noire, une casquette à carreaux et un pantalon rayé noir et blanc joue de la guitare électrique sur une scène avec un chanteur et un autre guitariste en arrêtère-plan.
GT-ENA man in a black jacket and checkered hat wearing black and white striped pants plays an electric guitar on a stage with a singer and another guitar player in the background.
XLMA man with a black jacket, striped headband and light brown and white shirt plays an electric guitar on a stage with a singer and another guitarist in the background.
OursA man with a black jacket, checkered cap and black and white striped pants plays an electric guitar on a stage with a singer and another guitarist in the background.
+ +Table 3: Qualitative results of the proposed model. GT denotes ground truth. Words highlighted in red are affected by the problem of lexical confusion. + +![](images/7040d306fc6f2b72cc36bf158842215c74fd5849719144a37c7ddd37e6629ac2.jpg) +Figure 3: 1-gram precision scores on Flickr2016 and Flickr2017 data sets. + +sults in some lexical confusion errors. For instance, in Example 1, "Fußballspiel" is translated to "baseball" instead of "soccer", and in Example 2, "orangefarbenen" is translated to "orange-red" instead of "orange". The representation of these words is similar, which makes it difficult for the model to differentiate between them. However, when we incorporate image information, the words obtain additional information from the image representation, which helps alleviate this problem to some extent. In Table 3, we can see that these words are correctly translated when we add image information. Therefore, our approach can effectively address the issue of lexical confusion in UNMT. + +# 4.2 1-gram Precision Scores + +To better observe how our method can alleviate the issue of lexical confusion, we report the 1-gram precision scores in Figure 3. The accuracy of the model's word translations can be reflected intuitively in the 1-gram precision scores metric. We report the results of our approach with and without + +incorporating images for various translation directions on the Flickr2016 and Flickr2017 test sets and find that our method with image incorporation outperforms the method without images in all directions. This demonstrates that our method can indeed provide more accurate word translations. + +# 4.3 Performance on WMT and LDC Dataset + +Not all sentences have a corresponding image that perfectly matches the text. On the other hand, generally, each sentence has some words that can be matched with images, making the use of word-level images more flexible and less costly to collect. Hence, we also conduct experiments on a more common news dataset. We conduct experiments on En-Fr, En-Cs, and En-Zh translation directions using the WMT 5M monolingual data for En-Fr and En-Cs, and the LDC dataset for En-Zh. We split the LDC bilingual data into two non-overlapping parts, each containing 0.75 million monolingual sentences. The news dataset is more abstract in content compared to the Multi30k dataset, and it is difficult to find an image that perfectly matches a sentence. This makes it challenging or almost impossible to apply previous methods that rely on sentence-level images to this dataset. As shown in Table 4, our method can be applied to the news dataset, and it shows improvements compared to the method that only uses text. The improvement on the WMT dataset was not significant, and we speculate that this may be due to the sufficient size of the WMT dataset and the lower frequency of noun entities with images compared to Multi30k. + +
ApproachZh-EnEn-ZhCs-EnEn-CsFr-EnEn-Fr
XLM8.437.9125.1417.7933.7335.77
Our Method9.879.3625.5118.2434.1436.46
Δ+1.44+1.45+0.37+0.45+0.41+0.69
+ +Table 4: BLUE scores on WMT and LDC test sets. + +![](images/008d6c4d02a5b500a1f50daca9f812157920c343553bc8a301e17dbd4d20fa7a.jpg) +Figure 4: BLEU scores on En-De with different ways and numbers of adding image. + +# 4.4 Quantity of Images + +We are also curious about how the method of adding images and image quantity affects the performance. Therefore, we test several ways of adding images, including selecting the best one, the top two, randomly selecting one, randomly selecting two, and randomly selecting three. We treat all images corresponding to a single word as a class and selected the image with the highest average similarity to the other images in the same class as the best image. We report the average BLEU score of En-De on the Flickr2016 and Flickr2017 test sets as illustrated in figure 4. We find that the fixed-best method yielded lower performance compared to randomly adding images. This could be due to the limitation of the fixed-best method in capturing the full range of visual variations that may exist for a given word. Among all the random selection methods, we find that adding two images randomly yielded the best performance. This could be because adding three images may introduce more noise, while adding only one image may not provide enough randomness. + +# 5 Related Work + +Unsupervised Machine Translation. Unlike common machine translation (Sutskever et al., 2014; Vaswani et al., 2017; Feng et al., 2020), UNMT don't use parallel corpus. Recent advancements (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a; Conneau and Lample, 2019; + +Song et al., 2019; Han et al., 2021) in UNMT have achieved remarkable milestones. However, there are still certain issues associated with solely utilizing monolingual data for translation, such as domain mismatch problem (Marchisio et al., 2020; Kim et al., 2020), poor performance on distance language pairs (Kim et al., 2020; Chronopoulou et al., 2020; Sun et al., 2021), translationese (He et al., 2022), and lexical confusion (Bapna et al., 2022; Jones et al., 2023). Our work aims to address the issue of lexical confusion by utilizing word-level image data, thereby eliminating the need for costly bilingual dictionary annotations. + +Multimodal Machine Translation. Multimodal machine translation involves incorporating image information into a machine translation model to enhance its performance. While most current multimodal machine translation systems (Caglayan et al., 2016; Yao and Wan, 2020; Yin et al., 2020; Caglayan et al., 2021; Zhang et al., 2020; Fang and Feng, 2022; Li et al., 2022) focus on supervised machine translation, it has been observed that visual information is generally ignored when parallel corpora are plentiful (Caglayan et al., 2019; Wu et al., 2021). Yang et al. (2022) proposed the use of image information in situations where parallel corpora are not abundant. On the other hand, Su et al. (2019) and Huang et al. (2020) utilized image information in an unsupervised setting. However, these approaches were based on sentence-level image information, which can be challenging to obtain since not all sentences have corresponding images. In contrast, word-level image information is easier to obtain and is more flexible and lightweight. Therefore, our work aims to use word-level image information to enhance the performance of UNMT. + +# 6 Conclusion and Future Work + +In this paper, we propose a method to alleviate the problem of lexical confusion in unsupervised machine translation by utilizing word-level images and provide a concise yet potent framework for integrating them into an unsupervised neural machine translation model. Through extensive experiments, we demonstrate that our method can effectively mitigate the issue of word confusion and even outperform the use of costly bilingual dictionaries in some directions. Furthermore, we release an open-source word-image dataset that covers 5 languages and involves over 300,000 images in total. In the future, we will endeavor to explore the + +potential of utilizing images for augmenting the translation quality of low-resource languages in a large-language-model-based translation system. + +# Limitations + +One limitation of our work is that we have not been able to effectively handle the issue of polysemy. The word-image dataset is obtained through direct web scraping of words, and the images collected may not all correspond to the same sense of the word. Moreover, when adding images, we have not been able to accurately select images that correspond to the intended meaning of the word in the sentence. However, we have attempted to alleviate this issue to some extent by randomly selecting images during the process of adding images, which has enabled the model to encounter the correct images to a certain degree. In the future, we plan to improve our approach by incorporating sentence-level information when inserting images. + +# Ethics Statement + +In our experimental endeavors, we have amassed a substantial collection of images, which naturally prompts discussions pertaining to these issues. It is imperative to clarify that we do not possess the copyright for these images. Instead, they will be exclusively available to researchers and educators seeking to utilize the dataset for non-commercial research and/or educational purposes. + +# Acknowledgements + +Xiaocheng Feng is the corresponding author of this work. We thank the anonymous reviewers for their insightful comments. This work was supported by the National Key RD Program of China via grant No. 2021ZD0112905, National Natural Science Foundation of China (NSFC) via grant 62276078, the Key RD Program of Heilongjiang via grant 2022ZX01A32, the International Cooperation Project of PCLPCL2022D01 and the Fundamental Research Funds for the Central Universities (Grant No.HIT.OCEF.2023018). + +# References + +Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2018. Unsupervised neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. + +Ankur Bapna, Isaac Caswell, Julia Kreutzer, Orhan Firat, Daan van Esch, Aditya Siddhant, Mengmeng Niu, Pallavi Baljekar, Xavier Garcia, Wolfgang Macherey, et al. 2022. Building machine translation systems for the next thousand languages. arXiv preprint arXiv:2205.03983. +Loic Barrault, Fethi Bougares, Lucia Specia, Chiraag Lala, Desmond Elliott, and Stella Frank. 2018. Findings of the third shared task on multimodal machine translation. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 304-323. +Ozan Caglayan, Loic Barrault, and Fethi Bougares. 2016. Multimodal attention for neural machine translation. arXiv preprint arXiv:1609.03976. +Ozan Caglayan, Menekse Kuyu, Mustafa Sercan Amac, Pranava Madhyastha, Erkut Erdem, Aykut Erdem, and Lucia Specia. 2021. Cross-lingual visual pretraining for multimodal machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1317-1324, Online. Association for Computational Linguistics. +Ozan Caglayan, Pranava Madhyastha, Lucia Specia, and Loic Barrault. 2019. Probing the need for visual context in multimodal machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4159-4170, Minneapolis, Minnesota. Association for Computational Linguistics. +Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2020. Reusing a Pretrained Language Model on Languages with Limited Corpora for Unsupervised NMT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2703-2711, Online. Association for Computational Linguistics. +Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32. +Desmond Elliott, Stella Frank, Loic Barrault, Fethi Bougares, and Lucia Specia. 2017. Findings of the second shared task on multimodal machine translation and multilingual image description. In Proceedings of the Second Conference on Machine Translation, pages 215-233, Copenhagen, Denmark. Association for Computational Linguistics. +Desmond Elliott, Stella Frank, Khalil Sima'an, and Lucia Specia. 2016. Multi30K: Multilingual English-German image descriptions. In Proceedings of the 5th Workshop on Vision and Language, pages 70-74, Berlin, Germany. Association for Computational Linguistics. + +Qingkai Fang and Yang Feng. 2022. Neural machine translation with phrase-level universal visual representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5687-5698, Dublin, Ireland. Association for Computational Linguistics. +Xiaocheng Feng, Zhangyin Feng, Wanlong Zhao, Bing Qin, and Ting Liu. 2020. Enhanced neural machine translation by joint decoding with word and posttagging sequences. Mobile Networks and Applications, 25:1722-1728. +Jesse Michael Han, Igor Babuschkin, Harrison Edwards, Arvind Neelakantan, Tao Xu, Stanislas Polu, Alex Ray, Pranav Shyam, Aditya Ramesh, Alec Radford, et al. 2021. Unsupervised neural machine translation with generative language models only. arXiv preprint arXiv:2110.05448. +Zhiwei He, Xing Wang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2022. Bridging the data gap between training and inference for unsupervised neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6611-6623, Dublin, Ireland. Association for Computational Linguistics. +Po-Yao Huang, Junjie Hu, Xiaojun Chang, and Alexander Hauptmann. 2020. Unsupervised multimodal neural machine translation with pseudo visual pivoting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8226-8237, Online. Association for Computational Linguistics. +Alex Jones, Isaac Caswell, Ishank Saxena, and Orhan Firat. 2023. Bilex rx: Lexical data augmentation for massively multilingual machine translation. CoRR, abs/2303.15265. +Yunsu Kim, Miguel Graça, and Hermann Ney. 2020. When and why is unsupervised neural machine translation useless? In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 35-44, Lisboa, Portugal. European Association for Machine Translation. +Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018a. Unsupervised machine translation using monolingual corpora only. In International Conference on Learning Representations. +Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018b. Word translation without parallel data. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net. +Guillaume Lample, Myle Ott, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018c. Phrase-based & neural unsupervised machine translation. + +In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5039-5049, Brussels, Belgium. Association for Computational Linguistics. +Bei Li, Chuanhao Lv, Zefan Zhou, Tao Zhou, Tong Xiao, Anxiang Ma, and JingBo Zhu. 2022. On vision features in multimodal machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6327-6337, Dublin, Ireland. Association for Computational Linguistics. +Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. 2020. K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2901-2908. +Kelly Marchisio, Kevin Duh, and Philipp Koehn. 2020. When does unsupervised machine translation work? In Proceedings of the Fifth Conference on Machine Translation, pages 571-583, Online. Association for Computational Linguistics. +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. +Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR. +Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. Mass: Masked sequence to sequence pretraining for language generation. In International Conference on Machine Learning, pages 5926-5936. PMLR. +Yuanhang Su, Kai Fan, Nguyen Bach, C-C Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10482-10491. +Haipeng Sun, Rui Wang, Masao Utiyama, Benjamin Marie, Kehai Chen, Eiichiro Sumita, and Tiejun Zhao. 2021. Unsupervised neural machine translation for similar and distant language pairs: An empirical study. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 20(1):1-17. +Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. +Zhiyong Wu, Lingpeng Kong, Wei Bi, Xiang Li, and Ben Kao. 2021. Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pages 6153-6166, Online. Association for Computational Linguistics. +Zhe Yang, Qingkai Fang, and Yang Feng. 2022. Lowresource neural machine translation with cross-modal alignment. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10134-10146, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Shaowei Yao and Xiaojun Wan. 2020. Multimodal transformer for multimodal machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4346-4350, Online. Association for Computational Linguistics. +Yongjing Yin, Fandong Meng, Jinsong Su, Chulun Zhou, Zhengyuan Yang, Jie Zhou, and Jiebo Luo. 2020. A novel graph-based multi-modal fusion encoder for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3025-3035, Online. Association for Computational Linguistics. +Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Adversarial training for unsupervised bilingual lexicon induction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1959-1970, Vancouver, Canada. Association for Computational Linguistics. +Zhuosheng Zhang, Kehai Chen, Rui Wang, Masao Utiyama, Eiichiro Sumita, Zuchao Li, and Hai Zhao. 2020. Neural machine translation with universal visual representation. In International Conference on Learning Representations. + +# A The Implementation Details of the XLM and UNMT-CS Method + +The XLM and UNMT-CS methods are re-implemented on the Multi30k dataset in our paper. During training, both these two methods are trained on the 5 million sentences from the WMT dataset for 70 epochs and subsequently continued training for 70 epochs on the Multi30k dataset. For the bilingual dictionary in codeswitching method, we use + +the Ground-truth bilingual dictionaries provided in MUSE $^{8}$ (Lample et al., 2018b). When training, we substituted words in $50\%$ of the sentences, while the remaining sentences were left unchanged. All other parameters used in the XLM and UNMT-CS methods are identical to those used in our proposed method, as described in our paper. + +# B The Implicit Mapping Relationship Between Image Representations + +In fact, our approach doesn't explicitly provide the model with a mapping relationship. Rather, we introduce the (word-image) pairs relationship into the model, enabling it to extract information from similar images and thus enhancing the mapping relationship between words. + +We are also curious about the impact on results if the implicit mapping relationship between image representations was absent. Therefore, during the second stage of multi30k training, we replace all images from one language with randomly generated representations while keeping the other language's images unchanged. The En→De direction results for the Flickr2016 and Flickr2017 datasets experience a significant decline, decreasing from 40.32/35.20 to 39.17/34.29. This confirms the crucial importance of the mapping relationship between images for our approach. \ No newline at end of file diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/images.zip b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a95d9d42799632316aeac4912986840c43ae577 --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82807bf951294e175fca151493c650fcf51067b467b0013d84943588725333b2 +size 438940 diff --git a/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/layout.json b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..016c27f4ae90d3f8d1e08d419b634e968731baf5 --- /dev/null +++ b/2023/Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations/layout.json @@ -0,0 +1,7604 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 68, + 509, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 68, + 509, + 103 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 509, + 103 + ], + "type": "text", + "content": "Enabling Unsupervised Neural Machine Translation with Word-level Visual Representations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "spans": [ + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": "Chengpeng Fu" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Xiaocheng Feng" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Yichong Huang" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Wenshuai Huo" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Hui Wang" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Bin Qin" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "text", + "content": ", Ting Liu" + }, + { + "bbox": [ + 113, + 109, + 486, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 179, + 139, + 416, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 139, + 416, + 153 + ], + "spans": [ + { + "bbox": [ + 179, + 139, + 416, + 153 + ], + "type": "text", + "content": "1Harbin Institution of Technology, Harbin, China" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 153, + 399, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 153, + 399, + 166 + ], + "spans": [ + { + "bbox": [ + 197, + 153, + 399, + 166 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 153, + 399, + 166 + ], + "type": "text", + "content": "Pengcheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 167, + 443, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 167, + 443, + 179 + ], + "spans": [ + { + "bbox": [ + 154, + 167, + 443, + 179 + ], + "type": "text", + "content": "fuchp@stu.hit.edu.cn, {huowsh, wangh06}@pc1.ac.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 177, + 181, + 419, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 181, + 419, + 194 + ], + "spans": [ + { + "bbox": [ + 177, + 181, + 419, + 194 + ], + "type": "text", + "content": "{xcfeng,ychuang,qinb,tliu}@ir.hit.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "spans": [ + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "text", + "content": "Unsupervised neural machine translation has recently made remarkable strides, achieving impressive results with the exclusive use of monolingual corpora. Nonetheless, these methods still exhibit fundamental flaws, such as confusing similar words. A straightforward remedy to rectify this drawback is to employ bilingual dictionaries, however, high-quality bilingual dictionaries can be costly to obtain. To overcome this limitation, we propose a method that incorporates images at the word level to augment the lexical mappings. Specifically, our method inserts visual representations into the model, modifying the corresponding embedding layer information. Besides, a visible matrix is adopted to isolate the impact of images on other unrelated words. Experiments on the Multi30k dataset with over 300,000 self-collected images validate the effectiveness in generating more accurate word translation, achieving an improvement of up to " + }, + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "inline_equation", + "content": "+2.81" + }, + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "text", + "content": " BLEU score, which is comparable or even superior to using bilingual dictionaries." + }, + { + "bbox": [ + 84, + 235, + 274, + 509 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 520, + 154, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 520, + 154, + 533 + ], + "spans": [ + { + "bbox": [ + 68, + 520, + 154, + 533 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 541, + 290, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 541, + 290, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 541, + 290, + 703 + ], + "type": "text", + "content": "Unsupervised Neural Machine Translation (UNMT) (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a,b,c) seeks to avoid the necessity for high-quality bilingual data in conventional machine translation by exploring the possibility of completing translations relying solely on the monolingual corpus. Consisting of three main components, cross-lingual language model initialization, denoising auto-encoder, and iterative back translation, recent methods (Conneau and Lample, 2019; Song et al., 2019) have achieved promising results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 703, + 291, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 291, + 745 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 291, + 745 + ], + "type": "text", + "content": "Nonetheless, recent UNMT approaches have been observed to make particular errors in comparison to supervised machine translation, such" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 213, + 527, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 527, + 482 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 527, + 482 + ], + "type": "text", + "content": "as confusing nouns that pertain to the same semantic category. For instance, Bapna et al. (2022) demonstrates that their models translate \"lion\" to \"rabbit\" or \"snake\" despite getting good CHRF scores. In the context of supervised machine translation, pinpointing accurate translations of words is straightforward with aligned bilingual data. However, with only monolingual data, the unsupervised model must rely on large amounts of context for word mapping, which can lead to confusion between words with similar distributions, particularly when there is insufficient data. To address this issue, a direct solution is to integrate high-quality bilingual dictionaries into the UNMT process using codeswitching or spliced methods (Jones et al., 2023). High-quality bilingual dictionaries can be obtained through manual annotation and bilingual data extraction, yet these methods either involve high costs or deviate from the original objective of the UNMT task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 486, + 526, + 729 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 486, + 526, + 729 + ], + "spans": [ + { + "bbox": [ + 302, + 486, + 526, + 729 + ], + "type": "text", + "content": "In addition to utilizing explicit bilingual dictionaries for word mapping, implicit methods can also be employed. As the saying goes, \"A picture is worth a thousand words\". We can leverage the \"thousand words\" represented by corresponding images to learn the mapping between words in different languages. Visual content has the potential to enhance word mapping abilities in latent spaces, as the physical visual perception is congruent among speakers of diverse languages (Huang et al., 2020). Furthermore, the cost of annotating a bilingual dictionary is relatively high, whereas the cost of searching images with monolingual keywords is relatively low in the era of sophisticated search engines. Describing an image is also less demanding for humans compared to translating a word, as the former only requires mastery of one language while the latter requires at least two." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "content": "Drawing on these insights, we propose to integrate word-level image information into the UNMT process to mitigate the issue of lexical confusion." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 751, + 290, + 772 + ], + "type": "text", + "content": "1The word-image dataset and associated codes are publicly available at: https://github.com/Brianmax99/UNMT-WVR" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "12608" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "spans": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 12608-12618 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 327 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 327 + ], + "type": "text", + "content": "It's worth noting that we don't use sentence-level images like some multi-modal UNMT approaches (Su et al. (2019); Huang et al. (2020)). Instead, we focused on word-level images since they are more effective and flexible in addressing the challenge of lexical confusion and are easier to collect. Our approach involves augmenting the encoded images to their corresponding words, modifying the corresponding position and language encoding information, and utilizing a visible matrix to isolate the impact of images on other words. We conduct experiments utilizing over 300,000 self-collected images, and further analysis demonstrates that we have effectively mitigated lexical confusion to a certain extent. Furthermore, we achieve new state-of-the-art results for the UNMT task in certain language directions on the Multi30k dataset. The main contributions of this work can be summarized as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 81, + 342, + 290, + 490 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 81, + 342, + 290, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 342, + 290, + 396 + ], + "spans": [ + { + "bbox": [ + 81, + 342, + 290, + 396 + ], + "type": "text", + "content": "- We introduce a concise yet potent framework that integrates word-level images to alleviate the similar lexical confusion issue in UNMT, without utilizing bilingual dictionaries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 402, + 290, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 402, + 290, + 444 + ], + "spans": [ + { + "bbox": [ + 81, + 402, + 290, + 444 + ], + "type": "text", + "content": "- Our approach has achieved promising or even unparalleled unsupervised results on the Multi30K dataset for certain language pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 449, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 449, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 81, + 449, + 290, + 490 + ], + "type": "text", + "content": "- We release the word-image datasets for five languages as open-source, which can serve as a valuable resource to the community." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 68, + 502, + 134, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 502, + 134, + 515 + ], + "spans": [ + { + "bbox": [ + 68, + 502, + 134, + 515 + ], + "type": "text", + "content": "2 Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 526, + 291, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 526, + 291, + 594 + ], + "spans": [ + { + "bbox": [ + 67, + 526, + 291, + 594 + ], + "type": "text", + "content": "In this section, we first review the text-only UNMT process. Then, we present our proposed method for incorporating images, which involves the modified embedding layer and mask transformer. Finally, we introduce some training strategies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 605, + 290, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 605, + 290, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 605, + 290, + 619 + ], + "type": "text", + "content": "2.1 Unsupervised Neural Machine Translation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": "Successful UNMT systems typically employ an encoder-decoder architecture similar to that of supervised machine translation, sharing parameters between the source-to-target and target-to-source models. They share several common principles: initialization, iterative back-translation, and denoising autoencoder. Let " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "X = \\{X_{i}\\}_{i=1}^{M}" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " denote monolingual data in language " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "Y = \\{Y_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " in language " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " are the sentence number of monolingual data " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 624, + 292, + 773 + ], + "type": "text", + "content": " represents the UNMT model parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 206 + ], + "type": "text", + "content": "Initialization. Performing parameter initialization that is non-random can introduce cross-lingual prior knowledge into the model. There are currently two main initialization methods for UNMT models. The first method initializes the embedding layer of a UNMT model with word pretrained embeddings, while the second method uses a pretrained cross-lingual language model with the same structure as the UNMT encoder to initialize most of the neural network parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 213, + 526, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 361 + ], + "type": "text", + "content": "Iterative Back-translation. Back-translation is a method that leverages monolingual data in the target language and a reverse translation model to generate pseudo-source data, improving the quality of translations in the forward direction. In UNMT, where only monolingual data is available, back-translation is used to generate pseudo-parallel data, transforming the unsupervised problem into a weakly or self-supervised one. During training, back-translation is repeatedly applied in an iterative manner. The loss is defined as followed:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 323, + 375, + 525, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 375, + 525, + 437 + ], + "spans": [ + { + "bbox": [ + 323, + 375, + 525, + 437 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {B} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {2} \\rightarrow L _ {1}} \\left(X _ {i} \\mid S _ {L _ {1} \\rightarrow L _ {2}} \\left(X _ {i}, \\theta\\right), \\theta\\right) \\tag {1} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {1} \\rightarrow L _ {2}} (Y _ {j} | S _ {L _ {2} \\rightarrow L _ {1}} (Y _ {j}, \\theta), \\theta) \\\\ \\end{array}", + "image_path": "4d79a95959738364d67ffa61c0b5cf4bc275f8548f982d6a8f6e18c58d659dd8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "spans": [ + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "inline_equation", + "content": "P_{L_a\\to L_b}" + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "content": " denotes the translation probabilities between two languages and " + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "inline_equation", + "content": "S_{L_a\\rightarrow L_b}" + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "content": " denotes the translation processes from language " + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "inline_equation", + "content": "L_{b}" + }, + { + "bbox": [ + 302, + 442, + 524, + 483 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 489, + 525, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 489, + 525, + 584 + ], + "spans": [ + { + "bbox": [ + 302, + 489, + 525, + 584 + ], + "type": "text", + "content": "Denoising Autoencoder. The introduction of the denoising autoencoder aims to guide the model in producing fluent text output. This is achieved by adding noise to sentences through techniques such as replacement, deletion, and shuffling, and then allowing the model to output the correct and fluent original sentence. The loss is defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 597, + 525, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 597, + 525, + 659 + ], + "spans": [ + { + "bbox": [ + 339, + 597, + 525, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {N} \\left(X _ {i}\\right), \\theta\\right) \\tag {2} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {N} (Y _ {j}), \\theta\\right) \\\\ \\end{array}", + "image_path": "5387fd4a851618cfbe07a2f29642ccf8922c339713e5bd94f62d8fdbf8c53499.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\cdot)" + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "text", + "content": " refers to the noise functions. " + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "inline_equation", + "content": "P_{L_a\\to L_a}" + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "text", + "content": " denotes the reconstruction probabilities in " + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "inline_equation", + "content": "L_{a}" + }, + { + "bbox": [ + 302, + 665, + 524, + 692 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 692, + 497, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 692, + 497, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 692, + 497, + 704 + ], + "type": "text", + "content": "The final objective function of UNMT is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 373, + 712, + 525, + 725 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 712, + 525, + 725 + ], + "spans": [ + { + "bbox": [ + 373, + 712, + 525, + 725 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {B} + \\lambda_ {D} \\mathcal {L} _ {D} \\tag {3}", + "image_path": "672d4c8a75a3562c9ba8d022dc1e5bab725f27ec9959adacc9d910ceacc9201f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "inline_equation", + "content": "\\lambda_{D}" + }, + { + "bbox": [ + 302, + 733, + 526, + 772 + ], + "type": "text", + "content": " is a hyper-parameter that weights denoise autoencoder loss term and is gradually reduced during training." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12609" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 71, + 512, + 319 + ], + "blocks": [ + { + "bbox": [ + 81, + 71, + 512, + 319 + ], + "lines": [ + { + "bbox": [ + 81, + 71, + 512, + 319 + ], + "spans": [ + { + "bbox": [ + 81, + 71, + 512, + 319 + ], + "type": "image", + "image_path": "a27cac5cbce675af81757d07cbf656191f31d0d39c6ad2bffb58c815a869ec83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 328, + 493, + 342 + ], + "lines": [ + { + "bbox": [ + 98, + 328, + 493, + 342 + ], + "spans": [ + { + "bbox": [ + 98, + 328, + 493, + 342 + ], + "type": "text", + "content": "Figure 1: The process of converting an image-inserted sentence into an embedding representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 362, + 160, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 362, + 160, + 375 + ], + "spans": [ + { + "bbox": [ + 67, + 362, + 160, + 375 + ], + "type": "text", + "content": "2.2 Image Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": "In this section, we present how to incorporate word-level images into the UNMT process to tackle the issue of lexical confusion. The main concept behind our approach is that words can acquire information not only from contextual information but also from their corresponding images. Let " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "X_{i} = \\{x_{i}^{1}, x_{i}^{2}, \\dots, x_{i}^{K}\\}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": " denote a sentence from " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": " in language " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "x_{i}^{K}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": "-th word in sentence " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": ". We denote a lookup table consisting of word-image pairs as " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L_i} = \\{(word, I_{word})\\}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "inline_equation", + "content": "I_{word}" + }, + { + "bbox": [ + 67, + 380, + 290, + 541 + ], + "type": "text", + "content": " is an image set with one or multiple images that depict the meaning of the word." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": "As shown in Figure " + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "inline_equation", + "content": "1^{2}" + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": ", we segment an input sentence " + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": " at the word level and obtain the images by retrieving the words appearing in the lookup table " + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{L_i}" + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": ". Then, we concatenate the retrieved images to the corresponding words, resulting in an Image-Inserted Sentence denoted " + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "inline_equation", + "content": "IIS_{i}" + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": ". Each word can be concatenated with one or multiple images. Formally, We denote " + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "inline_equation", + "content": "IIS_{i}" + }, + { + "bbox": [ + 67, + 543, + 291, + 651 + ], + "type": "text", + "content": " as follow:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 660, + 290, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 660, + 290, + 681 + ], + "spans": [ + { + "bbox": [ + 93, + 660, + 290, + 681 + ], + "type": "interline_equation", + "content": "I I S _ {i} = \\left\\{x _ {i} ^ {1}, \\widehat {I} _ {x _ {i} ^ {1}}, x _ {i} ^ {2}, \\widehat {I} _ {x _ {i} ^ {2}}, \\dots , x _ {i} ^ {K}, \\widehat {I} _ {x _ {i} ^ {K}} \\right\\} \\quad (4)", + "image_path": "91ccf8fd72a67e9384247d3e6120cdf17479c64b9b947d37dde20711c9ba6ada.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "spans": [ + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "inline_equation", + "content": "\\widehat{I}_{x_i^K}" + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "text", + "content": " is the image subset of " + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "inline_equation", + "content": "I_{x_i^K}" + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "inline_equation", + "content": "\\widehat{I}_{x_i^K}" + }, + { + "bbox": [ + 67, + 690, + 290, + 733 + ], + "type": "text", + "content": " can be an empty set). We define this process of injecting images as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 383, + 377, + 525, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 377, + 525, + 391 + ], + "spans": [ + { + "bbox": [ + 383, + 377, + 525, + 391 + ], + "type": "interline_equation", + "content": "I I S = \\mathcal {I} (X) \\tag {5}", + "image_path": "eb1035b56b815af36e2b85eec8706f966cc44a7f068d55cabc9ba3da7247ba1b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 397, + 418, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 397, + 418, + 413 + ], + "spans": [ + { + "bbox": [ + 302, + 397, + 418, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 397, + 418, + 413 + ], + "type": "inline_equation", + "content": "IIS = \\{IIS_{i}\\}_{i = 1}^{M}" + }, + { + "bbox": [ + 302, + 397, + 418, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 413, + 525, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 413, + 525, + 466 + ], + "spans": [ + { + "bbox": [ + 302, + 413, + 525, + 466 + ], + "type": "text", + "content": "After obtaining the " + }, + { + "bbox": [ + 302, + 413, + 525, + 466 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 413, + 525, + 466 + ], + "type": "text", + "content": ", we need to integrate them into the UNMT model. To achieve this integration, we employ a modified embedding layer and a mask transformer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 476, + 468, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 476, + 468, + 490 + ], + "spans": [ + { + "bbox": [ + 302, + 476, + 468, + 490 + ], + "type": "text", + "content": "2.2.1 Modified Embedding Layer" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 493, + 525, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 493, + 525, + 548 + ], + "spans": [ + { + "bbox": [ + 301, + 493, + 525, + 548 + ], + "type": "text", + "content": "The first modification part is the embedding layer, which generally includes token embedding, position embedding, and, in some models, language embedding. An example can be seen in Figure 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": "Token Embedding. In the context of " + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": ", both textual words and images are present. Given that most current machine translation methods operate at the subword level, we proceed to tokenize the " + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": ", focusing solely on the textual component of " + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 556, + 526, + 773 + ], + "type": "text", + "content": ". For encoding textual tokens, we continue to utilize the embedding layer from the original text-only model. Regarding images, we leverage the CLIP (Radford et al., 2021) model as the image encoder to extract visual features. The CLIP model is a multi-modal model trained on a variety of (image, text) pairs, which possesses a powerful capability to represent images. To prevent negative impacts on the encoding of images during the UNMT process, we freeze the visual encoder. However, directly using these representations may lead" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 772 + ], + "type": "text", + "content": "2\"Blue\" is an adjective in the figure. In our image collection process, \"blue\" can also appear as a noun in other training samples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12610" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "content": "to a gap between the image encoding space and the semantic space in UNMT. Therefore, we employ an MLP layer to bridge this gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 118, + 290, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 118, + 290, + 241 + ], + "spans": [ + { + "bbox": [ + 67, + 118, + 290, + 241 + ], + "type": "text", + "content": "Position Embedding. Positional encoding is a crucial module that supplements token structural information. For textual information in " + }, + { + "bbox": [ + 67, + 118, + 290, + 241 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 67, + 118, + 290, + 241 + ], + "type": "text", + "content": ", we continue to encode the position in the original order of the input sentence to prevent interference with sentence structure learning when integrating images into the model. For images, we assign the positional encoding of the first subword in the tokenized result for the corresponding word." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 248, + 291, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 248, + 291, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 248, + 291, + 343 + ], + "type": "text", + "content": "Language Embedding. Some UNMT models incorporate language embedding to inform the model of the input language and the desired output language. For textual information, we continue to utilize the original language encoding. However, for image information, as it does not belong to any language, we introduce a new image identifier." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 349, + 193, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 349, + 193, + 361 + ], + "spans": [ + { + "bbox": [ + 68, + 349, + 193, + 361 + ], + "type": "text", + "content": "2.2.2 Mask Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 364, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 364, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 67, + 364, + 290, + 539 + ], + "type": "text", + "content": "Despite modifying the original embedding layer to allow " + }, + { + "bbox": [ + 67, + 364, + 290, + 539 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 67, + 364, + 290, + 539 + ], + "type": "text", + "content": " to be properly fed into the model, two issues still exist. Firstly, inserting images may affect other irrelevant words in the original sentence and potentially disrupt the structural relationships between them. Secondly, a single word-level image may correspond to multiple subwords, and during the annotation of position embedding, only the information of the first subword is utilized, resulting in a lack of relationship between the image and the remaining subwords it corresponds to. To address these issues, we introduce the method of mask transformer (Liu et al., 2020)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 540, + 291, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 540, + 291, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 540, + 291, + 608 + ], + "type": "text", + "content": "We create a visible matrix to shield the relationship between images and other textual information in the sentence and highlight the relationship between the image and its corresponding subwords. The visible matrix " + }, + { + "bbox": [ + 67, + 540, + 291, + 608 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 540, + 291, + 608 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 612, + 290, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 612, + 290, + 646 + ], + "spans": [ + { + "bbox": [ + 121, + 612, + 290, + 646 + ], + "type": "interline_equation", + "content": "M _ {i j} = \\left\\{ \\begin{array}{l l} - \\infty & w _ {p} \\bullet w _ {q} \\\\ 0 & w _ {p} \\circ w _ {q}. \\end{array} \\right. \\tag {6}", + "image_path": "4d51475be8015483a08cbef3ed3ab460d86b401eaef6b6ac51b07f014120df59.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{p}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{q}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " are elements in " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "IIS_{i}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{p} \\circ w_{q}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " denotes that " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{p}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{q}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " are visible to each other, while " + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "inline_equation", + "content": "w_{p} \\bullet w_{q}" + }, + { + "bbox": [ + 67, + 651, + 290, + 706 + ], + "type": "text", + "content": " is not. An example of a visible matrix is depicted in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 291, + 772 + ], + "type": "text", + "content": "The visible matrix can be added to the self-attention module to mask certain attention calculations. Specifically, the new masked self-attention is:" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 349, + 69, + 480, + 201 + ], + "blocks": [ + { + "bbox": [ + 349, + 69, + 480, + 201 + ], + "lines": [ + { + "bbox": [ + 349, + 69, + 480, + 201 + ], + "spans": [ + { + "bbox": [ + 349, + 69, + 480, + 201 + ], + "type": "image", + "image_path": "21012451b1d6a291a0d8c1850e6f6419c4639d9dc4addb6ec6ebd30cad1dbd66.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "lines": [ + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "type": "text", + "content": "Figure 2: An example of the visible matrix corresponding to the " + }, + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "type": "text", + "content": " in Figure 1. The red numerical labels indicate the absolute position of the images in the " + }, + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "type": "inline_equation", + "content": "IIS" + }, + { + "bbox": [ + 302, + 210, + 527, + 269 + ], + "type": "text", + "content": ". In this matrix, black represents \"visible\" and white represents \"invisible\"." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 298, + 525, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 298, + 525, + 331 + ], + "spans": [ + { + "bbox": [ + 344, + 298, + 525, + 331 + ], + "type": "interline_equation", + "content": "S ^ {i} = \\operatorname {s o f t m a x} \\left(\\frac {Q ^ {i} K ^ {i T} + M}{\\sqrt {d}}\\right) \\tag {7}", + "image_path": "da86572415f70466fb6c90099645d07d38648472775a61f992d65a909e7d5efa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "spans": [ + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "text", + "content": " are the query matrix and key matrix in the self-attention module. if two elements are invisible, the " + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 333, + 525, + 387 + ], + "type": "text", + "content": " will mask the attention score to 0, which means making no contribution to each other." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 396, + 414, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 396, + 414, + 409 + ], + "spans": [ + { + "bbox": [ + 302, + 396, + 414, + 409 + ], + "type": "text", + "content": "2.3 Training Strategy" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 413, + 525, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 413, + 525, + 467 + ], + "spans": [ + { + "bbox": [ + 302, + 413, + 525, + 467 + ], + "type": "text", + "content": "After initialization by cross-lingual pretrained language model, we train denoising autoencoder and iterative back-translation alternatively. The final loss of image fusion UNMT is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 479, + 525, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 479, + 525, + 496 + ], + "spans": [ + { + "bbox": [ + 370, + 479, + 525, + 496 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {I} = \\mathcal {L} _ {B} ^ {I} + \\lambda_ {D} \\mathcal {L} _ {D} ^ {I} \\tag {8}", + "image_path": "b4b273da7afb94822689c53db7d5f9e63cc6f9d5719cf1cd2974d40c40611b4f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "text", + "content": "specifically, " + }, + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_B^I" + }, + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D^I" + }, + { + "bbox": [ + 302, + 498, + 449, + 513 + ], + "type": "text", + "content": " refer to:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 309, + 519, + 525, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 519, + 525, + 592 + ], + "spans": [ + { + "bbox": [ + 309, + 519, + 525, + 592 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {B} ^ {I} = \\sum_ {i = 1} ^ {M} - l o g P _ {L _ {2} \\rightarrow L _ {1}} (X _ {i} | \\mathcal {I} (S _ {L _ {1} \\rightarrow L _ {2}} (\\mathcal {I} (X _ {i}), \\theta)), \\theta) \\\\ + \\sum_ {j = 1} ^ {N} - \\log P _ {L _ {1} \\rightarrow L _ {2}} \\left(Y _ {j} \\mid \\mathcal {I} \\left(S _ {L _ {2} \\rightarrow L _ {1}} \\left(\\mathcal {I} \\left(Y _ {j}\\right), \\theta\\right)\\right), \\theta\\right) \\tag {9} \\\\ \\end{array}", + "image_path": "7117c760d2bcf90c07c0b61fb39f4577e74a4ba1816bec844712a16949bf2ec8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 333, + 610, + 525, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 610, + 525, + 672 + ], + "spans": [ + { + "bbox": [ + 333, + 610, + 525, + 672 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D} ^ {I} = \\sum_ {i = 1} ^ {M} - \\log P _ {L _ {1} \\rightarrow L _ {1}} \\left(X _ {i} \\mid \\mathcal {I} \\left(N \\left(X _ {i}\\right)\\right), \\theta\\right) \\tag {10} \\\\ + \\sum_ {j = 1} ^ {N} - l o g P _ {L _ {2} \\rightarrow L _ {2}} (Y _ {j} | \\mathcal {I} (N (Y _ {j})), \\theta) \\\\ \\end{array}", + "image_path": "ecd51b6a185e34aa523b65eb8bbd8df615abd844bc38dd7b839d2d442b1f2783.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 525, + 772 + ], + "type": "text", + "content": "During the denoising autoencoder process, we first add noise to the original sentence and then integrate the images. During the iterative backtranslation process, for the first round of synthetic data generation, we do not add image information to achieve a smoother transition because the pretrained language model did not include images." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "12611" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 68, + 274, + 184 + ], + "blocks": [ + { + "bbox": [ + 85, + 68, + 274, + 184 + ], + "lines": [ + { + "bbox": [ + 85, + 68, + 274, + 184 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 274, + 184 + ], + "type": "table", + "html": "
Language#Word#ImageAverage
En4,97545,2289.09
Fr4,64636,9277.95
De10,38485,4198.23
Cs10,70672,9426.81
Zh10,80585,1447.88
Total41,516325,6607.84
", + "image_path": "900fb12f415c4082a747b2ad331710aef679470e06127327f56b460ceadcf698.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 191, + 289, + 214 + ], + "lines": [ + { + "bbox": [ + 67, + 191, + 289, + 214 + ], + "spans": [ + { + "bbox": [ + 67, + 191, + 289, + 214 + ], + "type": "text", + "content": "Table 1: The statistics of our constructed Word-Image dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 241, + 155, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 241, + 155, + 255 + ], + "spans": [ + { + "bbox": [ + 68, + 241, + 155, + 255 + ], + "type": "text", + "content": "3 Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 267, + 290, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 267, + 290, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 290, + 321 + ], + "type": "text", + "content": "In this section, we provide a detailed description of the datasets and experimental setup employed in our study. Subsequently, we perform an in-depth analysis of the results obtained." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 337, + 136, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 337, + 136, + 349 + ], + "spans": [ + { + "bbox": [ + 67, + 337, + 136, + 349 + ], + "type": "text", + "content": "3.1 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 359, + 291, + 724 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 359, + 291, + 724 + ], + "spans": [ + { + "bbox": [ + 67, + 359, + 291, + 724 + ], + "type": "text", + "content": "Word-Image Dataset. In our approach, the most crucial component of the experiment data is the image dataset at the word level. Since there is no off-the-shelf word-level image dataset, we build one for performance evaluation. We collect these images through the following steps: word extraction, image crawling, and image filtering. Firstly, extract the nouns from the monolingual dataset using part-of-speech analysis tools including Spacy3 and Stanford CoreNLP4. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images5 to crawl images for these nouns, limiting the image size to between 200-by-200 and 10,000-by-10,000 pixels and retrieving 10 to 15 images per noun. The image format is restricted to JPG, JPEG, and PNG. Finally, filter the crawled images. We consider all the images collected for a single noun as a cluster and encode them using the CLIP model. We then calculate pairwise similarities between all images and remove those with an average similarity below a threshold " + }, + { + "bbox": [ + 67, + 359, + 291, + 724 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 67, + 359, + 291, + 724 + ], + "type": "text", + "content": ", which we set to 0.6, as they are considered anomalous. The statistics information of the Word-Image dataset can be seen in Table 1. Our dataset covers a wide range of common nouns and can be reused for various purposes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 527, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 527, + 408 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 527, + 408 + ], + "type": "text", + "content": "Monolingual Dataset. The monolingual datasets used in our experiments are Multi30k dataset (Elliott et al., 2016, 2017; Barrault et al., 2018), LDC dataset and WMT News crawl dataset. The Multi30k dataset contains images and their captions in four languages: English(En), French(Fr), Germany(De), and Czech(Cs). Every language has 29000 sentences for training and 1014 for validation. It is worth noting that we only utilized the lingual data from Multi30k as the images are sentence-level. To prevent the model from encountering bilingual corpora, we randomly split the training and validation sets of the Multi30k dataset into two equal parts. Each part of the training set contains 14,500 sentences, and each part of the validation set contains 507 sentences. We report results in all test sets in Multi30k including Test2016, Test2017, Test2018, and MsCOCO. For the Cs-En task, only Test2016 and Test2018 are available. Furthermore, we expanded the dataset by incorporating monolingual data from WMT 2008-2022 News crawl and LDC. We also report our results on the common news domain test set, which includes WMT Newstest 2014 for En-Fr, Newstest 2021 for En-Cs, and NIST06 for En-Zh." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 418, + 430, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 418, + 430, + 431 + ], + "spans": [ + { + "bbox": [ + 302, + 418, + 430, + 431 + ], + "type": "text", + "content": "3.2 Experimental Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 435, + 527, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 435, + 527, + 652 + ], + "spans": [ + { + "bbox": [ + 302, + 435, + 527, + 652 + ], + "type": "text", + "content": "Baseline Models. We compare recent unsupervised text-only and multimodal MT baselines listed in the following: (1) XLM (Conneau and Lample, 2019) uses a masked language model to train a cross-lingual language model to initialize. (2) UMMT (Su et al., 2019) uses visual features for denoising autoencoder and back-translation. (3) PVP (Huang et al., 2020) employs multimodal back-translation and features pseudo visual pivoting. (4) UNMT-CS (Jones et al., 2023) uses the codeswitching method, where words in the source sentence are swapped out for their corresponding translation from bilingual dictionaries, to solve the lexical confusion issue in UNMT. It is worth noting that both UMMT and PVP datasets utilize sentence-level images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 660, + 527, + 755 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 660, + 527, + 755 + ], + "spans": [ + { + "bbox": [ + 302, + 660, + 527, + 755 + ], + "type": "text", + "content": "System Settings. We conducted our experiments based on the XLM codebase6. The model architecture employed is a 6-layers, 8-heads transformer, with a hidden layer representation dimension of 1024. The epoch size is 100K and the tokens-per-batch is 2000. we use Adam optimizer and 2000 warm-up updates. The learning rate is 1e-5. For" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 761, + 476, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 761, + 476, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 761, + 476, + 772 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 315, + 761, + 476, + 772 + ], + "type": "text", + "content": "https://github.com/facebookresearch/XLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 80, + 738, + 215, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 738, + 215, + 750 + ], + "spans": [ + { + "bbox": [ + 80, + 738, + 215, + 750 + ], + "type": "text", + "content": "3https://github.com/explosion/spaCy" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 750, + 233, + 761 + ], + "type": "text", + "content": "4https://github.com/stanfordnlp/CoreNLP" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 81, + 761, + 182, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 182, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 182, + 772 + ], + "type": "text", + "content": "5https://images.google.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12612" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 119, + 68, + 476, + 297 + ], + "blocks": [ + { + "bbox": [ + 119, + 68, + 476, + 297 + ], + "lines": [ + { + "bbox": [ + 119, + 68, + 476, + 297 + ], + "spans": [ + { + "bbox": [ + 119, + 68, + 476, + 297 + ], + "type": "table", + "html": "
En-FrEn-DeEn-Cs
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT39.80*---23.50*-----
PVP52.30*---33.90*-----
XLM53.0846.8235.3143.0931.8126.8325.6222.9525.1420.15
UNMT-CS54.2446.3036.0643.4933.4327.5126.6724.2726.4622.71
Ours54.0746.9536.2144.4433.0629.1227.3724.7526.3422.88
Δ+0.99+0.13+0.90+1.35+1.25+2.29+1.75+1.80+1.20+2.73
Fr-EnDe-EnCs-En
ApproachF16F17F18C17F16F17F18C17F16F18
UMMT40.50*---26.40*-----
PVP46.00*---31.60*-----
XLM48.0442.1437.0343.5837.5133.2130.7327.9833.1330.29
UNMT-CS49.2842.8637.1944.3938.8334.4831.4428.0733.9131.33
Ours49.1043.0137.4144.5740.3235.2032.7029.8134.2931.56
Δ+1.06+0.87+0.38+0.99+2.81+1.99+1.97+1.83+1.16+1.27
", + "image_path": "d6a6ae708a278decf5de5cf56e71852a2703cd247c68a10ef75a7c3b6b8a9efa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 305, + 525, + 341 + ], + "lines": [ + { + "bbox": [ + 67, + 305, + 525, + 341 + ], + "spans": [ + { + "bbox": [ + 67, + 305, + 525, + 341 + ], + "type": "text", + "content": "Table 2: BLEU scores on Multi30k En-Fr, En-De, and En-Cs UNMT tasks. F18, F17, F16, and C17 respectively denote the Flickr2018, Flickr2017, Flickr2016, and COCO2017 test sets in Multi30k. * is the results reported from previous papers. Δ represents the improvement of our method over the XLM method in terms of BLEU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 362, + 290, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 362, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 67, + 362, + 290, + 401 + ], + "type": "text", + "content": "evaluation, we use 4-gram BLEU (Papineni et al., 2002) scores by multi-bleu.pl in Moses7. We train all models on 4 NVIDIA 32GB V100 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 403, + 291, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 403, + 291, + 673 + ], + "spans": [ + { + "bbox": [ + 69, + 403, + 291, + 673 + ], + "type": "text", + "content": "During training, we first train the models on a 5M WMT News Crawl dataset for 70 epochs and subsequently continued training on the Multi30k dataset for another 70 epochs due to its relatively small size. As for the CLIP model, we use ViT-L/14@336px. When incorporating images, we randomly select two images from the corresponding image set for each word and concatenate them to the text. We only incorporate images during the iterative back-translation and denoising autoencoder stages, so our method can be used in conjunction with other pre-training enhancement methods. Regarding the cross-lingual models used for initialization, we directly employ high-quality open-source cross-lingual models in experiments involving En-Fr, En-De, and En-Zh. However, for the En-Cs, we train a model using WMT monolingual dataset, with 10 million English sentences and 10 million Czech sentences, due to the unavailability of any open-source models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 682, + 158, + 694 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 682, + 158, + 694 + ], + "spans": [ + { + "bbox": [ + 67, + 682, + 158, + 694 + ], + "type": "text", + "content": "3.3 Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 699, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 699, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 699, + 291, + 754 + ], + "type": "text", + "content": "Table 2 shows the main results on the Multi30k test set, including translating among English, French, German, and Czech. Comparing the results of XLM and our method, we observe that adding im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 301, + 362, + 526, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 362, + 526, + 510 + ], + "spans": [ + { + "bbox": [ + 301, + 362, + 526, + 510 + ], + "type": "text", + "content": "ages leads to performance improvements in all six translation directions, with a maximum increase of " + }, + { + "bbox": [ + 301, + 362, + 526, + 510 + ], + "type": "inline_equation", + "content": "+2.81" + }, + { + "bbox": [ + 301, + 362, + 526, + 510 + ], + "type": "text", + "content": " BLEU points, indicating the effectiveness of our approach. UNMT-CS is a method that utilizes bilingual dictionaries to address the problem of lexical confusion via code-switching. By comparing our method with the UNMT-CS approach, we observed comparable results, and in some directions, our method even outperformed the code-switching approach, indicating that ours is more efficient and cost-effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 513, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 513, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 513, + 526, + 687 + ], + "type": "text", + "content": "When compared to the best-performing method that incorporates sentence-level images, our approach still achieves comparable or better results in most translation directions. This is despite the matching degree between the Multi30k monolingual data and word-level images being slightly lower than that of sentence-level images, as the text data in Multi30k is generated based on the sentence-level images. In some directions, our approach even outperforms the previous best methods by up to " + }, + { + "bbox": [ + 302, + 513, + 526, + 687 + ], + "type": "inline_equation", + "content": "+4.22" + }, + { + "bbox": [ + 302, + 513, + 526, + 687 + ], + "type": "text", + "content": " BLEU points (as shown in Table 2, where our method is compared to the first two methods of the table)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 702, + 368, + 715 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 702, + 368, + 715 + ], + "spans": [ + { + "bbox": [ + 302, + 702, + 368, + 715 + ], + "type": "text", + "content": "4 Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 726, + 383, + 740 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 726, + 383, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 726, + 383, + 740 + ], + "type": "text", + "content": "4.1 Case Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "content": "In Table 3, we provide several examples of translations, where we observe that the XLM method re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 247, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 247, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 247, + 772 + ], + "type": "text", + "content": "7https://github.com/moses-smt/mosesdecoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12613" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 510, + 236 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 510, + 236 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 510, + 236 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 510, + 236 + ], + "type": "table", + "html": "
GT-DEMenschen an der Seitenlinie bei einem Fußballspiel.
GT-ENPeople on the sideline of a soccer match.
XLMPeople at the plate in a baseball game.
OursPeople at the stands in a soccer game.
GT-DEEin kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.
GT-ENA young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.
XLMA small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.
OursA small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.
GT-FRUn homme avec une veste noire, une casquette à carreaux et un pantalon rayé noir et blanc joue de la guitare électrique sur une scène avec un chanteur et un autre guitariste en arrêtère-plan.
GT-ENA man in a black jacket and checkered hat wearing black and white striped pants plays an electric guitar on a stage with a singer and another guitar player in the background.
XLMA man with a black jacket, striped headband and light brown and white shirt plays an electric guitar on a stage with a singer and another guitarist in the background.
OursA man with a black jacket, checkered cap and black and white striped pants plays an electric guitar on a stage with a singer and another guitarist in the background.
", + "image_path": "f47789201e1f12da963f97196b02a4075cb11c6db020d93e0f702a1ea6bd03dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 243, + 525, + 268 + ], + "lines": [ + { + "bbox": [ + 67, + 243, + 525, + 268 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 525, + 268 + ], + "type": "text", + "content": "Table 3: Qualitative results of the proposed model. GT denotes ground truth. Words highlighted in red are affected by the problem of lexical confusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 289, + 286, + 421 + ], + "blocks": [ + { + "bbox": [ + 70, + 289, + 286, + 421 + ], + "lines": [ + { + "bbox": [ + 70, + 289, + 286, + 421 + ], + "spans": [ + { + "bbox": [ + 70, + 289, + 286, + 421 + ], + "type": "image", + "image_path": "7040d306fc6f2b72cc36bf158842215c74fd5849719144a37c7ddd37e6629ac2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 428, + 289, + 451 + ], + "lines": [ + { + "bbox": [ + 67, + 428, + 289, + 451 + ], + "spans": [ + { + "bbox": [ + 67, + 428, + 289, + 451 + ], + "type": "text", + "content": "Figure 3: 1-gram precision scores on Flickr2016 and Flickr2017 data sets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 475, + 291, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 475, + 291, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 475, + 291, + 663 + ], + "type": "text", + "content": "sults in some lexical confusion errors. For instance, in Example 1, \"Fußballspiel\" is translated to \"baseball\" instead of \"soccer\", and in Example 2, \"orangefarbenen\" is translated to \"orange-red\" instead of \"orange\". The representation of these words is similar, which makes it difficult for the model to differentiate between them. However, when we incorporate image information, the words obtain additional information from the image representation, which helps alleviate this problem to some extent. In Table 3, we can see that these words are correctly translated when we add image information. Therefore, our approach can effectively address the issue of lexical confusion in UNMT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 673, + 209, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 673, + 209, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 673, + 209, + 687 + ], + "type": "text", + "content": "4.2 1-gram Precision Scores" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 290, + 772 + ], + "type": "text", + "content": "To better observe how our method can alleviate the issue of lexical confusion, we report the 1-gram precision scores in Figure 3. The accuracy of the model's word translations can be reflected intuitively in the 1-gram precision scores metric. We report the results of our approach with and without" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 289, + 526, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 289, + 526, + 370 + ], + "spans": [ + { + "bbox": [ + 302, + 289, + 526, + 370 + ], + "type": "text", + "content": "incorporating images for various translation directions on the Flickr2016 and Flickr2017 test sets and find that our method with image incorporation outperforms the method without images in all directions. This demonstrates that our method can indeed provide more accurate word translations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 403, + 521, + 416 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 403, + 521, + 416 + ], + "spans": [ + { + "bbox": [ + 302, + 403, + 521, + 416 + ], + "type": "text", + "content": "4.3 Performance on WMT and LDC Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 301, + 435, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 435, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 435, + 525, + 772 + ], + "type": "text", + "content": "Not all sentences have a corresponding image that perfectly matches the text. On the other hand, generally, each sentence has some words that can be matched with images, making the use of word-level images more flexible and less costly to collect. Hence, we also conduct experiments on a more common news dataset. We conduct experiments on En-Fr, En-Cs, and En-Zh translation directions using the WMT 5M monolingual data for En-Fr and En-Cs, and the LDC dataset for En-Zh. We split the LDC bilingual data into two non-overlapping parts, each containing 0.75 million monolingual sentences. The news dataset is more abstract in content compared to the Multi30k dataset, and it is difficult to find an image that perfectly matches a sentence. This makes it challenging or almost impossible to apply previous methods that rely on sentence-level images to this dataset. As shown in Table 4, our method can be applied to the news dataset, and it shows improvements compared to the method that only uses text. The improvement on the WMT dataset was not significant, and we speculate that this may be due to the sufficient size of the WMT dataset and the lower frequency of noun entities with images compared to Multi30k." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12614" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 69, + 289, + 121 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 289, + 121 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 289, + 121 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 289, + 121 + ], + "type": "table", + "html": "
ApproachZh-EnEn-ZhCs-EnEn-CsFr-EnEn-Fr
XLM8.437.9125.1417.7933.7335.77
Our Method9.879.3625.5118.2434.1436.46
Δ+1.44+1.45+0.37+0.45+0.41+0.69
", + "image_path": "0c3e9a0658dbe8c5909009083b0051e563e27a0be14fdeb736830d74520ca6d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 129, + 281, + 140 + ], + "lines": [ + { + "bbox": [ + 75, + 129, + 281, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 129, + 281, + 140 + ], + "type": "text", + "content": "Table 4: BLUE scores on WMT and LDC test sets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 71, + 155, + 285, + 287 + ], + "blocks": [ + { + "bbox": [ + 71, + 155, + 285, + 287 + ], + "lines": [ + { + "bbox": [ + 71, + 155, + 285, + 287 + ], + "spans": [ + { + "bbox": [ + 71, + 155, + 285, + 287 + ], + "type": "image", + "image_path": "008d6c4d02a5b500a1f50daca9f812157920c343553bc8a301e17dbd4d20fa7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 295, + 290, + 321 + ], + "lines": [ + { + "bbox": [ + 67, + 295, + 290, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 295, + 290, + 321 + ], + "type": "text", + "content": "Figure 4: BLEU scores on En-De with different ways and numbers of adding image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 342, + 186, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 342, + 186, + 355 + ], + "spans": [ + { + "bbox": [ + 67, + 342, + 186, + 355 + ], + "type": "text", + "content": "4.4 Quantity of Images" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 359, + 291, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 359, + 291, + 658 + ], + "spans": [ + { + "bbox": [ + 67, + 359, + 291, + 658 + ], + "type": "text", + "content": "We are also curious about how the method of adding images and image quantity affects the performance. Therefore, we test several ways of adding images, including selecting the best one, the top two, randomly selecting one, randomly selecting two, and randomly selecting three. We treat all images corresponding to a single word as a class and selected the image with the highest average similarity to the other images in the same class as the best image. We report the average BLEU score of En-De on the Flickr2016 and Flickr2017 test sets as illustrated in figure 4. We find that the fixed-best method yielded lower performance compared to randomly adding images. This could be due to the limitation of the fixed-best method in capturing the full range of visual variations that may exist for a given word. Among all the random selection methods, we find that adding two images randomly yielded the best performance. This could be because adding three images may introduce more noise, while adding only one image may not provide enough randomness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 669, + 160, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 669, + 160, + 682 + ], + "spans": [ + { + "bbox": [ + 67, + 669, + 160, + 682 + ], + "type": "text", + "content": "5 Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 773 + ], + "type": "text", + "content": "Unsupervised Machine Translation. Unlike common machine translation (Sutskever et al., 2014; Vaswani et al., 2017; Feng et al., 2020), UNMT don't use parallel corpus. Recent advancements (Zhang et al., 2017; Artetxe et al., 2018; Lample et al., 2018a; Conneau and Lample, 2019;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": "Song et al., 2019; Han et al., 2021) in UNMT have achieved remarkable milestones. However, there are still certain issues associated with solely utilizing monolingual data for translation, such as domain mismatch problem (Marchisio et al., 2020; Kim et al., 2020), poor performance on distance language pairs (Kim et al., 2020; Chronopoulou et al., 2020; Sun et al., 2021), translationese (He et al., 2022), and lexical confusion (Bapna et al., 2022; Jones et al., 2023). Our work aims to address the issue of lexical confusion by utilizing word-level image data, thereby eliminating the need for costly bilingual dictionary annotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 254, + 526, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 254, + 526, + 565 + ], + "spans": [ + { + "bbox": [ + 302, + 254, + 526, + 565 + ], + "type": "text", + "content": "Multimodal Machine Translation. Multimodal machine translation involves incorporating image information into a machine translation model to enhance its performance. While most current multimodal machine translation systems (Caglayan et al., 2016; Yao and Wan, 2020; Yin et al., 2020; Caglayan et al., 2021; Zhang et al., 2020; Fang and Feng, 2022; Li et al., 2022) focus on supervised machine translation, it has been observed that visual information is generally ignored when parallel corpora are plentiful (Caglayan et al., 2019; Wu et al., 2021). Yang et al. (2022) proposed the use of image information in situations where parallel corpora are not abundant. On the other hand, Su et al. (2019) and Huang et al. (2020) utilized image information in an unsupervised setting. However, these approaches were based on sentence-level image information, which can be challenging to obtain since not all sentences have corresponding images. In contrast, word-level image information is easier to obtain and is more flexible and lightweight. Therefore, our work aims to use word-level image information to enhance the performance of UNMT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 576, + 474, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 576, + 474, + 588 + ], + "spans": [ + { + "bbox": [ + 302, + 576, + 474, + 588 + ], + "type": "text", + "content": "6 Conclusion and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 526, + 772 + ], + "type": "text", + "content": "In this paper, we propose a method to alleviate the problem of lexical confusion in unsupervised machine translation by utilizing word-level images and provide a concise yet potent framework for integrating them into an unsupervised neural machine translation model. Through extensive experiments, we demonstrate that our method can effectively mitigate the issue of word confusion and even outperform the use of costly bilingual dictionaries in some directions. Furthermore, we release an open-source word-image dataset that covers 5 languages and involves over 300,000 images in total. In the future, we will endeavor to explore the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12615" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 113 + ], + "type": "text", + "content": "potential of utilizing images for augmenting the translation quality of low-resource languages in a large-language-model-based translation system." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 121, + 131, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 121, + 131, + 133 + ], + "spans": [ + { + "bbox": [ + 68, + 121, + 131, + 133 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 142, + 291, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 291, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 291, + 346 + ], + "type": "text", + "content": "One limitation of our work is that we have not been able to effectively handle the issue of polysemy. The word-image dataset is obtained through direct web scraping of words, and the images collected may not all correspond to the same sense of the word. Moreover, when adding images, we have not been able to accurately select images that correspond to the intended meaning of the word in the sentence. However, we have attempted to alleviate this issue to some extent by randomly selecting images during the process of adding images, which has enabled the model to encounter the correct images to a certain degree. In the future, we plan to improve our approach by incorporating sentence-level information when inserting images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 355, + 158, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 355, + 158, + 367 + ], + "spans": [ + { + "bbox": [ + 68, + 355, + 158, + 367 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 376, + 290, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 376, + 290, + 484 + ], + "spans": [ + { + "bbox": [ + 67, + 376, + 290, + 484 + ], + "type": "text", + "content": "In our experimental endeavors, we have amassed a substantial collection of images, which naturally prompts discussions pertaining to these issues. It is imperative to clarify that we do not possess the copyright for these images. Instead, they will be exclusively available to researchers and educators seeking to utilize the dataset for non-commercial research and/or educational purposes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 495, + 170, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 495, + 170, + 508 + ], + "spans": [ + { + "bbox": [ + 68, + 495, + 170, + 508 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 516, + 291, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 516, + 291, + 664 + ], + "spans": [ + { + "bbox": [ + 67, + 516, + 291, + 664 + ], + "type": "text", + "content": "Xiaocheng Feng is the corresponding author of this work. We thank the anonymous reviewers for their insightful comments. This work was supported by the National Key RD Program of China via grant No. 2021ZD0112905, National Natural Science Foundation of China (NSFC) via grant 62276078, the Key RD Program of Heilongjiang via grant 2022ZX01A32, the International Cooperation Project of PCLPCL2022D01 and the Fundamental Research Funds for the Central Universities (Grant No.HIT.OCEF.2023018)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "spans": [ + { + "bbox": [ + 68, + 687, + 127, + 699 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 705, + 291, + 772 + ], + "type": "text", + "content": "Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2018. Unsupervised neural machine translation. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 138 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 138 + ], + "type": "text", + "content": "Ankur Bapna, Isaac Caswell, Julia Kreutzer, Orhan Firat, Daan van Esch, Aditya Siddhant, Mengmeng Niu, Pallavi Baljekar, Xavier Garcia, Wolfgang Macherey, et al. 2022. Building machine translation systems for the next thousand languages. arXiv preprint arXiv:2205.03983." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 150, + 526, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 150, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 150, + 526, + 216 + ], + "type": "text", + "content": "Loic Barrault, Fethi Bougares, Lucia Specia, Chiraag Lala, Desmond Elliott, and Stella Frank. 2018. Findings of the third shared task on multimodal machine translation. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 304-323." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 227, + 526, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 526, + 261 + ], + "type": "text", + "content": "Ozan Caglayan, Loic Barrault, and Fethi Bougares. 2016. Multimodal attention for neural machine translation. arXiv preprint arXiv:1609.03976." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 273, + 526, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 526, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 526, + 361 + ], + "type": "text", + "content": "Ozan Caglayan, Menekse Kuyu, Mustafa Sercan Amac, Pranava Madhyastha, Erkut Erdem, Aykut Erdem, and Lucia Specia. 2021. Cross-lingual visual pretraining for multimodal machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 1317-1324, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 525, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 525, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 525, + 472 + ], + "type": "text", + "content": "Ozan Caglayan, Pranava Madhyastha, Lucia Specia, and Loic Barrault. 2019. Probing the need for visual context in multimodal machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4159-4170, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 525, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 525, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 525, + 560 + ], + "type": "text", + "content": "Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2020. Reusing a Pretrained Language Model on Languages with Limited Corpora for Unsupervised NMT. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2703-2711, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 572, + 525, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 572, + 525, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 525, + 606 + ], + "type": "text", + "content": "Alexis Conneau and Guillaume Lample. 2019. Crosslingual language model pretraining. Advances in neural information processing systems, 32." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 525, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 525, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 525, + 695 + ], + "type": "text", + "content": "Desmond Elliott, Stella Frank, Loic Barrault, Fethi Bougares, and Lucia Specia. 2017. Findings of the second shared task on multimodal machine translation and multilingual image description. In Proceedings of the Second Conference on Machine Translation, pages 215-233, Copenhagen, Denmark. Association for Computational Linguistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "text", + "content": "Desmond Elliott, Stella Frank, Khalil Sima'an, and Lucia Specia. 2016. Multi30K: Multilingual English-German image descriptions. In Proceedings of the 5th Workshop on Vision and Language, pages 70-74, Berlin, Germany. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "12616" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 291, + 139 + ], + "type": "text", + "content": "Qingkai Fang and Yang Feng. 2022. Neural machine translation with phrase-level universal visual representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5687-5698, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 146, + 290, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 146, + 290, + 201 + ], + "spans": [ + { + "bbox": [ + 69, + 146, + 290, + 201 + ], + "type": "text", + "content": "Xiaocheng Feng, Zhangyin Feng, Wanlong Zhao, Bing Qin, and Ting Liu. 2020. Enhanced neural machine translation by joint decoding with word and posttagging sequences. Mobile Networks and Applications, 25:1722-1728." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 210, + 290, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 210, + 290, + 275 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 290, + 275 + ], + "type": "text", + "content": "Jesse Michael Han, Igor Babuschkin, Harrison Edwards, Arvind Neelakantan, Tao Xu, Stanislas Polu, Alex Ray, Pranav Shyam, Aditya Ramesh, Alec Radford, et al. 2021. Unsupervised neural machine translation with generative language models only. arXiv preprint arXiv:2110.05448." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 283, + 290, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 283, + 290, + 372 + ], + "spans": [ + { + "bbox": [ + 69, + 283, + 290, + 372 + ], + "type": "text", + "content": "Zhiwei He, Xing Wang, Rui Wang, Shuming Shi, and Zhaopeng Tu. 2022. Bridging the data gap between training and inference for unsupervised neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6611-6623, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 380, + 290, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 380, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 380, + 290, + 456 + ], + "type": "text", + "content": "Po-Yao Huang, Junjie Hu, Xiaojun Chang, and Alexander Hauptmann. 2020. Unsupervised multimodal neural machine translation with pseudo visual pivoting. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8226-8237, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 464, + 290, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 464, + 290, + 508 + ], + "spans": [ + { + "bbox": [ + 69, + 464, + 290, + 508 + ], + "type": "text", + "content": "Alex Jones, Isaac Caswell, Ishank Saxena, and Orhan Firat. 2023. Bilex rx: Lexical data augmentation for massively multilingual machine translation. CoRR, abs/2303.15265." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 516, + 290, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 516, + 290, + 583 + ], + "spans": [ + { + "bbox": [ + 69, + 516, + 290, + 583 + ], + "type": "text", + "content": "Yunsu Kim, Miguel Graça, and Hermann Ney. 2020. When and why is unsupervised neural machine translation useless? In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation, pages 35-44, Lisboa, Portugal. European Association for Machine Translation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 591, + 290, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 591, + 290, + 645 + ], + "spans": [ + { + "bbox": [ + 69, + 591, + 290, + 645 + ], + "type": "text", + "content": "Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018a. Unsupervised machine translation using monolingual corpora only. In International Conference on Learning Representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 653, + 290, + 730 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 653, + 290, + 730 + ], + "spans": [ + { + "bbox": [ + 69, + 653, + 290, + 730 + ], + "type": "text", + "content": "Guillaume Lample, Alexis Conneau, Marc'Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2018b. Word translation without parallel data. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 738, + 290, + 772 + ], + "type": "text", + "content": "Guillaume Lample, Myle Ott, Alexis Conneau, Ludovic Denoyer, and Marc'Aurelio Ranzato. 2018c. Phrase-based & neural unsupervised machine translation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 525, + 117 + ], + "type": "text", + "content": "In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5039-5049, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 126, + 525, + 204 + ], + "type": "text", + "content": "Bei Li, Chuanhao Lv, Zefan Zhou, Tao Zhou, Tong Xiao, Anxiang Ma, and JingBo Zhu. 2022. On vision features in multimodal machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6327-6337, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 213, + 525, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 525, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 525, + 269 + ], + "type": "text", + "content": "Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang. 2020. K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 2901-2908." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 280, + 525, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 525, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 525, + 335 + ], + "type": "text", + "content": "Kelly Marchisio, Kevin Duh, and Philipp Koehn. 2020. When does unsupervised machine translation work? In Proceedings of the Fifth Conference on Machine Translation, pages 571-583, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 344, + 525, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 525, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 525, + 423 + ], + "type": "text", + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 432, + 525, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 525, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 525, + 499 + ], + "type": "text", + "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 508, + 525, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 525, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 525, + 564 + ], + "type": "text", + "content": "Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. Mass: Masked sequence to sequence pretraining for language generation. In International Conference on Machine Learning, pages 5926-5936. PMLR." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 574, + 525, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 525, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 525, + 630 + ], + "type": "text", + "content": "Yuanhang Su, Kai Fan, Nguyen Bach, C-C Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10482-10491." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 640, + 525, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 525, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 525, + 717 + ], + "type": "text", + "content": "Haipeng Sun, Rui Wang, Masao Utiyama, Benjamin Marie, Kehai Chen, Eiichiro Sumita, and Tiejun Zhao. 2021. Unsupervised neural machine translation for similar and distant language pairs: An empirical study. ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), 20(1):1-17." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 727, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 727, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 727, + 525, + 772 + ], + "type": "text", + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "12617" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 621 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 127 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 290, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 290, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 290, + 236 + ], + "type": "text", + "content": "Zhiyong Wu, Lingpeng Kong, Wei Bi, Xiang Li, and Ben Kao. 2021. Good for misconceived reasons: An empirical revisiting on the need for visual context in multimodal machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume I: Long Papers), pages 6153-6166, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 244, + 290, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 244, + 290, + 311 + ], + "spans": [ + { + "bbox": [ + 69, + 244, + 290, + 311 + ], + "type": "text", + "content": "Zhe Yang, Qingkai Fang, and Yang Feng. 2022. Lowresource neural machine translation with cross-modal alignment. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10134-10146, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 318, + 290, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 318, + 290, + 386 + ], + "spans": [ + { + "bbox": [ + 69, + 318, + 290, + 386 + ], + "type": "text", + "content": "Shaowei Yao and Xiaojun Wan. 2020. Multimodal transformer for multimodal machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4346-4350, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 394, + 290, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 394, + 290, + 471 + ], + "spans": [ + { + "bbox": [ + 69, + 394, + 290, + 471 + ], + "type": "text", + "content": "Yongjing Yin, Fandong Meng, Jinsong Su, Chulun Zhou, Zhengyuan Yang, Jie Zhou, and Jiebo Luo. 2020. A novel graph-based multi-modal fusion encoder for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3025-3035, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 480, + 290, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 480, + 290, + 557 + ], + "spans": [ + { + "bbox": [ + 69, + 480, + 290, + 557 + ], + "type": "text", + "content": "Meng Zhang, Yang Liu, Huanbo Luan, and Maosong Sun. 2017. Adversarial training for unsupervised bilingual lexicon induction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1959-1970, Vancouver, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 565, + 290, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 565, + 290, + 621 + ], + "spans": [ + { + "bbox": [ + 69, + 565, + 290, + 621 + ], + "type": "text", + "content": "Zhuosheng Zhang, Kehai Chen, Rui Wang, Masao Utiyama, Eiichiro Sumita, Zuchao Li, and Hai Zhao. 2020. Neural machine translation with universal visual representation. In International Conference on Learning Representations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 68, + 643, + 267, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 643, + 267, + 669 + ], + "spans": [ + { + "bbox": [ + 68, + 643, + 267, + 669 + ], + "type": "text", + "content": "A The Implementation Details of the XLM and UNMT-CS Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "content": "The XLM and UNMT-CS methods are re-implemented on the Multi30k dataset in our paper. During training, both these two methods are trained on the 5 million sentences from the WMT dataset for 70 epochs and subsequently continued training for 70 epochs on the Multi30k dataset. For the bilingual dictionary in codeswitching method, we use" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "text", + "content": "the Ground-truth bilingual dictionaries provided in MUSE" + }, + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "text", + "content": " (Lample et al., 2018b). When training, we substituted words in " + }, + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 71, + 525, + 166 + ], + "type": "text", + "content": " of the sentences, while the remaining sentences were left unchanged. All other parameters used in the XLM and UNMT-CS methods are identical to those used in our proposed method, as described in our paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 175, + 507, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 175, + 507, + 204 + ], + "spans": [ + { + "bbox": [ + 302, + 175, + 507, + 204 + ], + "type": "text", + "content": "B The Implicit Mapping Relationship Between Image Representations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 211, + 525, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 211, + 525, + 291 + ], + "spans": [ + { + "bbox": [ + 302, + 211, + 525, + 291 + ], + "type": "text", + "content": "In fact, our approach doesn't explicitly provide the model with a mapping relationship. Rather, we introduce the (word-image) pairs relationship into the model, enabling it to extract information from similar images and thus enhancing the mapping relationship between words." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 292, + 526, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 292, + 526, + 455 + ], + "spans": [ + { + "bbox": [ + 302, + 292, + 526, + 455 + ], + "type": "text", + "content": "We are also curious about the impact on results if the implicit mapping relationship between image representations was absent. Therefore, during the second stage of multi30k training, we replace all images from one language with randomly generated representations while keeping the other language's images unchanged. The En→De direction results for the Flickr2016 and Flickr2017 datasets experience a significant decline, decreasing from 40.32/35.20 to 39.17/34.29. This confirms the crucial importance of the mapping relationship between images for our approach." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 760, + 480, + 772 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 760, + 480, + 772 + ], + "spans": [ + { + "bbox": [ + 316, + 760, + 480, + 772 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 316, + 760, + 480, + 772 + ], + "type": "text", + "content": "https://github.com/facebookresearch/MUSE" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "12618" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_content_list.json b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f80a77f6c541313a1c9d628f9cd6e03591424f4b --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_content_list.json @@ -0,0 +1,2176 @@ +[ + { + "type": "text", + "text": "End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems", + "text_level": 1, + "bbox": [ + 211, + 87, + 786, + 129 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benjamin Towle1, Ke Zhou1,2", + "bbox": [ + 374, + 141, + 626, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Nottingham", + "bbox": [ + 393, + 158, + 608, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nokia Bell Labs", + "bbox": [ + 428, + 175, + 571, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{benjamin.towle, ke.zhou}@nottingham.ac.uk", + "bbox": [ + 250, + 193, + 754, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reply suggestion systems represent a staple component of many instant messaging and email systems. However, the requirement to produce sets of replies, rather than individual replies, makes the task poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. As a result, these system often rely on additional post-processing modules to diversify the outputs. However, these approaches are ultimately bottlenecked by the performance of the initial retriever, which in practice struggles to present a sufficiently diverse range of options to the downstream diversification module, leading to the suggestions being less relevant to the user. In this paper, we consider a novel approach that radically simplifies this pipeline through an autoregressive text-to-text retrieval model, that learns the smart reply task end-to-end from a dataset of (message, reply set) pairs obtained via bootstrapping. Empirical results show this method consistently outperforms a range of state-of-the-art baselines across three datasets, corresponding to a $5.1\\% - 17.9\\%$ improvement in relevance, and a $0.5\\% - 63.1\\%$ improvement in diversity compared to the best baseline approach. We make our code publicly available. $^{1}$", + "bbox": [ + 144, + 281, + 460, + 677 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 692, + 258, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reply suggestion, or smart reply (SR), systems are a staple component of many commercial applications such as Gmail, Skype, Outlook, Microsoft Teams, LinkedIn and Facebook Messenger. They help the user process chats and emails quicker by offering a set of canned replies which can be clicked without requiring manual typing. However, dialogue is known to be a one-to-many problem (Zhao et al., 2017; Towle and Zhou, 2022) – namely, for any given message, there are multiple possible replies. To reflect this uncertainty, systems should", + "bbox": [ + 112, + 718, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "present a diverse set of options to the user. For instance, given the message How are you?, an SR system could suggest: {I'm good; Ok; Not great}. Resultantly, the quality of a given reply depends not only on the message, but on the other replies in the reply set.", + "bbox": [ + 507, + 253, + 884, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Several prior works explore solutions to this problem such as removing near duplicates, penalising inter-reply similarity (Deb et al., 2019), clustering by intent (Henderson et al., 2017; Weng et al., 2019), learning latent variables (Deb et al., 2019, 2021), or model-based simulation (Towle and Zhou, 2023). However, these methods share a common design choice (Figure 1A): (1) a retrieval-based Matching model, which has learned a shared embedding space between messages and replies, returns a shortlist of top scoring replies; (2) this shortlist is refined through some diversification procedure to obtain the final reply set.", + "bbox": [ + 507, + 351, + 884, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, this assumes that the initial shortlist contains at least one good reply set. In practice, we find Matching models often search myopically, only retrieving candidates that are very similar to one another (Figure 1A). Thus, the chosen reply set often fails to reflect a diverse range of user intents, while latency constraints make more sophisticated diversification techniques or larger shortlists prohibitive (Deb et al., 2019).", + "bbox": [ + 507, + 563, + 885, + 707 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An intuitive, but – to the best of our knowledge – unexplored, solution to this problem is to conduct the retrieval autoregressively, with each reply conditioned on both the initial message and the previous replies in the set. Unfortunately, this approach encounters a second problem, namely, the lack of any datasets containing (message, reply set) pairs (Towle and Zhou, 2023). In practice, SR systems are trained on individual (message, reply) pairs obtained from conversation datasets, while the task of presenting multiple diverse replies to the user is outsourced to a separate diversification module.", + "bbox": [ + 505, + 709, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "'https://github.com/ BenjaminTowle/STAR", + "bbox": [ + 134, + 903, + 478, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7610", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7610-7622", + "bbox": [ + 216, + 945, + 779, + 957 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 277, + 957, + 719, + 971 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e50b983ca8bdf7fb68ad09f880c265e3c03c12402818f04d12ec540b52ff6509.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 115, + 84, + 880, + 186 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eff81584e09698eba5df1bd7394244c42071a8e9d49cc8fef307cc4b79ef12e4.jpg", + "image_caption": [ + "Figure 1: Previous methods [A] compared to our approach, STAR [B]. The example displayed is taken from the DailyDialog Test set, and compares the predictions of STAR with SimSR (Towle and Zhou, 2023), the next best method. Our method's suggestions present a diverse range of topics/intents to drive the conversation." + ], + "image_footnote": [], + "bbox": [ + 122, + 191, + 848, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To meet this dual need, we present both (i) a bootstrapping method for creating a high-quality dataset of (message, reply sets) and (ii) a novel autoregressive retrieval model which predicts sequences of replies. For solving (i), we observe how model-based planning algorithms have been known to serve as a powerful policy improvement operator (Silver et al., 2017; Schrittwieser et al., 2019), including in several NLP systems (Jang et al., 2020, 2021). Specifically, the outputs of a model-based planning algorithm can be used to bootstrap a SR system. Further, by conducting this planning offline we are able to leverage two key advantages: (1) the system is free of the latency constraints of online inference, and therefore can increase the search space coverage of the planning algorithm; (2) the system can leverage information that would not be available during inference, such as the ground-truth reply, to further guide the search process. For (ii) we unify both steps of the standard SR pipeline into a single end-to-end model, which mitigates the myopic search, and allows the model to learn to diversify its predictions in a principled way through gradient-based learning. To this end, we present STAR (Suggested replies with T5 and Autoregressive Retrieval) (Figure 1B). At a high level, STAR is a text-to-text model trained to output sequences of replies, where each reply is conditioned both on the initial message and the previous replies in the sequence. Concretely, we instantiate our method with the T5 pretrained model (Raffel et al., 2020). We expand T5's vocabulary by", + "bbox": [ + 112, + 400, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "treating each reply in the candidate pool as a novel token, and demonstrate a simple-yet-effective technique for initialising the new token embeddings, which leverages the model's existing semantic priors. Notably, by treating each reply as a token, we limit the number of autoregressive decoding steps required, keeping the model's efficiency comparable to other retrieval-based methods.", + "bbox": [ + 507, + 400, + 884, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirically, we evaluate our approach on three benchmarks: Reddit (Zhang et al., 2021), which is the only publicly-available SR benchmark, as well as PersonaChat (Zhang et al., 2018) and DailyDialog (Li et al., 2017) which are both widely-used in dialogue research more broadly (Zhang et al., 2019; Roller et al., 2020, inter alia), and share a similar conversational style with SR apps. We demonstrate superior performance over state-of-the-art baselines across all datasets, corresponding to a $5.1\\% - 17.9\\%$ improvement in relevance, and a $0.5\\% - 63.1\\%$ improvement in diversity compared to the best baseline approach. We further show comparable efficiency to previous methods, and perform a range of ablations to motivate our design choices.", + "bbox": [ + 507, + 539, + 882, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our key contributions are as follows: (1) an autoregressive retrieval architecture for sequentially predicting suggested replies; (2) a bootstrapping framework for generating high-quality data of (message, reply set) pairs; (3) detailed analysis of model behaviour and performance including a case study and ablation of key components.", + "bbox": [ + 507, + 806, + 884, + 919 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7611", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 114, + 83, + 270, + 98 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Smart reply The proprietary nature of data from email and chat applications has led several previous works to use publicly-available dialogue datasets (Zhang et al., 2021; Deb et al., 2021; Towle and Zhou, 2023) to benchmark SR methods, due to their analogous conversational nature. While early SR systems used generative models (Kannan et al., 2016), current production systems favour retrieval methods due to their greater controllability of outputs and superior latency (Deb et al., 2019). Increasing the diversity of reply suggestions is a key focus of previous work, which has been attempted by: (1) mapping replies to discrete intents / topics (Kannan et al., 2016; Chakravarthi and Pasternack, 2017; Weng et al., 2019); (2) re-weighting replies according to their similarity with other replies in the set (Carbonell and Goldstein-Stewart, 1998; Deb et al., 2019); (3) learning continuous latent variables to generate multiple queries (Zhao et al., 2017; Deb et al., 2019); (4) using model-based simulation to iteratively search and evaluate the relevance of candidate reply sets (Towle and Zhou, 2023). Our proposed method differs from all of these approaches in that our model learns to account for the interdependencies between replies through end-to-end backpropagation.", + "bbox": [ + 115, + 115, + 490, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autoregressive retrieval Integrating neural retrieval into the well-established paradigm of text-to-text models is of growing interest. Earlier work focuses on outputting a document ID given a query (Tay et al., 2022). Further work has extended this by considering alternate ways of representing the document IDs, such as through unique substrings (Bevilacqua et al., 2022). Another line of work has used autoregressive retrieval for the entity linking task (Cao et al., 2021a,b,c). There, the motivation is to reduce the large number of entities by relying on the text-to-text model's pre-existing vocabulary, rather than having to retrieve embeddings from a memory-intensive dense index. Our proposed method differs considerably from these previous works both in instantiation and motivation. Instantiation-wise, we generate multiple replies - critical to making this possible is the novel bootstrapping technique for creating the dataset of (message, reply set) pairs to train on. Motivation-wise, our goal is to be able to condition each reply on both the input message and previous replies in the set, enabling the model to learn to predict se", + "bbox": [ + 115, + 549, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quences of replies in a differentiable way.", + "bbox": [ + 509, + 84, + 821, + 99 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bootstrapping The idea of bootstrapping training data from limited resources has received significant recent interest in NLP, given the newly demonstrated few / zero-shot capabilities of many large language models (Brown et al., 2020). It has seen usage in few-shot shot text-classification (Schick and Schütze, 2021a), semantic similarity (Schick and Schütze, 2021b), tool-usage (Schick et al., 2023), retrieval (Izacard and Grave, 2021), sequence generation (He et al., 2020), and instruction-tuning (Honovich et al., 2023; Wang et al., 2023; Taori et al., 2023), amongst others. These techniques can also be seen as a form of knowledge distillation (Hinton et al., 2015), except that the training typically involves predicting the exact token targets, rather than using the soft probabilities of a teacher model. Although sometimes these techniques are used as an addition to supervised learning (He et al., 2020), in our case there are no datasets containing the ideal reply sets to suggest to the user. Instead, we must bootstrap this in a more unsupervised way, by transforming a dataset of (message, reply) pairs into a dataset of (message, reply set) pairs.", + "bbox": [ + 507, + 109, + 885, + 495 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 509, + 507, + 658, + 524 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first describe the model-based planning process used to obtain the bootstrapped dataset of (message, reply set) pairs (Section 3.1). Then, we show how the STAR architecture can be trained on this dataset (Section 3.2).", + "bbox": [ + 507, + 533, + 882, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Offline Dataset Creation", + "text_level": 1, + "bbox": [ + 509, + 625, + 747, + 639 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to transform a dialogue dataset $\\mathcal{D} = \\{(x,y)\\}$ of (message, reply) tuples, into a dataset $\\mathcal{D}^* = \\{(x,Y)\\}$ where $Y$ is the set of replies $\\{y_k\\}^K$ to be presented to the user. Algorithm 1 summarises this process. While our method is general to any arbitrary planning algorithm, we choose to instantiate our approach with a modified version of SimSR (Towle and Zhou, 2023), a recently released publicly available state-of-the-art SR method, that employs model-based simulation to predict reply sets. As the original algorithm was designed for online inference, we make several changes to benefit the offline nature of our version, and detail the full implementation below.", + "bbox": [ + 507, + 645, + 884, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The initial retrieval is conducted by a Matching model $\\Phi$ that separately encodes messages and replies into a shared latent space. Given an encoded", + "bbox": [ + 507, + 871, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7612", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Offline Dataset Creation. We use $N = 100$ , $M = 100$ , $\\alpha = 0.75$ and $\\lambda = 0.05$ as our default setting.", + "bbox": [ + 112, + 83, + 485, + 130 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input Matching model $\\Phi$ , message $x$ , precomputed reply vectors $\\{\\mathbf{y_r}\\}^R$ , number of candidates $N$ , number of simulations $M$ , final reply set size $K$ , query augmentation coefficient $\\alpha$ , redundancy penalty $\\lambda$ .", + "bbox": [ + 112, + 131, + 487, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output reply set $Y_{K}$", + "bbox": [ + 137, + 181, + 270, + 193 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {x}, \\mathbf {y} \\leftarrow \\Phi (x), \\Phi (y) \\\\ \\tilde {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 195, + 485, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\\\ Y _ {N} \\leftarrow N \\underset {r} {\\operatorname {a r g m a x}} (\\bar {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}}) \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 209, + 485, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {M} \\leftarrow M \\operatorname {a r g m a x} _ {\\mathbf {r}} (\\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}})\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 237, + 295, + 254 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q (y _ {m} | x) \\propto \\exp \\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {m}} \\\\ Y _ {G} \\leftarrow \\emptyset \\end{array} \\quad \\triangleright \\text {s o f t m a x o v e r t o p - M s c o r e s}\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 255, + 485, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c}\\textbf{for} k\\gets 0\\text{to} K\\textbf{do}\\\\ y_{k}\\leftarrow \\operatorname *{argmax}_{n}\\sum_{m}^{M}f(Y_{G}^{n},y_{m})q(y_{m}|x) - \\lambda f(Y_{G},y_{n}) \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 280, + 478, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {G} \\leftarrow Y _ {G} \\cup y _ {k}\n$$\n", + "text_format": "latex", + "bbox": [ + 131, + 317, + 250, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 131, + 329, + 179, + 338 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {K} \\leftarrow Y _ {G}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 338, + 196, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {r e t u r n} Y _ {K}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 351, + 198, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "message $\\mathbf{x} = \\Phi (x)$ , it retrieves the top $N$ candidates from a pool of pre-computed reply vectors $\\mathbf{Y}_{\\mathbf{R}} = \\{\\mathbf{y}_{\\mathbf{r}}\\}^{R}$ by combining their dot product similarity with a pre-computed language-model bias - a standard component of SR systems to downweight overly specific replies (Deb et al., 2019).", + "bbox": [ + 112, + 388, + 487, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {N} = N \\underset {r} {\\operatorname {a r g m a x}} \\left(\\mathbf {x} \\cdot \\mathbf {y} _ {\\mathbf {r}} + \\beta \\mathrm {L M} \\left(y _ {r}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 491, + 487, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then output the $K$ -tuple $Y_{i} \\in \\binom{Y_{N}}{K}$ that has the highest expected similarity with the human reply, according to some similarity function $f(\\cdot, \\cdot)$ .", + "bbox": [ + 112, + 521, + 487, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {i} {\\operatorname {a r g m a x}} \\mathbb {E} _ {y \\sim p (\\cdot | x)} \\left[ f (Y _ {i}, y) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 208, + 577, + 487, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the objective in SR is for at least one of the replies to be relevant, the similarity function is defined as a maximum over the sampled reply and each of the replies in the reply set, using term-level F1-score: $\\max_k\\mathrm{F1}(y_k,y)$", + "bbox": [ + 112, + 609, + 487, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume $y$ is sampled from the ground-truth human distribution $p(\\cdot |x)$ . As we do not have access to the true human distribution in practice, we instead use the same Matching model $q$ as a proxy for this, given it is trained on (message, reply) pairs. We then approximate the expectation by marginalising over the top- $M$ most likely replies:", + "bbox": [ + 112, + 694, + 487, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\approx \\underset {i} {\\operatorname {a r g m a x}} \\sum_ {m} ^ {M} f \\left(Y _ {i}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 813, + 487, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, it is intractable to evaluate every possible reply tuple, due to their combinatorial scaling. We therefore approximate this by greedily constructing the reply set one reply at a time. Formally,", + "bbox": [ + 112, + 854, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "let $Y_{G}$ be the set of currently selected replies, such that initially $Y_{G} = \\emptyset$ . Then, for each of $y_{n} \\in Y_{N}$ , we compute the expected similarity for the union of $Y_{G}$ and $y_{n}$ , termed $Y_{G}^{n} = Y_{G} \\cup y_{n}$ for brevity:", + "bbox": [ + 507, + 84, + 882, + 149 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 618, + 160, + 882, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We repeat this process for $K$ timesteps, each time appending the highest scoring reply to $Y_{G}$ , i.e. until $|Y_{G}| = K$ . Note that this greedy search process implicitly canonicalises the order of the replies, as selecting replies in this way causes them to be roughly ordered by individual message-reply relevance.", + "bbox": [ + 507, + 200, + 882, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1 Adjustments", + "text_level": 1, + "bbox": [ + 507, + 323, + 670, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scaling $N$ and $M$ The original SimSR algorithm was used only in an online setting (Towle and Zhou, 2023). Therefore, the size of the search parameters $N$ (number of replies in the shortlist) and $M$ (number of simulated user replies) is kept low (15 and 25 respectively in the original paper). As we only need to run this model offline however to obtain the dataset, we find setting $N$ and $M$ to much larger values improves relevance (we use 100 for both), enabling both a broader search (i.e. by increasing $N$ ) and a more accurate similarity function (i.e. by increasing $M$ ).", + "bbox": [ + 507, + 341, + 882, + 533 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Redundancy penalty Early testing showed that scaling the search parameters reduced diversity. We therefore introduce a redundancy penalty, which penalises the model for selecting replies that are similar to replies already in the set $Y_{G}$ . This is analogous to the inter-document similarity penalty used in the maximum marginal relevance IR (information retrieval) technique (Carbonell and Goldstein-Stewart, 1998).", + "bbox": [ + 507, + 543, + 882, + 687 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) - \\lambda f \\left(Y _ {G}, y _ {n}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 700, + 882, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Query augmentation Unlike during online inference, we also have access to the ground-truth reply $y$ when constructing the dataset. Previous work has found that models obtain greater representational capabilities when given access to posterior information (Paranjape et al., 2022; Towle and Zhou, 2022). We therefore use an augmented query to retrieve with the Matching model. This is obtained by interpolating between the message and ground-truth reply embeddings. This biases the model's predictions towards the observed ground-truth in", + "bbox": [ + 507, + 741, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7613", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the dataset, while still allowing it to benefit from its own learned distribution.", + "bbox": [ + 112, + 84, + 487, + 115 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} = \\alpha \\Phi (x) + (1 - \\alpha) \\Phi (y) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 211, + 128, + 487, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Proposed STAR Model", + "text_level": 1, + "bbox": [ + 114, + 159, + 341, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialise STAR with a T5-based text-to-text language model, which has previously been shown to be effective in autoregressive retrieval (Tay et al., 2022). While some autoregressive retrieval approaches identify their documents/replies through unique substrings (Bevilacqua et al., 2022) or constrained beam search (Cao et al., 2021b), we focus on approaches requiring only a limited number of autoregressive steps, to maintain competitive inference speeds to existing retrieval methods (Section 5.3). There are several alternatives for this such as treating each reply set as a unique token, or separately training on each (message, reply pair), but ultimately we opted for autoregressively treating each reply as a unique token in the vocabulary in order to exploit the compositionality of reply sets (Section 5.2 for performance comparison). Note that as the types of replies used in smart reply are usually quite short and concise, e.g. 'how are you', 'I'm fine thanks', 'yes, that's right' etc., systems in deployment only need to retrieve from a pool of $30\\mathrm{k}$ or so replies (Deb et al., 2019), in order to provide good coverage of possible user intents. As a result, we are able to keep the size of the vocabulary reasonable. Thus, our new vocabulary is defined as: $W_{tokens} \\cup W_{replies}$ . An obvious challenge to this approach is that by treating each reply as a previously unseen word, it removes any semantic priors the model might have about their meaning. To mitigate this, we employ a bag-of-words initialisation strategy. Hence, we define the embedding of the $t$ -th reply $E(y_{t})$ as the average over the embeddings of the individual words within $w_{n} \\in y_{t}$ .", + "bbox": [ + 112, + 181, + 489, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE \\left(y _ {t}\\right) = \\frac {1}{N} \\sum_ {n} ^ {N} E \\left(w _ {n}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 726, + 487, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, this ensures that the initial embeddings are close to the word embeddings of the original vocabulary, while also capturing some of the underlying semantics of the reply. We allow the weights to update during fine-tuning. Note that for T5 the output and input embedding layers share weights, and therefore this approach is used to initialise both layers. We train the model using cross-entropy loss to predict the next reply given the current sequence", + "bbox": [ + 112, + 774, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of replies and messages:", + "bbox": [ + 509, + 84, + 694, + 99 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N L L} = - \\sum_ {k} ^ {K} \\log p \\left(y _ {k} \\mid x, y _ {0}, \\dots , y _ {k - 1}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 116, + 882, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experimental Setup", + "text_level": 1, + "bbox": [ + 507, + 168, + 717, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Baselines", + "text_level": 1, + "bbox": [ + 507, + 198, + 628, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work has largely been closed-source and is therefore unavailable for direct comparison (Henderson et al., 2017; Weng et al., 2019; Deb et al., 2019). With the exception of SimSR, which has publicly available code $^{2}$ , we re-implement a variety of methods that cover the broad range of previous techniques. Due to its comparable size, all baselines apart from Seq2Seq are initialised with DistilBERT as the encoder backbone. These are summarised as follows:", + "bbox": [ + 507, + 221, + 884, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Seq2Seq is a generative encoder-decoder. While current production systems and the majority of related works use only retrieval models (Deb et al., 2019; Towle and Zhou, 2023), at least one related work includes a standard generative transformer as a baseline (Zhang et al., 2021), which we follow here. For maximum comparability with our method, we use the same t5-small model as a backbone. For each message, we sample $K$ responses independently.", + "bbox": [ + 507, + 397, + 884, + 557 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Matching represents the out-of-the-box encoder with no additional diversification strategy and was used as a baseline method by Zhang et al. (2021). It simply selects the top $K$ responses according to individual message-reply scores.", + "bbox": [ + 507, + 571, + 882, + 653 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Matching-Topic uses an out-of-the-box topic classifier to ensure no two replies share the same topic, similar to previous work (Henderson et al., 2017; Weng et al., 2019). The classifier is trained on Twitter (Antypas et al., 2022), due to their comparable short-form open-domain chat conversations.", + "bbox": [ + 507, + 665, + 882, + 778 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Maximum Marginal Relevance (MMR) (Carbonell and Goldstein-Stewart, 1998) is originally an IR technique, used in several previous SR works (Deb et al., 2019; Towle and Zhou, 2023), which re-weights reply scores as a linear combination of their message-reply and inter-reply similarity.", + "bbox": [ + 507, + 793, + 884, + 890 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/BenjaminTowle/SimSR", + "bbox": [ + 529, + 903, + 880, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7614", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/abceb955e66f2b977663f9a1bcdae84cb69b592dd458973a3ab5eb126d5b77dd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainValidTest|YR|
Reddit50k5k5k48k
PersonaChat66k8k8k64k
DailyDialog76k7k7k62k
", + "bbox": [ + 146, + 80, + 457, + 147 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1: Number of samples in the Train, Validation, Test sets and Candidate pool in the three datasets for evaluation. The Candidate pool comprises the Train set with duplicate responses removed.", + "bbox": [ + 112, + 155, + 489, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MCVAE (Deb et al., 2019) is a conditional variational autoencoder (Zhao et al., 2017) which learns to generate multiple query vectors from a single message embedding, representing the multiple possible reply intents. Candidates are scored via a voting process, whereby the $K$ most-selected replies are chosen.", + "bbox": [ + 112, + 240, + 489, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SimSR (Towle and Zhou, 2023) uses an iterative search and evaluation process to select possible reply sets and score them according to their expected similarity from a learned world model, which serves as a proxy for the user. To ensure comparability of SimSR with our method and the other baselines, we include the language-model bias in the scoring process (Equation 1), and also deduplicate the candidate pool. $^{3}$", + "bbox": [ + 112, + 365, + 489, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Datasets", + "text_level": 1, + "bbox": [ + 112, + 524, + 228, + 538 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed method across three datasets, summarised in Table 1. Below, we describe the datasets in more detail and motivate their inclusion. Note, other than Reddit, there are no publicly available SR datasets, due to their commercial nature (e.g. Henderson et al. (2017); Deb et al. (2019); Weng et al. (2019)). Therefore, we adopt several dialogue datasets, which is the closest alternative to conversations on proprietary chat applications.", + "bbox": [ + 112, + 546, + 489, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reddit (Zhang et al., 2021) was originally introduced for training multilingual SR systems, and is the only publicly available dataset specifically intended for SR purposes. As the original dataset is very large, we follow Towle and Zhou (2023) and use the reduced version of the dataset. Note, this version only contains English, as our aim is limited to the monolingual setting. Due to the organic nature of the dataset, conversations cover a very broad range of topics.", + "bbox": [ + 112, + 720, + 489, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PersonaChat (Zhang et al., 2018) is a crowdworker-sourced dataset comprising personagrounded conversations, in which each speaker is assigned a persona comprising a few short sentences. Following previous methods (Humeau et al., 2020), we concatenate the persona to the beginning of the message. The participants are instructed to chat naturally and to try to get to know one another.", + "bbox": [ + 507, + 84, + 884, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DailyDialog (Li et al., 2017) is a dataset created from English language learning websites and consists of a variety of high-quality dialogues in everyday scenarios. The dataset differs from the former two in that the conversations often involve real-life scenarios, such as asking for directions, and therefore captures a different variety of conversational skills.", + "bbox": [ + 507, + 237, + 884, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3 Metrics", + "text_level": 1, + "bbox": [ + 507, + 376, + 616, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on the same weighted ROUGE ensemble as previous methods (Lin, 2004; Deb et al., 2019, 2021), which is known to correlate well with click-through rate (Zhang et al., 2021):", + "bbox": [ + 507, + 397, + 882, + 461 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\text {R O U G E} - 1}{6} + \\frac {\\text {R O U G E} - 2}{3} + \\frac {\\text {R O U G E} - 3}{2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 579, + 470, + 882, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As the goal of SR systems it to ensure that at least one of the suggested replies is relevant to the user, we only record the maximum ROUGE score across each of the $K = 3$ suggested replies. We also evaluate the model on Self-ROUGE (Celikyilmaz et al., 2020): This is an unreferenced metric that measures the internal dissimilarity (i.e. diversity) within the reply set by treating one reply as the predicted reply and the other parts as the references. Note that a lower Self-ROUGE score indicates more diversity.", + "bbox": [ + 507, + 499, + 884, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4 Inference", + "text_level": 1, + "bbox": [ + 507, + 686, + 631, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For inference, we use the entire training set as the candidate pool for each respective dataset, with dedduplication to remove exact matches. For STAR, we greedily decode the next reply token until $K$ tokens have been decoded. Note, we only allow the model to output replies represented in the bootstrapped dataset, and also block non-replies, i.e. words from the original vocabulary, from being predicted.", + "bbox": [ + 507, + 706, + 884, + 850 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 Experimental Results", + "text_level": 1, + "bbox": [ + 507, + 862, + 729, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We focus our efforts on answering the following Research Questions: $(\\mathbf{RQ}_1)$ How does STAR com", + "bbox": [ + 507, + 887, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3Both changes lead to consistently improved accuracy and diversity across all datasets compared to the original paper.", + "bbox": [ + 112, + 891, + 487, + 919 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7615", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pare to existing state-of-the-art methods? (Section 5.1, 5.4); $(\\mathbf{RQ}_2)$ Which components of the data collection algorithm and fine-tuning have the largest impact on STAR's performance? (Section 5.2); $(\\mathbf{RQ}_3)$ How efficient is STAR in inference? (Section 5.3)", + "bbox": [ + 112, + 84, + 487, + 180 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Main Results", + "text_level": 1, + "bbox": [ + 112, + 191, + 265, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 compares the performance of different SR systems across the Reddit, PersonaChat and Dialy-Dialog datasets. In terms of relevance (ROUGE), STAR shows an especially large improvement in Reddit $(+17.9\\%)$ and DailyDialog $(+15.8\\%)$ . We hypothesise the gains in PersonaChat $(+5.1\\%)$ are more modest because the replies are more easily predicted due to the persona, which is concatenated to each message. This significantly reduces the noise during the initial retrieval for the baselines, as they only need to retrieve the messages relevant to that particular persona.", + "bbox": [ + 112, + 211, + 489, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For diversity (Self-ROUGE), the strongest gains were found in DailyDialog $(+63.1\\%)$ . For PersonaChat, STAR performs much better than the retrieval methods, only falling behind Seq2Seq, due to its altogether noisier outputs as evidenced by having the worst relevance score. The Reddit results were comparatively more modest $(+0.5\\%)$ - we hypothesise this is because the dataset is altogether more noisy, and so there are relatively few similar replies in the dataset, as shown by the Self-ROUGE scores being lower than the other two datasets. Overall, the consistent outperformance in both relevance and diversity metrics supports the benefits of the STAR approach.", + "bbox": [ + 110, + 405, + 489, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Ablation", + "text_level": 1, + "bbox": [ + 112, + 640, + 231, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we conduct ablations across two key axes: data collection and STAR training. The data collection ablations serve to investigate the benefits of the novel changes to the SimSR algorithm from Section 3.1.1. The STAR training ablations investigates the degree to which the improvements in performance are caused by the bootstrapped dataset or by STAR's architecture itself; we achieve this by considering several alternative variants of STAR. Our data collection ablations consider two features: (A) removing the query augmentation prevents the model from leveraging any ground truth information during prediction; (B) removing the redundancy penalty no longer explicitly penalises lack of diversity in predicted reply sets. For STAR training, we consider three alternative configurations:", + "bbox": [ + 112, + 661, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(C) we replace the bag-of-words embeddings with randomly initialised embeddings - this removes any priors about the meaning of replies and forces the model to learn them tabula rasa; (D) we treat each reply set as a unique token - this removes the compositional element from the task, constraining the model to only predicting previously seen reply sets, therefore testing whether the model is capable of learning to compose novel reply sets; (E) we remove the ability to account for interdependencies between replies, by restructuring each (message, reply set) data point into $K$ data points of (message, reply $k$ ), and then outputting the top- $K$ replies during inference - this investigates whether the benefit lies simply in the bootstrapped dataset being better suited to the SR task, rather than in STAR's ability to account for interdependencies between replies.", + "bbox": [ + 507, + 84, + 884, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of data collection ablations, we found removing the redundancy penalty significantly reduced the diversity of predictions, although in some cases offered slightly improved relevance; removing the query augmentation generally led to a worse relevance/diversity trade-off. For the variants of STAR training, we found that random embeddings consistently reduced relevance, while also led to less diverse predictions; reply sets as tokens led to the most competitive variant of STAR compared to our default setup: diversity was overall better, due to using preconstructed reply sets from the offline planning algorithm, but this came at the trade-off of reduced flexibility from being unable to construct novel reply sets when the context required it – resultantly, we saw a corresponding reduction in relevance. Finally, predicting replies separately expectedly harmed both relevance and diversity, demonstrating the importance of accounting for reply interdependencies.", + "bbox": [ + 507, + 381, + 884, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Figure 2, we further validated the individual results of our ablation by aggregating the results across datasets (applying an equal weighting to each dataset). This demonstrates the overall trend that the default STAR offers the superior trade-off between relevance and diversity, while treating reply sets as tokens offered the next best alternative. Nevertheless, we believe that keeping individual replies as tokens – thus allowing the model to construct reply sets dynamically – is likely to be an attractive property for deployed systems, enabling the overall vocabulary size to remain modest.", + "bbox": [ + 507, + 726, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7616", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fcc2b01350f85ab7e0093002d6745304619854fa9ddd950586b59954cfa75814.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRedditPersonaChatDailyDialog
ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
Generative models
Seq2Seq2.413.436.836.88*4.013.91
Retrieval models
Matching1.959.427.5121.476.5316.65
M-Topic1.813.947.1615.436.1411.11
M-MMR2.204.447.8114.576.138.63
M-CVAE2.305.027.4312.216.7810.49
SimSR42.792.189.0410.526.824.80
STAR3.29*2.179.50*7.747.90*1.77*
", + "bbox": [ + 136, + 80, + 860, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21861ba1b51ba9279542a9fb1e3c65069d56ce66744bcf3910596c4ea5239733.jpg", + "table_caption": [ + "Table 2: Performance of STAR across Reddit, PersonaChat and DailyDialog Test sets on relevance (ROUGE) and diversity (Self-ROUGE) metrics. Bold indicates best result, underline indicates second-best. * = statistically significant versus next best result on t-test with $p$ -value < 0.01." + ], + "table_footnote": [], + "table_body": "
ModelRedditPersonaChatDailyDialog
ConfigurationROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
STAR3.35*2.278.857.488.391.81
Data Collection Ablations
A: No Query Augmentation2.942.008.996.947.242.89
B: No Redundancy Penalty3.064.299.0317.268.98*5.90
STAR Training Variants
C: Random embeddings2.674.938.3910.976.844.45
D: Reply sets as tokens2.851.59*8.766.817.751.57*
E: Predict replies separately2.2026.618.0730.986.4320.50
", + "bbox": [ + 114, + 316, + 895, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Performance of STAR on the Reddit, PersonaChat and DailyDialog Validation sets under different model configurations. Ablations are applied separately. **Bold indicates best result, **underline indicates second-best. * = statistically significant versus next best result on t-test with $p$ -value < 0.01.", + "bbox": [ + 112, + 481, + 882, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 Run-time Efficiency", + "text_level": 1, + "bbox": [ + 112, + 550, + 317, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Beyond performance gains in relevance and diversity, a major advantage of an autoregressive retrieval model is the ability to leverage the scalability of GPU-based inference. Figure 3 compares the efficiency of STAR with the other baseline methods. We use an NVIDIA GeForce RTX 3060 Ti GPU and AMD Ryzen 7 5700G with Radeon Graphics CPU, with a batch size of 32. The results show that the methods can be broadly clustered into three groups. The slowest group is the generative method Seq2Seq, due to needing to generate each reply word-by-word. The middle group - SimSR, M-CVAE and M-MMR - is characterised by methods that comprise a more involved diversification pipeline. The final and fastest group includes STAR, M-Topic and Matching, where no additional post-hoc diversification is required (for M-Topic the topics can be pre-computed prior to inference).", + "bbox": [ + 112, + 575, + 489, + 865 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4 Case Study", + "text_level": 1, + "bbox": [ + 509, + 550, + 643, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4 presents a case study on the DailyDialog Test set. We compare our approach, STAR, with the top-performing baseline from Table 2, SimSR. In both examples we consistently find STAR is able to output a broader range of intents. Quantitatively, we consider the rank that each suggestion receives according to the initial retrieval of the Matching model that underlies SimSR. We see that STAR is able to perform a much more global search across the reply space, selecting replies from within the top 100 or so ranks. This would be difficult for the standard retrieve-and-erank approach to emulate, given 100 is usually too large a number to efficiently rerank (Deb et al., 2019). Qualitatively, SimSR's suggestions converge around common phrases, e.g. 'let's go', which would be difficult to deduplicate with a heuristic rule given only a limited number of overlapping words between the replies. Conversely, STAR is able to represent a broader range of intents, such as replying with a question in both examples. Further examples are", + "bbox": [ + 505, + 580, + 884, + 919 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Results surpass reported numbers in original paper due to inclusion of language-model bias and deduplicated candidate pool to support better comparability (Section 4.1).", + "bbox": [ + 112, + 879, + 487, + 919 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7617", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/080a6316394fb7b9c651b39cc33d1f2c91230cb059f3f066da7cf0303ddeff7c.jpg", + "image_caption": [ + "Figure 2: Comparison of overall relevance and diversity scores across ablations, obtained by averaging across all three datasets with equal weighting." + ], + "image_footnote": [], + "bbox": [ + 139, + 95, + 860, + 209 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/dd266042b070a64a7b1c683f7b0ba67d2a9c00c2ad4f53a0ce0094baedb945cb.jpg", + "image_caption": [ + "Figure 3: Comparison of run-time efficiency between STAR and the baseline methods. Results are calculated over the Reddit Validation set." + ], + "image_footnote": [], + "bbox": [ + 114, + 302, + 485, + 491 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/30886690c9f9c576da8d6ddbe44dd1769834beeaccaecd511314ec9ddf363d55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Message:Hi, Kenny. Let's go for a drink.
SimSR- let's go! [#9] - ok, let's go. [#3] - ok. let's get something to drink. [#1]
STAR- ok. let's go. [#5] - you want something to drink? [#89] - good idea. [#105]
Message:Of course! Let's go.
SimSR- let's go! [#1] - ok, let's go. [#5] - all right. let's go. [#12]
STAR- let's go! [#1] - where are we? [#43] - good idea! [#85]
", + "bbox": [ + 129, + 601, + 470, + 810 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 4: Example model outputs from the DailyDialog Test set, comparing STAR (ours) with the top-performing baseline method. Numbers in bold indicate the ranking the reply received according to the Matching model.", + "bbox": [ + 112, + 820, + 489, + 891 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "provided in Appendix C.", + "bbox": [ + 509, + 277, + 695, + 292 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 509, + 305, + 640, + 319 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduce STAR, an autoregressive retrieval system for SR, which is an end-to-end text-to-text model that sequentially predicts replies conditioned on an initial message. To train STAR, we demonstrate an approach to bootstrap a dataset of high-quality (message, reply set) pairs, from regular dialogue datasets containing only (message, reply) pairs. Empirically, our results show significant improvement over existing state-of-the-art SR baselines, across multiple datasets, corresponding to a $5.1\\% - 17.9\\%$ improvement in relevance, and a $0.5\\% - 63.1\\%$ improvement in diversity compared to the best baseline approach.", + "bbox": [ + 507, + 332, + 884, + 541 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Future work could extend these techniques to other set-prediction tasks: e.g., in IR the relevance of each document depends on the quantity of new information it contains compared to other documents in the set. In recommender systems, use cases include: tailoring a user's news feed requires that the news articles presented are not simply duplicates of the same story; designing a bespoke music playlist requires songs to be unified by common themes but also sufficiently distinct from one another to maintain the listener's interest. Other lines of future work include considering alternate strategies for initialising the reply embeddings, beyond the bag-of-words initialisation demonstrated in this paper.", + "bbox": [ + 507, + 542, + 882, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 796, + 680, + 812 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank the reviewers for their helpful feedback and suggestions. This work is partly supported by the EPSRC DTP Studentship program. The opinions expressed in this paper are the authors', and are not necessarily shared/endorsed by their employers and/or sponsors.", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7618", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 98 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Although our work shows that STAR is able to absorb sufficient information about the replies in its weights, this may become increasingly challenging when larger numbers of replies need to be embedded. One notable instance of this would be the multilingual setting, as many SR systems are deployed globally. In this case, each language typically has its own candidate pool. A naive implementation which creates separate reply vectors for each language would incur a significant increase in model size. In this case, we hypothesise techniques around weight-sharing between reply embeddings between languages may be beneficial, e.g. 'how are you' (en) and 'ça va' (fr) sharing the same vector. Further, our techniques are only demonstrated in publicly available datasets, whereas proprietary conversations in chat and email applications may have unique features not accounted for here (e.g. timestamps, cc and bcc information, and file attachments). Our technique also requires a planning algorithm to create the initial dataset. This theoretically creates an upper bound to the overall performance of STAR, as it is limited to cloning the behaviour of the offline planning algorithm.", + "bbox": [ + 115, + 110, + 489, + 495 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 521, + 213, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor Silva, Leonardo Neves, and Francesco Barbieri. 2022. Twitter topic classification. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3386-3400, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.", + "bbox": [ + 112, + 545, + 487, + 638 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In NeurIPS.", + "bbox": [ + 112, + 646, + 489, + 701 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.", + "bbox": [ + 112, + 709, + 489, + 879 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021a. Highly parallel autoregressive entity linking with dis", + "bbox": [ + 112, + 891, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "criminative correction. In Conference on Empirical Methods in Natural Language Processing.", + "bbox": [ + 507, + 85, + 882, + 112 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021b. Autoregressive entity retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net.", + "bbox": [ + 507, + 120, + 884, + 187 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nicola De Cao, Ledell Yu Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2021c. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290.", + "bbox": [ + 507, + 195, + 884, + 274 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jaime G. Carbonell and Jade Goldstein-Stewart. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Annual International ACM SIGIR Conference on Research and Development in Information Retrieval.", + "bbox": [ + 507, + 282, + 884, + 349 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. 2020. Evaluation of text generation: A survey. ArXiv, abs/2006.14799.", + "bbox": [ + 507, + 357, + 884, + 395 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nimesh Chakravarthi and Jeff Pasternack. 2017. Building smart replies for member messages. press release. https://engineering.linkedin.com/blog/2017/10/building-smart-replies-for-member-messages.", + "bbox": [ + 507, + 405, + 884, + 472 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Budhaditya Deb, Peter Bailey, and Milad Shokouhi. 2019. Diversifying reply suggestions using a matching-conditional variational autoencoder. In North American Chapter of the Association for Computational Linguistics.", + "bbox": [ + 507, + 479, + 884, + 545 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A conditional generative matching model for multi-lingual reply suggestion. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1553–1568, Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "bbox": [ + 507, + 554, + 884, + 646 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2020. Revisiting self-training for neural sequence generation. In Proceedings of ICLR.", + "bbox": [ + 507, + 653, + 884, + 695 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Matthew L. Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. Efficient natural language response suggestion for smart reply. CoRR, abs/1705.00652.", + "bbox": [ + 507, + 703, + 884, + 769 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531.", + "bbox": [ + 507, + 777, + 884, + 816 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409-14428, Toronto, Canada. Association for Computational Linguistics.", + "bbox": [ + 507, + 826, + 884, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7619", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multisentence scoring. In International Conference on Learning Representations.", + "bbox": [ + 112, + 85, + 489, + 154 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Gautier Izacard and Edouard Grave. 2021. Distilling knowledge from reader to retriever for question answering. In International Conference on Learning Representations.", + "bbox": [ + 112, + 162, + 489, + 214 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Youngsoo Jang, Jongmin Lee, and Kee-Eung Kim. 2020. Bayes-adaptive monte-carlo planning and learning for goal-oriented dialogues. In AAAI Conference on Artificial Intelligence.", + "bbox": [ + 112, + 225, + 489, + 279 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Youngsoo Jang, Seokin Seo, Jongmin Lee, and Kee-Eung Kim. 2021. Monte-carlo planning and learning with language action value estimates. In International Conference on Learning Representations.", + "bbox": [ + 112, + 288, + 489, + 343 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Gregory S. Corrado, Laszlo Lukács, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart reply: Automated response suggestion for email. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.", + "bbox": [ + 112, + 353, + 489, + 447 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In International Joint Conference on Natural Language Processing.", + "bbox": [ + 112, + 456, + 487, + 510 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics.", + "bbox": [ + 112, + 520, + 487, + 561 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.", + "bbox": [ + 112, + 571, + 489, + 612 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In International Conference on Learning Representations.", + "bbox": [ + 112, + 621, + 489, + 688 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1).", + "bbox": [ + 112, + 697, + 489, + 765 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric Michael Smith, Y.-Lan Boureau, and Jason Weston. 2020. Recipes for building an open-domain chatbot. In Conference of the European Chapter of the Association for Computational Linguistics.", + "bbox": [ + 112, + 775, + 489, + 854 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108.", + "bbox": [ + 112, + 865, + 489, + 917 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761.", + "bbox": [ + 507, + 85, + 884, + 151 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Timo Schick and Hinrich Schütze. 2021a. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online. Association for Computational Linguistics.", + "bbox": [ + 507, + 164, + 884, + 256 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Timo Schick and Hinrich Schütze. 2021b. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "bbox": [ + 507, + 269, + 884, + 350 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, L. Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy P. Lillicrap, and David Silver. 2019. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588:604 - 609.", + "bbox": [ + 507, + 363, + 884, + 442 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lancot, L. Sifre, Dharshan Kumaran, Thore Graepel, Timothy P. Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. ArXiv, abs/1712.01815.", + "bbox": [ + 507, + 455, + 884, + 546 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.", + "bbox": [ + 507, + 561, + 884, + 629 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems.", + "bbox": [ + 507, + 640, + 884, + 720 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Benjamin Towle and Ke Zhou. 2022. Learn what is possible, then choose what is best: Disentangling one-to-many relations in language through text-based games. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4955–4965, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "bbox": [ + 507, + 733, + 884, + 826 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Benjamin Towle and Ke Zhou. 2023. Model-based simulation for optimising smart reply. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12030-12043, Toronto, Canada. Association for Computational Linguistics.", + "bbox": [ + 507, + 839, + 884, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7620", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics.", + "bbox": [ + 112, + 85, + 489, + 191 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yue Weng, Huaixiu Zheng, Franziska Bell, and Gokhan Tur. 2019. Occ: A smart reply system for efficient in-app communications. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining.", + "bbox": [ + 112, + 199, + 489, + 267 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Kilian Wenker. 2023. Who wrote this? how smart replies impact language and agency in the workplace. *Telematics and Informatics Reports*, 10:100062.", + "bbox": [ + 112, + 275, + 489, + 317 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mozhi Zhang, Wei Wang, Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A dataset and baselines for multilingual reply suggestion. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1207-1220, Online. Association for Computational Linguistics.", + "bbox": [ + 112, + 324, + 489, + 444 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur D. Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Annual Meeting of the Association for Computational Linguistics.", + "bbox": [ + 112, + 451, + 489, + 520 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B. Dolan. 2019. Dialogpt: Large-scale generative pre-training for conversational response generation. In Annual Meeting of the Association for Computational Linguistics.", + "bbox": [ + 112, + 527, + 489, + 608 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In ACL.", + "bbox": [ + 112, + 615, + 489, + 657 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A Ethical Considerations", + "text_level": 1, + "bbox": [ + 112, + 668, + 351, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Controlling the outputs of dialogue models is a forefront issue in ethics research for AI, particularly with the impact of recent gains in LLM capabilities. We believe the risks in the case of SR systems have several mitigants compared to this: the replies can be vetted by humans before deployment; the replies are usually in short-form, rather containing complex information the user may rely on; ultimately, a user must select one of the options, rather than the system being able to reply without user oversight. Conversely, there are some risks more unique to SR systems that should be mentioned. Particularly, the suggestions presented by the system can have subtle priming effects on user behaviour. Notably,", + "bbox": [ + 112, + 694, + 489, + 919 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "users have been shown to be slightly more positive in the sentiment of their emails when shown suggested replies (Wenker, 2023). SR systems are on the whole known to produce more positive sentiment messages than the human distribution (Kannan et al., 2016). We see this as an extension of the broader trend of LLMs to be overly obsequious.", + "bbox": [ + 507, + 84, + 884, + 197 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B Implementation Details", + "text_level": 1, + "bbox": [ + 507, + 208, + 752, + 225 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For constructing the training dataset, we use the following hyperparameters: SimSR is initialised from the distilbert-base-uncased checkpoint (Sanh et al., 2019). We set the search parameters to $N = 100$ and $M = 100$ . We use a redundancy penalty of 0.05 and a blending alpha of 0.75 for query augmentation. Both parameters provided a good trade-off between relevance and diversity in early testing, so we did not search hyperparameters further (see Section 5.2 for ablations).", + "bbox": [ + 507, + 233, + 884, + 394 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For training STAR, we initialise our model with the t5-small checkpoint. Note that this version of T5 has a comparable parameter count to the baselines which use DistilBERT (60M versus 66M). We tokenise the dataset with a maximum message length of 64 tokens. We train our model for up to 100k steps, with a warmup of 1k steps. In practice, the model typically converged around 20k steps. We use the AdamW optimiser (Loshchilov and Hutter, 2019) with an initial learning rate of 5e-4 and linear decay. We evaluate every 2k steps, by taking the ROUGE and Self-ROUGE scores on the validation set (this uses the ground-truth from the original dataset, not from the bootstrapped dataset), and employ early stopping once both metrics have ceased to improve. Note, we found these metrics were a much more reliable stopping point than crossentropy loss, which typically converged much earlier.", + "bbox": [ + 507, + 395, + 884, + 684 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "C Further Case Studies", + "text_level": 1, + "bbox": [ + 507, + 696, + 732, + 712 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 5 displays further examples of STAR's predictions versus SimSR, taken from the PersonaChat Test set.", + "bbox": [ + 507, + 721, + 884, + 769 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7621", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/11d01728c7d182b2d20c3b1a8664d6d73e85809f44cdcea4466a397b0475ad66.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Persona:my favorite tv show is the office.
i like jazz music.
i do not drive because i live in new york.
i am jewish.
i do not eat anything sweet.
Message:i was but am now divorced
SimSR- i am sorry to hear that .
- oh . i am sorry to hear that .
- i am sorry to hear that . do you have any kids ?
STAR- oh i am sorry to hear that
- do you like music ?
- like music ? i love jazz
Persona:hey there my name is jordan and i am a veterinarian.
love to read drama books.
i love playing video games.
i am also a musician on the weekends.
i am originally from california but i live in florida.
Message:sometimes . i listen to a lot of music . do you read a lot ?
SimSR- hi how are you ?
- hi how are you today
- i do not like music at all
STAR- i listen to music do you
- no not really i like all music .
- yes i do my favorite is country
", + "bbox": [ + 231, + 263, + 764, + 694 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 5: Example model outputs from the PersonaChat Test set, comparing STAR (ours) with the top-performing baseline method. STAR is able to capture a broader range of intents through its end-to-end autoregressive retrieval.", + "bbox": [ + 112, + 703, + 882, + 733 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "7622", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_model.json b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ca7646fb8e82560c89ab5b0a8cf7a4e3e615c7b --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_model.json @@ -0,0 +1,2173 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.212, + 0.089, + 0.787, + 0.13 + ], + "angle": 0, + "content": "End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.142, + 0.628, + 0.159 + ], + "angle": 0, + "content": "Benjamin Towle1, Ke Zhou1,2" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.159, + 0.61, + 0.176 + ], + "angle": 0, + "content": "1University of Nottingham" + }, + { + "type": "text", + "bbox": [ + 0.43, + 0.177, + 0.573, + 0.192 + ], + "angle": 0, + "content": "\\(^{2}\\)Nokia Bell Labs" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.194, + 0.756, + 0.21 + ], + "angle": 0, + "content": "{benjamin.towle, ke.zhou}@nottingham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.282, + 0.461, + 0.678 + ], + "angle": 0, + "content": "Reply suggestion systems represent a staple component of many instant messaging and email systems. However, the requirement to produce sets of replies, rather than individual replies, makes the task poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. As a result, these system often rely on additional post-processing modules to diversify the outputs. However, these approaches are ultimately bottlenecked by the performance of the initial retriever, which in practice struggles to present a sufficiently diverse range of options to the downstream diversification module, leading to the suggestions being less relevant to the user. In this paper, we consider a novel approach that radically simplifies this pipeline through an autoregressive text-to-text retrieval model, that learns the smart reply task end-to-end from a dataset of (message, reply set) pairs obtained via bootstrapping. Empirical results show this method consistently outperforms a range of state-of-the-art baselines across three datasets, corresponding to a \\(5.1\\% - 17.9\\%\\) improvement in relevance, and a \\(0.5\\% - 63.1\\%\\) improvement in diversity compared to the best baseline approach. We make our code publicly available.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.693, + 0.26, + 0.708 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.719, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Reply suggestion, or smart reply (SR), systems are a staple component of many commercial applications such as Gmail, Skype, Outlook, Microsoft Teams, LinkedIn and Facebook Messenger. They help the user process chats and emails quicker by offering a set of canned replies which can be clicked without requiring manual typing. However, dialogue is known to be a one-to-many problem (Zhao et al., 2017; Towle and Zhou, 2022) – namely, for any given message, there are multiple possible replies. To reflect this uncertainty, systems should" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.254, + 0.885, + 0.35 + ], + "angle": 0, + "content": "present a diverse set of options to the user. For instance, given the message How are you?, an SR system could suggest: {I'm good; Ok; Not great}. Resultantly, the quality of a given reply depends not only on the message, but on the other replies in the reply set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.352, + 0.885, + 0.561 + ], + "angle": 0, + "content": "Several prior works explore solutions to this problem such as removing near duplicates, penalising inter-reply similarity (Deb et al., 2019), clustering by intent (Henderson et al., 2017; Weng et al., 2019), learning latent variables (Deb et al., 2019, 2021), or model-based simulation (Towle and Zhou, 2023). However, these methods share a common design choice (Figure 1A): (1) a retrieval-based Matching model, which has learned a shared embedding space between messages and replies, returns a shortlist of top scoring replies; (2) this shortlist is refined through some diversification procedure to obtain the final reply set." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.564, + 0.886, + 0.708 + ], + "angle": 0, + "content": "Unfortunately, this assumes that the initial shortlist contains at least one good reply set. In practice, we find Matching models often search myopically, only retrieving candidates that are very similar to one another (Figure 1A). Thus, the chosen reply set often fails to reflect a diverse range of user intents, while latency constraints make more sophisticated diversification techniques or larger shortlists prohibitive (Deb et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.71, + 0.885, + 0.919 + ], + "angle": 0, + "content": "An intuitive, but – to the best of our knowledge – unexplored, solution to this problem is to conduct the retrieval autoregressively, with each reply conditioned on both the initial message and the previous replies in the set. Unfortunately, this approach encounters a second problem, namely, the lack of any datasets containing (message, reply set) pairs (Towle and Zhou, 2023). In practice, SR systems are trained on individual (message, reply) pairs obtained from conversation datasets, while the task of presenting multiple diverse replies to the user is outsourced to a separate diversification module." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.479, + 0.919 + ], + "angle": 0, + "content": "'https://github.com/ BenjaminTowle/STAR" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7610" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.78, + 0.958 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7610-7622" + }, + { + "type": "footer", + "bbox": [ + 0.278, + 0.958, + 0.721, + 0.972 + ], + "angle": 0, + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.085, + 0.882, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.192, + 0.85, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.333, + 0.884, + 0.378 + ], + "angle": 0, + "content": "Figure 1: Previous methods [A] compared to our approach, STAR [B]. The example displayed is taken from the DailyDialog Test set, and compares the predictions of STAR with SimSR (Towle and Zhou, 2023), the next best method. Our method's suggestions present a diverse range of topics/intents to drive the conversation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.401, + 0.49, + 0.918 + ], + "angle": 0, + "content": "To meet this dual need, we present both (i) a bootstrapping method for creating a high-quality dataset of (message, reply sets) and (ii) a novel autoregressive retrieval model which predicts sequences of replies. For solving (i), we observe how model-based planning algorithms have been known to serve as a powerful policy improvement operator (Silver et al., 2017; Schrittwieser et al., 2019), including in several NLP systems (Jang et al., 2020, 2021). Specifically, the outputs of a model-based planning algorithm can be used to bootstrap a SR system. Further, by conducting this planning offline we are able to leverage two key advantages: (1) the system is free of the latency constraints of online inference, and therefore can increase the search space coverage of the planning algorithm; (2) the system can leverage information that would not be available during inference, such as the ground-truth reply, to further guide the search process. For (ii) we unify both steps of the standard SR pipeline into a single end-to-end model, which mitigates the myopic search, and allows the model to learn to diversify its predictions in a principled way through gradient-based learning. To this end, we present STAR (Suggested replies with T5 and Autoregressive Retrieval) (Figure 1B). At a high level, STAR is a text-to-text model trained to output sequences of replies, where each reply is conditioned both on the initial message and the previous replies in the sequence. Concretely, we instantiate our method with the T5 pretrained model (Raffel et al., 2020). We expand T5's vocabulary by" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.401, + 0.885, + 0.529 + ], + "angle": 0, + "content": "treating each reply in the candidate pool as a novel token, and demonstrate a simple-yet-effective technique for initialising the new token embeddings, which leverages the model's existing semantic priors. Notably, by treating each reply as a token, we limit the number of autoregressive decoding steps required, keeping the model's efficiency comparable to other retrieval-based methods." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.54, + 0.884, + 0.797 + ], + "angle": 0, + "content": "Empirically, we evaluate our approach on three benchmarks: Reddit (Zhang et al., 2021), which is the only publicly-available SR benchmark, as well as PersonaChat (Zhang et al., 2018) and DailyDialog (Li et al., 2017) which are both widely-used in dialogue research more broadly (Zhang et al., 2019; Roller et al., 2020, inter alia), and share a similar conversational style with SR apps. We demonstrate superior performance over state-of-the-art baselines across all datasets, corresponding to a \\(5.1\\% - 17.9\\%\\) improvement in relevance, and a \\(0.5\\% - 63.1\\%\\) improvement in diversity compared to the best baseline approach. We further show comparable efficiency to previous methods, and perform a range of ablations to motivate our design choices." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.885, + 0.92 + ], + "angle": 0, + "content": "In summary, our key contributions are as follows: (1) an autoregressive retrieval architecture for sequentially predicting suggested replies; (2) a bootstrapping framework for generating high-quality data of (message, reply set) pairs; (3) detailed analysis of model behaviour and performance including a case study and ablation of key components." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "7611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.084, + 0.271, + 0.099 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.116, + 0.491, + 0.533 + ], + "angle": 0, + "content": "Smart reply The proprietary nature of data from email and chat applications has led several previous works to use publicly-available dialogue datasets (Zhang et al., 2021; Deb et al., 2021; Towle and Zhou, 2023) to benchmark SR methods, due to their analogous conversational nature. While early SR systems used generative models (Kannan et al., 2016), current production systems favour retrieval methods due to their greater controllability of outputs and superior latency (Deb et al., 2019). Increasing the diversity of reply suggestions is a key focus of previous work, which has been attempted by: (1) mapping replies to discrete intents / topics (Kannan et al., 2016; Chakravarthi and Pasternack, 2017; Weng et al., 2019); (2) re-weighting replies according to their similarity with other replies in the set (Carbonell and Goldstein-Stewart, 1998; Deb et al., 2019); (3) learning continuous latent variables to generate multiple queries (Zhao et al., 2017; Deb et al., 2019); (4) using model-based simulation to iteratively search and evaluate the relevance of candidate reply sets (Towle and Zhou, 2023). Our proposed method differs from all of these approaches in that our model learns to account for the interdependencies between replies through end-to-end backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.55, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Autoregressive retrieval Integrating neural retrieval into the well-established paradigm of text-to-text models is of growing interest. Earlier work focuses on outputting a document ID given a query (Tay et al., 2022). Further work has extended this by considering alternate ways of representing the document IDs, such as through unique substrings (Bevilacqua et al., 2022). Another line of work has used autoregressive retrieval for the entity linking task (Cao et al., 2021a,b,c). There, the motivation is to reduce the large number of entities by relying on the text-to-text model's pre-existing vocabulary, rather than having to retrieve embeddings from a memory-intensive dense index. Our proposed method differs considerably from these previous works both in instantiation and motivation. Instantiation-wise, we generate multiple replies - critical to making this possible is the novel bootstrapping technique for creating the dataset of (message, reply set) pairs to train on. Motivation-wise, our goal is to be able to condition each reply on both the input message and previous replies in the set, enabling the model to learn to predict se" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.822, + 0.101 + ], + "angle": 0, + "content": "quences of replies in a differentiable way." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.11, + 0.886, + 0.497 + ], + "angle": 0, + "content": "Bootstrapping The idea of bootstrapping training data from limited resources has received significant recent interest in NLP, given the newly demonstrated few / zero-shot capabilities of many large language models (Brown et al., 2020). It has seen usage in few-shot shot text-classification (Schick and Schütze, 2021a), semantic similarity (Schick and Schütze, 2021b), tool-usage (Schick et al., 2023), retrieval (Izacard and Grave, 2021), sequence generation (He et al., 2020), and instruction-tuning (Honovich et al., 2023; Wang et al., 2023; Taori et al., 2023), amongst others. These techniques can also be seen as a form of knowledge distillation (Hinton et al., 2015), except that the training typically involves predicting the exact token targets, rather than using the soft probabilities of a teacher model. Although sometimes these techniques are used as an addition to supervised learning (He et al., 2020), in our case there are no datasets containing the ideal reply sets to suggest to the user. Instead, we must bootstrap this in a more unsupervised way, by transforming a dataset of (message, reply) pairs into a dataset of (message, reply set) pairs." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.508, + 0.659, + 0.525 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.534, + 0.884, + 0.614 + ], + "angle": 0, + "content": "In this section, we first describe the model-based planning process used to obtain the bootstrapped dataset of (message, reply set) pairs (Section 3.1). Then, we show how the STAR architecture can be trained on this dataset (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.626, + 0.749, + 0.64 + ], + "angle": 0, + "content": "3.1 Offline Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.646, + 0.885, + 0.87 + ], + "angle": 0, + "content": "Our goal is to transform a dialogue dataset \\(\\mathcal{D} = \\{(x,y)\\}\\) of (message, reply) tuples, into a dataset \\(\\mathcal{D}^* = \\{(x,Y)\\}\\) where \\(Y\\) is the set of replies \\(\\{y_k\\}^K\\) to be presented to the user. Algorithm 1 summarises this process. While our method is general to any arbitrary planning algorithm, we choose to instantiate our approach with a modified version of SimSR (Towle and Zhou, 2023), a recently released publicly available state-of-the-art SR method, that employs model-based simulation to predict reply sets. As the original algorithm was designed for online inference, we make several changes to benefit the offline nature of our version, and detail the full implementation below." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.872, + 0.884, + 0.919 + ], + "angle": 0, + "content": "The initial retrieval is conducted by a Matching model \\(\\Phi\\) that separately encodes messages and replies into a shared latent space. Given an encoded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7612" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.084, + 0.487, + 0.131 + ], + "angle": 0, + "content": "Algorithm 1 Offline Dataset Creation. We use \\( N = 100 \\), \\( M = 100 \\), \\( \\alpha = 0.75 \\) and \\( \\lambda = 0.05 \\) as our default setting." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.133, + 0.488, + 0.182 + ], + "angle": 0, + "content": "Input Matching model \\(\\Phi\\), message \\(x\\), precomputed reply vectors \\(\\{\\mathbf{y_r}\\}^R\\), number of candidates \\(N\\), number of simulations \\(M\\), final reply set size \\(K\\), query augmentation coefficient \\(\\alpha\\), redundancy penalty \\(\\lambda\\)." + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.182, + 0.272, + 0.194 + ], + "angle": 0, + "content": "Output reply set \\(Y_{K}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.196, + 0.486, + 0.219 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {x}, \\mathbf {y} \\leftarrow \\Phi (x), \\Phi (y) \\\\ \\tilde {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.21, + 0.487, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\\\ Y _ {N} \\leftarrow N \\underset {r} {\\operatorname {a r g m a x}} (\\bar {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}}) \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.238, + 0.297, + 0.255 + ], + "angle": 0, + "content": "\\[\nY _ {M} \\leftarrow M \\operatorname {a r g m a x} _ {\\mathbf {r}} (\\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.256, + 0.487, + 0.28 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q (y _ {m} | x) \\propto \\exp \\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {m}} \\\\ Y _ {G} \\leftarrow \\emptyset \\end{array} \\quad \\triangleright \\text {s o f t m a x o v e r t o p - M s c o r e s}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.281, + 0.48, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c}\\textbf{for} k\\gets 0\\text{to} K\\textbf{do}\\\\ y_{k}\\leftarrow \\operatorname *{argmax}_{n}\\sum_{m}^{M}f(Y_{G}^{n},y_{m})q(y_{m}|x) - \\lambda f(Y_{G},y_{n}) \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.318, + 0.252, + 0.33 + ], + "angle": 0, + "content": "\\[\nY _ {G} \\leftarrow Y _ {G} \\cup y _ {k}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.33, + 0.181, + 0.339 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.34, + 0.198, + 0.352 + ], + "angle": 0, + "content": "\\[\nY _ {K} \\leftarrow Y _ {G}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.353, + 0.199, + 0.364 + ], + "angle": 0, + "content": "\\[\n\\text {r e t u r n} Y _ {K}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.39, + 0.489, + 0.485 + ], + "angle": 0, + "content": "message \\(\\mathbf{x} = \\Phi (x)\\), it retrieves the top \\(N\\) candidates from a pool of pre-computed reply vectors \\(\\mathbf{Y}_{\\mathbf{R}} = \\{\\mathbf{y}_{\\mathbf{r}}\\}^{R}\\) by combining their dot product similarity with a pre-computed language-model bias - a standard component of SR systems to downweight overly specific replies (Deb et al., 2019)." + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.492, + 0.488, + 0.511 + ], + "angle": 0, + "content": "\\[\nY _ {N} = N \\underset {r} {\\operatorname {a r g m a x}} \\left(\\mathbf {x} \\cdot \\mathbf {y} _ {\\mathbf {r}} + \\beta \\mathrm {L M} \\left(y _ {r}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.522, + 0.489, + 0.572 + ], + "angle": 0, + "content": "We then output the \\(K\\)-tuple \\(Y_{i} \\in \\binom{Y_{N}}{K}\\) that has the highest expected similarity with the human reply, according to some similarity function \\(f(\\cdot, \\cdot)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.578, + 0.488, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\underset {i} {\\operatorname {a r g m a x}} \\mathbb {E} _ {y \\sim p (\\cdot | x)} \\left[ f (Y _ {i}, y) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.61, + 0.489, + 0.695 + ], + "angle": 0, + "content": "Given the objective in SR is for at least one of the replies to be relevant, the similarity function is defined as a maximum over the sampled reply and each of the replies in the reply set, using term-level F1-score: \\(\\max_k\\mathrm{F1}(y_k,y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.489, + 0.806 + ], + "angle": 0, + "content": "We assume \\(y\\) is sampled from the ground-truth human distribution \\(p(\\cdot |x)\\). As we do not have access to the true human distribution in practice, we instead use the same Matching model \\(q\\) as a proxy for this, given it is trained on (message, reply) pairs. We then approximate the expectation by marginalising over the top-\\(M\\) most likely replies:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.814, + 0.488, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\approx \\underset {i} {\\operatorname {a r g m a x}} \\sum_ {m} ^ {M} f \\left(Y _ {i}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.489, + 0.919 + ], + "angle": 0, + "content": "In practice, it is intractable to evaluate every possible reply tuple, due to their combinatorial scaling. We therefore approximate this by greedily constructing the reply set one reply at a time. Formally," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.883, + 0.15 + ], + "angle": 0, + "content": "let \\(Y_{G}\\) be the set of currently selected replies, such that initially \\(Y_{G} = \\emptyset\\). Then, for each of \\(y_{n} \\in Y_{N}\\), we compute the expected similarity for the union of \\(Y_{G}\\) and \\(y_{n}\\), termed \\(Y_{G}^{n} = Y_{G} \\cup y_{n}\\) for brevity:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.161, + 0.883, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.201, + 0.884, + 0.312 + ], + "angle": 0, + "content": "We repeat this process for \\(K\\) timesteps, each time appending the highest scoring reply to \\(Y_{G}\\), i.e. until \\(|Y_{G}| = K\\). Note that this greedy search process implicitly canonicalises the order of the replies, as selecting replies in this way causes them to be roughly ordered by individual message-reply relevance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.324, + 0.671, + 0.339 + ], + "angle": 0, + "content": "3.1.1 Adjustments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.342, + 0.884, + 0.535 + ], + "angle": 0, + "content": "Scaling \\(N\\) and \\(M\\) The original SimSR algorithm was used only in an online setting (Towle and Zhou, 2023). Therefore, the size of the search parameters \\(N\\) (number of replies in the shortlist) and \\(M\\) (number of simulated user replies) is kept low (15 and 25 respectively in the original paper). As we only need to run this model offline however to obtain the dataset, we find setting \\(N\\) and \\(M\\) to much larger values improves relevance (we use 100 for both), enabling both a broader search (i.e. by increasing \\(N\\)) and a more accurate similarity function (i.e. by increasing \\(M\\))." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.544, + 0.884, + 0.688 + ], + "angle": 0, + "content": "Redundancy penalty Early testing showed that scaling the search parameters reduced diversity. We therefore introduce a redundancy penalty, which penalises the model for selecting replies that are similar to replies already in the set \\( Y_{G} \\). This is analogous to the inter-document similarity penalty used in the maximum marginal relevance IR (information retrieval) technique (Carbonell and Goldstein-Stewart, 1998)." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.701, + 0.883, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) - \\lambda f \\left(Y _ {G}, y _ {n}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Query augmentation Unlike during online inference, we also have access to the ground-truth reply \\(y\\) when constructing the dataset. Previous work has found that models obtain greater representational capabilities when given access to posterior information (Paranjape et al., 2022; Towle and Zhou, 2022). We therefore use an augmented query to retrieve with the Matching model. This is obtained by interpolating between the message and ground-truth reply embeddings. This biases the model's predictions towards the observed ground-truth in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7613" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.488, + 0.116 + ], + "angle": 0, + "content": "the dataset, while still allowing it to benefit from its own learned distribution." + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.129, + 0.488, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} = \\alpha \\Phi (x) + (1 - \\alpha) \\Phi (y) \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.16, + 0.342, + 0.176 + ], + "angle": 0, + "content": "3.2 Proposed STAR Model" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.182, + 0.49, + 0.716 + ], + "angle": 0, + "content": "We initialise STAR with a T5-based text-to-text language model, which has previously been shown to be effective in autoregressive retrieval (Tay et al., 2022). While some autoregressive retrieval approaches identify their documents/replies through unique substrings (Bevilacqua et al., 2022) or constrained beam search (Cao et al., 2021b), we focus on approaches requiring only a limited number of autoregressive steps, to maintain competitive inference speeds to existing retrieval methods (Section 5.3). There are several alternatives for this such as treating each reply set as a unique token, or separately training on each (message, reply pair), but ultimately we opted for autoregressively treating each reply as a unique token in the vocabulary in order to exploit the compositionality of reply sets (Section 5.2 for performance comparison). Note that as the types of replies used in smart reply are usually quite short and concise, e.g. 'how are you', 'I'm fine thanks', 'yes, that's right' etc., systems in deployment only need to retrieve from a pool of \\(30\\mathrm{k}\\) or so replies (Deb et al., 2019), in order to provide good coverage of possible user intents. As a result, we are able to keep the size of the vocabulary reasonable. Thus, our new vocabulary is defined as: \\(W_{tokens} \\cup W_{replies}\\). An obvious challenge to this approach is that by treating each reply as a previously unseen word, it removes any semantic priors the model might have about their meaning. To mitigate this, we employ a bag-of-words initialisation strategy. Hence, we define the embedding of the \\(t\\)-th reply \\(E(y_{t})\\) as the average over the embeddings of the individual words within \\(w_{n} \\in y_{t}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.727, + 0.489, + 0.763 + ], + "angle": 0, + "content": "\\[\nE \\left(y _ {t}\\right) = \\frac {1}{N} \\sum_ {n} ^ {N} E \\left(w _ {n}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Intuitively, this ensures that the initial embeddings are close to the word embeddings of the original vocabulary, while also capturing some of the underlying semantics of the reply. We allow the weights to update during fine-tuning. Note that for T5 the output and input embedding layers share weights, and therefore this approach is used to initialise both layers. We train the model using cross-entropy loss to predict the next reply given the current sequence" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.695, + 0.101 + ], + "angle": 0, + "content": "of replies and messages:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.117, + 0.884, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N L L} = - \\sum_ {k} ^ {K} \\log p \\left(y _ {k} \\mid x, y _ {0}, \\dots , y _ {k - 1}\\right) \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.169, + 0.719, + 0.187 + ], + "angle": 0, + "content": "4 Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.199, + 0.63, + 0.213 + ], + "angle": 0, + "content": "4.1 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.222, + 0.885, + 0.382 + ], + "angle": 0, + "content": "Previous work has largely been closed-source and is therefore unavailable for direct comparison (Henderson et al., 2017; Weng et al., 2019; Deb et al., 2019). With the exception of SimSR, which has publicly available code \\(^{2}\\), we re-implement a variety of methods that cover the broad range of previous techniques. Due to its comparable size, all baselines apart from Seq2Seq are initialised with DistilBERT as the encoder backbone. These are summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.398, + 0.885, + 0.558 + ], + "angle": 0, + "content": "Seq2Seq is a generative encoder-decoder. While current production systems and the majority of related works use only retrieval models (Deb et al., 2019; Towle and Zhou, 2023), at least one related work includes a standard generative transformer as a baseline (Zhang et al., 2021), which we follow here. For maximum comparability with our method, we use the same t5-small model as a backbone. For each message, we sample \\(K\\) responses independently." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.573, + 0.884, + 0.654 + ], + "angle": 0, + "content": "Matching represents the out-of-the-box encoder with no additional diversification strategy and was used as a baseline method by Zhang et al. (2021). It simply selects the top \\(K\\) responses according to individual message-reply scores." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.667, + 0.884, + 0.779 + ], + "angle": 0, + "content": "Matching-Topic uses an out-of-the-box topic classifier to ensure no two replies share the same topic, similar to previous work (Henderson et al., 2017; Weng et al., 2019). The classifier is trained on Twitter (Antypas et al., 2022), due to their comparable short-form open-domain chat conversations." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.794, + 0.885, + 0.891 + ], + "angle": 0, + "content": "Maximum Marginal Relevance (MMR) (Carbonell and Goldstein-Stewart, 1998) is originally an IR technique, used in several previous SR works (Deb et al., 2019; Towle and Zhou, 2023), which re-weights reply scores as a linear combination of their message-reply and inter-reply similarity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.882, + 0.919 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/BenjaminTowle/SimSR" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7614" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.147, + 0.082, + 0.458, + 0.148 + ], + "angle": 0, + "content": "
TrainValidTest|YR|
Reddit50k5k5k48k
PersonaChat66k8k8k64k
DailyDialog76k7k7k62k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.156, + 0.49, + 0.214 + ], + "angle": 0, + "content": "Table 1: Number of samples in the Train, Validation, Test sets and Candidate pool in the three datasets for evaluation. The Candidate pool comprises the Train set with duplicate responses removed." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.241, + 0.49, + 0.352 + ], + "angle": 0, + "content": "MCVAE (Deb et al., 2019) is a conditional variational autoencoder (Zhao et al., 2017) which learns to generate multiple query vectors from a single message embedding, representing the multiple possible reply intents. Candidates are scored via a voting process, whereby the \\(K\\) most-selected replies are chosen." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.366, + 0.49, + 0.512 + ], + "angle": 0, + "content": "SimSR (Towle and Zhou, 2023) uses an iterative search and evaluation process to select possible reply sets and score them according to their expected similarity from a learned world model, which serves as a proxy for the user. To ensure comparability of SimSR with our method and the other baselines, we include the language-model bias in the scoring process (Equation 1), and also deduplicate the candidate pool.\\(^{3}\\)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.525, + 0.229, + 0.539 + ], + "angle": 0, + "content": "4.2 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.548, + 0.49, + 0.709 + ], + "angle": 0, + "content": "We evaluate our proposed method across three datasets, summarised in Table 1. Below, we describe the datasets in more detail and motivate their inclusion. Note, other than Reddit, there are no publicly available SR datasets, due to their commercial nature (e.g. Henderson et al. (2017); Deb et al. (2019); Weng et al. (2019)). Therefore, we adopt several dialogue datasets, which is the closest alternative to conversations on proprietary chat applications." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.721, + 0.49, + 0.882 + ], + "angle": 0, + "content": "Reddit (Zhang et al., 2021) was originally introduced for training multilingual SR systems, and is the only publicly available dataset specifically intended for SR purposes. As the original dataset is very large, we follow Towle and Zhou (2023) and use the reduced version of the dataset. Note, this version only contains English, as our aim is limited to the monolingual setting. Due to the organic nature of the dataset, conversations cover a very broad range of topics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.228 + ], + "angle": 0, + "content": "PersonaChat (Zhang et al., 2018) is a crowdworker-sourced dataset comprising personagrounded conversations, in which each speaker is assigned a persona comprising a few short sentences. Following previous methods (Humeau et al., 2020), we concatenate the persona to the beginning of the message. The participants are instructed to chat naturally and to try to get to know one another." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.238, + 0.885, + 0.365 + ], + "angle": 0, + "content": "DailyDialog (Li et al., 2017) is a dataset created from English language learning websites and consists of a variety of high-quality dialogues in everyday scenarios. The dataset differs from the former two in that the conversations often involve real-life scenarios, such as asking for directions, and therefore captures a different variety of conversational skills." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.377, + 0.618, + 0.391 + ], + "angle": 0, + "content": "4.3 Metrics" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.398, + 0.884, + 0.462 + ], + "angle": 0, + "content": "We evaluate our method on the same weighted ROUGE ensemble as previous methods (Lin, 2004; Deb et al., 2019, 2021), which is known to correlate well with click-through rate (Zhang et al., 2021):" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.472, + 0.884, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\frac {\\text {R O U G E} - 1}{6} + \\frac {\\text {R O U G E} - 2}{3} + \\frac {\\text {R O U G E} - 3}{2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.5, + 0.885, + 0.676 + ], + "angle": 0, + "content": "As the goal of SR systems it to ensure that at least one of the suggested replies is relevant to the user, we only record the maximum ROUGE score across each of the \\( K = 3 \\) suggested replies. We also evaluate the model on Self-ROUGE (Celikyilmaz et al., 2020): This is an unreferenced metric that measures the internal dissimilarity (i.e. diversity) within the reply set by treating one reply as the predicted reply and the other parts as the references. Note that a lower Self-ROUGE score indicates more diversity." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.687, + 0.633, + 0.7 + ], + "angle": 0, + "content": "4.4 Inference" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.707, + 0.885, + 0.851 + ], + "angle": 0, + "content": "For inference, we use the entire training set as the candidate pool for each respective dataset, with dedduplication to remove exact matches. For STAR, we greedily decode the next reply token until \\( K \\) tokens have been decoded. Note, we only allow the model to output replies represented in the bootstrapped dataset, and also block non-replies, i.e. words from the original vocabulary, from being predicted." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.863, + 0.731, + 0.879 + ], + "angle": 0, + "content": "5 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We focus our efforts on answering the following Research Questions: \\((\\mathbf{RQ}_1)\\) How does STAR com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.892, + 0.488, + 0.92 + ], + "angle": 0, + "content": "3Both changes lead to consistently improved accuracy and diversity across all datasets compared to the original paper." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.181 + ], + "angle": 0, + "content": "pare to existing state-of-the-art methods? (Section 5.1, 5.4); \\((\\mathbf{RQ}_2)\\) Which components of the data collection algorithm and fine-tuning have the largest impact on STAR's performance? (Section 5.2); \\((\\mathbf{RQ}_3)\\) How efficient is STAR in inference? (Section 5.3)" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.192, + 0.266, + 0.206 + ], + "angle": 0, + "content": "5.1 Main Results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.212, + 0.49, + 0.405 + ], + "angle": 0, + "content": "Table 2 compares the performance of different SR systems across the Reddit, PersonaChat and Dialy-Dialog datasets. In terms of relevance (ROUGE), STAR shows an especially large improvement in Reddit \\((+17.9\\%)\\) and DailyDialog \\((+15.8\\%)\\). We hypothesise the gains in PersonaChat \\((+5.1\\%)\\) are more modest because the replies are more easily predicted due to the persona, which is concatenated to each message. This significantly reduces the noise during the initial retrieval for the baselines, as they only need to retrieve the messages relevant to that particular persona." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.406, + 0.49, + 0.631 + ], + "angle": 0, + "content": "For diversity (Self-ROUGE), the strongest gains were found in DailyDialog \\((+63.1\\%)\\). For PersonaChat, STAR performs much better than the retrieval methods, only falling behind Seq2Seq, due to its altogether noisier outputs as evidenced by having the worst relevance score. The Reddit results were comparatively more modest \\((+0.5\\%)\\) - we hypothesise this is because the dataset is altogether more noisy, and so there are relatively few similar replies in the dataset, as shown by the Self-ROUGE scores being lower than the other two datasets. Overall, the consistent outperformance in both relevance and diversity metrics supports the benefits of the STAR approach." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.642, + 0.232, + 0.656 + ], + "angle": 0, + "content": "5.2 Ablation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.919 + ], + "angle": 0, + "content": "In Table 3, we conduct ablations across two key axes: data collection and STAR training. The data collection ablations serve to investigate the benefits of the novel changes to the SimSR algorithm from Section 3.1.1. The STAR training ablations investigates the degree to which the improvements in performance are caused by the bootstrapped dataset or by STAR's architecture itself; we achieve this by considering several alternative variants of STAR. Our data collection ablations consider two features: (A) removing the query augmentation prevents the model from leveraging any ground truth information during prediction; (B) removing the redundancy penalty no longer explicitly penalises lack of diversity in predicted reply sets. For STAR training, we consider three alternative configurations:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.358 + ], + "angle": 0, + "content": "(C) we replace the bag-of-words embeddings with randomly initialised embeddings - this removes any priors about the meaning of replies and forces the model to learn them tabula rasa; (D) we treat each reply set as a unique token - this removes the compositional element from the task, constraining the model to only predicting previously seen reply sets, therefore testing whether the model is capable of learning to compose novel reply sets; (E) we remove the ability to account for interdependencies between replies, by restructuring each (message, reply set) data point into \\(K\\) data points of (message, reply \\(k\\)), and then outputting the top-\\(K\\) replies during inference - this investigates whether the benefit lies simply in the bootstrapped dataset being better suited to the SR task, rather than in STAR's ability to account for interdependencies between replies." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.382, + 0.885, + 0.703 + ], + "angle": 0, + "content": "In terms of data collection ablations, we found removing the redundancy penalty significantly reduced the diversity of predictions, although in some cases offered slightly improved relevance; removing the query augmentation generally led to a worse relevance/diversity trade-off. For the variants of STAR training, we found that random embeddings consistently reduced relevance, while also led to less diverse predictions; reply sets as tokens led to the most competitive variant of STAR compared to our default setup: diversity was overall better, due to using preconstructed reply sets from the offline planning algorithm, but this came at the trade-off of reduced flexibility from being unable to construct novel reply sets when the context required it – resultantly, we saw a corresponding reduction in relevance. Finally, predicting replies separately expectedly harmed both relevance and diversity, demonstrating the importance of accounting for reply interdependencies." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.727, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In Figure 2, we further validated the individual results of our ablation by aggregating the results across datasets (applying an equal weighting to each dataset). This demonstrates the overall trend that the default STAR offers the superior trade-off between relevance and diversity, while treating reply sets as tokens offered the next best alternative. Nevertheless, we believe that keeping individual replies as tokens – thus allowing the model to construct reply sets dynamically – is likely to be an attractive property for deployed systems, enabling the overall vocabulary size to remain modest." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7616" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.082, + 0.861, + 0.251 + ], + "angle": 0, + "content": "
MethodRedditPersonaChatDailyDialog
ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
Generative models
Seq2Seq2.413.436.836.88*4.013.91
Retrieval models
Matching1.959.427.5121.476.5316.65
M-Topic1.813.947.1615.436.1411.11
M-MMR2.204.447.8114.576.138.63
M-CVAE2.305.027.4312.216.7810.49
SimSR42.792.189.0410.526.824.80
STAR3.29*2.179.50*7.747.90*1.77*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.261, + 0.885, + 0.306 + ], + "angle": 0, + "content": "Table 2: Performance of STAR across Reddit, PersonaChat and DailyDialog Test sets on relevance (ROUGE) and diversity (Self-ROUGE) metrics. Bold indicates best result, underline indicates second-best. * = statistically significant versus next best result on t-test with \\( p \\)-value < 0.01." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.317, + 0.896, + 0.475 + ], + "angle": 0, + "content": "
ModelRedditPersonaChatDailyDialog
ConfigurationROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
STAR3.35*2.278.857.488.391.81
Data Collection Ablations
A: No Query Augmentation2.942.008.996.947.242.89
B: No Redundancy Penalty3.064.299.0317.268.98*5.90
STAR Training Variants
C: Random embeddings2.674.938.3910.976.844.45
D: Reply sets as tokens2.851.59*8.766.817.751.57*
E: Predict replies separately2.2026.618.0730.986.4320.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.482, + 0.884, + 0.527 + ], + "angle": 0, + "content": "Table 3: Performance of STAR on the Reddit, PersonaChat and DailyDialog Validation sets under different model configurations. Ablations are applied separately. **Bold indicates best result, **underline indicates second-best. * = statistically significant versus next best result on t-test with \\( p \\)-value < 0.01." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.551, + 0.318, + 0.567 + ], + "angle": 0, + "content": "5.3 Run-time Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.576, + 0.49, + 0.866 + ], + "angle": 0, + "content": "Beyond performance gains in relevance and diversity, a major advantage of an autoregressive retrieval model is the ability to leverage the scalability of GPU-based inference. Figure 3 compares the efficiency of STAR with the other baseline methods. We use an NVIDIA GeForce RTX 3060 Ti GPU and AMD Ryzen 7 5700G with Radeon Graphics CPU, with a batch size of 32. The results show that the methods can be broadly clustered into three groups. The slowest group is the generative method Seq2Seq, due to needing to generate each reply word-by-word. The middle group - SimSR, M-CVAE and M-MMR - is characterised by methods that comprise a more involved diversification pipeline. The final and fastest group includes STAR, M-Topic and Matching, where no additional post-hoc diversification is required (for M-Topic the topics can be pre-computed prior to inference)." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.551, + 0.645, + 0.567 + ], + "angle": 0, + "content": "5.4 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.581, + 0.885, + 0.92 + ], + "angle": 0, + "content": "Table 4 presents a case study on the DailyDialog Test set. We compare our approach, STAR, with the top-performing baseline from Table 2, SimSR. In both examples we consistently find STAR is able to output a broader range of intents. Quantitatively, we consider the rank that each suggestion receives according to the initial retrieval of the Matching model that underlies SimSR. We see that STAR is able to perform a much more global search across the reply space, selecting replies from within the top 100 or so ranks. This would be difficult for the standard retrieve-and-erank approach to emulate, given 100 is usually too large a number to efficiently rerank (Deb et al., 2019). Qualitatively, SimSR's suggestions converge around common phrases, e.g. 'let's go', which would be difficult to deduplicate with a heuristic rule given only a limited number of overlapping words between the replies. Conversely, STAR is able to represent a broader range of intents, such as replying with a question in both examples. Further examples are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.881, + 0.489, + 0.92 + ], + "angle": 0, + "content": "Results surpass reported numbers in original paper due to inclusion of language-model bias and deduplicated candidate pool to support better comparability (Section 4.1)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.096, + 0.862, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.223, + 0.883, + 0.253 + ], + "angle": 0, + "content": "Figure 2: Comparison of overall relevance and diversity scores across ablations, obtained by averaging across all three datasets with equal weighting." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.303, + 0.486, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.505, + 0.489, + 0.547 + ], + "angle": 0, + "content": "Figure 3: Comparison of run-time efficiency between STAR and the baseline methods. Results are calculated over the Reddit Validation set." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.602, + 0.472, + 0.811 + ], + "angle": 0, + "content": "
Message:Hi, Kenny. Let's go for a drink.
SimSR- let's go! [#9] - ok, let's go. [#3] - ok. let's get something to drink. [#1]
STAR- ok. let's go. [#5] - you want something to drink? [#89] - good idea. [#105]
Message:Of course! Let's go.
SimSR- let's go! [#1] - ok, let's go. [#5] - all right. let's go. [#12]
STAR- let's go! [#1] - where are we? [#43] - good idea! [#85]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.821, + 0.49, + 0.892 + ], + "angle": 0, + "content": "Table 4: Example model outputs from the DailyDialog Test set, comparing STAR (ours) with the top-performing baseline method. Numbers in bold indicate the ranking the reply received according to the Matching model." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.278, + 0.697, + 0.293 + ], + "angle": 0, + "content": "provided in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.306, + 0.642, + 0.321 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.333, + 0.885, + 0.542 + ], + "angle": 0, + "content": "We introduce STAR, an autoregressive retrieval system for SR, which is an end-to-end text-to-text model that sequentially predicts replies conditioned on an initial message. To train STAR, we demonstrate an approach to bootstrap a dataset of high-quality (message, reply set) pairs, from regular dialogue datasets containing only (message, reply) pairs. Empirically, our results show significant improvement over existing state-of-the-art SR baselines, across multiple datasets, corresponding to a \\(5.1\\% - 17.9\\%\\) improvement in relevance, and a \\(0.5\\% - 63.1\\%\\) improvement in diversity compared to the best baseline approach." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.543, + 0.884, + 0.785 + ], + "angle": 0, + "content": "Future work could extend these techniques to other set-prediction tasks: e.g., in IR the relevance of each document depends on the quantity of new information it contains compared to other documents in the set. In recommender systems, use cases include: tailoring a user's news feed requires that the news articles presented are not simply duplicates of the same story; designing a bespoke music playlist requires songs to be unified by common themes but also sufficiently distinct from one another to maintain the listener's interest. Other lines of future work include considering alternate strategies for initialising the reply embeddings, beyond the bag-of-words initialisation demonstrated in this paper." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.797, + 0.682, + 0.813 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We thank the reviewers for their helpful feedback and suggestions. This work is partly supported by the EPSRC DTP Studentship program. The opinions expressed in this paper are the authors', and are not necessarily shared/endorsed by their employers and/or sponsors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7618" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.099 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.111, + 0.49, + 0.496 + ], + "angle": 0, + "content": "Although our work shows that STAR is able to absorb sufficient information about the replies in its weights, this may become increasingly challenging when larger numbers of replies need to be embedded. One notable instance of this would be the multilingual setting, as many SR systems are deployed globally. In this case, each language typically has its own candidate pool. A naive implementation which creates separate reply vectors for each language would incur a significant increase in model size. In this case, we hypothesise techniques around weight-sharing between reply embeddings between languages may be beneficial, e.g. 'how are you' (en) and 'ça va' (fr) sharing the same vector. Further, our techniques are only demonstrated in publicly available datasets, whereas proprietary conversations in chat and email applications may have unique features not accounted for here (e.g. timestamps, cc and bcc information, and file attachments). Our technique also requires a planning algorithm to create the initial dataset. This theoretically creates an upper bound to the overall performance of STAR, as it is limited to cloning the behaviour of the offline planning algorithm." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.523, + 0.214, + 0.538 + ], + "angle": 0, + "content": "References" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.546, + 0.489, + 0.639 + ], + "angle": 0, + "content": "Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor Silva, Leonardo Neves, and Francesco Barbieri. 2022. Twitter topic classification. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3386-3400, Gyeongju, Republic of Korea. International Committee on Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.648, + 0.49, + 0.702 + ], + "angle": 0, + "content": "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In NeurIPS." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.711, + 0.49, + 0.881 + ], + "angle": 0, + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.892, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021a. Highly parallel autoregressive entity linking with dis" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.086, + 0.883, + 0.114 + ], + "angle": 0, + "content": "criminative correction. In Conference on Empirical Methods in Natural Language Processing." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.121, + 0.885, + 0.188 + ], + "angle": 0, + "content": "Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021b. Autoregressive entity retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.196, + 0.885, + 0.275 + ], + "angle": 0, + "content": "Nicola De Cao, Ledell Yu Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2021c. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.283, + 0.885, + 0.35 + ], + "angle": 0, + "content": "Jaime G. Carbonell and Jade Goldstein-Stewart. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Annual International ACM SIGIR Conference on Research and Development in Information Retrieval." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.358, + 0.885, + 0.397 + ], + "angle": 0, + "content": "Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. 2020. Evaluation of text generation: A survey. ArXiv, abs/2006.14799." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.406, + 0.885, + 0.473 + ], + "angle": 0, + "content": "Nimesh Chakravarthi and Jeff Pasternack. 2017. Building smart replies for member messages. press release. https://engineering.linkedin.com/blog/2017/10/building-smart-replies-for-member-messages." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.48, + 0.885, + 0.546 + ], + "angle": 0, + "content": "Budhaditya Deb, Peter Bailey, and Milad Shokouhi. 2019. Diversifying reply suggestions using a matching-conditional variational autoencoder. In North American Chapter of the Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.555, + 0.885, + 0.648 + ], + "angle": 0, + "content": "Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A conditional generative matching model for multi-lingual reply suggestion. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1553–1568, Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.655, + 0.885, + 0.696 + ], + "angle": 0, + "content": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2020. Revisiting self-training for neural sequence generation. In Proceedings of ICLR." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.704, + 0.885, + 0.77 + ], + "angle": 0, + "content": "Matthew L. Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. Efficient natural language response suggestion for smart reply. CoRR, abs/1705.00652." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.778, + 0.885, + 0.818 + ], + "angle": 0, + "content": "Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.827, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409-14428, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.086, + 0.49, + 0.155 + ], + "angle": 0, + "content": "Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multisentence scoring. In International Conference on Learning Representations." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.163, + 0.49, + 0.215 + ], + "angle": 0, + "content": "Gautier Izacard and Edouard Grave. 2021. Distilling knowledge from reader to retriever for question answering. In International Conference on Learning Representations." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.226, + 0.49, + 0.28 + ], + "angle": 0, + "content": "Youngsoo Jang, Jongmin Lee, and Kee-Eung Kim. 2020. Bayes-adaptive monte-carlo planning and learning for goal-oriented dialogues. In AAAI Conference on Artificial Intelligence." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.29, + 0.49, + 0.344 + ], + "angle": 0, + "content": "Youngsoo Jang, Seokin Seo, Jongmin Lee, and Kee-Eung Kim. 2021. Monte-carlo planning and learning with language action value estimates. In International Conference on Learning Representations." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.354, + 0.49, + 0.448 + ], + "angle": 0, + "content": "Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Gregory S. Corrado, Laszlo Lukács, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart reply: Automated response suggestion for email. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.457, + 0.489, + 0.511 + ], + "angle": 0, + "content": "Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In International Joint Conference on Natural Language Processing." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.521, + 0.489, + 0.562 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.572, + 0.49, + 0.613 + ], + "angle": 0, + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.623, + 0.49, + 0.689 + ], + "angle": 0, + "content": "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In International Conference on Learning Representations." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.699, + 0.49, + 0.766 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.776, + 0.49, + 0.856 + ], + "angle": 0, + "content": "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric Michael Smith, Y.-Lan Boureau, and Jason Weston. 2020. Recipes for building an open-domain chatbot. In Conference of the European Chapter of the Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.866, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.086, + 0.885, + 0.152 + ], + "angle": 0, + "content": "Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.165, + 0.885, + 0.258 + ], + "angle": 0, + "content": "Timo Schick and Hinrich Schütze. 2021a. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.271, + 0.885, + 0.351 + ], + "angle": 0, + "content": "Timo Schick and Hinrich Schütze. 2021b. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.364, + 0.885, + 0.443 + ], + "angle": 0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, L. Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy P. Lillicrap, and David Silver. 2019. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588:604 - 609." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.456, + 0.885, + 0.548 + ], + "angle": 0, + "content": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lancot, L. Sifre, Dharshan Kumaran, Thore Graepel, Timothy P. Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. ArXiv, abs/1712.01815." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.562, + 0.885, + 0.63 + ], + "angle": 0, + "content": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.641, + 0.885, + 0.721 + ], + "angle": 0, + "content": "Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.734, + 0.885, + 0.827 + ], + "angle": 0, + "content": "Benjamin Towle and Ke Zhou. 2022. Learn what is possible, then choose what is best: Disentangling one-to-many relations in language through text-based games. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4955–4965, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Benjamin Towle and Ke Zhou. 2023. Model-based simulation for optimising smart reply. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12030-12043, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7620" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.086, + 0.49, + 0.192 + ], + "angle": 0, + "content": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.2, + 0.49, + 0.268 + ], + "angle": 0, + "content": "Yue Weng, Huaixiu Zheng, Franziska Bell, and Gokhan Tur. 2019. Occ: A smart reply system for efficient in-app communications. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.276, + 0.49, + 0.318 + ], + "angle": 0, + "content": "Kilian Wenker. 2023. Who wrote this? how smart replies impact language and agency in the workplace. *Telematics and Informatics Reports*, 10:100062." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.325, + 0.49, + 0.445 + ], + "angle": 0, + "content": "Mozhi Zhang, Wei Wang, Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A dataset and baselines for multilingual reply suggestion. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1207-1220, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.453, + 0.49, + 0.521 + ], + "angle": 0, + "content": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur D. Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.528, + 0.49, + 0.609 + ], + "angle": 0, + "content": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B. Dolan. 2019. Dialogpt: Large-scale generative pre-training for conversational response generation. In Annual Meeting of the Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.617, + 0.49, + 0.658 + ], + "angle": 0, + "content": "Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In ACL." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.669, + 0.352, + 0.685 + ], + "angle": 0, + "content": "A Ethical Considerations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Controlling the outputs of dialogue models is a forefront issue in ethics research for AI, particularly with the impact of recent gains in LLM capabilities. We believe the risks in the case of SR systems have several mitigants compared to this: the replies can be vetted by humans before deployment; the replies are usually in short-form, rather containing complex information the user may rely on; ultimately, a user must select one of the options, rather than the system being able to reply without user oversight. Conversely, there are some risks more unique to SR systems that should be mentioned. Particularly, the suggestions presented by the system can have subtle priming effects on user behaviour. Notably," + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.198 + ], + "angle": 0, + "content": "users have been shown to be slightly more positive in the sentiment of their emails when shown suggested replies (Wenker, 2023). SR systems are on the whole known to produce more positive sentiment messages than the human distribution (Kannan et al., 2016). We see this as an extension of the broader trend of LLMs to be overly obsequious." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.209, + 0.753, + 0.226 + ], + "angle": 0, + "content": "B Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.234, + 0.885, + 0.395 + ], + "angle": 0, + "content": "For constructing the training dataset, we use the following hyperparameters: SimSR is initialised from the distilbert-base-uncased checkpoint (Sanh et al., 2019). We set the search parameters to \\(N = 100\\) and \\(M = 100\\). We use a redundancy penalty of 0.05 and a blending alpha of 0.75 for query augmentation. Both parameters provided a good trade-off between relevance and diversity in early testing, so we did not search hyperparameters further (see Section 5.2 for ablations)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.396, + 0.885, + 0.685 + ], + "angle": 0, + "content": "For training STAR, we initialise our model with the t5-small checkpoint. Note that this version of T5 has a comparable parameter count to the baselines which use DistilBERT (60M versus 66M). We tokenise the dataset with a maximum message length of 64 tokens. We train our model for up to 100k steps, with a warmup of 1k steps. In practice, the model typically converged around 20k steps. We use the AdamW optimiser (Loshchilov and Hutter, 2019) with an initial learning rate of 5e-4 and linear decay. We evaluate every 2k steps, by taking the ROUGE and Self-ROUGE scores on the validation set (this uses the ground-truth from the original dataset, not from the bootstrapped dataset), and employ early stopping once both metrics have ceased to improve. Note, we found these metrics were a much more reliable stopping point than crossentropy loss, which typically converged much earlier." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.697, + 0.733, + 0.713 + ], + "angle": 0, + "content": "C Further Case Studies" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.722, + 0.885, + 0.77 + ], + "angle": 0, + "content": "Table 5 displays further examples of STAR's predictions versus SimSR, taken from the PersonaChat Test set." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "7621" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.233, + 0.264, + 0.766, + 0.695 + ], + "angle": 0, + "content": "
Persona:my favorite tv show is the office.
i like jazz music.
i do not drive because i live in new york.
i am jewish.
i do not eat anything sweet.
Message:i was but am now divorced
SimSR- i am sorry to hear that .
- oh . i am sorry to hear that .
- i am sorry to hear that . do you have any kids ?
STAR- oh i am sorry to hear that
- do you like music ?
- like music ? i love jazz
Persona:hey there my name is jordan and i am a veterinarian.
love to read drama books.
i love playing video games.
i am also a musician on the weekends.
i am originally from california but i live in florida.
Message:sometimes . i listen to a lot of music . do you read a lot ?
SimSR- hi how are you ?
- hi how are you today
- i do not like music at all
STAR- i listen to music do you
- no not really i like all music .
- yes i do my favorite is country
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.705, + 0.884, + 0.734 + ], + "angle": 0, + "content": "Table 5: Example model outputs from the PersonaChat Test set, comparing STAR (ours) with the top-performing baseline method. STAR is able to capture a broader range of intents through its end-to-end autoregressive retrieval." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7622" + } + ] +] \ No newline at end of file diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_origin.pdf b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a1a58027bf03f9109ae445abbf2ff0be5fd0d386 --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/bf60ac95-a87c-4b98-a7d6-066f76db4f85_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f76aab952400396502db41a72976c5d17c900d1d79c34ee8f8206f997fdca4 +size 713369 diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/full.md b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f333542124ab0efc47a230c5c8bd30e2420e180 --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/full.md @@ -0,0 +1,382 @@ +# End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems + +Benjamin Towle1, Ke Zhou1,2 + +1University of Nottingham + +$^{2}$ Nokia Bell Labs + +{benjamin.towle, ke.zhou}@nottingham.ac.uk + +# Abstract + +Reply suggestion systems represent a staple component of many instant messaging and email systems. However, the requirement to produce sets of replies, rather than individual replies, makes the task poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. As a result, these system often rely on additional post-processing modules to diversify the outputs. However, these approaches are ultimately bottlenecked by the performance of the initial retriever, which in practice struggles to present a sufficiently diverse range of options to the downstream diversification module, leading to the suggestions being less relevant to the user. In this paper, we consider a novel approach that radically simplifies this pipeline through an autoregressive text-to-text retrieval model, that learns the smart reply task end-to-end from a dataset of (message, reply set) pairs obtained via bootstrapping. Empirical results show this method consistently outperforms a range of state-of-the-art baselines across three datasets, corresponding to a $5.1\% - 17.9\%$ improvement in relevance, and a $0.5\% - 63.1\%$ improvement in diversity compared to the best baseline approach. We make our code publicly available. $^{1}$ + +# 1 Introduction + +Reply suggestion, or smart reply (SR), systems are a staple component of many commercial applications such as Gmail, Skype, Outlook, Microsoft Teams, LinkedIn and Facebook Messenger. They help the user process chats and emails quicker by offering a set of canned replies which can be clicked without requiring manual typing. However, dialogue is known to be a one-to-many problem (Zhao et al., 2017; Towle and Zhou, 2022) – namely, for any given message, there are multiple possible replies. To reflect this uncertainty, systems should + +present a diverse set of options to the user. For instance, given the message How are you?, an SR system could suggest: {I'm good; Ok; Not great}. Resultantly, the quality of a given reply depends not only on the message, but on the other replies in the reply set. + +Several prior works explore solutions to this problem such as removing near duplicates, penalising inter-reply similarity (Deb et al., 2019), clustering by intent (Henderson et al., 2017; Weng et al., 2019), learning latent variables (Deb et al., 2019, 2021), or model-based simulation (Towle and Zhou, 2023). However, these methods share a common design choice (Figure 1A): (1) a retrieval-based Matching model, which has learned a shared embedding space between messages and replies, returns a shortlist of top scoring replies; (2) this shortlist is refined through some diversification procedure to obtain the final reply set. + +Unfortunately, this assumes that the initial shortlist contains at least one good reply set. In practice, we find Matching models often search myopically, only retrieving candidates that are very similar to one another (Figure 1A). Thus, the chosen reply set often fails to reflect a diverse range of user intents, while latency constraints make more sophisticated diversification techniques or larger shortlists prohibitive (Deb et al., 2019). + +An intuitive, but – to the best of our knowledge – unexplored, solution to this problem is to conduct the retrieval autoregressively, with each reply conditioned on both the initial message and the previous replies in the set. Unfortunately, this approach encounters a second problem, namely, the lack of any datasets containing (message, reply set) pairs (Towle and Zhou, 2023). In practice, SR systems are trained on individual (message, reply) pairs obtained from conversation datasets, while the task of presenting multiple diverse replies to the user is outsourced to a separate diversification module. + +![](images/e50b983ca8bdf7fb68ad09f880c265e3c03c12402818f04d12ec540b52ff6509.jpg) + +![](images/eff81584e09698eba5df1bd7394244c42071a8e9d49cc8fef307cc4b79ef12e4.jpg) +Figure 1: Previous methods [A] compared to our approach, STAR [B]. The example displayed is taken from the DailyDialog Test set, and compares the predictions of STAR with SimSR (Towle and Zhou, 2023), the next best method. Our method's suggestions present a diverse range of topics/intents to drive the conversation. + +To meet this dual need, we present both (i) a bootstrapping method for creating a high-quality dataset of (message, reply sets) and (ii) a novel autoregressive retrieval model which predicts sequences of replies. For solving (i), we observe how model-based planning algorithms have been known to serve as a powerful policy improvement operator (Silver et al., 2017; Schrittwieser et al., 2019), including in several NLP systems (Jang et al., 2020, 2021). Specifically, the outputs of a model-based planning algorithm can be used to bootstrap a SR system. Further, by conducting this planning offline we are able to leverage two key advantages: (1) the system is free of the latency constraints of online inference, and therefore can increase the search space coverage of the planning algorithm; (2) the system can leverage information that would not be available during inference, such as the ground-truth reply, to further guide the search process. For (ii) we unify both steps of the standard SR pipeline into a single end-to-end model, which mitigates the myopic search, and allows the model to learn to diversify its predictions in a principled way through gradient-based learning. To this end, we present STAR (Suggested replies with T5 and Autoregressive Retrieval) (Figure 1B). At a high level, STAR is a text-to-text model trained to output sequences of replies, where each reply is conditioned both on the initial message and the previous replies in the sequence. Concretely, we instantiate our method with the T5 pretrained model (Raffel et al., 2020). We expand T5's vocabulary by + +treating each reply in the candidate pool as a novel token, and demonstrate a simple-yet-effective technique for initialising the new token embeddings, which leverages the model's existing semantic priors. Notably, by treating each reply as a token, we limit the number of autoregressive decoding steps required, keeping the model's efficiency comparable to other retrieval-based methods. + +Empirically, we evaluate our approach on three benchmarks: Reddit (Zhang et al., 2021), which is the only publicly-available SR benchmark, as well as PersonaChat (Zhang et al., 2018) and DailyDialog (Li et al., 2017) which are both widely-used in dialogue research more broadly (Zhang et al., 2019; Roller et al., 2020, inter alia), and share a similar conversational style with SR apps. We demonstrate superior performance over state-of-the-art baselines across all datasets, corresponding to a $5.1\% - 17.9\%$ improvement in relevance, and a $0.5\% - 63.1\%$ improvement in diversity compared to the best baseline approach. We further show comparable efficiency to previous methods, and perform a range of ablations to motivate our design choices. + +In summary, our key contributions are as follows: (1) an autoregressive retrieval architecture for sequentially predicting suggested replies; (2) a bootstrapping framework for generating high-quality data of (message, reply set) pairs; (3) detailed analysis of model behaviour and performance including a case study and ablation of key components. + +# 2 Related Work + +Smart reply The proprietary nature of data from email and chat applications has led several previous works to use publicly-available dialogue datasets (Zhang et al., 2021; Deb et al., 2021; Towle and Zhou, 2023) to benchmark SR methods, due to their analogous conversational nature. While early SR systems used generative models (Kannan et al., 2016), current production systems favour retrieval methods due to their greater controllability of outputs and superior latency (Deb et al., 2019). Increasing the diversity of reply suggestions is a key focus of previous work, which has been attempted by: (1) mapping replies to discrete intents / topics (Kannan et al., 2016; Chakravarthi and Pasternack, 2017; Weng et al., 2019); (2) re-weighting replies according to their similarity with other replies in the set (Carbonell and Goldstein-Stewart, 1998; Deb et al., 2019); (3) learning continuous latent variables to generate multiple queries (Zhao et al., 2017; Deb et al., 2019); (4) using model-based simulation to iteratively search and evaluate the relevance of candidate reply sets (Towle and Zhou, 2023). Our proposed method differs from all of these approaches in that our model learns to account for the interdependencies between replies through end-to-end backpropagation. + +Autoregressive retrieval Integrating neural retrieval into the well-established paradigm of text-to-text models is of growing interest. Earlier work focuses on outputting a document ID given a query (Tay et al., 2022). Further work has extended this by considering alternate ways of representing the document IDs, such as through unique substrings (Bevilacqua et al., 2022). Another line of work has used autoregressive retrieval for the entity linking task (Cao et al., 2021a,b,c). There, the motivation is to reduce the large number of entities by relying on the text-to-text model's pre-existing vocabulary, rather than having to retrieve embeddings from a memory-intensive dense index. Our proposed method differs considerably from these previous works both in instantiation and motivation. Instantiation-wise, we generate multiple replies - critical to making this possible is the novel bootstrapping technique for creating the dataset of (message, reply set) pairs to train on. Motivation-wise, our goal is to be able to condition each reply on both the input message and previous replies in the set, enabling the model to learn to predict se + +quences of replies in a differentiable way. + +Bootstrapping The idea of bootstrapping training data from limited resources has received significant recent interest in NLP, given the newly demonstrated few / zero-shot capabilities of many large language models (Brown et al., 2020). It has seen usage in few-shot shot text-classification (Schick and Schütze, 2021a), semantic similarity (Schick and Schütze, 2021b), tool-usage (Schick et al., 2023), retrieval (Izacard and Grave, 2021), sequence generation (He et al., 2020), and instruction-tuning (Honovich et al., 2023; Wang et al., 2023; Taori et al., 2023), amongst others. These techniques can also be seen as a form of knowledge distillation (Hinton et al., 2015), except that the training typically involves predicting the exact token targets, rather than using the soft probabilities of a teacher model. Although sometimes these techniques are used as an addition to supervised learning (He et al., 2020), in our case there are no datasets containing the ideal reply sets to suggest to the user. Instead, we must bootstrap this in a more unsupervised way, by transforming a dataset of (message, reply) pairs into a dataset of (message, reply set) pairs. + +# 3 Methodology + +In this section, we first describe the model-based planning process used to obtain the bootstrapped dataset of (message, reply set) pairs (Section 3.1). Then, we show how the STAR architecture can be trained on this dataset (Section 3.2). + +# 3.1 Offline Dataset Creation + +Our goal is to transform a dialogue dataset $\mathcal{D} = \{(x,y)\}$ of (message, reply) tuples, into a dataset $\mathcal{D}^* = \{(x,Y)\}$ where $Y$ is the set of replies $\{y_k\}^K$ to be presented to the user. Algorithm 1 summarises this process. While our method is general to any arbitrary planning algorithm, we choose to instantiate our approach with a modified version of SimSR (Towle and Zhou, 2023), a recently released publicly available state-of-the-art SR method, that employs model-based simulation to predict reply sets. As the original algorithm was designed for online inference, we make several changes to benefit the offline nature of our version, and detail the full implementation below. + +The initial retrieval is conducted by a Matching model $\Phi$ that separately encodes messages and replies into a shared latent space. Given an encoded + +Algorithm 1 Offline Dataset Creation. We use $N = 100$ , $M = 100$ , $\alpha = 0.75$ and $\lambda = 0.05$ as our default setting. + +Input Matching model $\Phi$ , message $x$ , precomputed reply vectors $\{\mathbf{y_r}\}^R$ , number of candidates $N$ , number of simulations $M$ , final reply set size $K$ , query augmentation coefficient $\alpha$ , redundancy penalty $\lambda$ . + +Output reply set $Y_{K}$ + +$$ +\begin{array}{l} \mathbf {x}, \mathbf {y} \leftarrow \Phi (x), \Phi (y) \\ \tilde {\mathbf {x}} \leftarrow \alpha \mathbf {x} + (1 - \alpha) \mathbf {y} \end{array} \quad \triangleright \text {q u e r y a u g m e n t a t i o n} +$$ + +$$ +\begin{array}{l} \bar {\mathbf {x}} \leftarrow \alpha \mathbf {x} + (1 - \alpha) \mathbf {y} \\ Y _ {N} \leftarrow N \underset {r} {\operatorname {a r g m a x}} (\bar {\mathbf {x}} \cdot \mathbf {y} _ {\mathbf {r}}) \end{array} \quad \triangleright \text {q u e r y a u g m e n t a t i o n} +$$ + +$$ +Y _ {M} \leftarrow M \operatorname {a r g m a x} _ {\mathbf {r}} (\tilde {\mathbf {x}} \cdot \mathbf {y} _ {\mathbf {r}}) +$$ + +$$ +\begin{array}{l} q (y _ {m} | x) \propto \exp \tilde {\mathbf {x}} \cdot \mathbf {y} _ {\mathbf {m}} \\ Y _ {G} \leftarrow \emptyset \end{array} \quad \triangleright \text {s o f t m a x o v e r t o p - M s c o r e s} +$$ + +$$ +\begin{array}{c}\textbf{for} k\gets 0\text{to} K\textbf{do}\\ y_{k}\leftarrow \operatorname *{argmax}_{n}\sum_{m}^{M}f(Y_{G}^{n},y_{m})q(y_{m}|x) - \lambda f(Y_{G},y_{n}) \end{array} +$$ + +$$ +Y _ {G} \leftarrow Y _ {G} \cup y _ {k} +$$ + +end for + +$$ +Y _ {K} \leftarrow Y _ {G} +$$ + +$$ +\text {r e t u r n} Y _ {K} +$$ + +message $\mathbf{x} = \Phi (x)$ , it retrieves the top $N$ candidates from a pool of pre-computed reply vectors $\mathbf{Y}_{\mathbf{R}} = \{\mathbf{y}_{\mathbf{r}}\}^{R}$ by combining their dot product similarity with a pre-computed language-model bias - a standard component of SR systems to downweight overly specific replies (Deb et al., 2019). + +$$ +Y _ {N} = N \underset {r} {\operatorname {a r g m a x}} \left(\mathbf {x} \cdot \mathbf {y} _ {\mathbf {r}} + \beta \mathrm {L M} \left(y _ {r}\right)\right) \tag {1} +$$ + +We then output the $K$ -tuple $Y_{i} \in \binom{Y_{N}}{K}$ that has the highest expected similarity with the human reply, according to some similarity function $f(\cdot, \cdot)$ . + +$$ +\underset {i} {\operatorname {a r g m a x}} \mathbb {E} _ {y \sim p (\cdot | x)} \left[ f (Y _ {i}, y) \right] \tag {2} +$$ + +Given the objective in SR is for at least one of the replies to be relevant, the similarity function is defined as a maximum over the sampled reply and each of the replies in the reply set, using term-level F1-score: $\max_k\mathrm{F1}(y_k,y)$ + +We assume $y$ is sampled from the ground-truth human distribution $p(\cdot |x)$ . As we do not have access to the true human distribution in practice, we instead use the same Matching model $q$ as a proxy for this, given it is trained on (message, reply) pairs. We then approximate the expectation by marginalising over the top- $M$ most likely replies: + +$$ +\approx \underset {i} {\operatorname {a r g m a x}} \sum_ {m} ^ {M} f \left(Y _ {i}, y _ {m}\right) q \left(y _ {m} \mid x\right) \tag {3} +$$ + +In practice, it is intractable to evaluate every possible reply tuple, due to their combinatorial scaling. We therefore approximate this by greedily constructing the reply set one reply at a time. Formally, + +let $Y_{G}$ be the set of currently selected replies, such that initially $Y_{G} = \emptyset$ . Then, for each of $y_{n} \in Y_{N}$ , we compute the expected similarity for the union of $Y_{G}$ and $y_{n}$ , termed $Y_{G}^{n} = Y_{G} \cup y_{n}$ for brevity: + +$$ +\sum_ {m} ^ {M} f \left(Y _ {G} ^ {n}, y _ {m}\right) q \left(y _ {m} \mid x\right) \tag {4} +$$ + +We repeat this process for $K$ timesteps, each time appending the highest scoring reply to $Y_{G}$ , i.e. until $|Y_{G}| = K$ . Note that this greedy search process implicitly canonicalises the order of the replies, as selecting replies in this way causes them to be roughly ordered by individual message-reply relevance. + +# 3.1.1 Adjustments + +Scaling $N$ and $M$ The original SimSR algorithm was used only in an online setting (Towle and Zhou, 2023). Therefore, the size of the search parameters $N$ (number of replies in the shortlist) and $M$ (number of simulated user replies) is kept low (15 and 25 respectively in the original paper). As we only need to run this model offline however to obtain the dataset, we find setting $N$ and $M$ to much larger values improves relevance (we use 100 for both), enabling both a broader search (i.e. by increasing $N$ ) and a more accurate similarity function (i.e. by increasing $M$ ). + +Redundancy penalty Early testing showed that scaling the search parameters reduced diversity. We therefore introduce a redundancy penalty, which penalises the model for selecting replies that are similar to replies already in the set $Y_{G}$ . This is analogous to the inter-document similarity penalty used in the maximum marginal relevance IR (information retrieval) technique (Carbonell and Goldstein-Stewart, 1998). + +$$ +\sum_ {m} ^ {M} f \left(Y _ {G} ^ {n}, y _ {m}\right) q \left(y _ {m} \mid x\right) - \lambda f \left(Y _ {G}, y _ {n}\right) \tag {5} +$$ + +Query augmentation Unlike during online inference, we also have access to the ground-truth reply $y$ when constructing the dataset. Previous work has found that models obtain greater representational capabilities when given access to posterior information (Paranjape et al., 2022; Towle and Zhou, 2022). We therefore use an augmented query to retrieve with the Matching model. This is obtained by interpolating between the message and ground-truth reply embeddings. This biases the model's predictions towards the observed ground-truth in + +the dataset, while still allowing it to benefit from its own learned distribution. + +$$ +\tilde {\mathbf {x}} = \alpha \Phi (x) + (1 - \alpha) \Phi (y) \tag {6} +$$ + +# 3.2 Proposed STAR Model + +We initialise STAR with a T5-based text-to-text language model, which has previously been shown to be effective in autoregressive retrieval (Tay et al., 2022). While some autoregressive retrieval approaches identify their documents/replies through unique substrings (Bevilacqua et al., 2022) or constrained beam search (Cao et al., 2021b), we focus on approaches requiring only a limited number of autoregressive steps, to maintain competitive inference speeds to existing retrieval methods (Section 5.3). There are several alternatives for this such as treating each reply set as a unique token, or separately training on each (message, reply pair), but ultimately we opted for autoregressively treating each reply as a unique token in the vocabulary in order to exploit the compositionality of reply sets (Section 5.2 for performance comparison). Note that as the types of replies used in smart reply are usually quite short and concise, e.g. 'how are you', 'I'm fine thanks', 'yes, that's right' etc., systems in deployment only need to retrieve from a pool of $30\mathrm{k}$ or so replies (Deb et al., 2019), in order to provide good coverage of possible user intents. As a result, we are able to keep the size of the vocabulary reasonable. Thus, our new vocabulary is defined as: $W_{tokens} \cup W_{replies}$ . An obvious challenge to this approach is that by treating each reply as a previously unseen word, it removes any semantic priors the model might have about their meaning. To mitigate this, we employ a bag-of-words initialisation strategy. Hence, we define the embedding of the $t$ -th reply $E(y_{t})$ as the average over the embeddings of the individual words within $w_{n} \in y_{t}$ . + +$$ +E \left(y _ {t}\right) = \frac {1}{N} \sum_ {n} ^ {N} E \left(w _ {n}\right) \tag {7} +$$ + +Intuitively, this ensures that the initial embeddings are close to the word embeddings of the original vocabulary, while also capturing some of the underlying semantics of the reply. We allow the weights to update during fine-tuning. Note that for T5 the output and input embedding layers share weights, and therefore this approach is used to initialise both layers. We train the model using cross-entropy loss to predict the next reply given the current sequence + +of replies and messages: + +$$ +\mathcal {L} _ {N L L} = - \sum_ {k} ^ {K} \log p \left(y _ {k} \mid x, y _ {0}, \dots , y _ {k - 1}\right) \tag {8} +$$ + +# 4 Experimental Setup + +# 4.1 Baselines + +Previous work has largely been closed-source and is therefore unavailable for direct comparison (Henderson et al., 2017; Weng et al., 2019; Deb et al., 2019). With the exception of SimSR, which has publicly available code $^{2}$ , we re-implement a variety of methods that cover the broad range of previous techniques. Due to its comparable size, all baselines apart from Seq2Seq are initialised with DistilBERT as the encoder backbone. These are summarised as follows: + +Seq2Seq is a generative encoder-decoder. While current production systems and the majority of related works use only retrieval models (Deb et al., 2019; Towle and Zhou, 2023), at least one related work includes a standard generative transformer as a baseline (Zhang et al., 2021), which we follow here. For maximum comparability with our method, we use the same t5-small model as a backbone. For each message, we sample $K$ responses independently. + +Matching represents the out-of-the-box encoder with no additional diversification strategy and was used as a baseline method by Zhang et al. (2021). It simply selects the top $K$ responses according to individual message-reply scores. + +Matching-Topic uses an out-of-the-box topic classifier to ensure no two replies share the same topic, similar to previous work (Henderson et al., 2017; Weng et al., 2019). The classifier is trained on Twitter (Antypas et al., 2022), due to their comparable short-form open-domain chat conversations. + +Maximum Marginal Relevance (MMR) (Carbonell and Goldstein-Stewart, 1998) is originally an IR technique, used in several previous SR works (Deb et al., 2019; Towle and Zhou, 2023), which re-weights reply scores as a linear combination of their message-reply and inter-reply similarity. + +
TrainValidTest|YR|
Reddit50k5k5k48k
PersonaChat66k8k8k64k
DailyDialog76k7k7k62k
+ +Table 1: Number of samples in the Train, Validation, Test sets and Candidate pool in the three datasets for evaluation. The Candidate pool comprises the Train set with duplicate responses removed. + +MCVAE (Deb et al., 2019) is a conditional variational autoencoder (Zhao et al., 2017) which learns to generate multiple query vectors from a single message embedding, representing the multiple possible reply intents. Candidates are scored via a voting process, whereby the $K$ most-selected replies are chosen. + +SimSR (Towle and Zhou, 2023) uses an iterative search and evaluation process to select possible reply sets and score them according to their expected similarity from a learned world model, which serves as a proxy for the user. To ensure comparability of SimSR with our method and the other baselines, we include the language-model bias in the scoring process (Equation 1), and also deduplicate the candidate pool. $^{3}$ + +# 4.2 Datasets + +We evaluate our proposed method across three datasets, summarised in Table 1. Below, we describe the datasets in more detail and motivate their inclusion. Note, other than Reddit, there are no publicly available SR datasets, due to their commercial nature (e.g. Henderson et al. (2017); Deb et al. (2019); Weng et al. (2019)). Therefore, we adopt several dialogue datasets, which is the closest alternative to conversations on proprietary chat applications. + +Reddit (Zhang et al., 2021) was originally introduced for training multilingual SR systems, and is the only publicly available dataset specifically intended for SR purposes. As the original dataset is very large, we follow Towle and Zhou (2023) and use the reduced version of the dataset. Note, this version only contains English, as our aim is limited to the monolingual setting. Due to the organic nature of the dataset, conversations cover a very broad range of topics. + +PersonaChat (Zhang et al., 2018) is a crowdworker-sourced dataset comprising personagrounded conversations, in which each speaker is assigned a persona comprising a few short sentences. Following previous methods (Humeau et al., 2020), we concatenate the persona to the beginning of the message. The participants are instructed to chat naturally and to try to get to know one another. + +DailyDialog (Li et al., 2017) is a dataset created from English language learning websites and consists of a variety of high-quality dialogues in everyday scenarios. The dataset differs from the former two in that the conversations often involve real-life scenarios, such as asking for directions, and therefore captures a different variety of conversational skills. + +# 4.3 Metrics + +We evaluate our method on the same weighted ROUGE ensemble as previous methods (Lin, 2004; Deb et al., 2019, 2021), which is known to correlate well with click-through rate (Zhang et al., 2021): + +$$ +\frac {\text {R O U G E} - 1}{6} + \frac {\text {R O U G E} - 2}{3} + \frac {\text {R O U G E} - 3}{2} \tag {9} +$$ + +As the goal of SR systems it to ensure that at least one of the suggested replies is relevant to the user, we only record the maximum ROUGE score across each of the $K = 3$ suggested replies. We also evaluate the model on Self-ROUGE (Celikyilmaz et al., 2020): This is an unreferenced metric that measures the internal dissimilarity (i.e. diversity) within the reply set by treating one reply as the predicted reply and the other parts as the references. Note that a lower Self-ROUGE score indicates more diversity. + +# 4.4 Inference + +For inference, we use the entire training set as the candidate pool for each respective dataset, with dedduplication to remove exact matches. For STAR, we greedily decode the next reply token until $K$ tokens have been decoded. Note, we only allow the model to output replies represented in the bootstrapped dataset, and also block non-replies, i.e. words from the original vocabulary, from being predicted. + +# 5 Experimental Results + +We focus our efforts on answering the following Research Questions: $(\mathbf{RQ}_1)$ How does STAR com + +pare to existing state-of-the-art methods? (Section 5.1, 5.4); $(\mathbf{RQ}_2)$ Which components of the data collection algorithm and fine-tuning have the largest impact on STAR's performance? (Section 5.2); $(\mathbf{RQ}_3)$ How efficient is STAR in inference? (Section 5.3) + +# 5.1 Main Results + +Table 2 compares the performance of different SR systems across the Reddit, PersonaChat and Dialy-Dialog datasets. In terms of relevance (ROUGE), STAR shows an especially large improvement in Reddit $(+17.9\%)$ and DailyDialog $(+15.8\%)$ . We hypothesise the gains in PersonaChat $(+5.1\%)$ are more modest because the replies are more easily predicted due to the persona, which is concatenated to each message. This significantly reduces the noise during the initial retrieval for the baselines, as they only need to retrieve the messages relevant to that particular persona. + +For diversity (Self-ROUGE), the strongest gains were found in DailyDialog $(+63.1\%)$ . For PersonaChat, STAR performs much better than the retrieval methods, only falling behind Seq2Seq, due to its altogether noisier outputs as evidenced by having the worst relevance score. The Reddit results were comparatively more modest $(+0.5\%)$ - we hypothesise this is because the dataset is altogether more noisy, and so there are relatively few similar replies in the dataset, as shown by the Self-ROUGE scores being lower than the other two datasets. Overall, the consistent outperformance in both relevance and diversity metrics supports the benefits of the STAR approach. + +# 5.2 Ablation + +In Table 3, we conduct ablations across two key axes: data collection and STAR training. The data collection ablations serve to investigate the benefits of the novel changes to the SimSR algorithm from Section 3.1.1. The STAR training ablations investigates the degree to which the improvements in performance are caused by the bootstrapped dataset or by STAR's architecture itself; we achieve this by considering several alternative variants of STAR. Our data collection ablations consider two features: (A) removing the query augmentation prevents the model from leveraging any ground truth information during prediction; (B) removing the redundancy penalty no longer explicitly penalises lack of diversity in predicted reply sets. For STAR training, we consider three alternative configurations: + +(C) we replace the bag-of-words embeddings with randomly initialised embeddings - this removes any priors about the meaning of replies and forces the model to learn them tabula rasa; (D) we treat each reply set as a unique token - this removes the compositional element from the task, constraining the model to only predicting previously seen reply sets, therefore testing whether the model is capable of learning to compose novel reply sets; (E) we remove the ability to account for interdependencies between replies, by restructuring each (message, reply set) data point into $K$ data points of (message, reply $k$ ), and then outputting the top- $K$ replies during inference - this investigates whether the benefit lies simply in the bootstrapped dataset being better suited to the SR task, rather than in STAR's ability to account for interdependencies between replies. + +In terms of data collection ablations, we found removing the redundancy penalty significantly reduced the diversity of predictions, although in some cases offered slightly improved relevance; removing the query augmentation generally led to a worse relevance/diversity trade-off. For the variants of STAR training, we found that random embeddings consistently reduced relevance, while also led to less diverse predictions; reply sets as tokens led to the most competitive variant of STAR compared to our default setup: diversity was overall better, due to using preconstructed reply sets from the offline planning algorithm, but this came at the trade-off of reduced flexibility from being unable to construct novel reply sets when the context required it – resultantly, we saw a corresponding reduction in relevance. Finally, predicting replies separately expectedly harmed both relevance and diversity, demonstrating the importance of accounting for reply interdependencies. + +In Figure 2, we further validated the individual results of our ablation by aggregating the results across datasets (applying an equal weighting to each dataset). This demonstrates the overall trend that the default STAR offers the superior trade-off between relevance and diversity, while treating reply sets as tokens offered the next best alternative. Nevertheless, we believe that keeping individual replies as tokens – thus allowing the model to construct reply sets dynamically – is likely to be an attractive property for deployed systems, enabling the overall vocabulary size to remain modest. + +
MethodRedditPersonaChatDailyDialog
ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
Generative models
Seq2Seq2.413.436.836.88*4.013.91
Retrieval models
Matching1.959.427.5121.476.5316.65
M-Topic1.813.947.1615.436.1411.11
M-MMR2.204.447.8114.576.138.63
M-CVAE2.305.027.4312.216.7810.49
SimSR42.792.189.0410.526.824.80
STAR3.29*2.179.50*7.747.90*1.77*
+ +Table 2: Performance of STAR across Reddit, PersonaChat and DailyDialog Test sets on relevance (ROUGE) and diversity (Self-ROUGE) metrics. Bold indicates best result, underline indicates second-best. * = statistically significant versus next best result on t-test with $p$ -value < 0.01. + +
ModelRedditPersonaChatDailyDialog
ConfigurationROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
STAR3.35*2.278.857.488.391.81
Data Collection Ablations
A: No Query Augmentation2.942.008.996.947.242.89
B: No Redundancy Penalty3.064.299.0317.268.98*5.90
STAR Training Variants
C: Random embeddings2.674.938.3910.976.844.45
D: Reply sets as tokens2.851.59*8.766.817.751.57*
E: Predict replies separately2.2026.618.0730.986.4320.50
+ +Table 3: Performance of STAR on the Reddit, PersonaChat and DailyDialog Validation sets under different model configurations. Ablations are applied separately. **Bold indicates best result, **underline indicates second-best. * = statistically significant versus next best result on t-test with $p$ -value < 0.01. + +# 5.3 Run-time Efficiency + +Beyond performance gains in relevance and diversity, a major advantage of an autoregressive retrieval model is the ability to leverage the scalability of GPU-based inference. Figure 3 compares the efficiency of STAR with the other baseline methods. We use an NVIDIA GeForce RTX 3060 Ti GPU and AMD Ryzen 7 5700G with Radeon Graphics CPU, with a batch size of 32. The results show that the methods can be broadly clustered into three groups. The slowest group is the generative method Seq2Seq, due to needing to generate each reply word-by-word. The middle group - SimSR, M-CVAE and M-MMR - is characterised by methods that comprise a more involved diversification pipeline. The final and fastest group includes STAR, M-Topic and Matching, where no additional post-hoc diversification is required (for M-Topic the topics can be pre-computed prior to inference). + +# 5.4 Case Study + +Table 4 presents a case study on the DailyDialog Test set. We compare our approach, STAR, with the top-performing baseline from Table 2, SimSR. In both examples we consistently find STAR is able to output a broader range of intents. Quantitatively, we consider the rank that each suggestion receives according to the initial retrieval of the Matching model that underlies SimSR. We see that STAR is able to perform a much more global search across the reply space, selecting replies from within the top 100 or so ranks. This would be difficult for the standard retrieve-and-erank approach to emulate, given 100 is usually too large a number to efficiently rerank (Deb et al., 2019). Qualitatively, SimSR's suggestions converge around common phrases, e.g. 'let's go', which would be difficult to deduplicate with a heuristic rule given only a limited number of overlapping words between the replies. Conversely, STAR is able to represent a broader range of intents, such as replying with a question in both examples. Further examples are + +![](images/080a6316394fb7b9c651b39cc33d1f2c91230cb059f3f066da7cf0303ddeff7c.jpg) +Figure 2: Comparison of overall relevance and diversity scores across ablations, obtained by averaging across all three datasets with equal weighting. + +![](images/dd266042b070a64a7b1c683f7b0ba67d2a9c00c2ad4f53a0ce0094baedb945cb.jpg) +Figure 3: Comparison of run-time efficiency between STAR and the baseline methods. Results are calculated over the Reddit Validation set. + +
Message:Hi, Kenny. Let's go for a drink.
SimSR- let's go! [#9] - ok, let's go. [#3] - ok. let's get something to drink. [#1]
STAR- ok. let's go. [#5] - you want something to drink? [#89] - good idea. [#105]
Message:Of course! Let's go.
SimSR- let's go! [#1] - ok, let's go. [#5] - all right. let's go. [#12]
STAR- let's go! [#1] - where are we? [#43] - good idea! [#85]
+ +Table 4: Example model outputs from the DailyDialog Test set, comparing STAR (ours) with the top-performing baseline method. Numbers in bold indicate the ranking the reply received according to the Matching model. + +provided in Appendix C. + +# 6 Conclusion + +We introduce STAR, an autoregressive retrieval system for SR, which is an end-to-end text-to-text model that sequentially predicts replies conditioned on an initial message. To train STAR, we demonstrate an approach to bootstrap a dataset of high-quality (message, reply set) pairs, from regular dialogue datasets containing only (message, reply) pairs. Empirically, our results show significant improvement over existing state-of-the-art SR baselines, across multiple datasets, corresponding to a $5.1\% - 17.9\%$ improvement in relevance, and a $0.5\% - 63.1\%$ improvement in diversity compared to the best baseline approach. + +Future work could extend these techniques to other set-prediction tasks: e.g., in IR the relevance of each document depends on the quantity of new information it contains compared to other documents in the set. In recommender systems, use cases include: tailoring a user's news feed requires that the news articles presented are not simply duplicates of the same story; designing a bespoke music playlist requires songs to be unified by common themes but also sufficiently distinct from one another to maintain the listener's interest. Other lines of future work include considering alternate strategies for initialising the reply embeddings, beyond the bag-of-words initialisation demonstrated in this paper. + +# Acknowledgements + +We thank the reviewers for their helpful feedback and suggestions. This work is partly supported by the EPSRC DTP Studentship program. The opinions expressed in this paper are the authors', and are not necessarily shared/endorsed by their employers and/or sponsors. + +# Limitations + +Although our work shows that STAR is able to absorb sufficient information about the replies in its weights, this may become increasingly challenging when larger numbers of replies need to be embedded. One notable instance of this would be the multilingual setting, as many SR systems are deployed globally. In this case, each language typically has its own candidate pool. A naive implementation which creates separate reply vectors for each language would incur a significant increase in model size. In this case, we hypothesise techniques around weight-sharing between reply embeddings between languages may be beneficial, e.g. 'how are you' (en) and 'ça va' (fr) sharing the same vector. Further, our techniques are only demonstrated in publicly available datasets, whereas proprietary conversations in chat and email applications may have unique features not accounted for here (e.g. timestamps, cc and bcc information, and file attachments). Our technique also requires a planning algorithm to create the initial dataset. This theoretically creates an upper bound to the overall performance of STAR, as it is limited to cloning the behaviour of the offline planning algorithm. + +# References + +Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor Silva, Leonardo Neves, and Francesco Barbieri. 2022. Twitter topic classification. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3386-3400, Gyeongju, Republic of Korea. International Committee on Computational Linguistics. + +Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In NeurIPS. + +Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc. + +Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021a. Highly parallel autoregressive entity linking with dis + +criminative correction. In Conference on Empirical Methods in Natural Language Processing. + +Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021b. Autoregressive entity retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net. + +Nicola De Cao, Ledell Yu Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2021c. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290. + +Jaime G. Carbonell and Jade Goldstein-Stewart. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. + +Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. 2020. Evaluation of text generation: A survey. ArXiv, abs/2006.14799. + +Nimesh Chakravarthi and Jeff Pasternack. 2017. Building smart replies for member messages. press release. https://engineering.linkedin.com/blog/2017/10/building-smart-replies-for-member-messages. + +Budhaditya Deb, Peter Bailey, and Milad Shokouhi. 2019. Diversifying reply suggestions using a matching-conditional variational autoencoder. In North American Chapter of the Association for Computational Linguistics. + +Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A conditional generative matching model for multi-lingual reply suggestion. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1553–1568, Punta Cana, Dominican Republic. Association for Computational Linguistics. + +Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2020. Revisiting self-training for neural sequence generation. In Proceedings of ICLR. + +Matthew L. Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. Efficient natural language response suggestion for smart reply. CoRR, abs/1705.00652. + +Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531. + +Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409-14428, Toronto, Canada. Association for Computational Linguistics. + +Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multisentence scoring. In International Conference on Learning Representations. + +Gautier Izacard and Edouard Grave. 2021. Distilling knowledge from reader to retriever for question answering. In International Conference on Learning Representations. + +Youngsoo Jang, Jongmin Lee, and Kee-Eung Kim. 2020. Bayes-adaptive monte-carlo planning and learning for goal-oriented dialogues. In AAAI Conference on Artificial Intelligence. + +Youngsoo Jang, Seokin Seo, Jongmin Lee, and Kee-Eung Kim. 2021. Monte-carlo planning and learning with language action value estimates. In International Conference on Learning Representations. + +Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Gregory S. Corrado, Laszlo Lukács, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart reply: Automated response suggestion for email. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. + +Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In International Joint Conference on Natural Language Processing. + +Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics. + +Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations. + +Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In International Conference on Learning Representations. + +Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1). + +Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric Michael Smith, Y.-Lan Boureau, and Jason Weston. 2020. Recipes for building an open-domain chatbot. In Conference of the European Chapter of the Association for Computational Linguistics. + +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108. + +Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761. + +Timo Schick and Hinrich Schütze. 2021a. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online. Association for Computational Linguistics. + +Timo Schick and Hinrich Schütze. 2021b. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. + +Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, L. Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy P. Lillicrap, and David Silver. 2019. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588:604 - 609. + +David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lancot, L. Sifre, Dharshan Kumaran, Thore Graepel, Timothy P. Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. ArXiv, abs/1712.01815. + +Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca. + +Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems. + +Benjamin Towle and Ke Zhou. 2022. Learn what is possible, then choose what is best: Disentangling one-to-many relations in language through text-based games. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4955–4965, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. + +Benjamin Towle and Ke Zhou. 2023. Model-based simulation for optimising smart reply. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12030-12043, Toronto, Canada. Association for Computational Linguistics. + +Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics. + +Yue Weng, Huaixiu Zheng, Franziska Bell, and Gokhan Tur. 2019. Occ: A smart reply system for efficient in-app communications. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. + +Kilian Wenker. 2023. Who wrote this? how smart replies impact language and agency in the workplace. *Telematics and Informatics Reports*, 10:100062. + +Mozhi Zhang, Wei Wang, Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A dataset and baselines for multilingual reply suggestion. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1207-1220, Online. Association for Computational Linguistics. + +Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur D. Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Annual Meeting of the Association for Computational Linguistics. + +Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B. Dolan. 2019. Dialogpt: Large-scale generative pre-training for conversational response generation. In Annual Meeting of the Association for Computational Linguistics. + +Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In ACL. + +# A Ethical Considerations + +Controlling the outputs of dialogue models is a forefront issue in ethics research for AI, particularly with the impact of recent gains in LLM capabilities. We believe the risks in the case of SR systems have several mitigants compared to this: the replies can be vetted by humans before deployment; the replies are usually in short-form, rather containing complex information the user may rely on; ultimately, a user must select one of the options, rather than the system being able to reply without user oversight. Conversely, there are some risks more unique to SR systems that should be mentioned. Particularly, the suggestions presented by the system can have subtle priming effects on user behaviour. Notably, + +users have been shown to be slightly more positive in the sentiment of their emails when shown suggested replies (Wenker, 2023). SR systems are on the whole known to produce more positive sentiment messages than the human distribution (Kannan et al., 2016). We see this as an extension of the broader trend of LLMs to be overly obsequious. + +# B Implementation Details + +For constructing the training dataset, we use the following hyperparameters: SimSR is initialised from the distilbert-base-uncased checkpoint (Sanh et al., 2019). We set the search parameters to $N = 100$ and $M = 100$ . We use a redundancy penalty of 0.05 and a blending alpha of 0.75 for query augmentation. Both parameters provided a good trade-off between relevance and diversity in early testing, so we did not search hyperparameters further (see Section 5.2 for ablations). + +For training STAR, we initialise our model with the t5-small checkpoint. Note that this version of T5 has a comparable parameter count to the baselines which use DistilBERT (60M versus 66M). We tokenise the dataset with a maximum message length of 64 tokens. We train our model for up to 100k steps, with a warmup of 1k steps. In practice, the model typically converged around 20k steps. We use the AdamW optimiser (Loshchilov and Hutter, 2019) with an initial learning rate of 5e-4 and linear decay. We evaluate every 2k steps, by taking the ROUGE and Self-ROUGE scores on the validation set (this uses the ground-truth from the original dataset, not from the bootstrapped dataset), and employ early stopping once both metrics have ceased to improve. Note, we found these metrics were a much more reliable stopping point than crossentropy loss, which typically converged much earlier. + +# C Further Case Studies + +Table 5 displays further examples of STAR's predictions versus SimSR, taken from the PersonaChat Test set. + +
Persona:my favorite tv show is the office.
i like jazz music.
i do not drive because i live in new york.
i am jewish.
i do not eat anything sweet.
Message:i was but am now divorced
SimSR- i am sorry to hear that .
- oh . i am sorry to hear that .
- i am sorry to hear that . do you have any kids ?
STAR- oh i am sorry to hear that
- do you like music ?
- like music ? i love jazz
Persona:hey there my name is jordan and i am a veterinarian.
love to read drama books.
i love playing video games.
i am also a musician on the weekends.
i am originally from california but i live in florida.
Message:sometimes . i listen to a lot of music . do you read a lot ?
SimSR- hi how are you ?
- hi how are you today
- i do not like music at all
STAR- i listen to music do you
- no not really i like all music .
- yes i do my favorite is country
+ +Table 5: Example model outputs from the PersonaChat Test set, comparing STAR (ours) with the top-performing baseline method. STAR is able to capture a broader range of intents through its end-to-end autoregressive retrieval. \ No newline at end of file diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/images.zip b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ed850a60bcab049bac4f6f71cbadc0106dfb4ce8 --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b66b5288dd2fad127b62e53ea90d8c66cc887e12e937c126f22ebc0d4e253119 +size 414473 diff --git a/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/layout.json b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9eb9d99257e190b3cf2bc4b48e29790220ca6cc1 --- /dev/null +++ b/2023/End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems/layout.json @@ -0,0 +1,8334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 74, + 468, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 74, + 468, + 109 + ], + "spans": [ + { + "bbox": [ + 126, + 74, + 468, + 109 + ], + "type": "text", + "content": "End-to-End Autoregressive Retrieval via Bootstrapping for Smart Reply Systems" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 223, + 119, + 373, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 119, + 373, + 133 + ], + "spans": [ + { + "bbox": [ + 223, + 119, + 373, + 133 + ], + "type": "text", + "content": "Benjamin Towle1, Ke Zhou1,2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 234, + 133, + 362, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 133, + 362, + 148 + ], + "spans": [ + { + "bbox": [ + 234, + 133, + 362, + 148 + ], + "type": "text", + "content": "1University of Nottingham" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 255, + 148, + 340, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 148, + 340, + 161 + ], + "spans": [ + { + "bbox": [ + 255, + 148, + 340, + 161 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 255, + 148, + 340, + 161 + ], + "type": "text", + "content": "Nokia Bell Labs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 163, + 449, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 163, + 449, + 176 + ], + "spans": [ + { + "bbox": [ + 149, + 163, + 449, + 176 + ], + "type": "text", + "content": "{benjamin.towle, ke.zhou}@nottingham.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "spans": [ + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "text", + "content": "Reply suggestion systems represent a staple component of many instant messaging and email systems. However, the requirement to produce sets of replies, rather than individual replies, makes the task poorly suited for out-of-the-box retrieval architectures, which only consider individual message-reply similarity. As a result, these system often rely on additional post-processing modules to diversify the outputs. However, these approaches are ultimately bottlenecked by the performance of the initial retriever, which in practice struggles to present a sufficiently diverse range of options to the downstream diversification module, leading to the suggestions being less relevant to the user. In this paper, we consider a novel approach that radically simplifies this pipeline through an autoregressive text-to-text retrieval model, that learns the smart reply task end-to-end from a dataset of (message, reply set) pairs obtained via bootstrapping. Empirical results show this method consistently outperforms a range of state-of-the-art baselines across three datasets, corresponding to a " + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "inline_equation", + "content": "5.1\\% - 17.9\\%" + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "text", + "content": " improvement in relevance, and a " + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "inline_equation", + "content": "0.5\\% - 63.1\\%" + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "text", + "content": " improvement in diversity compared to the best baseline approach. We make our code publicly available." + }, + { + "bbox": [ + 86, + 237, + 274, + 570 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 582, + 154, + 595 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 582, + 154, + 595 + ], + "spans": [ + { + "bbox": [ + 68, + 582, + 154, + 595 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 604, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 604, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 604, + 291, + 754 + ], + "type": "text", + "content": "Reply suggestion, or smart reply (SR), systems are a staple component of many commercial applications such as Gmail, Skype, Outlook, Microsoft Teams, LinkedIn and Facebook Messenger. They help the user process chats and emails quicker by offering a set of canned replies which can be clicked without requiring manual typing. However, dialogue is known to be a one-to-many problem (Zhao et al., 2017; Towle and Zhou, 2022) – namely, for any given message, there are multiple possible replies. To reflect this uncertainty, systems should" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 213, + 526, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 526, + 294 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 526, + 294 + ], + "type": "text", + "content": "present a diverse set of options to the user. For instance, given the message How are you?, an SR system could suggest: {I'm good; Ok; Not great}. Resultantly, the quality of a given reply depends not only on the message, but on the other replies in the reply set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 296, + 526, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 296, + 526, + 471 + ], + "spans": [ + { + "bbox": [ + 302, + 296, + 526, + 471 + ], + "type": "text", + "content": "Several prior works explore solutions to this problem such as removing near duplicates, penalising inter-reply similarity (Deb et al., 2019), clustering by intent (Henderson et al., 2017; Weng et al., 2019), learning latent variables (Deb et al., 2019, 2021), or model-based simulation (Towle and Zhou, 2023). However, these methods share a common design choice (Figure 1A): (1) a retrieval-based Matching model, which has learned a shared embedding space between messages and replies, returns a shortlist of top scoring replies; (2) this shortlist is refined through some diversification procedure to obtain the final reply set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 474, + 527, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 474, + 527, + 595 + ], + "spans": [ + { + "bbox": [ + 302, + 474, + 527, + 595 + ], + "type": "text", + "content": "Unfortunately, this assumes that the initial shortlist contains at least one good reply set. In practice, we find Matching models often search myopically, only retrieving candidates that are very similar to one another (Figure 1A). Thus, the chosen reply set often fails to reflect a diverse range of user intents, while latency constraints make more sophisticated diversification techniques or larger shortlists prohibitive (Deb et al., 2019)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 301, + 597, + 526, + 772 + ], + "type": "text", + "content": "An intuitive, but – to the best of our knowledge – unexplored, solution to this problem is to conduct the retrieval autoregressively, with each reply conditioned on both the initial message and the previous replies in the set. Unfortunately, this approach encounters a second problem, namely, the lack of any datasets containing (message, reply set) pairs (Towle and Zhou, 2023). In practice, SR systems are trained on individual (message, reply) pairs obtained from conversation datasets, while the task of presenting multiple diverse replies to the user is outsourced to a separate diversification module." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 760, + 285, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 760, + 285, + 772 + ], + "spans": [ + { + "bbox": [ + 80, + 760, + 285, + 772 + ], + "type": "text", + "content": "'https://github.com/ BenjaminTowle/STAR" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 310, + 791 + ], + "type": "text", + "content": "7610" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 805 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7610-7622" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 165, + 805, + 428, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 805, + 428, + 817 + ], + "spans": [ + { + "bbox": [ + 165, + 805, + 428, + 817 + ], + "type": "text", + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 524, + 157 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 524, + 157 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 524, + 157 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 524, + 157 + ], + "type": "image", + "image_path": "e50b983ca8bdf7fb68ad09f880c265e3c03c12402818f04d12ec540b52ff6509.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 161, + 505, + 269 + ], + "blocks": [ + { + "bbox": [ + 73, + 161, + 505, + 269 + ], + "lines": [ + { + "bbox": [ + 73, + 161, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 73, + 161, + 505, + 269 + ], + "type": "image", + "image_path": "eff81584e09698eba5df1bd7394244c42071a8e9d49cc8fef307cc4b79ef12e4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 280, + 525, + 317 + ], + "lines": [ + { + "bbox": [ + 67, + 280, + 525, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 280, + 525, + 317 + ], + "type": "text", + "content": "Figure 1: Previous methods [A] compared to our approach, STAR [B]. The example displayed is taken from the DailyDialog Test set, and compares the predictions of STAR with SimSR (Towle and Zhou, 2023), the next best method. Our method's suggestions present a diverse range of topics/intents to drive the conversation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 337, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 337, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 337, + 291, + 772 + ], + "type": "text", + "content": "To meet this dual need, we present both (i) a bootstrapping method for creating a high-quality dataset of (message, reply sets) and (ii) a novel autoregressive retrieval model which predicts sequences of replies. For solving (i), we observe how model-based planning algorithms have been known to serve as a powerful policy improvement operator (Silver et al., 2017; Schrittwieser et al., 2019), including in several NLP systems (Jang et al., 2020, 2021). Specifically, the outputs of a model-based planning algorithm can be used to bootstrap a SR system. Further, by conducting this planning offline we are able to leverage two key advantages: (1) the system is free of the latency constraints of online inference, and therefore can increase the search space coverage of the planning algorithm; (2) the system can leverage information that would not be available during inference, such as the ground-truth reply, to further guide the search process. For (ii) we unify both steps of the standard SR pipeline into a single end-to-end model, which mitigates the myopic search, and allows the model to learn to diversify its predictions in a principled way through gradient-based learning. To this end, we present STAR (Suggested replies with T5 and Autoregressive Retrieval) (Figure 1B). At a high level, STAR is a text-to-text model trained to output sequences of replies, where each reply is conditioned both on the initial message and the previous replies in the sequence. Concretely, we instantiate our method with the T5 pretrained model (Raffel et al., 2020). We expand T5's vocabulary by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 337, + 526, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 337, + 526, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 337, + 526, + 444 + ], + "type": "text", + "content": "treating each reply in the candidate pool as a novel token, and demonstrate a simple-yet-effective technique for initialising the new token embeddings, which leverages the model's existing semantic priors. Notably, by treating each reply as a token, we limit the number of autoregressive decoding steps required, keeping the model's efficiency comparable to other retrieval-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "spans": [ + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "text", + "content": "Empirically, we evaluate our approach on three benchmarks: Reddit (Zhang et al., 2021), which is the only publicly-available SR benchmark, as well as PersonaChat (Zhang et al., 2018) and DailyDialog (Li et al., 2017) which are both widely-used in dialogue research more broadly (Zhang et al., 2019; Roller et al., 2020, inter alia), and share a similar conversational style with SR apps. We demonstrate superior performance over state-of-the-art baselines across all datasets, corresponding to a " + }, + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "inline_equation", + "content": "5.1\\% - 17.9\\%" + }, + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "text", + "content": " improvement in relevance, and a " + }, + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "inline_equation", + "content": "0.5\\% - 63.1\\%" + }, + { + "bbox": [ + 302, + 454, + 525, + 670 + ], + "type": "text", + "content": " improvement in diversity compared to the best baseline approach. We further show comparable efficiency to previous methods, and perform a range of ablations to motivate our design choices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 678, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 526, + 773 + ], + "type": "text", + "content": "In summary, our key contributions are as follows: (1) an autoregressive retrieval architecture for sequentially predicting suggested replies; (2) a bootstrapping framework for generating high-quality data of (message, reply set) pairs; (3) detailed analysis of model behaviour and performance including a case study and ablation of key components." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "7611" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 161, + 83 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 97, + 292, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 97, + 292, + 448 + ], + "spans": [ + { + "bbox": [ + 69, + 97, + 292, + 448 + ], + "type": "text", + "content": "Smart reply The proprietary nature of data from email and chat applications has led several previous works to use publicly-available dialogue datasets (Zhang et al., 2021; Deb et al., 2021; Towle and Zhou, 2023) to benchmark SR methods, due to their analogous conversational nature. While early SR systems used generative models (Kannan et al., 2016), current production systems favour retrieval methods due to their greater controllability of outputs and superior latency (Deb et al., 2019). Increasing the diversity of reply suggestions is a key focus of previous work, which has been attempted by: (1) mapping replies to discrete intents / topics (Kannan et al., 2016; Chakravarthi and Pasternack, 2017; Weng et al., 2019); (2) re-weighting replies according to their similarity with other replies in the set (Carbonell and Goldstein-Stewart, 1998; Deb et al., 2019); (3) learning continuous latent variables to generate multiple queries (Zhao et al., 2017; Deb et al., 2019); (4) using model-based simulation to iteratively search and evaluate the relevance of candidate reply sets (Towle and Zhou, 2023). Our proposed method differs from all of these approaches in that our model learns to account for the interdependencies between replies through end-to-end backpropagation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 462, + 291, + 772 + ], + "type": "text", + "content": "Autoregressive retrieval Integrating neural retrieval into the well-established paradigm of text-to-text models is of growing interest. Earlier work focuses on outputting a document ID given a query (Tay et al., 2022). Further work has extended this by considering alternate ways of representing the document IDs, such as through unique substrings (Bevilacqua et al., 2022). Another line of work has used autoregressive retrieval for the entity linking task (Cao et al., 2021a,b,c). There, the motivation is to reduce the large number of entities by relying on the text-to-text model's pre-existing vocabulary, rather than having to retrieve embeddings from a memory-intensive dense index. Our proposed method differs considerably from these previous works both in instantiation and motivation. Instantiation-wise, we generate multiple replies - critical to making this possible is the novel bootstrapping technique for creating the dataset of (message, reply set) pairs to train on. Motivation-wise, our goal is to be able to condition each reply on both the input message and previous replies in the set, enabling the model to learn to predict se" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 303, + 71, + 489, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 489, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 489, + 84 + ], + "type": "text", + "content": "quences of replies in a differentiable way." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 92, + 527, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 92, + 527, + 417 + ], + "spans": [ + { + "bbox": [ + 302, + 92, + 527, + 417 + ], + "type": "text", + "content": "Bootstrapping The idea of bootstrapping training data from limited resources has received significant recent interest in NLP, given the newly demonstrated few / zero-shot capabilities of many large language models (Brown et al., 2020). It has seen usage in few-shot shot text-classification (Schick and Schütze, 2021a), semantic similarity (Schick and Schütze, 2021b), tool-usage (Schick et al., 2023), retrieval (Izacard and Grave, 2021), sequence generation (He et al., 2020), and instruction-tuning (Honovich et al., 2023; Wang et al., 2023; Taori et al., 2023), amongst others. These techniques can also be seen as a form of knowledge distillation (Hinton et al., 2015), except that the training typically involves predicting the exact token targets, rather than using the soft probabilities of a teacher model. Although sometimes these techniques are used as an addition to supervised learning (He et al., 2020), in our case there are no datasets containing the ideal reply sets to suggest to the user. Instead, we must bootstrap this in a more unsupervised way, by transforming a dataset of (message, reply) pairs into a dataset of (message, reply set) pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 303, + 427, + 392, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 427, + 392, + 441 + ], + "spans": [ + { + "bbox": [ + 303, + 427, + 392, + 441 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 449, + 525, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 449, + 525, + 516 + ], + "spans": [ + { + "bbox": [ + 302, + 449, + 525, + 516 + ], + "type": "text", + "content": "In this section, we first describe the model-based planning process used to obtain the bootstrapped dataset of (message, reply set) pairs (Section 3.1). Then, we show how the STAR architecture can be trained on this dataset (Section 3.2)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 526, + 445, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 526, + 445, + 538 + ], + "spans": [ + { + "bbox": [ + 303, + 526, + 445, + 538 + ], + "type": "text", + "content": "3.1 Offline Dataset Creation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "spans": [ + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "content": "Our goal is to transform a dialogue dataset " + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(x,y)\\}" + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "content": " of (message, reply) tuples, into a dataset " + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^* = \\{(x,Y)\\}" + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "content": " is the set of replies " + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "inline_equation", + "content": "\\{y_k\\}^K" + }, + { + "bbox": [ + 302, + 543, + 526, + 731 + ], + "type": "text", + "content": " to be presented to the user. Algorithm 1 summarises this process. While our method is general to any arbitrary planning algorithm, we choose to instantiate our approach with a modified version of SimSR (Towle and Zhou, 2023), a recently released publicly available state-of-the-art SR method, that employs model-based simulation to predict reply sets. As the original algorithm was designed for online inference, we make several changes to benefit the offline nature of our version, and detail the full implementation below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": "The initial retrieval is conducted by a Matching model " + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": " that separately encodes messages and replies into a shared latent space. Given an encoded" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "7612" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "content": "Algorithm 1 Offline Dataset Creation. We use " + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "inline_equation", + "content": "M = 100" + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "inline_equation", + "content": "\\alpha = 0.75" + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 67, + 70, + 289, + 110 + ], + "type": "text", + "content": " as our default setting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "spans": [ + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": "Input Matching model " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", message " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", precomputed reply vectors " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y_r}\\}^R" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", number of candidates " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", number of simulations " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", final reply set size " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", query augmentation coefficient " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": ", redundancy penalty " + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 67, + 111, + 290, + 153 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 153, + 161, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 153, + 161, + 163 + ], + "spans": [ + { + "bbox": [ + 82, + 153, + 161, + 163 + ], + "type": "text", + "content": "Output reply set " + }, + { + "bbox": [ + 82, + 153, + 161, + 163 + ], + "type": "inline_equation", + "content": "Y_{K}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 164, + 289, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 164, + 289, + 184 + ], + "spans": [ + { + "bbox": [ + 78, + 164, + 289, + 184 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {x}, \\mathbf {y} \\leftarrow \\Phi (x), \\Phi (y) \\\\ \\tilde {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}", + "image_path": "bfbac6451aea77bb93e038b1580d4f6481b53ee5b50741bbc19a08a4a70c198e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 176, + 289, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 176, + 289, + 199 + ], + "spans": [ + { + "bbox": [ + 78, + 176, + 289, + 199 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {\\mathbf {x}} \\leftarrow \\alpha \\mathbf {x} + (1 - \\alpha) \\mathbf {y} \\\\ Y _ {N} \\leftarrow N \\underset {r} {\\operatorname {a r g m a x}} (\\bar {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}}) \\end{array} \\quad \\triangleright \\text {q u e r y a u g m e n t a t i o n}", + "image_path": "b904e197fc0a7d194459fe19fc893934d1c8ea4abf1c1d181b32bac18c00ffc0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 200, + 176, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 200, + 176, + 214 + ], + "spans": [ + { + "bbox": [ + 78, + 200, + 176, + 214 + ], + "type": "interline_equation", + "content": "Y _ {M} \\leftarrow M \\operatorname {a r g m a x} _ {\\mathbf {r}} (\\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {r}})", + "image_path": "5ea8ef5fbef1aee5fae9c963666f6a08858a6a6750cc4e8d4d23a60a38a21af5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 215, + 289, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 215, + 289, + 235 + ], + "spans": [ + { + "bbox": [ + 78, + 215, + 289, + 235 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q (y _ {m} | x) \\propto \\exp \\tilde {\\mathbf {x}} \\cdot \\mathbf {y} _ {\\mathbf {m}} \\\\ Y _ {G} \\leftarrow \\emptyset \\end{array} \\quad \\triangleright \\text {s o f t m a x o v e r t o p - M s c o r e s}", + "image_path": "3a97f270fbace3b987bf22b2c9844244ebc3bfa527d8c7cfdecbe58542b1c0df.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 236, + 285, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 236, + 285, + 266 + ], + "spans": [ + { + "bbox": [ + 78, + 236, + 285, + 266 + ], + "type": "interline_equation", + "content": "\\begin{array}{c}\\textbf{for} k\\gets 0\\text{to} K\\textbf{do}\\\\ y_{k}\\leftarrow \\operatorname *{argmax}_{n}\\sum_{m}^{M}f(Y_{G}^{n},y_{m})q(y_{m}|x) - \\lambda f(Y_{G},y_{n}) \\end{array}", + "image_path": "ba9201fa265a4a4a657ddbab34a825ec451c84c97e1a26f0033d1b5849dc2703.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 267, + 149, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 267, + 149, + 277 + ], + "spans": [ + { + "bbox": [ + 78, + 267, + 149, + 277 + ], + "type": "interline_equation", + "content": "Y _ {G} \\leftarrow Y _ {G} \\cup y _ {k}", + "image_path": "0eb53edcd11f3063acbab52f567d7104181f30bc459410969eaa819eba2650be.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 277, + 107, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 277, + 107, + 285 + ], + "spans": [ + { + "bbox": [ + 78, + 277, + 107, + 285 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 285, + 117, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 285, + 117, + 296 + ], + "spans": [ + { + "bbox": [ + 79, + 285, + 117, + 296 + ], + "type": "interline_equation", + "content": "Y _ {K} \\leftarrow Y _ {G}", + "image_path": "4f0ba53ede59a84a7075ffb2d38401ac39ac03c4060a1cf8d706c3efce175279.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 79, + 296, + 118, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 296, + 118, + 306 + ], + "spans": [ + { + "bbox": [ + 79, + 296, + 118, + 306 + ], + "type": "interline_equation", + "content": "\\text {r e t u r n} Y _ {K}", + "image_path": "3ae113e2f416d083062618d60dcd43966f248cd268d20608937579ed38ded12b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "content": "message " + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\Phi (x)" + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "content": ", it retrieves the top " + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "content": " candidates from a pool of pre-computed reply vectors " + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathbf{R}} = \\{\\mathbf{y}_{\\mathbf{r}}\\}^{R}" + }, + { + "bbox": [ + 67, + 327, + 290, + 407 + ], + "type": "text", + "content": " by combining their dot product similarity with a pre-computed language-model bias - a standard component of SR systems to downweight overly specific replies (Deb et al., 2019)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 413, + 290, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 413, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 107, + 413, + 290, + 429 + ], + "type": "interline_equation", + "content": "Y _ {N} = N \\underset {r} {\\operatorname {a r g m a x}} \\left(\\mathbf {x} \\cdot \\mathbf {y} _ {\\mathbf {r}} + \\beta \\mathrm {L M} \\left(y _ {r}\\right)\\right) \\tag {1}", + "image_path": "8c43b24482389293c0a94febe7d465f81c9654f0b2dc06c799e28383a62bd2da.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "spans": [ + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "text", + "content": "We then output the " + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "text", + "content": "-tuple " + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "inline_equation", + "content": "Y_{i} \\in \\binom{Y_{N}}{K}" + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "text", + "content": " that has the highest expected similarity with the human reply, according to some similarity function " + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "inline_equation", + "content": "f(\\cdot, \\cdot)" + }, + { + "bbox": [ + 67, + 439, + 290, + 481 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 124, + 486, + 290, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 486, + 290, + 506 + ], + "spans": [ + { + "bbox": [ + 124, + 486, + 290, + 506 + ], + "type": "interline_equation", + "content": "\\underset {i} {\\operatorname {a r g m a x}} \\mathbb {E} _ {y \\sim p (\\cdot | x)} \\left[ f (Y _ {i}, y) \\right] \\tag {2}", + "image_path": "597b36a560c90cd646b87a24d530648e8212c876c9115d93337af068bbd943ef.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 513, + 290, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 513, + 290, + 584 + ], + "spans": [ + { + "bbox": [ + 67, + 513, + 290, + 584 + ], + "type": "text", + "content": "Given the objective in SR is for at least one of the replies to be relevant, the similarity function is defined as a maximum over the sampled reply and each of the replies in the reply set, using term-level F1-score: " + }, + { + "bbox": [ + 67, + 513, + 290, + 584 + ], + "type": "inline_equation", + "content": "\\max_k\\mathrm{F1}(y_k,y)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "content": "We assume " + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "content": " is sampled from the ground-truth human distribution " + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "inline_equation", + "content": "p(\\cdot |x)" + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "content": ". As we do not have access to the true human distribution in practice, we instead use the same Matching model " + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "content": " as a proxy for this, given it is trained on (message, reply) pairs. We then approximate the expectation by marginalising over the top-" + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 67, + 584, + 290, + 677 + ], + "type": "text", + "content": " most likely replies:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 684, + 290, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 684, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 114, + 684, + 290, + 714 + ], + "type": "interline_equation", + "content": "\\approx \\underset {i} {\\operatorname {a r g m a x}} \\sum_ {m} ^ {M} f \\left(Y _ {i}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {3}", + "image_path": "395a142a1df3e2d283f2d7c514bc73a82b25e651729e023fa37ac13377f66b1d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "content": "In practice, it is intractable to evaluate every possible reply tuple, due to their combinatorial scaling. We therefore approximate this by greedily constructing the reply set one reply at a time. Formally," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": "let " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "Y_{G}" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": " be the set of currently selected replies, such that initially " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "Y_{G} = \\emptyset" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": ". Then, for each of " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "y_{n} \\in Y_{N}" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": ", we compute the expected similarity for the union of " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "Y_{G}" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "y_{n}" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": ", termed " + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "inline_equation", + "content": "Y_{G}^{n} = Y_{G} \\cup y_{n}" + }, + { + "bbox": [ + 302, + 71, + 525, + 126 + ], + "type": "text", + "content": " for brevity:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 368, + 135, + 525, + 164 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 135, + 525, + 164 + ], + "spans": [ + { + "bbox": [ + 368, + 135, + 525, + 164 + ], + "type": "interline_equation", + "content": "\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) \\tag {4}", + "image_path": "53d6d66289a18e428ad1f72243ca705fc6ad47d5d68898f03cc6829deead2bf3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "spans": [ + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "text", + "content": "We repeat this process for " + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "text", + "content": " timesteps, each time appending the highest scoring reply to " + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "inline_equation", + "content": "Y_{G}" + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "text", + "content": ", i.e. until " + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "inline_equation", + "content": "|Y_{G}| = K" + }, + { + "bbox": [ + 302, + 169, + 525, + 262 + ], + "type": "text", + "content": ". Note that this greedy search process implicitly canonicalises the order of the replies, as selecting replies in this way causes them to be roughly ordered by individual message-reply relevance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 272, + 399, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 272, + 399, + 285 + ], + "spans": [ + { + "bbox": [ + 302, + 272, + 399, + 285 + ], + "type": "text", + "content": "3.1.1 Adjustments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "spans": [ + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": "Scaling " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " The original SimSR algorithm was used only in an online setting (Towle and Zhou, 2023). Therefore, the size of the search parameters " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " (number of replies in the shortlist) and " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " (number of simulated user replies) is kept low (15 and 25 respectively in the original paper). As we only need to run this model offline however to obtain the dataset, we find setting " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": " to much larger values improves relevance (we use 100 for both), enabling both a broader search (i.e. by increasing " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": ") and a more accurate similarity function (i.e. by increasing " + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 302, + 287, + 525, + 449 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 302, + 457, + 525, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 457, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 302, + 457, + 525, + 578 + ], + "type": "text", + "content": "Redundancy penalty Early testing showed that scaling the search parameters reduced diversity. We therefore introduce a redundancy penalty, which penalises the model for selecting replies that are similar to replies already in the set " + }, + { + "bbox": [ + 302, + 457, + 525, + 578 + ], + "type": "inline_equation", + "content": "Y_{G}" + }, + { + "bbox": [ + 302, + 457, + 525, + 578 + ], + "type": "text", + "content": ". This is analogous to the inter-document similarity penalty used in the maximum marginal relevance IR (information retrieval) technique (Carbonell and Goldstein-Stewart, 1998)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 341, + 589, + 525, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 589, + 525, + 618 + ], + "spans": [ + { + "bbox": [ + 341, + 589, + 525, + 618 + ], + "type": "interline_equation", + "content": "\\sum_ {m} ^ {M} f \\left(Y _ {G} ^ {n}, y _ {m}\\right) q \\left(y _ {m} \\mid x\\right) - \\lambda f \\left(Y _ {G}, y _ {n}\\right) \\tag {5}", + "image_path": "acfab98ceb12439c8058d49940a258dce1de854974f91e48e0237f263b5e6627.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "content": "Query augmentation Unlike during online inference, we also have access to the ground-truth reply " + }, + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 302, + 624, + 525, + 772 + ], + "type": "text", + "content": " when constructing the dataset. Previous work has found that models obtain greater representational capabilities when given access to posterior information (Paranjape et al., 2022; Towle and Zhou, 2022). We therefore use an augmented query to retrieve with the Matching model. This is obtained by interpolating between the message and ground-truth reply embeddings. This biases the model's predictions towards the observed ground-truth in" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "7613" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 97 + ], + "type": "text", + "content": "the dataset, while still allowing it to benefit from its own learned distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 108, + 290, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 108, + 290, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 108, + 290, + 121 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} = \\alpha \\Phi (x) + (1 - \\alpha) \\Phi (y) \\tag {6}", + "image_path": "a4eb4483177f11158ed6fb89e04f766f7585196618ad3cb9ee4821f7c004a48f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 134, + 203, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 134, + 203, + 148 + ], + "spans": [ + { + "bbox": [ + 68, + 134, + 203, + 148 + ], + "type": "text", + "content": "3.2 Proposed STAR Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": "We initialise STAR with a T5-based text-to-text language model, which has previously been shown to be effective in autoregressive retrieval (Tay et al., 2022). While some autoregressive retrieval approaches identify their documents/replies through unique substrings (Bevilacqua et al., 2022) or constrained beam search (Cao et al., 2021b), we focus on approaches requiring only a limited number of autoregressive steps, to maintain competitive inference speeds to existing retrieval methods (Section 5.3). There are several alternatives for this such as treating each reply set as a unique token, or separately training on each (message, reply pair), but ultimately we opted for autoregressively treating each reply as a unique token in the vocabulary in order to exploit the compositionality of reply sets (Section 5.2 for performance comparison). Note that as the types of replies used in smart reply are usually quite short and concise, e.g. 'how are you', 'I'm fine thanks', 'yes, that's right' etc., systems in deployment only need to retrieve from a pool of " + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": " or so replies (Deb et al., 2019), in order to provide good coverage of possible user intents. As a result, we are able to keep the size of the vocabulary reasonable. Thus, our new vocabulary is defined as: " + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "inline_equation", + "content": "W_{tokens} \\cup W_{replies}" + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": ". An obvious challenge to this approach is that by treating each reply as a previously unseen word, it removes any semantic priors the model might have about their meaning. To mitigate this, we employ a bag-of-words initialisation strategy. Hence, we define the embedding of the " + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": "-th reply " + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "inline_equation", + "content": "E(y_{t})" + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": " as the average over the embeddings of the individual words within " + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "inline_equation", + "content": "w_{n} \\in y_{t}" + }, + { + "bbox": [ + 67, + 153, + 291, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 611, + 290, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 611, + 290, + 641 + ], + "spans": [ + { + "bbox": [ + 133, + 611, + 290, + 641 + ], + "type": "interline_equation", + "content": "E \\left(y _ {t}\\right) = \\frac {1}{N} \\sum_ {n} ^ {N} E \\left(w _ {n}\\right) \\tag {7}", + "image_path": "48c81e829bbc48d0719262f201d48b3a90854ace94e4308eeb39fa0feaf6dfbd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 291, + 773 + ], + "type": "text", + "content": "Intuitively, this ensures that the initial embeddings are close to the word embeddings of the original vocabulary, while also capturing some of the underlying semantics of the reply. We allow the weights to update during fine-tuning. Note that for T5 the output and input embedding layers share weights, and therefore this approach is used to initialise both layers. We train the model using cross-entropy loss to predict the next reply given the current sequence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 71, + 413, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 413, + 84 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 413, + 84 + ], + "type": "text", + "content": "of replies and messages:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 337, + 98, + 525, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 98, + 525, + 128 + ], + "spans": [ + { + "bbox": [ + 337, + 98, + 525, + 128 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N L L} = - \\sum_ {k} ^ {K} \\log p \\left(y _ {k} \\mid x, y _ {0}, \\dots , y _ {k - 1}\\right) \\tag {8}", + "image_path": "16364b36d33dc53895df951c91da919fc5064cdb36e2050062d21de3dadf099f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 142, + 427, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 142, + 427, + 157 + ], + "spans": [ + { + "bbox": [ + 302, + 142, + 427, + 157 + ], + "type": "text", + "content": "4 Experimental Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 167, + 374, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 167, + 374, + 179 + ], + "spans": [ + { + "bbox": [ + 302, + 167, + 374, + 179 + ], + "type": "text", + "content": "4.1 Baselines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 186, + 526, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 186, + 526, + 321 + ], + "spans": [ + { + "bbox": [ + 302, + 186, + 526, + 321 + ], + "type": "text", + "content": "Previous work has largely been closed-source and is therefore unavailable for direct comparison (Henderson et al., 2017; Weng et al., 2019; Deb et al., 2019). With the exception of SimSR, which has publicly available code " + }, + { + "bbox": [ + 302, + 186, + 526, + 321 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 302, + 186, + 526, + 321 + ], + "type": "text", + "content": ", we re-implement a variety of methods that cover the broad range of previous techniques. Due to its comparable size, all baselines apart from Seq2Seq are initialised with DistilBERT as the encoder backbone. These are summarised as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 334, + 526, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 334, + 526, + 469 + ], + "spans": [ + { + "bbox": [ + 302, + 334, + 526, + 469 + ], + "type": "text", + "content": "Seq2Seq is a generative encoder-decoder. While current production systems and the majority of related works use only retrieval models (Deb et al., 2019; Towle and Zhou, 2023), at least one related work includes a standard generative transformer as a baseline (Zhang et al., 2021), which we follow here. For maximum comparability with our method, we use the same t5-small model as a backbone. For each message, we sample " + }, + { + "bbox": [ + 302, + 334, + 526, + 469 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 334, + 526, + 469 + ], + "type": "text", + "content": " responses independently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 481, + 525, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 481, + 525, + 550 + ], + "spans": [ + { + "bbox": [ + 302, + 481, + 525, + 550 + ], + "type": "text", + "content": "Matching represents the out-of-the-box encoder with no additional diversification strategy and was used as a baseline method by Zhang et al. (2021). It simply selects the top " + }, + { + "bbox": [ + 302, + 481, + 525, + 550 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 481, + 525, + 550 + ], + "type": "text", + "content": " responses according to individual message-reply scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 560, + 525, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 560, + 525, + 655 + ], + "spans": [ + { + "bbox": [ + 302, + 560, + 525, + 655 + ], + "type": "text", + "content": "Matching-Topic uses an out-of-the-box topic classifier to ensure no two replies share the same topic, similar to previous work (Henderson et al., 2017; Weng et al., 2019). The classifier is trained on Twitter (Antypas et al., 2022), due to their comparable short-form open-domain chat conversations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 667, + 526, + 749 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 667, + 526, + 749 + ], + "spans": [ + { + "bbox": [ + 302, + 667, + 526, + 749 + ], + "type": "text", + "content": "Maximum Marginal Relevance (MMR) (Carbonell and Goldstein-Stewart, 1998) is originally an IR technique, used in several previous SR works (Deb et al., 2019; Towle and Zhou, 2023), which re-weights reply scores as a linear combination of their message-reply and inter-reply similarity." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 315, + 760, + 524, + 772 + ], + "type": "text", + "content": "https://github.com/BenjaminTowle/SimSR" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "7614" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 68, + 272, + 124 + ], + "blocks": [ + { + "bbox": [ + 87, + 68, + 272, + 124 + ], + "lines": [ + { + "bbox": [ + 87, + 68, + 272, + 124 + ], + "spans": [ + { + "bbox": [ + 87, + 68, + 272, + 124 + ], + "type": "table", + "html": "
TrainValidTest|YR|
Reddit50k5k5k48k
PersonaChat66k8k8k64k
DailyDialog76k7k7k62k
", + "image_path": "abceb955e66f2b977663f9a1bcdae84cb69b592dd458973a3ab5eb126d5b77dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 131, + 291, + 179 + ], + "lines": [ + { + "bbox": [ + 67, + 131, + 291, + 179 + ], + "spans": [ + { + "bbox": [ + 67, + 131, + 291, + 179 + ], + "type": "text", + "content": "Table 1: Number of samples in the Train, Validation, Test sets and Candidate pool in the three datasets for evaluation. The Candidate pool comprises the Train set with duplicate responses removed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 202, + 291, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 202, + 291, + 296 + ], + "spans": [ + { + "bbox": [ + 67, + 202, + 291, + 296 + ], + "type": "text", + "content": "MCVAE (Deb et al., 2019) is a conditional variational autoencoder (Zhao et al., 2017) which learns to generate multiple query vectors from a single message embedding, representing the multiple possible reply intents. Candidates are scored via a voting process, whereby the " + }, + { + "bbox": [ + 67, + 202, + 291, + 296 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 202, + 291, + 296 + ], + "type": "text", + "content": " most-selected replies are chosen." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 307, + 291, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 307, + 291, + 430 + ], + "spans": [ + { + "bbox": [ + 67, + 307, + 291, + 430 + ], + "type": "text", + "content": "SimSR (Towle and Zhou, 2023) uses an iterative search and evaluation process to select possible reply sets and score them according to their expected similarity from a learned world model, which serves as a proxy for the user. To ensure comparability of SimSR with our method and the other baselines, we include the language-model bias in the scoring process (Equation 1), and also deduplicate the candidate pool." + }, + { + "bbox": [ + 67, + 307, + 291, + 430 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 441, + 136, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 441, + 136, + 453 + ], + "spans": [ + { + "bbox": [ + 67, + 441, + 136, + 453 + ], + "type": "text", + "content": "4.2 Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 460, + 291, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 460, + 291, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 460, + 291, + 596 + ], + "type": "text", + "content": "We evaluate our proposed method across three datasets, summarised in Table 1. Below, we describe the datasets in more detail and motivate their inclusion. Note, other than Reddit, there are no publicly available SR datasets, due to their commercial nature (e.g. Henderson et al. (2017); Deb et al. (2019); Weng et al. (2019)). Therefore, we adopt several dialogue datasets, which is the closest alternative to conversations on proprietary chat applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 606, + 291, + 741 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 606, + 291, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 606, + 291, + 741 + ], + "type": "text", + "content": "Reddit (Zhang et al., 2021) was originally introduced for training multilingual SR systems, and is the only publicly available dataset specifically intended for SR purposes. As the original dataset is very large, we follow Towle and Zhou (2023) and use the reduced version of the dataset. Note, this version only contains English, as our aim is limited to the monolingual setting. Due to the organic nature of the dataset, conversations cover a very broad range of topics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 191 + ], + "type": "text", + "content": "PersonaChat (Zhang et al., 2018) is a crowdworker-sourced dataset comprising personagrounded conversations, in which each speaker is assigned a persona comprising a few short sentences. Following previous methods (Humeau et al., 2020), we concatenate the persona to the beginning of the message. The participants are instructed to chat naturally and to try to get to know one another." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 200, + 526, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 200, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 302, + 200, + 526, + 306 + ], + "type": "text", + "content": "DailyDialog (Li et al., 2017) is a dataset created from English language learning websites and consists of a variety of high-quality dialogues in everyday scenarios. The dataset differs from the former two in that the conversations often involve real-life scenarios, such as asking for directions, and therefore captures a different variety of conversational skills." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 317, + 367, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 317, + 367, + 328 + ], + "spans": [ + { + "bbox": [ + 302, + 317, + 367, + 328 + ], + "type": "text", + "content": "4.3 Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 334, + 525, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 334, + 525, + 388 + ], + "spans": [ + { + "bbox": [ + 302, + 334, + 525, + 388 + ], + "type": "text", + "content": "We evaluate our method on the same weighted ROUGE ensemble as previous methods (Lin, 2004; Deb et al., 2019, 2021), which is known to correlate well with click-through rate (Zhang et al., 2021):" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 345, + 396, + 525, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 396, + 525, + 418 + ], + "spans": [ + { + "bbox": [ + 345, + 396, + 525, + 418 + ], + "type": "interline_equation", + "content": "\\frac {\\text {R O U G E} - 1}{6} + \\frac {\\text {R O U G E} - 2}{3} + \\frac {\\text {R O U G E} - 3}{2} \\tag {9}", + "image_path": "f8ed275342e8fb5ab137d216f7062d2df71fe29e5b4488795f18e58839e2a764.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 420, + 526, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 420, + 526, + 568 + ], + "spans": [ + { + "bbox": [ + 302, + 420, + 526, + 568 + ], + "type": "text", + "content": "As the goal of SR systems it to ensure that at least one of the suggested replies is relevant to the user, we only record the maximum ROUGE score across each of the " + }, + { + "bbox": [ + 302, + 420, + 526, + 568 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 302, + 420, + 526, + 568 + ], + "type": "text", + "content": " suggested replies. We also evaluate the model on Self-ROUGE (Celikyilmaz et al., 2020): This is an unreferenced metric that measures the internal dissimilarity (i.e. diversity) within the reply set by treating one reply as the predicted reply and the other parts as the references. Note that a lower Self-ROUGE score indicates more diversity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 577, + 376, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 577, + 376, + 588 + ], + "spans": [ + { + "bbox": [ + 302, + 577, + 376, + 588 + ], + "type": "text", + "content": "4.4 Inference" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 594, + 526, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 594, + 526, + 715 + ], + "spans": [ + { + "bbox": [ + 302, + 594, + 526, + 715 + ], + "type": "text", + "content": "For inference, we use the entire training set as the candidate pool for each respective dataset, with dedduplication to remove exact matches. For STAR, we greedily decode the next reply token until " + }, + { + "bbox": [ + 302, + 594, + 526, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 594, + 526, + 715 + ], + "type": "text", + "content": " tokens have been decoded. Note, we only allow the model to output replies represented in the bootstrapped dataset, and also block non-replies, i.e. words from the original vocabulary, from being predicted." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 725, + 434, + 739 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 725, + 434, + 739 + ], + "spans": [ + { + "bbox": [ + 302, + 725, + 434, + 739 + ], + "type": "text", + "content": "5 Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": "We focus our efforts on answering the following Research Questions: " + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "inline_equation", + "content": "(\\mathbf{RQ}_1)" + }, + { + "bbox": [ + 302, + 746, + 525, + 772 + ], + "type": "text", + "content": " How does STAR com" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 750, + 290, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 290, + 773 + ], + "type": "text", + "content": "3Both changes lead to consistently improved accuracy and diversity across all datasets compared to the original paper." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "7615" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "content": "pare to existing state-of-the-art methods? (Section 5.1, 5.4); " + }, + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "inline_equation", + "content": "(\\mathbf{RQ}_2)" + }, + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "content": " Which components of the data collection algorithm and fine-tuning have the largest impact on STAR's performance? (Section 5.2); " + }, + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "inline_equation", + "content": "(\\mathbf{RQ}_3)" + }, + { + "bbox": [ + 67, + 71, + 290, + 152 + ], + "type": "text", + "content": " How efficient is STAR in inference? (Section 5.3)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 161, + 158, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 161, + 158, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 161, + 158, + 173 + ], + "type": "text", + "content": "5.1 Main Results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "spans": [ + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "text", + "content": "Table 2 compares the performance of different SR systems across the Reddit, PersonaChat and Dialy-Dialog datasets. In terms of relevance (ROUGE), STAR shows an especially large improvement in Reddit " + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "inline_equation", + "content": "(+17.9\\%)" + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "text", + "content": " and DailyDialog " + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "inline_equation", + "content": "(+15.8\\%)" + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "text", + "content": ". We hypothesise the gains in PersonaChat " + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "inline_equation", + "content": "(+5.1\\%)" + }, + { + "bbox": [ + 67, + 178, + 291, + 340 + ], + "type": "text", + "content": " are more modest because the replies are more easily predicted due to the persona, which is concatenated to each message. This significantly reduces the noise during the initial retrieval for the baselines, as they only need to retrieve the messages relevant to that particular persona." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "text", + "content": "For diversity (Self-ROUGE), the strongest gains were found in DailyDialog " + }, + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "inline_equation", + "content": "(+63.1\\%)" + }, + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "text", + "content": ". For PersonaChat, STAR performs much better than the retrieval methods, only falling behind Seq2Seq, due to its altogether noisier outputs as evidenced by having the worst relevance score. The Reddit results were comparatively more modest " + }, + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "inline_equation", + "content": "(+0.5\\%)" + }, + { + "bbox": [ + 66, + 341, + 291, + 530 + ], + "type": "text", + "content": " - we hypothesise this is because the dataset is altogether more noisy, and so there are relatively few similar replies in the dataset, as shown by the Self-ROUGE scores being lower than the other two datasets. Overall, the consistent outperformance in both relevance and diversity metrics supports the benefits of the STAR approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 539, + 138, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 539, + 138, + 551 + ], + "spans": [ + { + "bbox": [ + 67, + 539, + 138, + 551 + ], + "type": "text", + "content": "5.2 Ablation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "content": "In Table 3, we conduct ablations across two key axes: data collection and STAR training. The data collection ablations serve to investigate the benefits of the novel changes to the SimSR algorithm from Section 3.1.1. The STAR training ablations investigates the degree to which the improvements in performance are caused by the bootstrapped dataset or by STAR's architecture itself; we achieve this by considering several alternative variants of STAR. Our data collection ablations consider two features: (A) removing the query augmentation prevents the model from leveraging any ground truth information during prediction; (B) removing the redundancy penalty no longer explicitly penalises lack of diversity in predicted reply sets. For STAR training, we consider three alternative configurations:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": "(C) we replace the bag-of-words embeddings with randomly initialised embeddings - this removes any priors about the meaning of replies and forces the model to learn them tabula rasa; (D) we treat each reply set as a unique token - this removes the compositional element from the task, constraining the model to only predicting previously seen reply sets, therefore testing whether the model is capable of learning to compose novel reply sets; (E) we remove the ability to account for interdependencies between replies, by restructuring each (message, reply set) data point into " + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": " data points of (message, reply " + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": "), and then outputting the top-" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 302, + 71, + 526, + 301 + ], + "type": "text", + "content": " replies during inference - this investigates whether the benefit lies simply in the bootstrapped dataset being better suited to the SR task, rather than in STAR's ability to account for interdependencies between replies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 321, + 526, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 321, + 526, + 591 + ], + "spans": [ + { + "bbox": [ + 302, + 321, + 526, + 591 + ], + "type": "text", + "content": "In terms of data collection ablations, we found removing the redundancy penalty significantly reduced the diversity of predictions, although in some cases offered slightly improved relevance; removing the query augmentation generally led to a worse relevance/diversity trade-off. For the variants of STAR training, we found that random embeddings consistently reduced relevance, while also led to less diverse predictions; reply sets as tokens led to the most competitive variant of STAR compared to our default setup: diversity was overall better, due to using preconstructed reply sets from the offline planning algorithm, but this came at the trade-off of reduced flexibility from being unable to construct novel reply sets when the context required it – resultantly, we saw a corresponding reduction in relevance. Finally, predicting replies separately expectedly harmed both relevance and diversity, demonstrating the importance of accounting for reply interdependencies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 611, + 526, + 772 + ], + "type": "text", + "content": "In Figure 2, we further validated the individual results of our ablation by aggregating the results across datasets (applying an equal weighting to each dataset). This demonstrates the overall trend that the default STAR offers the superior trade-off between relevance and diversity, while treating reply sets as tokens offered the next best alternative. Nevertheless, we believe that keeping individual replies as tokens – thus allowing the model to construct reply sets dynamically – is likely to be an attractive property for deployed systems, enabling the overall vocabulary size to remain modest." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "7616" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 68, + 512, + 211 + ], + "blocks": [ + { + "bbox": [ + 81, + 68, + 512, + 211 + ], + "lines": [ + { + "bbox": [ + 81, + 68, + 512, + 211 + ], + "spans": [ + { + "bbox": [ + 81, + 68, + 512, + 211 + ], + "type": "table", + "html": "
MethodRedditPersonaChatDailyDialog
ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
Generative models
Seq2Seq2.413.436.836.88*4.013.91
Retrieval models
Matching1.959.427.5121.476.5316.65
M-Topic1.813.947.1615.436.1411.11
M-MMR2.204.447.8114.576.138.63
M-CVAE2.305.027.4312.216.7810.49
SimSR42.792.189.0410.526.824.80
STAR3.29*2.179.50*7.747.90*1.77*
", + "image_path": "fcc2b01350f85ab7e0093002d6745304619854fa9ddd950586b59954cfa75814.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 68, + 266, + 533, + 399 + ], + "blocks": [ + { + "bbox": [ + 67, + 219, + 526, + 257 + ], + "lines": [ + { + "bbox": [ + 67, + 219, + 526, + 257 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 526, + 257 + ], + "type": "text", + "content": "Table 2: Performance of STAR across Reddit, PersonaChat and DailyDialog Test sets on relevance (ROUGE) and diversity (Self-ROUGE) metrics. Bold indicates best result, underline indicates second-best. * = statistically significant versus next best result on t-test with " + }, + { + "bbox": [ + 67, + 219, + 526, + 257 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 219, + 526, + 257 + ], + "type": "text", + "content": "-value < 0.01." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 68, + 266, + 533, + 399 + ], + "lines": [ + { + "bbox": [ + 68, + 266, + 533, + 399 + ], + "spans": [ + { + "bbox": [ + 68, + 266, + 533, + 399 + ], + "type": "table", + "html": "
ModelRedditPersonaChatDailyDialog
ConfigurationROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓ROUGE ↑Self-ROUGE ↓
STAR3.35*2.278.857.488.391.81
Data Collection Ablations
A: No Query Augmentation2.942.008.996.947.242.89
B: No Redundancy Penalty3.064.299.0317.268.98*5.90
STAR Training Variants
C: Random embeddings2.674.938.3910.976.844.45
D: Reply sets as tokens2.851.59*8.766.817.751.57*
E: Predict replies separately2.2026.618.0730.986.4320.50
", + "image_path": "21861ba1b51ba9279542a9fb1e3c65069d56ce66744bcf3910596c4ea5239733.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 405, + 525, + 443 + ], + "lines": [ + { + "bbox": [ + 67, + 405, + 525, + 443 + ], + "spans": [ + { + "bbox": [ + 67, + 405, + 525, + 443 + ], + "type": "text", + "content": "Table 3: Performance of STAR on the Reddit, PersonaChat and DailyDialog Validation sets under different model configurations. Ablations are applied separately. **Bold indicates best result, **underline indicates second-best. * = statistically significant versus next best result on t-test with " + }, + { + "bbox": [ + 67, + 405, + 525, + 443 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 67, + 405, + 525, + 443 + ], + "type": "text", + "content": "-value < 0.01." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 463, + 189, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 463, + 189, + 476 + ], + "spans": [ + { + "bbox": [ + 67, + 463, + 189, + 476 + ], + "type": "text", + "content": "5.3 Run-time Efficiency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 484, + 291, + 728 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 484, + 291, + 728 + ], + "spans": [ + { + "bbox": [ + 67, + 484, + 291, + 728 + ], + "type": "text", + "content": "Beyond performance gains in relevance and diversity, a major advantage of an autoregressive retrieval model is the ability to leverage the scalability of GPU-based inference. Figure 3 compares the efficiency of STAR with the other baseline methods. We use an NVIDIA GeForce RTX 3060 Ti GPU and AMD Ryzen 7 5700G with Radeon Graphics CPU, with a batch size of 32. The results show that the methods can be broadly clustered into three groups. The slowest group is the generative method Seq2Seq, due to needing to generate each reply word-by-word. The middle group - SimSR, M-CVAE and M-MMR - is characterised by methods that comprise a more involved diversification pipeline. The final and fastest group includes STAR, M-Topic and Matching, where no additional post-hoc diversification is required (for M-Topic the topics can be pre-computed prior to inference)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 463, + 383, + 476 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 463, + 383, + 476 + ], + "spans": [ + { + "bbox": [ + 303, + 463, + 383, + 476 + ], + "type": "text", + "content": "5.4 Case Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 488, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 488, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 301, + 488, + 526, + 773 + ], + "type": "text", + "content": "Table 4 presents a case study on the DailyDialog Test set. We compare our approach, STAR, with the top-performing baseline from Table 2, SimSR. In both examples we consistently find STAR is able to output a broader range of intents. Quantitatively, we consider the rank that each suggestion receives according to the initial retrieval of the Matching model that underlies SimSR. We see that STAR is able to perform a much more global search across the reply space, selecting replies from within the top 100 or so ranks. This would be difficult for the standard retrieve-and-erank approach to emulate, given 100 is usually too large a number to efficiently rerank (Deb et al., 2019). Qualitatively, SimSR's suggestions converge around common phrases, e.g. 'let's go', which would be difficult to deduplicate with a heuristic rule given only a limited number of overlapping words between the replies. Conversely, STAR is able to represent a broader range of intents, such as replying with a question in both examples. Further examples are" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 740, + 290, + 773 + ], + "type": "text", + "content": "Results surpass reported numbers in original paper due to inclusion of language-model bias and deduplicated candidate pool to support better comparability (Section 4.1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "7617" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 80, + 512, + 176 + ], + "blocks": [ + { + "bbox": [ + 83, + 80, + 512, + 176 + ], + "lines": [ + { + "bbox": [ + 83, + 80, + 512, + 176 + ], + "spans": [ + { + "bbox": [ + 83, + 80, + 512, + 176 + ], + "type": "image", + "image_path": "080a6316394fb7b9c651b39cc33d1f2c91230cb059f3f066da7cf0303ddeff7c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 187, + 525, + 212 + ], + "lines": [ + { + "bbox": [ + 67, + 187, + 525, + 212 + ], + "spans": [ + { + "bbox": [ + 67, + 187, + 525, + 212 + ], + "type": "text", + "content": "Figure 2: Comparison of overall relevance and diversity scores across ablations, obtained by averaging across all three datasets with equal weighting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 68, + 254, + 289, + 413 + ], + "blocks": [ + { + "bbox": [ + 68, + 254, + 289, + 413 + ], + "lines": [ + { + "bbox": [ + 68, + 254, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 68, + 254, + 289, + 413 + ], + "type": "image", + "image_path": "dd266042b070a64a7b1c683f7b0ba67d2a9c00c2ad4f53a0ce0094baedb945cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 424, + 290, + 460 + ], + "lines": [ + { + "bbox": [ + 67, + 424, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 424, + 290, + 460 + ], + "type": "text", + "content": "Figure 3: Comparison of run-time efficiency between STAR and the baseline methods. Results are calculated over the Reddit Validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 77, + 506, + 280, + 682 + ], + "blocks": [ + { + "bbox": [ + 77, + 506, + 280, + 682 + ], + "lines": [ + { + "bbox": [ + 77, + 506, + 280, + 682 + ], + "spans": [ + { + "bbox": [ + 77, + 506, + 280, + 682 + ], + "type": "table", + "html": "
Message:Hi, Kenny. Let's go for a drink.
SimSR- let's go! [#9] - ok, let's go. [#3] - ok. let's get something to drink. [#1]
STAR- ok. let's go. [#5] - you want something to drink? [#89] - good idea. [#105]
Message:Of course! Let's go.
SimSR- let's go! [#1] - ok, let's go. [#5] - all right. let's go. [#12]
STAR- let's go! [#1] - where are we? [#43] - good idea! [#85]
", + "image_path": "30886690c9f9c576da8d6ddbe44dd1769834beeaccaecd511314ec9ddf363d55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 690, + 291, + 750 + ], + "lines": [ + { + "bbox": [ + 67, + 690, + 291, + 750 + ], + "spans": [ + { + "bbox": [ + 67, + 690, + 291, + 750 + ], + "type": "text", + "content": "Table 4: Example model outputs from the DailyDialog Test set, comparing STAR (ours) with the top-performing baseline method. Numbers in bold indicate the ranking the reply received according to the Matching model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 303, + 233, + 414, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 233, + 414, + 246 + ], + "spans": [ + { + "bbox": [ + 303, + 233, + 414, + 246 + ], + "type": "text", + "content": "provided in Appendix C." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 303, + 257, + 381, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 257, + 381, + 269 + ], + "spans": [ + { + "bbox": [ + 303, + 257, + 381, + 269 + ], + "type": "text", + "content": "6 Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "spans": [ + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "text", + "content": "We introduce STAR, an autoregressive retrieval system for SR, which is an end-to-end text-to-text model that sequentially predicts replies conditioned on an initial message. To train STAR, we demonstrate an approach to bootstrap a dataset of high-quality (message, reply set) pairs, from regular dialogue datasets containing only (message, reply) pairs. Empirically, our results show significant improvement over existing state-of-the-art SR baselines, across multiple datasets, corresponding to a " + }, + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "inline_equation", + "content": "5.1\\% - 17.9\\%" + }, + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "text", + "content": " improvement in relevance, and a " + }, + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "inline_equation", + "content": "0.5\\% - 63.1\\%" + }, + { + "bbox": [ + 302, + 280, + 526, + 455 + ], + "type": "text", + "content": " improvement in diversity compared to the best baseline approach." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 456, + 525, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 456, + 525, + 660 + ], + "spans": [ + { + "bbox": [ + 302, + 456, + 525, + 660 + ], + "type": "text", + "content": "Future work could extend these techniques to other set-prediction tasks: e.g., in IR the relevance of each document depends on the quantity of new information it contains compared to other documents in the set. In recommender systems, use cases include: tailoring a user's news feed requires that the news articles presented are not simply duplicates of the same story; designing a bespoke music playlist requires songs to be unified by common themes but also sufficiently distinct from one another to maintain the listener's interest. Other lines of future work include considering alternate strategies for initialising the reply embeddings, beyond the bag-of-words initialisation demonstrated in this paper." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 303, + 670, + 405, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 670, + 405, + 683 + ], + "spans": [ + { + "bbox": [ + 303, + 670, + 405, + 683 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "We thank the reviewers for their helpful feedback and suggestions. This work is partly supported by the EPSRC DTP Studentship program. The opinions expressed in this paper are the authors', and are not necessarily shared/endorsed by their employers and/or sponsors." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "7618" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 83 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 93, + 291, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 93, + 291, + 417 + ], + "spans": [ + { + "bbox": [ + 69, + 93, + 291, + 417 + ], + "type": "text", + "content": "Although our work shows that STAR is able to absorb sufficient information about the replies in its weights, this may become increasingly challenging when larger numbers of replies need to be embedded. One notable instance of this would be the multilingual setting, as many SR systems are deployed globally. In this case, each language typically has its own candidate pool. A naive implementation which creates separate reply vectors for each language would incur a significant increase in model size. In this case, we hypothesise techniques around weight-sharing between reply embeddings between languages may be beneficial, e.g. 'how are you' (en) and 'ça va' (fr) sharing the same vector. Further, our techniques are only demonstrated in publicly available datasets, whereas proprietary conversations in chat and email applications may have unique features not accounted for here (e.g. timestamps, cc and bcc information, and file attachments). Our technique also requires a planning algorithm to create the initial dataset. This theoretically creates an upper bound to the overall performance of STAR, as it is limited to cloning the behaviour of the offline planning algorithm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 439, + 127, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 439, + 127, + 452 + ], + "spans": [ + { + "bbox": [ + 68, + 439, + 127, + 452 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 459, + 290, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 459, + 290, + 537 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 290, + 537 + ], + "type": "text", + "content": "Dimosthenis Antypas, Asahi Ushio, Jose Camacho-Collados, Vitor Silva, Leonardo Neves, and Francesco Barbieri. 2022. Twitter topic classification. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3386-3400, Gyeongju, Republic of Korea. International Committee on Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 544, + 291, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 291, + 590 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 291, + 590 + ], + "type": "text", + "content": "Michele Bevilacqua, Giuseppe Ottaviano, Patrick Lewis, Wen tau Yih, Sebastian Riedel, and Fabio Petroni. 2022. Autoregressive search engines: Generating substrings as document identifiers. In NeurIPS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 597, + 291, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 740 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 740 + ], + "type": "text", + "content": "Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 750, + 291, + 772 + ], + "type": "text", + "content": "Nicola De Cao, Wilker Aziz, and Ivan Titov. 2021a. Highly parallel autoregressive entity linking with dis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 72, + 525, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 72, + 525, + 95 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 525, + 95 + ], + "type": "text", + "content": "criminative correction. In Conference on Empirical Methods in Natural Language Processing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 101, + 526, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 101, + 526, + 158 + ], + "spans": [ + { + "bbox": [ + 302, + 101, + 526, + 158 + ], + "type": "text", + "content": "Nicola De Cao, Gautier Izacard, Sebastian Riedel, and Fabio Petroni. 2021b. Autoregressive entity retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 164, + 526, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 164, + 526, + 231 + ], + "spans": [ + { + "bbox": [ + 302, + 164, + 526, + 231 + ], + "type": "text", + "content": "Nicola De Cao, Ledell Yu Wu, Kashyap Popat, Mikel Artetxe, Naman Goyal, Mikhail Plekhanov, Luke Zettlemoyer, Nicola Cancedda, Sebastian Riedel, and Fabio Petroni. 2021c. Multilingual autoregressive entity linking. Transactions of the Association for Computational Linguistics, 10:274-290." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 238, + 526, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 238, + 526, + 294 + ], + "spans": [ + { + "bbox": [ + 302, + 238, + 526, + 294 + ], + "type": "text", + "content": "Jaime G. Carbonell and Jade Goldstein-Stewart. 1998. The use of mmr, diversity-based reranking for reordering documents and producing summaries. In Annual International ACM SIGIR Conference on Research and Development in Information Retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 301, + 526, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 301, + 526, + 333 + ], + "spans": [ + { + "bbox": [ + 302, + 301, + 526, + 333 + ], + "type": "text", + "content": "Asli Celikyilmaz, Elizabeth Clark, and Jianfeng Gao. 2020. Evaluation of text generation: A survey. ArXiv, abs/2006.14799." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 341, + 526, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 341, + 526, + 397 + ], + "spans": [ + { + "bbox": [ + 302, + 341, + 526, + 397 + ], + "type": "text", + "content": "Nimesh Chakravarthi and Jeff Pasternack. 2017. Building smart replies for member messages. press release. https://engineering.linkedin.com/blog/2017/10/building-smart-replies-for-member-messages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 403, + 526, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 403, + 526, + 459 + ], + "spans": [ + { + "bbox": [ + 302, + 403, + 526, + 459 + ], + "type": "text", + "content": "Budhaditya Deb, Peter Bailey, and Milad Shokouhi. 2019. Diversifying reply suggestions using a matching-conditional variational autoencoder. In North American Chapter of the Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 466, + 526, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 466, + 526, + 544 + ], + "spans": [ + { + "bbox": [ + 302, + 466, + 526, + 544 + ], + "type": "text", + "content": "Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A conditional generative matching model for multi-lingual reply suggestion. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 1553–1568, Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 550, + 526, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 550, + 526, + 585 + ], + "spans": [ + { + "bbox": [ + 302, + 550, + 526, + 585 + ], + "type": "text", + "content": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc'Aurelio Ranzato. 2020. Revisiting self-training for neural sequence generation. In Proceedings of ICLR." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 592, + 526, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 592, + 526, + 647 + ], + "spans": [ + { + "bbox": [ + 302, + 592, + 526, + 647 + ], + "type": "text", + "content": "Matthew L. Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, Laszlo Lukacs, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. 2017. Efficient natural language response suggestion for smart reply. CoRR, abs/1705.00652." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 654, + 526, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 654, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 654, + 526, + 687 + ], + "type": "text", + "content": "Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 695, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 695, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 695, + 526, + 772 + ], + "type": "text", + "content": "Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. 2023. Unnatural instructions: Tuning language models with (almost) no human labor. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409-14428, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "7619" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 291, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 291, + 130 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 291, + 130 + ], + "type": "text", + "content": "Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and pre-training strategies for fast and accurate multisentence scoring. In International Conference on Learning Representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 137, + 291, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 137, + 291, + 180 + ], + "spans": [ + { + "bbox": [ + 67, + 137, + 291, + 180 + ], + "type": "text", + "content": "Gautier Izacard and Edouard Grave. 2021. Distilling knowledge from reader to retriever for question answering. In International Conference on Learning Representations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 190, + 291, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 291, + 235 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 291, + 235 + ], + "type": "text", + "content": "Youngsoo Jang, Jongmin Lee, and Kee-Eung Kim. 2020. Bayes-adaptive monte-carlo planning and learning for goal-oriented dialogues. In AAAI Conference on Artificial Intelligence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 243, + 291, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 243, + 291, + 289 + ], + "spans": [ + { + "bbox": [ + 67, + 243, + 291, + 289 + ], + "type": "text", + "content": "Youngsoo Jang, Seokin Seo, Jongmin Lee, and Kee-Eung Kim. 2021. Monte-carlo planning and learning with language action value estimates. In International Conference on Learning Representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 297, + 291, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 297, + 291, + 376 + ], + "spans": [ + { + "bbox": [ + 67, + 297, + 291, + 376 + ], + "type": "text", + "content": "Anjuli Kannan, Karol Kurach, Sujith Ravi, Tobias Kaufmann, Andrew Tomkins, Balint Miklos, Gregory S. Corrado, Laszlo Lukács, Marina Ganea, Peter Young, and Vivek Ramavajjala. 2016. Smart reply: Automated response suggestion for email. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 384, + 290, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 384, + 290, + 429 + ], + "spans": [ + { + "bbox": [ + 67, + 384, + 290, + 429 + ], + "type": "text", + "content": "Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. Dailydialog: A manually labelled multi-turn dialogue dataset. In International Joint Conference on Natural Language Processing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 438, + 290, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 438, + 290, + 472 + ], + "spans": [ + { + "bbox": [ + 67, + 438, + 290, + 472 + ], + "type": "text", + "content": "Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 481, + 291, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 291, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 291, + 515 + ], + "type": "text", + "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 523, + 291, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 579 + ], + "type": "text", + "content": "Ashwin Paranjape, Omar Khattab, Christopher Potts, Matei Zaharia, and Christopher D Manning. 2022. Hindsight: Posterior-guided training of retrievers for improved open-ended generation. In International Conference on Learning Representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 587, + 291, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 291, + 644 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 291, + 644 + ], + "type": "text", + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 652, + 291, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 291, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 291, + 719 + ], + "type": "text", + "content": "Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric Michael Smith, Y.-Lan Boureau, and Jason Weston. 2020. Recipes for building an open-domain chatbot. In Conference of the European Chapter of the Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 728, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 728, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 728, + 291, + 772 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. ArXiv, abs/1910.01108." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 72, + 526, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 72, + 526, + 127 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 526, + 127 + ], + "type": "text", + "content": "Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. ArXiv, abs/2302.04761." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 138, + 526, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 138, + 526, + 216 + ], + "spans": [ + { + "bbox": [ + 302, + 138, + 526, + 216 + ], + "type": "text", + "content": "Timo Schick and Hinrich Schütze. 2021a. Exploiting cloze-questions for few-shot text classification and natural language inference. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 255–269, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 227, + 526, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 227, + 526, + 295 + ], + "spans": [ + { + "bbox": [ + 302, + 227, + 526, + 295 + ], + "type": "text", + "content": "Timo Schick and Hinrich Schütze. 2021b. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 306, + 526, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 306, + 526, + 372 + ], + "spans": [ + { + "bbox": [ + 302, + 306, + 526, + 372 + ], + "type": "text", + "content": "Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, L. Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, Timothy P. Lillicrap, and David Silver. 2019. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588:604 - 609." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 383, + 526, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 383, + 526, + 460 + ], + "spans": [ + { + "bbox": [ + 302, + 383, + 526, + 460 + ], + "type": "text", + "content": "David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lancot, L. Sifre, Dharshan Kumaran, Thore Graepel, Timothy P. Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. ArXiv, abs/1712.01815." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 472, + 526, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 472, + 526, + 529 + ], + "spans": [ + { + "bbox": [ + 302, + 472, + 526, + 529 + ], + "type": "text", + "content": "Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 539, + 526, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 539, + 526, + 606 + ], + "spans": [ + { + "bbox": [ + 302, + 539, + 526, + 606 + ], + "type": "text", + "content": "Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 617, + 526, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 617, + 526, + 695 + ], + "spans": [ + { + "bbox": [ + 302, + 617, + 526, + 695 + ], + "type": "text", + "content": "Benjamin Towle and Ke Zhou. 2022. Learn what is possible, then choose what is best: Disentangling one-to-many relations in language through text-based games. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 4955–4965, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 706, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 526, + 772 + ], + "type": "text", + "content": "Benjamin Towle and Ke Zhou. 2023. Model-based simulation for optimising smart reply. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12030-12043, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "7620" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 291, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 291, + 161 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 291, + 161 + ], + "type": "text", + "content": "Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13484-13508, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 168, + 291, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 168, + 291, + 225 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 291, + 225 + ], + "type": "text", + "content": "Yue Weng, Huaixiu Zheng, Franziska Bell, and Gokhan Tur. 2019. Occ: A smart reply system for efficient in-app communications. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 232, + 291, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 232, + 291, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 232, + 291, + 267 + ], + "type": "text", + "content": "Kilian Wenker. 2023. Who wrote this? how smart replies impact language and agency in the workplace. *Telematics and Informatics Reports*, 10:100062." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 273, + 291, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 273, + 291, + 374 + ], + "spans": [ + { + "bbox": [ + 67, + 273, + 291, + 374 + ], + "type": "text", + "content": "Mozhi Zhang, Wei Wang, Budhaditya Deb, Guoqing Zheng, Milad Shokouhi, and Ahmed Hassan Awadallah. 2021. A dataset and baselines for multilingual reply suggestion. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1207-1220, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 380, + 291, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 291, + 438 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 291, + 438 + ], + "type": "text", + "content": "Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur D. Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? In Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 444, + 291, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 444, + 291, + 512 + ], + "spans": [ + { + "bbox": [ + 67, + 444, + 291, + 512 + ], + "type": "text", + "content": "Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and William B. Dolan. 2019. Dialogpt: Large-scale generative pre-training for conversational response generation. In Annual Meeting of the Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 518, + 291, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 518, + 291, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 518, + 291, + 553 + ], + "type": "text", + "content": "Tiancheng Zhao, Ran Zhao, and Maxine Eskenazi. 2017. Learning discourse-level diversity for neural dialog models using conditional variational autoencoders. In ACL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 562, + 209, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 562, + 209, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 562, + 209, + 576 + ], + "type": "text", + "content": "A Ethical Considerations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 584, + 291, + 773 + ], + "type": "text", + "content": "Controlling the outputs of dialogue models is a forefront issue in ethics research for AI, particularly with the impact of recent gains in LLM capabilities. We believe the risks in the case of SR systems have several mitigants compared to this: the replies can be vetted by humans before deployment; the replies are usually in short-form, rather containing complex information the user may rely on; ultimately, a user must select one of the options, rather than the system being able to reply without user oversight. Conversely, there are some risks more unique to SR systems that should be mentioned. Particularly, the suggestions presented by the system can have subtle priming effects on user behaviour. Notably," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 166 + ], + "type": "text", + "content": "users have been shown to be slightly more positive in the sentiment of their emails when shown suggested replies (Wenker, 2023). SR systems are on the whole known to produce more positive sentiment messages than the human distribution (Kannan et al., 2016). We see this as an extension of the broader trend of LLMs to be overly obsequious." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 175, + 448, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 175, + 448, + 190 + ], + "spans": [ + { + "bbox": [ + 302, + 175, + 448, + 190 + ], + "type": "text", + "content": "B Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "spans": [ + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "text", + "content": "For constructing the training dataset, we use the following hyperparameters: SimSR is initialised from the distilbert-base-uncased checkpoint (Sanh et al., 2019). We set the search parameters to " + }, + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "inline_equation", + "content": "M = 100" + }, + { + "bbox": [ + 302, + 196, + 526, + 332 + ], + "type": "text", + "content": ". We use a redundancy penalty of 0.05 and a blending alpha of 0.75 for query augmentation. Both parameters provided a good trade-off between relevance and diversity in early testing, so we did not search hyperparameters further (see Section 5.2 for ablations)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 333, + 526, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 333, + 526, + 576 + ], + "spans": [ + { + "bbox": [ + 302, + 333, + 526, + 576 + ], + "type": "text", + "content": "For training STAR, we initialise our model with the t5-small checkpoint. Note that this version of T5 has a comparable parameter count to the baselines which use DistilBERT (60M versus 66M). We tokenise the dataset with a maximum message length of 64 tokens. We train our model for up to 100k steps, with a warmup of 1k steps. In practice, the model typically converged around 20k steps. We use the AdamW optimiser (Loshchilov and Hutter, 2019) with an initial learning rate of 5e-4 and linear decay. We evaluate every 2k steps, by taking the ROUGE and Self-ROUGE scores on the validation set (this uses the ground-truth from the original dataset, not from the bootstrapped dataset), and employ early stopping once both metrics have ceased to improve. Note, we found these metrics were a much more reliable stopping point than crossentropy loss, which typically converged much earlier." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 586, + 436, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 586, + 436, + 599 + ], + "spans": [ + { + "bbox": [ + 302, + 586, + 436, + 599 + ], + "type": "text", + "content": "C Further Case Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 607, + 526, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 607, + 526, + 647 + ], + "spans": [ + { + "bbox": [ + 302, + 607, + 526, + 647 + ], + "type": "text", + "content": "Table 5 displays further examples of STAR's predictions versus SimSR, taken from the PersonaChat Test set." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "7621" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 138, + 222, + 455, + 584 + ], + "blocks": [ + { + "bbox": [ + 138, + 222, + 455, + 584 + ], + "lines": [ + { + "bbox": [ + 138, + 222, + 455, + 584 + ], + "spans": [ + { + "bbox": [ + 138, + 222, + 455, + 584 + ], + "type": "table", + "html": "
Persona:my favorite tv show is the office.
i like jazz music.
i do not drive because i live in new york.
i am jewish.
i do not eat anything sweet.
Message:i was but am now divorced
SimSR- i am sorry to hear that .
- oh . i am sorry to hear that .
- i am sorry to hear that . do you have any kids ?
STAR- oh i am sorry to hear that
- do you like music ?
- like music ? i love jazz
Persona:hey there my name is jordan and i am a veterinarian.
love to read drama books.
i love playing video games.
i am also a musician on the weekends.
i am originally from california but i live in florida.
Message:sometimes . i listen to a lot of music . do you read a lot ?
SimSR- hi how are you ?
- hi how are you today
- i do not like music at all
STAR- i listen to music do you
- no not really i like all music .
- yes i do my favorite is country
", + "image_path": "11d01728c7d182b2d20c3b1a8664d6d73e85809f44cdcea4466a397b0475ad66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 592, + 525, + 617 + ], + "lines": [ + { + "bbox": [ + 67, + 592, + 525, + 617 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 525, + 617 + ], + "type": "text", + "content": "Table 5: Example model outputs from the PersonaChat Test set, comparing STAR (ours) with the top-performing baseline method. STAR is able to capture a broader range of intents through its end-to-end autoregressive retrieval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "7622" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_content_list.json b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02bfbfa2f5932dd0f1e547cabba4a794e6280b96 --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "End-to-end Adversarial Sample Generation for Data Augmentation", + "text_level": 1, + "bbox": [ + 149, + 89, + 847, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyuan Liu and Yuqing Sun*", + "bbox": [ + 364, + 141, + 643, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shandong University", + "bbox": [ + 415, + 160, + 586, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zodiacg@foxmail.com sun_yuqing@sdu.edu.cn", + "bbox": [ + 295, + 177, + 705, + 192 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial samples pose a significant challenge to neural inference models. In this paper, we propose a novel enhancing approach $\\mathbf{A}^3$ for the robustness of the neural NLP models, which combines the adversarial training and data augmentation. We propose an adversarial sample generator that consists of a conditioned paraphrasing model and a condition generator. The latter aims to generate conditions which guides the paraphrasing model to generate adversarial samples. A pretrained discriminator is introduced to help the adversarial sample generator adapt to the data characteristics for different tasks. We adopt a weighted loss to incorporate the generated adversarial samples with the original samples for augmented training. Compared to existing methods, our approach is much efficient since the generation process is independent to the target model and the generated samples are reusable for different models. Experimental results on several tasks show that our approach improves the overall performance of the trained model. Specially, the enhanced model is robust for various attacking techniques.", + "bbox": [ + 144, + 280, + 460, + 634 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 646, + 258, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial samples refer to the inputs that intentionally designed to fool deep learning models (Morris et al., 2020a), which are also called the attack samples (Szegedy et al., 2013). The methods for generating attack samples have deceived many natural language processing (NLP) models, such as text classification (Ebrahimi et al., 2017; Ren et al., 2019), machine translation (Cheng et al., 2020) and etc (Goyal et al., 2022). Along with the application of neural models in various areas, the adversarial samples have drawn much attention.", + "bbox": [ + 112, + 673, + 489, + 848 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generally, the adversarial samples are generated by perturbations on the original sample. To enhance the robustness of NLP neural methods, the", + "bbox": [ + 112, + 850, + 489, + 897 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adversarial training approach uses the adversarial samples together with the original samples for training model (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) proposed the first NLP oriented adversarial training method by applying noise on the word representations. But the perturbed result is in the latent space and does not map to a meaningful sentence. Then many works turned to generating adversarial samples as a means of adversarial training. The often adopted method is the lexical substitution that is formulated as a combinatorial optimization problem and solved by heuristic search algorithms (Yoo and Qi, 2021). Since it requires much interaction with the target model for identifying the important words and selecting the appropriate replacement words, it is time-consuming. Due to the lack of supervised training data, so far there is not applicable end-to-end mode for adversarial generation.", + "bbox": [ + 507, + 252, + 884, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another kind of approaches to improve model robustness is data augmentation, which aims to generate the similar samples to the original data for increasing the amount of training data. The traditional methods on data augmentation include backtranslation (Sennrich et al., 2015), rule-based modification (Wei and Zou, 2019), paraphrasing (Kumar et al., 2019) and other techniques (Li et al., 2022). These methods are based on the distribution of training data, which are model agnostic. However, with these data, the models are still vulnerable to adversarial samples as the attacking techniques are designed for finding the weakness of the target model (Yoo and Qi, 2021).", + "bbox": [ + 507, + 562, + 885, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To solve the above challenges, in this paper, we designed an adversarial sample generation method for data augmentation $\\mathbf{A}^3$ . We propose an end-to-end adversarial sample generation model. Given an original sample, a condition generator is designed to output a condition which guides a conditioned paraphrasing model to generate a potential adversarial sample. By pretraining an adversarial dis", + "bbox": [ + 507, + 790, + 885, + 919 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Author", + "bbox": [ + 139, + 904, + 284, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11359", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11359-11368 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 210, + 945, + 786, + 972 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "criminator to be utilized as the training signal, it can learn the task-specific data characteristics for finding the potential vulnerabilities. Since the generation process doesn't require interaction with the target model, it is much efficient compared with the current search based generation methods. Besides, the generated samples can be reused for different models rather than for a specific model as the traditional way. Then we propose a weighted loss for incorporating the generated adversarial samples with the original data for the target task. We evaluate our method on several NLP tasks and compare it to both the data augmentation and adversarial training methods. Experimental results demonstrate that our approach not only improves the overall performance of the trained model but also enhances its robustness against multiple attacking techniques. Moreover, the computational time is significantly reduced compared to existing adversarial sample generation methods.", + "bbox": [ + 112, + 84, + 492, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related works", + "text_level": 1, + "bbox": [ + 112, + 419, + 273, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Adversarial Sample Generation", + "text_level": 1, + "bbox": [ + 112, + 445, + 410, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the NLP area, the disturbance on the original sample is often adopted in adversarial sample generation due to the discrete character of text, such as word replacement. For example, TextBugger (Li et al., 2019) generates the adversarial samples by randomly replacing the words in a sample text for gathering the candidate samples that may flip the model prediction. Then it chooses the reasonable adversarial samples with the similar semantics with the original text. PWWS (Ren et al., 2019) uses synonym for word substitution with the help of WordNet. With the development of pretrained language models (PLM), it helps various NLP tasks to gain improvement on the performance. Jin et al. proposed the TextFooler (Jin et al., 2020) method for evaluating the robustness of BERT. Meanwhile, several works utilize PLM to generate the candidate adversarial texts with similar semantics to original ones (Li et al., 2020, 2021). Recently, a survey on word substitution based adversarial sample generation (Chiang and Lee, 2022) shows that despite the effectiveness of such attacks, the generated samples have flaws on fluency and syntactical correctness and can be identified easily.", + "bbox": [ + 112, + 467, + 489, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 Adversarial Training", + "text_level": 1, + "bbox": [ + 112, + 865, + 327, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the proximity assumption, researchers have explored methods to enhance the robustness", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and effectiveness of trained models by introducing disturbances to the original training samples, commonly known as adversarial training (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) first proposed a NLP-oriented method that adds noise to the word embeddings of the input sample for adversarial training in text classification task. Similar approaches have been applied to adversarial training for pretrained language models (Liu et al., 2020) and other NLP tasks (Li and Qiu, 2021), showing improvements in model performance. Despite their effectiveness, disturbance in the continuous spaces lack clear meaning in actual language, making them less interpretable.", + "bbox": [ + 507, + 84, + 884, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works in adversarial training for NLP tasks consider modifications at the lexical level to make the perturbed samples more consistent with real samples. Early works use synonym substitution to explore proximate samples in the sample space. Huang et al. (Huang et al., 2019) and Dong et al. (Dong et al., 2021) used different sample space exploration methods to find samples that effectively enhance model robustness in adversarial training. The SSMBA method proposed by Ng et al. (Ng et al., 2020) generates perturbed samples by using pretrained language models to complete corrupted texts. Recently, researchers consider using adversarial sample generation methods for adversarial training. Previous work (Carlini et al., 2019) has shown that using samples generated by a specific adversarial sample generation method for augmented training generally only improves the robustness against that method. Yoo et al. designed the A2T method (Yoo and Qi, 2021), which introduced a fast lexical substitution adversarial sample generation method into the model training process, achieving improved robustness against multiple model attacking methods. Current adversarial training methods for NLP models are mostly based on constrained searching in the sample space, and lack the ability to self learn and adapt to models or datasets. Moreover, the lexical substitution based generation methods have potential problems regarding semantic soundness and syntactic correctness. (Chiang and Lee, 2022)", + "bbox": [ + 507, + 310, + 884, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 The End-to-end Adversarial Sample Based Data Augmentation", + "text_level": 1, + "bbox": [ + 507, + 821, + 857, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Framework", + "text_level": 1, + "bbox": [ + 507, + 865, + 650, + 879 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we design the end-to-end adversarial sample generation method for data augmentation.", + "bbox": [ + 507, + 887, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11360", + "bbox": [ + 477, + 927, + 527, + 940 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ce257ad75998fd282d32545e6a2196b1198fdc4fe824b9ca452e97531e858f7.jpg", + "image_caption": [ + "Figure 1: The framework for the end-to-end adversarial sample based data augmentation" + ], + "image_footnote": [], + "bbox": [ + 115, + 80, + 487, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a given sample $\\mathbf{x}$ , the condition generator gives the condition $\\mathbf{c}$ , based on which the paraphrasing model generates $\\mathbf{x}'$ that is a potential adversarial sample. The whole architecture consists of three parts, as shown in Fig.1. The first condition generator is designed to generate the appropriate conditions for guiding the generated paraphrase being an adversarial sample. The second is the pretrained conditioned paraphrasing model that generates the paraphrases satisfying the given condition.", + "bbox": [ + 112, + 269, + 487, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since there are no supervised data for training the condition generator, we propose an adversarial discriminator to provide the proper training signals, similar to the structure of a generative adversarial network (GAN) (Goodfellow et al., 2014a). The third part is the adversarial discriminator that can be a classifier or regression model for the target task, which checks the flip of predication.", + "bbox": [ + 112, + 431, + 487, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then for the target task, the generated adversarial samples are incorporated with the original data for data augmentation, which are used to train a model for this task.", + "bbox": [ + 112, + 561, + 489, + 624 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Condition Generator", + "text_level": 1, + "bbox": [ + 112, + 639, + 326, + 653 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The condition generator takes an original sample as input and aims to generate a condition $\\mathbf{c}$ for guiding the paraphrasing model to generate an adversarial sample. Let $G$ denote the condition generator.", + "bbox": [ + 112, + 662, + 487, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = G (\\mathbf {x}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 260, + 741, + 487, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are many candidate forms of condition $\\mathbf{c}$ , such as the ratio of changed words, the sentence structure or the sentiment. Following the normal way of generating adversarial samples, we adopt the lexical substitution indicator as the condition, which is in the form of binary sequence corresponding to the words in $\\mathbf{x}$ , denoted by $\\mathbf{c} \\in \\{0,1\\}^{|\\mathbf{x}|}$ . $c_{i} = 1$ indicates $x_{i}$ should be changed in the paraphrase $\\mathbf{x}'$ , $c_{i} = 0$ for otherwise.", + "bbox": [ + 112, + 774, + 490, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper we adopt a multi-layer bidirectional recurrent neural network as the basic network. The output hidden vector at each step is projected to a scalar. The output of the condition generator is fed into the conditioned paraphrasing model. Since the condition $\\mathbf{c}$ is defined as a sequence of binary value, while conventional activation function such as sigmoid outputs a contiguous value between 0 and 1. To maintain a reasonable form of $\\mathbf{c}$ and achieve a stable training, we adopt the straight-through estimator (Bengio et al., 2013) method as the activation function, such that the output of the condition generator will be binary.", + "bbox": [ + 507, + 84, + 884, + 294 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = S T E (G (\\mathbf {x})) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 626, + 306, + 882, + 323 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Pretraining the Conditioned Paraphrasing Model", + "text_level": 1, + "bbox": [ + 507, + 336, + 882, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The general purpose of paraphrasing is to generate a sample with the same meaning but different expression with the given text. Here we introduce the conditioned paraphrasing model for guiding the paraphrase. For a given text $\\mathbf{x}$ and a condition $\\mathbf{c}$ , it generates a paraphrase $\\mathbf{x}'$ with the similar semantics to $\\mathbf{x}$ and satisfying $\\mathbf{c}$ , formally:", + "bbox": [ + 507, + 373, + 882, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 498, + 882, + 514 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt a multi-layer bidirectional recurrent neural network for the conditioned paraphrasing model. The pretraining data are automatically generated, where the input is a pair of a text $\\mathbf{x}$ and the randomly selected positions as the condition c. As the paraphrase $\\mathbf{x}'$ , there are many candidate methods for choosing the replacement words, such as the synonym vocabulary or the masked language model. To ensure the semantic similarity and the syntactical correctness, we adopt the word com- posable knowledge (Liu et al., 2023) for choosing the words to form a new text $\\mathbf{x}'$ as the supervised target.", + "bbox": [ + 507, + 529, + 882, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since only a small proportion of the text are changed, the simple negative log-likelihood loss would lead the model to just copy the input. Thus, we apply a larger weight $\\omega_{p} > 0$ for the replacement positions $c_{i} = 1$ :", + "bbox": [ + 507, + 738, + 882, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {P} = - \\sum_ {i = 1} ^ {| \\mathbf {x} |} \\left(1 + c _ {i} \\omega_ {p}\\right) \\log p \\left(\\hat {x} _ {i} ^ {\\prime} = x _ {i} ^ {\\prime} | \\mathbf {x}, \\mathbf {c}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 829, + 882, + 873 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{x}_i^{\\prime}$ is the output of the model and $x_{i}^{\\prime}$ is the supervised paraphrase.", + "bbox": [ + 507, + 885, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11361", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4 Adversarial Discriminator", + "text_level": 1, + "bbox": [ + 112, + 84, + 369, + 98 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train the end-to-end adversarial sample generation model, we pretrain an adversarial discriminator to provide training signals, denote by $D$ . It is a trained classifier or a regression model for the target task.", + "bbox": [ + 112, + 105, + 489, + 184 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = D (\\mathbf {x}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 258, + 186, + 487, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretically the paraphrasing model should generate the semantically similar text with the original one. If the prediction $\\hat{y}^\\prime$ by $D$ for the generated sample $x^{\\prime}$ is flipped compared to the ground truth $y$ , $x^{\\prime}$ is a potentially adversarial sample.", + "bbox": [ + 112, + 208, + 487, + 288 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5 End-to-End Adversarial Sample Generation Training", + "text_level": 1, + "bbox": [ + 112, + 299, + 413, + 331 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The core of the adversarial sample generation model is to generate the appropriate condition for paraphrasing. Since the discriminator can be seen as a white-box victim model, we can pass the gradient through the neural network structure by utilizing the tricks like Gumbel softmax (Jang et al., 2017). This allows an end-to-end training mode for the whole adversarial sample generation method.", + "bbox": [ + 112, + 336, + 487, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the training, the condition generator, the conditioned paraphrasing model and the adversarial discriminator are connected in a pipeline way. For a given sample $\\mathbf{x}$ , the pipeline outputs a prediction $\\hat{y}$ for a perturbed sample. The conditioned paraphrasing model $F(\\mathbf{x},\\mathbf{c})$ outputs a sequence of word probabilities. We adopt Gumbel softmax to convert the probabilities to word embeddings that can be directly used by $D$ .", + "bbox": [ + 112, + 464, + 489, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = G (\\mathbf {x})\n$$\n", + "text_format": "latex", + "bbox": [ + 255, + 620, + 332, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 638, + 485, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = D (\\mathbf {x} ^ {\\prime})\n$$\n", + "text_format": "latex", + "bbox": [ + 257, + 658, + 339, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, the parameters of the condition generator and the conditioned paraphrasing model are tuned while the adversarial discriminator remains frozen. Training the condition paraphrasing model allows the model to be trained adaptively to generate words more suitable for the target task. To train the adversarial sample generation model, we used the following losses.", + "bbox": [ + 112, + 684, + 487, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial loss is the main goal of the model, namely to generate a sample that can make the target model flip its prediction. Since previous works on model distilling show the effectiveness and simplicity of the MSE loss function (Hinton et al., 2015), thus it is adopted in our work. More", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "advanced distances between probabilities such as KL-distance can be an alternative.", + "bbox": [ + 507, + 84, + 880, + 115 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {a d v} = 1 - \\left| D \\left(\\mathbf {x} ^ {\\prime}\\right) - D (\\mathbf {x}) \\right| ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 128, + 882, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perturbation ratio loss To maintain the similar semantics between the perturbed sample and the original sample, the ratio of perturbed words in text are constrained, namely the number of $c_{i} = 1$ cases in the generated condition $\\mathbf{c}$ . Let $\\rho$ denote the expected ratio, $L$ is the length of sample, the perturbation ratio loss is calculated as follows.", + "bbox": [ + 507, + 160, + 882, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c o n} = \\left| \\rho - \\| \\mathbf {c} \\| _ {1} / L \\right| ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 284, + 882, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioned paraphrasing losses The conditioned paraphrasing model is fine-tuned for different expectation by a few loss functions. The first is the reconstruction loss for the words not perturbed in $\\mathbf{c}$ , i.e. on the positions $c_{i} = 0$ , denoted by $\\mathbf{c}(0)$ . It is a negative likelihood loss for the original words such that the paraphrasing model is trained to reconstruct the original input.", + "bbox": [ + 507, + 316, + 882, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {r e c} = - \\sum_ {i \\in \\mathbf {c} (0)} \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 564, + 457, + 882, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the perturbed positions, the probability of the original word should be lower than any other words. Inspired by the loss function for the negative samples in the noise contrastive estimation (Gutmann and Hyvarinen, 2010), we design the loss of punishing repeating the original word. Let $\\mathbf{c}(1)$ denote the positions of the 1s in the condition $\\mathbf{c}$ .", + "bbox": [ + 507, + 504, + 882, + 615 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {p u n} = - \\sum_ {i \\in \\mathbf {c} (1)} \\sigma (- \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right)) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 629, + 882, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics similarity and syntactic correctness losses To maintain the semantics of the generated paraphrase similar with the original sample, we adopt a word-level similarity loss based on cosine similarities for the word embeddings between the generated word and the original word.", + "bbox": [ + 507, + 686, + 880, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {s e m} = 1 - \\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 579, + 793, + 756, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) = \\frac {1}{| \\mathbf {x} |} \\sum_ {i = 1} ^ {| x |} \\left(1 - \\cos \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i} ^ {\\prime}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 815, + 880, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The word embeddings similarity cannot guarantee a proper syntactical role of the replacement word. Thus, we also adopt the word composable", + "bbox": [ + 507, + 871, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11362", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1 The Training Process for End-to-End Adversarial Sample Generator", + "Input: Training set $D_{t}$" + ], + "code_body": "Output: Condition generator $G$ , conditioned paraphrasing model $F$", + "guess_lang": "txt", + "bbox": [ + 115, + 135, + 487, + 167 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: //Pretrain conditioned paraphrasing model", + "2: Generate paraphrasing training data $D_{p}$ based on $D_{t}$", + "3: Pretrain $F$ with $D_{p}$ according to Eq.4", + "4: //Pretrain adversarial discriminator", + "5: Pretrain $D$ with $D_{t}$", + "6: //Train the end-to-end adversarial sample generation model", + "7: Freeze the parameters of $D$", + "8: for sample $\\mathbf{x}, y$ in $D_{t}$ do", + "9: Let $\\mathbf{c} = G(\\mathbf{x})$", + "10: Calculate $\\ell_{con}$ according to Eq.8", + "11: Let $\\mathbf{x}' = F(\\mathbf{x},\\mathbf{c})$", + "12: Calculate $\\ell_{rec},\\ell_{pun},\\ell_{sem},\\ell_{syn}$ according to Eq.9-12", + "13: Let $\\hat{y} = D(\\mathbf{x}^{\\prime})$", + "14: Calculate $\\ell_{adv}$ according to Eq.7", + "15: Calculate $L_{adv}$ according to Eq.13", + "16: Update $G,F$ according to $L_{adv}$", + "17: end for", + "18: return Condition generator $G$ , conditioned paraphrasing model $F$" + ], + "bbox": [ + 119, + 168, + 487, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "knowledge based loss (Liu et al., 2023) to provide syntactical correctness guidance for the generated sample. Given a dependency relationship between two words $< u, r, v >$ , the word composable knowledge estimates the syntactical soundness of this combination, denoted by $p_s(u, r, v)$ . It relies on the pretrained word embeddings and thus is easy to integrate with other neural network structures. Let $s_x = (i, r, j)$ denote the dependency parsing tree of $\\mathbf{x}$ , where $(i, r, j)$ represents the dependency relation between $\\mathbf{x}_j$ and its head word $\\mathbf{x}_i$ is $r$ . The syntactic correctness loss is calculated as follows:", + "bbox": [ + 112, + 551, + 489, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {s y n} = - \\sum_ {(i, r, j) \\in s _ {x}} \\log p _ {s} \\left(\\mathbf {x} _ {i} ^ {\\prime}, r, \\mathbf {x} _ {j} ^ {\\prime}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 154, + 757, + 487, + 793 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The complete loss function for training the adaptive adversarial sample generation model is the weighted sum of the above losses.", + "bbox": [ + 112, + 806, + 487, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {a d v} = w _ {a} \\ell_ {a d v} + w _ {c} \\ell_ {c o n} + w _ {r} \\ell_ {r e c} + w _ {p} \\ell_ {p u n} \\\\ + w _ {s e m} \\ell_ {s e m} + w _ {s y n} \\ell_ {s y n} \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 868, + 485, + 919 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c52d956a59da1b610a511a48ec3034e9dd70abeb75f0c5c14ba30f0691f493d9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Class#Train#TestAve.Len.
IMDB22500025000270.7
Quora23636604040425.33
SNLI35501521000022.4
", + "bbox": [ + 510, + 80, + 894, + 149 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Dataset statistics", + "bbox": [ + 606, + 160, + 783, + 173 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training process of the end-to-end adversarial sample generation is listed in Alg.1.", + "bbox": [ + 507, + 199, + 880, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6 Data Augmentation", + "text_level": 1, + "bbox": [ + 507, + 241, + 709, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having the above end-to-end adversarial sample generator, we can apply data augmentation on the training data for the target task. For the given original training set $D_{t}$ , the corresponding generated adversarial samples is denoted by $D_{t}^{\\prime}$ . Then the model for the target task can be trained on $D_{t} \\cup D_{t}^{\\prime}$ for improving the model robustness. Since the adversarial samples are more likely to introduce noises, their training loss is assigned a lower weight. Let $L_{t}$ denotes the loss of the target task, $\\lambda_{i}$ denotes the loss weight for sample $s_{i}$ . $\\lambda_{i} = 1$ for an original sample while $\\lambda_{i} < 1$ for a generated sample.", + "bbox": [ + 507, + 261, + 882, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t a s k} = \\sum_ {s _ {i} \\in D _ {t} \\bigcup D _ {t} ^ {\\prime}} \\lambda_ {i} L _ {t} (s _ {i}) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 464, + 882, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The previous work has shown the adversarial samples generated based on one model are potentially adversarial to other models (Szegedy et al., 2013).", + "bbox": [ + 507, + 508, + 882, + 557 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 507, + 568, + 655, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the proposed method, four aspects should be considered and evaluated in the subsections 4.2-4.5.", + "bbox": [ + 507, + 594, + 882, + 640 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "A1: To check the attacking effectiveness of generated adversarial sample.", + "A2: To check whether the model performance is improved after augmented training.", + "A3: To check whether the robustness of the target model is improved.", + "A4: To check the efficiency of our method." + ], + "bbox": [ + 507, + 650, + 882, + 788 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Experiment Setups", + "text_level": 1, + "bbox": [ + 507, + 799, + 705, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Tasks", + "text_level": 1, + "bbox": [ + 507, + 819, + 613, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three tasks are considered as evaluation tasks to cover a wider range of NLP tasks such that the effectiveness of data augmentation can be thoroughly evaluated in different scenarios. The first is a sentiment analysis task on the dataset IMDB. The IMDB", + "bbox": [ + 507, + 838, + 882, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11363", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6bb7a4770d226dfc2bf62cf2ca79206e096a5e9856c45ba8a11937a66ad445d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelBERTDistilBERTPerturbed %Bert Score
DatasetOri.Acc.Att.Acc.ΔOri.Acc.Att.Acc.Δ
IMDB93.9%66.58%-27.32%92.96%61.76%-31.2%9.95%95.98
Quora84.84%75.95%-8.89%84.32%76.44%-7.88%16.6%95.96
SNLI88.95%55.64%-33.31%88.04%55.20%-32.84%14.73%95.62
", + "bbox": [ + 122, + 80, + 873, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset consists of long movie reviews which are classified as positive of negative sentiment polarity. The second is a text classification task on the dataset Quora Question Pairs(QQP), which consists of question pairs from the Quora site classified as duplicate or not. The last is a natural language inference task on the dataset SNLI. Each sample in SNLI consists a hypothesis sentence and a premise sentence, the logical relationship between them is determined as entailment, neutral or contradiction. The statistics of the datasets are listed in Table.1.", + "bbox": [ + 112, + 233, + 489, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 Base Models", + "text_level": 1, + "bbox": [ + 112, + 420, + 273, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Recently, pretrained language models are widely used as a fundamental part for NLP methods. In the following experiments, BERT (Kenton and Toutanova, 2019) and DistilBERT (Sanh et al., 2019) are chosen as the base model for the tasks. BERT is a widely used pretrained language model, and achieves improvements on various NLP tasks. DistilBERT is a compressed version of BERT model based on model distilling technique, making it easier to be used on devices with lower computation power. Both models will be finetuned on the original train set or the augmented train set for the target task and evaluated afterwards. Their results will be listed as Baseline in the following experiments.", + "bbox": [ + 112, + 439, + 489, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.3 Implementation", + "text_level": 1, + "bbox": [ + 112, + 690, + 302, + 706 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the experiments are conducted on a single NVIDIA Geforce RTX 2080Ti. Our method is implemented with PyTorch 1.12.0. The code for our proposed method is shared at https://splab.sdu.edu.cn/xscg/sjjydm.htm. The compared methods and adversarial sample generation are implemented by TextAttack (Morris et al., 2020b). The size of hidden vectors in our model is 300. We adopt GloVe (Pennington et al., 2014) as the word embeddings and reused the pretrained parameters of word composable knowledge released in SynPara (Liu et al., 2023). For pretraining the paraphrasing model, the ratio of replaced words is set", + "bbox": [ + 112, + 709, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cbc1faa007bc3cd97bc3826084c43ea1ad9982f81d85d515cc32d4fbc935263.jpg", + "table_caption": [ + "Table 2: Accuracy comparison between accuracies on the original testing set and adversarial testing set." + ], + "table_footnote": [], + "table_body": "
Base ModelMethodDataset
IMDBQuoraSNLI
BERTBaseline93.984.8488.95
SSMBA-0.41+3.03-2.88
A2T-0.03+1.4-0.5
A3+1+4+0.1
Distil-BERTBaseline91.8784.3388.04
SSMBA+1.09+3.79-0.12
A2T+1.27+4.39+1.38
A3+1.93+4.89+0.22
", + "bbox": [ + 514, + 230, + 878, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Accuracy comparison on the original testing set. The values for the augmentation methods are the differences with the accuracy of the base model.", + "bbox": [ + 507, + 405, + 882, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$15\\%$ when generating training data. For training the complete model, the weights in the final loss function are set as $w_{a} = 5.0$ , $w_{c} = 3.0$ , $w_{r} = 2.0$ , $w_{p} = 5.0$ , $w_{sem} = 2.0$ , $w_{syn} = 0.1$ in most cases. The ratio of perturbed words $\\rho$ for the condition generator is set as 0.15 for SNLI and 0.1 for IMDB and Quora unless stated otherwise.", + "bbox": [ + 507, + 473, + 882, + 586 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.4 Compared methods", + "text_level": 1, + "bbox": [ + 507, + 595, + 724, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the comparison methods, we consider the following methods. SSMBA (Ng et al., 2020) is a data augmentation method that uses the corruption function similar to the masked language model training and reconstructs the samples with BERT. A2T (Yoo and Qi, 2021) is an adversarial training method, where a fast adversarial sample generation method uses the gradients for identifying important words in a sample and replaces them based on word embeddings or MLM models. It iteratively trains the task model and augment the training data set with the generated adversarial samples.", + "bbox": [ + 505, + 614, + 882, + 807 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Attack Effectiveness", + "text_level": 1, + "bbox": [ + 507, + 818, + 717, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the generated adversarial samples, we compared the performances of the baseline models on the original testing data and on the adversarial data. The results are shown in Table.2, where the ratio of perturbed words in", + "bbox": [ + 507, + 839, + 882, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11364", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/119e918b24bad9a9f349391ff8ab646e760f1a109cfe713af6d6a0b3e325cb2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelBERTDistilBERT
DatasetsIMDBQuoraSNLIIMDBQuoraSNLI
Attack MethodAugment Method%Δ%Δ%Δ%Δ%Δ%Δ
A3Baseline69.377.358.364.180.159.8
SSMBA71.6+2.377.7+0.458.7+0.461.6-2.578.8-1.362.9+3.1
A2T73.7+4.476.9-0.459.7+1.466.1+278.6-1.560.3+0.5
A392.2+22.989.3+1286.2+27.993.8+29.790.9+10.886.5+26.7
A2TBaseline38.840.144.130.650.341.1
SSMBA38.9+148.1+849.3+5.227.9-2.750.7+0.445.1+4
A2T44+5.276.2+36.179.4+35.354.5+23.976.1+25.875.6+34.5
A345.3+6.560.7+20.648+3.936.1+5.560.2+9.943.3+2.2
Text-FooterBaseline1.723.60.90.110.30.47
SSMBA5.3+3.612.5-11.10.4-0.50-0.112.1+1.80.73+0.26
A2T0.8-0.914.3-9.35.1+4.21.3+1.28-2.33.93+3.46
A312+10.327.1+3.51.9+11.6+1.515.2+4.90.67+0.2
BAEBaseline3139.322.124.939.323.87
SSMBA30.6-0.439-0.325+2.921-3.939.7+0.423.93+0.06
A2T29.1-1.943.4+4.123.5+1.426.3+1.444.2+4.925.5+1.63
A337.9+6.940.8+1.527.6+5.527.7+2.845.7+6.424.27+0.4
PWWSBaseline125.11.20.822.11.07
SSMBA2.2+1.223.1-22.3+1.10.1-0.722.9+0.81.2+0.13
A2T0.3-0.726.3+1.20.7-0.50.6-0.223+0.90.87-0.2
A37.3+6.327.7+2.62.4+1.20.9+0.124.1+22.4+1.33
", + "bbox": [ + 115, + 80, + 884, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4: Comparison of attacking methods against augmentation methods. Column % denotes the model accuracy under attacking. Column $\\Delta$ denotes the difference between augmented model with the original model.", + "bbox": [ + 112, + 386, + 882, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the adversarial samples is listed in column Perturbed%. A lower ratio indicates more structure information of the text preserved. BertScore (Zhang et al., 2020) is used for quantifying the semantics similarity between the original sample and the adversarial sample. The higher the BertScore, the closer their semantics. We can see that the model performances of both baseline models drop significantly on the adversarial samples, which show that the generated adversarial samples are effective on attacking the baseline models.", + "bbox": [ + 112, + 439, + 489, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Performance Improvement", + "text_level": 1, + "bbox": [ + 112, + 636, + 371, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We then evaluated the model performance on the target tasks before and after training with augmented data. By using our method and SSMBA, the data is augmented to doubled size of the original training set, respectively. The augmented dataset is used for training the base models. For A2T, it iteratively trains the model and then augments the training set by generating adversarial samples. The results in Table.3 show that our method improves the performance on all settings and are better than the compared methods in most cases. The compared methods are unstable on performances, namely in some cases performances are the improved while some are not. For example, SSMBA has negative impact on SNLI dataset for both base models. This is probably because SSMBA doesn't take into con", + "bbox": [ + 112, + 661, + 489, + 917 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sideration the task-specific characteristics, such as the logical connections between the hypothesis and premise in SNLI. It should be also noted that while the original performance of DistilBERT is slightly lower than BERT, it improves a lot after data augmentation, which confirmed previous studies regarding the effectiveness of data augmentation for distilled models (Das et al., 2020).", + "bbox": [ + 507, + 439, + 884, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 Robustness Improvement", + "text_level": 1, + "bbox": [ + 507, + 588, + 756, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the robustness of base model against different attacking methods, we choose three representative attacking methods: TextFooler (Jin et al., 2020), BAE (Garg and Ramakrishnan, 2020) and PWWS (Ren et al., 2019). We adopt the notion $m1 - m2$ denoting the attacking method $m1$ against augmentation method $m2$ . For example, BAE-A2T means using BAE to attack a model trained using A2T. Similar to the settings of previous works (Yoo and Qi, 2021), 1000 samples are selected from the original testing set of each dataset (1500 samples from SNLI for three labels), where the label distributions are balanced. To better show the impact of an attacking method, the selected samples can be predicted correctly by the base models before data augmentation. The attack methods generate the adversarial samples based on the selected samples and the accuracy on these adversarial samples of the trained model would be compared. A2T and", + "bbox": [ + 507, + 613, + 884, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11365", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d3b32ca86f4160b89f1e76a25b93ac24f3cd01c012ea471653d8ac4438e6527.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsDataset
IMDBQuoraSNLI
SSMBA10m38m1h2m
A2T36h(3)111h(2)70h(2)
A34m27m21m
", + "bbox": [ + 146, + 80, + 455, + 165 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5: Time for augmenting the dataset. It should be noted that the time for A2T is the cumulated time for multiple training epochs (number of epochs is listed in the brackets).", + "bbox": [ + 112, + 175, + 487, + 231 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a019b2b244879cc0e9658e0a2870775cec70d24b3a78a269374655290b3cb541.jpg", + "image_caption": [ + "Figure 2: The influence of $\\rho$ on the perturbed ratio and the attack effectiveness (denoted by $\\Delta \\mathrm{Acc}$ )" + ], + "image_footnote": [], + "bbox": [ + 144, + 244, + 458, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our proposed method are tested as they are also adversarial sample generation methods.", + "bbox": [ + 112, + 450, + 485, + 480 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are listed in Table.4, where the column $\\%$ denotes the model accuracy under attacking and column $\\Delta$ shows the difference between the performance of the augmented model with the baseline. As explored in a previous research (Carlini et al., 2019), the robustness against a certain attack method will be improved if trained with the corresponding adversarial samples. From the row $\\mathrm{A}^3\\mathrm{-A}^3$ and A2T-A2T, we can see that both methods significantly improve the model robustness. For other representative attacking methods, the augmented models with our generated data show better robustness than the compared methods in most settings.", + "bbox": [ + 112, + 482, + 489, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 Efficiency Comparison", + "text_level": 1, + "bbox": [ + 112, + 703, + 339, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another vital aspect of adversarial training is the efficiency. Most adversarial sample generation methods involves constantly interaction with the target model. Such process is both computational heavy and time-consuming. We recorded the time used for augmenting the training set of different methods. Both SSMBA and our proposed method are one-time augmentation thus the time taken is similar. Our method is slightly faster than SSMBA since SSMBA uses BERT as the reconstruction model. The A2T method, on the other hand, is significantly slower than our method due to two possible rea", + "bbox": [ + 112, + 726, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "son. First reason is that the attacking process is conducted multiple times. Secondly, the method is based on attacking the target model which is not only time-consuming, but the time taken will also increase over time since the model is trained better after each epoch. Comparably, our proposed method generates model-agnostic adversarial samples, such that the time for augmented training is significantly shortened.", + "bbox": [ + 507, + 84, + 884, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6 Parameter Analysis", + "text_level": 1, + "bbox": [ + 507, + 243, + 710, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the effect of parameter $\\rho$ , the targeted perturbed ratio. In our method for a $\\rho$ setting, the actual perturbed ratio in the generated adversarial samples may be different with the initial value. Thus, we analyze the actual perturbed ratio and its impact on the attacking effectiveness, shown in Fig.2. The blue and orange lines represent the set $\\rho$ and the actual perturbed ratio respectively, corresponding to the left Y-axis. The $\\Delta \\mathrm{Acc}$ is the accuracy difference after the attack, corresponding to the right Y-axis. Higher value of $\\Delta \\mathrm{Acc}$ means more effective attack. We can see that the actual perturbed ratio approaches the $\\rho$ setting. A higher $\\rho$ may cause the predicated label flipped but may lead to a larger semantics difference. Thus, the parameter $\\rho$ should be set properly to achieve a balanced performance.", + "bbox": [ + 507, + 265, + 884, + 538 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 552, + 640, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a method named $\\mathrm{A}^3$ that combines the advantages of adversarial training and data augmentation to enhance the robustness of NLP models. We introduce an end-to-end adversarial sample generation model which consists of a condition generator and a conditioned paraphrasing model. A pretrained adversarial discriminator is proposed to train the adversarial sample generation model for capturing the task-specific data characteristics. This method is much efficient without the need for extensive model interactions. The generated adversarial samples are incorporated with the original training samples by a weighted loss for augmented training downstream models. Experiments are conducted on multiple NLP tasks and demonstrate the effectiveness and efficiency of our proposed method. Compared to existing methods, the performances of the base models have improved more after our data augmentation, as well as their robustness enhanced against multiple attacking techniques.", + "bbox": [ + 507, + 580, + 884, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11366", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 99 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Firstly, the current form of condition limits the diversity and fine-grained control over the generation process. For example, a condition can be a subtree from the dependency syntax structure, which allows paraphrases with potential structure reform and correct syntax (Liu et al., 2023). Secondly, a condition may also contain the direction of replacement word for the paraphrasing model. This would allow reusing the same conditioned paraphrasing model across different tasks, thereby reducing the training overhead for adversarial sample generation. Lastly, in future research, we aim to explore the feasibility of directly using the gradient from the target model for training the adversarial sample generation model.", + "bbox": [ + 112, + 110, + 492, + 351 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 114, + 366, + 278, + 382 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (62376138) and the Innovative Development Joint Fund Key Projects of Shandong NSF (ZR2022LZH007).", + "bbox": [ + 112, + 393, + 489, + 458 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 486, + 213, + 502 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432.", + "Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. 2019. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705.", + "Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan Zhang, and Cho-Jui Hsieh. 2020. Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3601-3608.", + "Cheng-Han Chiang and Hung-yi Lee. 2022. How far are we from real synonym substitution attacks? arXiv preprint arXiv:2210.02844.", + "Deepan Das, Haley Massa, Abhimanyu Kulkarni, and Theodoros Rekatsinas. 2020. An empirical analysis of the impact of data augmentation on knowledge distillation. CoRR, abs/2006.03810.", + "Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. 2021. Towards robustness against natural language word substitutions. arXiv preprint arXiv:2107.13541." + ], + "bbox": [ + 115, + 510, + 489, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Javid Ebrahimi, Anyi Rao, Daniel Lowd, and De- jing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751.", + "Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6174-6181.", + "Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. 2014a. Generative adversarial networks. CoRR, abs/1406.2661.", + "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014b. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572.", + "Shreya Goyal, Sumanth Doddapaneni, Mitesh M Khapra, and Balaraman Ravindran. 2022. A survey of adversarial defences and robustness in nlp. ACM Computing Surveys.", + "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297-304. JMLR Workshop and Conference Proceedings.", + "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.", + "Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, and Pushmeet Kohli. 2019. Achieving verified robustness to symbol substitutions via interval bound propagation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4083-4093.", + "Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017). OpenReview.net.", + "Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025.", + "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.", + "Ashutosh Kumar, Satwik Bhattamishra, Manik Bhandari, and Partha Talukdar. 2019. Submodular" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11367", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "optimization-based diverse paraphrasing and its effectiveness in data augmentation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3609-3619.", + "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open, 3:71-90.", + "Dianqi Li, Yizhe Zhang, Hao Peng, Liquin Chen, Chris Brockett, Ming-Ting Sun, and William B Dolan. 2021. Contextualized perturbation for textual adversarial attack. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5053-5069.", + "J Li, S Ji, T Du, B Li, and T Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium.", + "Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020. Bert-attack: Adversarial attack against using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6193–6202.", + "Linyang Li and Xipeng Qiu. 2021. Token-aware virtual adversarial training in natural language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8410-8418.", + "Tianyuan Liu, Yuqing Sun, Jiaqi Wu, Xu Xi, Yuchen Han, Cheng Li, and Bin Gong. 2023. Unsupervised paraphrasing under syntax knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 37.", + "Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994.", + "Takeru Miyato, Andrew M Dai, and Ian Goodfellow. 2016. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725.", + "John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, and Yanjun Qi. 2020a. Reevaluating adversarial examples in natural language. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3829-3839.", + "John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020b. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nathan Ng, Kyunghyun Cho, and Marzyeh Ghassemi. 2020. Ssmba: Self-supervised manifold based data augmentation for improving out-of-domain robustness. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1268-1283.", + "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics.", + "Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. 2019. Generating natural language adversarial examples through probability weighted word saliency. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 1085-1097.", + "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.", + "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709.", + "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199.", + "Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196.", + "Jin Yong Yoo and Yanjun Qi. 2021. Towards improving adversarial training of nlp models. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 945-956.", + "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net." + ], + "bbox": [ + 510, + 85, + 882, + 730 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11368", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_model.json b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4946eeebbcce07bd98c621d04fbacb783543e23e --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.151, + 0.09, + 0.848, + 0.112 + ], + "angle": 0, + "content": "End-to-end Adversarial Sample Generation for Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.142, + 0.644, + 0.159 + ], + "angle": 0, + "content": "Tianyuan Liu and Yuqing Sun*" + }, + { + "type": "text", + "bbox": [ + 0.416, + 0.161, + 0.588, + 0.176 + ], + "angle": 0, + "content": "Shandong University" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.178, + 0.706, + 0.193 + ], + "angle": 0, + "content": "zodiacg@foxmail.com sun_yuqing@sdu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.281, + 0.461, + 0.636 + ], + "angle": 0, + "content": "Adversarial samples pose a significant challenge to neural inference models. In this paper, we propose a novel enhancing approach \\(\\mathbf{A}^3\\) for the robustness of the neural NLP models, which combines the adversarial training and data augmentation. We propose an adversarial sample generator that consists of a conditioned paraphrasing model and a condition generator. The latter aims to generate conditions which guides the paraphrasing model to generate adversarial samples. A pretrained discriminator is introduced to help the adversarial sample generator adapt to the data characteristics for different tasks. We adopt a weighted loss to incorporate the generated adversarial samples with the original samples for augmented training. Compared to existing methods, our approach is much efficient since the generation process is independent to the target model and the generated samples are reusable for different models. Experimental results on several tasks show that our approach improves the overall performance of the trained model. Specially, the enhanced model is robust for various attacking techniques." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.648, + 0.26, + 0.663 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.674, + 0.49, + 0.849 + ], + "angle": 0, + "content": "Adversarial samples refer to the inputs that intentionally designed to fool deep learning models (Morris et al., 2020a), which are also called the attack samples (Szegedy et al., 2013). The methods for generating attack samples have deceived many natural language processing (NLP) models, such as text classification (Ebrahimi et al., 2017; Ren et al., 2019), machine translation (Cheng et al., 2020) and etc (Goyal et al., 2022). Along with the application of neural models in various areas, the adversarial samples have drawn much attention." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.851, + 0.49, + 0.898 + ], + "angle": 0, + "content": "Generally, the adversarial samples are generated by perturbations on the original sample. To enhance the robustness of NLP neural methods, the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.253, + 0.885, + 0.559 + ], + "angle": 0, + "content": "adversarial training approach uses the adversarial samples together with the original samples for training model (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) proposed the first NLP oriented adversarial training method by applying noise on the word representations. But the perturbed result is in the latent space and does not map to a meaningful sentence. Then many works turned to generating adversarial samples as a means of adversarial training. The often adopted method is the lexical substitution that is formulated as a combinatorial optimization problem and solved by heuristic search algorithms (Yoo and Qi, 2021). Since it requires much interaction with the target model for identifying the important words and selecting the appropriate replacement words, it is time-consuming. Due to the lack of supervised training data, so far there is not applicable end-to-end mode for adversarial generation." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.563, + 0.886, + 0.788 + ], + "angle": 0, + "content": "Another kind of approaches to improve model robustness is data augmentation, which aims to generate the similar samples to the original data for increasing the amount of training data. The traditional methods on data augmentation include backtranslation (Sennrich et al., 2015), rule-based modification (Wei and Zou, 2019), paraphrasing (Kumar et al., 2019) and other techniques (Li et al., 2022). These methods are based on the distribution of training data, which are model agnostic. However, with these data, the models are still vulnerable to adversarial samples as the attacking techniques are designed for finding the weakness of the target model (Yoo and Qi, 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.886, + 0.92 + ], + "angle": 0, + "content": "To solve the above challenges, in this paper, we designed an adversarial sample generation method for data augmentation \\(\\mathbf{A}^3\\). We propose an end-to-end adversarial sample generation model. Given an original sample, a condition generator is designed to output a condition which guides a conditioned paraphrasing model to generate a potential adversarial sample. By pretraining an adversarial dis" + }, + { + "type": "page_footnote", + "bbox": [ + 0.141, + 0.905, + 0.285, + 0.919 + ], + "angle": 0, + "content": "Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11359" + }, + { + "type": "footer", + "bbox": [ + 0.211, + 0.946, + 0.788, + 0.973 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11359-11368 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.408 + ], + "angle": 0, + "content": "criminator to be utilized as the training signal, it can learn the task-specific data characteristics for finding the potential vulnerabilities. Since the generation process doesn't require interaction with the target model, it is much efficient compared with the current search based generation methods. Besides, the generated samples can be reused for different models rather than for a specific model as the traditional way. Then we propose a weighted loss for incorporating the generated adversarial samples with the original data for the target task. We evaluate our method on several NLP tasks and compare it to both the data augmentation and adversarial training methods. Experimental results demonstrate that our approach not only improves the overall performance of the trained model but also enhances its robustness against multiple attacking techniques. Moreover, the computational time is significantly reduced compared to existing adversarial sample generation methods." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.42, + 0.275, + 0.435 + ], + "angle": 0, + "content": "2 Related works" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.446, + 0.411, + 0.461 + ], + "angle": 0, + "content": "2.1 Adversarial Sample Generation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.468, + 0.49, + 0.855 + ], + "angle": 0, + "content": "In the NLP area, the disturbance on the original sample is often adopted in adversarial sample generation due to the discrete character of text, such as word replacement. For example, TextBugger (Li et al., 2019) generates the adversarial samples by randomly replacing the words in a sample text for gathering the candidate samples that may flip the model prediction. Then it chooses the reasonable adversarial samples with the similar semantics with the original text. PWWS (Ren et al., 2019) uses synonym for word substitution with the help of WordNet. With the development of pretrained language models (PLM), it helps various NLP tasks to gain improvement on the performance. Jin et al. proposed the TextFooler (Jin et al., 2020) method for evaluating the robustness of BERT. Meanwhile, several works utilize PLM to generate the candidate adversarial texts with similar semantics to original ones (Li et al., 2020, 2021). Recently, a survey on word substitution based adversarial sample generation (Chiang and Lee, 2022) shows that despite the effectiveness of such attacks, the generated samples have flaws on fluency and syntactical correctness and can be identified easily." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.866, + 0.329, + 0.882 + ], + "angle": 0, + "content": "2.2 Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Based on the proximity assumption, researchers have explored methods to enhance the robustness" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.31 + ], + "angle": 0, + "content": "and effectiveness of trained models by introducing disturbances to the original training samples, commonly known as adversarial training (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) first proposed a NLP-oriented method that adds noise to the word embeddings of the input sample for adversarial training in text classification task. Similar approaches have been applied to adversarial training for pretrained language models (Liu et al., 2020) and other NLP tasks (Li and Qiu, 2021), showing improvements in model performance. Despite their effectiveness, disturbance in the continuous spaces lack clear meaning in actual language, making them less interpretable." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.311, + 0.885, + 0.811 + ], + "angle": 0, + "content": "Recent works in adversarial training for NLP tasks consider modifications at the lexical level to make the perturbed samples more consistent with real samples. Early works use synonym substitution to explore proximate samples in the sample space. Huang et al. (Huang et al., 2019) and Dong et al. (Dong et al., 2021) used different sample space exploration methods to find samples that effectively enhance model robustness in adversarial training. The SSMBA method proposed by Ng et al. (Ng et al., 2020) generates perturbed samples by using pretrained language models to complete corrupted texts. Recently, researchers consider using adversarial sample generation methods for adversarial training. Previous work (Carlini et al., 2019) has shown that using samples generated by a specific adversarial sample generation method for augmented training generally only improves the robustness against that method. Yoo et al. designed the A2T method (Yoo and Qi, 2021), which introduced a fast lexical substitution adversarial sample generation method into the model training process, achieving improved robustness against multiple model attacking methods. Current adversarial training methods for NLP models are mostly based on constrained searching in the sample space, and lack the ability to self learn and adapt to models or datasets. Moreover, the lexical substitution based generation methods have potential problems regarding semantic soundness and syntactic correctness. (Chiang and Lee, 2022)" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.822, + 0.858, + 0.856 + ], + "angle": 0, + "content": "3 The End-to-end Adversarial Sample Based Data Augmentation" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.866, + 0.651, + 0.88 + ], + "angle": 0, + "content": "3.1 Framework" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.888, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this paper, we design the end-to-end adversarial sample generation method for data augmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.528, + 0.941 + ], + "angle": 0, + "content": "11360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.081, + 0.488, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.214, + 0.49, + 0.245 + ], + "angle": 0, + "content": "Figure 1: The framework for the end-to-end adversarial sample based data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.271, + 0.489, + 0.431 + ], + "angle": 0, + "content": "For a given sample \\(\\mathbf{x}\\), the condition generator gives the condition \\(\\mathbf{c}\\), based on which the paraphrasing model generates \\(\\mathbf{x}'\\) that is a potential adversarial sample. The whole architecture consists of three parts, as shown in Fig.1. The first condition generator is designed to generate the appropriate conditions for guiding the generated paraphrase being an adversarial sample. The second is the pretrained conditioned paraphrasing model that generates the paraphrases satisfying the given condition." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.432, + 0.489, + 0.56 + ], + "angle": 0, + "content": "Since there are no supervised data for training the condition generator, we propose an adversarial discriminator to provide the proper training signals, similar to the structure of a generative adversarial network (GAN) (Goodfellow et al., 2014a). The third part is the adversarial discriminator that can be a classifier or regression model for the target task, which checks the flip of predication." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.562, + 0.49, + 0.625 + ], + "angle": 0, + "content": "Then for the target task, the generated adversarial samples are incorporated with the original data for data augmentation, which are used to train a model for this task." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.64, + 0.327, + 0.654 + ], + "angle": 0, + "content": "3.2 Condition Generator" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.663, + 0.489, + 0.727 + ], + "angle": 0, + "content": "The condition generator takes an original sample as input and aims to generate a condition \\( \\mathbf{c} \\) for guiding the paraphrasing model to generate an adversarial sample. Let \\( G \\) denote the condition generator." + }, + { + "type": "equation", + "bbox": [ + 0.261, + 0.742, + 0.488, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = G (\\mathbf {x}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.775, + 0.491, + 0.919 + ], + "angle": 0, + "content": "There are many candidate forms of condition \\(\\mathbf{c}\\), such as the ratio of changed words, the sentence structure or the sentiment. Following the normal way of generating adversarial samples, we adopt the lexical substitution indicator as the condition, which is in the form of binary sequence corresponding to the words in \\(\\mathbf{x}\\), denoted by \\(\\mathbf{c} \\in \\{0,1\\}^{|\\mathbf{x}|}\\). \\(c_{i} = 1\\) indicates \\(x_{i}\\) should be changed in the paraphrase \\(\\mathbf{x}'\\), \\(c_{i} = 0\\) for otherwise." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.295 + ], + "angle": 0, + "content": "In this paper we adopt a multi-layer bidirectional recurrent neural network as the basic network. The output hidden vector at each step is projected to a scalar. The output of the condition generator is fed into the conditioned paraphrasing model. Since the condition \\( \\mathbf{c} \\) is defined as a sequence of binary value, while conventional activation function such as sigmoid outputs a contiguous value between 0 and 1. To maintain a reasonable form of \\( \\mathbf{c} \\) and achieve a stable training, we adopt the straight-through estimator (Bengio et al., 2013) method as the activation function, such that the output of the condition generator will be binary." + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.307, + 0.884, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = S T E (G (\\mathbf {x})) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.337, + 0.883, + 0.367 + ], + "angle": 0, + "content": "3.3 Pretraining the Conditioned Paraphrasing Model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.374, + 0.884, + 0.487 + ], + "angle": 0, + "content": "The general purpose of paraphrasing is to generate a sample with the same meaning but different expression with the given text. Here we introduce the conditioned paraphrasing model for guiding the paraphrase. For a given text \\(\\mathbf{x}\\) and a condition \\(\\mathbf{c}\\), it generates a paraphrase \\(\\mathbf{x}'\\) with the similar semantics to \\(\\mathbf{x}\\) and satisfying \\(\\mathbf{c}\\), formally:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.499, + 0.883, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.53, + 0.884, + 0.738 + ], + "angle": 0, + "content": "We adopt a multi-layer bidirectional recurrent neural network for the conditioned paraphrasing model. The pretraining data are automatically generated, where the input is a pair of a text \\( \\mathbf{x} \\) and the randomly selected positions as the condition c. As the paraphrase \\( \\mathbf{x}' \\), there are many candidate methods for choosing the replacement words, such as the synonym vocabulary or the masked language model. To ensure the semantic similarity and the syntactical correctness, we adopt the word com- posable knowledge (Liu et al., 2023) for choosing the words to form a new text \\( \\mathbf{x}' \\) as the supervised target." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.739, + 0.884, + 0.819 + ], + "angle": 0, + "content": "Since only a small proportion of the text are changed, the simple negative log-likelihood loss would lead the model to just copy the input. Thus, we apply a larger weight \\( \\omega_{p} > 0 \\) for the replacement positions \\( c_{i} = 1 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.831, + 0.883, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\ell_ {P} = - \\sum_ {i = 1} ^ {| \\mathbf {x} |} \\left(1 + c _ {i} \\omega_ {p}\\right) \\log p \\left(\\hat {x} _ {i} ^ {\\prime} = x _ {i} ^ {\\prime} | \\mathbf {x}, \\mathbf {c}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.887, + 0.883, + 0.919 + ], + "angle": 0, + "content": "where \\(\\hat{x}_i^{\\prime}\\) is the output of the model and \\(x_{i}^{\\prime}\\) is the supervised paraphrase." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "11361" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.085, + 0.37, + 0.099 + ], + "angle": 0, + "content": "3.4 Adversarial Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.106, + 0.49, + 0.185 + ], + "angle": 0, + "content": "To train the end-to-end adversarial sample generation model, we pretrain an adversarial discriminator to provide training signals, denote by \\(D\\). It is a trained classifier or a regression model for the target task." + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.187, + 0.488, + 0.203 + ], + "angle": 0, + "content": "\\[\ny = D (\\mathbf {x}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.209, + 0.489, + 0.29 + ], + "angle": 0, + "content": "Theoretically the paraphrasing model should generate the semantically similar text with the original one. If the prediction \\(\\hat{y}^\\prime\\) by \\(D\\) for the generated sample \\(x^{\\prime}\\) is flipped compared to the ground truth \\(y\\), \\(x^{\\prime}\\) is a potentially adversarial sample." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.3, + 0.415, + 0.332 + ], + "angle": 0, + "content": "3.5 End-to-End Adversarial Sample Generation Training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.337, + 0.489, + 0.464 + ], + "angle": 0, + "content": "The core of the adversarial sample generation model is to generate the appropriate condition for paraphrasing. Since the discriminator can be seen as a white-box victim model, we can pass the gradient through the neural network structure by utilizing the tricks like Gumbel softmax (Jang et al., 2017). This allows an end-to-end training mode for the whole adversarial sample generation method." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.466, + 0.49, + 0.61 + ], + "angle": 0, + "content": "During the training, the condition generator, the conditioned paraphrasing model and the adversarial discriminator are connected in a pipeline way. For a given sample \\(\\mathbf{x}\\), the pipeline outputs a prediction \\(\\hat{y}\\) for a perturbed sample. The conditioned paraphrasing model \\(F(\\mathbf{x},\\mathbf{c})\\) outputs a sequence of word probabilities. We adopt Gumbel softmax to convert the probabilities to word embeddings that can be directly used by \\(D\\)." + }, + { + "type": "equation", + "bbox": [ + 0.257, + 0.621, + 0.334, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = G (\\mathbf {x})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.639, + 0.487, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.258, + 0.659, + 0.34, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = D (\\mathbf {x} ^ {\\prime})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.686, + 0.489, + 0.814 + ], + "angle": 0, + "content": "During training, the parameters of the condition generator and the conditioned paraphrasing model are tuned while the adversarial discriminator remains frozen. Training the condition paraphrasing model allows the model to be trained adaptively to generate words more suitable for the target task. To train the adversarial sample generation model, we used the following losses." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.823, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Adversarial loss is the main goal of the model, namely to generate a sample that can make the target model flip its prediction. Since previous works on model distilling show the effectiveness and simplicity of the MSE loss function (Hinton et al., 2015), thus it is adopted in our work. More" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.085, + 0.882, + 0.116 + ], + "angle": 0, + "content": "advanced distances between probabilities such as KL-distance can be an alternative." + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.129, + 0.883, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\ell_ {a d v} = 1 - \\left| D \\left(\\mathbf {x} ^ {\\prime}\\right) - D (\\mathbf {x}) \\right| ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.161, + 0.884, + 0.273 + ], + "angle": 0, + "content": "Perturbation ratio loss To maintain the similar semantics between the perturbed sample and the original sample, the ratio of perturbed words in text are constrained, namely the number of \\( c_{i} = 1 \\) cases in the generated condition \\( \\mathbf{c} \\). Let \\( \\rho \\) denote the expected ratio, \\( L \\) is the length of sample, the perturbation ratio loss is calculated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.285, + 0.883, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c o n} = \\left| \\rho - \\| \\mathbf {c} \\| _ {1} / L \\right| ^ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.317, + 0.884, + 0.446 + ], + "angle": 0, + "content": "Conditioned paraphrasing losses The conditioned paraphrasing model is fine-tuned for different expectation by a few loss functions. The first is the reconstruction loss for the words not perturbed in \\(\\mathbf{c}\\), i.e. on the positions \\(c_{i} = 0\\), denoted by \\(\\mathbf{c}(0)\\). It is a negative likelihood loss for the original words such that the paraphrasing model is trained to reconstruct the original input." + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.458, + 0.883, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\ell_ {r e c} = - \\sum_ {i \\in \\mathbf {c} (0)} \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.505, + 0.884, + 0.617 + ], + "angle": 0, + "content": "For the perturbed positions, the probability of the original word should be lower than any other words. Inspired by the loss function for the negative samples in the noise contrastive estimation (Gutmann and Hyvarinen, 2010), we design the loss of punishing repeating the original word. Let \\( \\mathbf{c}(1) \\) denote the positions of the 1s in the condition \\( \\mathbf{c} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.63, + 0.883, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\ell_ {p u n} = - \\sum_ {i \\in \\mathbf {c} (1)} \\sigma (- \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right)) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.687, + 0.882, + 0.782 + ], + "angle": 0, + "content": "Semantics similarity and syntactic correctness losses To maintain the semantics of the generated paraphrase similar with the original sample, we adopt a word-level similarity loss based on cosine similarities for the word embeddings between the generated word and the original word." + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.794, + 0.757, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\ell_ {s e m} = 1 - \\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.816, + 0.882, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) = \\frac {1}{| \\mathbf {x} |} \\sum_ {i = 1} ^ {| x |} \\left(1 - \\cos \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i} ^ {\\prime}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.872, + 0.884, + 0.919 + ], + "angle": 0, + "content": "The word embeddings similarity cannot guarantee a proper syntactical role of the replacement word. Thus, we also adopt the word composable" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11362" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.115, + 0.083, + 0.486, + 0.116 + ], + "angle": 0, + "content": "Algorithm 1 The Training Process for End-to-End Adversarial Sample Generator" + }, + { + "type": "code_caption", + "bbox": [ + 0.117, + 0.12, + 0.292, + 0.135 + ], + "angle": 0, + "content": "Input: Training set \\(D_{t}\\)" + }, + { + "type": "code", + "bbox": [ + 0.117, + 0.136, + 0.489, + 0.168 + ], + "angle": 0, + "content": "Output: Condition generator \\(G\\), conditioned paraphrasing model \\(F\\)" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.17, + 0.468, + 0.184 + ], + "angle": 0, + "content": "1: //Pretrain conditioned paraphrasing model" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.185, + 0.486, + 0.215 + ], + "angle": 0, + "content": "2: Generate paraphrasing training data \\(D_{p}\\) based on \\(D_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.217, + 0.427, + 0.232 + ], + "angle": 0, + "content": "3: Pretrain \\(F\\) with \\(D_{p}\\) according to Eq.4" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.233, + 0.416, + 0.247 + ], + "angle": 0, + "content": "4: //Pretrain adversarial discriminator" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.249, + 0.292, + 0.263 + ], + "angle": 0, + "content": "5: Pretrain \\(D\\) with \\(D_{t}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.265, + 0.489, + 0.295 + ], + "angle": 0, + "content": "6: //Train the end-to-end adversarial sample generation model" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.297, + 0.352, + 0.312 + ], + "angle": 0, + "content": "7: Freeze the parameters of \\( D \\)" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.314, + 0.333, + 0.329 + ], + "angle": 0, + "content": "8: for sample \\(\\mathbf{x}, y\\) in \\(D_{t}\\) do" + }, + { + "type": "text", + "bbox": [ + 0.127, + 0.33, + 0.28, + 0.344 + ], + "angle": 0, + "content": "9: Let \\(\\mathbf{c} = G(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.346, + 0.417, + 0.361 + ], + "angle": 0, + "content": "10: Calculate \\(\\ell_{con}\\) according to Eq.8" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.362, + 0.304, + 0.377 + ], + "angle": 0, + "content": "11: Let \\(\\mathbf{x}' = F(\\mathbf{x},\\mathbf{c})\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.378, + 0.486, + 0.408 + ], + "angle": 0, + "content": "12: Calculate \\(\\ell_{rec},\\ell_{pun},\\ell_{sem},\\ell_{syn}\\) according to Eq.9-12" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.41, + 0.286, + 0.425 + ], + "angle": 0, + "content": "13: Let \\(\\hat{y} = D(\\mathbf{x}^{\\prime})\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.426, + 0.417, + 0.441 + ], + "angle": 0, + "content": "14: Calculate \\(\\ell_{adv}\\) according to Eq.7" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.443, + 0.431, + 0.458 + ], + "angle": 0, + "content": "15: Calculate \\(L_{adv}\\) according to Eq.13" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.459, + 0.408, + 0.473 + ], + "angle": 0, + "content": "16: Update \\(G,F\\) according to \\(L_{adv}\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.475, + 0.207, + 0.488 + ], + "angle": 0, + "content": "17: end for" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.49, + 0.486, + 0.522 + ], + "angle": 0, + "content": "18: return Condition generator \\( G \\), conditioned paraphrasing model \\( F \\)" + }, + { + "type": "list", + "bbox": [ + 0.121, + 0.17, + 0.489, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.552, + 0.49, + 0.745 + ], + "angle": 0, + "content": "knowledge based loss (Liu et al., 2023) to provide syntactical correctness guidance for the generated sample. Given a dependency relationship between two words \\(< u, r, v >\\), the word composable knowledge estimates the syntactical soundness of this combination, denoted by \\(p_s(u, r, v)\\). It relies on the pretrained word embeddings and thus is easy to integrate with other neural network structures. Let \\(s_x = (i, r, j)\\) denote the dependency parsing tree of \\(\\mathbf{x}\\), where \\((i, r, j)\\) represents the dependency relation between \\(\\mathbf{x}_j\\) and its head word \\(\\mathbf{x}_i\\) is \\(r\\). The syntactic correctness loss is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.758, + 0.488, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\ell_ {s y n} = - \\sum_ {(i, r, j) \\in s _ {x}} \\log p _ {s} \\left(\\mathbf {x} _ {i} ^ {\\prime}, r, \\mathbf {x} _ {j} ^ {\\prime}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.489, + 0.855 + ], + "angle": 0, + "content": "The complete loss function for training the adaptive adversarial sample generation model is the weighted sum of the above losses." + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.869, + 0.487, + 0.92 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {a d v} = w _ {a} \\ell_ {a d v} + w _ {c} \\ell_ {c o n} + w _ {r} \\ell_ {r e c} + w _ {p} \\ell_ {p u n} \\\\ + w _ {s e m} \\ell_ {s e m} + w _ {s y n} \\ell_ {s y n} \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.082, + 0.895, + 0.151 + ], + "angle": 0, + "content": "
Dataset#Class#Train#TestAve.Len.
IMDB22500025000270.7
Quora23636604040425.33
SNLI35501521000022.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.608, + 0.161, + 0.784, + 0.174 + ], + "angle": 0, + "content": "Table 1: Dataset statistics" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.2, + 0.882, + 0.232 + ], + "angle": 0, + "content": "The overall training process of the end-to-end adversarial sample generation is listed in Alg.1." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.242, + 0.71, + 0.258 + ], + "angle": 0, + "content": "3.6 Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.262, + 0.884, + 0.456 + ], + "angle": 0, + "content": "Having the above end-to-end adversarial sample generator, we can apply data augmentation on the training data for the target task. For the given original training set \\( D_{t} \\), the corresponding generated adversarial samples is denoted by \\( D_{t}^{\\prime} \\). Then the model for the target task can be trained on \\( D_{t} \\cup D_{t}^{\\prime} \\) for improving the model robustness. Since the adversarial samples are more likely to introduce noises, their training loss is assigned a lower weight. Let \\( L_{t} \\) denotes the loss of the target task, \\( \\lambda_{i} \\) denotes the loss weight for sample \\( s_{i} \\). \\( \\lambda_{i} = 1 \\) for an original sample while \\( \\lambda_{i} < 1 \\) for a generated sample." + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.465, + 0.884, + 0.501 + ], + "angle": 0, + "content": "\\[\nL _ {t a s k} = \\sum_ {s _ {i} \\in D _ {t} \\bigcup D _ {t} ^ {\\prime}} \\lambda_ {i} L _ {t} (s _ {i}) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.51, + 0.884, + 0.558 + ], + "angle": 0, + "content": "The previous work has shown the adversarial samples generated based on one model are potentially adversarial to other models (Szegedy et al., 2013)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.569, + 0.656, + 0.586 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.595, + 0.884, + 0.642 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the proposed method, four aspects should be considered and evaluated in the subsections 4.2-4.5." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.651, + 0.884, + 0.684 + ], + "angle": 0, + "content": "A1: To check the attacking effectiveness of generated adversarial sample." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.692, + 0.882, + 0.724 + ], + "angle": 0, + "content": "A2: To check whether the model performance is improved after augmented training." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.732, + 0.882, + 0.765 + ], + "angle": 0, + "content": "A3: To check whether the robustness of the target model is improved." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.773, + 0.843, + 0.789 + ], + "angle": 0, + "content": "A4: To check the efficiency of our method." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.651, + 0.884, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.8, + 0.706, + 0.816 + ], + "angle": 0, + "content": "4.1 Experiment Setups" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.82, + 0.615, + 0.834 + ], + "angle": 0, + "content": "4.1.1 Tasks" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.839, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Three tasks are considered as evaluation tasks to cover a wider range of NLP tasks such that the effectiveness of data augmentation can be thoroughly evaluated in different scenarios. The first is a sentiment analysis task on the dataset IMDB. The IMDB" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "11363" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.082, + 0.875, + 0.184 + ], + "angle": 0, + "content": "
Base ModelBERTDistilBERTPerturbed %Bert Score
DatasetOri.Acc.Att.Acc.ΔOri.Acc.Att.Acc.Δ
IMDB93.9%66.58%-27.32%92.96%61.76%-31.2%9.95%95.98
Quora84.84%75.95%-8.89%84.32%76.44%-7.88%16.6%95.96
SNLI88.95%55.64%-33.31%88.04%55.20%-32.84%14.73%95.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.194, + 0.848, + 0.209 + ], + "angle": 0, + "content": "Table 2: Accuracy comparison between accuracies on the original testing set and adversarial testing set." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.234, + 0.49, + 0.41 + ], + "angle": 0, + "content": "dataset consists of long movie reviews which are classified as positive of negative sentiment polarity. The second is a text classification task on the dataset Quora Question Pairs(QQP), which consists of question pairs from the Quora site classified as duplicate or not. The last is a natural language inference task on the dataset SNLI. Each sample in SNLI consists a hypothesis sentence and a premise sentence, the logical relationship between them is determined as entailment, neutral or contradiction. The statistics of the datasets are listed in Table.1." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.421, + 0.275, + 0.435 + ], + "angle": 0, + "content": "4.1.2 Base Models" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.44, + 0.49, + 0.681 + ], + "angle": 0, + "content": "Recently, pretrained language models are widely used as a fundamental part for NLP methods. In the following experiments, BERT (Kenton and Toutanova, 2019) and DistilBERT (Sanh et al., 2019) are chosen as the base model for the tasks. BERT is a widely used pretrained language model, and achieves improvements on various NLP tasks. DistilBERT is a compressed version of BERT model based on model distilling technique, making it easier to be used on devices with lower computation power. Both models will be finetuned on the original train set or the augmented train set for the target task and evaluated afterwards. Their results will be listed as Baseline in the following experiments." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.691, + 0.303, + 0.707 + ], + "angle": 0, + "content": "4.1.3 Implementation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.71, + 0.49, + 0.919 + ], + "angle": 0, + "content": "All the experiments are conducted on a single NVIDIA Geforce RTX 2080Ti. Our method is implemented with PyTorch 1.12.0. The code for our proposed method is shared at https://splab.sdu.edu.cn/xscg/sjjydm.htm. The compared methods and adversarial sample generation are implemented by TextAttack (Morris et al., 2020b). The size of hidden vectors in our model is 300. We adopt GloVe (Pennington et al., 2014) as the word embeddings and reused the pretrained parameters of word composable knowledge released in SynPara (Liu et al., 2023). For pretraining the paraphrasing model, the ratio of replaced words is set" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.231, + 0.88, + 0.396 + ], + "angle": 0, + "content": "
Base ModelMethodDataset
IMDBQuoraSNLI
BERTBaseline93.984.8488.95
SSMBA-0.41+3.03-2.88
A2T-0.03+1.4-0.5
A3+1+4+0.1
Distil-BERTBaseline91.8784.3388.04
SSMBA+1.09+3.79-0.12
A2T+1.27+4.39+1.38
A3+1.93+4.89+0.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.406, + 0.884, + 0.45 + ], + "angle": 0, + "content": "Table 3: Accuracy comparison on the original testing set. The values for the augmentation methods are the differences with the accuracy of the base model." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.474, + 0.884, + 0.587 + ], + "angle": 0, + "content": "\\(15\\%\\) when generating training data. For training the complete model, the weights in the final loss function are set as \\(w_{a} = 5.0\\), \\(w_{c} = 3.0\\), \\(w_{r} = 2.0\\), \\(w_{p} = 5.0\\), \\(w_{sem} = 2.0\\), \\(w_{syn} = 0.1\\) in most cases. The ratio of perturbed words \\(\\rho\\) for the condition generator is set as 0.15 for SNLI and 0.1 for IMDB and Quora unless stated otherwise." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.596, + 0.725, + 0.611 + ], + "angle": 0, + "content": "4.1.4 Compared methods" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.615, + 0.884, + 0.808 + ], + "angle": 0, + "content": "For the comparison methods, we consider the following methods. SSMBA (Ng et al., 2020) is a data augmentation method that uses the corruption function similar to the masked language model training and reconstructs the samples with BERT. A2T (Yoo and Qi, 2021) is an adversarial training method, where a fast adversarial sample generation method uses the gradients for identifying important words in a sample and replaces them based on word embeddings or MLM models. It iteratively trains the task model and augment the training data set with the generated adversarial samples." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.819, + 0.718, + 0.833 + ], + "angle": 0, + "content": "4.2 Attack Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.84, + 0.884, + 0.919 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the generated adversarial samples, we compared the performances of the baseline models on the original testing data and on the adversarial data. The results are shown in Table.2, where the ratio of perturbed words in" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11364" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.081, + 0.885, + 0.378 + ], + "angle": 0, + "content": "
Base ModelBERTDistilBERT
DatasetsIMDBQuoraSNLIIMDBQuoraSNLI
Attack MethodAugment Method%Δ%Δ%Δ%Δ%Δ%Δ
A3Baseline69.377.358.364.180.159.8
SSMBA71.6+2.377.7+0.458.7+0.461.6-2.578.8-1.362.9+3.1
A2T73.7+4.476.9-0.459.7+1.466.1+278.6-1.560.3+0.5
A392.2+22.989.3+1286.2+27.993.8+29.790.9+10.886.5+26.7
A2TBaseline38.840.144.130.650.341.1
SSMBA38.9+148.1+849.3+5.227.9-2.750.7+0.445.1+4
A2T44+5.276.2+36.179.4+35.354.5+23.976.1+25.875.6+34.5
A345.3+6.560.7+20.648+3.936.1+5.560.2+9.943.3+2.2
Text-FooterBaseline1.723.60.90.110.30.47
SSMBA5.3+3.612.5-11.10.4-0.50-0.112.1+1.80.73+0.26
A2T0.8-0.914.3-9.35.1+4.21.3+1.28-2.33.93+3.46
A312+10.327.1+3.51.9+11.6+1.515.2+4.90.67+0.2
BAEBaseline3139.322.124.939.323.87
SSMBA30.6-0.439-0.325+2.921-3.939.7+0.423.93+0.06
A2T29.1-1.943.4+4.123.5+1.426.3+1.444.2+4.925.5+1.63
A337.9+6.940.8+1.527.6+5.527.7+2.845.7+6.424.27+0.4
PWWSBaseline125.11.20.822.11.07
SSMBA2.2+1.223.1-22.3+1.10.1-0.722.9+0.81.2+0.13
A2T0.3-0.726.3+1.20.7-0.50.6-0.223+0.90.87-0.2
A37.3+6.327.7+2.62.4+1.20.9+0.124.1+22.4+1.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.387, + 0.884, + 0.418 + ], + "angle": 0, + "content": "Table 4: Comparison of attacking methods against augmentation methods. Column % denotes the model accuracy under attacking. Column \\(\\Delta\\) denotes the difference between augmented model with the original model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.441, + 0.49, + 0.618 + ], + "angle": 0, + "content": "the adversarial samples is listed in column Perturbed%. A lower ratio indicates more structure information of the text preserved. BertScore (Zhang et al., 2020) is used for quantifying the semantics similarity between the original sample and the adversarial sample. The higher the BertScore, the closer their semantics. We can see that the model performances of both baseline models drop significantly on the adversarial samples, which show that the generated adversarial samples are effective on attacking the baseline models." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.637, + 0.373, + 0.652 + ], + "angle": 0, + "content": "4.3 Performance Improvement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.662, + 0.49, + 0.919 + ], + "angle": 0, + "content": "We then evaluated the model performance on the target tasks before and after training with augmented data. By using our method and SSMBA, the data is augmented to doubled size of the original training set, respectively. The augmented dataset is used for training the base models. For A2T, it iteratively trains the model and then augments the training set by generating adversarial samples. The results in Table.3 show that our method improves the performance on all settings and are better than the compared methods in most cases. The compared methods are unstable on performances, namely in some cases performances are the improved while some are not. For example, SSMBA has negative impact on SNLI dataset for both base models. This is probably because SSMBA doesn't take into con" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.441, + 0.885, + 0.57 + ], + "angle": 0, + "content": "sideration the task-specific characteristics, such as the logical connections between the hypothesis and premise in SNLI. It should be also noted that while the original performance of DistilBERT is slightly lower than BERT, it improves a lot after data augmentation, which confirmed previous studies regarding the effectiveness of data augmentation for distilled models (Das et al., 2020)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.589, + 0.757, + 0.604 + ], + "angle": 0, + "content": "4.4 Robustness Improvement" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.885, + 0.919 + ], + "angle": 0, + "content": "To evaluate the robustness of base model against different attacking methods, we choose three representative attacking methods: TextFooler (Jin et al., 2020), BAE (Garg and Ramakrishnan, 2020) and PWWS (Ren et al., 2019). We adopt the notion \\( m1 - m2 \\) denoting the attacking method \\( m1 \\) against augmentation method \\( m2 \\). For example, BAE-A2T means using BAE to attack a model trained using A2T. Similar to the settings of previous works (Yoo and Qi, 2021), 1000 samples are selected from the original testing set of each dataset (1500 samples from SNLI for three labels), where the label distributions are balanced. To better show the impact of an attacking method, the selected samples can be predicted correctly by the base models before data augmentation. The attack methods generate the adversarial samples based on the selected samples and the accuracy on these adversarial samples of the trained model would be compared. A2T and" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11365" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.147, + 0.082, + 0.457, + 0.166 + ], + "angle": 0, + "content": "
MethodsDataset
IMDBQuoraSNLI
SSMBA10m38m1h2m
A2T36h(3)111h(2)70h(2)
A34m27m21m
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.176, + 0.489, + 0.233 + ], + "angle": 0, + "content": "Table 5: Time for augmenting the dataset. It should be noted that the time for A2T is the cumulated time for multiple training epochs (number of epochs is listed in the brackets)." + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.246, + 0.459, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.395, + 0.487, + 0.424 + ], + "angle": 0, + "content": "Figure 2: The influence of \\(\\rho\\) on the perturbed ratio and the attack effectiveness (denoted by \\(\\Delta \\mathrm{Acc}\\))" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.451, + 0.486, + 0.481 + ], + "angle": 0, + "content": "our proposed method are tested as they are also adversarial sample generation methods." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.483, + 0.49, + 0.692 + ], + "angle": 0, + "content": "The results are listed in Table.4, where the column \\(\\%\\) denotes the model accuracy under attacking and column \\(\\Delta\\) shows the difference between the performance of the augmented model with the baseline. As explored in a previous research (Carlini et al., 2019), the robustness against a certain attack method will be improved if trained with the corresponding adversarial samples. From the row \\(\\mathrm{A}^3\\mathrm{-A}^3\\) and A2T-A2T, we can see that both methods significantly improve the model robustness. For other representative attacking methods, the augmented models with our generated data show better robustness than the compared methods in most settings." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.705, + 0.34, + 0.721 + ], + "angle": 0, + "content": "4.5 Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.727, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Another vital aspect of adversarial training is the efficiency. Most adversarial sample generation methods involves constantly interaction with the target model. Such process is both computational heavy and time-consuming. We recorded the time used for augmenting the training set of different methods. Both SSMBA and our proposed method are one-time augmentation thus the time taken is similar. Our method is slightly faster than SSMBA since SSMBA uses BERT as the reconstruction model. The A2T method, on the other hand, is significantly slower than our method due to two possible rea" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.23 + ], + "angle": 0, + "content": "son. First reason is that the attacking process is conducted multiple times. Secondly, the method is based on attacking the target model which is not only time-consuming, but the time taken will also increase over time since the model is trained better after each epoch. Comparably, our proposed method generates model-agnostic adversarial samples, such that the time for augmented training is significantly shortened." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.244, + 0.712, + 0.259 + ], + "angle": 0, + "content": "4.6 Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.266, + 0.885, + 0.539 + ], + "angle": 0, + "content": "We analyze the effect of parameter \\(\\rho\\), the targeted perturbed ratio. In our method for a \\(\\rho\\) setting, the actual perturbed ratio in the generated adversarial samples may be different with the initial value. Thus, we analyze the actual perturbed ratio and its impact on the attacking effectiveness, shown in Fig.2. The blue and orange lines represent the set \\(\\rho\\) and the actual perturbed ratio respectively, corresponding to the left Y-axis. The \\(\\Delta \\mathrm{Acc}\\) is the accuracy difference after the attack, corresponding to the right Y-axis. Higher value of \\(\\Delta \\mathrm{Acc}\\) means more effective attack. We can see that the actual perturbed ratio approaches the \\(\\rho\\) setting. A higher \\(\\rho\\) may cause the predicated label flipped but may lead to a larger semantics difference. Thus, the parameter \\(\\rho\\) should be set properly to achieve a balanced performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.554, + 0.642, + 0.569 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.581, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In this paper, we present a method named \\(\\mathrm{A}^3\\) that combines the advantages of adversarial training and data augmentation to enhance the robustness of NLP models. We introduce an end-to-end adversarial sample generation model which consists of a condition generator and a conditioned paraphrasing model. A pretrained adversarial discriminator is proposed to train the adversarial sample generation model for capturing the task-specific data characteristics. This method is much efficient without the need for extensive model interactions. The generated adversarial samples are incorporated with the original training samples by a weighted loss for augmented training downstream models. Experiments are conducted on multiple NLP tasks and demonstrate the effectiveness and efficiency of our proposed method. Compared to existing methods, the performances of the base models have improved more after our data augmentation, as well as their robustness enhanced against multiple attacking techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11366" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.1 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.111, + 0.493, + 0.353 + ], + "angle": 0, + "content": "Firstly, the current form of condition limits the diversity and fine-grained control over the generation process. For example, a condition can be a subtree from the dependency syntax structure, which allows paraphrases with potential structure reform and correct syntax (Liu et al., 2023). Secondly, a condition may also contain the direction of replacement word for the paraphrasing model. This would allow reusing the same conditioned paraphrasing model across different tasks, thereby reducing the training overhead for adversarial sample generation. Lastly, in future research, we aim to explore the feasibility of directly using the gradient from the target model for training the adversarial sample generation model." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.367, + 0.279, + 0.384 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.394, + 0.49, + 0.459 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (62376138) and the Innovative Development Joint Fund Key Projects of Shandong NSF (ZR2022LZH007)." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.487, + 0.214, + 0.503 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.511, + 0.49, + 0.566 + ], + "angle": 0, + "content": "Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.578, + 0.49, + 0.644 + ], + "angle": 0, + "content": "Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. 2019. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.656, + 0.49, + 0.735 + ], + "angle": 0, + "content": "Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan Zhang, and Cho-Jui Hsieh. 2020. Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3601-3608." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.748, + 0.489, + 0.788 + ], + "angle": 0, + "content": "Cheng-Han Chiang and Hung-yi Lee. 2022. How far are we from real synonym substitution attacks? arXiv preprint arXiv:2210.02844." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.8, + 0.489, + 0.853 + ], + "angle": 0, + "content": "Deepan Das, Haley Massa, Abhimanyu Kulkarni, and Theodoros Rekatsinas. 2020. An empirical analysis of the impact of data augmentation on knowledge distillation. CoRR, abs/2006.03810." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.49, + 0.918 + ], + "angle": 0, + "content": "Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. 2021. Towards robustness against natural language word substitutions. arXiv preprint arXiv:2107.13541." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.511, + 0.49, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.139 + ], + "angle": 0, + "content": "Javid Ebrahimi, Anyi Rao, Daniel Lowd, and De- jing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.885, + 0.216 + ], + "angle": 0, + "content": "Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6174-6181." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.227, + 0.885, + 0.279 + ], + "angle": 0, + "content": "Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. 2014a. Generative adversarial networks. CoRR, abs/1406.2661." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.29, + 0.885, + 0.331 + ], + "angle": 0, + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014b. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.341, + 0.885, + 0.394 + ], + "angle": 0, + "content": "Shreya Goyal, Sumanth Doddapaneni, Mitesh M Khapra, and Balaraman Ravindran. 2022. A survey of adversarial defences and robustness in nlp. ACM Computing Surveys." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.405, + 0.885, + 0.485 + ], + "angle": 0, + "content": "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297-304. JMLR Workshop and Conference Proceedings." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.495, + 0.885, + 0.535 + ], + "angle": 0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.546, + 0.885, + 0.664 + ], + "angle": 0, + "content": "Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, and Pushmeet Kohli. 2019. Achieving verified robustness to symbol substitutions via interval bound propagation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4083-4093." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.674, + 0.885, + 0.728 + ], + "angle": 0, + "content": "Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017). OpenReview.net." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.738, + 0.885, + 0.817 + ], + "angle": 0, + "content": "Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.828, + 0.885, + 0.881 + ], + "angle": 0, + "content": "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.892, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Ashutosh Kumar, Satwik Bhattamishra, Manik Bhandari, and Partha Talukdar. 2019. Submodular" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.885, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "11367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.166 + ], + "angle": 0, + "content": "optimization-based diverse paraphrasing and its effectiveness in data augmentation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3609-3619." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.177, + 0.489, + 0.218 + ], + "angle": 0, + "content": "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open, 3:71-90." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.229, + 0.489, + 0.321 + ], + "angle": 0, + "content": "Dianqi Li, Yizhe Zhang, Hao Peng, Liquin Chen, Chris Brockett, Ming-Ting Sun, and William B Dolan. 2021. Contextualized perturbation for textual adversarial attack. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5053-5069." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.333, + 0.489, + 0.386 + ], + "angle": 0, + "content": "J Li, S Ji, T Du, B Li, and T Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.398, + 0.489, + 0.464 + ], + "angle": 0, + "content": "Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020. Bert-attack: Adversarial attack against using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6193–6202." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.476, + 0.489, + 0.529 + ], + "angle": 0, + "content": "Linyang Li and Xipeng Qiu. 2021. Token-aware virtual adversarial training in natural language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8410-8418." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.541, + 0.489, + 0.606 + ], + "angle": 0, + "content": "Tianyuan Liu, Yuqing Sun, Jiaqi Wu, Xu Xi, Yuchen Han, Cheng Li, and Bin Gong. 2023. Unsupervised paraphrasing under syntax knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 37." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.618, + 0.489, + 0.672 + ], + "angle": 0, + "content": "Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.684, + 0.489, + 0.736 + ], + "angle": 0, + "content": "Takeru Miyato, Andrew M Dai, and Ian Goodfellow. 2016. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.748, + 0.489, + 0.815 + ], + "angle": 0, + "content": "John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, and Yanjun Qi. 2020a. Reevaluating adversarial examples in natural language. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3829-3839." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.489, + 0.919 + ], + "angle": 0, + "content": "John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020b. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.166 + ], + "angle": 0, + "content": "Nathan Ng, Kyunghyun Cho, and Marzyeh Ghassemi. 2020. Ssmba: Self-supervised manifold based data augmentation for improving out-of-domain robustness. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1268-1283." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.175, + 0.883, + 0.254 + ], + "angle": 0, + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.263, + 0.883, + 0.342 + ], + "angle": 0, + "content": "Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. 2019. Generating natural language adversarial examples through probability weighted word saliency. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 1085-1097." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.352, + 0.883, + 0.405 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.415, + 0.883, + 0.468 + ], + "angle": 0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.477, + 0.883, + 0.531 + ], + "angle": 0, + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.54, + 0.883, + 0.581 + ], + "angle": 0, + "content": "Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.59, + 0.883, + 0.643 + ], + "angle": 0, + "content": "Jin Yong Yoo and Yanjun Qi. 2021. Towards improving adversarial training of nlp models. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 945-956." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.653, + 0.883, + 0.731 + ], + "angle": 0, + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "11368" + } + ] +] \ No newline at end of file diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_origin.pdf b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e60c091219a82dd1c76699d25f57491b9f0d8ec5 --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/e8d461ca-bdc3-491b-9c8e-624c2e748899_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02164a89c0e243c02226010550d3f2848efcfed7f2ed1fe632074f284fce288c +size 391276 diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/full.md b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25eb57c0cd77a2245852cd450b35bb8544456f66 --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/full.md @@ -0,0 +1,345 @@ +# End-to-end Adversarial Sample Generation for Data Augmentation + +Tianyuan Liu and Yuqing Sun* + +Shandong University + +zodiacg@foxmail.com sun_yuqing@sdu.edu.cn + +# Abstract + +Adversarial samples pose a significant challenge to neural inference models. In this paper, we propose a novel enhancing approach $\mathbf{A}^3$ for the robustness of the neural NLP models, which combines the adversarial training and data augmentation. We propose an adversarial sample generator that consists of a conditioned paraphrasing model and a condition generator. The latter aims to generate conditions which guides the paraphrasing model to generate adversarial samples. A pretrained discriminator is introduced to help the adversarial sample generator adapt to the data characteristics for different tasks. We adopt a weighted loss to incorporate the generated adversarial samples with the original samples for augmented training. Compared to existing methods, our approach is much efficient since the generation process is independent to the target model and the generated samples are reusable for different models. Experimental results on several tasks show that our approach improves the overall performance of the trained model. Specially, the enhanced model is robust for various attacking techniques. + +# 1 Introduction + +Adversarial samples refer to the inputs that intentionally designed to fool deep learning models (Morris et al., 2020a), which are also called the attack samples (Szegedy et al., 2013). The methods for generating attack samples have deceived many natural language processing (NLP) models, such as text classification (Ebrahimi et al., 2017; Ren et al., 2019), machine translation (Cheng et al., 2020) and etc (Goyal et al., 2022). Along with the application of neural models in various areas, the adversarial samples have drawn much attention. + +Generally, the adversarial samples are generated by perturbations on the original sample. To enhance the robustness of NLP neural methods, the + +adversarial training approach uses the adversarial samples together with the original samples for training model (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) proposed the first NLP oriented adversarial training method by applying noise on the word representations. But the perturbed result is in the latent space and does not map to a meaningful sentence. Then many works turned to generating adversarial samples as a means of adversarial training. The often adopted method is the lexical substitution that is formulated as a combinatorial optimization problem and solved by heuristic search algorithms (Yoo and Qi, 2021). Since it requires much interaction with the target model for identifying the important words and selecting the appropriate replacement words, it is time-consuming. Due to the lack of supervised training data, so far there is not applicable end-to-end mode for adversarial generation. + +Another kind of approaches to improve model robustness is data augmentation, which aims to generate the similar samples to the original data for increasing the amount of training data. The traditional methods on data augmentation include backtranslation (Sennrich et al., 2015), rule-based modification (Wei and Zou, 2019), paraphrasing (Kumar et al., 2019) and other techniques (Li et al., 2022). These methods are based on the distribution of training data, which are model agnostic. However, with these data, the models are still vulnerable to adversarial samples as the attacking techniques are designed for finding the weakness of the target model (Yoo and Qi, 2021). + +To solve the above challenges, in this paper, we designed an adversarial sample generation method for data augmentation $\mathbf{A}^3$ . We propose an end-to-end adversarial sample generation model. Given an original sample, a condition generator is designed to output a condition which guides a conditioned paraphrasing model to generate a potential adversarial sample. By pretraining an adversarial dis + +criminator to be utilized as the training signal, it can learn the task-specific data characteristics for finding the potential vulnerabilities. Since the generation process doesn't require interaction with the target model, it is much efficient compared with the current search based generation methods. Besides, the generated samples can be reused for different models rather than for a specific model as the traditional way. Then we propose a weighted loss for incorporating the generated adversarial samples with the original data for the target task. We evaluate our method on several NLP tasks and compare it to both the data augmentation and adversarial training methods. Experimental results demonstrate that our approach not only improves the overall performance of the trained model but also enhances its robustness against multiple attacking techniques. Moreover, the computational time is significantly reduced compared to existing adversarial sample generation methods. + +# 2 Related works + +# 2.1 Adversarial Sample Generation + +In the NLP area, the disturbance on the original sample is often adopted in adversarial sample generation due to the discrete character of text, such as word replacement. For example, TextBugger (Li et al., 2019) generates the adversarial samples by randomly replacing the words in a sample text for gathering the candidate samples that may flip the model prediction. Then it chooses the reasonable adversarial samples with the similar semantics with the original text. PWWS (Ren et al., 2019) uses synonym for word substitution with the help of WordNet. With the development of pretrained language models (PLM), it helps various NLP tasks to gain improvement on the performance. Jin et al. proposed the TextFooler (Jin et al., 2020) method for evaluating the robustness of BERT. Meanwhile, several works utilize PLM to generate the candidate adversarial texts with similar semantics to original ones (Li et al., 2020, 2021). Recently, a survey on word substitution based adversarial sample generation (Chiang and Lee, 2022) shows that despite the effectiveness of such attacks, the generated samples have flaws on fluency and syntactical correctness and can be identified easily. + +# 2.2 Adversarial Training + +Based on the proximity assumption, researchers have explored methods to enhance the robustness + +and effectiveness of trained models by introducing disturbances to the original training samples, commonly known as adversarial training (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) first proposed a NLP-oriented method that adds noise to the word embeddings of the input sample for adversarial training in text classification task. Similar approaches have been applied to adversarial training for pretrained language models (Liu et al., 2020) and other NLP tasks (Li and Qiu, 2021), showing improvements in model performance. Despite their effectiveness, disturbance in the continuous spaces lack clear meaning in actual language, making them less interpretable. + +Recent works in adversarial training for NLP tasks consider modifications at the lexical level to make the perturbed samples more consistent with real samples. Early works use synonym substitution to explore proximate samples in the sample space. Huang et al. (Huang et al., 2019) and Dong et al. (Dong et al., 2021) used different sample space exploration methods to find samples that effectively enhance model robustness in adversarial training. The SSMBA method proposed by Ng et al. (Ng et al., 2020) generates perturbed samples by using pretrained language models to complete corrupted texts. Recently, researchers consider using adversarial sample generation methods for adversarial training. Previous work (Carlini et al., 2019) has shown that using samples generated by a specific adversarial sample generation method for augmented training generally only improves the robustness against that method. Yoo et al. designed the A2T method (Yoo and Qi, 2021), which introduced a fast lexical substitution adversarial sample generation method into the model training process, achieving improved robustness against multiple model attacking methods. Current adversarial training methods for NLP models are mostly based on constrained searching in the sample space, and lack the ability to self learn and adapt to models or datasets. Moreover, the lexical substitution based generation methods have potential problems regarding semantic soundness and syntactic correctness. (Chiang and Lee, 2022) + +# 3 The End-to-end Adversarial Sample Based Data Augmentation + +# 3.1 Framework + +In this paper, we design the end-to-end adversarial sample generation method for data augmentation. + +![](images/0ce257ad75998fd282d32545e6a2196b1198fdc4fe824b9ca452e97531e858f7.jpg) +Figure 1: The framework for the end-to-end adversarial sample based data augmentation + +For a given sample $\mathbf{x}$ , the condition generator gives the condition $\mathbf{c}$ , based on which the paraphrasing model generates $\mathbf{x}'$ that is a potential adversarial sample. The whole architecture consists of three parts, as shown in Fig.1. The first condition generator is designed to generate the appropriate conditions for guiding the generated paraphrase being an adversarial sample. The second is the pretrained conditioned paraphrasing model that generates the paraphrases satisfying the given condition. + +Since there are no supervised data for training the condition generator, we propose an adversarial discriminator to provide the proper training signals, similar to the structure of a generative adversarial network (GAN) (Goodfellow et al., 2014a). The third part is the adversarial discriminator that can be a classifier or regression model for the target task, which checks the flip of predication. + +Then for the target task, the generated adversarial samples are incorporated with the original data for data augmentation, which are used to train a model for this task. + +# 3.2 Condition Generator + +The condition generator takes an original sample as input and aims to generate a condition $\mathbf{c}$ for guiding the paraphrasing model to generate an adversarial sample. Let $G$ denote the condition generator. + +$$ +\mathbf {c} = G (\mathbf {x}) \tag {1} +$$ + +There are many candidate forms of condition $\mathbf{c}$ , such as the ratio of changed words, the sentence structure or the sentiment. Following the normal way of generating adversarial samples, we adopt the lexical substitution indicator as the condition, which is in the form of binary sequence corresponding to the words in $\mathbf{x}$ , denoted by $\mathbf{c} \in \{0,1\}^{|\mathbf{x}|}$ . $c_{i} = 1$ indicates $x_{i}$ should be changed in the paraphrase $\mathbf{x}'$ , $c_{i} = 0$ for otherwise. + +In this paper we adopt a multi-layer bidirectional recurrent neural network as the basic network. The output hidden vector at each step is projected to a scalar. The output of the condition generator is fed into the conditioned paraphrasing model. Since the condition $\mathbf{c}$ is defined as a sequence of binary value, while conventional activation function such as sigmoid outputs a contiguous value between 0 and 1. To maintain a reasonable form of $\mathbf{c}$ and achieve a stable training, we adopt the straight-through estimator (Bengio et al., 2013) method as the activation function, such that the output of the condition generator will be binary. + +$$ +\mathbf {c} = S T E (G (\mathbf {x})) \tag {2} +$$ + +# 3.3 Pretraining the Conditioned Paraphrasing Model + +The general purpose of paraphrasing is to generate a sample with the same meaning but different expression with the given text. Here we introduce the conditioned paraphrasing model for guiding the paraphrase. For a given text $\mathbf{x}$ and a condition $\mathbf{c}$ , it generates a paraphrase $\mathbf{x}'$ with the similar semantics to $\mathbf{x}$ and satisfying $\mathbf{c}$ , formally: + +$$ +\mathbf {x} ^ {\prime} = F (\mathbf {x}, \mathbf {c}) \tag {3} +$$ + +We adopt a multi-layer bidirectional recurrent neural network for the conditioned paraphrasing model. The pretraining data are automatically generated, where the input is a pair of a text $\mathbf{x}$ and the randomly selected positions as the condition c. As the paraphrase $\mathbf{x}'$ , there are many candidate methods for choosing the replacement words, such as the synonym vocabulary or the masked language model. To ensure the semantic similarity and the syntactical correctness, we adopt the word com- posable knowledge (Liu et al., 2023) for choosing the words to form a new text $\mathbf{x}'$ as the supervised target. + +Since only a small proportion of the text are changed, the simple negative log-likelihood loss would lead the model to just copy the input. Thus, we apply a larger weight $\omega_{p} > 0$ for the replacement positions $c_{i} = 1$ : + +$$ +\ell_ {P} = - \sum_ {i = 1} ^ {| \mathbf {x} |} \left(1 + c _ {i} \omega_ {p}\right) \log p \left(\hat {x} _ {i} ^ {\prime} = x _ {i} ^ {\prime} | \mathbf {x}, \mathbf {c}\right) \tag {4} +$$ + +where $\hat{x}_i^{\prime}$ is the output of the model and $x_{i}^{\prime}$ is the supervised paraphrase. + +# 3.4 Adversarial Discriminator + +To train the end-to-end adversarial sample generation model, we pretrain an adversarial discriminator to provide training signals, denote by $D$ . It is a trained classifier or a regression model for the target task. + +$$ +y = D (\mathbf {x}) \tag {5} +$$ + +Theoretically the paraphrasing model should generate the semantically similar text with the original one. If the prediction $\hat{y}^\prime$ by $D$ for the generated sample $x^{\prime}$ is flipped compared to the ground truth $y$ , $x^{\prime}$ is a potentially adversarial sample. + +# 3.5 End-to-End Adversarial Sample Generation Training + +The core of the adversarial sample generation model is to generate the appropriate condition for paraphrasing. Since the discriminator can be seen as a white-box victim model, we can pass the gradient through the neural network structure by utilizing the tricks like Gumbel softmax (Jang et al., 2017). This allows an end-to-end training mode for the whole adversarial sample generation method. + +During the training, the condition generator, the conditioned paraphrasing model and the adversarial discriminator are connected in a pipeline way. For a given sample $\mathbf{x}$ , the pipeline outputs a prediction $\hat{y}$ for a perturbed sample. The conditioned paraphrasing model $F(\mathbf{x},\mathbf{c})$ outputs a sequence of word probabilities. We adopt Gumbel softmax to convert the probabilities to word embeddings that can be directly used by $D$ . + +$$ +\mathbf {c} = G (\mathbf {x}) +$$ + +$$ +\mathbf {x} ^ {\prime} = F (\mathbf {x}, \mathbf {c}) \tag {6} +$$ + +$$ +\hat {y} = D (\mathbf {x} ^ {\prime}) +$$ + +During training, the parameters of the condition generator and the conditioned paraphrasing model are tuned while the adversarial discriminator remains frozen. Training the condition paraphrasing model allows the model to be trained adaptively to generate words more suitable for the target task. To train the adversarial sample generation model, we used the following losses. + +Adversarial loss is the main goal of the model, namely to generate a sample that can make the target model flip its prediction. Since previous works on model distilling show the effectiveness and simplicity of the MSE loss function (Hinton et al., 2015), thus it is adopted in our work. More + +advanced distances between probabilities such as KL-distance can be an alternative. + +$$ +\ell_ {a d v} = 1 - \left| D \left(\mathbf {x} ^ {\prime}\right) - D (\mathbf {x}) \right| ^ {2} \tag {7} +$$ + +Perturbation ratio loss To maintain the similar semantics between the perturbed sample and the original sample, the ratio of perturbed words in text are constrained, namely the number of $c_{i} = 1$ cases in the generated condition $\mathbf{c}$ . Let $\rho$ denote the expected ratio, $L$ is the length of sample, the perturbation ratio loss is calculated as follows. + +$$ +\ell_ {c o n} = \left| \rho - \| \mathbf {c} \| _ {1} / L \right| ^ {2} \tag {8} +$$ + +Conditioned paraphrasing losses The conditioned paraphrasing model is fine-tuned for different expectation by a few loss functions. The first is the reconstruction loss for the words not perturbed in $\mathbf{c}$ , i.e. on the positions $c_{i} = 0$ , denoted by $\mathbf{c}(0)$ . It is a negative likelihood loss for the original words such that the paraphrasing model is trained to reconstruct the original input. + +$$ +\ell_ {r e c} = - \sum_ {i \in \mathbf {c} (0)} \log p \left(\mathbf {x} _ {i} ^ {\prime} = \mathbf {x} _ {i} \mid \mathbf {x}, \mathbf {c}\right) \tag {9} +$$ + +For the perturbed positions, the probability of the original word should be lower than any other words. Inspired by the loss function for the negative samples in the noise contrastive estimation (Gutmann and Hyvarinen, 2010), we design the loss of punishing repeating the original word. Let $\mathbf{c}(1)$ denote the positions of the 1s in the condition $\mathbf{c}$ . + +$$ +\ell_ {p u n} = - \sum_ {i \in \mathbf {c} (1)} \sigma (- \log p \left(\mathbf {x} _ {i} ^ {\prime} = \mathbf {x} _ {i} \mid \mathbf {x}, \mathbf {c}\right)) \tag {10} +$$ + +Semantics similarity and syntactic correctness losses To maintain the semantics of the generated paraphrase similar with the original sample, we adopt a word-level similarity loss based on cosine similarities for the word embeddings between the generated word and the original word. + +$$ +\ell_ {s e m} = 1 - \operatorname {s i m} \left(\mathbf {x}, \mathbf {x} ^ {\prime}\right) +$$ + +$$ +\operatorname {s i m} \left(\mathbf {x}, \mathbf {x} ^ {\prime}\right) = \frac {1}{| \mathbf {x} |} \sum_ {i = 1} ^ {| x |} \left(1 - \cos \left(\mathbf {x} _ {i}, \mathbf {x} _ {i} ^ {\prime}\right)\right) \tag {11} +$$ + +The word embeddings similarity cannot guarantee a proper syntactical role of the replacement word. Thus, we also adopt the word composable + +Algorithm 1 The Training Process for End-to-End Adversarial Sample Generator +Input: Training set $D_{t}$ +```txt +Output: Condition generator $G$ , conditioned paraphrasing model $F$ +``` + +1: //Pretrain conditioned paraphrasing model +2: Generate paraphrasing training data $D_{p}$ based on $D_{t}$ +3: Pretrain $F$ with $D_{p}$ according to Eq.4 +4: //Pretrain adversarial discriminator +5: Pretrain $D$ with $D_{t}$ +6: //Train the end-to-end adversarial sample generation model +7: Freeze the parameters of $D$ +8: for sample $\mathbf{x}, y$ in $D_{t}$ do +9: Let $\mathbf{c} = G(\mathbf{x})$ +10: Calculate $\ell_{con}$ according to Eq.8 +11: Let $\mathbf{x}' = F(\mathbf{x},\mathbf{c})$ +12: Calculate $\ell_{rec},\ell_{pun},\ell_{sem},\ell_{syn}$ according to Eq.9-12 +13: Let $\hat{y} = D(\mathbf{x}^{\prime})$ +14: Calculate $\ell_{adv}$ according to Eq.7 +15: Calculate $L_{adv}$ according to Eq.13 +16: Update $G,F$ according to $L_{adv}$ +17: end for +18: return Condition generator $G$ , conditioned paraphrasing model $F$ + +knowledge based loss (Liu et al., 2023) to provide syntactical correctness guidance for the generated sample. Given a dependency relationship between two words $< u, r, v >$ , the word composable knowledge estimates the syntactical soundness of this combination, denoted by $p_s(u, r, v)$ . It relies on the pretrained word embeddings and thus is easy to integrate with other neural network structures. Let $s_x = (i, r, j)$ denote the dependency parsing tree of $\mathbf{x}$ , where $(i, r, j)$ represents the dependency relation between $\mathbf{x}_j$ and its head word $\mathbf{x}_i$ is $r$ . The syntactic correctness loss is calculated as follows: + +$$ +\ell_ {s y n} = - \sum_ {(i, r, j) \in s _ {x}} \log p _ {s} \left(\mathbf {x} _ {i} ^ {\prime}, r, \mathbf {x} _ {j} ^ {\prime}\right) \tag {12} +$$ + +The complete loss function for training the adaptive adversarial sample generation model is the weighted sum of the above losses. + +$$ +\begin{array}{l} L _ {a d v} = w _ {a} \ell_ {a d v} + w _ {c} \ell_ {c o n} + w _ {r} \ell_ {r e c} + w _ {p} \ell_ {p u n} \\ + w _ {s e m} \ell_ {s e m} + w _ {s y n} \ell_ {s y n} \tag {13} \\ \end{array} +$$ + +
Dataset#Class#Train#TestAve.Len.
IMDB22500025000270.7
Quora23636604040425.33
SNLI35501521000022.4
+ +Table 1: Dataset statistics + +The overall training process of the end-to-end adversarial sample generation is listed in Alg.1. + +# 3.6 Data Augmentation + +Having the above end-to-end adversarial sample generator, we can apply data augmentation on the training data for the target task. For the given original training set $D_{t}$ , the corresponding generated adversarial samples is denoted by $D_{t}^{\prime}$ . Then the model for the target task can be trained on $D_{t} \cup D_{t}^{\prime}$ for improving the model robustness. Since the adversarial samples are more likely to introduce noises, their training loss is assigned a lower weight. Let $L_{t}$ denotes the loss of the target task, $\lambda_{i}$ denotes the loss weight for sample $s_{i}$ . $\lambda_{i} = 1$ for an original sample while $\lambda_{i} < 1$ for a generated sample. + +$$ +L _ {t a s k} = \sum_ {s _ {i} \in D _ {t} \bigcup D _ {t} ^ {\prime}} \lambda_ {i} L _ {t} (s _ {i}) \tag {14} +$$ + +The previous work has shown the adversarial samples generated based on one model are potentially adversarial to other models (Szegedy et al., 2013). + +# 4 Experiments + +To evaluate the effectiveness of the proposed method, four aspects should be considered and evaluated in the subsections 4.2-4.5. + +A1: To check the attacking effectiveness of generated adversarial sample. +A2: To check whether the model performance is improved after augmented training. +A3: To check whether the robustness of the target model is improved. +A4: To check the efficiency of our method. + +# 4.1 Experiment Setups + +# 4.1.1 Tasks + +Three tasks are considered as evaluation tasks to cover a wider range of NLP tasks such that the effectiveness of data augmentation can be thoroughly evaluated in different scenarios. The first is a sentiment analysis task on the dataset IMDB. The IMDB + +
Base ModelBERTDistilBERTPerturbed %Bert Score
DatasetOri.Acc.Att.Acc.ΔOri.Acc.Att.Acc.Δ
IMDB93.9%66.58%-27.32%92.96%61.76%-31.2%9.95%95.98
Quora84.84%75.95%-8.89%84.32%76.44%-7.88%16.6%95.96
SNLI88.95%55.64%-33.31%88.04%55.20%-32.84%14.73%95.62
+ +dataset consists of long movie reviews which are classified as positive of negative sentiment polarity. The second is a text classification task on the dataset Quora Question Pairs(QQP), which consists of question pairs from the Quora site classified as duplicate or not. The last is a natural language inference task on the dataset SNLI. Each sample in SNLI consists a hypothesis sentence and a premise sentence, the logical relationship between them is determined as entailment, neutral or contradiction. The statistics of the datasets are listed in Table.1. + +# 4.1.2 Base Models + +Recently, pretrained language models are widely used as a fundamental part for NLP methods. In the following experiments, BERT (Kenton and Toutanova, 2019) and DistilBERT (Sanh et al., 2019) are chosen as the base model for the tasks. BERT is a widely used pretrained language model, and achieves improvements on various NLP tasks. DistilBERT is a compressed version of BERT model based on model distilling technique, making it easier to be used on devices with lower computation power. Both models will be finetuned on the original train set or the augmented train set for the target task and evaluated afterwards. Their results will be listed as Baseline in the following experiments. + +# 4.1.3 Implementation + +All the experiments are conducted on a single NVIDIA Geforce RTX 2080Ti. Our method is implemented with PyTorch 1.12.0. The code for our proposed method is shared at https://splab.sdu.edu.cn/xscg/sjjydm.htm. The compared methods and adversarial sample generation are implemented by TextAttack (Morris et al., 2020b). The size of hidden vectors in our model is 300. We adopt GloVe (Pennington et al., 2014) as the word embeddings and reused the pretrained parameters of word composable knowledge released in SynPara (Liu et al., 2023). For pretraining the paraphrasing model, the ratio of replaced words is set + +Table 2: Accuracy comparison between accuracies on the original testing set and adversarial testing set. + +
Base ModelMethodDataset
IMDBQuoraSNLI
BERTBaseline93.984.8488.95
SSMBA-0.41+3.03-2.88
A2T-0.03+1.4-0.5
A3+1+4+0.1
Distil-BERTBaseline91.8784.3388.04
SSMBA+1.09+3.79-0.12
A2T+1.27+4.39+1.38
A3+1.93+4.89+0.22
+ +Table 3: Accuracy comparison on the original testing set. The values for the augmentation methods are the differences with the accuracy of the base model. + +$15\%$ when generating training data. For training the complete model, the weights in the final loss function are set as $w_{a} = 5.0$ , $w_{c} = 3.0$ , $w_{r} = 2.0$ , $w_{p} = 5.0$ , $w_{sem} = 2.0$ , $w_{syn} = 0.1$ in most cases. The ratio of perturbed words $\rho$ for the condition generator is set as 0.15 for SNLI and 0.1 for IMDB and Quora unless stated otherwise. + +# 4.1.4 Compared methods + +For the comparison methods, we consider the following methods. SSMBA (Ng et al., 2020) is a data augmentation method that uses the corruption function similar to the masked language model training and reconstructs the samples with BERT. A2T (Yoo and Qi, 2021) is an adversarial training method, where a fast adversarial sample generation method uses the gradients for identifying important words in a sample and replaces them based on word embeddings or MLM models. It iteratively trains the task model and augment the training data set with the generated adversarial samples. + +# 4.2 Attack Effectiveness + +To evaluate the effectiveness of the generated adversarial samples, we compared the performances of the baseline models on the original testing data and on the adversarial data. The results are shown in Table.2, where the ratio of perturbed words in + +
Base ModelBERTDistilBERT
DatasetsIMDBQuoraSNLIIMDBQuoraSNLI
Attack MethodAugment Method%Δ%Δ%Δ%Δ%Δ%Δ
A3Baseline69.377.358.364.180.159.8
SSMBA71.6+2.377.7+0.458.7+0.461.6-2.578.8-1.362.9+3.1
A2T73.7+4.476.9-0.459.7+1.466.1+278.6-1.560.3+0.5
A392.2+22.989.3+1286.2+27.993.8+29.790.9+10.886.5+26.7
A2TBaseline38.840.144.130.650.341.1
SSMBA38.9+148.1+849.3+5.227.9-2.750.7+0.445.1+4
A2T44+5.276.2+36.179.4+35.354.5+23.976.1+25.875.6+34.5
A345.3+6.560.7+20.648+3.936.1+5.560.2+9.943.3+2.2
Text-FooterBaseline1.723.60.90.110.30.47
SSMBA5.3+3.612.5-11.10.4-0.50-0.112.1+1.80.73+0.26
A2T0.8-0.914.3-9.35.1+4.21.3+1.28-2.33.93+3.46
A312+10.327.1+3.51.9+11.6+1.515.2+4.90.67+0.2
BAEBaseline3139.322.124.939.323.87
SSMBA30.6-0.439-0.325+2.921-3.939.7+0.423.93+0.06
A2T29.1-1.943.4+4.123.5+1.426.3+1.444.2+4.925.5+1.63
A337.9+6.940.8+1.527.6+5.527.7+2.845.7+6.424.27+0.4
PWWSBaseline125.11.20.822.11.07
SSMBA2.2+1.223.1-22.3+1.10.1-0.722.9+0.81.2+0.13
A2T0.3-0.726.3+1.20.7-0.50.6-0.223+0.90.87-0.2
A37.3+6.327.7+2.62.4+1.20.9+0.124.1+22.4+1.33
+ +Table 4: Comparison of attacking methods against augmentation methods. Column % denotes the model accuracy under attacking. Column $\Delta$ denotes the difference between augmented model with the original model. + +the adversarial samples is listed in column Perturbed%. A lower ratio indicates more structure information of the text preserved. BertScore (Zhang et al., 2020) is used for quantifying the semantics similarity between the original sample and the adversarial sample. The higher the BertScore, the closer their semantics. We can see that the model performances of both baseline models drop significantly on the adversarial samples, which show that the generated adversarial samples are effective on attacking the baseline models. + +# 4.3 Performance Improvement + +We then evaluated the model performance on the target tasks before and after training with augmented data. By using our method and SSMBA, the data is augmented to doubled size of the original training set, respectively. The augmented dataset is used for training the base models. For A2T, it iteratively trains the model and then augments the training set by generating adversarial samples. The results in Table.3 show that our method improves the performance on all settings and are better than the compared methods in most cases. The compared methods are unstable on performances, namely in some cases performances are the improved while some are not. For example, SSMBA has negative impact on SNLI dataset for both base models. This is probably because SSMBA doesn't take into con + +sideration the task-specific characteristics, such as the logical connections between the hypothesis and premise in SNLI. It should be also noted that while the original performance of DistilBERT is slightly lower than BERT, it improves a lot after data augmentation, which confirmed previous studies regarding the effectiveness of data augmentation for distilled models (Das et al., 2020). + +# 4.4 Robustness Improvement + +To evaluate the robustness of base model against different attacking methods, we choose three representative attacking methods: TextFooler (Jin et al., 2020), BAE (Garg and Ramakrishnan, 2020) and PWWS (Ren et al., 2019). We adopt the notion $m1 - m2$ denoting the attacking method $m1$ against augmentation method $m2$ . For example, BAE-A2T means using BAE to attack a model trained using A2T. Similar to the settings of previous works (Yoo and Qi, 2021), 1000 samples are selected from the original testing set of each dataset (1500 samples from SNLI for three labels), where the label distributions are balanced. To better show the impact of an attacking method, the selected samples can be predicted correctly by the base models before data augmentation. The attack methods generate the adversarial samples based on the selected samples and the accuracy on these adversarial samples of the trained model would be compared. A2T and + +
MethodsDataset
IMDBQuoraSNLI
SSMBA10m38m1h2m
A2T36h(3)111h(2)70h(2)
A34m27m21m
+ +Table 5: Time for augmenting the dataset. It should be noted that the time for A2T is the cumulated time for multiple training epochs (number of epochs is listed in the brackets). + +![](images/a019b2b244879cc0e9658e0a2870775cec70d24b3a78a269374655290b3cb541.jpg) +Figure 2: The influence of $\rho$ on the perturbed ratio and the attack effectiveness (denoted by $\Delta \mathrm{Acc}$ ) + +our proposed method are tested as they are also adversarial sample generation methods. + +The results are listed in Table.4, where the column $\%$ denotes the model accuracy under attacking and column $\Delta$ shows the difference between the performance of the augmented model with the baseline. As explored in a previous research (Carlini et al., 2019), the robustness against a certain attack method will be improved if trained with the corresponding adversarial samples. From the row $\mathrm{A}^3\mathrm{-A}^3$ and A2T-A2T, we can see that both methods significantly improve the model robustness. For other representative attacking methods, the augmented models with our generated data show better robustness than the compared methods in most settings. + +# 4.5 Efficiency Comparison + +Another vital aspect of adversarial training is the efficiency. Most adversarial sample generation methods involves constantly interaction with the target model. Such process is both computational heavy and time-consuming. We recorded the time used for augmenting the training set of different methods. Both SSMBA and our proposed method are one-time augmentation thus the time taken is similar. Our method is slightly faster than SSMBA since SSMBA uses BERT as the reconstruction model. The A2T method, on the other hand, is significantly slower than our method due to two possible rea + +son. First reason is that the attacking process is conducted multiple times. Secondly, the method is based on attacking the target model which is not only time-consuming, but the time taken will also increase over time since the model is trained better after each epoch. Comparably, our proposed method generates model-agnostic adversarial samples, such that the time for augmented training is significantly shortened. + +# 4.6 Parameter Analysis + +We analyze the effect of parameter $\rho$ , the targeted perturbed ratio. In our method for a $\rho$ setting, the actual perturbed ratio in the generated adversarial samples may be different with the initial value. Thus, we analyze the actual perturbed ratio and its impact on the attacking effectiveness, shown in Fig.2. The blue and orange lines represent the set $\rho$ and the actual perturbed ratio respectively, corresponding to the left Y-axis. The $\Delta \mathrm{Acc}$ is the accuracy difference after the attack, corresponding to the right Y-axis. Higher value of $\Delta \mathrm{Acc}$ means more effective attack. We can see that the actual perturbed ratio approaches the $\rho$ setting. A higher $\rho$ may cause the predicated label flipped but may lead to a larger semantics difference. Thus, the parameter $\rho$ should be set properly to achieve a balanced performance. + +# 5 Conclusion + +In this paper, we present a method named $\mathrm{A}^3$ that combines the advantages of adversarial training and data augmentation to enhance the robustness of NLP models. We introduce an end-to-end adversarial sample generation model which consists of a condition generator and a conditioned paraphrasing model. A pretrained adversarial discriminator is proposed to train the adversarial sample generation model for capturing the task-specific data characteristics. This method is much efficient without the need for extensive model interactions. The generated adversarial samples are incorporated with the original training samples by a weighted loss for augmented training downstream models. Experiments are conducted on multiple NLP tasks and demonstrate the effectiveness and efficiency of our proposed method. Compared to existing methods, the performances of the base models have improved more after our data augmentation, as well as their robustness enhanced against multiple attacking techniques. + +# Limitations + +Firstly, the current form of condition limits the diversity and fine-grained control over the generation process. For example, a condition can be a subtree from the dependency syntax structure, which allows paraphrases with potential structure reform and correct syntax (Liu et al., 2023). Secondly, a condition may also contain the direction of replacement word for the paraphrasing model. This would allow reusing the same conditioned paraphrasing model across different tasks, thereby reducing the training overhead for adversarial sample generation. Lastly, in future research, we aim to explore the feasibility of directly using the gradient from the target model for training the adversarial sample generation model. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (62376138) and the Innovative Development Joint Fund Key Projects of Shandong NSF (ZR2022LZH007). + +# References + +Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432. +Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. 2019. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705. +Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan Zhang, and Cho-Jui Hsieh. 2020. Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3601-3608. +Cheng-Han Chiang and Hung-yi Lee. 2022. How far are we from real synonym substitution attacks? arXiv preprint arXiv:2210.02844. +Deepan Das, Haley Massa, Abhimanyu Kulkarni, and Theodoros Rekatsinas. 2020. An empirical analysis of the impact of data augmentation on knowledge distillation. CoRR, abs/2006.03810. +Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. 2021. Towards robustness against natural language word substitutions. arXiv preprint arXiv:2107.13541. + +Javid Ebrahimi, Anyi Rao, Daniel Lowd, and De- jing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751. +Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6174-6181. +Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. 2014a. Generative adversarial networks. CoRR, abs/1406.2661. +Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014b. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572. +Shreya Goyal, Sumanth Doddapaneni, Mitesh M Khapra, and Balaraman Ravindran. 2022. A survey of adversarial defences and robustness in nlp. ACM Computing Surveys. +Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297-304. JMLR Workshop and Conference Proceedings. +Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531. +Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, and Pushmeet Kohli. 2019. Achieving verified robustness to symbol substitutions via interval bound propagation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4083-4093. +Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017). OpenReview.net. +Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025. +Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186. +Ashutosh Kumar, Satwik Bhattamishra, Manik Bhandari, and Partha Talukdar. 2019. Submodular + +optimization-based diverse paraphrasing and its effectiveness in data augmentation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3609-3619. +Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open, 3:71-90. +Dianqi Li, Yizhe Zhang, Hao Peng, Liquin Chen, Chris Brockett, Ming-Ting Sun, and William B Dolan. 2021. Contextualized perturbation for textual adversarial attack. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5053-5069. +J Li, S Ji, T Du, B Li, and T Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium. +Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020. Bert-attack: Adversarial attack against using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6193–6202. +Linyang Li and Xipeng Qiu. 2021. Token-aware virtual adversarial training in natural language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8410-8418. +Tianyuan Liu, Yuqing Sun, Jiaqi Wu, Xu Xi, Yuchen Han, Cheng Li, and Bin Gong. 2023. Unsupervised paraphrasing under syntax knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 37. +Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994. +Takeru Miyato, Andrew M Dai, and Ian Goodfellow. 2016. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725. +John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, and Yanjun Qi. 2020a. Reevaluating adversarial examples in natural language. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3829-3839. +John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020b. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126. + +Nathan Ng, Kyunghyun Cho, and Marzyeh Ghassemi. 2020. Ssmba: Self-supervised manifold based data augmentation for improving out-of-domain robustness. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1268-1283. +Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics. +Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. 2019. Generating natural language adversarial examples through probability weighted word saliency. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 1085-1097. +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. +Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709. +Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199. +Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196. +Jin Yong Yoo and Yanjun Qi. 2021. Towards improving adversarial training of nlp models. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 945-956. +Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net. \ No newline at end of file diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/images.zip b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..380b0526402070d21a0d2baf41161529a5dbfccd --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c61546f8282b7a36d72ab948d880a7504850a2ab1db5ee3489d051d313e2dee +size 367364 diff --git a/2023/End-to-end Adversarial Sample Generation for Data Augmentation/layout.json b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..447fcd88acc43406bef379676eb6eefc6c346031 --- /dev/null +++ b/2023/End-to-end Adversarial Sample Generation for Data Augmentation/layout.json @@ -0,0 +1,8918 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 75, + 504, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 75, + 504, + 94 + ], + "spans": [ + { + "bbox": [ + 89, + 75, + 504, + 94 + ], + "type": "text", + "content": "End-to-end Adversarial Sample Generation for Data Augmentation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 217, + 119, + 383, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 119, + 383, + 133 + ], + "spans": [ + { + "bbox": [ + 217, + 119, + 383, + 133 + ], + "type": "text", + "content": "Tianyuan Liu and Yuqing Sun*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 247, + 135, + 349, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 135, + 349, + 148 + ], + "spans": [ + { + "bbox": [ + 247, + 135, + 349, + 148 + ], + "type": "text", + "content": "Shandong University" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 176, + 149, + 420, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 149, + 420, + 162 + ], + "spans": [ + { + "bbox": [ + 176, + 149, + 420, + 162 + ], + "type": "text", + "content": "zodiacg@foxmail.com sun_yuqing@sdu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 236, + 274, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 236, + 274, + 534 + ], + "spans": [ + { + "bbox": [ + 86, + 236, + 274, + 534 + ], + "type": "text", + "content": "Adversarial samples pose a significant challenge to neural inference models. In this paper, we propose a novel enhancing approach " + }, + { + "bbox": [ + 86, + 236, + 274, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^3" + }, + { + "bbox": [ + 86, + 236, + 274, + 534 + ], + "type": "text", + "content": " for the robustness of the neural NLP models, which combines the adversarial training and data augmentation. We propose an adversarial sample generator that consists of a conditioned paraphrasing model and a condition generator. The latter aims to generate conditions which guides the paraphrasing model to generate adversarial samples. A pretrained discriminator is introduced to help the adversarial sample generator adapt to the data characteristics for different tasks. We adopt a weighted loss to incorporate the generated adversarial samples with the original samples for augmented training. Compared to existing methods, our approach is much efficient since the generation process is independent to the target model and the generated samples are reusable for different models. Experimental results on several tasks show that our approach improves the overall performance of the trained model. Specially, the enhanced model is robust for various attacking techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 544, + 154, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 544, + 154, + 557 + ], + "spans": [ + { + "bbox": [ + 68, + 544, + 154, + 557 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 566, + 291, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 566, + 291, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 566, + 291, + 714 + ], + "type": "text", + "content": "Adversarial samples refer to the inputs that intentionally designed to fool deep learning models (Morris et al., 2020a), which are also called the attack samples (Szegedy et al., 2013). The methods for generating attack samples have deceived many natural language processing (NLP) models, such as text classification (Ebrahimi et al., 2017; Ren et al., 2019), machine translation (Cheng et al., 2020) and etc (Goyal et al., 2022). Along with the application of neural models in various areas, the adversarial samples have drawn much attention." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 715, + 291, + 755 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 715, + 291, + 755 + ], + "spans": [ + { + "bbox": [ + 67, + 715, + 291, + 755 + ], + "type": "text", + "content": "Generally, the adversarial samples are generated by perturbations on the original sample. To enhance the robustness of NLP neural methods, the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 212, + 526, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 212, + 526, + 470 + ], + "spans": [ + { + "bbox": [ + 302, + 212, + 526, + 470 + ], + "type": "text", + "content": "adversarial training approach uses the adversarial samples together with the original samples for training model (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) proposed the first NLP oriented adversarial training method by applying noise on the word representations. But the perturbed result is in the latent space and does not map to a meaningful sentence. Then many works turned to generating adversarial samples as a means of adversarial training. The often adopted method is the lexical substitution that is formulated as a combinatorial optimization problem and solved by heuristic search algorithms (Yoo and Qi, 2021). Since it requires much interaction with the target model for identifying the important words and selecting the appropriate replacement words, it is time-consuming. Due to the lack of supervised training data, so far there is not applicable end-to-end mode for adversarial generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 473, + 527, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 473, + 527, + 662 + ], + "spans": [ + { + "bbox": [ + 302, + 473, + 527, + 662 + ], + "type": "text", + "content": "Another kind of approaches to improve model robustness is data augmentation, which aims to generate the similar samples to the original data for increasing the amount of training data. The traditional methods on data augmentation include backtranslation (Sennrich et al., 2015), rule-based modification (Wei and Zou, 2019), paraphrasing (Kumar et al., 2019) and other techniques (Li et al., 2022). These methods are based on the distribution of training data, which are model agnostic. However, with these data, the models are still vulnerable to adversarial samples as the attacking techniques are designed for finding the weakness of the target model (Yoo and Qi, 2021)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 665, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 665, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 665, + 527, + 773 + ], + "type": "text", + "content": "To solve the above challenges, in this paper, we designed an adversarial sample generation method for data augmentation " + }, + { + "bbox": [ + 302, + 665, + 527, + 773 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^3" + }, + { + "bbox": [ + 302, + 665, + 527, + 773 + ], + "type": "text", + "content": ". We propose an end-to-end adversarial sample generation model. Given an original sample, a condition generator is designed to output a condition which guides a conditioned paraphrasing model to generate a potential adversarial sample. By pretraining an adversarial dis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 83, + 761, + 169, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 761, + 169, + 772 + ], + "spans": [ + { + "bbox": [ + 83, + 761, + 169, + 772 + ], + "type": "text", + "content": "Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "11359" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "spans": [ + { + "bbox": [ + 125, + 795, + 468, + 818 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11359-11368 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 343 + ], + "type": "text", + "content": "criminator to be utilized as the training signal, it can learn the task-specific data characteristics for finding the potential vulnerabilities. Since the generation process doesn't require interaction with the target model, it is much efficient compared with the current search based generation methods. Besides, the generated samples can be reused for different models rather than for a specific model as the traditional way. Then we propose a weighted loss for incorporating the generated adversarial samples with the original data for the target task. We evaluate our method on several NLP tasks and compare it to both the data augmentation and adversarial training methods. Experimental results demonstrate that our approach not only improves the overall performance of the trained model but also enhances its robustness against multiple attacking techniques. Moreover, the computational time is significantly reduced compared to existing adversarial sample generation methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 353, + 163, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 353, + 163, + 365 + ], + "spans": [ + { + "bbox": [ + 67, + 353, + 163, + 365 + ], + "type": "text", + "content": "2 Related works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 375, + 244, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 375, + 244, + 387 + ], + "spans": [ + { + "bbox": [ + 67, + 375, + 244, + 387 + ], + "type": "text", + "content": "2.1 Adversarial Sample Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 393, + 291, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 393, + 291, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 291, + 719 + ], + "type": "text", + "content": "In the NLP area, the disturbance on the original sample is often adopted in adversarial sample generation due to the discrete character of text, such as word replacement. For example, TextBugger (Li et al., 2019) generates the adversarial samples by randomly replacing the words in a sample text for gathering the candidate samples that may flip the model prediction. Then it chooses the reasonable adversarial samples with the similar semantics with the original text. PWWS (Ren et al., 2019) uses synonym for word substitution with the help of WordNet. With the development of pretrained language models (PLM), it helps various NLP tasks to gain improvement on the performance. Jin et al. proposed the TextFooler (Jin et al., 2020) method for evaluating the robustness of BERT. Meanwhile, several works utilize PLM to generate the candidate adversarial texts with similar semantics to original ones (Li et al., 2020, 2021). Recently, a survey on word substitution based adversarial sample generation (Chiang and Lee, 2022) shows that despite the effectiveness of such attacks, the generated samples have flaws on fluency and syntactical correctness and can be identified easily." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 728, + 195, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 728, + 195, + 741 + ], + "spans": [ + { + "bbox": [ + 67, + 728, + 195, + 741 + ], + "type": "text", + "content": "2.2 Adversarial Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 290, + 772 + ], + "type": "text", + "content": "Based on the proximity assumption, researchers have explored methods to enhance the robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 260 + ], + "type": "text", + "content": "and effectiveness of trained models by introducing disturbances to the original training samples, commonly known as adversarial training (Goodfellow et al., 2014b). Miyato et al. (Miyato et al., 2016) first proposed a NLP-oriented method that adds noise to the word embeddings of the input sample for adversarial training in text classification task. Similar approaches have been applied to adversarial training for pretrained language models (Liu et al., 2020) and other NLP tasks (Li and Qiu, 2021), showing improvements in model performance. Despite their effectiveness, disturbance in the continuous spaces lack clear meaning in actual language, making them less interpretable." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 261, + 526, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 261, + 526, + 682 + ], + "spans": [ + { + "bbox": [ + 302, + 261, + 526, + 682 + ], + "type": "text", + "content": "Recent works in adversarial training for NLP tasks consider modifications at the lexical level to make the perturbed samples more consistent with real samples. Early works use synonym substitution to explore proximate samples in the sample space. Huang et al. (Huang et al., 2019) and Dong et al. (Dong et al., 2021) used different sample space exploration methods to find samples that effectively enhance model robustness in adversarial training. The SSMBA method proposed by Ng et al. (Ng et al., 2020) generates perturbed samples by using pretrained language models to complete corrupted texts. Recently, researchers consider using adversarial sample generation methods for adversarial training. Previous work (Carlini et al., 2019) has shown that using samples generated by a specific adversarial sample generation method for augmented training generally only improves the robustness against that method. Yoo et al. designed the A2T method (Yoo and Qi, 2021), which introduced a fast lexical substitution adversarial sample generation method into the model training process, achieving improved robustness against multiple model attacking methods. Current adversarial training methods for NLP models are mostly based on constrained searching in the sample space, and lack the ability to self learn and adapt to models or datasets. Moreover, the lexical substitution based generation methods have potential problems regarding semantic soundness and syntactic correctness. (Chiang and Lee, 2022)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 691, + 510, + 719 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 691, + 510, + 719 + ], + "spans": [ + { + "bbox": [ + 302, + 691, + 510, + 719 + ], + "type": "text", + "content": "3 The End-to-end Adversarial Sample Based Data Augmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 728, + 387, + 740 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 728, + 387, + 740 + ], + "spans": [ + { + "bbox": [ + 302, + 728, + 387, + 740 + ], + "type": "text", + "content": "3.1 Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 746, + 526, + 772 + ], + "type": "text", + "content": "In this paper, we design the end-to-end adversarial sample generation method for data augmentation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 314, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 314, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 314, + 791 + ], + "type": "text", + "content": "11360" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 290, + 173 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 290, + 173 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 290, + 173 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 290, + 173 + ], + "type": "image", + "image_path": "0ce257ad75998fd282d32545e6a2196b1198fdc4fe824b9ca452e97531e858f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 179, + 291, + 206 + ], + "lines": [ + { + "bbox": [ + 67, + 179, + 291, + 206 + ], + "spans": [ + { + "bbox": [ + 67, + 179, + 291, + 206 + ], + "type": "text", + "content": "Figure 1: The framework for the end-to-end adversarial sample based data augmentation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "text", + "content": "For a given sample " + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "text", + "content": ", the condition generator gives the condition " + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "text", + "content": ", based on which the paraphrasing model generates " + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 67, + 227, + 290, + 362 + ], + "type": "text", + "content": " that is a potential adversarial sample. The whole architecture consists of three parts, as shown in Fig.1. The first condition generator is designed to generate the appropriate conditions for guiding the generated paraphrase being an adversarial sample. The second is the pretrained conditioned paraphrasing model that generates the paraphrases satisfying the given condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 363, + 290, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 363, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 67, + 363, + 290, + 470 + ], + "type": "text", + "content": "Since there are no supervised data for training the condition generator, we propose an adversarial discriminator to provide the proper training signals, similar to the structure of a generative adversarial network (GAN) (Goodfellow et al., 2014a). The third part is the adversarial discriminator that can be a classifier or regression model for the target task, which checks the flip of predication." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 472, + 291, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 472, + 291, + 525 + ], + "spans": [ + { + "bbox": [ + 67, + 472, + 291, + 525 + ], + "type": "text", + "content": "Then for the target task, the generated adversarial samples are incorporated with the original data for data augmentation, which are used to train a model for this task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 538, + 194, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 194, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 194, + 550 + ], + "type": "text", + "content": "3.2 Condition Generator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "text", + "content": "The condition generator takes an original sample as input and aims to generate a condition " + }, + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "text", + "content": " for guiding the paraphrasing model to generate an adversarial sample. Let " + }, + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 67, + 557, + 290, + 611 + ], + "type": "text", + "content": " denote the condition generator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 624, + 290, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 624, + 290, + 638 + ], + "spans": [ + { + "bbox": [ + 155, + 624, + 290, + 638 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = G (\\mathbf {x}) \\tag {1}", + "image_path": "fded484a5d04dff83f5d86159d981cfa70f8ced56f96f266f6ba20c3904b64ae.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": "There are many candidate forms of condition " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": ", such as the ratio of changed words, the sentence structure or the sentiment. Following the normal way of generating adversarial samples, we adopt the lexical substitution indicator as the condition, which is in the form of binary sequence corresponding to the words in " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\{0,1\\}^{|\\mathbf{x}|}" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "c_{i} = 1" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": " indicates " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": " should be changed in the paraphrase " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "inline_equation", + "content": "c_{i} = 0" + }, + { + "bbox": [ + 67, + 651, + 292, + 772 + ], + "type": "text", + "content": " for otherwise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": "In this paper we adopt a multi-layer bidirectional recurrent neural network as the basic network. The output hidden vector at each step is projected to a scalar. The output of the condition generator is fed into the conditioned paraphrasing model. Since the condition " + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": " is defined as a sequence of binary value, while conventional activation function such as sigmoid outputs a contiguous value between 0 and 1. To maintain a reasonable form of " + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 71, + 526, + 248 + ], + "type": "text", + "content": " and achieve a stable training, we adopt the straight-through estimator (Bengio et al., 2013) method as the activation function, such that the output of the condition generator will be binary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 258, + 525, + 272 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 258, + 525, + 272 + ], + "spans": [ + { + "bbox": [ + 373, + 258, + 525, + 272 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = S T E (G (\\mathbf {x})) \\tag {2}", + "image_path": "50b51f9c572e3b85add27a736b820c7b565d15c25e9f871461b23aac89983836.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 283, + 525, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 283, + 525, + 308 + ], + "spans": [ + { + "bbox": [ + 302, + 283, + 525, + 308 + ], + "type": "text", + "content": "3.3 Pretraining the Conditioned Paraphrasing Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "spans": [ + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": "The general purpose of paraphrasing is to generate a sample with the same meaning but different expression with the given text. Here we introduce the conditioned paraphrasing model for guiding the paraphrase. For a given text " + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": " and a condition " + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": ", it generates a paraphrase " + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": " with the similar semantics to " + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": " and satisfying " + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 314, + 525, + 409 + ], + "type": "text", + "content": ", formally:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 419, + 525, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 419, + 525, + 433 + ], + "spans": [ + { + "bbox": [ + 378, + 419, + 525, + 433 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {3}", + "image_path": "87ecc63c0ac096a252f5955026a7756f089c4d0ee27770db63e085420476c821.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "spans": [ + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "text", + "content": "We adopt a multi-layer bidirectional recurrent neural network for the conditioned paraphrasing model. The pretraining data are automatically generated, where the input is a pair of a text " + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "text", + "content": " and the randomly selected positions as the condition c. As the paraphrase " + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "text", + "content": ", there are many candidate methods for choosing the replacement words, such as the synonym vocabulary or the masked language model. To ensure the semantic similarity and the syntactical correctness, we adopt the word com- posable knowledge (Liu et al., 2023) for choosing the words to form a new text " + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 302, + 445, + 525, + 620 + ], + "type": "text", + "content": " as the supervised target." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "spans": [ + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "text", + "content": "Since only a small proportion of the text are changed, the simple negative log-likelihood loss would lead the model to just copy the input. Thus, we apply a larger weight " + }, + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "inline_equation", + "content": "\\omega_{p} > 0" + }, + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "text", + "content": " for the replacement positions " + }, + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "inline_equation", + "content": "c_{i} = 1" + }, + { + "bbox": [ + 302, + 621, + 525, + 688 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 698, + 525, + 735 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 698, + 525, + 735 + ], + "spans": [ + { + "bbox": [ + 313, + 698, + 525, + 735 + ], + "type": "interline_equation", + "content": "\\ell_ {P} = - \\sum_ {i = 1} ^ {| \\mathbf {x} |} \\left(1 + c _ {i} \\omega_ {p}\\right) \\log p \\left(\\hat {x} _ {i} ^ {\\prime} = x _ {i} ^ {\\prime} | \\mathbf {x}, \\mathbf {c}\\right) \\tag {4}", + "image_path": "86d13b27c39c92388622cfa1efa27b2dffba40dfada6d95aa7c379e8c76ccd33.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\hat{x}_i^{\\prime}" + }, + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "text", + "content": " is the output of the model and " + }, + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 302, + 745, + 525, + 772 + ], + "type": "text", + "content": " is the supervised paraphrase." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "11361" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 220, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 220, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 220, + 83 + ], + "type": "text", + "content": "3.4 Adversarial Discriminator" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 89, + 291, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 89, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 291, + 155 + ], + "type": "text", + "content": "To train the end-to-end adversarial sample generation model, we pretrain an adversarial discriminator to provide training signals, denote by " + }, + { + "bbox": [ + 67, + 89, + 291, + 155 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 89, + 291, + 155 + ], + "type": "text", + "content": ". It is a trained classifier or a regression model for the target task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 154, + 157, + 290, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 157, + 290, + 170 + ], + "spans": [ + { + "bbox": [ + 154, + 157, + 290, + 170 + ], + "type": "interline_equation", + "content": "y = D (\\mathbf {x}) \\tag {5}", + "image_path": "ff488c5b8113948e09867e4ae12913b456ba6d6a621dbde5f7ea855b80d92810.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": "Theoretically the paraphrasing model should generate the semantically similar text with the original one. If the prediction " + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "inline_equation", + "content": "\\hat{y}^\\prime" + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": " for the generated sample " + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": " is flipped compared to the ground truth " + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 67, + 175, + 290, + 243 + ], + "type": "text", + "content": " is a potentially adversarial sample." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 252, + 246, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 252, + 246, + 279 + ], + "spans": [ + { + "bbox": [ + 67, + 252, + 246, + 279 + ], + "type": "text", + "content": "3.5 End-to-End Adversarial Sample Generation Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 283, + 290, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 283, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 67, + 283, + 290, + 390 + ], + "type": "text", + "content": "The core of the adversarial sample generation model is to generate the appropriate condition for paraphrasing. Since the discriminator can be seen as a white-box victim model, we can pass the gradient through the neural network structure by utilizing the tricks like Gumbel softmax (Jang et al., 2017). This allows an end-to-end training mode for the whole adversarial sample generation method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "content": "During the training, the condition generator, the conditioned paraphrasing model and the adversarial discriminator are connected in a pipeline way. For a given sample " + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "content": ", the pipeline outputs a prediction " + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "content": " for a perturbed sample. The conditioned paraphrasing model " + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "inline_equation", + "content": "F(\\mathbf{x},\\mathbf{c})" + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "content": " outputs a sequence of word probabilities. We adopt Gumbel softmax to convert the probabilities to word embeddings that can be directly used by " + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 67, + 391, + 291, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 522, + 198, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 522, + 198, + 535 + ], + "spans": [ + { + "bbox": [ + 152, + 522, + 198, + 535 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = G (\\mathbf {x})", + "image_path": "a5e468512b57002dcec53d30de9d7301a0bba7ee60290618a82516a222a52b7c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 149, + 537, + 289, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 537, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 149, + 537, + 289, + 551 + ], + "type": "interline_equation", + "content": "\\mathbf {x} ^ {\\prime} = F (\\mathbf {x}, \\mathbf {c}) \\tag {6}", + "image_path": "426276596b047af9903d6d298661cf180abf3ee6e62e2f63a55ad9a9fb503335.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 153, + 554, + 202, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 554, + 202, + 569 + ], + "spans": [ + { + "bbox": [ + 153, + 554, + 202, + 569 + ], + "type": "interline_equation", + "content": "\\hat {y} = D (\\mathbf {x} ^ {\\prime})", + "image_path": "2c8aa34a78998ced2fa7ad612aac0c23436ade42b19bb392be0dd562593482ce.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 576, + 290, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 576, + 290, + 684 + ], + "spans": [ + { + "bbox": [ + 67, + 576, + 290, + 684 + ], + "type": "text", + "content": "During training, the parameters of the condition generator and the conditioned paraphrasing model are tuned while the adversarial discriminator remains frozen. Training the condition paraphrasing model allows the model to be trained adaptively to generate words more suitable for the target task. To train the adversarial sample generation model, we used the following losses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "Adversarial loss is the main goal of the model, namely to generate a sample that can make the target model flip its prediction. Since previous works on model distilling show the effectiveness and simplicity of the MSE loss function (Hinton et al., 2015), thus it is adopted in our work. More" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "content": "advanced distances between probabilities such as KL-distance can be an alternative." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 108, + 525, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 108, + 525, + 124 + ], + "spans": [ + { + "bbox": [ + 348, + 108, + 525, + 124 + ], + "type": "interline_equation", + "content": "\\ell_ {a d v} = 1 - \\left| D \\left(\\mathbf {x} ^ {\\prime}\\right) - D (\\mathbf {x}) \\right| ^ {2} \\tag {7}", + "image_path": "77175bcf38abaadda12f8f6bf2be91ae18ec20603d279767c341e941b15d4fa6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "spans": [ + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "content": "Perturbation ratio loss To maintain the similar semantics between the perturbed sample and the original sample, the ratio of perturbed words in text are constrained, namely the number of " + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "inline_equation", + "content": "c_{i} = 1" + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "content": " cases in the generated condition " + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "content": " denote the expected ratio, " + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 302, + 135, + 525, + 229 + ], + "type": "text", + "content": " is the length of sample, the perturbation ratio loss is calculated as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 239, + 525, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 239, + 525, + 255 + ], + "spans": [ + { + "bbox": [ + 364, + 239, + 525, + 255 + ], + "type": "interline_equation", + "content": "\\ell_ {c o n} = \\left| \\rho - \\| \\mathbf {c} \\| _ {1} / L \\right| ^ {2} \\tag {8}", + "image_path": "30642270165eddf5ff036fc0947f22afd735db05a62e483c52619dbca43939af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "spans": [ + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "text", + "content": "Conditioned paraphrasing losses The conditioned paraphrasing model is fine-tuned for different expectation by a few loss functions. The first is the reconstruction loss for the words not perturbed in " + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "text", + "content": ", i.e. on the positions " + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "inline_equation", + "content": "c_{i} = 0" + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(0)" + }, + { + "bbox": [ + 302, + 266, + 525, + 375 + ], + "type": "text", + "content": ". It is a negative likelihood loss for the original words such that the paraphrasing model is trained to reconstruct the original input." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 385, + 525, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 385, + 525, + 413 + ], + "spans": [ + { + "bbox": [ + 336, + 385, + 525, + 413 + ], + "type": "interline_equation", + "content": "\\ell_ {r e c} = - \\sum_ {i \\in \\mathbf {c} (0)} \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right) \\tag {9}", + "image_path": "3ece6873cdda86820fb7871c9bac4697a9a683c3d9fc276c78d7ea5805584f4d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "spans": [ + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "text", + "content": "For the perturbed positions, the probability of the original word should be lower than any other words. Inspired by the loss function for the negative samples in the noise contrastive estimation (Gutmann and Hyvarinen, 2010), we design the loss of punishing repeating the original word. Let " + }, + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{c}(1)" + }, + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "text", + "content": " denote the positions of the 1s in the condition " + }, + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 302, + 424, + 525, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 525, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 525, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 525, + 559 + ], + "type": "interline_equation", + "content": "\\ell_ {p u n} = - \\sum_ {i \\in \\mathbf {c} (1)} \\sigma (- \\log p \\left(\\mathbf {x} _ {i} ^ {\\prime} = \\mathbf {x} _ {i} \\mid \\mathbf {x}, \\mathbf {c}\\right)) \\tag {10}", + "image_path": "ee883575cf1b4c8cc38f09245a933cd074c08551ca69addb553e0446da446e96.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 577, + 524, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 577, + 524, + 657 + ], + "spans": [ + { + "bbox": [ + 302, + 577, + 524, + 657 + ], + "type": "text", + "content": "Semantics similarity and syntactic correctness losses To maintain the semantics of the generated paraphrase similar with the original sample, we adopt a word-level similarity loss based on cosine similarities for the word embeddings between the generated word and the original word." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 667, + 450, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 667, + 450, + 682 + ], + "spans": [ + { + "bbox": [ + 345, + 667, + 450, + 682 + ], + "type": "interline_equation", + "content": "\\ell_ {s e m} = 1 - \\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right)", + "image_path": "76868b16b6b5b4fa0f92d22a48bfa8cbcd98066ee8fe5c4ef4d30c6a5dbc8ee8.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 686, + 524, + 723 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 686, + 524, + 723 + ], + "spans": [ + { + "bbox": [ + 318, + 686, + 524, + 723 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\mathbf {x}, \\mathbf {x} ^ {\\prime}\\right) = \\frac {1}{| \\mathbf {x} |} \\sum_ {i = 1} ^ {| x |} \\left(1 - \\cos \\left(\\mathbf {x} _ {i}, \\mathbf {x} _ {i} ^ {\\prime}\\right)\\right) \\tag {11}", + "image_path": "0791fc864d3d8c35f16482ec39dac9bfbdb1425eabbe9f0aa836ace7086f2525.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 733, + 525, + 772 + ], + "type": "text", + "content": "The word embeddings similarity cannot guarantee a proper syntactical role of the replacement word. Thus, we also adopt the word composable" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11362" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 69, + 114, + 290, + 141 + ], + "blocks": [ + { + "bbox": [ + 68, + 69, + 289, + 97 + ], + "lines": [ + { + "bbox": [ + 68, + 69, + 289, + 97 + ], + "spans": [ + { + "bbox": [ + 68, + 69, + 289, + 97 + ], + "type": "text", + "content": "Algorithm 1 The Training Process for End-to-End Adversarial Sample Generator" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 69, + 100, + 173, + 113 + ], + "lines": [ + { + "bbox": [ + 69, + 100, + 173, + 113 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 173, + 113 + ], + "type": "text", + "content": "Input: Training set " + }, + { + "bbox": [ + 69, + 100, + 173, + 113 + ], + "type": "inline_equation", + "content": "D_{t}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "lines": [ + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "spans": [ + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "type": "text", + "content": "Output: Condition generator " + }, + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "type": "text", + "content": ", conditioned paraphrasing model " + }, + { + "bbox": [ + 69, + 114, + 290, + 141 + ], + "type": "inline_equation", + "content": "F" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 71, + 142, + 290, + 439 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 76, + 142, + 278, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 142, + 278, + 154 + ], + "spans": [ + { + "bbox": [ + 76, + 142, + 278, + 154 + ], + "type": "text", + "content": "1: //Pretrain conditioned paraphrasing model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "type": "text", + "content": "2: Generate paraphrasing training data " + }, + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 74, + 155, + 289, + 180 + ], + "type": "inline_equation", + "content": "D_{t}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "spans": [ + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "text", + "content": "3: Pretrain " + }, + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 74, + 182, + 254, + 195 + ], + "type": "text", + "content": " according to Eq.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 195, + 247, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 195, + 247, + 207 + ], + "spans": [ + { + "bbox": [ + 74, + 195, + 247, + 207 + ], + "type": "text", + "content": "4: //Pretrain adversarial discriminator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "spans": [ + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "type": "text", + "content": "5: Pretrain " + }, + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 75, + 209, + 173, + 221 + ], + "type": "inline_equation", + "content": "D_{t}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 222, + 290, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 222, + 290, + 248 + ], + "spans": [ + { + "bbox": [ + 74, + 222, + 290, + 248 + ], + "type": "text", + "content": "6: //Train the end-to-end adversarial sample generation model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 75, + 249, + 209, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 249, + 209, + 262 + ], + "spans": [ + { + "bbox": [ + 75, + 249, + 209, + 262 + ], + "type": "text", + "content": "7: Freeze the parameters of " + }, + { + "bbox": [ + 75, + 249, + 209, + 262 + ], + "type": "inline_equation", + "content": "D" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "spans": [ + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "text", + "content": "8: for sample " + }, + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "inline_equation", + "content": "\\mathbf{x}, y" + }, + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 75, + 264, + 198, + 276 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 277, + 166, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 277, + 166, + 289 + ], + "spans": [ + { + "bbox": [ + 75, + 277, + 166, + 289 + ], + "type": "text", + "content": "9: Let " + }, + { + "bbox": [ + 75, + 277, + 166, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = G(\\mathbf{x})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 71, + 290, + 248, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 290, + 248, + 303 + ], + "spans": [ + { + "bbox": [ + 71, + 290, + 248, + 303 + ], + "type": "text", + "content": "10: Calculate " + }, + { + "bbox": [ + 71, + 290, + 248, + 303 + ], + "type": "inline_equation", + "content": "\\ell_{con}" + }, + { + "bbox": [ + 71, + 290, + 248, + 303 + ], + "type": "text", + "content": " according to Eq.8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 71, + 304, + 180, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 304, + 180, + 317 + ], + "spans": [ + { + "bbox": [ + 71, + 304, + 180, + 317 + ], + "type": "text", + "content": "11: Let " + }, + { + "bbox": [ + 71, + 304, + 180, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{x}' = F(\\mathbf{x},\\mathbf{c})" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 317, + 289, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 317, + 289, + 343 + ], + "spans": [ + { + "bbox": [ + 71, + 317, + 289, + 343 + ], + "type": "text", + "content": "12: Calculate " + }, + { + "bbox": [ + 71, + 317, + 289, + 343 + ], + "type": "inline_equation", + "content": "\\ell_{rec},\\ell_{pun},\\ell_{sem},\\ell_{syn}" + }, + { + "bbox": [ + 71, + 317, + 289, + 343 + ], + "type": "text", + "content": " according to Eq.9-12" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 71, + 344, + 170, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 344, + 170, + 357 + ], + "spans": [ + { + "bbox": [ + 71, + 344, + 170, + 357 + ], + "type": "text", + "content": "13: Let " + }, + { + "bbox": [ + 71, + 344, + 170, + 357 + ], + "type": "inline_equation", + "content": "\\hat{y} = D(\\mathbf{x}^{\\prime})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 71, + 358, + 248, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 358, + 248, + 370 + ], + "spans": [ + { + "bbox": [ + 71, + 358, + 248, + 370 + ], + "type": "text", + "content": "14: Calculate " + }, + { + "bbox": [ + 71, + 358, + 248, + 370 + ], + "type": "inline_equation", + "content": "\\ell_{adv}" + }, + { + "bbox": [ + 71, + 358, + 248, + 370 + ], + "type": "text", + "content": " according to Eq.7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 71, + 372, + 256, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 372, + 256, + 385 + ], + "spans": [ + { + "bbox": [ + 71, + 372, + 256, + 385 + ], + "type": "text", + "content": "15: Calculate " + }, + { + "bbox": [ + 71, + 372, + 256, + 385 + ], + "type": "inline_equation", + "content": "L_{adv}" + }, + { + "bbox": [ + 71, + 372, + 256, + 385 + ], + "type": "text", + "content": " according to Eq.13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "spans": [ + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "type": "text", + "content": "16: Update " + }, + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "type": "inline_equation", + "content": "G,F" + }, + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "type": "text", + "content": " according to " + }, + { + "bbox": [ + 71, + 386, + 242, + 397 + ], + "type": "inline_equation", + "content": "L_{adv}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 71, + 399, + 123, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 399, + 123, + 410 + ], + "spans": [ + { + "bbox": [ + 71, + 399, + 123, + 410 + ], + "type": "text", + "content": "17: end for" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "type": "text", + "content": "18: return Condition generator " + }, + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "type": "text", + "content": ", conditioned paraphrasing model " + }, + { + "bbox": [ + 71, + 412, + 289, + 439 + ], + "type": "inline_equation", + "content": "F" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "spans": [ + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": "knowledge based loss (Liu et al., 2023) to provide syntactical correctness guidance for the generated sample. Given a dependency relationship between two words " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "< u, r, v >" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": ", the word composable knowledge estimates the syntactical soundness of this combination, denoted by " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "p_s(u, r, v)" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": ". It relies on the pretrained word embeddings and thus is easy to integrate with other neural network structures. Let " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "s_x = (i, r, j)" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": " denote the dependency parsing tree of " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "(i, r, j)" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": " represents the dependency relation between " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": " and its head word " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 67, + 464, + 291, + 626 + ], + "type": "text", + "content": ". The syntactic correctness loss is calculated as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 92, + 637, + 290, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 637, + 290, + 667 + ], + "spans": [ + { + "bbox": [ + 92, + 637, + 290, + 667 + ], + "type": "interline_equation", + "content": "\\ell_ {s y n} = - \\sum_ {(i, r, j) \\in s _ {x}} \\log p _ {s} \\left(\\mathbf {x} _ {i} ^ {\\prime}, r, \\mathbf {x} _ {j} ^ {\\prime}\\right) \\tag {12}", + "image_path": "3d0f33118fd1dd4051f26d874f63ebbd8300285a2b436a00e0db79d6a0fa55e5.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 67, + 678, + 290, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 719 + ], + "type": "text", + "content": "The complete loss function for training the adaptive adversarial sample generation model is the weighted sum of the above losses." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 79, + 730, + 289, + 773 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 730, + 289, + 773 + ], + "spans": [ + { + "bbox": [ + 79, + 730, + 289, + 773 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {a d v} = w _ {a} \\ell_ {a d v} + w _ {c} \\ell_ {c o n} + w _ {r} \\ell_ {r e c} + w _ {p} \\ell_ {p u n} \\\\ + w _ {s e m} \\ell_ {s e m} + w _ {s y n} \\ell_ {s y n} \\tag {13} \\\\ \\end{array}", + "image_path": "b972de79eac150c94c972b52279126300e54e2fcb519a1f11fa5f34de8c174cf.jpg" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 304, + 68, + 532, + 126 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 532, + 126 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 532, + 126 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 532, + 126 + ], + "type": "table", + "html": "
Dataset#Class#Train#TestAve.Len.
IMDB22500025000270.7
Quora23636604040425.33
SNLI35501521000022.4
", + "image_path": "c52d956a59da1b610a511a48ec3034e9dd70abeb75f0c5c14ba30f0691f493d9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 361, + 135, + 466, + 146 + ], + "lines": [ + { + "bbox": [ + 361, + 135, + 466, + 146 + ], + "spans": [ + { + "bbox": [ + 361, + 135, + 466, + 146 + ], + "type": "text", + "content": "Table 1: Dataset statistics" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 168, + 524, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 168, + 524, + 195 + ], + "spans": [ + { + "bbox": [ + 302, + 168, + 524, + 195 + ], + "type": "text", + "content": "The overall training process of the end-to-end adversarial sample generation is listed in Alg.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 302, + 203, + 422, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 203, + 422, + 216 + ], + "spans": [ + { + "bbox": [ + 302, + 203, + 422, + 216 + ], + "type": "text", + "content": "3.6 Data Augmentation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": "Having the above end-to-end adversarial sample generator, we can apply data augmentation on the training data for the target task. For the given original training set " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": ", the corresponding generated adversarial samples is denoted by " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "D_{t}^{\\prime}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": ". Then the model for the target task can be trained on " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "D_{t} \\cup D_{t}^{\\prime}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": " for improving the model robustness. Since the adversarial samples are more likely to introduce noises, their training loss is assigned a lower weight. Let " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "L_{t}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": " denotes the loss of the target task, " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": " denotes the loss weight for sample " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "\\lambda_{i} = 1" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": " for an original sample while " + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "inline_equation", + "content": "\\lambda_{i} < 1" + }, + { + "bbox": [ + 302, + 220, + 525, + 383 + ], + "type": "text", + "content": " for a generated sample." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 351, + 391, + 525, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 391, + 525, + 421 + ], + "spans": [ + { + "bbox": [ + 351, + 391, + 525, + 421 + ], + "type": "interline_equation", + "content": "L _ {t a s k} = \\sum_ {s _ {i} \\in D _ {t} \\bigcup D _ {t} ^ {\\prime}} \\lambda_ {i} L _ {t} (s _ {i}) \\tag {14}", + "image_path": "aba7f516a602926bd801ddf4777035f5f4b3cf6f6ac4947dcd77440e763744db.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 302, + 428, + 525, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 428, + 525, + 469 + ], + "spans": [ + { + "bbox": [ + 302, + 428, + 525, + 469 + ], + "type": "text", + "content": "The previous work has shown the adversarial samples generated based on one model are potentially adversarial to other models (Szegedy et al., 2013)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 302, + 478, + 390, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 478, + 390, + 492 + ], + "spans": [ + { + "bbox": [ + 302, + 478, + 390, + 492 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 302, + 500, + 525, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 500, + 525, + 539 + ], + "spans": [ + { + "bbox": [ + 302, + 500, + 525, + 539 + ], + "type": "text", + "content": "To evaluate the effectiveness of the proposed method, four aspects should be considered and evaluated in the subsections 4.2-4.5." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 302, + 547, + 525, + 663 + ], + "type": "list", + "angle": 0, + "index": 39, + "blocks": [ + { + "bbox": [ + 302, + 547, + 525, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 547, + 525, + 575 + ], + "spans": [ + { + "bbox": [ + 302, + 547, + 525, + 575 + ], + "type": "text", + "content": "A1: To check the attacking effectiveness of generated adversarial sample." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 302, + 581, + 524, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 581, + 524, + 608 + ], + "spans": [ + { + "bbox": [ + 302, + 581, + 524, + 608 + ], + "type": "text", + "content": "A2: To check whether the model performance is improved after augmented training." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 302, + 615, + 524, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 615, + 524, + 643 + ], + "spans": [ + { + "bbox": [ + 302, + 615, + 524, + 643 + ], + "type": "text", + "content": "A3: To check whether the robustness of the target model is improved." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 302, + 650, + 501, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 650, + 501, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 650, + 501, + 663 + ], + "type": "text", + "content": "A4: To check the efficiency of our method." + } + ] + } + ], + "index": 38 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 302, + 672, + 420, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 672, + 420, + 686 + ], + "spans": [ + { + "bbox": [ + 302, + 672, + 420, + 686 + ], + "type": "text", + "content": "4.1 Experiment Setups" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 302, + 689, + 365, + 701 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 689, + 365, + 701 + ], + "spans": [ + { + "bbox": [ + 302, + 689, + 365, + 701 + ], + "type": "text", + "content": "4.1.1 Tasks" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 705, + 525, + 772 + ], + "type": "text", + "content": "Three tasks are considered as evaluation tasks to cover a wider range of NLP tasks such that the effectiveness of data augmentation can be thoroughly evaluated in different scenarios. The first is a sentiment analysis task on the dataset IMDB. The IMDB" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11363" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 68, + 520, + 154 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 520, + 154 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 520, + 154 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 520, + 154 + ], + "type": "table", + "html": "
Base ModelBERTDistilBERTPerturbed %Bert Score
DatasetOri.Acc.Att.Acc.ΔOri.Acc.Att.Acc.Δ
IMDB93.9%66.58%-27.32%92.96%61.76%-31.2%9.95%95.98
Quora84.84%75.95%-8.89%84.32%76.44%-7.88%16.6%95.96
SNLI88.95%55.64%-33.31%88.04%55.20%-32.84%14.73%95.62
", + "image_path": "6bb7a4770d226dfc2bf62cf2ca79206e096a5e9856c45ba8a11937a66ad445d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 196, + 291, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 196, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 67, + 196, + 291, + 344 + ], + "type": "text", + "content": "dataset consists of long movie reviews which are classified as positive of negative sentiment polarity. The second is a text classification task on the dataset Quora Question Pairs(QQP), which consists of question pairs from the Quora site classified as duplicate or not. The last is a natural language inference task on the dataset SNLI. Each sample in SNLI consists a hypothesis sentence and a premise sentence, the logical relationship between them is determined as entailment, neutral or contradiction. The statistics of the datasets are listed in Table.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 354, + 163, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 163, + 365 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 163, + 365 + ], + "type": "text", + "content": "4.1.2 Base Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 370, + 291, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 370, + 291, + 572 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 291, + 572 + ], + "type": "text", + "content": "Recently, pretrained language models are widely used as a fundamental part for NLP methods. In the following experiments, BERT (Kenton and Toutanova, 2019) and DistilBERT (Sanh et al., 2019) are chosen as the base model for the tasks. BERT is a widely used pretrained language model, and achieves improvements on various NLP tasks. DistilBERT is a compressed version of BERT model based on model distilling technique, making it easier to be used on devices with lower computation power. Both models will be finetuned on the original train set or the augmented train set for the target task and evaluated afterwards. Their results will be listed as Baseline in the following experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 581, + 180, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 180, + 594 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 180, + 594 + ], + "type": "text", + "content": "4.1.3 Implementation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 291, + 772 + ], + "type": "text", + "content": "All the experiments are conducted on a single NVIDIA Geforce RTX 2080Ti. Our method is implemented with PyTorch 1.12.0. The code for our proposed method is shared at https://splab.sdu.edu.cn/xscg/sjjydm.htm. The compared methods and adversarial sample generation are implemented by TextAttack (Morris et al., 2020b). The size of hidden vectors in our model is 300. We adopt GloVe (Pennington et al., 2014) as the word embeddings and reused the pretrained parameters of word composable knowledge released in SynPara (Liu et al., 2023). For pretraining the paraphrasing model, the ratio of replaced words is set" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 306, + 194, + 523, + 333 + ], + "blocks": [ + { + "bbox": [ + 88, + 163, + 504, + 175 + ], + "lines": [ + { + "bbox": [ + 88, + 163, + 504, + 175 + ], + "spans": [ + { + "bbox": [ + 88, + 163, + 504, + 175 + ], + "type": "text", + "content": "Table 2: Accuracy comparison between accuracies on the original testing set and adversarial testing set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 194, + 523, + 333 + ], + "lines": [ + { + "bbox": [ + 306, + 194, + 523, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 523, + 333 + ], + "type": "table", + "html": "
Base ModelMethodDataset
IMDBQuoraSNLI
BERTBaseline93.984.8488.95
SSMBA-0.41+3.03-2.88
A2T-0.03+1.4-0.5
A3+1+4+0.1
Distil-BERTBaseline91.8784.3388.04
SSMBA+1.09+3.79-0.12
A2T+1.27+4.39+1.38
A3+1.93+4.89+0.22
", + "image_path": "7cbc1faa007bc3cd97bc3826084c43ea1ad9982f81d85d515cc32d4fbc935263.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 341, + 525, + 378 + ], + "lines": [ + { + "bbox": [ + 302, + 341, + 525, + 378 + ], + "spans": [ + { + "bbox": [ + 302, + 341, + 525, + 378 + ], + "type": "text", + "content": "Table 3: Accuracy comparison on the original testing set. The values for the augmentation methods are the differences with the accuracy of the base model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "spans": [ + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": " when generating training data. For training the complete model, the weights in the final loss function are set as " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{a} = 5.0" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{c} = 3.0" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{r} = 2.0" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{p} = 5.0" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{sem} = 2.0" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "w_{syn} = 0.1" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": " in most cases. The ratio of perturbed words " + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 398, + 525, + 493 + ], + "type": "text", + "content": " for the condition generator is set as 0.15 for SNLI and 0.1 for IMDB and Quora unless stated otherwise." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 501, + 431, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 501, + 431, + 513 + ], + "spans": [ + { + "bbox": [ + 302, + 501, + 431, + 513 + ], + "type": "text", + "content": "4.1.4 Compared methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 301, + 517, + 525, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 517, + 525, + 679 + ], + "spans": [ + { + "bbox": [ + 301, + 517, + 525, + 679 + ], + "type": "text", + "content": "For the comparison methods, we consider the following methods. SSMBA (Ng et al., 2020) is a data augmentation method that uses the corruption function similar to the masked language model training and reconstructs the samples with BERT. A2T (Yoo and Qi, 2021) is an adversarial training method, where a fast adversarial sample generation method uses the gradients for identifying important words in a sample and replaces them based on word embeddings or MLM models. It iteratively trains the task model and augment the training data set with the generated adversarial samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 688, + 427, + 700 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 688, + 427, + 700 + ], + "spans": [ + { + "bbox": [ + 302, + 688, + 427, + 700 + ], + "type": "text", + "content": "4.2 Attack Effectiveness" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 706, + 525, + 772 + ], + "type": "text", + "content": "To evaluate the effectiveness of the generated adversarial samples, we compared the performances of the baseline models on the original testing data and on the adversarial data. The results are shown in Table.2, where the ratio of perturbed words in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11364" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 68, + 526, + 317 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 526, + 317 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 526, + 317 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 526, + 317 + ], + "type": "table", + "html": "
Base ModelBERTDistilBERT
DatasetsIMDBQuoraSNLIIMDBQuoraSNLI
Attack MethodAugment Method%Δ%Δ%Δ%Δ%Δ%Δ
A3Baseline69.377.358.364.180.159.8
SSMBA71.6+2.377.7+0.458.7+0.461.6-2.578.8-1.362.9+3.1
A2T73.7+4.476.9-0.459.7+1.466.1+278.6-1.560.3+0.5
A392.2+22.989.3+1286.2+27.993.8+29.790.9+10.886.5+26.7
A2TBaseline38.840.144.130.650.341.1
SSMBA38.9+148.1+849.3+5.227.9-2.750.7+0.445.1+4
A2T44+5.276.2+36.179.4+35.354.5+23.976.1+25.875.6+34.5
A345.3+6.560.7+20.648+3.936.1+5.560.2+9.943.3+2.2
Text-FooterBaseline1.723.60.90.110.30.47
SSMBA5.3+3.612.5-11.10.4-0.50-0.112.1+1.80.73+0.26
A2T0.8-0.914.3-9.35.1+4.21.3+1.28-2.33.93+3.46
A312+10.327.1+3.51.9+11.6+1.515.2+4.90.67+0.2
BAEBaseline3139.322.124.939.323.87
SSMBA30.6-0.439-0.325+2.921-3.939.7+0.423.93+0.06
A2T29.1-1.943.4+4.123.5+1.426.3+1.444.2+4.925.5+1.63
A337.9+6.940.8+1.527.6+5.527.7+2.845.7+6.424.27+0.4
PWWSBaseline125.11.20.822.11.07
SSMBA2.2+1.223.1-22.3+1.10.1-0.722.9+0.81.2+0.13
A2T0.3-0.726.3+1.20.7-0.50.6-0.223+0.90.87-0.2
A37.3+6.327.7+2.62.4+1.20.9+0.124.1+22.4+1.33
", + "image_path": "119e918b24bad9a9f349391ff8ab646e760f1a109cfe713af6d6a0b3e325cb2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 325, + 525, + 351 + ], + "lines": [ + { + "bbox": [ + 67, + 325, + 525, + 351 + ], + "spans": [ + { + "bbox": [ + 67, + 325, + 525, + 351 + ], + "type": "text", + "content": "Table 4: Comparison of attacking methods against augmentation methods. Column % denotes the model accuracy under attacking. Column " + }, + { + "bbox": [ + 67, + 325, + 525, + 351 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 67, + 325, + 525, + 351 + ], + "type": "text", + "content": " denotes the difference between augmented model with the original model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 370, + 291, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 370, + 291, + 519 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 291, + 519 + ], + "type": "text", + "content": "the adversarial samples is listed in column Perturbed%. A lower ratio indicates more structure information of the text preserved. BertScore (Zhang et al., 2020) is used for quantifying the semantics similarity between the original sample and the adversarial sample. The higher the BertScore, the closer their semantics. We can see that the model performances of both baseline models drop significantly on the adversarial samples, which show that the generated adversarial samples are effective on attacking the baseline models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 535, + 221, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 535, + 221, + 548 + ], + "spans": [ + { + "bbox": [ + 67, + 535, + 221, + 548 + ], + "type": "text", + "content": "4.3 Performance Improvement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 556, + 291, + 772 + ], + "type": "text", + "content": "We then evaluated the model performance on the target tasks before and after training with augmented data. By using our method and SSMBA, the data is augmented to doubled size of the original training set, respectively. The augmented dataset is used for training the base models. For A2T, it iteratively trains the model and then augments the training set by generating adversarial samples. The results in Table.3 show that our method improves the performance on all settings and are better than the compared methods in most cases. The compared methods are unstable on performances, namely in some cases performances are the improved while some are not. For example, SSMBA has negative impact on SNLI dataset for both base models. This is probably because SSMBA doesn't take into con" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 370, + 526, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 370, + 526, + 479 + ], + "spans": [ + { + "bbox": [ + 302, + 370, + 526, + 479 + ], + "type": "text", + "content": "sideration the task-specific characteristics, such as the logical connections between the hypothesis and premise in SNLI. It should be also noted that while the original performance of DistilBERT is slightly lower than BERT, it improves a lot after data augmentation, which confirmed previous studies regarding the effectiveness of data augmentation for distilled models (Das et al., 2020)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 495, + 450, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 495, + 450, + 507 + ], + "spans": [ + { + "bbox": [ + 302, + 495, + 450, + 507 + ], + "type": "text", + "content": "4.4 Robustness Improvement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "content": "To evaluate the robustness of base model against different attacking methods, we choose three representative attacking methods: TextFooler (Jin et al., 2020), BAE (Garg and Ramakrishnan, 2020) and PWWS (Ren et al., 2019). We adopt the notion " + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "inline_equation", + "content": "m1 - m2" + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "content": " denoting the attacking method " + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "inline_equation", + "content": "m1" + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "content": " against augmentation method " + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "inline_equation", + "content": "m2" + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "content": ". For example, BAE-A2T means using BAE to attack a model trained using A2T. Similar to the settings of previous works (Yoo and Qi, 2021), 1000 samples are selected from the original testing set of each dataset (1500 samples from SNLI for three labels), where the label distributions are balanced. To better show the impact of an attacking method, the selected samples can be predicted correctly by the base models before data augmentation. The attack methods generate the adversarial samples based on the selected samples and the accuracy on these adversarial samples of the trained model would be compared. A2T and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11365" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 68, + 271, + 139 + ], + "blocks": [ + { + "bbox": [ + 87, + 68, + 271, + 139 + ], + "lines": [ + { + "bbox": [ + 87, + 68, + 271, + 139 + ], + "spans": [ + { + "bbox": [ + 87, + 68, + 271, + 139 + ], + "type": "table", + "html": "
MethodsDataset
IMDBQuoraSNLI
SSMBA10m38m1h2m
A2T36h(3)111h(2)70h(2)
A34m27m21m
", + "image_path": "3d3b32ca86f4160b89f1e76a25b93ac24f3cd01c012ea471653d8ac4438e6527.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 148, + 290, + 195 + ], + "lines": [ + { + "bbox": [ + 67, + 148, + 290, + 195 + ], + "spans": [ + { + "bbox": [ + 67, + 148, + 290, + 195 + ], + "type": "text", + "content": "Table 5: Time for augmenting the dataset. It should be noted that the time for A2T is the cumulated time for multiple training epochs (number of epochs is listed in the brackets)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 86, + 206, + 273, + 324 + ], + "blocks": [ + { + "bbox": [ + 86, + 206, + 273, + 324 + ], + "lines": [ + { + "bbox": [ + 86, + 206, + 273, + 324 + ], + "spans": [ + { + "bbox": [ + 86, + 206, + 273, + 324 + ], + "type": "image", + "image_path": "a019b2b244879cc0e9658e0a2870775cec70d24b3a78a269374655290b3cb541.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "lines": [ + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "spans": [ + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "type": "text", + "content": "Figure 2: The influence of " + }, + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "type": "text", + "content": " on the perturbed ratio and the attack effectiveness (denoted by " + }, + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{Acc}" + }, + { + "bbox": [ + 67, + 332, + 289, + 356 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 379, + 289, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 379, + 289, + 404 + ], + "spans": [ + { + "bbox": [ + 67, + 379, + 289, + 404 + ], + "type": "text", + "content": "our proposed method are tested as they are also adversarial sample generation methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "text", + "content": "The results are listed in Table.4, where the column " + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "text", + "content": " denotes the model accuracy under attacking and column " + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "text", + "content": " shows the difference between the performance of the augmented model with the baseline. As explored in a previous research (Carlini et al., 2019), the robustness against a certain attack method will be improved if trained with the corresponding adversarial samples. From the row " + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^3\\mathrm{-A}^3" + }, + { + "bbox": [ + 67, + 406, + 291, + 581 + ], + "type": "text", + "content": " and A2T-A2T, we can see that both methods significantly improve the model robustness. For other representative attacking methods, the augmented models with our generated data show better robustness than the compared methods in most settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 592, + 202, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 592, + 202, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 202, + 606 + ], + "type": "text", + "content": "4.5 Efficiency Comparison" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 611, + 291, + 772 + ], + "type": "text", + "content": "Another vital aspect of adversarial training is the efficiency. Most adversarial sample generation methods involves constantly interaction with the target model. Such process is both computational heavy and time-consuming. We recorded the time used for augmenting the training set of different methods. Both SSMBA and our proposed method are one-time augmentation thus the time taken is similar. Our method is slightly faster than SSMBA since SSMBA uses BERT as the reconstruction model. The A2T method, on the other hand, is significantly slower than our method due to two possible rea" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 193 + ], + "type": "text", + "content": "son. First reason is that the attacking process is conducted multiple times. Secondly, the method is based on attacking the target model which is not only time-consuming, but the time taken will also increase over time since the model is trained better after each epoch. Comparably, our proposed method generates model-agnostic adversarial samples, such that the time for augmented training is significantly shortened." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 205, + 423, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 205, + 423, + 217 + ], + "spans": [ + { + "bbox": [ + 302, + 205, + 423, + 217 + ], + "type": "text", + "content": "4.6 Parameter Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": "We analyze the effect of parameter " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": ", the targeted perturbed ratio. In our method for a " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " setting, the actual perturbed ratio in the generated adversarial samples may be different with the initial value. Thus, we analyze the actual perturbed ratio and its impact on the attacking effectiveness, shown in Fig.2. The blue and orange lines represent the set " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " and the actual perturbed ratio respectively, corresponding to the left Y-axis. The " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{Acc}" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " is the accuracy difference after the attack, corresponding to the right Y-axis. Higher value of " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{Acc}" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " means more effective attack. We can see that the actual perturbed ratio approaches the " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " setting. A higher " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " may cause the predicated label flipped but may lead to a larger semantics difference. Thus, the parameter " + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 302, + 223, + 526, + 453 + ], + "type": "text", + "content": " should be set properly to achieve a balanced performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 465, + 381, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 465, + 381, + 478 + ], + "spans": [ + { + "bbox": [ + 302, + 465, + 381, + 478 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 488, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 488, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 488, + 526, + 772 + ], + "type": "text", + "content": "In this paper, we present a method named " + }, + { + "bbox": [ + 302, + 488, + 526, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^3" + }, + { + "bbox": [ + 302, + 488, + 526, + 772 + ], + "type": "text", + "content": " that combines the advantages of adversarial training and data augmentation to enhance the robustness of NLP models. We introduce an end-to-end adversarial sample generation model which consists of a condition generator and a conditioned paraphrasing model. A pretrained adversarial discriminator is proposed to train the adversarial sample generation model for capturing the task-specific data characteristics. This method is much efficient without the need for extensive model interactions. The generated adversarial samples are incorporated with the original training samples by a weighted loss for augmented training downstream models. Experiments are conducted on multiple NLP tasks and demonstrate the effectiveness and efficiency of our proposed method. Compared to existing methods, the performances of the base models have improved more after our data augmentation, as well as their robustness enhanced against multiple attacking techniques." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11366" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 93, + 293, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 93, + 293, + 296 + ], + "spans": [ + { + "bbox": [ + 67, + 93, + 293, + 296 + ], + "type": "text", + "content": "Firstly, the current form of condition limits the diversity and fine-grained control over the generation process. For example, a condition can be a subtree from the dependency syntax structure, which allows paraphrases with potential structure reform and correct syntax (Liu et al., 2023). Secondly, a condition may also contain the direction of replacement word for the paraphrasing model. This would allow reusing the same conditioned paraphrasing model across different tasks, thereby reducing the training overhead for adversarial sample generation. Lastly, in future research, we aim to explore the feasibility of directly using the gradient from the target model for training the adversarial sample generation model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 308, + 166, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 308, + 166, + 322 + ], + "spans": [ + { + "bbox": [ + 68, + 308, + 166, + 322 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 331, + 291, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 331, + 291, + 386 + ], + "spans": [ + { + "bbox": [ + 67, + 331, + 291, + 386 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (62376138) and the Innovative Development Joint Fund Key Projects of Shandong NSF (ZR2022LZH007)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 409, + 127, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 409, + 127, + 423 + ], + "spans": [ + { + "bbox": [ + 68, + 409, + 127, + 423 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 429, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 69, + 429, + 291, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 429, + 291, + 476 + ], + "spans": [ + { + "bbox": [ + 69, + 429, + 291, + 476 + ], + "type": "text", + "content": "Yoshua Bengio, Nicholas Léonard, and Aaron Courville. 2013. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 486, + 291, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 486, + 291, + 541 + ], + "spans": [ + { + "bbox": [ + 69, + 486, + 291, + 541 + ], + "type": "text", + "content": "Nicholas Carlini, Anish Athalye, Nicolas Papernot, Wieland Brendel, Jonas Rauber, Dimitris Tsipras, Ian Goodfellow, Aleksander Madry, and Alexey Kurakin. 2019. On evaluating adversarial robustness. arXiv preprint arXiv:1902.06705." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 551, + 291, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 551, + 291, + 618 + ], + "spans": [ + { + "bbox": [ + 69, + 551, + 291, + 618 + ], + "type": "text", + "content": "Minhao Cheng, Jinfeng Yi, Pin-Yu Chen, Huan Zhang, and Cho-Jui Hsieh. 2020. Seq2sick: Evaluating the robustness of sequence-to-sequence models with adversarial examples. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3601-3608." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 629, + 290, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 629, + 290, + 662 + ], + "spans": [ + { + "bbox": [ + 69, + 629, + 290, + 662 + ], + "type": "text", + "content": "Cheng-Han Chiang and Hung-yi Lee. 2022. How far are we from real synonym substitution attacks? arXiv preprint arXiv:2210.02844." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 672, + 290, + 717 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 672, + 290, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 672, + 290, + 717 + ], + "type": "text", + "content": "Deepan Das, Haley Massa, Abhimanyu Kulkarni, and Theodoros Rekatsinas. 2020. An empirical analysis of the impact of data augmentation on knowledge distillation. CoRR, abs/2006.03810." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 291, + 772 + ], + "type": "text", + "content": "Xinshuai Dong, Anh Tuan Luu, Rongrong Ji, and Hong Liu. 2021. Towards robustness against natural language word substitutions. arXiv preprint arXiv:2107.13541." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 526, + 116 + ], + "type": "text", + "content": "Javid Ebrahimi, Anyi Rao, Daniel Lowd, and De- jing Dou. 2017. Hotflip: White-box adversarial examples for text classification. arXiv preprint arXiv:1712.06751." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 125, + 526, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 181 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 181 + ], + "type": "text", + "content": "Siddhant Garg and Goutham Ramakrishnan. 2020. Bae: Bert-based adversarial examples for text classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6174-6181." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 190, + 526, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 526, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 526, + 234 + ], + "type": "text", + "content": "Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. 2014a. Generative adversarial networks. CoRR, abs/1406.2661." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 243, + 526, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 526, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 526, + 278 + ], + "type": "text", + "content": "Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014b. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 286, + 526, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 526, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 526, + 331 + ], + "type": "text", + "content": "Shreya Goyal, Sumanth Doddapaneni, Mitesh M Khapra, and Balaraman Ravindran. 2022. A survey of adversarial defences and robustness in nlp. ACM Computing Surveys." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 340, + 526, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 526, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 526, + 407 + ], + "type": "text", + "content": "Michael Gutmann and Aapo Hyvarinen. 2010. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297-304. JMLR Workshop and Conference Proceedings." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 416, + 526, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 526, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 526, + 449 + ], + "type": "text", + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 459, + 526, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 526, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 526, + 558 + ], + "type": "text", + "content": "Po-Sen Huang, Robert Stanforth, Johannes Welbl, Chris Dyer, Dani Yogatama, Sven Gowal, Krishnamurthy Dvijotham, and Pushmeet Kohli. 2019. Achieving verified robustness to symbol substitutions via interval bound propagation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 4083-4093." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 566, + 526, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 526, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 526, + 612 + ], + "type": "text", + "content": "Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017). OpenReview.net." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 526, + 687 + ], + "type": "text", + "content": "Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is bert really robust? a strong baseline for natural language attack on text classification and entailment. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 8018-8025." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 696, + 526, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 696, + 526, + 740 + ], + "spans": [ + { + "bbox": [ + 304, + 696, + 526, + 740 + ], + "type": "text", + "content": "Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 750, + 526, + 772 + ], + "type": "text", + "content": "Ashutosh Kumar, Satwik Bhattamishra, Manik Bhandari, and Partha Talukdar. 2019. Submodular" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "11367" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 139 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 139 + ], + "type": "text", + "content": "optimization-based diverse paraphrasing and its effectiveness in data augmentation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3609-3619." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 148, + 290, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 148, + 290, + 183 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 290, + 183 + ], + "type": "text", + "content": "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. AI Open, 3:71-90." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 192, + 290, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 192, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 69, + 192, + 290, + 269 + ], + "type": "text", + "content": "Dianqi Li, Yizhe Zhang, Hao Peng, Liquin Chen, Chris Brockett, Ming-Ting Sun, and William B Dolan. 2021. Contextualized perturbation for textual adversarial attack. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5053-5069." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 280, + 290, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 280, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 290, + 324 + ], + "type": "text", + "content": "J Li, S Ji, T Du, B Li, and T Wang. 2019. Textbugger: Generating adversarial text against real-world applications. In 26th Annual Network and Distributed System Security Symposium." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 334, + 290, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 334, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 290, + 390 + ], + "type": "text", + "content": "Linyang Li, Ruotian Ma, Qipeng Guo, Xiangyang Xue, and Xipeng Qiu. 2020. Bert-attack: Adversarial attack against using bert. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6193–6202." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 400, + 290, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 400, + 290, + 444 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 290, + 444 + ], + "type": "text", + "content": "Linyang Li and Xipeng Qiu. 2021. Token-aware virtual adversarial training in natural language understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 8410-8418." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 454, + 290, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 454, + 290, + 509 + ], + "spans": [ + { + "bbox": [ + 69, + 454, + 290, + 509 + ], + "type": "text", + "content": "Tianyuan Liu, Yuqing Sun, Jiaqi Wu, Xu Xi, Yuchen Han, Cheng Li, and Bin Gong. 2023. Unsupervised paraphrasing under syntax knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 37." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 519, + 290, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 519, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 69, + 519, + 290, + 565 + ], + "type": "text", + "content": "Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. 2020. Adversarial training for large neural language models. arXiv preprint arXiv:2004.08994." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 575, + 290, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 575, + 290, + 618 + ], + "spans": [ + { + "bbox": [ + 69, + 575, + 290, + 618 + ], + "type": "text", + "content": "Takeru Miyato, Andrew M Dai, and Ian Goodfellow. 2016. Adversarial training methods for semi-supervised text classification. arXiv preprint arXiv:1605.07725." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 629, + 290, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 629, + 290, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 629, + 290, + 685 + ], + "type": "text", + "content": "John Morris, Eli Lifland, Jack Lanchantin, Yangfeng Ji, and Yanjun Qi. 2020a. Reevaluating adversarial examples in natural language. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3829-3839." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020b. Textattack: A framework for adversarial attacks, data augmentation, and adversarial training in nlp. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119-126." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 614 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 139 + ], + "type": "text", + "content": "Nathan Ng, Kyunghyun Cho, and Marzyeh Ghassemi. 2020. Ssmba: Self-supervised manifold based data augmentation for improving out-of-domain robustness. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1268-1283." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 147, + 525, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 525, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 525, + 213 + ], + "type": "text", + "content": "Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, Doha, Qatar. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 221, + 525, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 525, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 525, + 287 + ], + "type": "text", + "content": "Shuhuai Ren, Yihe Deng, Kun He, and Wanxiang Che. 2019. Generating natural language adversarial examples through probability weighted word saliency. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 1085-1097." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 296, + 525, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 525, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 525, + 340 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 349, + 525, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 525, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 525, + 393 + ], + "type": "text", + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 401, + 525, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 525, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 525, + 446 + ], + "type": "text", + "content": "Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. 2013. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 454, + 525, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 525, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 525, + 488 + ], + "type": "text", + "content": "Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 496, + 525, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 525, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 525, + 540 + ], + "type": "text", + "content": "Jin Yong Yoo and Yanjun Qi. 2021. Towards improving adversarial training of nlp models. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 945-956." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 549, + 525, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 525, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 525, + 614 + ], + "type": "text", + "content": "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with BERT. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "11368" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_content_list.json b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c13f9b63efad2a4dab48a4360983b44ea8673964 --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_content_list.json @@ -0,0 +1,1727 @@ +[ + { + "type": "text", + "text": "Energy and Carbon Considerations of Fine-Tuning BERT", + "text_level": 1, + "bbox": [ + 200, + 83, + 800, + 104 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaorong Wang $^{1*}$", + "bbox": [ + 263, + 114, + 416, + 131 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "swangxr@gmail.com", + "bbox": [ + 247, + 133, + 421, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clara Na $^{2*}$", + "bbox": [ + 616, + 114, + 712, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "csna@cs.cmu.edu", + "bbox": [ + 589, + 133, + 742, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emma Strubell $^{2,3}$", + "bbox": [ + 169, + 162, + 322, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strubell@cmu.edu", + "bbox": [ + 166, + 181, + 329, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sorelle A. Friedler1", + "bbox": [ + 413, + 164, + 581, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sorelle@cs.haverford.edu", + "bbox": [ + 376, + 181, + 620, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sasha Luccioni4", + "bbox": [ + 677, + 164, + 818, + 179 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sasha.luccioni@hf.co", + "bbox": [ + 648, + 181, + 850, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Haverford College, $^{2}$ Carnegie Mellon University, $^{3}$ Allen Institute for AI, $^{4}$ Hugging Face", + "bbox": [ + 136, + 212, + 858, + 230 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the popularity of the pre-train then finetune paradigm in the NLP community, existing work quantifying energy costs and associated carbon emissions has largely focused on language model pre-training. Although a single pre-training run draws substantially more energy than fine-tuning, fine-tuning is performed more frequently by many more individual actors, and thus must be accounted for when considering the energy and carbon footprint of NLP. In order to better characterize the role of finetuning in the landscape of energy and carbon emissions in NLP, we perform a careful empirical study of the computational costs of finetuning across tasks, datasets, hardware infrastructure and measurement modalities. Our experimental results allow us to place fine-tuning energy and carbon costs into perspective with respect to pre-training and inference, and outline recommendations to NLP researchers and practitioners who wish to improve their finetuning energy efficiency.", + "bbox": [ + 141, + 282, + 460, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 609, + 258, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fine-tuning pre-trained language models is a frequent occurrence in natural language processing (NLP) research and practice, yet the vast majority of work quantifying the energy and carbon footprints of NLP workloads has focused on pretraining (Strubell et al., 2019; Dodge et al., 2022; Luccioni et al., 2023) or inference (Desislavov et al., 2023; Luccioni et al., 2023). The typical lifecycle of an NLP model includes data ingestion, pre-training, fine-tuning and inference, all of which contribute non-trivially to energy use and corresponding emissions (Patterson et al., 2021; Wu et al., 2022). Better understanding of the role that each phase plays in overall energy and carbon footprint is vital to inform policy decisions, yet we still lack basic data quantifying the rela", + "bbox": [ + 112, + 637, + 489, + 895 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tive contributions due to different aspects of model development and use (Kaack et al., 2022).", + "bbox": [ + 507, + 253, + 880, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we perform an empirical study to quantify the energy requirements of language model fine-tuning, including in the context of pretraining and inference energy requirements. While this may seem like it should be a straightforward calculation, there are several variables that can influence compute time and energy consumption, ranging from: (1) the type of hardware used for both pre-training and fine-tuning (since this usually differs between the two), (2) the type of task and the type of computation required to carry it out, and (3) intrinsic characteristics of the dataset, such as average sequence length, its similarity with the pre-training dataset, etc.", + "bbox": [ + 507, + 287, + 884, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to isolate the factors that have the most influence on fine-tuning dynamics, we compare fine-tuning energy use across a suite of common supervised NLP datasets including the tasks of entailment, sentiment analysis, question answering, and named entity recognition, and with training data sizes ranging from 6K to 400K examples. We also measure energy use across two different sets of hardware, using the CodeCarbon (Schmidt et al., 2021) software package and a physical energy measurement device at the wall, to quantify variance due to physical factors. To enable carefully controlled comparison of the roles of pre-training and fine-tuning in NLP model lifecycle energy use, we additionally pre-train BERT variants from scratch on the same hardware. We find that pre-training BERT is equivalent to anywhere from 400 (MNLI) to 45,000 (RTE) fine-tuning runs depending on the dataset size, and that number of training tokens1 is a reasonable heuristic for estimating fine-tuning energy use. Further comparison of fine-tuning in", + "bbox": [ + 507, + 514, + 884, + 853 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1The \"true\" number of training tokens seen, accounting for dynamic padding of sequences to the maximum length in a batch, is a better predictor than relying on to mean or median number of tokens per example.", + "bbox": [ + 507, + 869, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Denotes equal contribution.", + "bbox": [ + 142, + 904, + 317, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9058", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9058-9069", + "bbox": [ + 216, + 945, + 779, + 958 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 277, + 958, + 719, + 971 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ference energy intensity across tasks confirms that example sequence length holds a much stronger influence on energy intensity in the fine-tuning phase than in the inference phase, in alignment with expectations from previous work (Zhou et al., 2021). Together, our observations contextualize the energy and carbon requirements of fine-tuning in the broader model lifecycle and highlight the need to study fine-tuning energy efficiency separately from pre-training and inference workloads in NLP models. We hope that our careful measurement of the relative costs of different NLP workloads will serve as a valuable datapoint informing decision-making both within and beyond the NLP community.", + "bbox": [ + 112, + 84, + 492, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 112, + 325, + 270, + 341 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Measurement of energy consumption and carbon emissions of NLP models has become an active area of research since it was first identified that modern state-of-the-art models based on deep learning can produce substantial greenhouse gas emissions due to the energy required to train them (Strubell et al., 2019; Schwartz et al., 2020). These measurements have mostly focused on two research directions. First, there has been a series of empirical studies on different model architectures, focused on estimating the carbon emissions generated by their training process and the relative efficiency of different methods (Naidu et al., 2021; Patterson et al., 2021, 2022). Recent work by Luccioni et al. has built upon this, aiming to encompass the embodied emissions of manufacturing computing hardware as well as those produced via the inference process (2023). There has also been complementary work measuring the energy used by Transformer models during inference and ways of predicting those costs for different architectures and models (Cao et al., 2021; Ang et al., 2022).", + "bbox": [ + 112, + 355, + 490, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Closest to our work, the HULK benchmark (Zhou et al., 2021) was proposed to measure the relative efficiency-accuracy trade-offs of different pre-trained models, measuring the wall-clock time for different pre-trained models to reach a target accuracy on one of three fine-tuning tasks. Different from Zhou et al. (2021), our work explicitly focuses on the energy and carbon required for fine-tuning (theirs uses time and financial cost as proxies), evaluates a wider variety of fine-tuning tasks and hardware settings in order to elucidate the factors that predict fine-tuning energy requirements, and further contextualizes fine-tuning energy re", + "bbox": [ + 112, + 709, + 492, + 919 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "requirements in the bigger picture of ML model lifecycle emissions.", + "bbox": [ + 507, + 84, + 884, + 115 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another related direction of research examines the dynamics of pre-training and fine-tuning of language models and the influence of factors like random seeds and early stopping (Dodge et al., 2020), scaling (Tay et al., 2022) and learning dynamics (Hao et al., 2020). While all of these studies have shed important light on these processes, in practice most of the decisions made remain empirical, with practitioners either referring to previous work (when hyperparameters and other training details are reported), or using techniques such as grid or random search (Bergstra and Bengio, 2012) to converge on optimal parameter values. Our own work builds upon both of these research directions. We study both the pre-training and fine-tuning process, and our experiments for studying their energy intensity are based on the works cited above.", + "bbox": [ + 507, + 116, + 885, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 Methodology", + "text_level": 1, + "bbox": [ + 507, + 401, + 658, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Full training details can be found in Appendix A.1. We release code for replicating our measurements.2 We encourage others to run our code on their hardware to add to a repository of measurements across different hardware platforms.", + "bbox": [ + 507, + 428, + 885, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1 Pre-training BERT", + "text_level": 1, + "bbox": [ + 507, + 520, + 705, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work we are interested in measuring the energy consumption of fine-tuning, as it compares to other stages of model use: pre-training and inference. In order to establish a comparable baseline for the energy consumption and carbon emissions of pre-training, we pre-trained a BERT-base model (Devlin et al., 2019) from scratch on the same hardware that we use for fine-tuning (Section 3.3). Although in practice pre-training and fine-tuning are often done separately and on different hardware, we fixed the machine for both sets of experiments in order to aid direct comparability of energy usage measurements. Following Devlin et al. (2019), we pre-train our model on the BookCorpus (Zhu et al., 2015) and the 2020 version of Wikipedia (Foundation, 2020), both downloaded from HuggingFace Datasets (Lhoest et al., 2021).", + "bbox": [ + 507, + 541, + 885, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our precise pre-training methodology differs slightly from Devlin et al. (2019): our data necessarily differs slightly because the original training corpus was not released along with the model, and we only use the masked language modeling", + "bbox": [ + 507, + 815, + 885, + 896 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://github.com/swangxr/FT-energy.git", + "bbox": [ + 529, + 903, + 845, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9059", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(MLM) objective without next sentence prediction (NSP) following Liu et al. (2019), who found that removing NSP did not substantially impact endtask performance.", + "bbox": [ + 112, + 84, + 487, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to assess the relative impact of using a more efficient pre-trained BERT variant, we also followed the DistilBERT (Sanh et al., 2019) recipe, performing knowledge distillation on our pre-trained BERT-base model.", + "bbox": [ + 112, + 149, + 487, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 Fine-tuning BERT", + "text_level": 1, + "bbox": [ + 112, + 239, + 307, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We evaluate the energy consumption and carbon emissions of the fine-tuning process on the tasks in Table 1. We deliberately chose this selection of end-tasks in order to vary fine-tuning dataset size, task type, and sequence length, while also aligning with tasks commonly used in NLP applications.", + "bbox": [ + 112, + 260, + 487, + 356 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ef8e05d8a07fff594948bedfce455772696e2e63ce02598badf7fb28843d3ee8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTaskExamplesSeq. length
Med.Batch
Wiki+BooksMLM43M128128
RTENLI6K56128
MNLINLI433K3790
SQuADv1QA98K159336
SQuADv2QA142K158336
IMDBSent.50K128128
SST2Sent.70K1040
CoNLL2003NER21K1553
CoNLL2012NER143K2065
", + "bbox": [ + 127, + 366, + 473, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1: Pre-training and fine-tuning dataset descriptions. We report two sequence length statistics: Median tokens per sequence in the dataset, and Batch, the average maximum sequence length per batch as predicted by simulated sampling. The latter is included as a more direct predictor of computation cost for dynamic padding.", + "bbox": [ + 112, + 530, + 489, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All models are fine-tuned on the BERT models described in §3.1. For each fine-tuning task, we use typical fine-tuning hyperparameters specific to the task or user-reported hyperparameters on current fine-tuned models on HuggingFace, in order to mimic the common real-world use cases. For each task, we dynamically pad sequences to the maximum length in each batch. All fine-tuning hyperparameters are reported in Appendix A.1.", + "bbox": [ + 112, + 631, + 487, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also report average per-example energy use for inference. All the inference tasks are performed on 1 GPU with batch size 1 on the same machines.", + "bbox": [ + 112, + 776, + 487, + 822 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3 Hardware Platforms", + "text_level": 1, + "bbox": [ + 112, + 834, + 322, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ensure reproducibility and measure variability across hardware platforms, we replicate experiments across two hardware platforms: One A100 machine and one RTX 8000 machine, where each", + "bbox": [ + 112, + 854, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "machine had four GPUs. Pre-training experiments used all 4 GPUs in each machine.", + "bbox": [ + 507, + 84, + 880, + 115 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All fine-tuning tasks were performed on the same machines, but using only one GPU. This reflects the typical scenario where fine-tuning is done on a single GPU even if the machine itself has more GPUs. To better compare the energy usage results across pre-training and fine-tuning, we also report an energy usage estimate for BERT-base pretraining on 1 RTX 8000 GPU with hyperparameters equivalent to training on 4 GPUs, extrapolated from a 12-hour partial training run. Details are recorded in Appendix A.1.", + "bbox": [ + 507, + 117, + 882, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4 Measuring Energy and Carbon", + "text_level": 1, + "bbox": [ + 507, + 303, + 801, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To measure the energy consumed, we use the software tool CodeCarbon (Schmidt et al., 2021). Recent work has found that the existing libraries and code for estimating the carbon emissions of NLP techniques vary in terms of their accuracy and generalizability to different types of hardware (Bannour et al., 2021). To compensate for this, we calibrate the programmatic energy usage readings with a physical energy meter, with which we record energy readings directly from the compute node during experiments. Subsequently, we calculate a coefficient of expected power loss, 1.059, as the average proportion (over runs across fine-tuning tasks) of physical energy reading vs. programmatic energy measurement. Full results are given in Appendix A.2. Thus, the energy consumed in kWh, denoted as $E$ , is determined via the formula:", + "bbox": [ + 507, + 323, + 882, + 596 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE (\\mathrm {k W h}) = 1. 0 5 9 \\cdot \\text {c o d e c a r b o n} \\mathrm {k W h}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 604, + 835, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Converting the power loss adjusted values to $\\mathrm{CO}_{2}$ emissions is done through CodeCarbon using a coefficient specific to the energy grid based on the location of the server from the most recent EPA Power Profiler Data. The conversion factor for the server's location (Pittsburgh, PA) in Table 2 is 1046.1 lbs/MWh, while the factor for the second server's location (Haverford, PA) is 672.8 lbs/MWh. The total kilograms of $\\mathrm{CO}_{2}$ emitted, denoted as $C$ , is then determined via:", + "bbox": [ + 507, + 626, + 882, + 785 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = E \\times 1 0 4 6. 1 \\frac {\\mathrm {l b}}{\\mathrm {M W h}} \\times \\frac {1 \\mathrm {M W h}}{1 0 0 0 \\mathrm {k W h}} \\times \\frac {1 \\mathrm {k g}}{2 . 2 0 4 6 2 \\mathrm {l b}}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 789, + 880, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We convert the $\\mathrm{CO}_{2}$ emissions result to human understandable reference values using the EPA Greenhouse Gas Equivalencies Calculator; in Table 2, we also show the equivalent $\\mathrm{CO}_{2}$ emissions of miles driven by an average gasoline-powered passenger vehicle.", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9060", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/09b794f912e76e30c995fc3aa36df1d53b851f63b9c3999047e54287ad42b091.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDatasetTraining / Fine-tuningInference
Energy (kWh)Emissions (kg CO2)TimeEquiv. # PT runsEquiv. milesEnergy (kWh) / 1000 ex.Equiv. # FT runs
MLM4 GPU (1m @ len 128)270.9128.627 hr1.358330——
1 GPU (1m @ len 128)419.6199.1673 hr0.646510——
4 GPU (100k @ len 512)124.158.90114 hr2.965151——
Total 4 GPU368.0174.6357 hr1.000408——
NLIRTE0.0080.00459 s451090.010.794e-31k
MNLI0.9380.4456700 s3921.147.349e-3127k
QASQuAD v10.5370.2553780 s6850.652.157e-3249k
SQuAD v20.7950.3775604 s4630.972.471e-3322k
Sent.IMDB0.1510.0721074 s24410.180.849e-3178k
SST20.0810.038587 s45400.100.691e-312k
NERCoNLL20030.0210.010149 s178860.030.808e-33k
CoNLL20120.2070.0981487 s17780.250.878e-324k
", + "bbox": [ + 122, + 80, + 873, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 Results and Discussion", + "text_level": 1, + "bbox": [ + 112, + 420, + 344, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2 shows energy, carbon, and wall-clock time required to fine-tune BERT-base models on the RTX8000 machine. Results on the A100 machine are recorded in Appendix A.3 in Tables 8 and 9.", + "bbox": [ + 112, + 445, + 485, + 510 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/b4a5b5a6b6781ce7f3c64e5b4c66a768dcf7ea46e4f92a9bf3182f9e1eed190b.jpg", + "table_caption": [ + "Table 2: Energy consumption of pre-training BERT and fine-tuning on the RTX8000 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent $\\mathrm{CO}_{2}$ emissions. 1 GPU pre-training costs are extrapolated from a shorter pre-training run lasting only a few hours. Total cost of fine-tuning is derived from 900k steps of pre-training on sequences of length 128 followed by 100k steps on sequences of length 512, following (Devlin et al., 2019). Energy consumption for inference is calculated using single-example batches." + ], + "table_footnote": [], + "table_body": "
TaskDatasetEnergy (kWh)TimeEmiss. (kg CO2)
Distil.Wiki+Books187.74175.5 hr89.08
NLIRTE0.00430 s0.002
MNLI0.4813447 s0.228
QASQuAD v10.2761954 s0.131
SQuAD v20.4122916 s0.196
Sent.IMDB0.077549 s0.037
SST0.042306 s0.020
NERCoNLL20030.01178 s0.005
CoNLL20120.107762 s0.051
", + "bbox": [ + 117, + 521, + 485, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3: Energy consumption of training (distillation) and fine-tuning DistilBERT on the RTX8000 machine. Energy calculations are the same as in Table 2. Distillation is performed on a pre-trained model, and so the true \"total\" cost includes the pre-training cost as well.", + "bbox": [ + 112, + 686, + 487, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1 Pre-training and Distillation", + "text_level": 1, + "bbox": [ + 112, + 785, + 381, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that it requires an additional $50\\%$ of the energy cost of pre-training in order to perform knowledge distillation, but it takes nearly $50\\%$ less energy to fine-tune on the same tasks using DistilBERT vs. normal BERT (see Table 3). By our estimate, one can fully amortize the up-front cost of distillation within anywhere from 86 fine-tuning", + "bbox": [ + 112, + 806, + 487, + 917 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "runs of an MNLI-like task, to 47k fine-tunings on an RTE-like task. DistilBERT fine-tuning results on the A100 machine are in Appendix A.3.", + "bbox": [ + 507, + 420, + 880, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Comparing and Predicting Fine-tuning Emissions", + "text_level": 1, + "bbox": [ + 507, + 494, + 862, + 524 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We find that, controlling for hardware, energy consumption scales most predictably with wall clock time and number of tokens encountered during training (including the pad tokens added to sequences to match the maximum sequence length in a batch). The linear relationship between energy consumption and total number of tokens holds similarly on both machines (see Figure 1). Additionally, we observe a consistently higher energy consumption in the RTX 8000 GPU machine. This is likely due to the higher energy overhead and the (in)efficiency of the hardware compared to the A100 GPUs. Other figures in Appendix A.3 illustrate that, in contrast, energy requirements as a function of optimization steps or even number of examples in the dataset can vary significantly across datasets and tasks.", + "bbox": [ + 505, + 538, + 882, + 810 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3Note that cheaper inference is often the primary goal of knowledge distillation. Inference is much cheaper than training and therefore requires more to amortize the initial cost of distillation, but inference also occurs much more frequently than training. Models running inference at scale are typically highly optimized with respect to specific deployment settings, so our estimates approximate a lower bound.", + "bbox": [ + 507, + 832, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9061", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5cc94fe344b22ab526108be9eb3d4542409fbbb51295da988f13c0f7a05264c.jpg", + "image_caption": [ + "Figure 1: Total energy consumed (kWh) is strongly related to number of tokens seen for BERT models on both A100 and RTX 8000 GPU machines, although the relationship is more predictive on the RTX 8000 machine and energy usage is less consistent with DistilBERT. Note that both axes are in log scale. An alternative view of similar data in Figure 4 distinguishes pre-training workloads' energy consumption slightly from that of fine-tuning tasks." + ], + "image_footnote": [], + "bbox": [ + 119, + 84, + 495, + 267 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1085d6f94d0c10db282f6f816c9f26fafe1527c50b5c4555746981f04826cd54.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4: Inference vs. fine-tuning energy requirements across end tasks. We see that fine-tuning energy usage varies according to sequence length much more widely than inference energy usage." + ], + "table_body": "
DatasetSeq. lengthkWh / 1k ex.
Med.BatchInf.FT
RTE (NLI)561280.75e-31.09e-3
MNLI (NLI)37900.70e-30.80e-3
SQUADv1 (QA)1593361.08e-33.04e-3
SQUADv2 (QA)1583361.24e-33.02e-3
IMDB (Sent.)1281280.80e-31.21e-3
SST2 (Sent.)10400.65e-30.40e-3
CoNLL2003 (NER)15530.76e-30.49e-3
CoNLL2012 (NER)20650.83e-30.60e-3
", + "bbox": [ + 115, + 424, + 485, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3 Fine-tuning vs Pre-training", + "text_level": 1, + "bbox": [ + 112, + 655, + 376, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Even for the more reliable predictors of energy consumption and carbon emissions (duration of training and number of tokens processed), the energy cost profiles of pre-training vs. fine-tuning are different, likely due to differences in training infrastructure, training objectives, and sequence lengths typically seen in pre-training vs. fine-tuning (see Figures 4, 3, 5, and 6). Pre-training in general is almost always performed over multiple GPUs which incurs energy costs from communication between GPUs, and often also with gradient accumulation to accommodate large batches. Moreover, sequences are packed together such that batches consist largely or entirely of sequences of identical length equal to the maximum sequence length for", + "bbox": [ + 112, + 677, + 489, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the model.", + "bbox": [ + 509, + 84, + 591, + 98 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, there are many types of finetuning tasks where examples consist of sequences of varying lengths significantly shorter than the maximum length that the model has been trained on, as shown in Table 1. Since, effective sequence lengths are determined dynamically at training time (where sequences are padded to the maximum length in each given batch), total training time is not as simple to extrapolate from standard measures of dataset size as in pre-training.", + "bbox": [ + 507, + 99, + 882, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4 Fine-tuning vs. Inference", + "text_level": 1, + "bbox": [ + 507, + 272, + 754, + 288 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although we do observe that per-example inference efficiency costs are related to sequence lengths, there is overall less variation across datasets and tasks in inference costs compared to fine-tuning costs (see Table 4). This mirrors an observation noted in the HULK benchmark (Zhou et al., 2021), though to the best of our knowledge ours is the first to explicitly draw comparisons across task types and different aspects of dataset size (i.e. number of examples and examples' sequence lengths).", + "bbox": [ + 507, + 292, + 884, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5 Single vs. Multiple GPUs", + "text_level": 1, + "bbox": [ + 507, + 464, + 752, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general, typical hardware and data settings for pre-training and fine-tuning tend to differ. Though to the best of our knowledge it is less common to fine-tune causal LMs of this scale on multiple GPUs, we present additional results from multi-GPU fine-tuning on the $4 \\times$ RTX8000 machine with the same fine-tuning tasks in Table 10 in Appendix A.3. Our recommendations from an energy efficiency standpoint align with common rules of thumb for effective utilization of hardware; if the resources would be idle otherwise, one could reasonably consider increasing batch size and learning rate to saturate the available hardware for both time- and energy-efficient training.", + "bbox": [ + 507, + 485, + 882, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 721, + 640, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We share a procedure for rigorous measurement of energy consumption from causal LM fine-tuning given multiple concrete hardware settings. We hope our work is useful to researchers and practitioners who are interested in obtaining measurements for their own specific hardware, gaining intuitions about factors affecting relative energy costs of different types of fine-tuning tasks, or understanding these energy costs in context of the model lifecycle.", + "bbox": [ + 507, + 747, + 884, + 907 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9062", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 114, + 84, + 220, + 99 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While our work provides important first steps towards a clearer understanding of model finetuning's impact on the environment, we note that our experimentation is limited to various token classification, sequence classification, and question answering tasks with BERT-base and DistilBERT-base models. We do not make claims or extrapolations about much larger language models, or models with different architectures, as well as other types of tasks such as summarization. Future work in this direction can expand the number of tasks that we consider as well as feature different architectures such as RoBERTa (Liu et al., 2019).", + "bbox": [ + 112, + 111, + 489, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, the on-premises hardware infrastructure used for our experimentation is realistic and typical of compute resources in academic settings, but we provide no firsthand evidence of fine-tuning emissions profiles expected from either local model training (where the impracticality of pretraining makes direct comparisons with fine-tuning emissions difficult) or fine-tuning on hardware that is part of much larger scale infrastructure such as on a public cloud. Furthermore, we expect that use of specialized hardware such as TPUs (as opposed to GPUs, which we use) would be associated with different emissions profiles.", + "bbox": [ + 115, + 321, + 489, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 114, + 544, + 265, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and deploying ML models and systems requires large amounts of energy, the production of which results in the emission of greenhouse gases. While the goal of our research is to contribute towards a better understanding of the factors that influence these emissions, by carrying out our experiments, we were ourselves responsible for the emission of $350\\mathrm{kg}$ of carbon equivalents. We release our code and the data generated by our experiments to maximize the transparency and reproducibility of our work.", + "bbox": [ + 112, + 571, + 489, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 114, + 763, + 285, + 778 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We are grateful to Victor Sanh for his time and patience in answering questions about BERT and DistilBERT pre-training. We would also like to thank our anonymous reviewers for taking the time to provide helpful feedback.", + "bbox": [ + 112, + 789, + 487, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This work was supported in part by a grant from the National Science Foundation Graduate Research Fellowship Program under Grant No.", + "bbox": [ + 112, + 871, + 489, + 917 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DGE2140739. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors.", + "bbox": [ + 507, + 84, + 885, + 149 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 510, + 175, + 608, + 190 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Phyllis Ang, Bhuwan Dhingra, and Lisa Wu Wills. 2022. Characterizing the efficiency vs. accuracy trade-off for long-context NLP models. In Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP, pages 113-121, Dublin, Ireland. Association for Computational Linguistics.", + "Nesrine Bannour, Sahar Ghannay, Aurélie Néveol, and Anne-Laure Ligozat. 2021. Evaluating the carbon footprint of NLP methods: a survey and analysis of existing tools. In EMNLP, Workshop SustainNLP.", + "James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(10):281-305.", + "Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. 2021. IrEne: Interpretable energy prediction for transformers. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2145-2157, Online. Association for Computational Linguistics.", + "Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. 2023. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. *Sustainable Computing: Informatics and Systems*, 38:100857.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. 2020. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305.", + "Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the carbon intensity of ai in cloud instances. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1877-1894.", + "Wikipedia Foundation. 2020. Wikipedia downloads." + ], + "bbox": [ + 510, + 197, + 885, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9063", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yaru Hao, Li Dong, Furu Wei, and Ke Xu. 2020. Investigating learning dynamics of bert fine-tuning. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 87-92.", + "Lynn H. Kaack, Priya L. Donti, Emma Strubell, George Kamiya, Felix Creutzig, and David Rolnick. 2022. Aligning artificial intelligence with climate change mitigation. Nature Climate Change, 12(6):518-527.", + "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach.", + "Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. 2023. Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253):1-15.", + "Rakshit Naidu, Harshita Diddee, Ajinkya Mulay, Aleti Vardhan, Krithika Ramesh, and Ahmed Zamzam. 2021. Towards quantifying the carbon emissions of differentially private machine learning. arXiv preprint arXiv:2107.06946.", + "David Patterson, Joseph Gonzalez, Urs Holzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. The carbon footprint of machine learning training will plateau, then shrink. Computer, 55(7):18-28.", + "David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350.", + "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + ], + "bbox": [ + 115, + 85, + 487, + 917 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Ziyao Wang, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luccioni. 2021. Codecarbon: estimate and track carbon emissions from machine learning computing.", + "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2020. Green AI. Communications of the ACM, 63(12):54-63.", + "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics.", + "Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2022. Scale efficiently: Insights from pretraining and finetuning transformers. In International Conference on Learning Representations.", + "C.-J. Wu et al. 2022. Sustainable AI: Environmental implications, challenges, and opportunities. In $MLSys$ .", + "Xiyou Zhou, Zhiyu Chen, Xiaoyong Jin, and William Yang Wang. 2021. HULK: An energy efficiency benchmark platform for responsible natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 329-336, Online. Association for Computational Linguistics.", + "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV)." + ], + "bbox": [ + 510, + 85, + 880, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 510, + 653, + 631, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A.1 Pre-training and Fine-tuning Details", + "text_level": 1, + "bbox": [ + 509, + 681, + 843, + 697 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For pre-training BERT with MLM, we mainly follow what is listed in Devlin et al. (2019). The specific hyperparameters used for pre-training are listed in Table 5 and were used on both machines.", + "bbox": [ + 507, + 703, + 880, + 766 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8fb56ce7f7d25c63b5549d17b6e36b2ee7ca994a7eb1c4daa16edb30c778efdd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyperparameters
maximum steps1000000
training batch size per device64
evaluation batch size per device64
maximum sequence length128
learning rate1e-4
warmup steps10000
weight decay0.01
", + "bbox": [ + 527, + 777, + 818, + 890 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5: Hyperparameters used for pre-training BERT", + "bbox": [ + 510, + 898, + 878, + 914 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9064", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 6, we list the set of hyperparameters used to fine-tune each task. All fine-tuning tasks were run on both machines.", + "bbox": [ + 112, + 84, + 487, + 131 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hardware details The A100 machine is located in Haverford, Pennsylvania, USA, and has 4x NVIDIA A100 GPUs with 40GB GDDR SDRAM, 376GB main memory and 32 Intel Xeon processors. The RTX 8000 machine is located in Pittsburgh, Pennsylvania, USA, and has 4x NVIDIA Quadro RTX 8000 GPUs with 48GB GDDR SDRAM, 36 Intel Xeon processors and 251GB RAM.", + "bbox": [ + 112, + 149, + 489, + 278 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.2 Kill-A-Watt Measurements", + "text_level": 1, + "bbox": [ + 112, + 297, + 378, + 312 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Energy is lost during the process of transferring energy from a power source to the machine. The coefficient for the loss is acquired through the readings from Kill-A-Watt devices. The device measures the instantaneous Watt reading extracted from the wall and displays on the monitor. The measurements for the A100 machine were recorded only on the single node of the cluster containing it (the RTX8000 machine is not part of a larger cluster). The GPU node is connected to two outlets, and we plug separate Kill-A-Watt devices into both. For each instantaneous reading, we read off of and sum up the readings on both Kill-A-Watt devices. To best compare between the package reading and the wall reading, we read off of the device at the same 15 second interval that CodeCarbon records the energy consumed. To convert each instantaneous Watts readings into Kilowatt-Hours, we follow the formula:", + "bbox": [ + 112, + 323, + 489, + 627 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {k W h} = \\frac {1 5 \\text {s e c o n d s}}{3 6 0 0 \\text {s e c o n d s}} \\times \\frac {\\text {w a t t s} \\times \\text {h o u r s}}{1 0 0 0}\n$$\n", + "text_format": "latex", + "bbox": [ + 149, + 643, + 450, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We sum up all the calculations for the entire run of a fine-tuning experiment. Then, we can compare the sum of the wall readings with the Code Carbon energy consumption recording. We divide the wall reading over the package reading to get a coefficient, measuring the more realistic energy used. The setup of an instantaneous reading on the A100 machine is shown in Figure 2. The recordings of the wall readings on the A100 GPU machine are recorded in Table 7.", + "bbox": [ + 112, + 690, + 489, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ran these wall reading experiments for both machines and obtained a coefficient of 1.09 on the A100 machine, and a coefficient of 1.05 on the RTX machine.", + "bbox": [ + 112, + 854, + 487, + 917 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23cc7a984d5ac1e8681ad75dba598a06618c597141aee5221814227c57bce57f.jpg", + "image_caption": [ + "Figure 2: Kill-A-Watt wall reading measurement setup" + ], + "image_footnote": [], + "bbox": [ + 510, + 80, + 885, + 156 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A.3 Additional Results", + "text_level": 1, + "bbox": [ + 507, + 205, + 705, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "BERT pre-training and fine-tuning results Table 8 records the energy consumption results from the A100 GPU machine, located in Haverford, PA. The conversion factor for this location is 672.8 lbs $\\mathrm{CO}_{2} / \\mathrm{MWh}$ . We then calculate the emissions using raw output from CodeCarbon, which is listed in Table 8.", + "bbox": [ + 507, + 225, + 884, + 336 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DistilBERT results Table 9 records the energy consumption and emission on the A100 GPU machine. Distillation is not done on this machine, and all the fine-tunings are done on the DistilBERT trained on the RTX8000 machine. The results on the A100 machine show a similar trend that fine-tunings on DistilBERT takes around $50\\%$ less energy than on BERT base models.", + "bbox": [ + 507, + 348, + 885, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8747192e9dcea7336391a18c5da5097b74439f1090a7ab3d2a05676346772234.jpg", + "image_caption": [ + "Figure 3: Energy consumed (kWh) against total tokens seen on A100 machine" + ], + "image_footnote": [], + "bbox": [ + 515, + 491, + 890, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training duration and energy consumption From Figure 5 and 6, we see that there is a strong correlation between training time and energy consumption, which holds across our models and hardware settings (Figure 7). However, similarly to tokens seen, we observe that pre-training exhibits a slightly different energy consumption profile, as do question answering tasks on the A100 machine (which have the longest sequences out of our finetuning tasks). When training time estimates are not feasible in advance (such as in certain hyperparam", + "bbox": [ + 507, + 741, + 885, + 919 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9065", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1b804e445f59f54cd7022bbe897ffc35609000e2fe76a50125d3d5611f2462f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RTEMNLIIMDBSST2SQuAD v1SQuAD v2CoNLL 2003CoNLL 2012
epochs33532233
train bs3232163232323232
eval bs3232163232323232
max seq len128128128128384384128128
doc stride128128
lr2e-52e-52e-52e-53e-53e-55e-55e-5
", + "bbox": [ + 141, + 80, + 857, + 227 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e9a9d1959441b1d862867b3e6c53725a3f3789be8a1499bd881beee54194c2ce.jpg", + "table_caption": [ + "Table 6: Hyperparameters used for fine-tuning tasks" + ], + "table_footnote": [], + "table_body": "
TaskRandom seedWall reading converted (kWh)Code Carbon reading (kWh)Power Loss Co-efficient
SST2 (no AC)420.1030.0881.170
SST21230.0590.0551.057
IMDB420.0700.0611.134
IMDB420.0790.0711.122
IMDB1230.07930.0731.092
RTE420.006410.006021.065
", + "bbox": [ + 200, + 263, + 800, + 388 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/1594d6e89606593a7a75284f38d3a87508dafecfb78750a33285ba8925ddfaee.jpg", + "table_caption": [ + "Table 7: Kill-A-Watt experiment recordings for the A100 GPU machine" + ], + "table_footnote": [], + "table_body": "
TaskDatasetEnergy (kWh)Emissions (kg CO2)Training time (s)Equiv. # PT runsEquivalent emissions of miles driven
MLMWiki + Books146.15044.6027265531144
fine-tuningRTE0.0020.00219584420.005
MNLI0.3560.23924004060.613
SQuAD v10.1710.11510618460.295
SQuAD v20.2630.17715805490.454
IMDB0.0620.019047821140.048
SST20.0510.03437728580.087
CoNLL20030.0110.00778134110.018
CoNLL20120.0910.06165415840.156
", + "bbox": [ + 136, + 424, + 860, + 590 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/a97a18319bf5047ab66da40d676a4c1643f9ce02ccb37454545aca1b6ac32dd2.jpg", + "table_caption": [ + "Table 8: Energy consumption of pre-training BERT and fine-tuning on the A100 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent $\\mathrm{CO}_{2}$ emissions. Note that here, unlike in Table 2, the \"baseline\" pre-training cost is fixed to the cost of 1 million steps of masked language modeling of sequences of length 128." + ], + "table_footnote": [], + "table_body": "
TaskDatasetEnergyTimeEmiss.
EntailmentRTE0.001100.001
MNLI0.21015390.141
QASQuAD v10.0915660.061
SQuAD v20.1419150.095
SentimentIMDB0.0453410.012
SST0.0342620.023
NERCoNLL20030.007570.005
CoNLL20120.0594480.040
", + "bbox": [ + 119, + 694, + 482, + 818 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 9: Energy consumption of training (distillation) and fine-tuning DistilBERT on the A100 machine. Units and energy calculations are the same as given in Table 8.", + "bbox": [ + 112, + 827, + 489, + 871 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "eter sweeps), we recommend that researchers and practitioners use token counts estimates (including dynamic padding tokens) if they have reasonable knowledge of their data.", + "bbox": [ + 507, + 697, + 882, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As indicated in Figure 8 and Figure 9, energy increases as the optimization steps increase. This is not surprising given the correlation between training time and energy consumption. However, we see that for different tasks, the energy required for each step is very different. Each step of pre-training takes a longer time, likely due to the higher batch size than all fine-tuning tasks. For QA tasks, the per-step energy consumption is higher than other", + "bbox": [ + 507, + 774, + 884, + 919 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9066", + "bbox": [ + 480, + 928, + 521, + 940 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/e32afc811c7479f199676d9b1914dd4389d41d2b25501b5d722d5edd3cebb7d0.jpg", + "image_caption": [ + "Figure 4: Energy consumed (kWh) against total tokens seen on RTX8000 machine" + ], + "image_footnote": [], + "bbox": [ + 119, + 84, + 495, + 266 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/3c98d08b317fc611f443cdc3e699e81fb24739f221ce843e4c60e38586d646e9.jpg", + "image_caption": [ + "Figure 7: Total training time is strongly predictive of total energy consumed (kWh) for both BERT and DistilBERT models, on both A100 and RTX 8000 GPU machines. Note that both axes are in log scale. Different from token counts, the trendlines themselves are similar as well across models and hardware settings." + ], + "image_footnote": [], + "bbox": [ + 119, + 329, + 495, + 512 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/c9aea71968f5a63e944731649adfcb202f8fa9da9fc67b3758e83eca27f240c1.jpg", + "image_caption": [ + "Figure 5: Energy consumed (kWh) against training time (s) on A100 machine, showing the first 30 minutes" + ], + "image_footnote": [], + "bbox": [ + 119, + 573, + 485, + 766 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "tasks. This is likely due to the maximum sequence length of 384 being higher than for the other tasks.", + "bbox": [ + 112, + 837, + 487, + 869 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 10 shows the number of examples of the task and the energy consumed in log scale. Similar to Figure 1, we see a direct correlation between", + "bbox": [ + 112, + 871, + 487, + 917 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f002ed9f7b98d360afc2f3150ee0f3e9031c9fff200615ceea9a4b79653b4005.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 515, + 85, + 890, + 267 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f42988e7016a356666720e5b891c9f3f1ab238a3f11b4c21ad8dee8e78e6c64a.jpg", + "image_caption": [ + "Figure 8: Energy consumed (kWh) against optimization steps on A100 machine machine" + ], + "image_footnote": [], + "bbox": [ + 521, + 385, + 870, + 579 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/fe29ab73b8c948511d25e7107c86a087e409a5522fed2563a589be714454e0d4.jpg", + "image_caption": [ + "Figure 6: Energy consumed (kWh) against training time (s) on RTX8000 machine, showing the first 30 minutes", + "Figure 9: Energy consumed (kWh) against optimization steps on RTX8000 machine" + ], + "image_footnote": [], + "bbox": [ + 519, + 639, + 872, + 832 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9067", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2f26e3758d0114bec7d6f2f0e0062c9ba1dce8585899e7e842bd698bf6648e64.jpg", + "image_caption": [ + "Figure 10: Energy consumed (kWh) against total number of examples seen on both A100 and RTX8000 machines. Comparing with Figure 1, we see that merely counting number of training examples is much less predictive of energy consumption than accounting for example sequence lengths along with batch size (which affects the maximum sequence length in each batch)" + ], + "image_footnote": [], + "bbox": [ + 119, + 84, + 495, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "higher number of training examples and higher energy usage on both machines.", + "bbox": [ + 112, + 406, + 487, + 437 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/12fafe6f47ae862e264f38b522c97bacfcf9ce2919844c5c644a9abf412dbe7c.jpg", + "image_caption": [ + "Figure 11: $\\mathrm{CO}_{2}$ emissions (kg) against task accuracy for SST-2 fine-tuning on RTX 8000 GPU machine" + ], + "image_footnote": [], + "bbox": [ + 124, + 451, + 473, + 642 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Figure 11 shows the accuracy of SST2 task as $\\mathrm{CO}_{2}$ emissions (kg) increases. As shown previously in Figure 6, energy consumption increases as time increases. Generally, as emission increases, accuracy increases. We can see that as emissions goes up, accuracy is trending towards converging.", + "bbox": [ + 112, + 701, + 487, + 797 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Single- vs. multi-GPU fine-tuning We fine-tune BERT-base and DistilBERT-base models on the RTX8000 machine as well. If the same hyperparameters as the single-GPU setting are used naively (i.e. the same batch size is split over 4 GPUs), fine-tuning using 4 GPUs can (but does not always) take even longer than using just 1 GPU, and can (but", + "bbox": [ + 112, + 806, + 489, + 917 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/8262f4339c02c4b224cb879747f1fac398eaf5ca7a2fb0c1cdaa8d9c56baf70f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDuration (s)Energy (kWh)
1GPU4GPU1GPU4GPU
BERT-base
RTE (NLI)59278.16e-38.52e-3
MNLI (NLI)670030149.38e-18.64e-1
SQuADv1 (QA)378013565.37e-14.28e-1
SQuADv2 (QA)560420117.95e-16.16e-1
IMDB (Sent.)10744721.51e-11.32e-1
SST2 (Sent.)5873238.11e-28.39e-2
CoNLL2003 (NER)1497372.06e-22.05e-1
CoNLL2012 (NER)14877372.07e-12.14e-1
DistilBERT
RTE (NLI)30184.34e-34.72e-3
MNLI (NLI)344717604.81e-14.98e-1
SQuADv1 (QA)19547922.76e-12.40e-1
SQuADv2 (QA)291611704.12e-13.46e-1
IMDB (Sent.)5492727.75e-27.46e-2
SST2 (Sent.)3061954.21e-25.49e-2
CoNLL2003 (NER)784361.10e-21.18e-1
CoNLL2012 (NER)7624341.07e-11.19e-1
", + "bbox": [ + 510, + 80, + 890, + 348 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Table 10: Single-GPU vs. BS and LR-optimized 4-GPU fine-tuning energy requirements across end tasks on the RTX8000 machine, for BERT-base and DistilBERT.", + "bbox": [ + 507, + 357, + 882, + 400 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "does not always) use about twice as much energy in both BERT-base and DistilBERT. If we increase batch size x 4 with 4 GPUs (and adjust learning rate accordingly), however, and compare single-GPU fine-tuning with multi-GPU fine-tuning (see Table 10), we observe that energy cost is typically similar or even less than when using 1 GPU, while taking around half as much time or less. In both the \"naive\" and \"optimized\" multi-GPU settings, the single-vs-multi-GPU difference in energy cost and job duration seems to be related to dataset sequence lengths. Tasks with longer sequences (such as QA tasks, and, to a lesser extent, IMDB and RTE) tend to exhibit more consistent and dramatic energy and time efficiency gains than the other tasks when using 4 GPUs. On the other hand, tasks with shorter sequences (such as NER) tend to require more energy with 4 GPUs, even if the wall-clock efficiency may be improved. One way one might interpret this is a large-enough per-device batch size and typical sequence length is necessary for multi-GPU training to be \"worth\" the overhead of communication between GPUs.", + "bbox": [ + 505, + 432, + 884, + 802 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In light of these observations, our general recommendation is that, if one owns a machine with multiple GPUs, one should consider using all available (idle) GPUs for energy- and time-efficient finetuning. Although it is often sufficient to use a single GPU when fine-tuning models of scale similar to ours, and instantaneous energy usage may be", + "bbox": [ + 507, + 806, + 885, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9068", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "higher using more GPUs, total energy used may end up being less while also requiring less time, especially if the training sequences tend to be longer. On the other hand, tasks with short sequences are likely best kept to a single GPU.", + "bbox": [ + 115, + 84, + 489, + 162 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9069", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_model.json b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d674a457029d83ff7b6d969e0c8a8d88906e82e1 --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_model.json @@ -0,0 +1,2061 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.2, + 0.084, + 0.8, + 0.105 + ], + "angle": 0, + "content": "Energy and Carbon Considerations of Fine-Tuning BERT" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.115, + 0.418, + 0.132 + ], + "angle": 0, + "content": "Xiaorong Wang\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.134, + 0.422, + 0.148 + ], + "angle": 0, + "content": "swangxr@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.115, + 0.713, + 0.131 + ], + "angle": 0, + "content": "Clara Na\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.134, + 0.743, + 0.147 + ], + "angle": 0, + "content": "csna@cs.cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.164, + 0.324, + 0.18 + ], + "angle": 0, + "content": "Emma Strubell\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.182, + 0.33, + 0.196 + ], + "angle": 0, + "content": "strubell@cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.165, + 0.582, + 0.18 + ], + "angle": 0, + "content": "Sorelle A. Friedler1" + }, + { + "type": "text", + "bbox": [ + 0.378, + 0.182, + 0.621, + 0.196 + ], + "angle": 0, + "content": "sorelle@cs.haverford.edu" + }, + { + "type": "text", + "bbox": [ + 0.678, + 0.165, + 0.82, + 0.18 + ], + "angle": 0, + "content": "Sasha Luccioni4" + }, + { + "type": "text", + "bbox": [ + 0.649, + 0.182, + 0.852, + 0.196 + ], + "angle": 0, + "content": "sasha.luccioni@hf.co" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.213, + 0.859, + 0.231 + ], + "angle": 0, + "content": "\\(^{1}\\)Haverford College, \\(^{2}\\)Carnegie Mellon University, \\(^{3}\\)Allen Institute for AI, \\(^{4}\\)Hugging Face" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.283, + 0.461, + 0.596 + ], + "angle": 0, + "content": "Despite the popularity of the pre-train then finetune paradigm in the NLP community, existing work quantifying energy costs and associated carbon emissions has largely focused on language model pre-training. Although a single pre-training run draws substantially more energy than fine-tuning, fine-tuning is performed more frequently by many more individual actors, and thus must be accounted for when considering the energy and carbon footprint of NLP. In order to better characterize the role of finetuning in the landscape of energy and carbon emissions in NLP, we perform a careful empirical study of the computational costs of finetuning across tasks, datasets, hardware infrastructure and measurement modalities. Our experimental results allow us to place fine-tuning energy and carbon costs into perspective with respect to pre-training and inference, and outline recommendations to NLP researchers and practitioners who wish to improve their finetuning energy efficiency." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.611, + 0.26, + 0.627 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.638, + 0.49, + 0.896 + ], + "angle": 0, + "content": "Fine-tuning pre-trained language models is a frequent occurrence in natural language processing (NLP) research and practice, yet the vast majority of work quantifying the energy and carbon footprints of NLP workloads has focused on pretraining (Strubell et al., 2019; Dodge et al., 2022; Luccioni et al., 2023) or inference (Desislavov et al., 2023; Luccioni et al., 2023). The typical lifecycle of an NLP model includes data ingestion, pre-training, fine-tuning and inference, all of which contribute non-trivially to energy use and corresponding emissions (Patterson et al., 2021; Wu et al., 2022). Better understanding of the role that each phase plays in overall energy and carbon footprint is vital to inform policy decisions, yet we still lack basic data quantifying the rela" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.254, + 0.882, + 0.285 + ], + "angle": 0, + "content": "tive contributions due to different aspects of model development and use (Kaack et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.288, + 0.885, + 0.513 + ], + "angle": 0, + "content": "In this work we perform an empirical study to quantify the energy requirements of language model fine-tuning, including in the context of pretraining and inference energy requirements. While this may seem like it should be a straightforward calculation, there are several variables that can influence compute time and energy consumption, ranging from: (1) the type of hardware used for both pre-training and fine-tuning (since this usually differs between the two), (2) the type of task and the type of computation required to carry it out, and (3) intrinsic characteristics of the dataset, such as average sequence length, its similarity with the pre-training dataset, etc." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.516, + 0.885, + 0.854 + ], + "angle": 0, + "content": "In order to isolate the factors that have the most influence on fine-tuning dynamics, we compare fine-tuning energy use across a suite of common supervised NLP datasets including the tasks of entailment, sentiment analysis, question answering, and named entity recognition, and with training data sizes ranging from 6K to 400K examples. We also measure energy use across two different sets of hardware, using the CodeCarbon (Schmidt et al., 2021) software package and a physical energy measurement device at the wall, to quantify variance due to physical factors. To enable carefully controlled comparison of the roles of pre-training and fine-tuning in NLP model lifecycle energy use, we additionally pre-train BERT variants from scratch on the same hardware. We find that pre-training BERT is equivalent to anywhere from 400 (MNLI) to 45,000 (RTE) fine-tuning runs depending on the dataset size, and that number of training tokens1 is a reasonable heuristic for estimating fine-tuning energy use. Further comparison of fine-tuning in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.87, + 0.884, + 0.919 + ], + "angle": 0, + "content": "1The \"true\" number of training tokens seen, accounting for dynamic padding of sequences to the maximum length in a batch, is a better predictor than relying on to mean or median number of tokens per example." + }, + { + "type": "page_footnote", + "bbox": [ + 0.143, + 0.905, + 0.319, + 0.918 + ], + "angle": 0, + "content": "*Denotes equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9058" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.78, + 0.959 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9058-9069" + }, + { + "type": "footer", + "bbox": [ + 0.278, + 0.959, + 0.72, + 0.972 + ], + "angle": 0, + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.493, + 0.312 + ], + "angle": 0, + "content": "ference energy intensity across tasks confirms that example sequence length holds a much stronger influence on energy intensity in the fine-tuning phase than in the inference phase, in alignment with expectations from previous work (Zhou et al., 2021). Together, our observations contextualize the energy and carbon requirements of fine-tuning in the broader model lifecycle and highlight the need to study fine-tuning energy efficiency separately from pre-training and inference workloads in NLP models. We hope that our careful measurement of the relative costs of different NLP workloads will serve as a valuable datapoint informing decision-making both within and beyond the NLP community." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.326, + 0.271, + 0.342 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.356, + 0.492, + 0.709 + ], + "angle": 0, + "content": "Measurement of energy consumption and carbon emissions of NLP models has become an active area of research since it was first identified that modern state-of-the-art models based on deep learning can produce substantial greenhouse gas emissions due to the energy required to train them (Strubell et al., 2019; Schwartz et al., 2020). These measurements have mostly focused on two research directions. First, there has been a series of empirical studies on different model architectures, focused on estimating the carbon emissions generated by their training process and the relative efficiency of different methods (Naidu et al., 2021; Patterson et al., 2021, 2022). Recent work by Luccioni et al. has built upon this, aiming to encompass the embodied emissions of manufacturing computing hardware as well as those produced via the inference process (2023). There has also been complementary work measuring the energy used by Transformer models during inference and ways of predicting those costs for different architectures and models (Cao et al., 2021; Ang et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.711, + 0.493, + 0.92 + ], + "angle": 0, + "content": "Closest to our work, the HULK benchmark (Zhou et al., 2021) was proposed to measure the relative efficiency-accuracy trade-offs of different pre-trained models, measuring the wall-clock time for different pre-trained models to reach a target accuracy on one of three fine-tuning tasks. Different from Zhou et al. (2021), our work explicitly focuses on the energy and carbon required for fine-tuning (theirs uses time and financial cost as proxies), evaluates a wider variety of fine-tuning tasks and hardware settings in order to elucidate the factors that predict fine-tuning energy requirements, and further contextualizes fine-tuning energy re" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.116 + ], + "angle": 0, + "content": "requirements in the bigger picture of ML model lifecycle emissions." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.117, + 0.887, + 0.392 + ], + "angle": 0, + "content": "Another related direction of research examines the dynamics of pre-training and fine-tuning of language models and the influence of factors like random seeds and early stopping (Dodge et al., 2020), scaling (Tay et al., 2022) and learning dynamics (Hao et al., 2020). While all of these studies have shed important light on these processes, in practice most of the decisions made remain empirical, with practitioners either referring to previous work (when hyperparameters and other training details are reported), or using techniques such as grid or random search (Bergstra and Bengio, 2012) to converge on optimal parameter values. Our own work builds upon both of these research directions. We study both the pre-training and fine-tuning process, and our experiments for studying their energy intensity are based on the works cited above." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.403, + 0.659, + 0.42 + ], + "angle": 0, + "content": "3 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.429, + 0.886, + 0.51 + ], + "angle": 0, + "content": "Full training details can be found in Appendix A.1. We release code for replicating our measurements.2 We encourage others to run our code on their hardware to add to a repository of measurements across different hardware platforms." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.521, + 0.707, + 0.537 + ], + "angle": 0, + "content": "3.1 Pre-training BERT" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.542, + 0.886, + 0.815 + ], + "angle": 0, + "content": "In this work we are interested in measuring the energy consumption of fine-tuning, as it compares to other stages of model use: pre-training and inference. In order to establish a comparable baseline for the energy consumption and carbon emissions of pre-training, we pre-trained a BERT-base model (Devlin et al., 2019) from scratch on the same hardware that we use for fine-tuning (Section 3.3). Although in practice pre-training and fine-tuning are often done separately and on different hardware, we fixed the machine for both sets of experiments in order to aid direct comparability of energy usage measurements. Following Devlin et al. (2019), we pre-train our model on the BookCorpus (Zhu et al., 2015) and the 2020 version of Wikipedia (Foundation, 2020), both downloaded from HuggingFace Datasets (Lhoest et al., 2021)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.816, + 0.886, + 0.897 + ], + "angle": 0, + "content": "Our precise pre-training methodology differs slightly from Devlin et al. (2019): our data necessarily differs slightly because the original training corpus was not released along with the model, and we only use the masked language modeling" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.904, + 0.846, + 0.919 + ], + "angle": 0, + "content": "2https://github.com/swangxr/FT-energy.git" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9059" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.488, + 0.149 + ], + "angle": 0, + "content": "(MLM) objective without next sentence prediction (NSP) following Liu et al. (2019), who found that removing NSP did not substantially impact endtask performance." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.15, + 0.488, + 0.23 + ], + "angle": 0, + "content": "In order to assess the relative impact of using a more efficient pre-trained BERT variant, we also followed the DistilBERT (Sanh et al., 2019) recipe, performing knowledge distillation on our pre-trained BERT-base model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.24, + 0.308, + 0.256 + ], + "angle": 0, + "content": "3.2 Fine-tuning BERT" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.261, + 0.488, + 0.357 + ], + "angle": 0, + "content": "We evaluate the energy consumption and carbon emissions of the fine-tuning process on the tasks in Table 1. We deliberately chose this selection of end-tasks in order to vary fine-tuning dataset size, task type, and sequence length, while also aligning with tasks commonly used in NLP applications." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.367, + 0.475, + 0.521 + ], + "angle": 0, + "content": "
DatasetTaskExamplesSeq. length
Med.Batch
Wiki+BooksMLM43M128128
RTENLI6K56128
MNLINLI433K3790
SQuADv1QA98K159336
SQuADv2QA142K158336
IMDBSent.50K128128
SST2Sent.70K1040
CoNLL2003NER21K1553
CoNLL2012NER143K2065
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.531, + 0.49, + 0.618 + ], + "angle": 0, + "content": "Table 1: Pre-training and fine-tuning dataset descriptions. We report two sequence length statistics: Median tokens per sequence in the dataset, and Batch, the average maximum sequence length per batch as predicted by simulated sampling. The latter is included as a more direct predictor of computation cost for dynamic padding." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.632, + 0.488, + 0.776 + ], + "angle": 0, + "content": "All models are fine-tuned on the BERT models described in §3.1. For each fine-tuning task, we use typical fine-tuning hyperparameters specific to the task or user-reported hyperparameters on current fine-tuned models on HuggingFace, in order to mimic the common real-world use cases. For each task, we dynamically pad sequences to the maximum length in each batch. All fine-tuning hyperparameters are reported in Appendix A.1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.777, + 0.488, + 0.824 + ], + "angle": 0, + "content": "We also report average per-example energy use for inference. All the inference tasks are performed on 1 GPU with batch size 1 on the same machines." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.835, + 0.323, + 0.849 + ], + "angle": 0, + "content": "3.3 Hardware Platforms" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.488, + 0.919 + ], + "angle": 0, + "content": "To ensure reproducibility and measure variability across hardware platforms, we replicate experiments across two hardware platforms: One A100 machine and one RTX 8000 machine, where each" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.881, + 0.116 + ], + "angle": 0, + "content": "machine had four GPUs. Pre-training experiments used all 4 GPUs in each machine." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.118, + 0.884, + 0.295 + ], + "angle": 0, + "content": "All fine-tuning tasks were performed on the same machines, but using only one GPU. This reflects the typical scenario where fine-tuning is done on a single GPU even if the machine itself has more GPUs. To better compare the energy usage results across pre-training and fine-tuning, we also report an energy usage estimate for BERT-base pretraining on 1 RTX 8000 GPU with hyperparameters equivalent to training on 4 GPUs, extrapolated from a 12-hour partial training run. Details are recorded in Appendix A.1." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.304, + 0.803, + 0.32 + ], + "angle": 0, + "content": "3.4 Measuring Energy and Carbon" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.324, + 0.884, + 0.597 + ], + "angle": 0, + "content": "To measure the energy consumed, we use the software tool CodeCarbon (Schmidt et al., 2021). Recent work has found that the existing libraries and code for estimating the carbon emissions of NLP techniques vary in terms of their accuracy and generalizability to different types of hardware (Bannour et al., 2021). To compensate for this, we calibrate the programmatic energy usage readings with a physical energy meter, with which we record energy readings directly from the compute node during experiments. Subsequently, we calculate a coefficient of expected power loss, 1.059, as the average proportion (over runs across fine-tuning tasks) of physical energy reading vs. programmatic energy measurement. Full results are given in Appendix A.2. Thus, the energy consumed in kWh, denoted as \\( E \\), is determined via the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.605, + 0.836, + 0.62 + ], + "angle": 0, + "content": "\\[\nE (\\mathrm {k W h}) = 1. 0 5 9 \\cdot \\text {c o d e c a r b o n} \\mathrm {k W h}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.627, + 0.884, + 0.786 + ], + "angle": 0, + "content": "Converting the power loss adjusted values to \\(\\mathrm{CO}_{2}\\) emissions is done through CodeCarbon using a coefficient specific to the energy grid based on the location of the server from the most recent EPA Power Profiler Data. The conversion factor for the server's location (Pittsburgh, PA) in Table 2 is 1046.1 lbs/MWh, while the factor for the second server's location (Haverford, PA) is 672.8 lbs/MWh. The total kilograms of \\(\\mathrm{CO}_{2}\\) emitted, denoted as \\(C\\), is then determined via:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.79, + 0.881, + 0.821 + ], + "angle": 0, + "content": "\\[\nC = E \\times 1 0 4 6. 1 \\frac {\\mathrm {l b}}{\\mathrm {M W h}} \\times \\frac {1 \\mathrm {M W h}}{1 0 0 0 \\mathrm {k W h}} \\times \\frac {1 \\mathrm {k g}}{2 . 2 0 4 6 2 \\mathrm {l b}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.824, + 0.884, + 0.919 + ], + "angle": 0, + "content": "We convert the \\(\\mathrm{CO}_{2}\\) emissions result to human understandable reference values using the EPA Greenhouse Gas Equivalencies Calculator; in Table 2, we also show the equivalent \\(\\mathrm{CO}_{2}\\) emissions of miles driven by an average gasoline-powered passenger vehicle." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9060" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.082, + 0.875, + 0.301 + ], + "angle": 0, + "content": "
TaskDatasetTraining / Fine-tuningInference
Energy (kWh)Emissions (kg CO2)TimeEquiv. # PT runsEquiv. milesEnergy (kWh) / 1000 ex.Equiv. # FT runs
MLM4 GPU (1m @ len 128)270.9128.627 hr1.358330——
1 GPU (1m @ len 128)419.6199.1673 hr0.646510——
4 GPU (100k @ len 512)124.158.90114 hr2.965151——
Total 4 GPU368.0174.6357 hr1.000408——
NLIRTE0.0080.00459 s451090.010.794e-31k
MNLI0.9380.4456700 s3921.147.349e-3127k
QASQuAD v10.5370.2553780 s6850.652.157e-3249k
SQuAD v20.7950.3775604 s4630.972.471e-3322k
Sent.IMDB0.1510.0721074 s24410.180.849e-3178k
SST20.0810.038587 s45400.100.691e-312k
NERCoNLL20030.0210.010149 s178860.030.808e-33k
CoNLL20120.2070.0981487 s17780.250.878e-324k
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.311, + 0.883, + 0.398 + ], + "angle": 0, + "content": "Table 2: Energy consumption of pre-training BERT and fine-tuning on the RTX8000 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent \\(\\mathrm{CO}_{2}\\) emissions. 1 GPU pre-training costs are extrapolated from a shorter pre-training run lasting only a few hours. Total cost of fine-tuning is derived from 900k steps of pre-training on sequences of length 128 followed by 100k steps on sequences of length 512, following (Devlin et al., 2019). Energy consumption for inference is calculated using single-example batches." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.422, + 0.345, + 0.436 + ], + "angle": 0, + "content": "4 Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.447, + 0.487, + 0.511 + ], + "angle": 0, + "content": "Table 2 shows energy, carbon, and wall-clock time required to fine-tune BERT-base models on the RTX8000 machine. Results on the A100 machine are recorded in Appendix A.3 in Tables 8 and 9." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.523, + 0.486, + 0.677 + ], + "angle": 0, + "content": "
TaskDatasetEnergy (kWh)TimeEmiss. (kg CO2)
Distil.Wiki+Books187.74175.5 hr89.08
NLIRTE0.00430 s0.002
MNLI0.4813447 s0.228
QASQuAD v10.2761954 s0.131
SQuAD v20.4122916 s0.196
Sent.IMDB0.077549 s0.037
SST0.042306 s0.020
NERCoNLL20030.01178 s0.005
CoNLL20120.107762 s0.051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.687, + 0.489, + 0.759 + ], + "angle": 0, + "content": "Table 3: Energy consumption of training (distillation) and fine-tuning DistilBERT on the RTX8000 machine. Energy calculations are the same as in Table 2. Distillation is performed on a pre-trained model, and so the true \"total\" cost includes the pre-training cost as well." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.786, + 0.383, + 0.801 + ], + "angle": 0, + "content": "4.1 Pre-training and Distillation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We observe that it requires an additional \\(50\\%\\) of the energy cost of pre-training in order to perform knowledge distillation, but it takes nearly \\(50\\%\\) less energy to fine-tune on the same tasks using DistilBERT vs. normal BERT (see Table 3). By our estimate, one can fully amortize the up-front cost of distillation within anywhere from 86 fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.422, + 0.882, + 0.47 + ], + "angle": 0, + "content": "runs of an MNLI-like task, to 47k fine-tunings on an RTE-like task. DistilBERT fine-tuning results on the A100 machine are in Appendix A.3." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.495, + 0.863, + 0.525 + ], + "angle": 0, + "content": "4.2 Comparing and Predicting Fine-tuning Emissions" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.539, + 0.884, + 0.812 + ], + "angle": 0, + "content": "We find that, controlling for hardware, energy consumption scales most predictably with wall clock time and number of tokens encountered during training (including the pad tokens added to sequences to match the maximum sequence length in a batch). The linear relationship between energy consumption and total number of tokens holds similarly on both machines (see Figure 1). Additionally, we observe a consistently higher energy consumption in the RTX 8000 GPU machine. This is likely due to the higher energy overhead and the (in)efficiency of the hardware compared to the A100 GPUs. Other figures in Appendix A.3 illustrate that, in contrast, energy requirements as a function of optimization steps or even number of examples in the dataset can vary significantly across datasets and tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.833, + 0.883, + 0.919 + ], + "angle": 0, + "content": "3Note that cheaper inference is often the primary goal of knowledge distillation. Inference is much cheaper than training and therefore requires more to amortize the initial cost of distillation, but inference also occurs much more frequently than training. Models running inference at scale are typically highly optimized with respect to specific deployment settings, so our estimates approximate a lower bound." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "9061" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.085, + 0.496, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.282, + 0.49, + 0.41 + ], + "angle": 0, + "content": "Figure 1: Total energy consumed (kWh) is strongly related to number of tokens seen for BERT models on both A100 and RTX 8000 GPU machines, although the relationship is more predictive on the RTX 8000 machine and energy usage is less consistent with DistilBERT. Note that both axes are in log scale. An alternative view of similar data in Figure 4 distinguishes pre-training workloads' energy consumption slightly from that of fine-tuning tasks." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.425, + 0.487, + 0.561 + ], + "angle": 0, + "content": "
DatasetSeq. lengthkWh / 1k ex.
Med.BatchInf.FT
RTE (NLI)561280.75e-31.09e-3
MNLI (NLI)37900.70e-30.80e-3
SQUADv1 (QA)1593361.08e-33.04e-3
SQUADv2 (QA)1583361.24e-33.02e-3
IMDB (Sent.)1281280.80e-31.21e-3
SST2 (Sent.)10400.65e-30.40e-3
CoNLL2003 (NER)15530.76e-30.49e-3
CoNLL2012 (NER)20650.83e-30.60e-3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.113, + 0.57, + 0.488, + 0.629 + ], + "angle": 0, + "content": "Table 4: Inference vs. fine-tuning energy requirements across end tasks. We see that fine-tuning energy usage varies according to sequence length much more widely than inference energy usage." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.656, + 0.377, + 0.672 + ], + "angle": 0, + "content": "4.3 Fine-tuning vs Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.678, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Even for the more reliable predictors of energy consumption and carbon emissions (duration of training and number of tokens processed), the energy cost profiles of pre-training vs. fine-tuning are different, likely due to differences in training infrastructure, training objectives, and sequence lengths typically seen in pre-training vs. fine-tuning (see Figures 4, 3, 5, and 6). Pre-training in general is almost always performed over multiple GPUs which incurs energy costs from communication between GPUs, and often also with gradient accumulation to accommodate large batches. Moreover, sequences are packed together such that batches consist largely or entirely of sequences of identical length equal to the maximum sequence length for" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.085, + 0.593, + 0.099 + ], + "angle": 0, + "content": "the model." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.101, + 0.884, + 0.262 + ], + "angle": 0, + "content": "On the other hand, there are many types of finetuning tasks where examples consist of sequences of varying lengths significantly shorter than the maximum length that the model has been trained on, as shown in Table 1. Since, effective sequence lengths are determined dynamically at training time (where sequences are padded to the maximum length in each given batch), total training time is not as simple to extrapolate from standard measures of dataset size as in pre-training." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.273, + 0.755, + 0.289 + ], + "angle": 0, + "content": "4.4 Fine-tuning vs. Inference" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.293, + 0.885, + 0.455 + ], + "angle": 0, + "content": "Although we do observe that per-example inference efficiency costs are related to sequence lengths, there is overall less variation across datasets and tasks in inference costs compared to fine-tuning costs (see Table 4). This mirrors an observation noted in the HULK benchmark (Zhou et al., 2021), though to the best of our knowledge ours is the first to explicitly draw comparisons across task types and different aspects of dataset size (i.e. number of examples and examples' sequence lengths)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.465, + 0.754, + 0.48 + ], + "angle": 0, + "content": "4.5 Single vs. Multiple GPUs" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.486, + 0.884, + 0.71 + ], + "angle": 0, + "content": "In general, typical hardware and data settings for pre-training and fine-tuning tend to differ. Though to the best of our knowledge it is less common to fine-tune causal LMs of this scale on multiple GPUs, we present additional results from multi-GPU fine-tuning on the \\(4 \\times\\) RTX8000 machine with the same fine-tuning tasks in Table 10 in Appendix A.3. Our recommendations from an energy efficiency standpoint align with common rules of thumb for effective utilization of hardware; if the resources would be idle otherwise, one could reasonably consider increasing batch size and learning rate to saturate the available hardware for both time- and energy-efficient training." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.722, + 0.642, + 0.737 + ], + "angle": 0, + "content": "5 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.748, + 0.885, + 0.908 + ], + "angle": 0, + "content": "We share a procedure for rigorous measurement of energy consumption from causal LM fine-tuning given multiple concrete hardware settings. We hope our work is useful to researchers and practitioners who are interested in obtaining measurements for their own specific hardware, gaining intuitions about factors affecting relative energy costs of different types of fine-tuning tasks, or understanding these energy costs in context of the model lifecycle." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9062" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.115, + 0.085, + 0.221, + 0.1 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.112, + 0.49, + 0.319 + ], + "angle": 0, + "content": "While our work provides important first steps towards a clearer understanding of model finetuning's impact on the environment, we note that our experimentation is limited to various token classification, sequence classification, and question answering tasks with BERT-base and DistilBERT-base models. We do not make claims or extrapolations about much larger language models, or models with different architectures, as well as other types of tasks such as summarization. Future work in this direction can expand the number of tasks that we consider as well as feature different architectures such as RoBERTa (Liu et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.322, + 0.49, + 0.531 + ], + "angle": 0, + "content": "Additionally, the on-premises hardware infrastructure used for our experimentation is realistic and typical of compute resources in academic settings, but we provide no firsthand evidence of fine-tuning emissions profiles expected from either local model training (where the impracticality of pretraining makes direct comparisons with fine-tuning emissions difficult) or fine-tuning on hardware that is part of much larger scale infrastructure such as on a public cloud. Furthermore, we expect that use of specialized hardware such as TPUs (as opposed to GPUs, which we use) would be associated with different emissions profiles." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.545, + 0.267, + 0.56 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.572, + 0.49, + 0.749 + ], + "angle": 0, + "content": "Training and deploying ML models and systems requires large amounts of energy, the production of which results in the emission of greenhouse gases. While the goal of our research is to contribute towards a better understanding of the factors that influence these emissions, by carrying out our experiments, we were ourselves responsible for the emission of \\(350\\mathrm{kg}\\) of carbon equivalents. We release our code and the data generated by our experiments to maximize the transparency and reproducibility of our work." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.764, + 0.287, + 0.78 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.79, + 0.489, + 0.87 + ], + "angle": 0, + "content": "We are grateful to Victor Sanh for his time and patience in answering questions about BERT and DistilBERT pre-training. We would also like to thank our anonymous reviewers for taking the time to provide helpful feedback." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.872, + 0.49, + 0.919 + ], + "angle": 0, + "content": "This work was supported in part by a grant from the National Science Foundation Graduate Research Fellowship Program under Grant No." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.886, + 0.15 + ], + "angle": 0, + "content": "DGE2140739. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.176, + 0.61, + 0.191 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.198, + 0.886, + 0.278 + ], + "angle": 0, + "content": "Phyllis Ang, Bhuwan Dhingra, and Lisa Wu Wills. 2022. Characterizing the efficiency vs. accuracy trade-off for long-context NLP models. In Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP, pages 113-121, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.286, + 0.884, + 0.34 + ], + "angle": 0, + "content": "Nesrine Bannour, Sahar Ghannay, Aurélie Néveol, and Anne-Laure Ligozat. 2021. Evaluating the carbon footprint of NLP methods: a survey and analysis of existing tools. In EMNLP, Workshop SustainNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.349, + 0.885, + 0.389 + ], + "angle": 0, + "content": "James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(10):281-305." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.398, + 0.886, + 0.517 + ], + "angle": 0, + "content": "Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. 2021. IrEne: Interpretable energy prediction for transformers. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2145-2157, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.525, + 0.886, + 0.593 + ], + "angle": 0, + "content": "Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. 2023. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. *Sustainable Computing: Informatics and Systems*, 38:100857." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.601, + 0.886, + 0.72 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.728, + 0.886, + 0.795 + ], + "angle": 0, + "content": "Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. 2020. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.803, + 0.886, + 0.897 + ], + "angle": 0, + "content": "Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the carbon intensity of ai in cloud instances. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1877-1894." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.904, + 0.876, + 0.919 + ], + "angle": 0, + "content": "Wikipedia Foundation. 2020. Wikipedia downloads." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.198, + 0.886, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9063" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.166 + ], + "angle": 0, + "content": "Yaru Hao, Li Dong, Furu Wei, and Ke Xu. 2020. Investigating learning dynamics of bert fine-tuning. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 87-92." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.177, + 0.488, + 0.23 + ], + "angle": 0, + "content": "Lynn H. Kaack, Priya L. Donti, Emma Strubell, George Kamiya, Felix Creutzig, and David Rolnick. 2022. Aligning artificial intelligence with climate change mitigation. Nature Climate Change, 12(6):518-527." + }, + { + "type": "ref_text", + "bbox": [ + 0.116, + 0.242, + 0.488, + 0.464 + ], + "angle": 0, + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.476, + 0.488, + 0.543 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.554, + 0.488, + 0.607 + ], + "angle": 0, + "content": "Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. 2023. Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253):1-15." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.619, + 0.488, + 0.684 + ], + "angle": 0, + "content": "Rakshit Naidu, Harshita Diddee, Ajinkya Mulay, Aleti Vardhan, Krithika Ramesh, and Ahmed Zamzam. 2021. Towards quantifying the carbon emissions of differentially private machine learning. arXiv preprint arXiv:2107.06946." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.697, + 0.488, + 0.774 + ], + "angle": 0, + "content": "David Patterson, Joseph Gonzalez, Urs Holzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. The carbon footprint of machine learning training will plateau, then shrink. Computer, 55(7):18-28." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.788, + 0.488, + 0.854 + ], + "angle": 0, + "content": "David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.866, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + }, + { + "type": "list", + "bbox": [ + 0.116, + 0.086, + 0.488, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.152 + ], + "angle": 0, + "content": "Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Ziyao Wang, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luccioni. 2021. Codecarbon: estimate and track carbon emissions from machine learning computing." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.165, + 0.882, + 0.204 + ], + "angle": 0, + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2020. Green AI. Communications of the ACM, 63(12):54-63." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.218, + 0.882, + 0.297 + ], + "angle": 0, + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.31, + 0.882, + 0.389 + ], + "angle": 0, + "content": "Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2022. Scale efficiently: Insights from pretraining and finetuning transformers. In International Conference on Learning Representations." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.402, + 0.882, + 0.429 + ], + "angle": 0, + "content": "C.-J. Wu et al. 2022. Sustainable AI: Environmental implications, challenges, and opportunities. In \\(MLSys\\)." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.441, + 0.882, + 0.546 + ], + "angle": 0, + "content": "Xiyou Zhou, Zhiyu Chen, Xiaoyong Jin, and William Yang Wang. 2021. HULK: An energy efficiency benchmark platform for responsible natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 329-336, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.56, + 0.882, + 0.638 + ], + "angle": 0, + "content": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV)." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.655, + 0.633, + 0.671 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.682, + 0.845, + 0.698 + ], + "angle": 0, + "content": "A.1 Pre-training and Fine-tuning Details" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.705, + 0.882, + 0.768 + ], + "angle": 0, + "content": "For pre-training BERT with MLM, we mainly follow what is listed in Devlin et al. (2019). The specific hyperparameters used for pre-training are listed in Table 5 and were used on both machines." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.778, + 0.819, + 0.891 + ], + "angle": 0, + "content": "
Hyperparameters
maximum steps1000000
training batch size per device64
evaluation batch size per device64
maximum sequence length128
learning rate1e-4
warmup steps10000
weight decay0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.511, + 0.9, + 0.879, + 0.915 + ], + "angle": 0, + "content": "Table 5: Hyperparameters used for pre-training BERT" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.488, + 0.133 + ], + "angle": 0, + "content": "In Table 6, we list the set of hyperparameters used to fine-tune each task. All fine-tuning tasks were run on both machines." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.15, + 0.49, + 0.279 + ], + "angle": 0, + "content": "Hardware details The A100 machine is located in Haverford, Pennsylvania, USA, and has 4x NVIDIA A100 GPUs with 40GB GDDR SDRAM, 376GB main memory and 32 Intel Xeon processors. The RTX 8000 machine is located in Pittsburgh, Pennsylvania, USA, and has 4x NVIDIA Quadro RTX 8000 GPUs with 48GB GDDR SDRAM, 36 Intel Xeon processors and 251GB RAM." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.298, + 0.379, + 0.313 + ], + "angle": 0, + "content": "A.2 Kill-A-Watt Measurements" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.324, + 0.49, + 0.628 + ], + "angle": 0, + "content": "Energy is lost during the process of transferring energy from a power source to the machine. The coefficient for the loss is acquired through the readings from Kill-A-Watt devices. The device measures the instantaneous Watt reading extracted from the wall and displays on the monitor. The measurements for the A100 machine were recorded only on the single node of the cluster containing it (the RTX8000 machine is not part of a larger cluster). The GPU node is connected to two outlets, and we plug separate Kill-A-Watt devices into both. For each instantaneous reading, we read off of and sum up the readings on both Kill-A-Watt devices. To best compare between the package reading and the wall reading, we read off of the device at the same 15 second interval that CodeCarbon records the energy consumed. To convert each instantaneous Watts readings into Kilowatt-Hours, we follow the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.644, + 0.452, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathrm {k W h} = \\frac {1 5 \\text {s e c o n d s}}{3 6 0 0 \\text {s e c o n d s}} \\times \\frac {\\text {w a t t s} \\times \\text {h o u r s}}{1 0 0 0}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.692, + 0.49, + 0.852 + ], + "angle": 0, + "content": "We sum up all the calculations for the entire run of a fine-tuning experiment. Then, we can compare the sum of the wall readings with the Code Carbon energy consumption recording. We divide the wall reading over the package reading to get a coefficient, measuring the more realistic energy used. The setup of an instantaneous reading on the A100 machine is shown in Figure 2. The recordings of the wall readings on the A100 GPU machine are recorded in Table 7." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.855, + 0.489, + 0.919 + ], + "angle": 0, + "content": "We ran these wall reading experiments for both machines and obtained a coefficient of 1.09 on the A100 machine, and a coefficient of 1.05 on the RTX machine." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.082, + 0.886, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.165, + 0.883, + 0.181 + ], + "angle": 0, + "content": "Figure 2: Kill-A-Watt wall reading measurement setup" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.206, + 0.707, + 0.22 + ], + "angle": 0, + "content": "A.3 Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.227, + 0.885, + 0.337 + ], + "angle": 0, + "content": "BERT pre-training and fine-tuning results Table 8 records the energy consumption results from the A100 GPU machine, located in Haverford, PA. The conversion factor for this location is 672.8 lbs \\(\\mathrm{CO}_{2} / \\mathrm{MWh}\\) . We then calculate the emissions using raw output from CodeCarbon, which is listed in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.349, + 0.886, + 0.477 + ], + "angle": 0, + "content": "DistilBERT results Table 9 records the energy consumption and emission on the A100 GPU machine. Distillation is not done on this machine, and all the fine-tunings are done on the DistilBERT trained on the RTX8000 machine. The results on the A100 machine show a similar trend that fine-tunings on DistilBERT takes around \\(50\\%\\) less energy than on BERT base models." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.492, + 0.891, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.688, + 0.883, + 0.716 + ], + "angle": 0, + "content": "Figure 3: Energy consumed (kWh) against total tokens seen on A100 machine" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.886, + 0.92 + ], + "angle": 0, + "content": "Training duration and energy consumption From Figure 5 and 6, we see that there is a strong correlation between training time and energy consumption, which holds across our models and hardware settings (Figure 7). However, similarly to tokens seen, we observe that pre-training exhibits a slightly different energy consumption profile, as do question answering tasks on the A100 machine (which have the longest sequences out of our finetuning tasks). When training time estimates are not feasible in advance (such as in certain hyperparam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9065" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.142, + 0.082, + 0.858, + 0.228 + ], + "angle": 0, + "content": "
RTEMNLIIMDBSST2SQuAD v1SQuAD v2CoNLL 2003CoNLL 2012
epochs33532233
train bs3232163232323232
eval bs3232163232323232
max seq len128128128128384384128128
doc stride128128
lr2e-52e-52e-52e-53e-53e-55e-55e-5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.321, + 0.237, + 0.677, + 0.251 + ], + "angle": 0, + "content": "Table 6: Hyperparameters used for fine-tuning tasks" + }, + { + "type": "table", + "bbox": [ + 0.2, + 0.264, + 0.8, + 0.389 + ], + "angle": 0, + "content": "
TaskRandom seedWall reading converted (kWh)Code Carbon reading (kWh)Power Loss Co-efficient
SST2 (no AC)420.1030.0881.170
SST21230.0590.0551.057
IMDB420.0700.0611.134
IMDB420.0790.0711.122
IMDB1230.07930.0731.092
RTE420.006410.006021.065
" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.398, + 0.744, + 0.412 + ], + "angle": 0, + "content": "Table 7: Kill-A-Watt experiment recordings for the A100 GPU machine" + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.425, + 0.861, + 0.591 + ], + "angle": 0, + "content": "
TaskDatasetEnergy (kWh)Emissions (kg CO2)Training time (s)Equiv. # PT runsEquivalent emissions of miles driven
MLMWiki + Books146.15044.6027265531144
fine-tuningRTE0.0020.00219584420.005
MNLI0.3560.23924004060.613
SQuAD v10.1710.11510618460.295
SQuAD v20.2630.17715805490.454
IMDB0.0620.019047821140.048
SST20.0510.03437728580.087
CoNLL20030.0110.00778134110.018
CoNLL20120.0910.06165415840.156
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.601, + 0.884, + 0.675 + ], + "angle": 0, + "content": "Table 8: Energy consumption of pre-training BERT and fine-tuning on the A100 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent \\(\\mathrm{CO}_{2}\\) emissions. Note that here, unlike in Table 2, the \"baseline\" pre-training cost is fixed to the cost of 1 million steps of masked language modeling of sequences of length 128." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.695, + 0.484, + 0.819 + ], + "angle": 0, + "content": "
TaskDatasetEnergyTimeEmiss.
EntailmentRTE0.001100.001
MNLI0.21015390.141
QASQuAD v10.0915660.061
SQuAD v20.1419150.095
SentimentIMDB0.0453410.012
SST0.0342620.023
NERCoNLL20030.007570.005
CoNLL20120.0594480.040
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.828, + 0.49, + 0.872 + ], + "angle": 0, + "content": "Table 9: Energy consumption of training (distillation) and fine-tuning DistilBERT on the A100 machine. Units and energy calculations are the same as given in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.698, + 0.883, + 0.763 + ], + "angle": 0, + "content": "eter sweeps), we recommend that researchers and practitioners use token counts estimates (including dynamic padding tokens) if they have reasonable knowledge of their data." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.775, + 0.885, + 0.92 + ], + "angle": 0, + "content": "As indicated in Figure 8 and Figure 9, energy increases as the optimization steps increase. This is not surprising given the correlation between training time and energy consumption. However, we see that for different tasks, the energy required for each step is very different. Each step of pre-training takes a longer time, likely due to the higher batch size than all fine-tuning tasks. For QA tasks, the per-step energy consumption is higher than other" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.522, + 0.941 + ], + "angle": 0, + "content": "9066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.085, + 0.496, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.281, + 0.489, + 0.31 + ], + "angle": 0, + "content": "Figure 4: Energy consumed (kWh) against total tokens seen on RTX8000 machine" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.33, + 0.496, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.526, + 0.489, + 0.556 + ], + "angle": 0, + "content": "Figure 5: Energy consumed (kWh) against training time (s) on A100 machine, showing the first 30 minutes" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.574, + 0.486, + 0.768 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.781, + 0.489, + 0.81 + ], + "angle": 0, + "content": "Figure 6: Energy consumed (kWh) against training time (s) on RTX8000 machine, showing the first 30 minutes" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.838, + 0.489, + 0.87 + ], + "angle": 0, + "content": "tasks. This is likely due to the maximum sequence length of 384 being higher than for the other tasks." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.872, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Figure 10 shows the number of examples of the task and the energy consumed in log scale. Similar to Figure 1, we see a direct correlation between" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.086, + 0.891, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.282, + 0.886, + 0.369 + ], + "angle": 0, + "content": "Figure 7: Total training time is strongly predictive of total energy consumed (kWh) for both BERT and DistilBERT models, on both A100 and RTX 8000 GPU machines. Note that both axes are in log scale. Different from token counts, the trendlines themselves are similar as well across models and hardware settings." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.386, + 0.871, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.593, + 0.883, + 0.622 + ], + "angle": 0, + "content": "Figure 8: Energy consumed (kWh) against optimization steps on A100 machine machine" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.64, + 0.873, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.847, + 0.882, + 0.876 + ], + "angle": 0, + "content": "Figure 9: Energy consumed (kWh) against optimization steps on RTX8000 machine" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.085, + 0.496, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.282, + 0.49, + 0.382 + ], + "angle": 0, + "content": "Figure 10: Energy consumed (kWh) against total number of examples seen on both A100 and RTX8000 machines. Comparing with Figure 1, we see that merely counting number of training examples is much less predictive of energy consumption than accounting for example sequence lengths along with batch size (which affects the maximum sequence length in each batch)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.407, + 0.489, + 0.438 + ], + "angle": 0, + "content": "higher number of training examples and higher energy usage on both machines." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.453, + 0.475, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.657, + 0.488, + 0.686 + ], + "angle": 0, + "content": "Figure 11: \\(\\mathrm{CO}_{2}\\) emissions (kg) against task accuracy for SST-2 fine-tuning on RTX 8000 GPU machine" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.702, + 0.489, + 0.799 + ], + "angle": 0, + "content": "Figure 11 shows the accuracy of SST2 task as \\(\\mathrm{CO}_{2}\\) emissions (kg) increases. As shown previously in Figure 6, energy consumption increases as time increases. Generally, as emission increases, accuracy increases. We can see that as emissions goes up, accuracy is trending towards converging." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Single- vs. multi-GPU fine-tuning We fine-tune BERT-base and DistilBERT-base models on the RTX8000 machine as well. If the same hyperparameters as the single-GPU setting are used naively (i.e. the same batch size is split over 4 GPUs), fine-tuning using 4 GPUs can (but does not always) take even longer than using just 1 GPU, and can (but" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.082, + 0.892, + 0.349 + ], + "angle": 0, + "content": "
DatasetDuration (s)Energy (kWh)
1GPU4GPU1GPU4GPU
BERT-base
RTE (NLI)59278.16e-38.52e-3
MNLI (NLI)670030149.38e-18.64e-1
SQuADv1 (QA)378013565.37e-14.28e-1
SQuADv2 (QA)560420117.95e-16.16e-1
IMDB (Sent.)10744721.51e-11.32e-1
SST2 (Sent.)5873238.11e-28.39e-2
CoNLL2003 (NER)1497372.06e-22.05e-1
CoNLL2012 (NER)14877372.07e-12.14e-1
DistilBERT
RTE (NLI)30184.34e-34.72e-3
MNLI (NLI)344717604.81e-14.98e-1
SQuADv1 (QA)19547922.76e-12.40e-1
SQuADv2 (QA)291611704.12e-13.46e-1
IMDB (Sent.)5492727.75e-27.46e-2
SST2 (Sent.)3061954.21e-25.49e-2
CoNLL2003 (NER)784361.10e-21.18e-1
CoNLL2012 (NER)7624341.07e-11.19e-1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.358, + 0.883, + 0.401 + ], + "angle": 0, + "content": "Table 10: Single-GPU vs. BS and LR-optimized 4-GPU fine-tuning energy requirements across end tasks on the RTX8000 machine, for BERT-base and DistilBERT." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.434, + 0.885, + 0.803 + ], + "angle": 0, + "content": "does not always) use about twice as much energy in both BERT-base and DistilBERT. If we increase batch size x 4 with 4 GPUs (and adjust learning rate accordingly), however, and compare single-GPU fine-tuning with multi-GPU fine-tuning (see Table 10), we observe that energy cost is typically similar or even less than when using 1 GPU, while taking around half as much time or less. In both the \"naive\" and \"optimized\" multi-GPU settings, the single-vs-multi-GPU difference in energy cost and job duration seems to be related to dataset sequence lengths. Tasks with longer sequences (such as QA tasks, and, to a lesser extent, IMDB and RTE) tend to exhibit more consistent and dramatic energy and time efficiency gains than the other tasks when using 4 GPUs. On the other hand, tasks with shorter sequences (such as NER) tend to require more energy with 4 GPUs, even if the wall-clock efficiency may be improved. One way one might interpret this is a large-enough per-device batch size and typical sequence length is necessary for multi-GPU training to be \"worth\" the overhead of communication between GPUs." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.886, + 0.919 + ], + "angle": 0, + "content": "In light of these observations, our general recommendation is that, if one owns a machine with multiple GPUs, one should consider using all available (idle) GPUs for energy- and time-efficient finetuning. Although it is often sufficient to use a single GPU when fine-tuning models of scale similar to ours, and instantaneous energy usage may be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.117, + 0.085, + 0.49, + 0.164 + ], + "angle": 0, + "content": "higher using more GPUs, total energy used may end up being less while also requiring less time, especially if the training sequences tend to be longer. On the other hand, tasks with short sequences are likely best kept to a single GPU." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "9069" + } + ] +] \ No newline at end of file diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_origin.pdf b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f72e0e5307df92c93f19ad37b00eb3adf6c63f53 --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/5c30e66f-a910-426f-bfe5-a165820d39b3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f81d396ca2d927f52d1517981a66ccbfdc17ff79f9e27233b7daeb2c4af594b +size 1412740 diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/full.md b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88832f53e8618f093a6b46de46d9b2e202e49e0f --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/full.md @@ -0,0 +1,300 @@ +# Energy and Carbon Considerations of Fine-Tuning BERT + +Xiaorong Wang $^{1*}$ + +swangxr@gmail.com + +Clara Na $^{2*}$ + +csna@cs.cmu.edu + +Emma Strubell $^{2,3}$ + +strubell@cmu.edu + +Sorelle A. Friedler1 + +sorelle@cs.haverford.edu + +Sasha Luccioni4 + +sasha.luccioni@hf.co + +$^{1}$ Haverford College, $^{2}$ Carnegie Mellon University, $^{3}$ Allen Institute for AI, $^{4}$ Hugging Face + +# Abstract + +Despite the popularity of the pre-train then finetune paradigm in the NLP community, existing work quantifying energy costs and associated carbon emissions has largely focused on language model pre-training. Although a single pre-training run draws substantially more energy than fine-tuning, fine-tuning is performed more frequently by many more individual actors, and thus must be accounted for when considering the energy and carbon footprint of NLP. In order to better characterize the role of finetuning in the landscape of energy and carbon emissions in NLP, we perform a careful empirical study of the computational costs of finetuning across tasks, datasets, hardware infrastructure and measurement modalities. Our experimental results allow us to place fine-tuning energy and carbon costs into perspective with respect to pre-training and inference, and outline recommendations to NLP researchers and practitioners who wish to improve their finetuning energy efficiency. + +# 1 Introduction + +Fine-tuning pre-trained language models is a frequent occurrence in natural language processing (NLP) research and practice, yet the vast majority of work quantifying the energy and carbon footprints of NLP workloads has focused on pretraining (Strubell et al., 2019; Dodge et al., 2022; Luccioni et al., 2023) or inference (Desislavov et al., 2023; Luccioni et al., 2023). The typical lifecycle of an NLP model includes data ingestion, pre-training, fine-tuning and inference, all of which contribute non-trivially to energy use and corresponding emissions (Patterson et al., 2021; Wu et al., 2022). Better understanding of the role that each phase plays in overall energy and carbon footprint is vital to inform policy decisions, yet we still lack basic data quantifying the rela + +tive contributions due to different aspects of model development and use (Kaack et al., 2022). + +In this work we perform an empirical study to quantify the energy requirements of language model fine-tuning, including in the context of pretraining and inference energy requirements. While this may seem like it should be a straightforward calculation, there are several variables that can influence compute time and energy consumption, ranging from: (1) the type of hardware used for both pre-training and fine-tuning (since this usually differs between the two), (2) the type of task and the type of computation required to carry it out, and (3) intrinsic characteristics of the dataset, such as average sequence length, its similarity with the pre-training dataset, etc. + +In order to isolate the factors that have the most influence on fine-tuning dynamics, we compare fine-tuning energy use across a suite of common supervised NLP datasets including the tasks of entailment, sentiment analysis, question answering, and named entity recognition, and with training data sizes ranging from 6K to 400K examples. We also measure energy use across two different sets of hardware, using the CodeCarbon (Schmidt et al., 2021) software package and a physical energy measurement device at the wall, to quantify variance due to physical factors. To enable carefully controlled comparison of the roles of pre-training and fine-tuning in NLP model lifecycle energy use, we additionally pre-train BERT variants from scratch on the same hardware. We find that pre-training BERT is equivalent to anywhere from 400 (MNLI) to 45,000 (RTE) fine-tuning runs depending on the dataset size, and that number of training tokens1 is a reasonable heuristic for estimating fine-tuning energy use. Further comparison of fine-tuning in + +ference energy intensity across tasks confirms that example sequence length holds a much stronger influence on energy intensity in the fine-tuning phase than in the inference phase, in alignment with expectations from previous work (Zhou et al., 2021). Together, our observations contextualize the energy and carbon requirements of fine-tuning in the broader model lifecycle and highlight the need to study fine-tuning energy efficiency separately from pre-training and inference workloads in NLP models. We hope that our careful measurement of the relative costs of different NLP workloads will serve as a valuable datapoint informing decision-making both within and beyond the NLP community. + +# 2 Related Work + +Measurement of energy consumption and carbon emissions of NLP models has become an active area of research since it was first identified that modern state-of-the-art models based on deep learning can produce substantial greenhouse gas emissions due to the energy required to train them (Strubell et al., 2019; Schwartz et al., 2020). These measurements have mostly focused on two research directions. First, there has been a series of empirical studies on different model architectures, focused on estimating the carbon emissions generated by their training process and the relative efficiency of different methods (Naidu et al., 2021; Patterson et al., 2021, 2022). Recent work by Luccioni et al. has built upon this, aiming to encompass the embodied emissions of manufacturing computing hardware as well as those produced via the inference process (2023). There has also been complementary work measuring the energy used by Transformer models during inference and ways of predicting those costs for different architectures and models (Cao et al., 2021; Ang et al., 2022). + +Closest to our work, the HULK benchmark (Zhou et al., 2021) was proposed to measure the relative efficiency-accuracy trade-offs of different pre-trained models, measuring the wall-clock time for different pre-trained models to reach a target accuracy on one of three fine-tuning tasks. Different from Zhou et al. (2021), our work explicitly focuses on the energy and carbon required for fine-tuning (theirs uses time and financial cost as proxies), evaluates a wider variety of fine-tuning tasks and hardware settings in order to elucidate the factors that predict fine-tuning energy requirements, and further contextualizes fine-tuning energy re + +requirements in the bigger picture of ML model lifecycle emissions. + +Another related direction of research examines the dynamics of pre-training and fine-tuning of language models and the influence of factors like random seeds and early stopping (Dodge et al., 2020), scaling (Tay et al., 2022) and learning dynamics (Hao et al., 2020). While all of these studies have shed important light on these processes, in practice most of the decisions made remain empirical, with practitioners either referring to previous work (when hyperparameters and other training details are reported), or using techniques such as grid or random search (Bergstra and Bengio, 2012) to converge on optimal parameter values. Our own work builds upon both of these research directions. We study both the pre-training and fine-tuning process, and our experiments for studying their energy intensity are based on the works cited above. + +# 3 Methodology + +Full training details can be found in Appendix A.1. We release code for replicating our measurements.2 We encourage others to run our code on their hardware to add to a repository of measurements across different hardware platforms. + +# 3.1 Pre-training BERT + +In this work we are interested in measuring the energy consumption of fine-tuning, as it compares to other stages of model use: pre-training and inference. In order to establish a comparable baseline for the energy consumption and carbon emissions of pre-training, we pre-trained a BERT-base model (Devlin et al., 2019) from scratch on the same hardware that we use for fine-tuning (Section 3.3). Although in practice pre-training and fine-tuning are often done separately and on different hardware, we fixed the machine for both sets of experiments in order to aid direct comparability of energy usage measurements. Following Devlin et al. (2019), we pre-train our model on the BookCorpus (Zhu et al., 2015) and the 2020 version of Wikipedia (Foundation, 2020), both downloaded from HuggingFace Datasets (Lhoest et al., 2021). + +Our precise pre-training methodology differs slightly from Devlin et al. (2019): our data necessarily differs slightly because the original training corpus was not released along with the model, and we only use the masked language modeling + +(MLM) objective without next sentence prediction (NSP) following Liu et al. (2019), who found that removing NSP did not substantially impact endtask performance. + +In order to assess the relative impact of using a more efficient pre-trained BERT variant, we also followed the DistilBERT (Sanh et al., 2019) recipe, performing knowledge distillation on our pre-trained BERT-base model. + +# 3.2 Fine-tuning BERT + +We evaluate the energy consumption and carbon emissions of the fine-tuning process on the tasks in Table 1. We deliberately chose this selection of end-tasks in order to vary fine-tuning dataset size, task type, and sequence length, while also aligning with tasks commonly used in NLP applications. + +
DatasetTaskExamplesSeq. length
Med.Batch
Wiki+BooksMLM43M128128
RTENLI6K56128
MNLINLI433K3790
SQuADv1QA98K159336
SQuADv2QA142K158336
IMDBSent.50K128128
SST2Sent.70K1040
CoNLL2003NER21K1553
CoNLL2012NER143K2065
+ +Table 1: Pre-training and fine-tuning dataset descriptions. We report two sequence length statistics: Median tokens per sequence in the dataset, and Batch, the average maximum sequence length per batch as predicted by simulated sampling. The latter is included as a more direct predictor of computation cost for dynamic padding. + +All models are fine-tuned on the BERT models described in §3.1. For each fine-tuning task, we use typical fine-tuning hyperparameters specific to the task or user-reported hyperparameters on current fine-tuned models on HuggingFace, in order to mimic the common real-world use cases. For each task, we dynamically pad sequences to the maximum length in each batch. All fine-tuning hyperparameters are reported in Appendix A.1. + +We also report average per-example energy use for inference. All the inference tasks are performed on 1 GPU with batch size 1 on the same machines. + +# 3.3 Hardware Platforms + +To ensure reproducibility and measure variability across hardware platforms, we replicate experiments across two hardware platforms: One A100 machine and one RTX 8000 machine, where each + +machine had four GPUs. Pre-training experiments used all 4 GPUs in each machine. + +All fine-tuning tasks were performed on the same machines, but using only one GPU. This reflects the typical scenario where fine-tuning is done on a single GPU even if the machine itself has more GPUs. To better compare the energy usage results across pre-training and fine-tuning, we also report an energy usage estimate for BERT-base pretraining on 1 RTX 8000 GPU with hyperparameters equivalent to training on 4 GPUs, extrapolated from a 12-hour partial training run. Details are recorded in Appendix A.1. + +# 3.4 Measuring Energy and Carbon + +To measure the energy consumed, we use the software tool CodeCarbon (Schmidt et al., 2021). Recent work has found that the existing libraries and code for estimating the carbon emissions of NLP techniques vary in terms of their accuracy and generalizability to different types of hardware (Bannour et al., 2021). To compensate for this, we calibrate the programmatic energy usage readings with a physical energy meter, with which we record energy readings directly from the compute node during experiments. Subsequently, we calculate a coefficient of expected power loss, 1.059, as the average proportion (over runs across fine-tuning tasks) of physical energy reading vs. programmatic energy measurement. Full results are given in Appendix A.2. Thus, the energy consumed in kWh, denoted as $E$ , is determined via the formula: + +$$ +E (\mathrm {k W h}) = 1. 0 5 9 \cdot \text {c o d e c a r b o n} \mathrm {k W h} +$$ + +Converting the power loss adjusted values to $\mathrm{CO}_{2}$ emissions is done through CodeCarbon using a coefficient specific to the energy grid based on the location of the server from the most recent EPA Power Profiler Data. The conversion factor for the server's location (Pittsburgh, PA) in Table 2 is 1046.1 lbs/MWh, while the factor for the second server's location (Haverford, PA) is 672.8 lbs/MWh. The total kilograms of $\mathrm{CO}_{2}$ emitted, denoted as $C$ , is then determined via: + +$$ +C = E \times 1 0 4 6. 1 \frac {\mathrm {l b}}{\mathrm {M W h}} \times \frac {1 \mathrm {M W h}}{1 0 0 0 \mathrm {k W h}} \times \frac {1 \mathrm {k g}}{2 . 2 0 4 6 2 \mathrm {l b}} +$$ + +We convert the $\mathrm{CO}_{2}$ emissions result to human understandable reference values using the EPA Greenhouse Gas Equivalencies Calculator; in Table 2, we also show the equivalent $\mathrm{CO}_{2}$ emissions of miles driven by an average gasoline-powered passenger vehicle. + +
TaskDatasetTraining / Fine-tuningInference
Energy (kWh)Emissions (kg CO2)TimeEquiv. # PT runsEquiv. milesEnergy (kWh) / 1000 ex.Equiv. # FT runs
MLM4 GPU (1m @ len 128)270.9128.627 hr1.358330——
1 GPU (1m @ len 128)419.6199.1673 hr0.646510——
4 GPU (100k @ len 512)124.158.90114 hr2.965151——
Total 4 GPU368.0174.6357 hr1.000408——
NLIRTE0.0080.00459 s451090.010.794e-31k
MNLI0.9380.4456700 s3921.147.349e-3127k
QASQuAD v10.5370.2553780 s6850.652.157e-3249k
SQuAD v20.7950.3775604 s4630.972.471e-3322k
Sent.IMDB0.1510.0721074 s24410.180.849e-3178k
SST20.0810.038587 s45400.100.691e-312k
NERCoNLL20030.0210.010149 s178860.030.808e-33k
CoNLL20120.2070.0981487 s17780.250.878e-324k
+ +# 4 Results and Discussion + +Table 2 shows energy, carbon, and wall-clock time required to fine-tune BERT-base models on the RTX8000 machine. Results on the A100 machine are recorded in Appendix A.3 in Tables 8 and 9. + +Table 2: Energy consumption of pre-training BERT and fine-tuning on the RTX8000 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent $\mathrm{CO}_{2}$ emissions. 1 GPU pre-training costs are extrapolated from a shorter pre-training run lasting only a few hours. Total cost of fine-tuning is derived from 900k steps of pre-training on sequences of length 128 followed by 100k steps on sequences of length 512, following (Devlin et al., 2019). Energy consumption for inference is calculated using single-example batches. + +
TaskDatasetEnergy (kWh)TimeEmiss. (kg CO2)
Distil.Wiki+Books187.74175.5 hr89.08
NLIRTE0.00430 s0.002
MNLI0.4813447 s0.228
QASQuAD v10.2761954 s0.131
SQuAD v20.4122916 s0.196
Sent.IMDB0.077549 s0.037
SST0.042306 s0.020
NERCoNLL20030.01178 s0.005
CoNLL20120.107762 s0.051
+ +Table 3: Energy consumption of training (distillation) and fine-tuning DistilBERT on the RTX8000 machine. Energy calculations are the same as in Table 2. Distillation is performed on a pre-trained model, and so the true "total" cost includes the pre-training cost as well. + +# 4.1 Pre-training and Distillation + +We observe that it requires an additional $50\%$ of the energy cost of pre-training in order to perform knowledge distillation, but it takes nearly $50\%$ less energy to fine-tune on the same tasks using DistilBERT vs. normal BERT (see Table 3). By our estimate, one can fully amortize the up-front cost of distillation within anywhere from 86 fine-tuning + +runs of an MNLI-like task, to 47k fine-tunings on an RTE-like task. DistilBERT fine-tuning results on the A100 machine are in Appendix A.3. + +# 4.2 Comparing and Predicting Fine-tuning Emissions + +We find that, controlling for hardware, energy consumption scales most predictably with wall clock time and number of tokens encountered during training (including the pad tokens added to sequences to match the maximum sequence length in a batch). The linear relationship between energy consumption and total number of tokens holds similarly on both machines (see Figure 1). Additionally, we observe a consistently higher energy consumption in the RTX 8000 GPU machine. This is likely due to the higher energy overhead and the (in)efficiency of the hardware compared to the A100 GPUs. Other figures in Appendix A.3 illustrate that, in contrast, energy requirements as a function of optimization steps or even number of examples in the dataset can vary significantly across datasets and tasks. + +![](images/f5cc94fe344b22ab526108be9eb3d4542409fbbb51295da988f13c0f7a05264c.jpg) +Figure 1: Total energy consumed (kWh) is strongly related to number of tokens seen for BERT models on both A100 and RTX 8000 GPU machines, although the relationship is more predictive on the RTX 8000 machine and energy usage is less consistent with DistilBERT. Note that both axes are in log scale. An alternative view of similar data in Figure 4 distinguishes pre-training workloads' energy consumption slightly from that of fine-tuning tasks. + +
DatasetSeq. lengthkWh / 1k ex.
Med.BatchInf.FT
RTE (NLI)561280.75e-31.09e-3
MNLI (NLI)37900.70e-30.80e-3
SQUADv1 (QA)1593361.08e-33.04e-3
SQUADv2 (QA)1583361.24e-33.02e-3
IMDB (Sent.)1281280.80e-31.21e-3
SST2 (Sent.)10400.65e-30.40e-3
CoNLL2003 (NER)15530.76e-30.49e-3
CoNLL2012 (NER)20650.83e-30.60e-3
+ +Table 4: Inference vs. fine-tuning energy requirements across end tasks. We see that fine-tuning energy usage varies according to sequence length much more widely than inference energy usage. + +# 4.3 Fine-tuning vs Pre-training + +Even for the more reliable predictors of energy consumption and carbon emissions (duration of training and number of tokens processed), the energy cost profiles of pre-training vs. fine-tuning are different, likely due to differences in training infrastructure, training objectives, and sequence lengths typically seen in pre-training vs. fine-tuning (see Figures 4, 3, 5, and 6). Pre-training in general is almost always performed over multiple GPUs which incurs energy costs from communication between GPUs, and often also with gradient accumulation to accommodate large batches. Moreover, sequences are packed together such that batches consist largely or entirely of sequences of identical length equal to the maximum sequence length for + +the model. + +On the other hand, there are many types of finetuning tasks where examples consist of sequences of varying lengths significantly shorter than the maximum length that the model has been trained on, as shown in Table 1. Since, effective sequence lengths are determined dynamically at training time (where sequences are padded to the maximum length in each given batch), total training time is not as simple to extrapolate from standard measures of dataset size as in pre-training. + +# 4.4 Fine-tuning vs. Inference + +Although we do observe that per-example inference efficiency costs are related to sequence lengths, there is overall less variation across datasets and tasks in inference costs compared to fine-tuning costs (see Table 4). This mirrors an observation noted in the HULK benchmark (Zhou et al., 2021), though to the best of our knowledge ours is the first to explicitly draw comparisons across task types and different aspects of dataset size (i.e. number of examples and examples' sequence lengths). + +# 4.5 Single vs. Multiple GPUs + +In general, typical hardware and data settings for pre-training and fine-tuning tend to differ. Though to the best of our knowledge it is less common to fine-tune causal LMs of this scale on multiple GPUs, we present additional results from multi-GPU fine-tuning on the $4 \times$ RTX8000 machine with the same fine-tuning tasks in Table 10 in Appendix A.3. Our recommendations from an energy efficiency standpoint align with common rules of thumb for effective utilization of hardware; if the resources would be idle otherwise, one could reasonably consider increasing batch size and learning rate to saturate the available hardware for both time- and energy-efficient training. + +# 5 Conclusion + +We share a procedure for rigorous measurement of energy consumption from causal LM fine-tuning given multiple concrete hardware settings. We hope our work is useful to researchers and practitioners who are interested in obtaining measurements for their own specific hardware, gaining intuitions about factors affecting relative energy costs of different types of fine-tuning tasks, or understanding these energy costs in context of the model lifecycle. + +# Limitations + +While our work provides important first steps towards a clearer understanding of model finetuning's impact on the environment, we note that our experimentation is limited to various token classification, sequence classification, and question answering tasks with BERT-base and DistilBERT-base models. We do not make claims or extrapolations about much larger language models, or models with different architectures, as well as other types of tasks such as summarization. Future work in this direction can expand the number of tasks that we consider as well as feature different architectures such as RoBERTa (Liu et al., 2019). + +Additionally, the on-premises hardware infrastructure used for our experimentation is realistic and typical of compute resources in academic settings, but we provide no firsthand evidence of fine-tuning emissions profiles expected from either local model training (where the impracticality of pretraining makes direct comparisons with fine-tuning emissions difficult) or fine-tuning on hardware that is part of much larger scale infrastructure such as on a public cloud. Furthermore, we expect that use of specialized hardware such as TPUs (as opposed to GPUs, which we use) would be associated with different emissions profiles. + +# Ethics Statement + +Training and deploying ML models and systems requires large amounts of energy, the production of which results in the emission of greenhouse gases. While the goal of our research is to contribute towards a better understanding of the factors that influence these emissions, by carrying out our experiments, we were ourselves responsible for the emission of $350\mathrm{kg}$ of carbon equivalents. We release our code and the data generated by our experiments to maximize the transparency and reproducibility of our work. + +# Acknowledgements + +We are grateful to Victor Sanh for his time and patience in answering questions about BERT and DistilBERT pre-training. We would also like to thank our anonymous reviewers for taking the time to provide helpful feedback. + +This work was supported in part by a grant from the National Science Foundation Graduate Research Fellowship Program under Grant No. + +DGE2140739. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. + +# References + +Phyllis Ang, Bhuwan Dhingra, and Lisa Wu Wills. 2022. Characterizing the efficiency vs. accuracy trade-off for long-context NLP models. In Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP, pages 113-121, Dublin, Ireland. Association for Computational Linguistics. +Nesrine Bannour, Sahar Ghannay, Aurélie Néveol, and Anne-Laure Ligozat. 2021. Evaluating the carbon footprint of NLP methods: a survey and analysis of existing tools. In EMNLP, Workshop SustainNLP. +James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(10):281-305. +Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. 2021. IrEne: Interpretable energy prediction for transformers. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2145-2157, Online. Association for Computational Linguistics. +Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. 2023. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. *Sustainable Computing: Informatics and Systems*, 38:100857. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. 2020. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305. +Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the carbon intensity of ai in cloud instances. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1877-1894. +Wikipedia Foundation. 2020. Wikipedia downloads. + +Yaru Hao, Li Dong, Furu Wei, and Ke Xu. 2020. Investigating learning dynamics of bert fine-tuning. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 87-92. +Lynn H. Kaack, Priya L. Donti, Emma Strubell, George Kamiya, Felix Creutzig, and David Rolnick. 2022. Aligning artificial intelligence with climate change mitigation. Nature Climate Change, 12(6):518-527. +Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. +Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. 2023. Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253):1-15. +Rakshit Naidu, Harshita Diddee, Ajinkya Mulay, Aleti Vardhan, Krithika Ramesh, and Ahmed Zamzam. 2021. Towards quantifying the carbon emissions of differentially private machine learning. arXiv preprint arXiv:2107.06946. +David Patterson, Joseph Gonzalez, Urs Holzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. The carbon footprint of machine learning training will plateau, then shrink. Computer, 55(7):18-28. +David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350. +Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108. + +Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Ziyao Wang, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luccioni. 2021. Codecarbon: estimate and track carbon emissions from machine learning computing. +Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2020. Green AI. Communications of the ACM, 63(12):54-63. +Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics. +Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2022. Scale efficiently: Insights from pretraining and finetuning transformers. In International Conference on Learning Representations. +C.-J. Wu et al. 2022. Sustainable AI: Environmental implications, challenges, and opportunities. In $MLSys$ . +Xiyou Zhou, Zhiyu Chen, Xiaoyong Jin, and William Yang Wang. 2021. HULK: An energy efficiency benchmark platform for responsible natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 329-336, Online. Association for Computational Linguistics. +Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV). + +# A Appendix + +# A.1 Pre-training and Fine-tuning Details + +For pre-training BERT with MLM, we mainly follow what is listed in Devlin et al. (2019). The specific hyperparameters used for pre-training are listed in Table 5 and were used on both machines. + +
Hyperparameters
maximum steps1000000
training batch size per device64
evaluation batch size per device64
maximum sequence length128
learning rate1e-4
warmup steps10000
weight decay0.01
+ +Table 5: Hyperparameters used for pre-training BERT + +In Table 6, we list the set of hyperparameters used to fine-tune each task. All fine-tuning tasks were run on both machines. + +Hardware details The A100 machine is located in Haverford, Pennsylvania, USA, and has 4x NVIDIA A100 GPUs with 40GB GDDR SDRAM, 376GB main memory and 32 Intel Xeon processors. The RTX 8000 machine is located in Pittsburgh, Pennsylvania, USA, and has 4x NVIDIA Quadro RTX 8000 GPUs with 48GB GDDR SDRAM, 36 Intel Xeon processors and 251GB RAM. + +# A.2 Kill-A-Watt Measurements + +Energy is lost during the process of transferring energy from a power source to the machine. The coefficient for the loss is acquired through the readings from Kill-A-Watt devices. The device measures the instantaneous Watt reading extracted from the wall and displays on the monitor. The measurements for the A100 machine were recorded only on the single node of the cluster containing it (the RTX8000 machine is not part of a larger cluster). The GPU node is connected to two outlets, and we plug separate Kill-A-Watt devices into both. For each instantaneous reading, we read off of and sum up the readings on both Kill-A-Watt devices. To best compare between the package reading and the wall reading, we read off of the device at the same 15 second interval that CodeCarbon records the energy consumed. To convert each instantaneous Watts readings into Kilowatt-Hours, we follow the formula: + +$$ +\mathrm {k W h} = \frac {1 5 \text {s e c o n d s}}{3 6 0 0 \text {s e c o n d s}} \times \frac {\text {w a t t s} \times \text {h o u r s}}{1 0 0 0} +$$ + +We sum up all the calculations for the entire run of a fine-tuning experiment. Then, we can compare the sum of the wall readings with the Code Carbon energy consumption recording. We divide the wall reading over the package reading to get a coefficient, measuring the more realistic energy used. The setup of an instantaneous reading on the A100 machine is shown in Figure 2. The recordings of the wall readings on the A100 GPU machine are recorded in Table 7. + +We ran these wall reading experiments for both machines and obtained a coefficient of 1.09 on the A100 machine, and a coefficient of 1.05 on the RTX machine. + +![](images/23cc7a984d5ac1e8681ad75dba598a06618c597141aee5221814227c57bce57f.jpg) +Figure 2: Kill-A-Watt wall reading measurement setup + +# A.3 Additional Results + +BERT pre-training and fine-tuning results Table 8 records the energy consumption results from the A100 GPU machine, located in Haverford, PA. The conversion factor for this location is 672.8 lbs $\mathrm{CO}_{2} / \mathrm{MWh}$ . We then calculate the emissions using raw output from CodeCarbon, which is listed in Table 8. + +DistilBERT results Table 9 records the energy consumption and emission on the A100 GPU machine. Distillation is not done on this machine, and all the fine-tunings are done on the DistilBERT trained on the RTX8000 machine. The results on the A100 machine show a similar trend that fine-tunings on DistilBERT takes around $50\%$ less energy than on BERT base models. + +![](images/8747192e9dcea7336391a18c5da5097b74439f1090a7ab3d2a05676346772234.jpg) +Figure 3: Energy consumed (kWh) against total tokens seen on A100 machine + +Training duration and energy consumption From Figure 5 and 6, we see that there is a strong correlation between training time and energy consumption, which holds across our models and hardware settings (Figure 7). However, similarly to tokens seen, we observe that pre-training exhibits a slightly different energy consumption profile, as do question answering tasks on the A100 machine (which have the longest sequences out of our finetuning tasks). When training time estimates are not feasible in advance (such as in certain hyperparam + +
RTEMNLIIMDBSST2SQuAD v1SQuAD v2CoNLL 2003CoNLL 2012
epochs33532233
train bs3232163232323232
eval bs3232163232323232
max seq len128128128128384384128128
doc stride128128
lr2e-52e-52e-52e-53e-53e-55e-55e-5
+ +Table 6: Hyperparameters used for fine-tuning tasks + +
TaskRandom seedWall reading converted (kWh)Code Carbon reading (kWh)Power Loss Co-efficient
SST2 (no AC)420.1030.0881.170
SST21230.0590.0551.057
IMDB420.0700.0611.134
IMDB420.0790.0711.122
IMDB1230.07930.0731.092
RTE420.006410.006021.065
+ +Table 7: Kill-A-Watt experiment recordings for the A100 GPU machine + +
TaskDatasetEnergy (kWh)Emissions (kg CO2)Training time (s)Equiv. # PT runsEquivalent emissions of miles driven
MLMWiki + Books146.15044.6027265531144
fine-tuningRTE0.0020.00219584420.005
MNLI0.3560.23924004060.613
SQuAD v10.1710.11510618460.295
SQuAD v20.2630.17715805490.454
IMDB0.0620.019047821140.048
SST20.0510.03437728580.087
CoNLL20030.0110.00778134110.018
CoNLL20120.0910.06165415840.156
+ +Table 8: Energy consumption of pre-training BERT and fine-tuning on the A100 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent $\mathrm{CO}_{2}$ emissions. Note that here, unlike in Table 2, the "baseline" pre-training cost is fixed to the cost of 1 million steps of masked language modeling of sequences of length 128. + +
TaskDatasetEnergyTimeEmiss.
EntailmentRTE0.001100.001
MNLI0.21015390.141
QASQuAD v10.0915660.061
SQuAD v20.1419150.095
SentimentIMDB0.0453410.012
SST0.0342620.023
NERCoNLL20030.007570.005
CoNLL20120.0594480.040
+ +Table 9: Energy consumption of training (distillation) and fine-tuning DistilBERT on the A100 machine. Units and energy calculations are the same as given in Table 8. + +eter sweeps), we recommend that researchers and practitioners use token counts estimates (including dynamic padding tokens) if they have reasonable knowledge of their data. + +As indicated in Figure 8 and Figure 9, energy increases as the optimization steps increase. This is not surprising given the correlation between training time and energy consumption. However, we see that for different tasks, the energy required for each step is very different. Each step of pre-training takes a longer time, likely due to the higher batch size than all fine-tuning tasks. For QA tasks, the per-step energy consumption is higher than other + +![](images/e32afc811c7479f199676d9b1914dd4389d41d2b25501b5d722d5edd3cebb7d0.jpg) +Figure 4: Energy consumed (kWh) against total tokens seen on RTX8000 machine + +![](images/3c98d08b317fc611f443cdc3e699e81fb24739f221ce843e4c60e38586d646e9.jpg) +Figure 7: Total training time is strongly predictive of total energy consumed (kWh) for both BERT and DistilBERT models, on both A100 and RTX 8000 GPU machines. Note that both axes are in log scale. Different from token counts, the trendlines themselves are similar as well across models and hardware settings. + +![](images/c9aea71968f5a63e944731649adfcb202f8fa9da9fc67b3758e83eca27f240c1.jpg) +Figure 5: Energy consumed (kWh) against training time (s) on A100 machine, showing the first 30 minutes + +tasks. This is likely due to the maximum sequence length of 384 being higher than for the other tasks. + +Figure 10 shows the number of examples of the task and the energy consumed in log scale. Similar to Figure 1, we see a direct correlation between + +![](images/f002ed9f7b98d360afc2f3150ee0f3e9031c9fff200615ceea9a4b79653b4005.jpg) + +![](images/f42988e7016a356666720e5b891c9f3f1ab238a3f11b4c21ad8dee8e78e6c64a.jpg) +Figure 8: Energy consumed (kWh) against optimization steps on A100 machine machine + +![](images/fe29ab73b8c948511d25e7107c86a087e409a5522fed2563a589be714454e0d4.jpg) +Figure 6: Energy consumed (kWh) against training time (s) on RTX8000 machine, showing the first 30 minutes +Figure 9: Energy consumed (kWh) against optimization steps on RTX8000 machine + +![](images/2f26e3758d0114bec7d6f2f0e0062c9ba1dce8585899e7e842bd698bf6648e64.jpg) +Figure 10: Energy consumed (kWh) against total number of examples seen on both A100 and RTX8000 machines. Comparing with Figure 1, we see that merely counting number of training examples is much less predictive of energy consumption than accounting for example sequence lengths along with batch size (which affects the maximum sequence length in each batch) + +higher number of training examples and higher energy usage on both machines. + +![](images/12fafe6f47ae862e264f38b522c97bacfcf9ce2919844c5c644a9abf412dbe7c.jpg) +Figure 11: $\mathrm{CO}_{2}$ emissions (kg) against task accuracy for SST-2 fine-tuning on RTX 8000 GPU machine + +Figure 11 shows the accuracy of SST2 task as $\mathrm{CO}_{2}$ emissions (kg) increases. As shown previously in Figure 6, energy consumption increases as time increases. Generally, as emission increases, accuracy increases. We can see that as emissions goes up, accuracy is trending towards converging. + +Single- vs. multi-GPU fine-tuning We fine-tune BERT-base and DistilBERT-base models on the RTX8000 machine as well. If the same hyperparameters as the single-GPU setting are used naively (i.e. the same batch size is split over 4 GPUs), fine-tuning using 4 GPUs can (but does not always) take even longer than using just 1 GPU, and can (but + +
DatasetDuration (s)Energy (kWh)
1GPU4GPU1GPU4GPU
BERT-base
RTE (NLI)59278.16e-38.52e-3
MNLI (NLI)670030149.38e-18.64e-1
SQuADv1 (QA)378013565.37e-14.28e-1
SQuADv2 (QA)560420117.95e-16.16e-1
IMDB (Sent.)10744721.51e-11.32e-1
SST2 (Sent.)5873238.11e-28.39e-2
CoNLL2003 (NER)1497372.06e-22.05e-1
CoNLL2012 (NER)14877372.07e-12.14e-1
DistilBERT
RTE (NLI)30184.34e-34.72e-3
MNLI (NLI)344717604.81e-14.98e-1
SQuADv1 (QA)19547922.76e-12.40e-1
SQuADv2 (QA)291611704.12e-13.46e-1
IMDB (Sent.)5492727.75e-27.46e-2
SST2 (Sent.)3061954.21e-25.49e-2
CoNLL2003 (NER)784361.10e-21.18e-1
CoNLL2012 (NER)7624341.07e-11.19e-1
+ +Table 10: Single-GPU vs. BS and LR-optimized 4-GPU fine-tuning energy requirements across end tasks on the RTX8000 machine, for BERT-base and DistilBERT. + +does not always) use about twice as much energy in both BERT-base and DistilBERT. If we increase batch size x 4 with 4 GPUs (and adjust learning rate accordingly), however, and compare single-GPU fine-tuning with multi-GPU fine-tuning (see Table 10), we observe that energy cost is typically similar or even less than when using 1 GPU, while taking around half as much time or less. In both the "naive" and "optimized" multi-GPU settings, the single-vs-multi-GPU difference in energy cost and job duration seems to be related to dataset sequence lengths. Tasks with longer sequences (such as QA tasks, and, to a lesser extent, IMDB and RTE) tend to exhibit more consistent and dramatic energy and time efficiency gains than the other tasks when using 4 GPUs. On the other hand, tasks with shorter sequences (such as NER) tend to require more energy with 4 GPUs, even if the wall-clock efficiency may be improved. One way one might interpret this is a large-enough per-device batch size and typical sequence length is necessary for multi-GPU training to be "worth" the overhead of communication between GPUs. + +In light of these observations, our general recommendation is that, if one owns a machine with multiple GPUs, one should consider using all available (idle) GPUs for energy- and time-efficient finetuning. Although it is often sufficient to use a single GPU when fine-tuning models of scale similar to ours, and instantaneous energy usage may be + +higher using more GPUs, total energy used may end up being less while also requiring less time, especially if the training sequences tend to be longer. On the other hand, tasks with short sequences are likely best kept to a single GPU. \ No newline at end of file diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/images.zip b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..843230d110c6bdb13fb720a7f44f6b38ae812f47 --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55d064893fe99c4b60638cc8ba97fa991d3e0c6683581c6b6f99ff0dbea90f6d +size 745040 diff --git a/2023/Energy and Carbon Considerations of Fine-Tuning BERT/layout.json b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71f4ddee6afd9a885021e1ffe4f24ad79bf8499a --- /dev/null +++ b/2023/Energy and Carbon Considerations of Fine-Tuning BERT/layout.json @@ -0,0 +1,6891 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 70, + 476, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 70, + 476, + 88 + ], + "spans": [ + { + "bbox": [ + 119, + 70, + 476, + 88 + ], + "type": "text", + "content": "Energy and Carbon Considerations of Fine-Tuning BERT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 157, + 96, + 248, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 96, + 248, + 111 + ], + "spans": [ + { + "bbox": [ + 157, + 96, + 248, + 111 + ], + "type": "text", + "content": "Xiaorong Wang" + }, + { + "bbox": [ + 157, + 96, + 248, + 111 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 147, + 112, + 251, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 112, + 251, + 124 + ], + "spans": [ + { + "bbox": [ + 147, + 112, + 251, + 124 + ], + "type": "text", + "content": "swangxr@gmail.com" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 367, + 96, + 424, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 96, + 424, + 110 + ], + "spans": [ + { + "bbox": [ + 367, + 96, + 424, + 110 + ], + "type": "text", + "content": "Clara Na" + }, + { + "bbox": [ + 367, + 96, + 424, + 110 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 351, + 112, + 442, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 112, + 442, + 123 + ], + "spans": [ + { + "bbox": [ + 351, + 112, + 442, + 123 + ], + "type": "text", + "content": "csna@cs.cmu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 137, + 192, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 137, + 192, + 151 + ], + "spans": [ + { + "bbox": [ + 101, + 137, + 192, + 151 + ], + "type": "text", + "content": "Emma Strubell" + }, + { + "bbox": [ + 101, + 137, + 192, + 151 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 153, + 196, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 153, + 196, + 164 + ], + "spans": [ + { + "bbox": [ + 99, + 153, + 196, + 164 + ], + "type": "text", + "content": "strubell@cmu.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 246, + 138, + 346, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 138, + 346, + 151 + ], + "spans": [ + { + "bbox": [ + 246, + 138, + 346, + 151 + ], + "type": "text", + "content": "Sorelle A. Friedler1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 224, + 153, + 369, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 224, + 153, + 369, + 164 + ], + "spans": [ + { + "bbox": [ + 224, + 153, + 369, + 164 + ], + "type": "text", + "content": "sorelle@cs.haverford.edu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 403, + 138, + 487, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 138, + 487, + 151 + ], + "spans": [ + { + "bbox": [ + 403, + 138, + 487, + 151 + ], + "type": "text", + "content": "Sasha Luccioni4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 386, + 153, + 506, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 153, + 506, + 164 + ], + "spans": [ + { + "bbox": [ + 386, + 153, + 506, + 164 + ], + "type": "text", + "content": "sasha.luccioni@hf.co" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "spans": [ + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "text", + "content": "Haverford College, " + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "text", + "content": "Carnegie Mellon University, " + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "text", + "content": "Allen Institute for AI, " + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 179, + 511, + 194 + ], + "type": "text", + "content": "Hugging Face" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 84, + 238, + 274, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 238, + 274, + 501 + ], + "spans": [ + { + "bbox": [ + 84, + 238, + 274, + 501 + ], + "type": "text", + "content": "Despite the popularity of the pre-train then finetune paradigm in the NLP community, existing work quantifying energy costs and associated carbon emissions has largely focused on language model pre-training. Although a single pre-training run draws substantially more energy than fine-tuning, fine-tuning is performed more frequently by many more individual actors, and thus must be accounted for when considering the energy and carbon footprint of NLP. In order to better characterize the role of finetuning in the landscape of energy and carbon emissions in NLP, we perform a careful empirical study of the computational costs of finetuning across tasks, datasets, hardware infrastructure and measurement modalities. Our experimental results allow us to place fine-tuning energy and carbon costs into perspective with respect to pre-training and inference, and outline recommendations to NLP researchers and practitioners who wish to improve their finetuning energy efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 68, + 513, + 154, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 513, + 154, + 527 + ], + "spans": [ + { + "bbox": [ + 68, + 513, + 154, + 527 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 536, + 291, + 753 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 536, + 291, + 753 + ], + "spans": [ + { + "bbox": [ + 67, + 536, + 291, + 753 + ], + "type": "text", + "content": "Fine-tuning pre-trained language models is a frequent occurrence in natural language processing (NLP) research and practice, yet the vast majority of work quantifying the energy and carbon footprints of NLP workloads has focused on pretraining (Strubell et al., 2019; Dodge et al., 2022; Luccioni et al., 2023) or inference (Desislavov et al., 2023; Luccioni et al., 2023). The typical lifecycle of an NLP model includes data ingestion, pre-training, fine-tuning and inference, all of which contribute non-trivially to energy use and corresponding emissions (Patterson et al., 2021; Wu et al., 2022). Better understanding of the role that each phase plays in overall energy and carbon footprint is vital to inform policy decisions, yet we still lack basic data quantifying the rela" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 213, + 524, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 213, + 524, + 239 + ], + "spans": [ + { + "bbox": [ + 302, + 213, + 524, + 239 + ], + "type": "text", + "content": "tive contributions due to different aspects of model development and use (Kaack et al., 2022)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 242, + 526, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 242, + 526, + 431 + ], + "spans": [ + { + "bbox": [ + 302, + 242, + 526, + 431 + ], + "type": "text", + "content": "In this work we perform an empirical study to quantify the energy requirements of language model fine-tuning, including in the context of pretraining and inference energy requirements. While this may seem like it should be a straightforward calculation, there are several variables that can influence compute time and energy consumption, ranging from: (1) the type of hardware used for both pre-training and fine-tuning (since this usually differs between the two), (2) the type of task and the type of computation required to carry it out, and (3) intrinsic characteristics of the dataset, such as average sequence length, its similarity with the pre-training dataset, etc." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 433, + 526, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 433, + 526, + 718 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 526, + 718 + ], + "type": "text", + "content": "In order to isolate the factors that have the most influence on fine-tuning dynamics, we compare fine-tuning energy use across a suite of common supervised NLP datasets including the tasks of entailment, sentiment analysis, question answering, and named entity recognition, and with training data sizes ranging from 6K to 400K examples. We also measure energy use across two different sets of hardware, using the CodeCarbon (Schmidt et al., 2021) software package and a physical energy measurement device at the wall, to quantify variance due to physical factors. To enable carefully controlled comparison of the roles of pre-training and fine-tuning in NLP model lifecycle energy use, we additionally pre-train BERT variants from scratch on the same hardware. We find that pre-training BERT is equivalent to anywhere from 400 (MNLI) to 45,000 (RTE) fine-tuning runs depending on the dataset size, and that number of training tokens1 is a reasonable heuristic for estimating fine-tuning energy use. Further comparison of fine-tuning in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 731, + 525, + 772 + ], + "type": "text", + "content": "1The \"true\" number of training tokens seen, accounting for dynamic padding of sequences to the maximum length in a batch, is a better predictor than relying on to mean or median number of tokens per example." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 85, + 761, + 189, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 761, + 189, + 772 + ], + "spans": [ + { + "bbox": [ + 85, + 761, + 189, + 772 + ], + "type": "text", + "content": "*Denotes equal contribution." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "9058" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 129, + 795, + 464, + 806 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 795, + 464, + 806 + ], + "spans": [ + { + "bbox": [ + 129, + 795, + 464, + 806 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9058-9069" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "spans": [ + { + "bbox": [ + 165, + 806, + 428, + 817 + ], + "type": "text", + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 293, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 293, + 262 + ], + "type": "text", + "content": "ference energy intensity across tasks confirms that example sequence length holds a much stronger influence on energy intensity in the fine-tuning phase than in the inference phase, in alignment with expectations from previous work (Zhou et al., 2021). Together, our observations contextualize the energy and carbon requirements of fine-tuning in the broader model lifecycle and highlight the need to study fine-tuning energy efficiency separately from pre-training and inference workloads in NLP models. We hope that our careful measurement of the relative costs of different NLP workloads will serve as a valuable datapoint informing decision-making both within and beyond the NLP community." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 274, + 161, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 274, + 161, + 287 + ], + "spans": [ + { + "bbox": [ + 67, + 274, + 161, + 287 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 299, + 292, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 299, + 292, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 299, + 292, + 596 + ], + "type": "text", + "content": "Measurement of energy consumption and carbon emissions of NLP models has become an active area of research since it was first identified that modern state-of-the-art models based on deep learning can produce substantial greenhouse gas emissions due to the energy required to train them (Strubell et al., 2019; Schwartz et al., 2020). These measurements have mostly focused on two research directions. First, there has been a series of empirical studies on different model architectures, focused on estimating the carbon emissions generated by their training process and the relative efficiency of different methods (Naidu et al., 2021; Patterson et al., 2021, 2022). Recent work by Luccioni et al. has built upon this, aiming to encompass the embodied emissions of manufacturing computing hardware as well as those produced via the inference process (2023). There has also been complementary work measuring the energy used by Transformer models during inference and ways of predicting those costs for different architectures and models (Cao et al., 2021; Ang et al., 2022)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 597, + 293, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 293, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 293, + 773 + ], + "type": "text", + "content": "Closest to our work, the HULK benchmark (Zhou et al., 2021) was proposed to measure the relative efficiency-accuracy trade-offs of different pre-trained models, measuring the wall-clock time for different pre-trained models to reach a target accuracy on one of three fine-tuning tasks. Different from Zhou et al. (2021), our work explicitly focuses on the energy and carbon required for fine-tuning (theirs uses time and financial cost as proxies), evaluates a wider variety of fine-tuning tasks and hardware settings in order to elucidate the factors that predict fine-tuning energy requirements, and further contextualizes fine-tuning energy re" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 302, + 71, + 526, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 97 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 97 + ], + "type": "text", + "content": "requirements in the bigger picture of ML model lifecycle emissions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 98, + 527, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 98, + 527, + 329 + ], + "spans": [ + { + "bbox": [ + 302, + 98, + 527, + 329 + ], + "type": "text", + "content": "Another related direction of research examines the dynamics of pre-training and fine-tuning of language models and the influence of factors like random seeds and early stopping (Dodge et al., 2020), scaling (Tay et al., 2022) and learning dynamics (Hao et al., 2020). While all of these studies have shed important light on these processes, in practice most of the decisions made remain empirical, with practitioners either referring to previous work (when hyperparameters and other training details are reported), or using techniques such as grid or random search (Bergstra and Bengio, 2012) to converge on optimal parameter values. Our own work builds upon both of these research directions. We study both the pre-training and fine-tuning process, and our experiments for studying their energy intensity are based on the works cited above." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 338, + 392, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 338, + 392, + 353 + ], + "spans": [ + { + "bbox": [ + 302, + 338, + 392, + 353 + ], + "type": "text", + "content": "3 Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 360, + 527, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 360, + 527, + 428 + ], + "spans": [ + { + "bbox": [ + 302, + 360, + 527, + 428 + ], + "type": "text", + "content": "Full training details can be found in Appendix A.1. We release code for replicating our measurements.2 We encourage others to run our code on their hardware to add to a repository of measurements across different hardware platforms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 438, + 420, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 438, + 420, + 451 + ], + "spans": [ + { + "bbox": [ + 302, + 438, + 420, + 451 + ], + "type": "text", + "content": "3.1 Pre-training BERT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 455, + 527, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 455, + 527, + 685 + ], + "spans": [ + { + "bbox": [ + 302, + 455, + 527, + 685 + ], + "type": "text", + "content": "In this work we are interested in measuring the energy consumption of fine-tuning, as it compares to other stages of model use: pre-training and inference. In order to establish a comparable baseline for the energy consumption and carbon emissions of pre-training, we pre-trained a BERT-base model (Devlin et al., 2019) from scratch on the same hardware that we use for fine-tuning (Section 3.3). Although in practice pre-training and fine-tuning are often done separately and on different hardware, we fixed the machine for both sets of experiments in order to aid direct comparability of energy usage measurements. Following Devlin et al. (2019), we pre-train our model on the BookCorpus (Zhu et al., 2015) and the 2020 version of Wikipedia (Foundation, 2020), both downloaded from HuggingFace Datasets (Lhoest et al., 2021)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 686, + 527, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 686, + 527, + 754 + ], + "spans": [ + { + "bbox": [ + 302, + 686, + 527, + 754 + ], + "type": "text", + "content": "Our precise pre-training methodology differs slightly from Devlin et al. (2019): our data necessarily differs slightly because the original training corpus was not released along with the model, and we only use the masked language modeling" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 760, + 503, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 760, + 503, + 772 + ], + "spans": [ + { + "bbox": [ + 315, + 760, + 503, + 772 + ], + "type": "text", + "content": "2https://github.com/swangxr/FT-energy.git" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 780, + 309, + 791 + ], + "type": "text", + "content": "9059" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 125 + ], + "type": "text", + "content": "(MLM) objective without next sentence prediction (NSP) following Liu et al. (2019), who found that removing NSP did not substantially impact endtask performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 126, + 290, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 126, + 290, + 193 + ], + "spans": [ + { + "bbox": [ + 67, + 126, + 290, + 193 + ], + "type": "text", + "content": "In order to assess the relative impact of using a more efficient pre-trained BERT variant, we also followed the DistilBERT (Sanh et al., 2019) recipe, performing knowledge distillation on our pre-trained BERT-base model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 201, + 183, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 201, + 183, + 215 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 183, + 215 + ], + "type": "text", + "content": "3.2 Fine-tuning BERT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 219, + 290, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 219, + 290, + 300 + ], + "spans": [ + { + "bbox": [ + 67, + 219, + 290, + 300 + ], + "type": "text", + "content": "We evaluate the energy consumption and carbon emissions of the fine-tuning process on the tasks in Table 1. We deliberately chose this selection of end-tasks in order to vary fine-tuning dataset size, task type, and sequence length, while also aligning with tasks commonly used in NLP applications." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 76, + 308, + 282, + 438 + ], + "blocks": [ + { + "bbox": [ + 76, + 308, + 282, + 438 + ], + "lines": [ + { + "bbox": [ + 76, + 308, + 282, + 438 + ], + "spans": [ + { + "bbox": [ + 76, + 308, + 282, + 438 + ], + "type": "table", + "html": "
DatasetTaskExamplesSeq. length
Med.Batch
Wiki+BooksMLM43M128128
RTENLI6K56128
MNLINLI433K3790
SQuADv1QA98K159336
SQuADv2QA142K158336
IMDBSent.50K128128
SST2Sent.70K1040
CoNLL2003NER21K1553
CoNLL2012NER143K2065
", + "image_path": "ef8e05d8a07fff594948bedfce455772696e2e63ce02598badf7fb28843d3ee8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 446, + 291, + 519 + ], + "lines": [ + { + "bbox": [ + 67, + 446, + 291, + 519 + ], + "spans": [ + { + "bbox": [ + 67, + 446, + 291, + 519 + ], + "type": "text", + "content": "Table 1: Pre-training and fine-tuning dataset descriptions. We report two sequence length statistics: Median tokens per sequence in the dataset, and Batch, the average maximum sequence length per batch as predicted by simulated sampling. The latter is included as a more direct predictor of computation cost for dynamic padding." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 531, + 290, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 531, + 290, + 652 + ], + "spans": [ + { + "bbox": [ + 67, + 531, + 290, + 652 + ], + "type": "text", + "content": "All models are fine-tuned on the BERT models described in §3.1. For each fine-tuning task, we use typical fine-tuning hyperparameters specific to the task or user-reported hyperparameters on current fine-tuned models on HuggingFace, in order to mimic the common real-world use cases. For each task, we dynamically pad sequences to the maximum length in each batch. All fine-tuning hyperparameters are reported in Appendix A.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 653, + 290, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 653, + 290, + 692 + ], + "spans": [ + { + "bbox": [ + 67, + 653, + 290, + 692 + ], + "type": "text", + "content": "We also report average per-example energy use for inference. All the inference tasks are performed on 1 GPU with batch size 1 on the same machines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 702, + 192, + 714 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 192, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 192, + 714 + ], + "type": "text", + "content": "3.3 Hardware Platforms" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "content": "To ensure reproducibility and measure variability across hardware platforms, we replicate experiments across two hardware platforms: One A100 machine and one RTX 8000 machine, where each" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 524, + 97 + ], + "type": "text", + "content": "machine had four GPUs. Pre-training experiments used all 4 GPUs in each machine." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 99, + 525, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 99, + 525, + 248 + ], + "spans": [ + { + "bbox": [ + 302, + 99, + 525, + 248 + ], + "type": "text", + "content": "All fine-tuning tasks were performed on the same machines, but using only one GPU. This reflects the typical scenario where fine-tuning is done on a single GPU even if the machine itself has more GPUs. To better compare the energy usage results across pre-training and fine-tuning, we also report an energy usage estimate for BERT-base pretraining on 1 RTX 8000 GPU with hyperparameters equivalent to training on 4 GPUs, extrapolated from a 12-hour partial training run. Details are recorded in Appendix A.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 255, + 477, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 255, + 477, + 269 + ], + "spans": [ + { + "bbox": [ + 302, + 255, + 477, + 269 + ], + "type": "text", + "content": "3.4 Measuring Energy and Carbon" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 272, + 525, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 272, + 525, + 502 + ], + "spans": [ + { + "bbox": [ + 302, + 272, + 525, + 502 + ], + "type": "text", + "content": "To measure the energy consumed, we use the software tool CodeCarbon (Schmidt et al., 2021). Recent work has found that the existing libraries and code for estimating the carbon emissions of NLP techniques vary in terms of their accuracy and generalizability to different types of hardware (Bannour et al., 2021). To compensate for this, we calibrate the programmatic energy usage readings with a physical energy meter, with which we record energy readings directly from the compute node during experiments. Subsequently, we calculate a coefficient of expected power loss, 1.059, as the average proportion (over runs across fine-tuning tasks) of physical energy reading vs. programmatic energy measurement. Full results are given in Appendix A.2. Thus, the energy consumed in kWh, denoted as " + }, + { + "bbox": [ + 302, + 272, + 525, + 502 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 302, + 272, + 525, + 502 + ], + "type": "text", + "content": ", is determined via the formula:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 508, + 497, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 508, + 497, + 521 + ], + "spans": [ + { + "bbox": [ + 332, + 508, + 497, + 521 + ], + "type": "interline_equation", + "content": "E (\\mathrm {k W h}) = 1. 0 5 9 \\cdot \\text {c o d e c a r b o n} \\mathrm {k W h}", + "image_path": "f2bf451a59c76a124cf9f5cd9601dd432fef1802d9600351ea674a18124dfc78.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "spans": [ + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "text", + "content": "Converting the power loss adjusted values to " + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "text", + "content": " emissions is done through CodeCarbon using a coefficient specific to the energy grid based on the location of the server from the most recent EPA Power Profiler Data. The conversion factor for the server's location (Pittsburgh, PA) in Table 2 is 1046.1 lbs/MWh, while the factor for the second server's location (Haverford, PA) is 672.8 lbs/MWh. The total kilograms of " + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "text", + "content": " emitted, denoted as " + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 302, + 527, + 525, + 661 + ], + "type": "text", + "content": ", is then determined via:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 664, + 524, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 664, + 524, + 690 + ], + "spans": [ + { + "bbox": [ + 302, + 664, + 524, + 690 + ], + "type": "interline_equation", + "content": "C = E \\times 1 0 4 6. 1 \\frac {\\mathrm {l b}}{\\mathrm {M W h}} \\times \\frac {1 \\mathrm {M W h}}{1 0 0 0 \\mathrm {k W h}} \\times \\frac {1 \\mathrm {k g}}{2 . 2 0 4 6 2 \\mathrm {l b}}", + "image_path": "c80404c1a04f29b362ae627cd38f305956689c75077e7aec2978728e1f8cc867.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "We convert the " + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": " emissions result to human understandable reference values using the EPA Greenhouse Gas Equivalencies Calculator; in Table 2, we also show the equivalent " + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": " emissions of miles driven by an average gasoline-powered passenger vehicle." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9060" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 68, + 520, + 253 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 520, + 253 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 520, + 253 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 520, + 253 + ], + "type": "table", + "html": "
TaskDatasetTraining / Fine-tuningInference
Energy (kWh)Emissions (kg CO2)TimeEquiv. # PT runsEquiv. milesEnergy (kWh) / 1000 ex.Equiv. # FT runs
MLM4 GPU (1m @ len 128)270.9128.627 hr1.358330——
1 GPU (1m @ len 128)419.6199.1673 hr0.646510——
4 GPU (100k @ len 512)124.158.90114 hr2.965151——
Total 4 GPU368.0174.6357 hr1.000408——
NLIRTE0.0080.00459 s451090.010.794e-31k
MNLI0.9380.4456700 s3921.147.349e-3127k
QASQuAD v10.5370.2553780 s6850.652.157e-3249k
SQuAD v20.7950.3775604 s4630.972.471e-3322k
Sent.IMDB0.1510.0721074 s24410.180.849e-3178k
SST20.0810.038587 s45400.100.691e-312k
NERCoNLL20030.0210.010149 s178860.030.808e-33k
CoNLL20120.2070.0981487 s17780.250.878e-324k
", + "image_path": "09b794f912e76e30c995fc3aa36df1d53b851f63b9c3999047e54287ad42b091.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 354, + 205, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 354, + 205, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 205, + 366 + ], + "type": "text", + "content": "4 Results and Discussion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 375, + 289, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 375, + 289, + 429 + ], + "spans": [ + { + "bbox": [ + 67, + 375, + 289, + 429 + ], + "type": "text", + "content": "Table 2 shows energy, carbon, and wall-clock time required to fine-tune BERT-base models on the RTX8000 machine. Results on the A100 machine are recorded in Appendix A.3 in Tables 8 and 9." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 70, + 439, + 289, + 569 + ], + "blocks": [ + { + "bbox": [ + 67, + 261, + 525, + 334 + ], + "lines": [ + { + "bbox": [ + 67, + 261, + 525, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 525, + 334 + ], + "type": "text", + "content": "Table 2: Energy consumption of pre-training BERT and fine-tuning on the RTX8000 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent " + }, + { + "bbox": [ + 67, + 261, + 525, + 334 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 67, + 261, + 525, + 334 + ], + "type": "text", + "content": " emissions. 1 GPU pre-training costs are extrapolated from a shorter pre-training run lasting only a few hours. Total cost of fine-tuning is derived from 900k steps of pre-training on sequences of length 128 followed by 100k steps on sequences of length 512, following (Devlin et al., 2019). Energy consumption for inference is calculated using single-example batches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 439, + 289, + 569 + ], + "lines": [ + { + "bbox": [ + 70, + 439, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 70, + 439, + 289, + 569 + ], + "type": "table", + "html": "
TaskDatasetEnergy (kWh)TimeEmiss. (kg CO2)
Distil.Wiki+Books187.74175.5 hr89.08
NLIRTE0.00430 s0.002
MNLI0.4813447 s0.228
QASQuAD v10.2761954 s0.131
SQuAD v20.4122916 s0.196
Sent.IMDB0.077549 s0.037
SST0.042306 s0.020
NERCoNLL20030.01178 s0.005
CoNLL20120.107762 s0.051
", + "image_path": "b4a5b5a6b6781ce7f3c64e5b4c66a768dcf7ea46e4f92a9bf3182f9e1eed190b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 577, + 290, + 638 + ], + "lines": [ + { + "bbox": [ + 67, + 577, + 290, + 638 + ], + "spans": [ + { + "bbox": [ + 67, + 577, + 290, + 638 + ], + "type": "text", + "content": "Table 3: Energy consumption of training (distillation) and fine-tuning DistilBERT on the RTX8000 machine. Energy calculations are the same as in Table 2. Distillation is performed on a pre-trained model, and so the true \"total\" cost includes the pre-training cost as well." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 67, + 661, + 227, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 661, + 227, + 673 + ], + "spans": [ + { + "bbox": [ + 67, + 661, + 227, + 673 + ], + "type": "text", + "content": "4.1 Pre-training and Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": "We observe that it requires an additional " + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": " of the energy cost of pre-training in order to perform knowledge distillation, but it takes nearly " + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 67, + 678, + 290, + 772 + ], + "type": "text", + "content": " less energy to fine-tune on the same tasks using DistilBERT vs. normal BERT (see Table 3). By our estimate, one can fully amortize the up-front cost of distillation within anywhere from 86 fine-tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 354, + 524, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 354, + 524, + 395 + ], + "spans": [ + { + "bbox": [ + 302, + 354, + 524, + 395 + ], + "type": "text", + "content": "runs of an MNLI-like task, to 47k fine-tunings on an RTE-like task. DistilBERT fine-tuning results on the A100 machine are in Appendix A.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 416, + 513, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 416, + 513, + 441 + ], + "spans": [ + { + "bbox": [ + 302, + 416, + 513, + 441 + ], + "type": "text", + "content": "4.2 Comparing and Predicting Fine-tuning Emissions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 301, + 453, + 525, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 453, + 525, + 682 + ], + "spans": [ + { + "bbox": [ + 301, + 453, + 525, + 682 + ], + "type": "text", + "content": "We find that, controlling for hardware, energy consumption scales most predictably with wall clock time and number of tokens encountered during training (including the pad tokens added to sequences to match the maximum sequence length in a batch). The linear relationship between energy consumption and total number of tokens holds similarly on both machines (see Figure 1). Additionally, we observe a consistently higher energy consumption in the RTX 8000 GPU machine. This is likely due to the higher energy overhead and the (in)efficiency of the hardware compared to the A100 GPUs. Other figures in Appendix A.3 illustrate that, in contrast, energy requirements as a function of optimization steps or even number of examples in the dataset can vary significantly across datasets and tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 700, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 700, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 700, + 525, + 772 + ], + "type": "text", + "content": "3Note that cheaper inference is often the primary goal of knowledge distillation. Inference is much cheaper than training and therefore requires more to amortize the initial cost of distillation, but inference also occurs much more frequently than training. Models running inference at scale are typically highly optimized with respect to specific deployment settings, so our estimates approximate a lower bound." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 308, + 791 + ], + "type": "text", + "content": "9061" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 295, + 225 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "type": "image", + "image_path": "f5cc94fe344b22ab526108be9eb3d4542409fbbb51295da988f13c0f7a05264c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 237, + 291, + 344 + ], + "lines": [ + { + "bbox": [ + 67, + 237, + 291, + 344 + ], + "spans": [ + { + "bbox": [ + 67, + 237, + 291, + 344 + ], + "type": "text", + "content": "Figure 1: Total energy consumed (kWh) is strongly related to number of tokens seen for BERT models on both A100 and RTX 8000 GPU machines, although the relationship is more predictive on the RTX 8000 machine and energy usage is less consistent with DistilBERT. Note that both axes are in log scale. An alternative view of similar data in Figure 4 distinguishes pre-training workloads' energy consumption slightly from that of fine-tuning tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 357, + 289, + 471 + ], + "blocks": [ + { + "bbox": [ + 69, + 357, + 289, + 471 + ], + "lines": [ + { + "bbox": [ + 69, + 357, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 69, + 357, + 289, + 471 + ], + "type": "table", + "html": "
DatasetSeq. lengthkWh / 1k ex.
Med.BatchInf.FT
RTE (NLI)561280.75e-31.09e-3
MNLI (NLI)37900.70e-30.80e-3
SQUADv1 (QA)1593361.08e-33.04e-3
SQUADv2 (QA)1583361.24e-33.02e-3
IMDB (Sent.)1281280.80e-31.21e-3
SST2 (Sent.)10400.65e-30.40e-3
CoNLL2003 (NER)15530.76e-30.49e-3
CoNLL2012 (NER)20650.83e-30.60e-3
", + "image_path": "1085d6f94d0c10db282f6f816c9f26fafe1527c50b5c4555746981f04826cd54.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 479, + 290, + 528 + ], + "lines": [ + { + "bbox": [ + 67, + 479, + 290, + 528 + ], + "spans": [ + { + "bbox": [ + 67, + 479, + 290, + 528 + ], + "type": "text", + "content": "Table 4: Inference vs. fine-tuning energy requirements across end tasks. We see that fine-tuning energy usage varies according to sequence length much more widely than inference energy usage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 551, + 224, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 551, + 224, + 565 + ], + "spans": [ + { + "bbox": [ + 67, + 551, + 224, + 565 + ], + "type": "text", + "content": "4.3 Fine-tuning vs Pre-training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 291, + 773 + ], + "type": "text", + "content": "Even for the more reliable predictors of energy consumption and carbon emissions (duration of training and number of tokens processed), the energy cost profiles of pre-training vs. fine-tuning are different, likely due to differences in training infrastructure, training objectives, and sequence lengths typically seen in pre-training vs. fine-tuning (see Figures 4, 3, 5, and 6). Pre-training in general is almost always performed over multiple GPUs which incurs energy costs from communication between GPUs, and often also with gradient accumulation to accommodate large batches. Moreover, sequences are packed together such that batches consist largely or entirely of sequences of identical length equal to the maximum sequence length for" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 303, + 71, + 352, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 71, + 352, + 83 + ], + "spans": [ + { + "bbox": [ + 303, + 71, + 352, + 83 + ], + "type": "text", + "content": "the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 84, + 525, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 84, + 525, + 220 + ], + "spans": [ + { + "bbox": [ + 302, + 84, + 525, + 220 + ], + "type": "text", + "content": "On the other hand, there are many types of finetuning tasks where examples consist of sequences of varying lengths significantly shorter than the maximum length that the model has been trained on, as shown in Table 1. Since, effective sequence lengths are determined dynamically at training time (where sequences are padded to the maximum length in each given batch), total training time is not as simple to extrapolate from standard measures of dataset size as in pre-training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 229, + 449, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 229, + 449, + 243 + ], + "spans": [ + { + "bbox": [ + 302, + 229, + 449, + 243 + ], + "type": "text", + "content": "4.4 Fine-tuning vs. Inference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 246, + 526, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 246, + 526, + 382 + ], + "spans": [ + { + "bbox": [ + 302, + 246, + 526, + 382 + ], + "type": "text", + "content": "Although we do observe that per-example inference efficiency costs are related to sequence lengths, there is overall less variation across datasets and tasks in inference costs compared to fine-tuning costs (see Table 4). This mirrors an observation noted in the HULK benchmark (Zhou et al., 2021), though to the best of our knowledge ours is the first to explicitly draw comparisons across task types and different aspects of dataset size (i.e. number of examples and examples' sequence lengths)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 391, + 448, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 391, + 448, + 403 + ], + "spans": [ + { + "bbox": [ + 302, + 391, + 448, + 403 + ], + "type": "text", + "content": "4.5 Single vs. Multiple GPUs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 408, + 525, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 408, + 525, + 597 + ], + "spans": [ + { + "bbox": [ + 302, + 408, + 525, + 597 + ], + "type": "text", + "content": "In general, typical hardware and data settings for pre-training and fine-tuning tend to differ. Though to the best of our knowledge it is less common to fine-tune causal LMs of this scale on multiple GPUs, we present additional results from multi-GPU fine-tuning on the " + }, + { + "bbox": [ + 302, + 408, + 525, + 597 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 302, + 408, + 525, + 597 + ], + "type": "text", + "content": " RTX8000 machine with the same fine-tuning tasks in Table 10 in Appendix A.3. Our recommendations from an energy efficiency standpoint align with common rules of thumb for effective utilization of hardware; if the resources would be idle otherwise, one could reasonably consider increasing batch size and learning rate to saturate the available hardware for both time- and energy-efficient training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 607, + 381, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 607, + 381, + 619 + ], + "spans": [ + { + "bbox": [ + 302, + 607, + 381, + 619 + ], + "type": "text", + "content": "5 Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 629, + 526, + 763 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 629, + 526, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 629, + 526, + 763 + ], + "type": "text", + "content": "We share a procedure for rigorous measurement of energy consumption from causal LM fine-tuning given multiple concrete hardware settings. We hope our work is useful to researchers and practitioners who are interested in obtaining measurements for their own specific hardware, gaining intuitions about factors affecting relative energy costs of different types of fine-tuning tasks, or understanding these energy costs in context of the model lifecycle." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9062" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 71, + 131, + 84 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 94, + 291, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 94, + 291, + 268 + ], + "spans": [ + { + "bbox": [ + 67, + 94, + 291, + 268 + ], + "type": "text", + "content": "While our work provides important first steps towards a clearer understanding of model finetuning's impact on the environment, we note that our experimentation is limited to various token classification, sequence classification, and question answering tasks with BERT-base and DistilBERT-base models. We do not make claims or extrapolations about much larger language models, or models with different architectures, as well as other types of tasks such as summarization. Future work in this direction can expand the number of tasks that we consider as well as feature different architectures such as RoBERTa (Liu et al., 2019)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 270, + 291, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 270, + 291, + 446 + ], + "spans": [ + { + "bbox": [ + 69, + 270, + 291, + 446 + ], + "type": "text", + "content": "Additionally, the on-premises hardware infrastructure used for our experimentation is realistic and typical of compute resources in academic settings, but we provide no firsthand evidence of fine-tuning emissions profiles expected from either local model training (where the impracticality of pretraining makes direct comparisons with fine-tuning emissions difficult) or fine-tuning on hardware that is part of much larger scale infrastructure such as on a public cloud. Furthermore, we expect that use of specialized hardware such as TPUs (as opposed to GPUs, which we use) would be associated with different emissions profiles." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 458, + 158, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 458, + 158, + 470 + ], + "spans": [ + { + "bbox": [ + 68, + 458, + 158, + 470 + ], + "type": "text", + "content": "Ethics Statement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 481, + 291, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 291, + 629 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 291, + 629 + ], + "type": "text", + "content": "Training and deploying ML models and systems requires large amounts of energy, the production of which results in the emission of greenhouse gases. While the goal of our research is to contribute towards a better understanding of the factors that influence these emissions, by carrying out our experiments, we were ourselves responsible for the emission of " + }, + { + "bbox": [ + 67, + 481, + 291, + 629 + ], + "type": "inline_equation", + "content": "350\\mathrm{kg}" + }, + { + "bbox": [ + 67, + 481, + 291, + 629 + ], + "type": "text", + "content": " of carbon equivalents. We release our code and the data generated by our experiments to maximize the transparency and reproducibility of our work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 642, + 170, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 642, + 170, + 655 + ], + "spans": [ + { + "bbox": [ + 68, + 642, + 170, + 655 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 664, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 290, + 731 + ], + "type": "text", + "content": "We are grateful to Victor Sanh for his time and patience in answering questions about BERT and DistilBERT pre-training. We would also like to thank our anonymous reviewers for taking the time to provide helpful feedback." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 291, + 772 + ], + "type": "text", + "content": "This work was supported in part by a grant from the National Science Foundation Graduate Research Fellowship Program under Grant No." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 71, + 527, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 527, + 126 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 527, + 126 + ], + "type": "text", + "content": "DGE2140739. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 148, + 362, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 362, + 160 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 362, + 160 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 166, + 527, + 772 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 304, + 166, + 527, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 166, + 527, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 166, + 527, + 233 + ], + "type": "text", + "content": "Phyllis Ang, Bhuwan Dhingra, and Lisa Wu Wills. 2022. Characterizing the efficiency vs. accuracy trade-off for long-context NLP models. In Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP, pages 113-121, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 525, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 525, + 285 + ], + "type": "text", + "content": "Nesrine Bannour, Sahar Ghannay, Aurélie Néveol, and Anne-Laure Ligozat. 2021. Evaluating the carbon footprint of NLP methods: a survey and analysis of existing tools. In EMNLP, Workshop SustainNLP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 293, + 526, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 526, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 526, + 327 + ], + "type": "text", + "content": "James Bergstra and Yoshua Bengio. 2012. Random search for hyper-parameter optimization. Journal of Machine Learning Research, 13(10):281-305." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 334, + 527, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 527, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 527, + 434 + ], + "type": "text", + "content": "Qingqing Cao, Yash Kumar Lal, Harsh Trivedi, Aruna Balasubramanian, and Niranjan Balasubramanian. 2021. IrEne: Interpretable energy prediction for transformers. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2145-2157, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 441, + 527, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 527, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 527, + 498 + ], + "type": "text", + "content": "Radosvet Desislavov, Fernando Martínez-Plumed, and José Hernández-Orallo. 2023. Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. *Sustainable Computing: Informatics and Systems*, 38:100857." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 505, + 527, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 527, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 527, + 605 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 612, + 527, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 527, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 527, + 668 + ], + "type": "text", + "content": "Jesse Dodge, Gabriel Ilharco, Roy Schwartz, Ali Farhadi, Hannaneh Hajishirzi, and Noah Smith. 2020. Fine-tuning pretrained language models: Weight initializations, data orders, and early stopping. arXiv preprint arXiv:2002.06305." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 675, + 527, + 754 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 527, + 754 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 527, + 754 + ], + "type": "text", + "content": "Jesse Dodge, Taylor Prewitt, Remi Tachet des Combes, Erika Odmark, Roy Schwartz, Emma Strubell, Alexandra Sasha Luccioni, Noah A Smith, Nicole DeCario, and Will Buchanan. 2022. Measuring the carbon intensity of ai in cloud instances. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1877-1894." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 760, + 521, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 760, + 521, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 760, + 521, + 772 + ], + "type": "text", + "content": "Wikipedia Foundation. 2020. Wikipedia downloads." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9063" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 772 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 290, + 139 + ], + "type": "text", + "content": "Yaru Hao, Li Dong, Furu Wei, and Ke Xu. 2020. Investigating learning dynamics of bert fine-tuning. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 87-92." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 148, + 290, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 148, + 290, + 193 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 290, + 193 + ], + "type": "text", + "content": "Lynn H. Kaack, Priya L. Donti, Emma Strubell, George Kamiya, Felix Creutzig, and David Rolnick. 2022. Aligning artificial intelligence with climate change mitigation. Nature Climate Change, 12(6):518-527." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 203, + 290, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 203, + 290, + 390 + ], + "spans": [ + { + "bbox": [ + 69, + 203, + 290, + 390 + ], + "type": "text", + "content": "Quentin Lhoest, Albert Villanova del Moral, Yacine Jernite, Abhishek Thakur, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Julien Plu, Lewis Tunstall, Joe Davison, Mario Šaško, Gunjan Chhablani, Bhavitvya Malik, Simon Brandeis, Teven Le Scao, Victor Sanh, Canwen Xu, Nicolas Patry, Angelina McMillan-Major, Philipp Schmid, Sylvain Gugger, Clément Delangue, Théo Matussière, Lysandre Debut, Stas Bekman, Pierrick Cistac, Thibault Goehringer, Victor Mustar, François Lagunas, Alexander Rush, and Thomas Wolf. 2021. Datasets: A community library for natural language processing. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 175-184, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 400, + 290, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 400, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 69, + 400, + 290, + 456 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 465, + 290, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 465, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 465, + 290, + 510 + ], + "type": "text", + "content": "Alexandra Sasha Luccioni, Sylvain Viguier, and Anne-Laure Ligozat. 2023. Estimating the carbon footprint of bloom, a 176b parameter language model. Journal of Machine Learning Research, 24(253):1-15." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 520, + 290, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 520, + 290, + 575 + ], + "spans": [ + { + "bbox": [ + 69, + 520, + 290, + 575 + ], + "type": "text", + "content": "Rakshit Naidu, Harshita Diddee, Ajinkya Mulay, Aleti Vardhan, Krithika Ramesh, and Ahmed Zamzam. 2021. Towards quantifying the carbon emissions of differentially private machine learning. arXiv preprint arXiv:2107.06946." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 586, + 290, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 586, + 290, + 650 + ], + "spans": [ + { + "bbox": [ + 69, + 586, + 290, + 650 + ], + "type": "text", + "content": "David Patterson, Joseph Gonzalez, Urs Holzle, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David R. So, Maud Texier, and Jeff Dean. 2022. The carbon footprint of machine learning training will plateau, then shrink. Computer, 55(7):18-28." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 662, + 290, + 718 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 662, + 290, + 718 + ], + "spans": [ + { + "bbox": [ + 69, + 662, + 290, + 718 + ], + "type": "text", + "content": "David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon emissions and large neural network training. arXiv preprint arXiv:2104.10350." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 728, + 290, + 772 + ], + "type": "text", + "content": "Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 524, + 536 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 524, + 127 + ], + "type": "text", + "content": "Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Ziyao Wang, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luccioni. 2021. Codecarbon: estimate and track carbon emissions from machine learning computing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 138, + 524, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 138, + 524, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 524, + 171 + ], + "type": "text", + "content": "Roy Schwartz, Jesse Dodge, Noah A Smith, and Oren Etzioni. 2020. Green AI. Communications of the ACM, 63(12):54-63." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 183, + 524, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 524, + 249 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 524, + 249 + ], + "type": "text", + "content": "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in NLP. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 260, + 524, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 524, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 524, + 327 + ], + "type": "text", + "content": "Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2022. Scale efficiently: Insights from pretraining and finetuning transformers. In International Conference on Learning Representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 524, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 524, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 524, + 360 + ], + "type": "text", + "content": "C.-J. Wu et al. 2022. Sustainable AI: Environmental implications, challenges, and opportunities. In " + }, + { + "bbox": [ + 304, + 338, + 524, + 360 + ], + "type": "inline_equation", + "content": "MLSys" + }, + { + "bbox": [ + 304, + 338, + 524, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 370, + 524, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 524, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 524, + 459 + ], + "type": "text", + "content": "Xiyou Zhou, Zhiyu Chen, Xiaoyong Jin, and William Yang Wang. 2021. HULK: An energy efficiency benchmark platform for responsible natural language processing. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 329-336, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 470, + 524, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 524, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 524, + 536 + ], + "type": "text", + "content": "Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In The IEEE International Conference on Computer Vision (ICCV)." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 550, + 376, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 376, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 376, + 564 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 303, + 573, + 502, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 573, + 502, + 587 + ], + "spans": [ + { + "bbox": [ + 303, + 573, + 502, + 587 + ], + "type": "text", + "content": "A.1 Pre-training and Fine-tuning Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 592, + 524, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 592, + 524, + 645 + ], + "spans": [ + { + "bbox": [ + 302, + 592, + 524, + 645 + ], + "type": "text", + "content": "For pre-training BERT with MLM, we mainly follow what is listed in Devlin et al. (2019). The specific hyperparameters used for pre-training are listed in Table 5 and were used on both machines." + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 314, + 654, + 487, + 749 + ], + "blocks": [ + { + "bbox": [ + 314, + 654, + 487, + 749 + ], + "lines": [ + { + "bbox": [ + 314, + 654, + 487, + 749 + ], + "spans": [ + { + "bbox": [ + 314, + 654, + 487, + 749 + ], + "type": "table", + "html": "
Hyperparameters
maximum steps1000000
training batch size per device64
evaluation batch size per device64
maximum sequence length128
learning rate1e-4
warmup steps10000
weight decay0.01
", + "image_path": "8fb56ce7f7d25c63b5549d17b6e36b2ee7ca994a7eb1c4daa16edb30c778efdd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 756, + 523, + 769 + ], + "lines": [ + { + "bbox": [ + 304, + 756, + 523, + 769 + ], + "spans": [ + { + "bbox": [ + 304, + 756, + 523, + 769 + ], + "type": "text", + "content": "Table 5: Hyperparameters used for pre-training BERT" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9064" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 111 + ], + "type": "text", + "content": "In Table 6, we list the set of hyperparameters used to fine-tune each task. All fine-tuning tasks were run on both machines." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 126, + 291, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 126, + 291, + 234 + ], + "spans": [ + { + "bbox": [ + 67, + 126, + 291, + 234 + ], + "type": "text", + "content": "Hardware details The A100 machine is located in Haverford, Pennsylvania, USA, and has 4x NVIDIA A100 GPUs with 40GB GDDR SDRAM, 376GB main memory and 32 Intel Xeon processors. The RTX 8000 machine is located in Pittsburgh, Pennsylvania, USA, and has 4x NVIDIA Quadro RTX 8000 GPUs with 48GB GDDR SDRAM, 36 Intel Xeon processors and 251GB RAM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 250, + 225, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 250, + 225, + 263 + ], + "spans": [ + { + "bbox": [ + 67, + 250, + 225, + 263 + ], + "type": "text", + "content": "A.2 Kill-A-Watt Measurements" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 272, + 291, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 272, + 291, + 528 + ], + "spans": [ + { + "bbox": [ + 67, + 272, + 291, + 528 + ], + "type": "text", + "content": "Energy is lost during the process of transferring energy from a power source to the machine. The coefficient for the loss is acquired through the readings from Kill-A-Watt devices. The device measures the instantaneous Watt reading extracted from the wall and displays on the monitor. The measurements for the A100 machine were recorded only on the single node of the cluster containing it (the RTX8000 machine is not part of a larger cluster). The GPU node is connected to two outlets, and we plug separate Kill-A-Watt devices into both. For each instantaneous reading, we read off of and sum up the readings on both Kill-A-Watt devices. To best compare between the package reading and the wall reading, we read off of the device at the same 15 second interval that CodeCarbon records the energy consumed. To convert each instantaneous Watts readings into Kilowatt-Hours, we follow the formula:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 541, + 268, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 541, + 268, + 569 + ], + "spans": [ + { + "bbox": [ + 89, + 541, + 268, + 569 + ], + "type": "interline_equation", + "content": "\\mathrm {k W h} = \\frac {1 5 \\text {s e c o n d s}}{3 6 0 0 \\text {s e c o n d s}} \\times \\frac {\\text {w a t t s} \\times \\text {h o u r s}}{1 0 0 0}", + "image_path": "1d11f3d244d5ee8d83e661088f0ba6e3b807b6e2c8191968ce6183c8deb5dbb3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 581, + 291, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 291, + 716 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 291, + 716 + ], + "type": "text", + "content": "We sum up all the calculations for the entire run of a fine-tuning experiment. Then, we can compare the sum of the wall readings with the Code Carbon energy consumption recording. We divide the wall reading over the package reading to get a coefficient, measuring the more realistic energy used. The setup of an instantaneous reading on the A100 machine is shown in Figure 2. The recordings of the wall readings on the A100 GPU machine are recorded in Table 7." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 719, + 290, + 772 + ], + "type": "text", + "content": "We ran these wall reading experiments for both machines and obtained a coefficient of 1.09 on the A100 machine, and a coefficient of 1.05 on the RTX machine." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 527, + 132 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 527, + 132 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 527, + 132 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 527, + 132 + ], + "type": "image", + "image_path": "23cc7a984d5ac1e8681ad75dba598a06618c597141aee5221814227c57bce57f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 138, + 525, + 152 + ], + "lines": [ + { + "bbox": [ + 302, + 138, + 525, + 152 + ], + "spans": [ + { + "bbox": [ + 302, + 138, + 525, + 152 + ], + "type": "text", + "content": "Figure 2: Kill-A-Watt wall reading measurement setup" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 173, + 420, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 173, + 420, + 185 + ], + "spans": [ + { + "bbox": [ + 302, + 173, + 420, + 185 + ], + "type": "text", + "content": "A.3 Additional Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 190, + 526, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 190, + 526, + 283 + ], + "spans": [ + { + "bbox": [ + 302, + 190, + 526, + 283 + ], + "type": "text", + "content": "BERT pre-training and fine-tuning results Table 8 records the energy consumption results from the A100 GPU machine, located in Haverford, PA. The conversion factor for this location is 672.8 lbs " + }, + { + "bbox": [ + 302, + 190, + 526, + 283 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2} / \\mathrm{MWh}" + }, + { + "bbox": [ + 302, + 190, + 526, + 283 + ], + "type": "text", + "content": " . We then calculate the emissions using raw output from CodeCarbon, which is listed in Table 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 293, + 527, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 293, + 527, + 401 + ], + "spans": [ + { + "bbox": [ + 302, + 293, + 527, + 401 + ], + "type": "text", + "content": "DistilBERT results Table 9 records the energy consumption and emission on the A100 GPU machine. Distillation is not done on this machine, and all the fine-tunings are done on the DistilBERT trained on the RTX8000 machine. The results on the A100 machine show a similar trend that fine-tunings on DistilBERT takes around " + }, + { + "bbox": [ + 302, + 293, + 527, + 401 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 302, + 293, + 527, + 401 + ], + "type": "text", + "content": " less energy than on BERT base models." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 413, + 530, + 568 + ], + "blocks": [ + { + "bbox": [ + 307, + 413, + 530, + 568 + ], + "lines": [ + { + "bbox": [ + 307, + 413, + 530, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 413, + 530, + 568 + ], + "type": "image", + "image_path": "8747192e9dcea7336391a18c5da5097b74439f1090a7ab3d2a05676346772234.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 578, + 525, + 602 + ], + "lines": [ + { + "bbox": [ + 302, + 578, + 525, + 602 + ], + "spans": [ + { + "bbox": [ + 302, + 578, + 525, + 602 + ], + "type": "text", + "content": "Figure 3: Energy consumed (kWh) against total tokens seen on A100 machine" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 624, + 527, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 527, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 527, + 773 + ], + "type": "text", + "content": "Training duration and energy consumption From Figure 5 and 6, we see that there is a strong correlation between training time and energy consumption, which holds across our models and hardware settings (Figure 7). However, similarly to tokens seen, we observe that pre-training exhibits a slightly different energy consumption profile, as do question answering tasks on the A100 machine (which have the longest sequences out of our finetuning tasks). When training time estimates are not feasible in advance (such as in certain hyperparam" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9065" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 84, + 68, + 510, + 191 + ], + "blocks": [ + { + "bbox": [ + 84, + 68, + 510, + 191 + ], + "lines": [ + { + "bbox": [ + 84, + 68, + 510, + 191 + ], + "spans": [ + { + "bbox": [ + 84, + 68, + 510, + 191 + ], + "type": "table", + "html": "
RTEMNLIIMDBSST2SQuAD v1SQuAD v2CoNLL 2003CoNLL 2012
epochs33532233
train bs3232163232323232
eval bs3232163232323232
max seq len128128128128384384128128
doc stride128128
lr2e-52e-52e-52e-53e-53e-55e-55e-5
", + "image_path": "1b804e445f59f54cd7022bbe897ffc35609000e2fe76a50125d3d5611f2462f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 119, + 222, + 476, + 327 + ], + "blocks": [ + { + "bbox": [ + 190, + 199, + 402, + 211 + ], + "lines": [ + { + "bbox": [ + 190, + 199, + 402, + 211 + ], + "spans": [ + { + "bbox": [ + 190, + 199, + 402, + 211 + ], + "type": "text", + "content": "Table 6: Hyperparameters used for fine-tuning tasks" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 119, + 222, + 476, + 327 + ], + "lines": [ + { + "bbox": [ + 119, + 222, + 476, + 327 + ], + "spans": [ + { + "bbox": [ + 119, + 222, + 476, + 327 + ], + "type": "table", + "html": "
TaskRandom seedWall reading converted (kWh)Code Carbon reading (kWh)Power Loss Co-efficient
SST2 (no AC)420.1030.0881.170
SST21230.0590.0551.057
IMDB420.0700.0611.134
IMDB420.0790.0711.122
IMDB1230.07930.0731.092
RTE420.006410.006021.065
", + "image_path": "e9a9d1959441b1d862867b3e6c53725a3f3789be8a1499bd881beee54194c2ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 81, + 357, + 512, + 497 + ], + "blocks": [ + { + "bbox": [ + 151, + 334, + 442, + 346 + ], + "lines": [ + { + "bbox": [ + 151, + 334, + 442, + 346 + ], + "spans": [ + { + "bbox": [ + 151, + 334, + 442, + 346 + ], + "type": "text", + "content": "Table 7: Kill-A-Watt experiment recordings for the A100 GPU machine" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 357, + 512, + 497 + ], + "lines": [ + { + "bbox": [ + 81, + 357, + 512, + 497 + ], + "spans": [ + { + "bbox": [ + 81, + 357, + 512, + 497 + ], + "type": "table", + "html": "
TaskDatasetEnergy (kWh)Emissions (kg CO2)Training time (s)Equiv. # PT runsEquivalent emissions of miles driven
MLMWiki + Books146.15044.6027265531144
fine-tuningRTE0.0020.00219584420.005
MNLI0.3560.23924004060.613
SQuAD v10.1710.11510618460.295
SQuAD v20.2630.17715805490.454
IMDB0.0620.019047821140.048
SST20.0510.03437728580.087
CoNLL20030.0110.00778134110.018
CoNLL20120.0910.06165415840.156
", + "image_path": "1594d6e89606593a7a75284f38d3a87508dafecfb78750a33285ba8925ddfaee.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 71, + 584, + 287, + 688 + ], + "blocks": [ + { + "bbox": [ + 67, + 505, + 525, + 567 + ], + "lines": [ + { + "bbox": [ + 67, + 505, + 525, + 567 + ], + "spans": [ + { + "bbox": [ + 67, + 505, + 525, + 567 + ], + "type": "text", + "content": "Table 8: Energy consumption of pre-training BERT and fine-tuning on the A100 GPU machine. Energy is computed as raw energy measured by CodeCarbon multiplied by a coefficient to correct for power loss (Eq. 3.4). Equiv. miles refers to the approximate number of vehicle miles driven resulting in equivalent " + }, + { + "bbox": [ + 67, + 505, + 525, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 67, + 505, + 525, + 567 + ], + "type": "text", + "content": " emissions. Note that here, unlike in Table 2, the \"baseline\" pre-training cost is fixed to the cost of 1 million steps of masked language modeling of sequences of length 128." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 584, + 287, + 688 + ], + "lines": [ + { + "bbox": [ + 71, + 584, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 71, + 584, + 287, + 688 + ], + "type": "table", + "html": "
TaskDatasetEnergyTimeEmiss.
EntailmentRTE0.001100.001
MNLI0.21015390.141
QASQuAD v10.0915660.061
SQuAD v20.1419150.095
SentimentIMDB0.0453410.012
SST0.0342620.023
NERCoNLL20030.007570.005
CoNLL20120.0594480.040
", + "image_path": "a97a18319bf5047ab66da40d676a4c1643f9ce02ccb37454545aca1b6ac32dd2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 696, + 291, + 733 + ], + "lines": [ + { + "bbox": [ + 67, + 696, + 291, + 733 + ], + "spans": [ + { + "bbox": [ + 67, + 696, + 291, + 733 + ], + "type": "text", + "content": "Table 9: Energy consumption of training (distillation) and fine-tuning DistilBERT on the A100 machine. Units and energy calculations are the same as given in Table 8." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 302, + 587, + 525, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 587, + 525, + 641 + ], + "spans": [ + { + "bbox": [ + 302, + 587, + 525, + 641 + ], + "type": "text", + "content": "eter sweeps), we recommend that researchers and practitioners use token counts estimates (including dynamic padding tokens) if they have reasonable knowledge of their data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "spans": [ + { + "bbox": [ + 302, + 651, + 526, + 773 + ], + "type": "text", + "content": "As indicated in Figure 8 and Figure 9, energy increases as the optimization steps increase. This is not surprising given the correlation between training time and energy consumption. However, we see that for different tasks, the energy required for each step is very different. Each step of pre-training takes a longer time, likely due to the higher batch size than all fine-tuning tasks. For QA tasks, the per-step energy consumption is higher than other" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 310, + 791 + ], + "type": "text", + "content": "9066" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 295, + 224 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 295, + 224 + ], + "type": "image", + "image_path": "e32afc811c7479f199676d9b1914dd4389d41d2b25501b5d722d5edd3cebb7d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 236, + 290, + 260 + ], + "lines": [ + { + "bbox": [ + 67, + 236, + 290, + 260 + ], + "spans": [ + { + "bbox": [ + 67, + 236, + 290, + 260 + ], + "type": "text", + "content": "Figure 4: Energy consumed (kWh) against total tokens seen on RTX8000 machine" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 277, + 295, + 431 + ], + "blocks": [ + { + "bbox": [ + 71, + 277, + 295, + 431 + ], + "lines": [ + { + "bbox": [ + 71, + 277, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 71, + 277, + 295, + 431 + ], + "type": "image", + "image_path": "3c98d08b317fc611f443cdc3e699e81fb24739f221ce843e4c60e38586d646e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 237, + 527, + 310 + ], + "lines": [ + { + "bbox": [ + 302, + 237, + 527, + 310 + ], + "spans": [ + { + "bbox": [ + 302, + 237, + 527, + 310 + ], + "type": "text", + "content": "Figure 7: Total training time is strongly predictive of total energy consumed (kWh) for both BERT and DistilBERT models, on both A100 and RTX 8000 GPU machines. Note that both axes are in log scale. Different from token counts, the trendlines themselves are similar as well across models and hardware settings." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 482, + 289, + 645 + ], + "blocks": [ + { + "bbox": [ + 67, + 442, + 290, + 467 + ], + "lines": [ + { + "bbox": [ + 67, + 442, + 290, + 467 + ], + "spans": [ + { + "bbox": [ + 67, + 442, + 290, + 467 + ], + "type": "text", + "content": "Figure 5: Energy consumed (kWh) against training time (s) on A100 machine, showing the first 30 minutes" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 482, + 289, + 645 + ], + "lines": [ + { + "bbox": [ + 71, + 482, + 289, + 645 + ], + "spans": [ + { + "bbox": [ + 71, + 482, + 289, + 645 + ], + "type": "image", + "image_path": "c9aea71968f5a63e944731649adfcb202f8fa9da9fc67b3758e83eca27f240c1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 704, + 290, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 290, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 290, + 731 + ], + "type": "text", + "content": "tasks. This is likely due to the maximum sequence length of 384 being higher than for the other tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 733, + 290, + 772 + ], + "type": "text", + "content": "Figure 10 shows the number of examples of the task and the energy consumed in log scale. Similar to Figure 1, we see a direct correlation between" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 530, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 530, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 530, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 530, + 225 + ], + "type": "image", + "image_path": "f002ed9f7b98d360afc2f3150ee0f3e9031c9fff200615ceea9a4b79653b4005.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 324, + 518, + 487 + ], + "blocks": [ + { + "bbox": [ + 310, + 324, + 518, + 487 + ], + "lines": [ + { + "bbox": [ + 310, + 324, + 518, + 487 + ], + "spans": [ + { + "bbox": [ + 310, + 324, + 518, + 487 + ], + "type": "image", + "image_path": "f42988e7016a356666720e5b891c9f3f1ab238a3f11b4c21ad8dee8e78e6c64a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 498, + 525, + 523 + ], + "lines": [ + { + "bbox": [ + 302, + 498, + 525, + 523 + ], + "spans": [ + { + "bbox": [ + 302, + 498, + 525, + 523 + ], + "type": "text", + "content": "Figure 8: Energy consumed (kWh) against optimization steps on A100 machine machine" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 538, + 519, + 700 + ], + "blocks": [ + { + "bbox": [ + 67, + 656, + 290, + 681 + ], + "lines": [ + { + "bbox": [ + 67, + 656, + 290, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 656, + 290, + 681 + ], + "type": "text", + "content": "Figure 6: Energy consumed (kWh) against training time (s) on RTX8000 machine, showing the first 30 minutes" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 538, + 519, + 700 + ], + "lines": [ + { + "bbox": [ + 309, + 538, + 519, + 700 + ], + "spans": [ + { + "bbox": [ + 309, + 538, + 519, + 700 + ], + "type": "image", + "image_path": "fe29ab73b8c948511d25e7107c86a087e409a5522fed2563a589be714454e0d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 712, + 524, + 736 + ], + "lines": [ + { + "bbox": [ + 302, + 712, + 524, + 736 + ], + "spans": [ + { + "bbox": [ + 302, + 712, + 524, + 736 + ], + "type": "text", + "content": "Figure 9: Energy consumed (kWh) against optimization steps on RTX8000 machine" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9067" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 295, + 225 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 295, + 225 + ], + "type": "image", + "image_path": "2f26e3758d0114bec7d6f2f0e0062c9ba1dce8585899e7e842bd698bf6648e64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 237, + 291, + 321 + ], + "lines": [ + { + "bbox": [ + 67, + 237, + 291, + 321 + ], + "spans": [ + { + "bbox": [ + 67, + 237, + 291, + 321 + ], + "type": "text", + "content": "Figure 10: Energy consumed (kWh) against total number of examples seen on both A100 and RTX8000 machines. Comparing with Figure 1, we see that merely counting number of training examples is much less predictive of energy consumption than accounting for example sequence lengths along with batch size (which affects the maximum sequence length in each batch)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 342, + 290, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 342, + 290, + 368 + ], + "spans": [ + { + "bbox": [ + 67, + 342, + 290, + 368 + ], + "type": "text", + "content": "higher number of training examples and higher energy usage on both machines." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 74, + 380, + 282, + 540 + ], + "blocks": [ + { + "bbox": [ + 74, + 380, + 282, + 540 + ], + "lines": [ + { + "bbox": [ + 74, + 380, + 282, + 540 + ], + "spans": [ + { + "bbox": [ + 74, + 380, + 282, + 540 + ], + "type": "image", + "image_path": "12fafe6f47ae862e264f38b522c97bacfcf9ce2919844c5c644a9abf412dbe7c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 552, + 290, + 576 + ], + "lines": [ + { + "bbox": [ + 67, + 552, + 290, + 576 + ], + "spans": [ + { + "bbox": [ + 67, + 552, + 290, + 576 + ], + "type": "text", + "content": "Figure 11: " + }, + { + "bbox": [ + 67, + 552, + 290, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 67, + 552, + 290, + 576 + ], + "type": "text", + "content": " emissions (kg) against task accuracy for SST-2 fine-tuning on RTX 8000 GPU machine" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 590, + 290, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 290, + 671 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 290, + 671 + ], + "type": "text", + "content": "Figure 11 shows the accuracy of SST2 task as " + }, + { + "bbox": [ + 67, + 590, + 290, + 671 + ], + "type": "inline_equation", + "content": "\\mathrm{CO}_{2}" + }, + { + "bbox": [ + 67, + 590, + 290, + 671 + ], + "type": "text", + "content": " emissions (kg) increases. As shown previously in Figure 6, energy consumption increases as time increases. Generally, as emission increases, accuracy increases. We can see that as emissions goes up, accuracy is trending towards converging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "content": "Single- vs. multi-GPU fine-tuning We fine-tune BERT-base and DistilBERT-base models on the RTX8000 machine as well. If the same hyperparameters as the single-GPU setting are used naively (i.e. the same batch size is split over 4 GPUs), fine-tuning using 4 GPUs can (but does not always) take even longer than using just 1 GPU, and can (but" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 304, + 68, + 530, + 293 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 530, + 293 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 530, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 530, + 293 + ], + "type": "table", + "html": "
DatasetDuration (s)Energy (kWh)
1GPU4GPU1GPU4GPU
BERT-base
RTE (NLI)59278.16e-38.52e-3
MNLI (NLI)670030149.38e-18.64e-1
SQuADv1 (QA)378013565.37e-14.28e-1
SQuADv2 (QA)560420117.95e-16.16e-1
IMDB (Sent.)10744721.51e-11.32e-1
SST2 (Sent.)5873238.11e-28.39e-2
CoNLL2003 (NER)1497372.06e-22.05e-1
CoNLL2012 (NER)14877372.07e-12.14e-1
DistilBERT
RTE (NLI)30184.34e-34.72e-3
MNLI (NLI)344717604.81e-14.98e-1
SQuADv1 (QA)19547922.76e-12.40e-1
SQuADv2 (QA)291611704.12e-13.46e-1
IMDB (Sent.)5492727.75e-27.46e-2
SST2 (Sent.)3061954.21e-25.49e-2
CoNLL2003 (NER)784361.10e-21.18e-1
CoNLL2012 (NER)7624341.07e-11.19e-1
", + "image_path": "8262f4339c02c4b224cb879747f1fac398eaf5ca7a2fb0c1cdaa8d9c56baf70f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 301, + 525, + 337 + ], + "lines": [ + { + "bbox": [ + 302, + 301, + 525, + 337 + ], + "spans": [ + { + "bbox": [ + 302, + 301, + 525, + 337 + ], + "type": "text", + "content": "Table 10: Single-GPU vs. BS and LR-optimized 4-GPU fine-tuning energy requirements across end tasks on the RTX8000 machine, for BERT-base and DistilBERT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 301, + 364, + 526, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 364, + 526, + 675 + ], + "spans": [ + { + "bbox": [ + 301, + 364, + 526, + 675 + ], + "type": "text", + "content": "does not always) use about twice as much energy in both BERT-base and DistilBERT. If we increase batch size x 4 with 4 GPUs (and adjust learning rate accordingly), however, and compare single-GPU fine-tuning with multi-GPU fine-tuning (see Table 10), we observe that energy cost is typically similar or even less than when using 1 GPU, while taking around half as much time or less. In both the \"naive\" and \"optimized\" multi-GPU settings, the single-vs-multi-GPU difference in energy cost and job duration seems to be related to dataset sequence lengths. Tasks with longer sequences (such as QA tasks, and, to a lesser extent, IMDB and RTE) tend to exhibit more consistent and dramatic energy and time efficiency gains than the other tasks when using 4 GPUs. On the other hand, tasks with shorter sequences (such as NER) tend to require more energy with 4 GPUs, even if the wall-clock efficiency may be improved. One way one might interpret this is a large-enough per-device batch size and typical sequence length is necessary for multi-GPU training to be \"worth\" the overhead of communication between GPUs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 678, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 527, + 772 + ], + "type": "text", + "content": "In light of these observations, our general recommendation is that, if one owns a machine with multiple GPUs, one should consider using all available (idle) GPUs for energy- and time-efficient finetuning. Although it is often sufficient to use a single GPU when fine-tuning models of scale similar to ours, and instantaneous energy usage may be" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9068" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 71, + 291, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 71, + 291, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 291, + 137 + ], + "type": "text", + "content": "higher using more GPUs, total energy used may end up being less while also requiring less time, especially if the training sequences tend to be longer. On the other hand, tasks with short sequences are likely best kept to a single GPU." + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "spans": [ + { + "bbox": [ + 286, + 781, + 309, + 791 + ], + "type": "text", + "content": "9069" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_content_list.json b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4fd1069175f492441300c0799ac2e486f0b4b24 --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_content_list.json @@ -0,0 +1,2671 @@ +[ + { + "type": "text", + "text": "Enhanced Simultaneous Machine Translation with Word-level Policies", + "text_level": 1, + "bbox": [ + 134, + 89, + 862, + 109 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kang Kim*, Hankyu Cho*", + "bbox": [ + 384, + 142, + 615, + 158 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "XL8 Inc.", + "bbox": [ + 462, + 160, + 537, + 173 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kai, matthew}@x18.ai", + "bbox": [ + 394, + 177, + 606, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have seen remarkable advances in the field of Simultaneous Machine Translation (SiMT) due to the introduction of innovative policies that dictate whether to READ or WRITE at each step of the translation process. However, a common assumption in many existing studies is that operations are carried out at the subword level, even though the standard unit for input and output in most practical scenarios is typically at the word level. This paper demonstrates that policies devised and validated at the subword level are surpassed by those operating at the word level, which process multiple subwords to form a complete word in a single step. Additionally, we suggest a method to boost SiMT models using language models (LMs), wherein the proposed word-level policy plays a vital role in addressing the subword disparity between LMs and SiMT models. Code is available at https://github.com/x18-ai/WordSiMT.", + "bbox": [ + 141, + 275, + 460, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 583, + 258, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous Machine Translation (SiMT) commences the translation process while simultaneously receiving the input, making it an effective approach for applications that require minimal latency such as simultaneous interpretation or live broadcast. The development of a novel policy is central to research efforts in SiMT. This policy dictates the translation process by determining whether to execute a READ or WRITE action at each step of the process.", + "bbox": [ + 112, + 607, + 489, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural SiMT models, like offline Neural Machine Translation (NMT) models, commonly employ Byte Pair Encoding (BPE) (Sennrich et al., 2016) or similar techniques to encode an input sentence into a sequence of tokens. Typically, a single READ or WRITE action of a SiMT policy is responsible for handling an encoded token, which may sometimes be a word but often a subword.", + "bbox": [ + 112, + 769, + 489, + 896 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/698d58ed3af2627eea3939c9cb293ed07e64a927e1c670a7484c4fa413657720.jpg", + "image_caption": [ + "Figure 1: Illustration of both token-level Wait-1 (red arrow lines) and word-level Wait-1 (green arrow lines) policies. The symbol \" $\\_$ \" at the end of a token indicates a word boundary." + ], + "image_footnote": [], + "bbox": [ + 510, + 249, + 884, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The development of BPE-based SiMT models and their policies has resulted in researchers focusing on working at the subword level. The performance analysis and implementation of many SiMT systems, to the best of our knowledge, have been carried out on encoded sequences of source and target subwords, rather than on the original source and target sentences1. This has led to two critical issues that need to be addressed.", + "bbox": [ + 507, + 456, + 882, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The first issue is the lack of a standardized tokenization and encoding scheme, meaning that different implementations may employ varying token sequences to encode identical text. This variability can impact latency evaluation results and complicate score comparisons across different systems.", + "bbox": [ + 507, + 601, + 882, + 697 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The second issue is the missed opportunity to process more source tokens before writing each target token without added latency. For a BPE-based SiMT model, the input must be received on a word-by-word basis to ensure proper encoding of each word into a sequence of subwords. Consequently, when the model encodes a word and performs a READ to process only a subword, it delays the reading of the remaining subwords without any benefit in actual latency², and may adversely impact", + "bbox": [ + 507, + 699, + 884, + 859 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We provide a list of reference works pertaining to this case in A.1", + "bbox": [ + 507, + 868, + 882, + 891 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Here we assume that the processing time of a READ is negligible, compared to the interval between receiving words.", + "bbox": [ + 509, + 892, + 882, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 136, + 904, + 262, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15616", + "bbox": [ + 475, + 927, + 524, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15616-15634", + "bbox": [ + 208, + 945, + 786, + 958 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 277, + 959, + 719, + 971 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/231157632db099fa480ad5e3f153a799abe77907d7c96e89f011863bd87a35c4.jpg", + "image_caption": [ + "Figure 2: An exemplary case depicting the difference between token-level and word-level Wait-k policies and their Average Lagging (AL) scores. The token-level model begins translating in the middle of reading \"Tremendously\" and fails to recover from the incorrectly translated target prefix. On the other hand, the word-level model processes \"Tremendously\" in a single READ action and produces a correct translation." + ], + "image_footnote": [], + "bbox": [ + 117, + 82, + 884, + 155 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "translation quality by causing the model to rely on incomplete representations extracted from partially-read words. Similarly, performing a WRITE to generate a subword earlier than the remaining subwords does not necessarily reduce latency, as the subword must wait for a complete word to be generated before it can be displayed.", + "bbox": [ + 110, + 249, + 487, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we show that establishing the unit of measuring and operating SiMT systems at the word level, rather than the subword level, offers a viable solution to these issues. Specifically, to tackle the first issue, we propose word-level latency metric calculation for measuring the latency of SiMT systems. This not only enables consistent comparisons between different systems but also provides a more accurate reflection of the actual latency experienced in SiMT applications that display translation results word by word.", + "bbox": [ + 110, + 365, + 487, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the second issue, we illustrate that an existing token-level policy can be transformed into a word-level policy that inherently overcomes the second issue, resulting in improved performance. Word-level policies take into consideration word boundaries and perform a READ or WRITE action to sequentially process a sequence of tokens that form a complete word. This conversion process can be applied to any token-level policies, and our experiments reveal that state-of-the-art fixed and adaptive policies exhibit significantly better performance when transformed into their word-level counterparts. Notably, these word-level policies often outperform token-level policies, even when evaluated using a token-level latency metric, due to their enhanced utilization of input and output tokens.", + "bbox": [ + 110, + 545, + 487, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, to boost translation accuracy, we suggest incorporating a pre-trained language model (LM) into a SiMT model, where the word-level policy plays a crucial role as a pivotal component. One of the major hurdles in utilizing an LM for SiMT is the vocabulary mismatch between the SiMT model", + "bbox": [ + 112, + 822, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and the LM. The difficulty of handling subword disparities when utilizing an LM for a downstream task is widely acknowledged (Liu et al., 2021a; Wang et al., 2022), and it becomes particularly problematic in SiMT, as inconsistent subwords between the LM and SiMT model make processing the same source or target prefix challenging. Our study demonstrates that our proposed word-level policy effectively tackles this challenge, enabling a successful integration of LMs into SiMT systems.", + "bbox": [ + 507, + 249, + 884, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 507, + 422, + 663, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Simultaneous Machine Translation", + "text_level": 1, + "bbox": [ + 507, + 447, + 830, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SiMT systems that employ a fixed policy utilize a pre-defined sequence of READ and WRITE operations for each source sentence. STATIC-RW (Dalvi et al., 2018) and Wait-k (Ma et al., 2019a) policies first read k source tokens, then alternate between reading and writing a single token. Elbayad et al. (2020) propose the multi-path training of a Wait-k model to train a single model that supports different k values at test time. Zhang et al. (2021) improve Wait-k policy using knowledge distillation from an offline MT model, while Zhang and Feng (2021) suggest a Mixture-of-Experts Wait-k Policy where predictions from multiple k values are combined inside a single model.", + "bbox": [ + 507, + 468, + 882, + 692 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, research efforts on adaptive policies focus on the development of dynamic decision-making processes for READ/WRITE actions. Cho and Esipova (2016) firstly introduce model-based adaptive criteria for Neural SiMT. Gu et al. (2017) propose to learn a policy by using reinforcement learning. Raffel et al. (2017) introduce Monotonic Attention that ensures monotonic alignment in the attention mechanism. Succeeding works improve it by extending the alignment window (Chiu and Raffel, 2017; Arivazhagan et al., 2019), extending it as monotonic multi-head attention (MMA) (Ma et al., 2019b) or learning transposed policies between the forward and backward models. (Zhang and Feng,", + "bbox": [ + 507, + 694, + 882, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15617", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2022c). Zheng et al. (2020) derive a policy by composing Wait-k models trained with different values of k. Zhang and Feng (2022a) model to predict the alignment between each target token and the source token. Zhang and Feng (2022b) measure accumulated information from source tokens to decide whether to write the current target token.", + "bbox": [ + 112, + 84, + 487, + 197 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite significant advancements, the impact of operating policies at the word level has not been thoroughly explored in existing works, which have mainly focused on developing and evaluating systems at the token level. In this paper, we address this gap by demonstrating that implementing various types of policies at the word level consistently outperforms their token-level counterparts.", + "bbox": [ + 112, + 199, + 489, + 326 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Utilizing pre-trained LM for MT", + "text_level": 1, + "bbox": [ + 112, + 344, + 416, + 360 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the successful utilization of Transformer-based LMs pre-trained on large text corpora for downstream NLP tasks (Devlin et al., 2019; Liu et al., 2019; Lample and Conneau, 2019), the utilization of these models for MT has become a significant research area. Several studies have demonstrated the effectiveness of incorporating encoder-only LMs into NMT models. Weng et al. (2020); Yang et al. (2020) combine the LM representations with the encoder's representation using a gating mechanism. Zhu et al. (2020) propose attention between BERT and both the encoder and decoder. Weng et al. (2022) leverage mBERT as an encoder and introduce a decoder that attends to grouped representations of the encoder output.", + "bbox": [ + 112, + 368, + 489, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another research direction focuses on developing LMs with the encoder-decoder architecture designed for NMT as the target downstream task (Lewis et al., 2020; Liu et al., 2020). These models show improvements particularly for low-resource language pairs. To enhance their adaptation for MT, various methods have been proposed, including fine-tuning specific parts of the LMs (Cooper Stickland et al., 2021), reducing domain mismatch and overestimation (Wang et al., 2022) and mitigating the copying behavior (Liu et al., 2021b).", + "bbox": [ + 112, + 611, + 489, + 788 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The integration of pre-trained LMs into SiMT remains an underexplored area of research. To date, Indurthi et al. (2022) is the only related study we are aware of. It improves MMA by integrating the LM's prediction of the next target token, which is encoded using their model's vocabulary before being inputted into the model. However, this approach sacrifices the semantic coherence of the", + "bbox": [ + 112, + 790, + 489, + 917 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "original tokens due to token fragmentation. Additionally, their approach falls under target-side LM integration, overlooking the potential advantages of source-side LM integration.", + "bbox": [ + 507, + 84, + 884, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we demonstrate an effective way of integrating a source-side LM into SiMT systems, offering a more versatile solution that can be integrated into most existing neural SiMT models. Building upon previous research conducted in offline MT (Zhu et al., 2020), we introduce essential modifications, with a particular focus on word-level policies as a pivotal component. The effective management of vocabulary mismatches between the LM and the SiMT model is contingent upon the successful implementation of a word-level SiMT policy, a key aspect that we address in our study.", + "bbox": [ + 507, + 149, + 885, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Proposed Methods", + "text_level": 1, + "bbox": [ + 507, + 356, + 705, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we propose the concept of employing a word-level latency metric and outline our conversion process for translating token-level policies into their word-level equivalents. Additionally, we present an integration of LM into SiMT, highlighting the advantages of utilizing word-level policies.", + "bbox": [ + 507, + 382, + 885, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Preliminaries", + "text_level": 1, + "bbox": [ + 507, + 492, + 662, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a source sentence $\\mathbf{x} = (x_{1}, x_{2}, \\ldots x_{n})$ , the goal of a SiMT model is to generate a target sentence of $\\mathbf{y} = (y_{1}, y_{2}, \\ldots y_{m})$ while minimizing latency metrics. A SiMT model's policy, represented by the variable $g_{i}$ , determines the number of source tokens to process before predicting target token $y_{i}$ . Then the probability of generating $\\mathbf{y}$ given $\\mathbf{x}$ is formulated as follows:", + "bbox": [ + 507, + 514, + 882, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {y} | \\mathbf {x}) = \\prod_ {i} ^ {| \\mathbf {y} |} p \\left(y _ {i} \\mid \\mathbf {x} _ {\\leq g _ {i}}, \\mathbf {y} _ {< i}; \\theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 657, + 882, + 701 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ is the model's parameters which are commonly optimized with a cross-entropy loss.", + "bbox": [ + 507, + 709, + 884, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer encoder-decoder model (Vaswani et al., 2017) is currently the most widely used architecture for SiMT. To avoid redundant encoding of the input sequence after each READ operation, the encoder is typically modified to encode the source tokens unidirectionally (Elbayad et al., 2020). Alternatively, more advanced techniques like the recurrent Linear Transformer (Kahardipraja et al., 2021) or Partial Bidirectional Encoding (Iranzo Sanchez et al., 2022) can be adopted to enhance the encoding capabilities further.", + "bbox": [ + 507, + 741, + 885, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15618", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the evaluation of SiMT systems, translation quality is commonly assessed in conjunction with the latency required for generating translations. Various metrics have been proposed to calculate latency scores, with Average Lagging (AL) (Ma et al., 2019a) being the most commonly used metric.", + "bbox": [ + 112, + 84, + 489, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 Measuring latency based on the word level", + "text_level": 1, + "bbox": [ + 112, + 195, + 485, + 211 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As detailed in A.1, a substantial body of prior research work assesses the performance of SiMT systems by utilizing a latency metric on encoded source tokens under different tokenization and encoding schemes. This practice results in each system being evaluated on non-identical token sequences for the same dataset, thereby making it challenging to accurately compare scores across different systems.", + "bbox": [ + 112, + 218, + 487, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose word-level latency score calculation by considering the word boundaries in token-level sequences. Specifically, when the first token of a source word is processed through a READ operation, we consider it as reading the corresponding word. Similarly, when the last token of a target word is written via a WRITE operation, we consider it as writing that word. By doing so, the latency scores are calculated consistently, regardless of the tokenization and encoding of the input. This ensures that results from different systems can be compared fairly.", + "bbox": [ + 112, + 363, + 489, + 557 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 Word-level SiMT policies", + "text_level": 1, + "bbox": [ + 112, + 571, + 359, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed word-level policy restricts a SiMT policy's transition from READ to WRITE or vice versa to occur exclusively at the boundaries of words. Any token-level policy can be transformed to operate at the word-level by following the conversion process we outline below.", + "bbox": [ + 112, + 595, + 487, + 690 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, we ensure that a word-level policy does not write a target token in the middle of reading a sequence of source tokens that make up a word. To accomplish this word-level READ, we delay $g_{i}$ until it reaches the nearest source word boundary. We formally define $r_{i}$ that has a refined value of $g_{i}$ based on the word boundaries in $\\mathbf{x}$ as follows:", + "bbox": [ + 112, + 693, + 487, + 820 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i} := \\min \\{j | j \\geq g _ {i} \\wedge j \\in B _ {S} \\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 838, + 487, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $B_{S}$ denotes the indices of the source words' last tokens. Substituting $r_i$ for $g_i$ as a policy transforms it into another policy that upholds the same", + "bbox": [ + 112, + 870, + 489, + 917 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9c360f553d97fd7c1279b3790fec24e9433423a6cc35154833a43eb39f6596e.jpg", + "image_caption": [ + "Token-level unidirectional encoding", + "Figure 3: Comparison of masking in the token-level unidirectional attention (left) and the intra-word bidirectional encoding (right). Word boundaries are represented by vertical/horizontal bars on each axis." + ], + "image_footnote": [], + "bbox": [ + 510, + 82, + 695, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eeb2ecb69355b1b06599dbab51b7beb5aab89743c0897a11dbf4adddc39a2098.jpg", + "image_caption": [ + "Intra-word bidirectional encoding" + ], + "image_footnote": [], + "bbox": [ + 702, + 80, + 880, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "decision-making criterion while ensuring uninterrupted reading of an entire word when initiating the reading of a token.", + "bbox": [ + 507, + 310, + 882, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly to the word-level READ, we design a word-level WRITE to balance READ and WRITE actions throughout the translation. To achieve this, we can modify $r_i$ such that it writes until it produces a token that ends with an end-of-word symbol. We define $w_i$ that satisfies this as follows:", + "bbox": [ + 507, + 359, + 882, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb _ {i} := \\min \\{j | j \\geq i \\wedge j \\in B _ {T} \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 579, + 473, + 882, + 489 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} := \\left\\{ \\begin{array}{l l} r _ {i}, & \\text {i f} i = 1 \\vee b _ {i - 1} \\neq b _ {i} \\\\ w _ {i - 1}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 498, + 882, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B_{T}$ denotes the indices of the target words' last tokens and $b_{i}$ represents the index of the final token in the word that includes $y_{i}$ . By employing $w_{i}$ in place of $r_{i}$ (or $g_{i}$ ), we ensure that the policy consistently composes entire words without interruptions from any READ actions. This approach effectively reduces latency by facilitating faster writing of certain tokens compared to the original policy, thereby compensating for the increased latency resulting from word-level READ operations. Figure 1 provides a visual comparison between word-level and token-level policies in the context of Wait-1, with the word-level policy encompassing both word-level READ and WRITE operations.", + "bbox": [ + 507, + 546, + 882, + 788 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4 Intra-word bidirectional encoding", + "text_level": 1, + "bbox": [ + 507, + 801, + 823, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unidirectional encoding in SiMT is vital for managing computational complexity and training efficiency. However, it has an inevitable consequence of weakening the source sequence representations compared to bidirectional encoding. This is an additional factor contributing to the lower translation", + "bbox": [ + 507, + 822, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15619", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a727ea4b5c0a6c130677d30ed24fcb7c203ad12fa20e003632cbfa99f9e280a6.jpg", + "image_caption": [ + "Figure 4: Illustration of LM-fused attention with the word-level Wait-1. Word and model activations processed at a specific stage are highlighted in red. When a source word is received, it is independently encoded by the LM's vocabulary and the in-domain vocabulary. The hidden activations of the word from the LM is then utilized in the encoder. The decoder generates a sequence of tokens for a word by using both the LM and encoder activations." + ], + "image_footnote": [], + "bbox": [ + 114, + 80, + 438, + 237 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff2cb4a8e818630be28676722bcc1420696bbff0092e9c83b3b5ecdea1197ad2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 80, + 884, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "quality of SiMT compared to offline models, along with the early translation from partial inputs.", + "bbox": [ + 112, + 328, + 485, + 360 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate this issue, we utilize a technique called intra-word bidirectional encoding. At the word level, this approach involves unidirectional encoding for each word in the input sentence, meaning past words cannot attend to future words. However, at the subword level, subwords within the same word can attend to each other, allowing past subwords to attend to future subwords within the same word. Since READ operates at the word level in word-level policies, this encoding does not require recomputation during each WRITE operation. It only necessitates a single forward pass, similar to token-level unidirectional encoding. However, it can produce a better encoded representation by enabling attention to more tokens. An example masking to enable intra-word bidirectional encoding is depicted in Figure 3.", + "bbox": [ + 112, + 362, + 489, + 634 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5 Integration of LM into SiMT through word-level policies", + "text_level": 1, + "bbox": [ + 112, + 652, + 455, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we showcase an additional benefit of word-level policies when integrating an LM into a SiMT system. One of the key challenges in this integration is the vocabulary mismatch between the LM and the SiMT model, which hinders ensuring that both models process an equal amount of input prefix at each translation step.", + "bbox": [ + 112, + 692, + 487, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One possible solution is to use the LM's vocabulary for the SiMT model. However, the LM's training data may not align well with the specific domain targeted by the SiMT system (Wang et al., 2022). This can result in suboptimal vocabulary for the SiMT model compared to a vocabulary obtained from in-domain data (Liu et al., 2021a). An", + "bbox": [ + 112, + 806, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "other option is to explore methods to bridge vocabulary gaps (Kim et al., 2019; Sato et al., 2020; Liu et al., 2021a), but they are either validated only in certain transfer learning scenarios or require an additional training phase to train adapters or finetuning the entire LM using pre-training data.", + "bbox": [ + 507, + 328, + 884, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we introduce a method for leveraging an LM in a manner that facilitates the integration of an off-the-shelf LM into a SiMT model, utilizing a word-level policy, regardless of vocabulary mismatches and the internal structure of the SiMT model. Specifically, we employ an LM fused attention for both the encoder and decoder, following the approach outlined in (Zhu et al., 2020), but with two notable modifications.", + "bbox": [ + 507, + 447, + 884, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, we replace BERT with a decoder-only auto-regressive LM (Radford et al., 2019; Lin et al., 2022) for unidirectional encoding of the input, aligning with SiMT models for efficient training and inference. Secondly, the attention between the SiMT model and the LM occurs when both models execute a word-level READ for an input word. This ensures they interact only when they process an equal amount of input prefix, naturally resolving the synchronization issue. Additionally, as they align at every word boundary, the SiMT model can operate independently with a vocabulary derived from in-domain data, while the LM continues to use its original vocabulary. Unlike methods targeting specific SiMT models (Indurthi et al., 2022), our approach can benefit any Neural SiMT model with any decoder-only LM. Figure 4 illustrates the proposed integration of the LM with word-level Wait-1.", + "bbox": [ + 507, + 613, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15620", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d5bcec75ccb8102db8d55b6645ec1b3a09e7c3c9ad3e216c6f2cdca793b7754b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 115, + 80, + 364, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/24e7a2f13f43e53f3d020453632e5e0a7c5f48531eac80496f7b5711a1de1be3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 376, + 80, + 620, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0743529918e4cac20bc0bb494314a3051c8894e58014c09ac047a402ac000152.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 80, + 880, + 214 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bcf1943b0977721a88373dcdf9a230257aa37eb2d928590877358cfcfef7a357.jpg", + "image_caption": [ + "(a) En-Fr Wait-k Transformer small", + "(d) De-En Wait-k Transformer big", + "Figure 5: Results of translation quality v.s. word-level AL." + ], + "image_footnote": [], + "bbox": [ + 115, + 231, + 361, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97ec0498c4715076be48915533852f0c18de28f3c468d5e5f253a81e911273a5.jpg", + "image_caption": [ + "(b) En-Fr MoE Transformer small", + "(e) De-En MoE Transformer big" + ], + "image_footnote": [], + "bbox": [ + 374, + 231, + 621, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33c00968ad7a1b50905ae7a5d41482ddc496a57786dd6a9eba79a39c63e3dd2c.jpg", + "image_caption": [ + "(c) En-Fr ITST Transformer small", + "(f) De-En ITST Transformer big" + ], + "image_footnote": [], + "bbox": [ + 633, + 231, + 880, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Experiments", + "text_level": 1, + "bbox": [ + 112, + 431, + 258, + 447 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Datasets", + "text_level": 1, + "bbox": [ + 112, + 458, + 226, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following two datasets for experiments: IWSLT 17 English to French (En-Fr) (Cettolo et al., 2017) consists of $242\\mathrm{k}$ pairs split into $233\\mathrm{k}$ training, 890 validation, and 8,597 test pairs.", + "bbox": [ + 112, + 481, + 487, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WMT 15 German to English (De-En) comprises 4.5 million pairs. We use newstest2013 (3,000 pairs) for validation and newstest2015 (2,169 pairs) for testing.", + "bbox": [ + 112, + 548, + 487, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Experimental settings", + "text_level": 1, + "bbox": [ + 112, + 627, + 331, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our methods are evaluated on the following systems, all of which are based on Transformer (Vaswani et al., 2017) with unidirectional encoding.", + "bbox": [ + 112, + 650, + 487, + 697 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Wait-k (Ma et al., 2019a): A model operating under Wait-k policy. A single model is trained for all k values by randomly sampling k during training (Elbayad et al., 2020). For the word-level Wait-k policy, we define it as reading the first k words and then alternating between reading one source word and writing one target word.", + "bbox": [ + 112, + 699, + 487, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MoE Wait-k (Zhang and Feng, 2021): A Mixture-of-Experts model initially trained with", + "bbox": [ + 112, + 812, + 487, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fixed Experts weights and then fine-tuned with dynamic Experts weights. The word-level Wait-k policy of different k is applied to each expert.", + "bbox": [ + 507, + 431, + 882, + 480 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ITST (Zhang and Feng, 2022b): A SoTA SiMT model equipped with Information-Transport-based policy that quantifies information weights from each source to the current target token. To implement word-level ITST, we convert the number of source tokens required for the first target token of each target word into the corresponding number of source words using Equation 2. Once the required number of source words is read, we complete the translation of the word. Additionally, we calculate the latency cost at the word level.", + "bbox": [ + 507, + 482, + 882, + 658 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare each system by training both token-level and word-level models, with and without an LM. For models with an LM, we use XGLM-564M (Lin et al., 2022) and employ the two-stage training in which we first train the SiMT model without the LM and initialize the encoder and decoder from the LM-fused model with the pre-trained weights (Zhu et al., 2020). We also tested the single stage training where all trainable parameters are trained jointly with the LM from scratch. The difference of these strategies are discussed in Section 5.3. We tokenized and encoded the input using sentencepiece (Kudo and Richardson, 2018) and applied BPE with a vocabulary size of $32k$ . We use sacreBLEU for BLEU calculation (Post, 2018). For models with token-level policies, we trained", + "bbox": [ + 507, + 661, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3Technically, the word-level policy derived from the token-level Wait-k through the conversion process in Section 3.3 can wait between 1 and $k$ tokens, depending on the input encoding. Therefore, it is not equivalent to the word-level Wait-k policy we define here, which always waits for $\\mathbf{k}$ words.", + "bbox": [ + 112, + 856, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15621", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models with the official implementations $^{4 5 6}$ , and implemented word-level policies based on these implementations. More training details are described in A.2.", + "bbox": [ + 112, + 83, + 489, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Main results", + "text_level": 1, + "bbox": [ + 112, + 160, + 260, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of each system is compared in Figure 5. Notably, when measuring latency at the word level, the word-level policy proves to be highly effective for all three systems across different latency levels and datasets, resulting in superior performance. The only exception is observed in the En-Fr ITST models that demonstrate similar levels of performance. The incorporation of an LM using the proposed LM-fused attention further enhances the performance for all word-level configurations. This observation highlights the suitability of word-level policies for the LM-fused attention approach and underscores the effectiveness of leveraging an LM to enhance SiMT systems.", + "bbox": [ + 112, + 181, + 489, + 405 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, as depicted in Figure 11, the word-level policies also outperform or compete with the token-level policies in token-level latency. This can be attributed to the enhanced token representations under the word-level policy, thanks to the contextual information provided by all other tokens belonging to the same word for each token.", + "bbox": [ + 112, + 407, + 489, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Analysis", + "text_level": 1, + "bbox": [ + 112, + 532, + 223, + 548 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of word-level policies from multiple angles, we conduct several analyses on various settings. All the experiments were conducted on WMT De.En with transformer-big unless specified otherwise.", + "bbox": [ + 112, + 558, + 487, + 639 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Ablation study", + "text_level": 1, + "bbox": [ + 112, + 650, + 275, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.1 Effects of Word-level READ and WRITE", + "text_level": 1, + "bbox": [ + 112, + 671, + 428, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gain insights into the functionality of word-level READ and WRITE actions, we trained Wait-k models with various policy settings and conducted a performance comparison. Specifically, we examined models with the following policy settings:", + "bbox": [ + 112, + 707, + 489, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WW: word-level READ and WRITE.", + "bbox": [ + 112, + 788, + 396, + 802 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TW: token-level READ and word-level WRITE.", + "bbox": [ + 112, + 804, + 477, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WT: word-level READ and token-level WRITE.", + "bbox": [ + 112, + 821, + 473, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TkTk: a simpler baseline policy which involves alternating reading k source tokens and writing k", + "bbox": [ + 112, + 837, + 485, + 869 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1617a6da950878273b5b9d7426a25edf9b884d66ba1fa0358f278f70b88bb03a.jpg", + "image_caption": [ + "(a)", + "Figure 6: Ablation studies for word-level policies. (a): Comparison of word-level Wait-k policies with different policies. (b): Comparison of word-level Wait-k models with and without the intra-word bidirectional encoding." + ], + "image_footnote": [], + "bbox": [ + 515, + 86, + 687, + 206 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c1e514dc8f8250f15e8c991565d6d7356e7701d4a87a47975224a0d4f40d773.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 700, + 86, + 873, + 206 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "target tokens without considering word boundaries.", + "bbox": [ + 507, + 323, + 882, + 338 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are presented in Figure 6 (a). The word-level policy (WW) consistently outperforms TW across all latency settings. This is attributed to its imbalance between the number of source and target prefixes processed in each step. Additionally, WT achieves a minimal AL of approximately 3.6, indicating that it is not well-suited for scenarios that require low latency. Lastly, TkTk shows significantly worse performance than WW, suggesting that reading or writing a few consecutive tokens without considering semantic boundaries offers no benefits, unlike word-level policies.", + "bbox": [ + 507, + 340, + 882, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.2 Effects of intra-word bidirectional encoding", + "text_level": 1, + "bbox": [ + 507, + 543, + 836, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to assess the impact of the proposed intraword bidirectional encoding, we trained word-level Wait-K models with and without it and compared the accuracy of the two models across different AL settings. The results are presented in Figure 6 (b).", + "bbox": [ + 507, + 579, + 882, + 658 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remarkably, the model equipped with the intraword bidirectional encoding consistently achieved higher BLEU scores compared to the model without it, across all tested latency settings. This provides strong evidence of the effectiveness of the intra-word bidirectional encoding in enhancing SiMT performance.", + "bbox": [ + 507, + 659, + 882, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Effectiveness of word-level policies for LM", + "text_level": 1, + "bbox": [ + 507, + 784, + 880, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we aim to explore the significance of word-level policies in leveraging LM for SiMT. We compare different configurations based on three factors:", + "bbox": [ + 507, + 806, + 882, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Word vs. Token: The policy type that the model operates with.", + "bbox": [ + 507, + 870, + 880, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In-domain vocab vs. LM vocab: Whether the", + "bbox": [ + 507, + 903, + 880, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "4Efficient Wait-k: https://github.com/elbayadm/attn2d", + "bbox": [ + 134, + 878, + 465, + 891 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "${}^{5}$ MoE Wait-k: https://github.com/ictnlp/MoE-Waitk", + "bbox": [ + 136, + 891, + 452, + 904 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{6}$ ITST: https://github.com/ictnlp/ITST", + "bbox": [ + 136, + 904, + 369, + 917 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15622", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/779b878c35c67f128b72cafc92caeeafcaf256e3dd6b25c7f9734de8b0e5742f.jpg", + "image_caption": [ + "Figure 7: Comparison of models with different LM integration." + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 455, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model uses an in-domain vocabulary obtained from the in-domain training data or uses the LM's vocabulary for the source language. Note that the use of \"in-domain vocab\" is specific to the \"Word\" configuration due to the vocabulary mismatch.", + "bbox": [ + 112, + 305, + 489, + 384 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LM-fused attention vs. LM embed: Whether the model incorporates an LM using the LM-fused attention or replacing the embedding layer of the encoder with the LM's embedding (Xu et al., 2021). The latter approach uses \"LM vocab\" by design.", + "bbox": [ + 112, + 385, + 489, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 showcases the results. The models with word-level policies consistently outperform those with token-level policies by a significant margin in both LM-fused attention and LM embedding settings, underscoring the importance of word-level policy adoption for effective LM integration. The top-performing configuration is the proposed LMAtn-In-domain vocab-Word, demonstrating that the highest translation accuracy is achieved when the SiMT model operates with an in-domain vocabulary. Additionally, it is evident that the \"LM embed\" approach performs notably worse than the proposed LM-fused attention, further affirming the latter's superiority.", + "bbox": [ + 112, + 467, + 489, + 692 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 Effects of LM-fused attention with various LMs and training configurations", + "text_level": 1, + "bbox": [ + 112, + 703, + 487, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the effectiveness and broad applicability of our proposed LM integration, we conducted experiments on the IWSLT17 En-Fr dataset with two decoder-only LMs of different sizes: the 137M parameter GPT-2 model (Radford et al., 2018) and the XGLM-564M model. Additionally, we explore the option of training models in the single training stage instead of the two-stage training. The results, presented in Figure 8, demonstrate that GPT-2 model also exhibits improved performance with LM-fused attention, although their impact is", + "bbox": [ + 112, + 741, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c99091fa070de565656017ad1894b99576ec056ee67b7d47ac0af000302a2a50.jpg", + "image_caption": [ + "(a)", + "Figure 8: Translation accuracy comparison of LM-fused attention models with different training configurations. (a): En-Fr Transformer small (b) De-En Transformer big" + ], + "image_footnote": [], + "bbox": [ + 515, + 86, + 690, + 206 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31ef32ad43fd802b73ec692f820117850df0cf34c92ee5f6260366c2e282f676.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 702, + 86, + 875, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "naturally less pronounced compared to XGLM due to the difference in model size. Moreover, although models trained using the single-stage training generally exhibit lower performance compared to those trained using the two-stage training, they still outperform models without the LM for most configurations. This indicates that the LM-fused attention is applicable to various types of LMs and remains effective even when using the single-stage training strategy. This flexibility allows users to choose a training approach and model configuration that aligns best with their desired accuracy goals and computational constraints.", + "bbox": [ + 507, + 322, + 884, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4 Policy quality comparison", + "text_level": 1, + "bbox": [ + 507, + 542, + 761, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the accuracy of policies in determining when to read or write, we adopt the methodology of estimating the quality of a policy in prior research (Zhang and Feng, 2022b,c; Guo et al., 2022). We measure the quality of a policy by analyzing the proportion of aligned source words received before translating on RWTH De-En alignment dataset7.", + "bbox": [ + 507, + 563, + 882, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To ensure accurate word-level alignment calculation, we consider an aligned source word is read before writing a ground truth (GT) target word if the last token of the source word is read before the first token of the target word is written. The results of this analysis are presented in Figure 9. It is observed that word-level policies, both for ITST and Wait-k, exhibit better alignments across most latency settings. This suggests that word-level policies contribute to revising premature WRITE actions by guiding the model to read the remaining tokens of the aligned word, without negatively impacting the model's latency.", + "bbox": [ + 507, + 676, + 882, + 885 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "7https://www-i6.informatik.rwth-aachen.de/goldAlignment/", + "bbox": [ + 509, + 891, + 739, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15623", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96dac2a388686619654ce0211c754f994a329e79389415a4a32950cde31ea028.jpg", + "image_caption": [ + "Figure 9: The percentage of aligned source words read before the translation of the target words started." + ], + "image_footnote": [], + "bbox": [ + 117, + 86, + 448, + 231 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Conclusions", + "text_level": 1, + "bbox": [ + 112, + 303, + 253, + 317 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper explores the potential benefits of word-level operations in SiMT systems. We propose word-level latency calculation to ensure fair and accurate latency comparisons. We introduce a conversion process that transforms token-level policies into word-level policies, enabling the processing of multiple subwords that form a word within a single READ or WRITE action. Additionally, we propose the integration of LM-fused attention, which combines an autoregressive LM model into SiMT models with word-level policies. Experimental results demonstrate the superiority of word-level policies compared to token-level policies, as well as the effectiveness of the LM integration. Our findings highlight the crucial role of word-level policies in the integration process.", + "bbox": [ + 112, + 328, + 489, + 586 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 Limitations", + "text_level": 1, + "bbox": [ + 112, + 596, + 248, + 612 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While the proposed word-level policy implementation is widely applicable to most existing SiMT systems, it is important to note that systems utilizing languages with a writing style that lacks spaces or other delimiters between words or sentences (e.g., Chinese) are unable to derive benefits from this approach. Furthermore, it is important to consider that while the proposed LM-fused attention proves effective in enhancing translation quality across all latency levels, integrating a large LM may necessitate a faster compute capability to fulfill the low-latency demands of the SiMT task.", + "bbox": [ + 112, + 621, + 489, + 815 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 843, + 213, + 857 + ], + "page_idx": 8 + }, + { + "type": "ref_text", + "text": "Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. 2019. Monotonic infinite lookback attention for simultaneous machine", + "bbox": [ + 114, + 865, + 487, + 917 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1313-1323, Florence, Italy. Association for Computational Linguistics.", + "Ozan Caglayan, Julia Ive, Veneta Haralampieva, Pranava Madhyastha, Loic Barrault, and Lucia Specia. 2020. Simultaneous machine translation with visual context. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2350-2361, Online. Association for Computational Linguistics.", + "Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuhito Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the IWSLT 2017 evaluation campaign. In Proceedings of the 14th International Conference on Spoken Language Translation, pages 2-14, Tokyo, Japan. International Workshop on Spoken Language Translation.", + "Chung-Cheng Chiu and Colin Raffel. 2017. Monotonic chunkwise attention. arXiv preprint arXiv:1712.05382.", + "Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? arXiv preprint arXiv:1606.02012.", + "Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3440-3453, Online. Association for Computational Linguistics.", + "Fahim Dalvi, Nadir Durrani, Hassan Sajjad, and Stephan Vogel. 2018. Incremental decoding and training methods for simultaneous translation in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 493-499, New Orleans, Louisiana. Association for Computational Linguistics.", + "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.", + "Maha Elbayad, Laurent Besacier, and Jakob Verbeek. 2020. Efficient Wait-k Models for Simultaneous Machine Translation. In Proc. Interspeech 2020, pages 1461-1465.", + "Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor O.K. Li. 2017. Learning to translate in real-time" + ], + "bbox": [ + 509, + 85, + 884, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15624", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with neural machine translation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1053-1062, Valencia, Spain. Association for Computational Linguistics.", + "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2022. Turning fixed to adaptive: Integrating post-evaluation into simultaneous machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2264–2278, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2023. Learning optimal policy for simultaneous machine translation via binary search. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2318-2333, Toronto, Canada. Association for Computational Linguistics.", + "Sathish Reddy Indurthi, Mohd Abbas Zaidi, Beomseok Lee, Nikhil Kumar Lakumarapu, and Sangha Kim. 2022. Language model augmented monotonic attention for simultaneous translation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 38-45, Seattle, United States. Association for Computational Linguistics.", + "Javier Iranzo Sanchez, Jorge Civera, and Alfons Juan-Ciscar. 2022. From simultaneous to streaming machine translation by leveraging streaming history. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6972-6985, Dublin, Ireland. Association for Computational Linguistics.", + "Patrick Kahardipraja, Brielen Madureira, and David Schlangen. 2021. Towards incremental transformers: An empirical analysis of transformer models for incremental NLU. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1178-1189, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1246–1257, Florence, Italy. Association for Computational Linguistics.", + "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint arXiv:1901.07291.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.", + "Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettle-moyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, and Xian Li. 2022. Few-shot learning with multilingual generative language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9019-9052, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Xin Liu, Baosong Yang, Dayiheng Liu, Haibo Zhang, Weihua Luo, Min Zhang, Haiying Zhang, and Jinsong Su. 2021a. Bridging subword gaps in pretrainfinetune paradigm for natural language generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6001-6011, Online. Association for Computational Linguistics.", + "Xuebo Liu, Longyue Wang, Derek F. Wong, Liang Ding, Lidia S. Chao, Shuming Shi, and Zhaopeng Tu. 2021b. On the copying behaviors of pre-training for neural machine translation. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 4265-4275, Online. Association for Computational Linguistics.", + "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.", + "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.", + "Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, Hua Wu, and Haifeng Wang. 2019a. STACL: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework. In Proceedings of the 57th Annual Meeting of the Association for" + ], + "bbox": [ + 510, + 85, + 884, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15625", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computational Linguistics, pages 3025-3036, Florence, Italy. Association for Computational Linguistics.", + "Xutai Ma, Juan Pino, James Cross, Liezl Puzon, and Jiatao Gu. 2019b. Monotonic multihead attention. arXiv preprint arXiv:1909.12406.", + "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics.", + "Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training.", + "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.", + "Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, and Douglas Eck. 2017. Online and linear-time attention by enforcing monotonic alignments. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 2837-2846. PMLR.", + "Shoetsu Sato, Jin Sakuma, Naoki Yoshinaga, Masashi Toyoda, and Masaru Kitsuregawa. 2020. Vocabulary adaptation for domain adaptation in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4269-4279, Online. Association for Computational Linguistics.", + "Felix Schneider and Alexander Waibel. 2020. Towards stream translation: Adaptive computation time for simultaneous machine translation. In Proceedings of the 17th International Conference on Spoken Language Translation, pages 228-236, Online. Association for Computational Linguistics.", + "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.", + "Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. 2023. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2334-2349, Toronto, Canada. Association for Computational Linguistics." + ], + "bbox": [ + 115, + 85, + 489, + 917 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2591-2600, Dublin, Ireland. Association for Computational Linguistics.", + "Rongxiang Weng, Heng Yu, Shujian Huang, Shanbo Cheng, and Weihua Luo. 2020. Acquiring knowledge from pre-trained model to neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9266-9273.", + "Rongxiang Weng, Heng Yu, Weihua Luo, and Min Zhang. 2022. Deep fusing pre-trained models into neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11468-11476.", + "Haoran Xu, Benjamin Van Durme, and Kenton Murray. 2021. BERT, mBERT, or BiBERT? a study on contextualized embeddings for neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6663-6675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Jiacheng Yang, Mingxuan Wang, Hao Zhou, Chengqi Zhao, Weinan Zhang, Yong Yu, and Lei Li. 2020. Towards making the most of bert in neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9378-9385.", + "Shaolei Zhang and Yang Feng. 2021. Universal simultaneous machine translation with mixture-of-experts wait-k policy. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7306-7317, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.", + "Shaolei Zhang and Yang Feng. 2022a. Gaussian multihead attention for simultaneous machine translation. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 3019–3030, Dublin, Ireland. Association for Computational Linguistics.", + "Shaolei Zhang and Yang Feng. 2022b. Information-transport-based policy for simultaneous translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 992-1013, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "Shaolei Zhang and Yang Feng. 2022c. Modeling dual read/write paths for simultaneous machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2461-2477, Dublin, Ireland. Association for Computational Linguistics." + ], + "bbox": [ + 510, + 85, + 882, + 917 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15626", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Shaolei Zhang and Yang Feng. 2023. Hidden markov transformer for simultaneous machine translation. arXiv preprint arXiv:2303.00257.", + "bbox": [ + 114, + 85, + 487, + 124 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shaolei Zhang, Yang Feng, and Liangyou Li. 2021. Future-guided incremental transformer for simultaneous translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14428-14436.", + "bbox": [ + 114, + 133, + 489, + 199 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Shaolei Zhang, Shoutao Guo, and Yang Feng. 2022. Wait-info policy: Balancing source and target at information level for simultaneous machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2249-2263, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.", + "bbox": [ + 114, + 208, + 489, + 300 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Baigong Zheng, Kaibo Liu, Renjie Zheng, Mingbo Ma, Hairong Liu, and Liang Huang. 2020. Simultaneous translation policies: From fixed to adaptive. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2847-2853, Online. Association for Computational Linguistics.", + "bbox": [ + 114, + 309, + 489, + 401 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. 2020. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823.", + "bbox": [ + 114, + 409, + 489, + 463 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A Appendix", + "text_level": 1, + "bbox": [ + 114, + 472, + 238, + 488 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.1 Token-level latency evaluation in previous works", + "text_level": 1, + "bbox": [ + 114, + 497, + 485, + 527 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Many previous works in the field of SiMT have employed token-level latency calculation as a fundamental metric for evaluating system performance. In the first group of papers, which includes (Arivazhagan et al., 2019; Ma et al., 2019b; Schneider and Waibel, 2020; Zheng et al., 2020; Zhang and Feng, 2022b), researchers explicitly stated their use of token-level latency calculation or incorporated token-level scores in their analyses. On the other hand, the second group, comprising (Zhang and Feng, 2022b; Elbayad et al., 2020; Zhang and Feng, 2022c, 2021, 2022a; Zhang et al., 2022; Guo et al., 2022; Zhang and Feng, 2023; Guo et al., 2023; Caglayan et al., 2020; Wang et al., 2023), can be identified by their publicly available official implementations, which incorporate token-level latency evaluation code. This underscores the prevalence of token-level latency calculation in assessing the effectiveness of SiMT systems in the cutting-edge SiMT research works.", + "bbox": [ + 112, + 533, + 489, + 854 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2 More training details", + "text_level": 1, + "bbox": [ + 114, + 866, + 331, + 881 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For Transformer model configurations, we follow the Efficient Wait-k (Elbayad et al., 2020)'s set", + "bbox": [ + 112, + 887, + 489, + 917 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "tings. $^{8}$ The details of hyperparameter settings for each model is in Table 1.", + "bbox": [ + 507, + 83, + 880, + 114 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "To ensure optimal performance of the ITST models, we conducted multiple training runs for each model configuration and dataset. We observed performance fluctuations across the training runs. To select the best-performing models for both token-level and word-level policies, we repeated the training process three times and selected the checkpoint with the lowest validation loss for each configuration.", + "bbox": [ + 507, + 116, + 884, + 259 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For LM integrated models, the intra-word bidirectional encoding was not applied to the LM to prevent the need for fine-tuning.", + "bbox": [ + 507, + 261, + 882, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.3 Examples and discussions", + "text_level": 1, + "bbox": [ + 507, + 319, + 761, + 335 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In Table 2, an illustrative case is presented to demonstrate the distinction between a token-level policy (token-level Wait-1) and a word-level policy (word-level Wait-1). At Step 2, the token-level model predicts the target token “B” after processing the subword “B” for the word “Beine” (which means “leg” in German). Subsequently, it fails to recover from this incorrect prediction and continues by predicting “ody_” as the next token, even after processing the remaining token “eine_”. In contrast, the word-level model accurately predicts “legs” (encoded as “leg s_”) after processing the complete word ‘Beine’ (encoded as ‘B eine_’). Furthermore, the token model also makes erroneous predictions for all steps when processing the word ‘blutüberströmmt.’ (encoded as ‘bl ut über ström t_’, meaning “covered in blood” in German), while the word-level model accurately predicts a correct word in a single step.", + "bbox": [ + 505, + 340, + 884, + 646 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Another case can be observed in Table 3, where the token-level model initiates the translation before fully reading the word \"Irgendetwas\" (encoded as \"Irgen det was\"). As a consequence, it produces an incorrect translation. On the other hand, the word-level model accurately translates the sentence by processing the complete word before making any predictions.", + "bbox": [ + 507, + 646, + 882, + 775 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.4 Table of main results", + "text_level": 1, + "bbox": [ + 507, + 785, + 722, + 800 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.5 Graphs of main results", + "text_level": 1, + "bbox": [ + 507, + 807, + 739, + 822 + ], + "page_idx": 11 + }, + { + "type": "page_footnote", + "text": "$^{8}$ https://github.com/elbayadm/attn2d", + "bbox": [ + 531, + 903, + 756, + 917 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15627", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/6f02bb8f188a8cb248e2712614dfedcb4bf155f86a411b5a423c84eac5ad4cd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hyperparametertransformer-smalltransformer-basetransformer-big
attention heads4816
embedding dime2565121024
ffn embedding dim102420484096
dropout0.30.30.3
encoder layers666
decoder layers666
lr5e-45e-45e-4
lr schedulerinverse sqrtinverse sqrtinverse sqrt
optimizerAdam (0.9, 0.98)Adam (0.9, 0.98)Adam (0.9, 0.98)
clip-norm000
warmup-updates400040004000
warmup-init-lr1e-71e-71e-7
weight decay000
label smoothing0.10.10.1
max tokens327685242881048576
", + "bbox": [ + 171, + 105, + 826, + 367 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8868b7b9646288d5fad9842cddb3ff3a7faf811aaff0945f1e1cafc0592865e6.jpg", + "table_caption": [ + "Table 1: Hyperparameters of each Transformer model." + ], + "table_footnote": [], + "table_body": "
Src:Meine_B eine_waren_bl ut über ström t.
Ref:My leg s_were Covered_in_blood.
StepStreaminig InputToken Wait-1 OutputWord Wait-1 Output
1MeineMyMy
2Meine_BB
3Meine_B eineodyleg s
4Meine_B eine_warenwaswere
5Meine_B eine_waren_blblue
6Meine_B eine_waren_bl utand
7Meine_B eine_waren_bl ut übermy
8Meine_B eine_waren_bl ut über strömleg
9Meine_B eine_waren_bl ut über ström ts
10Meine_B eine_waren_bl ut über ström t.were_blood - ri dden.blood dy.
", + "bbox": [ + 139, + 448, + 858, + 609 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/83390317b18088f0dcb2c19432051eafb4888477fbd11093c859a791f7a32385.jpg", + "table_caption": [ + "Table 2: An example case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.56, word-AL: 1.69) and Word-level Wait-1 (token-AL: 1.44, word-AL: 1.0)." + ], + "table_footnote": [], + "table_body": "
Src:Ir gen det was lag in der Luft .
Ref:Some thing was up .
StepStreaming InputToken Wait-1 OutputWord Wait-1 Output
1IrIr
2Ir gengen
3Ir gen detgen
4Ir gen det wasgenSome thing_
5Ir gen det was laggenlay_
6Ir gen det was lag insin_
7Ir gen det was lag in der,the_
8Ir gen det was lag in der Luftin
9Ir gen det was lag in der Luft .fact , was in the air .air .
", + "bbox": [ + 139, + 703, + 858, + 852 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Table 3: An Eexample case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.69, word-AL: 4.25) and Word-level Wait-1 (token-AL: 2.38, word-AL: 1.0).", + "bbox": [ + 112, + 860, + 882, + 890 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "15628", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/a7c2f1f08eb9e88dd9ef08d785d6c4d82ff39d39211e8d4a400c58abdefa40f8.jpg", + "image_caption": [ + "(a) De-En Wait-k Transformer base" + ], + "image_footnote": [], + "bbox": [ + 117, + 124, + 364, + 258 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/84c13e68369add4451739ed4985cf2fb0a7369e75084a4fb33c26288da6bc039.jpg", + "image_caption": [ + "(b) De-En MoE Transformer base", + "Figure 10: Results of translation quality v.s. word-level AL, De-En transformer-base." + ], + "image_footnote": [], + "bbox": [ + 376, + 124, + 620, + 258 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cd652d3f4e088a493be32d365fa0f641cc4dc5ab82a6105e750d9829cf91bf6b.jpg", + "image_caption": [ + "(c) De-En ITST Transformer base" + ], + "image_footnote": [], + "bbox": [ + 635, + 124, + 878, + 258 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0e4a8516a1f8797398d67d483ca155986b84acc7cc5ec468050494c9897920fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 395, + 361, + 527 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/0ada3edc3344d0d06ec3dbf9bd85e2fb68caf26a3742abbb66fb233aa6b1dbe3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 376, + 395, + 621, + 527 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/3f5c144dc9fe3d0c40307c39c86dab14519469c5de082683a35f9b04a5c995d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 395, + 878, + 527 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/323f89b5334a770fa7032530970985fae90748ef9da6d7a287e0d0b646cc03df.jpg", + "image_caption": [ + "(a) En-Fr Wait-k Transformer small" + ], + "image_footnote": [], + "bbox": [ + 117, + 546, + 361, + 678 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/aafe8835c201e63e967ad66c630b512e34ecebeabb9c769a7b3549bc2c37a95f.jpg", + "image_caption": [ + "(b) En-Fr MoE Transformer small" + ], + "image_footnote": [], + "bbox": [ + 376, + 546, + 621, + 678 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/eee39750e797615ae83ba316087c8f4da0123f2576d70ffd41df469544822c36.jpg", + "image_caption": [ + "(c) En-Fr ITST Transformer small" + ], + "image_footnote": [], + "bbox": [ + 635, + 546, + 878, + 678 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/cbf2f32a207e488330d3e6161e6f6e583f398e48e59c8ea6d91e5c204171cc12.jpg", + "image_caption": [ + "(d) De-En Wait-k Transformer base", + "(g) De-En Wait-k Transformer big", + "Figure 11: Results of translation quality v.s. token-level AL." + ], + "image_footnote": [], + "bbox": [ + 117, + 695, + 361, + 827 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/21ab652590c28f0f7a7cebd049f50ff4ba38d82570f5d7c63b42d018b98f631c.jpg", + "image_caption": [ + "(e) De-En MoE Transformer base", + "(h) De-En MoE Transformer big" + ], + "image_footnote": [], + "bbox": [ + 376, + 695, + 621, + 827 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/177f24907f5fd2547d3881b704fdf7ff430fdffd2d5f4ce05a256029d5feb0dd.jpg", + "image_caption": [ + "(f) De-En ITST Transformer base", + "(i) De-En ITST Transformer big" + ], + "image_footnote": [], + "bbox": [ + 635, + 695, + 878, + 827 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "15629", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/f45195fa04605265623e7451aab3c7af80b050b8d9992cc96066d26e898704ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token-level Wait-k
ktoken ALword ALBLEU
11.951.9132.32
33.833.4338.90
55.674.8240.68
77.516.2340.95
99.287.5841.06
1110.918.8241.14
1312.399.9541.12
Word-level Wait-k
ktoken ALword ALBLEU
12.241.7432.43
34.183.3239.36
56.555.1541.09
78.856.9441.48
910.928.5741.57
1112.759.9841.58
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.181.7433.47
34.233.3740.83
56.585.1942.78
78.866.9743.00
910.948.5843.08
1112.7610.0043.09
", + "bbox": [ + 157, + 273, + 445, + 683 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/5f1e40807cbfee63a2bc86b703237032e584713f25b2da20279d617de819fd42.jpg", + "table_caption": [ + "Table 4: Main results of transformer-small Waik-k models on IWLST17 En.Fr dataset." + ], + "table_footnote": [], + "table_body": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.901.8832.20
33.833.4238.92
55.694.8440.51
77.536.2540.83
99.317.6040.92
1110.948.8540.95
1312.429.9740.96
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.011.6032.01
34.183.3239.28
56.575.1740.77
78.866.9541.03
910.948.5841.11
1112.769.9941.10
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.171.7233.01
34.223.3740.46
56.565.1742.35
78.846.9542.61
910.918.5742.66
1112.749.9942.67
", + "bbox": [ + 552, + 273, + 838, + 682 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Table 5: Main results of transformer-small MoE Waik-k models on IWSLT17 En.Fr dataset.", + "bbox": [ + 507, + 692, + 880, + 720 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "15630", + "bbox": [ + 477, + 927, + 526, + 940 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/4cf5870a862751b5063aeb22225ccc9ed6ed18e72a9934510e3b4f2484cde731.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token-level ITST
δtoken ALword ALBLEU
0.23.333.0537.81
0.33.593.2538.64
0.44.143.7039.40
0.55.184.5940.19
0.66.715.9240.84
0.78.627.6341.27
0.759.568.4541.34
0.810.729.4141.42
0.8512.1110.5341.42
0.913.8411.8041.42
Word-level ITST
δtoken ALword ALBLEU
0.23.272.6437.25
0.33.462.8037.88
0.43.913.2138.69
0.54.673.9139.57
0.65.814.9640.44
0.77.296.2840.94
0.758.217.0541.13
0.89.418.0241.36
0.8511.199.3741.50
0.913.3611.0041.53
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.252.6138.30
0.33.302.6538.75
0.43.482.8139.61
0.53.943.2340.50
0.64.804.0441.35
0.76.025.1941.88
0.756.765.8642.09
0.87.726.6742.30
0.858.997.6842.56
0.910.909.1642.63
", + "bbox": [ + 149, + 184, + 450, + 770 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/de2dd378cf60b689e0f8a55f8a4a3cc518a59e6b6fffed493b522eb8737e3a14.jpg", + "table_caption": [ + "Table 6: Main results of transformer-small ITST models on IWSLT17 En.Fr dataset." + ], + "table_footnote": [], + "table_body": "
Token-level Wait-k
ktoken ALword ALBLEU
10.520.7517.63
32.161.9123.24
53.923.0526.14
75.774.2327.46
97.765.5028.77
119.646.6929.37
1311.497.8629.88
1513.188.9229.89
1714.799.9330.19
Word-level Wait-k
ktoken ALword ALBLEU
12.961.8424.02
35.333.4328.08
58.085.2329.35
710.827.0129.87
913.478.7230.17
1115.8810.2530.38
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.961.8424.71
35.273.3828.49
58.055.2029.95
710.817.0030.56
913.448.6930.74
1115.8510.2430.95
", + "bbox": [ + 552, + 256, + 838, + 699 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Table 7: Main results of transformer-base Waik-k models on WMT15 De.En dataset.", + "bbox": [ + 507, + 708, + 880, + 736 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "15631", + "bbox": [ + 477, + 927, + 522, + 940 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/a0a4c3007bbf594f661970f89be975077d28c64a72f69d6260fc80e77028437c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.171.1117.28
32.502.0923.29
54.173.1825.95
76.004.3627.86
97.815.5128.81
119.626.6629.53
1311.427.8029.66
1513.158.8929.81
1714.809.9129.79
", + "bbox": [ + 157, + 256, + 443, + 436 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/80a3f5ca7a97da89b89c9198454ce5e1a8d09b5641e6ca5b3d633faa4e98611a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
13.121.9123.59
35.293.3927.75
58.035.1929.29
710.796.9830.14
913.388.6530.39
1115.8110.230.6
", + "bbox": [ + 157, + 437, + 443, + 567 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/c20796ce6611dc73a8c4ffa97d571256d860ad733c34cfe6aca3c164e60769d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
13.081.9224.97
35.323.4428.66
58.085.2330.21
710.877.0430.63
913.478.7330.9
1115.8810.2631.0
", + "bbox": [ + 157, + 568, + 443, + 697 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/2fe04fde2cfb6d0d0088f385c3c2e60a7a4009023f7f335f64a116078aa5b797.jpg", + "table_caption": [ + "Table 8: Main results of transformer-base MoE Waik-k models on WMT15 De.En dataset." + ], + "table_footnote": [], + "table_body": "
Token-level ITST
δtoken ALword ALBLEU
0.22.452.0822.19
0.32.892.3824.15
0.43.822.9626.21
0.55.233.9228.19
0.67.285.5029.25
0.79.877.4829.47
0.7511.668.8129.76
0.813.7910.3829.79
0.8516.6012.1129.90
0.919.8113.8529.71
", + "bbox": [ + 544, + 185, + 845, + 381 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/45b1f2f452d95bc1ae4cfd6baf626f680655c71f178cce3c6bf282c890649ca6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level ITST
δtoken ALword ALBLEU
0.24.132.6426.81
0.34.192.6927.01
0.44.392.8127.47
0.55.093.3028.04
0.66.434.2828.87
0.78.605.9329.54
0.759.916.9229.54
0.811.548.0929.67
0.8513.839.6129.82
0.916.7511.3829.93
", + "bbox": [ + 544, + 382, + 845, + 574 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/e6a677d3aa2464e7579de5bfac355e06e292e3c7e02df482165b6c9f5575b58c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.24.152.6427.74
0.34.242.7027.99
0.44.552.9228.44
0.55.333.4929.15
0.66.774.6329.97
0.78.576.0830.49
0.759.516.8030.53
0.810.717.6730.60
0.8512.328.7230.68
0.914.289.9330.71
", + "bbox": [ + 544, + 576, + 845, + 770 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Table 9: Main results of transformer-base ITST models on WMT15 De.En dataset.", + "bbox": [ + 507, + 781, + 882, + 809 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "15632", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 16 + }, + { + "type": "table", + "img_path": "images/8341018f2ceaecb0a4c02bccbb7234fa7303c99ba3930b43e3175d70f36e4351.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token-level Wait-k
ktoken ALword ALBLEU
1-0.130.3919.20
31.801.7124.01
53.742.9526.86
75.814.2628.86
97.805.5329.96
119.726.7430.62
1311.517.8730.91
1513.208.9531.16
1714.849.9731.31
", + "bbox": [ + 157, + 256, + 443, + 437 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/576895b19625c6abe42f685905293d4a8237e8b631dbd4a18baae5f5c7d0174e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level Wait-k
ktoken ALword ALBLEU
12.551.5924.63
35.113.3028.82
57.955.1630.43
710.776.9931.35
913.378.67831.86
1115.8010.2232.03
", + "bbox": [ + 157, + 437, + 443, + 567 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/2dd2afc0415868650d3ced5ce1e32e9880bf3bb5124c1e7efb60e19bc69fb1f2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.891.7926.21
35.213.3630.16
57.985.1731.36
710.827.0132.29
913.428.7132.58
1115.7410.1632.45
", + "bbox": [ + 157, + 567, + 443, + 697 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/6488edd0bf33944d926ae4289a4578216ef5571f7ffa6461bca910f2a26efde6.jpg", + "table_caption": [ + "Table 10: Main results of transformer-big Waik-k models on WMT15 De.En dataset." + ], + "table_footnote": [], + "table_body": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
10.860.9418.44
32.191.9224.07
53.903.0426.70
75.704.1928.47
97.695.4729.59
119.606.6730.36
1311.447.8330.51
1513.148.9130.56
1714.759.9130.65
", + "bbox": [ + 552, + 256, + 838, + 437 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/d94c887699214c539e6755c3467e04adcbfb0332e0484828d815cff89dbc626b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.861.7624.79
35.243.3728.95
57.985.1530.26
710.776.9731.01
913.378.6631.47
1115.8010.2031.55
", + "bbox": [ + 552, + 437, + 838, + 567 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/5281560146f781c0093c98ae66e11f7eb04243b8577c17b9caac85c3cf526158.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.601.6025.26
35.193.3129.03
58.015.1630.79
710.836.9931.66
913.438.6731.82
1115.8310.2132.17
", + "bbox": [ + 552, + 567, + 838, + 697 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Table 11: Main results of transformer-big MoE Waik-k models on WMT15 De.En dataset.", + "bbox": [ + 507, + 708, + 882, + 736 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "15633", + "bbox": [ + 477, + 927, + 524, + 940 + ], + "page_idx": 17 + }, + { + "type": "table", + "img_path": "images/570f67e802876b0c8be79a328058a509cecb9bdb8fe48832da96736f62ee0cdb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Token-level ITST
δtoken ALword ALBLEU
0.21.491.5523.22
0.32.412.1224.75
0.43.933.0626.67
0.55.994.3928.23
0.69.126.5729.85
0.712.699.3231.02
0.7514.7210.7531.28
0.816.9512.2331.54
0.8519.0613.4931.70
0.921.2614.6931.84
Word-level ITST
δtoken ALword ALBLEU
0.23.692.3526.33
0.33.792.4226.62
0.43.942.5426.62
0.54.693.0227.55
0.66.604.2728.99
0.710.286.7730.41
0.7512.508.3430.95
0.814.229.6531.54
0.8515.9010.8131.56
0.918.1312.2331.61
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.972.5328.12
0.34.002.5528.13
0.44.102.6228.29
0.54.392.8228.92
0.65.373.4629.66
0.77.855.1331.18
0.759.736.4831.77
0.811.908.0432.10
0.8514.049.5732.34
0.916.1410.9732.62
", + "bbox": [ + 149, + 184, + 450, + 770 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Table 12: Main results of transformer-big ITST models on WMT15 De.En dataset.", + "bbox": [ + 112, + 781, + 489, + 809 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "15634", + "bbox": [ + 477, + 928, + 524, + 940 + ], + "page_idx": 18 + } +] \ No newline at end of file diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_model.json b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..afd0f7723e02cc7f3f95c0b3ac731c0ea060d390 --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_model.json @@ -0,0 +1,3406 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.136, + 0.09, + 0.863, + 0.11 + ], + "angle": 0, + "content": "Enhanced Simultaneous Machine Translation with Word-level Policies" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.143, + 0.616, + 0.159 + ], + "angle": 0, + "content": "Kang Kim*, Hankyu Cho*" + }, + { + "type": "text", + "bbox": [ + 0.463, + 0.161, + 0.538, + 0.174 + ], + "angle": 0, + "content": "XL8 Inc." + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.178, + 0.607, + 0.192 + ], + "angle": 0, + "content": "{kai, matthew}@x18.ai" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.277, + 0.461, + 0.575 + ], + "angle": 0, + "content": "Recent years have seen remarkable advances in the field of Simultaneous Machine Translation (SiMT) due to the introduction of innovative policies that dictate whether to READ or WRITE at each step of the translation process. However, a common assumption in many existing studies is that operations are carried out at the subword level, even though the standard unit for input and output in most practical scenarios is typically at the word level. This paper demonstrates that policies devised and validated at the subword level are surpassed by those operating at the word level, which process multiple subwords to form a complete word in a single step. Additionally, we suggest a method to boost SiMT models using language models (LMs), wherein the proposed word-level policy plays a vital role in addressing the subword disparity between LMs and SiMT models. Code is available at https://github.com/x18-ai/WordSiMT." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.584, + 0.26, + 0.599 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.608, + 0.49, + 0.768 + ], + "angle": 0, + "content": "Simultaneous Machine Translation (SiMT) commences the translation process while simultaneously receiving the input, making it an effective approach for applications that require minimal latency such as simultaneous interpretation or live broadcast. The development of a novel policy is central to research efforts in SiMT. This policy dictates the translation process by determining whether to execute a READ or WRITE action at each step of the process." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.77, + 0.49, + 0.897 + ], + "angle": 0, + "content": "Neural SiMT models, like offline Neural Machine Translation (NMT) models, commonly employ Byte Pair Encoding (BPE) (Sennrich et al., 2016) or similar techniques to encode an input sentence into a sequence of tokens. Typically, a single READ or WRITE action of a SiMT policy is responsible for handling an encoded token, which may sometimes be a word but often a subword." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.25, + 0.885, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.374, + 0.885, + 0.432 + ], + "angle": 0, + "content": "Figure 1: Illustration of both token-level Wait-1 (red arrow lines) and word-level Wait-1 (green arrow lines) policies. The symbol \" \\( \\_ \\) \" at the end of a token indicates a word boundary." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.457, + 0.884, + 0.6 + ], + "angle": 0, + "content": "The development of BPE-based SiMT models and their policies has resulted in researchers focusing on working at the subword level. The performance analysis and implementation of many SiMT systems, to the best of our knowledge, have been carried out on encoded sequences of source and target subwords, rather than on the original source and target sentences1. This has led to two critical issues that need to be addressed." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.602, + 0.884, + 0.698 + ], + "angle": 0, + "content": "The first issue is the lack of a standardized tokenization and encoding scheme, meaning that different implementations may employ varying token sequences to encode identical text. This variability can impact latency evaluation results and complicate score comparisons across different systems." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.7, + 0.885, + 0.86 + ], + "angle": 0, + "content": "The second issue is the missed opportunity to process more source tokens before writing each target token without added latency. For a BPE-based SiMT model, the input must be received on a word-by-word basis to ensure proper encoding of each word into a sequence of subwords. Consequently, when the model encodes a word and performs a READ to process only a subword, it delays the reading of the remaining subwords without any benefit in actual latency², and may adversely impact" + }, + { + "type": "page_footnote", + "bbox": [ + 0.509, + 0.869, + 0.883, + 0.892 + ], + "angle": 0, + "content": "1We provide a list of reference works pertaining to this case in A.1" + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.893, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Here we assume that the processing time of a READ is negligible, compared to the interval between receiving words." + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.869, + 0.884, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.263, + 0.919 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.477, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15616" + }, + { + "type": "footer", + "bbox": [ + 0.21, + 0.946, + 0.788, + 0.959 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15616-15634" + }, + { + "type": "footer", + "bbox": [ + 0.278, + 0.96, + 0.72, + 0.972 + ], + "angle": 0, + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.083, + 0.885, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.167, + 0.885, + 0.224 + ], + "angle": 0, + "content": "Figure 2: An exemplary case depicting the difference between token-level and word-level Wait-k policies and their Average Lagging (AL) scores. The token-level model begins translating in the middle of reading \"Tremendously\" and fails to recover from the incorrectly translated target prefix. On the other hand, the word-level model processes \"Tremendously\" in a single READ action and produces a correct translation." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.25, + 0.489, + 0.362 + ], + "angle": 0, + "content": "translation quality by causing the model to rely on incomplete representations extracted from partially-read words. Similarly, performing a WRITE to generate a subword earlier than the remaining subwords does not necessarily reduce latency, as the subword must wait for a complete word to be generated before it can be displayed." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.366, + 0.489, + 0.542 + ], + "angle": 0, + "content": "In this paper, we show that establishing the unit of measuring and operating SiMT systems at the word level, rather than the subword level, offers a viable solution to these issues. Specifically, to tackle the first issue, we propose word-level latency metric calculation for measuring the latency of SiMT systems. This not only enables consistent comparisons between different systems but also provides a more accurate reflection of the actual latency experienced in SiMT applications that display translation results word by word." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.546, + 0.489, + 0.819 + ], + "angle": 0, + "content": "To address the second issue, we illustrate that an existing token-level policy can be transformed into a word-level policy that inherently overcomes the second issue, resulting in improved performance. Word-level policies take into consideration word boundaries and perform a READ or WRITE action to sequentially process a sequence of tokens that form a complete word. This conversion process can be applied to any token-level policies, and our experiments reveal that state-of-the-art fixed and adaptive policies exhibit significantly better performance when transformed into their word-level counterparts. Notably, these word-level policies often outperform token-level policies, even when evaluated using a token-level latency metric, due to their enhanced utilization of input and output tokens." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.824, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Additionally, to boost translation accuracy, we suggest incorporating a pre-trained language model (LM) into a SiMT model, where the word-level policy plays a crucial role as a pivotal component. One of the major hurdles in utilizing an LM for SiMT is the vocabulary mismatch between the SiMT model" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.25, + 0.885, + 0.41 + ], + "angle": 0, + "content": "and the LM. The difficulty of handling subword disparities when utilizing an LM for a downstream task is widely acknowledged (Liu et al., 2021a; Wang et al., 2022), and it becomes particularly problematic in SiMT, as inconsistent subwords between the LM and SiMT model make processing the same source or target prefix challenging. Our study demonstrates that our proposed word-level policy effectively tackles this challenge, enabling a successful integration of LMs into SiMT systems." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.423, + 0.665, + 0.437 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.448, + 0.831, + 0.463 + ], + "angle": 0, + "content": "2.1 Simultaneous Machine Translation" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.469, + 0.884, + 0.693 + ], + "angle": 0, + "content": "SiMT systems that employ a fixed policy utilize a pre-defined sequence of READ and WRITE operations for each source sentence. STATIC-RW (Dalvi et al., 2018) and Wait-k (Ma et al., 2019a) policies first read k source tokens, then alternate between reading and writing a single token. Elbayad et al. (2020) propose the multi-path training of a Wait-k model to train a single model that supports different k values at test time. Zhang et al. (2021) improve Wait-k policy using knowledge distillation from an offline MT model, while Zhang and Feng (2021) suggest a Mixture-of-Experts Wait-k Policy where predictions from multiple k values are combined inside a single model." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.695, + 0.884, + 0.919 + ], + "angle": 0, + "content": "In contrast, research efforts on adaptive policies focus on the development of dynamic decision-making processes for READ/WRITE actions. Cho and Esipova (2016) firstly introduce model-based adaptive criteria for Neural SiMT. Gu et al. (2017) propose to learn a policy by using reinforcement learning. Raffel et al. (2017) introduce Monotonic Attention that ensures monotonic alignment in the attention mechanism. Succeeding works improve it by extending the alignment window (Chiu and Raffel, 2017; Arivazhagan et al., 2019), extending it as monotonic multi-head attention (MMA) (Ma et al., 2019b) or learning transposed policies between the forward and backward models. (Zhang and Feng," + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.489, + 0.198 + ], + "angle": 0, + "content": "2022c). Zheng et al. (2020) derive a policy by composing Wait-k models trained with different values of k. Zhang and Feng (2022a) model to predict the alignment between each target token and the source token. Zhang and Feng (2022b) measure accumulated information from source tokens to decide whether to write the current target token." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.2, + 0.49, + 0.328 + ], + "angle": 0, + "content": "Despite significant advancements, the impact of operating policies at the word level has not been thoroughly explored in existing works, which have mainly focused on developing and evaluating systems at the token level. In this paper, we address this gap by demonstrating that implementing various types of policies at the word level consistently outperforms their token-level counterparts." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.345, + 0.418, + 0.361 + ], + "angle": 0, + "content": "2.2 Utilizing pre-trained LM for MT" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.369, + 0.49, + 0.61 + ], + "angle": 0, + "content": "Since the successful utilization of Transformer-based LMs pre-trained on large text corpora for downstream NLP tasks (Devlin et al., 2019; Liu et al., 2019; Lample and Conneau, 2019), the utilization of these models for MT has become a significant research area. Several studies have demonstrated the effectiveness of incorporating encoder-only LMs into NMT models. Weng et al. (2020); Yang et al. (2020) combine the LM representations with the encoder's representation using a gating mechanism. Zhu et al. (2020) propose attention between BERT and both the encoder and decoder. Weng et al. (2022) leverage mBERT as an encoder and introduce a decoder that attends to grouped representations of the encoder output." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.612, + 0.49, + 0.789 + ], + "angle": 0, + "content": "Another research direction focuses on developing LMs with the encoder-decoder architecture designed for NMT as the target downstream task (Lewis et al., 2020; Liu et al., 2020). These models show improvements particularly for low-resource language pairs. To enhance their adaptation for MT, various methods have been proposed, including fine-tuning specific parts of the LMs (Cooper Stickland et al., 2021), reducing domain mismatch and overestimation (Wang et al., 2022) and mitigating the copying behavior (Liu et al., 2021b)." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.791, + 0.49, + 0.919 + ], + "angle": 0, + "content": "The integration of pre-trained LMs into SiMT remains an underexplored area of research. To date, Indurthi et al. (2022) is the only related study we are aware of. It improves MMA by integrating the LM's prediction of the next target token, which is encoded using their model's vocabulary before being inputted into the model. However, this approach sacrifices the semantic coherence of the" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.149 + ], + "angle": 0, + "content": "original tokens due to token fragmentation. Additionally, their approach falls under target-side LM integration, overlooking the potential advantages of source-side LM integration." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.15, + 0.886, + 0.343 + ], + "angle": 0, + "content": "In this paper, we demonstrate an effective way of integrating a source-side LM into SiMT systems, offering a more versatile solution that can be integrated into most existing neural SiMT models. Building upon previous research conducted in offline MT (Zhu et al., 2020), we introduce essential modifications, with a particular focus on word-level policies as a pivotal component. The effective management of vocabulary mismatches between the LM and the SiMT model is contingent upon the successful implementation of a word-level SiMT policy, a key aspect that we address in our study." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.357, + 0.706, + 0.373 + ], + "angle": 0, + "content": "3 Proposed Methods" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.384, + 0.886, + 0.48 + ], + "angle": 0, + "content": "In this section, we propose the concept of employing a word-level latency metric and outline our conversion process for translating token-level policies into their word-level equivalents. Additionally, we present an integration of LM into SiMT, highlighting the advantages of utilizing word-level policies." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.493, + 0.663, + 0.507 + ], + "angle": 0, + "content": "3.1 Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.515, + 0.884, + 0.643 + ], + "angle": 0, + "content": "Given a source sentence \\(\\mathbf{x} = (x_{1}, x_{2}, \\ldots x_{n})\\), the goal of a SiMT model is to generate a target sentence of \\(\\mathbf{y} = (y_{1}, y_{2}, \\ldots y_{m})\\) while minimizing latency metrics. A SiMT model's policy, represented by the variable \\(g_{i}\\), determines the number of source tokens to process before predicting target token \\(y_{i}\\). Then the probability of generating \\(\\mathbf{y}\\) given \\(\\mathbf{x}\\) is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.658, + 0.884, + 0.702 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {y} | \\mathbf {x}) = \\prod_ {i} ^ {| \\mathbf {y} |} p \\left(y _ {i} \\mid \\mathbf {x} _ {\\leq g _ {i}}, \\mathbf {y} _ {< i}; \\theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.71, + 0.885, + 0.742 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the model's parameters which are commonly optimized with a cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.886, + 0.919 + ], + "angle": 0, + "content": "Transformer encoder-decoder model (Vaswani et al., 2017) is currently the most widely used architecture for SiMT. To avoid redundant encoding of the input sequence after each READ operation, the encoder is typically modified to encode the source tokens unidirectionally (Elbayad et al., 2020). Alternatively, more advanced techniques like the recurrent Linear Transformer (Kahardipraja et al., 2021) or Partial Bidirectional Encoding (Iranzo Sanchez et al., 2022) can be adopted to enhance the encoding capabilities further." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15618" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.182 + ], + "angle": 0, + "content": "During the evaluation of SiMT systems, translation quality is commonly assessed in conjunction with the latency required for generating translations. Various metrics have been proposed to calculate latency scores, with Average Lagging (AL) (Ma et al., 2019a) being the most commonly used metric." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.196, + 0.486, + 0.212 + ], + "angle": 0, + "content": "3.2 Measuring latency based on the word level" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.219, + 0.489, + 0.363 + ], + "angle": 0, + "content": "As detailed in A.1, a substantial body of prior research work assesses the performance of SiMT systems by utilizing a latency metric on encoded source tokens under different tokenization and encoding schemes. This practice results in each system being evaluated on non-identical token sequences for the same dataset, thereby making it challenging to accurately compare scores across different systems." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.365, + 0.49, + 0.558 + ], + "angle": 0, + "content": "To address this, we propose word-level latency score calculation by considering the word boundaries in token-level sequences. Specifically, when the first token of a source word is processed through a READ operation, we consider it as reading the corresponding word. Similarly, when the last token of a target word is written via a WRITE operation, we consider it as writing that word. By doing so, the latency scores are calculated consistently, regardless of the tokenization and encoding of the input. This ensures that results from different systems can be compared fairly." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.573, + 0.36, + 0.589 + ], + "angle": 0, + "content": "3.3 Word-level SiMT policies" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.596, + 0.489, + 0.691 + ], + "angle": 0, + "content": "The proposed word-level policy restricts a SiMT policy's transition from READ to WRITE or vice versa to occur exclusively at the boundaries of words. Any token-level policy can be transformed to operate at the word-level by following the conversion process we outline below." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.694, + 0.489, + 0.821 + ], + "angle": 0, + "content": "Concretely, we ensure that a word-level policy does not write a target token in the middle of reading a sequence of source tokens that make up a word. To accomplish this word-level READ, we delay \\( g_{i} \\) until it reaches the nearest source word boundary. We formally define \\( r_{i} \\) that has a refined value of \\( g_{i} \\) based on the word boundaries in \\( \\mathbf{x} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.839, + 0.488, + 0.856 + ], + "angle": 0, + "content": "\\[\nr _ {i} := \\min \\{j | j \\geq g _ {i} \\wedge j \\in B _ {S} \\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.871, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Here, \\( B_{S} \\) denotes the indices of the source words' last tokens. Substituting \\( r_i \\) for \\( g_i \\) as a policy transforms it into another policy that upholds the same" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.083, + 0.696, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.205, + 0.696, + 0.217 + ], + "angle": 0, + "content": "Token-level unidirectional encoding" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.082, + 0.882, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.205, + 0.88, + 0.217 + ], + "angle": 0, + "content": "Intra-word bidirectional encoding" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.227, + 0.885, + 0.284 + ], + "angle": 0, + "content": "Figure 3: Comparison of masking in the token-level unidirectional attention (left) and the intra-word bidirectional encoding (right). Word boundaries are represented by vertical/horizontal bars on each axis." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.311, + 0.884, + 0.358 + ], + "angle": 0, + "content": "decision-making criterion while ensuring uninterrupted reading of an entire word when initiating the reading of a token." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.36, + 0.884, + 0.456 + ], + "angle": 0, + "content": "Similarly to the word-level READ, we design a word-level WRITE to balance READ and WRITE actions throughout the translation. To achieve this, we can modify \\( r_i \\) such that it writes until it produces a token that ends with an end-of-word symbol. We define \\( w_i \\) that satisfies this as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.474, + 0.884, + 0.491 + ], + "angle": 0, + "content": "\\[\nb _ {i} := \\min \\{j | j \\geq i \\wedge j \\in B _ {T} \\} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.499, + 0.883, + 0.542 + ], + "angle": 0, + "content": "\\[\nw _ {i} := \\left\\{ \\begin{array}{l l} r _ {i}, & \\text {i f} i = 1 \\vee b _ {i - 1} \\neq b _ {i} \\\\ w _ {i - 1}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.548, + 0.884, + 0.789 + ], + "angle": 0, + "content": "where \\( B_{T} \\) denotes the indices of the target words' last tokens and \\( b_{i} \\) represents the index of the final token in the word that includes \\( y_{i} \\). By employing \\( w_{i} \\) in place of \\( r_{i} \\) (or \\( g_{i} \\)), we ensure that the policy consistently composes entire words without interruptions from any READ actions. This approach effectively reduces latency by facilitating faster writing of certain tokens compared to the original policy, thereby compensating for the increased latency resulting from word-level READ operations. Figure 1 provides a visual comparison between word-level and token-level policies in the context of Wait-1, with the word-level policy encompassing both word-level READ and WRITE operations." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.802, + 0.824, + 0.817 + ], + "angle": 0, + "content": "3.4 Intra-word bidirectional encoding" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.823, + 0.884, + 0.919 + ], + "angle": 0, + "content": "Unidirectional encoding in SiMT is vital for managing computational complexity and training efficiency. However, it has an inevitable consequence of weakening the source sequence representations compared to bidirectional encoding. This is an additional factor contributing to the lower translation" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.081, + 0.44, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.081, + 0.885, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.247, + 0.885, + 0.304 + ], + "angle": 0, + "content": "Figure 4: Illustration of LM-fused attention with the word-level Wait-1. Word and model activations processed at a specific stage are highlighted in red. When a source word is received, it is independently encoded by the LM's vocabulary and the in-domain vocabulary. The hidden activations of the word from the LM is then utilized in the encoder. The decoder generates a sequence of tokens for a word by using both the LM and encoder activations." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.329, + 0.487, + 0.361 + ], + "angle": 0, + "content": "quality of SiMT compared to offline models, along with the early translation from partial inputs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.363, + 0.49, + 0.636 + ], + "angle": 0, + "content": "To mitigate this issue, we utilize a technique called intra-word bidirectional encoding. At the word level, this approach involves unidirectional encoding for each word in the input sentence, meaning past words cannot attend to future words. However, at the subword level, subwords within the same word can attend to each other, allowing past subwords to attend to future subwords within the same word. Since READ operates at the word level in word-level policies, this encoding does not require recomputation during each WRITE operation. It only necessitates a single forward pass, similar to token-level unidirectional encoding. However, it can produce a better encoded representation by enabling attention to more tokens. An example masking to enable intra-word bidirectional encoding is depicted in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.653, + 0.456, + 0.684 + ], + "angle": 0, + "content": "3.5 Integration of LM into SiMT through word-level policies" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.693, + 0.489, + 0.805 + ], + "angle": 0, + "content": "In this subsection, we showcase an additional benefit of word-level policies when integrating an LM into a SiMT system. One of the key challenges in this integration is the vocabulary mismatch between the LM and the SiMT model, which hinders ensuring that both models process an equal amount of input prefix at each translation step." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.919 + ], + "angle": 0, + "content": "One possible solution is to use the LM's vocabulary for the SiMT model. However, the LM's training data may not align well with the specific domain targeted by the SiMT system (Wang et al., 2022). This can result in suboptimal vocabulary for the SiMT model compared to a vocabulary obtained from in-domain data (Liu et al., 2021a). An" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.329, + 0.885, + 0.426 + ], + "angle": 0, + "content": "other option is to explore methods to bridge vocabulary gaps (Kim et al., 2019; Sato et al., 2020; Liu et al., 2021a), but they are either validated only in certain transfer learning scenarios or require an additional training phase to train adapters or finetuning the entire LM using pre-training data." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.448, + 0.885, + 0.591 + ], + "angle": 0, + "content": "In this paper, we introduce a method for leveraging an LM in a manner that facilitates the integration of an off-the-shelf LM into a SiMT model, utilizing a word-level policy, regardless of vocabulary mismatches and the internal structure of the SiMT model. Specifically, we employ an LM fused attention for both the encoder and decoder, following the approach outlined in (Zhu et al., 2020), but with two notable modifications." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.614, + 0.885, + 0.918 + ], + "angle": 0, + "content": "Firstly, we replace BERT with a decoder-only auto-regressive LM (Radford et al., 2019; Lin et al., 2022) for unidirectional encoding of the input, aligning with SiMT models for efficient training and inference. Secondly, the attention between the SiMT model and the LM occurs when both models execute a word-level READ for an input word. This ensures they interact only when they process an equal amount of input prefix, naturally resolving the synchronization issue. Additionally, as they align at every word boundary, the SiMT model can operate independently with a vocabulary derived from in-domain data, while the LM continues to use its original vocabulary. Unlike methods targeting specific SiMT models (Indurthi et al., 2022), our approach can benefit any Neural SiMT model with any decoder-only LM. Figure 4 illustrates the proposed integration of the LM with word-level Wait-1." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15620" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.082, + 0.365, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.219, + 0.35, + 0.231 + ], + "angle": 0, + "content": "(a) En-Fr Wait-k Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.082, + 0.621, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.219, + 0.604, + 0.231 + ], + "angle": 0, + "content": "(b) En-Fr MoE Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.082, + 0.882, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.219, + 0.863, + 0.231 + ], + "angle": 0, + "content": "(c) En-Fr ITST Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.232, + 0.363, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.368, + 0.345, + 0.382 + ], + "angle": 0, + "content": "(d) De-En Wait-k Transformer big" + }, + { + "type": "image", + "bbox": [ + 0.376, + 0.232, + 0.622, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.368, + 0.598, + 0.382 + ], + "angle": 0, + "content": "(e) De-En MoE Transformer big" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.232, + 0.882, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.368, + 0.858, + 0.382 + ], + "angle": 0, + "content": "(f) De-En ITST Transformer big" + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.392, + 0.698, + 0.407 + ], + "angle": 0, + "content": "Figure 5: Results of translation quality v.s. word-level AL." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.432, + 0.26, + 0.448 + ], + "angle": 0, + "content": "4 Experiments" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.46, + 0.228, + 0.473 + ], + "angle": 0, + "content": "4.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.482, + 0.489, + 0.547 + ], + "angle": 0, + "content": "We use the following two datasets for experiments: IWSLT 17 English to French (En-Fr) (Cettolo et al., 2017) consists of \\(242\\mathrm{k}\\) pairs split into \\(233\\mathrm{k}\\) training, 890 validation, and 8,597 test pairs." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.549, + 0.489, + 0.613 + ], + "angle": 0, + "content": "WMT 15 German to English (De-En) comprises 4.5 million pairs. We use newstest2013 (3,000 pairs) for validation and newstest2015 (2,169 pairs) for testing." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.628, + 0.332, + 0.643 + ], + "angle": 0, + "content": "4.2 Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.651, + 0.489, + 0.698 + ], + "angle": 0, + "content": "Our methods are evaluated on the following systems, all of which are based on Transformer (Vaswani et al., 2017) with unidirectional encoding." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.7, + 0.489, + 0.812 + ], + "angle": 0, + "content": "Wait-k (Ma et al., 2019a): A model operating under Wait-k policy. A single model is trained for all k values by randomly sampling k during training (Elbayad et al., 2020). For the word-level Wait-k policy, we define it as reading the first k words and then alternating between reading one source word and writing one target word." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.813, + 0.489, + 0.845 + ], + "angle": 0, + "content": "MoE Wait-k (Zhang and Feng, 2021): A Mixture-of-Experts model initially trained with" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.432, + 0.884, + 0.481 + ], + "angle": 0, + "content": "fixed Experts weights and then fine-tuned with dynamic Experts weights. The word-level Wait-k policy of different k is applied to each expert." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.483, + 0.884, + 0.659 + ], + "angle": 0, + "content": "ITST (Zhang and Feng, 2022b): A SoTA SiMT model equipped with Information-Transport-based policy that quantifies information weights from each source to the current target token. To implement word-level ITST, we convert the number of source tokens required for the first target token of each target word into the corresponding number of source words using Equation 2. Once the required number of source words is read, we complete the translation of the word. Additionally, we calculate the latency cost at the word level." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "We compare each system by training both token-level and word-level models, with and without an LM. For models with an LM, we use XGLM-564M (Lin et al., 2022) and employ the two-stage training in which we first train the SiMT model without the LM and initialize the encoder and decoder from the LM-fused model with the pre-trained weights (Zhu et al., 2020). We also tested the single stage training where all trainable parameters are trained jointly with the LM from scratch. The difference of these strategies are discussed in Section 5.3. We tokenized and encoded the input using sentencepiece (Kudo and Richardson, 2018) and applied BPE with a vocabulary size of \\(32k\\). We use sacreBLEU for BLEU calculation (Post, 2018). For models with token-level policies, we trained" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.857, + 0.489, + 0.919 + ], + "angle": 0, + "content": "3Technically, the word-level policy derived from the token-level Wait-k through the conversion process in Section 3.3 can wait between 1 and \\(k\\) tokens, depending on the input encoding. Therefore, it is not equivalent to the word-level Wait-k policy we define here, which always waits for \\(\\mathbf{k}\\) words." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "15621" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.084, + 0.49, + 0.149 + ], + "angle": 0, + "content": "models with the official implementations\\(^{4 5 6}\\), and implemented word-level policies based on these implementations. More training details are described in A.2." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.161, + 0.262, + 0.175 + ], + "angle": 0, + "content": "4.3 Main results" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.182, + 0.49, + 0.406 + ], + "angle": 0, + "content": "The performance of each system is compared in Figure 5. Notably, when measuring latency at the word level, the word-level policy proves to be highly effective for all three systems across different latency levels and datasets, resulting in superior performance. The only exception is observed in the En-Fr ITST models that demonstrate similar levels of performance. The incorporation of an LM using the proposed LM-fused attention further enhances the performance for all word-level configurations. This observation highlights the suitability of word-level policies for the LM-fused attention approach and underscores the effectiveness of leveraging an LM to enhance SiMT systems." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.408, + 0.49, + 0.519 + ], + "angle": 0, + "content": "Notably, as depicted in Figure 11, the word-level policies also outperform or compete with the token-level policies in token-level latency. This can be attributed to the enhanced token representations under the word-level policy, thanks to the contextual information provided by all other tokens belonging to the same word for each token." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.533, + 0.224, + 0.549 + ], + "angle": 0, + "content": "5 Analysis" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.559, + 0.489, + 0.64 + ], + "angle": 0, + "content": "To validate the effectiveness of word-level policies from multiple angles, we conduct several analyses on various settings. All the experiments were conducted on WMT De.En with transformer-big unless specified otherwise." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.651, + 0.277, + 0.667 + ], + "angle": 0, + "content": "5.1 Ablation study" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.672, + 0.43, + 0.703 + ], + "angle": 0, + "content": "5.1.1 Effects of Word-level READ and WRITE" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.708, + 0.49, + 0.788 + ], + "angle": 0, + "content": "To gain insights into the functionality of word-level READ and WRITE actions, we trained Wait-k models with various policy settings and conducted a performance comparison. Specifically, we examined models with the following policy settings:" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.789, + 0.397, + 0.803 + ], + "angle": 0, + "content": "WW: word-level READ and WRITE." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.805, + 0.478, + 0.819 + ], + "angle": 0, + "content": "TW: token-level READ and word-level WRITE." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.822, + 0.475, + 0.835 + ], + "angle": 0, + "content": "WT: word-level READ and token-level WRITE." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.838, + 0.486, + 0.87 + ], + "angle": 0, + "content": "TkTk: a simpler baseline policy which involves alternating reading k source tokens and writing k" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.087, + 0.689, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.217, + 0.613, + 0.229 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.087, + 0.875, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.217, + 0.8, + 0.229 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.24, + 0.884, + 0.299 + ], + "angle": 0, + "content": "Figure 6: Ablation studies for word-level policies. (a): Comparison of word-level Wait-k policies with different policies. (b): Comparison of word-level Wait-k models with and without the intra-word bidirectional encoding." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.324, + 0.884, + 0.34 + ], + "angle": 0, + "content": "target tokens without considering word boundaries." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.341, + 0.884, + 0.533 + ], + "angle": 0, + "content": "The results are presented in Figure 6 (a). The word-level policy (WW) consistently outperforms TW across all latency settings. This is attributed to its imbalance between the number of source and target prefixes processed in each step. Additionally, WT achieves a minimal AL of approximately 3.6, indicating that it is not well-suited for scenarios that require low latency. Lastly, TkTk shows significantly worse performance than WW, suggesting that reading or writing a few consecutive tokens without considering semantic boundaries offers no benefits, unlike word-level policies." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.544, + 0.838, + 0.576 + ], + "angle": 0, + "content": "5.1.2 Effects of intra-word bidirectional encoding" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.58, + 0.884, + 0.659 + ], + "angle": 0, + "content": "In order to assess the impact of the proposed intraword bidirectional encoding, we trained word-level Wait-K models with and without it and compared the accuracy of the two models across different AL settings. The results are presented in Figure 6 (b)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.661, + 0.884, + 0.773 + ], + "angle": 0, + "content": "Remarkably, the model equipped with the intraword bidirectional encoding consistently achieved higher BLEU scores compared to the model without it, across all tested latency settings. This provides strong evidence of the effectiveness of the intra-word bidirectional encoding in enhancing SiMT performance." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.785, + 0.882, + 0.8 + ], + "angle": 0, + "content": "5.2 Effectiveness of word-level policies for LM" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.807, + 0.884, + 0.869 + ], + "angle": 0, + "content": "In this subsection, we aim to explore the significance of word-level policies in leveraging LM for SiMT. We compare different configurations based on three factors:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.871, + 0.882, + 0.903 + ], + "angle": 0, + "content": "Word vs. Token: The policy type that the model operates with." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.904, + 0.882, + 0.918 + ], + "angle": 0, + "content": "In-domain vocab vs. LM vocab: Whether the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.879, + 0.466, + 0.892 + ], + "angle": 0, + "content": "4Efficient Wait-k: https://github.com/elbayadm/attn2d" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.892, + 0.453, + 0.905 + ], + "angle": 0, + "content": "\\( {}^{5} \\) MoE Wait-k: https://github.com/ictnlp/MoE-Waitk" + }, + { + "type": "page_footnote", + "bbox": [ + 0.137, + 0.905, + 0.37, + 0.918 + ], + "angle": 0, + "content": "\\(^{6}\\)ITST: https://github.com/ictnlp/ITST" + }, + { + "type": "list", + "bbox": [ + 0.136, + 0.879, + 0.466, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15622" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.088, + 0.456, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.25, + 0.49, + 0.281 + ], + "angle": 0, + "content": "Figure 7: Comparison of models with different LM integration." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.306, + 0.49, + 0.385 + ], + "angle": 0, + "content": "model uses an in-domain vocabulary obtained from the in-domain training data or uses the LM's vocabulary for the source language. Note that the use of \"in-domain vocab\" is specific to the \"Word\" configuration due to the vocabulary mismatch." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.386, + 0.49, + 0.466 + ], + "angle": 0, + "content": "LM-fused attention vs. LM embed: Whether the model incorporates an LM using the LM-fused attention or replacing the embedding layer of the encoder with the LM's embedding (Xu et al., 2021). The latter approach uses \"LM vocab\" by design." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.468, + 0.49, + 0.693 + ], + "angle": 0, + "content": "Figure 7 showcases the results. The models with word-level policies consistently outperform those with token-level policies by a significant margin in both LM-fused attention and LM embedding settings, underscoring the importance of word-level policy adoption for effective LM integration. The top-performing configuration is the proposed LMAtn-In-domain vocab-Word, demonstrating that the highest translation accuracy is achieved when the SiMT model operates with an in-domain vocabulary. Additionally, it is evident that the \"LM embed\" approach performs notably worse than the proposed LM-fused attention, further affirming the latter's superiority." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.705, + 0.489, + 0.738 + ], + "angle": 0, + "content": "5.3 Effects of LM-fused attention with various LMs and training configurations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.919 + ], + "angle": 0, + "content": "To assess the effectiveness and broad applicability of our proposed LM integration, we conducted experiments on the IWSLT17 En-Fr dataset with two decoder-only LMs of different sizes: the 137M parameter GPT-2 model (Radford et al., 2018) and the XGLM-564M model. Additionally, we explore the option of training models in the single training stage instead of the two-stage training. The results, presented in Figure 8, demonstrate that GPT-2 model also exhibits improved performance with LM-fused attention, although their impact is" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.691, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.217, + 0.613, + 0.229 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.087, + 0.877, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.217, + 0.8, + 0.229 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.241, + 0.885, + 0.298 + ], + "angle": 0, + "content": "Figure 8: Translation accuracy comparison of LM-fused attention models with different training configurations. (a): En-Fr Transformer small (b) De-En Transformer big" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.323, + 0.885, + 0.532 + ], + "angle": 0, + "content": "naturally less pronounced compared to XGLM due to the difference in model size. Moreover, although models trained using the single-stage training generally exhibit lower performance compared to those trained using the two-stage training, they still outperform models without the LM for most configurations. This indicates that the LM-fused attention is applicable to various types of LMs and remains effective even when using the single-stage training strategy. This flexibility allows users to choose a training approach and model configuration that aligns best with their desired accuracy goals and computational constraints." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.543, + 0.762, + 0.559 + ], + "angle": 0, + "content": "5.4 Policy quality comparison" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.564, + 0.883, + 0.675 + ], + "angle": 0, + "content": "To assess the accuracy of policies in determining when to read or write, we adopt the methodology of estimating the quality of a policy in prior research (Zhang and Feng, 2022b,c; Guo et al., 2022). We measure the quality of a policy by analyzing the proportion of aligned source words received before translating on RWTH De-En alignment dataset7." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.677, + 0.884, + 0.886 + ], + "angle": 0, + "content": "To ensure accurate word-level alignment calculation, we consider an aligned source word is read before writing a ground truth (GT) target word if the last token of the source word is read before the first token of the target word is written. The results of this analysis are presented in Figure 9. It is observed that word-level policies, both for ITST and Wait-k, exhibit better alignments across most latency settings. This suggests that word-level policies contribute to revising premature WRITE actions by guiding the model to read the remaining tokens of the aligned word, without negatively impacting the model's latency." + }, + { + "type": "page_footnote", + "bbox": [ + 0.51, + 0.892, + 0.741, + 0.919 + ], + "angle": 0, + "content": "7https://www-i6.informatik.rwth-aachen.de/goldAlignment/" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.087, + 0.45, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.25, + 0.486, + 0.279 + ], + "angle": 0, + "content": "Figure 9: The percentage of aligned source words read before the translation of the target words started." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.304, + 0.254, + 0.318 + ], + "angle": 0, + "content": "6 Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.329, + 0.49, + 0.587 + ], + "angle": 0, + "content": "This paper explores the potential benefits of word-level operations in SiMT systems. We propose word-level latency calculation to ensure fair and accurate latency comparisons. We introduce a conversion process that transforms token-level policies into word-level policies, enabling the processing of multiple subwords that form a word within a single READ or WRITE action. Additionally, we propose the integration of LM-fused attention, which combines an autoregressive LM model into SiMT models with word-level policies. Experimental results demonstrate the superiority of word-level policies compared to token-level policies, as well as the effectiveness of the LM integration. Our findings highlight the crucial role of word-level policies in the integration process." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.598, + 0.249, + 0.613 + ], + "angle": 0, + "content": "7 Limitations" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.623, + 0.49, + 0.816 + ], + "angle": 0, + "content": "While the proposed word-level policy implementation is widely applicable to most existing SiMT systems, it is important to note that systems utilizing languages with a writing style that lacks spaces or other delimiters between words or sentences (e.g., Chinese) are unable to derive benefits from this approach. Furthermore, it is important to consider that while the proposed LM-fused attention proves effective in enhancing translation quality across all latency levels, integrating a large LM may necessitate a faster compute capability to fulfill the low-latency demands of the SiMT task." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.844, + 0.214, + 0.858 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.866, + 0.488, + 0.918 + ], + "angle": 0, + "content": "Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. 2019. Monotonic infinite lookback attention for simultaneous machine" + }, + { + "type": "ref_text", + "bbox": [ + 0.527, + 0.086, + 0.883, + 0.14 + ], + "angle": 0, + "content": "translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1313-1323, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.885, + 0.241 + ], + "angle": 0, + "content": "Ozan Caglayan, Julia Ive, Veneta Haralampieva, Pranava Madhyastha, Loic Barrault, and Lucia Specia. 2020. Simultaneous machine translation with visual context. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2350-2361, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.252, + 0.885, + 0.356 + ], + "angle": 0, + "content": "Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuhito Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the IWSLT 2017 evaluation campaign. In Proceedings of the 14th International Conference on Spoken Language Translation, pages 2-14, Tokyo, Japan. International Workshop on Spoken Language Translation." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.367, + 0.884, + 0.407 + ], + "angle": 0, + "content": "Chung-Cheng Chiu and Colin Raffel. 2017. Monotonic chunkwise attention. arXiv preprint arXiv:1712.05382." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.418, + 0.884, + 0.459 + ], + "angle": 0, + "content": "Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? arXiv preprint arXiv:1606.02012." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.469, + 0.884, + 0.561 + ], + "angle": 0, + "content": "Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3440-3453, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.571, + 0.883, + 0.689 + ], + "angle": 0, + "content": "Fahim Dalvi, Nadir Durrani, Hassan Sajjad, and Stephan Vogel. 2018. Incremental decoding and training methods for simultaneous translation in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 493-499, New Orleans, Louisiana. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.699, + 0.884, + 0.818 + ], + "angle": 0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.828, + 0.883, + 0.88 + ], + "angle": 0, + "content": "Maha Elbayad, Laurent Besacier, and Jakob Verbeek. 2020. Efficient Wait-k Models for Simultaneous Machine Translation. In Proc. Interspeech 2020, pages 1461-1465." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.891, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor O.K. Li. 2017. Learning to translate in real-time" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.152 + ], + "angle": 0, + "content": "with neural machine translation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1053-1062, Valencia, Spain. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.164, + 0.49, + 0.256 + ], + "angle": 0, + "content": "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2022. Turning fixed to adaptive: Integrating post-evaluation into simultaneous machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2264–2278, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.268, + 0.49, + 0.36 + ], + "angle": 0, + "content": "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2023. Learning optimal policy for simultaneous machine translation via binary search. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2318-2333, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.372, + 0.49, + 0.49 + ], + "angle": 0, + "content": "Sathish Reddy Indurthi, Mohd Abbas Zaidi, Beomseok Lee, Nikhil Kumar Lakumarapu, and Sangha Kim. 2022. Language model augmented monotonic attention for simultaneous translation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 38-45, Seattle, United States. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.502, + 0.49, + 0.594 + ], + "angle": 0, + "content": "Javier Iranzo Sanchez, Jorge Civera, and Alfons Juan-Ciscar. 2022. From simultaneous to streaming machine translation by leveraging streaming history. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6972-6985, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.606, + 0.49, + 0.711 + ], + "angle": 0, + "content": "Patrick Kahardipraja, Brielen Madureira, and David Schlangen. 2021. Towards incremental transformers: An empirical analysis of transformer models for incremental NLU. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1178-1189, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.722, + 0.49, + 0.815 + ], + "angle": 0, + "content": "Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1246–1257, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.125 + ], + "angle": 0, + "content": "Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint arXiv:1901.07291." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.134, + 0.885, + 0.252 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.26, + 0.885, + 0.418 + ], + "angle": 0, + "content": "Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettle-moyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, and Xian Li. 2022. Few-shot learning with multilingual generative language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9019-9052, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.426, + 0.885, + 0.557 + ], + "angle": 0, + "content": "Xin Liu, Baosong Yang, Dayiheng Liu, Haibo Zhang, Weihua Luo, Min Zhang, Haiying Zhang, and Jinsong Su. 2021a. Bridging subword gaps in pretrainfinetune paradigm for natural language generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6001-6011, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.565, + 0.883, + 0.657 + ], + "angle": 0, + "content": "Xuebo Liu, Longyue Wang, Derek F. Wong, Liang Ding, Lidia S. Chao, Shuming Shi, and Zhaopeng Tu. 2021b. On the copying behaviors of pre-training for neural machine translation. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 4265-4275, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.665, + 0.883, + 0.743 + ], + "angle": 0, + "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.752, + 0.883, + 0.819 + ], + "angle": 0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.826, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, Hua Wu, and Haifeng Wang. 2019a. STACL: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework. In Proceedings of the 57th Annual Meeting of the Association for" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.885, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "15625" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.086, + 0.49, + 0.126 + ], + "angle": 0, + "content": "Computational Linguistics, pages 3025-3036, Florence, Italy. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.135, + 0.488, + 0.174 + ], + "angle": 0, + "content": "Xutai Ma, Juan Pino, James Cross, Liezl Puzon, and Jiatao Gu. 2019b. Monotonic multihead attention. arXiv preprint arXiv:1909.12406." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.182, + 0.488, + 0.248 + ], + "angle": 0, + "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.256, + 0.488, + 0.297 + ], + "angle": 0, + "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.304, + 0.488, + 0.357 + ], + "angle": 0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.365, + 0.488, + 0.456 + ], + "angle": 0, + "content": "Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, and Douglas Eck. 2017. Online and linear-time attention by enforcing monotonic alignments. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 2837-2846. PMLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.465, + 0.488, + 0.558 + ], + "angle": 0, + "content": "Shoetsu Sato, Jin Sakuma, Naoki Yoshinaga, Masashi Toyoda, and Masaru Kitsuregawa. 2020. Vocabulary adaptation for domain adaptation in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4269-4279, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.565, + 0.488, + 0.645 + ], + "angle": 0, + "content": "Felix Schneider and Alexander Waibel. 2020. Towards stream translation: Adaptive computation time for simultaneous machine translation. In Proceedings of the 17th International Conference on Spoken Language Translation, pages 228-236, Online. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.652, + 0.488, + 0.745 + ], + "angle": 0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.752, + 0.488, + 0.819 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.826, + 0.488, + 0.919 + ], + "angle": 0, + "content": "Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. 2023. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2334-2349, Toronto, Canada. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.086, + 0.883, + 0.191 + ], + "angle": 0, + "content": "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2591-2600, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.203, + 0.883, + 0.269 + ], + "angle": 0, + "content": "Rongxiang Weng, Heng Yu, Shujian Huang, Shanbo Cheng, and Weihua Luo. 2020. Acquiring knowledge from pre-trained model to neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9266-9273." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.281, + 0.883, + 0.347 + ], + "angle": 0, + "content": "Rongxiang Weng, Heng Yu, Weihua Luo, and Min Zhang. 2022. Deep fusing pre-trained models into neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11468-11476." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.359, + 0.883, + 0.464 + ], + "angle": 0, + "content": "Haoran Xu, Benjamin Van Durme, and Kenton Murray. 2021. BERT, mBERT, or BiBERT? a study on contextualized embeddings for neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6663-6675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.476, + 0.883, + 0.553 + ], + "angle": 0, + "content": "Jiacheng Yang, Mingxuan Wang, Hao Zhou, Chengqi Zhao, Weinan Zhang, Yong Yu, and Lei Li. 2020. Towards making the most of bert in neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9378-9385." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.567, + 0.883, + 0.658 + ], + "angle": 0, + "content": "Shaolei Zhang and Yang Feng. 2021. Universal simultaneous machine translation with mixture-of-experts wait-k policy. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7306-7317, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.671, + 0.883, + 0.737 + ], + "angle": 0, + "content": "Shaolei Zhang and Yang Feng. 2022a. Gaussian multihead attention for simultaneous machine translation. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 3019–3030, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.749, + 0.883, + 0.827 + ], + "angle": 0, + "content": "Shaolei Zhang and Yang Feng. 2022b. Information-transport-based policy for simultaneous translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 992-1013, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.84, + 0.883, + 0.919 + ], + "angle": 0, + "content": "Shaolei Zhang and Yang Feng. 2022c. Modeling dual read/write paths for simultaneous machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2461-2477, Dublin, Ireland. Association for Computational Linguistics." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.086, + 0.883, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.525, + 0.941 + ], + "angle": 0, + "content": "15626" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.115, + 0.086, + 0.489, + 0.126 + ], + "angle": 0, + "content": "Shaolei Zhang and Yang Feng. 2023. Hidden markov transformer for simultaneous machine translation. arXiv preprint arXiv:2303.00257." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.134, + 0.49, + 0.2 + ], + "angle": 0, + "content": "Shaolei Zhang, Yang Feng, and Liangyou Li. 2021. Future-guided incremental transformer for simultaneous translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14428-14436." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.209, + 0.49, + 0.302 + ], + "angle": 0, + "content": "Shaolei Zhang, Shoutao Guo, and Yang Feng. 2022. Wait-info policy: Balancing source and target at information level for simultaneous machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2249-2263, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.31, + 0.49, + 0.403 + ], + "angle": 0, + "content": "Baigong Zheng, Kaibo Liu, Renjie Zheng, Mingbo Ma, Hairong Liu, and Liang Huang. 2020. Simultaneous translation policies: From fixed to adaptive. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2847-2853, Online. Association for Computational Linguistics." + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.41, + 0.49, + 0.464 + ], + "angle": 0, + "content": "Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. 2020. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.473, + 0.239, + 0.489 + ], + "angle": 0, + "content": "A Appendix" + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.498, + 0.486, + 0.528 + ], + "angle": 0, + "content": "A.1 Token-level latency evaluation in previous works" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.535, + 0.49, + 0.856 + ], + "angle": 0, + "content": "Many previous works in the field of SiMT have employed token-level latency calculation as a fundamental metric for evaluating system performance. In the first group of papers, which includes (Arivazhagan et al., 2019; Ma et al., 2019b; Schneider and Waibel, 2020; Zheng et al., 2020; Zhang and Feng, 2022b), researchers explicitly stated their use of token-level latency calculation or incorporated token-level scores in their analyses. On the other hand, the second group, comprising (Zhang and Feng, 2022b; Elbayad et al., 2020; Zhang and Feng, 2022c, 2021, 2022a; Zhang et al., 2022; Guo et al., 2022; Zhang and Feng, 2023; Guo et al., 2023; Caglayan et al., 2020; Wang et al., 2023), can be identified by their publicly available official implementations, which incorporate token-level latency evaluation code. This underscores the prevalence of token-level latency calculation in assessing the effectiveness of SiMT systems in the cutting-edge SiMT research works." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.867, + 0.332, + 0.882 + ], + "angle": 0, + "content": "A.2 More training details" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.888, + 0.49, + 0.919 + ], + "angle": 0, + "content": "For Transformer model configurations, we follow the Efficient Wait-k (Elbayad et al., 2020)'s set" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.084, + 0.882, + 0.115 + ], + "angle": 0, + "content": "tings.\\(^{8}\\) The details of hyperparameter settings for each model is in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.117, + 0.885, + 0.26 + ], + "angle": 0, + "content": "To ensure optimal performance of the ITST models, we conducted multiple training runs for each model configuration and dataset. We observed performance fluctuations across the training runs. To select the best-performing models for both token-level and word-level policies, we repeated the training process three times and selected the checkpoint with the lowest validation loss for each configuration." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.262, + 0.884, + 0.31 + ], + "angle": 0, + "content": "For LM integrated models, the intra-word bidirectional encoding was not applied to the LM to prevent the need for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.321, + 0.763, + 0.336 + ], + "angle": 0, + "content": "A.3 Examples and discussions" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.341, + 0.885, + 0.647 + ], + "angle": 0, + "content": "In Table 2, an illustrative case is presented to demonstrate the distinction between a token-level policy (token-level Wait-1) and a word-level policy (word-level Wait-1). At Step 2, the token-level model predicts the target token “B” after processing the subword “B” for the word “Beine” (which means “leg” in German). Subsequently, it fails to recover from this incorrect prediction and continues by predicting “ody_” as the next token, even after processing the remaining token “eine_”. In contrast, the word-level model accurately predicts “legs” (encoded as “leg s_”) after processing the complete word ‘Beine’ (encoded as ‘B eine_’). Furthermore, the token model also makes erroneous predictions for all steps when processing the word ‘blutüberströmmt.’ (encoded as ‘bl ut über ström t_’, meaning “covered in blood” in German), while the word-level model accurately predicts a correct word in a single step." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.648, + 0.884, + 0.776 + ], + "angle": 0, + "content": "Another case can be observed in Table 3, where the token-level model initiates the translation before fully reading the word \"Irgendetwas\" (encoded as \"Irgen det was\"). As a consequence, it produces an incorrect translation. On the other hand, the word-level model accurately translates the sentence by processing the complete word before making any predictions." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.787, + 0.724, + 0.801 + ], + "angle": 0, + "content": "A.4 Table of main results" + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.808, + 0.741, + 0.823 + ], + "angle": 0, + "content": "A.5 Graphs of main results" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.904, + 0.757, + 0.919 + ], + "angle": 0, + "content": "\\(^{8}\\)https://github.com/elbayadm/attn2d" + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15627" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.172, + 0.106, + 0.827, + 0.368 + ], + "angle": 0, + "content": "
Hyperparametertransformer-smalltransformer-basetransformer-big
attention heads4816
embedding dime2565121024
ffn embedding dim102420484096
dropout0.30.30.3
encoder layers666
decoder layers666
lr5e-45e-45e-4
lr schedulerinverse sqrtinverse sqrtinverse sqrt
optimizerAdam (0.9, 0.98)Adam (0.9, 0.98)Adam (0.9, 0.98)
clip-norm000
warmup-updates400040004000
warmup-init-lr1e-71e-71e-7
weight decay000
label smoothing0.10.10.1
max tokens327685242881048576
" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.378, + 0.684, + 0.394 + ], + "angle": 0, + "content": "Table 1: Hyperparameters of each Transformer model." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.449, + 0.859, + 0.611 + ], + "angle": 0, + "content": "
Src:Meine_B eine_waren_bl ut über ström t.
Ref:My leg s_were Covered_in_blood.
StepStreaminig InputToken Wait-1 OutputWord Wait-1 Output
1MeineMyMy
2Meine_BB
3Meine_B eineodyleg s
4Meine_B eine_warenwaswere
5Meine_B eine_waren_blblue
6Meine_B eine_waren_bl utand
7Meine_B eine_waren_bl ut übermy
8Meine_B eine_waren_bl ut über strömleg
9Meine_B eine_waren_bl ut über ström ts
10Meine_B eine_waren_bl ut über ström t.were_blood - ri dden.blood dy.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.619, + 0.884, + 0.647 + ], + "angle": 0, + "content": "Table 2: An example case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.56, word-AL: 1.69) and Word-level Wait-1 (token-AL: 1.44, word-AL: 1.0)." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.705, + 0.86, + 0.853 + ], + "angle": 0, + "content": "
Src:Ir gen det was lag in der Luft .
Ref:Some thing was up .
StepStreaming InputToken Wait-1 OutputWord Wait-1 Output
1IrIr
2Ir gengen
3Ir gen detgen
4Ir gen det wasgenSome thing_
5Ir gen det was laggenlay_
6Ir gen det was lag insin_
7Ir gen det was lag in der,the_
8Ir gen det was lag in der Luftin
9Ir gen det was lag in der Luft .fact , was in the air .air .
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.862, + 0.884, + 0.891 + ], + "angle": 0, + "content": "Table 3: An Eexample case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.69, word-AL: 4.25) and Word-level Wait-1 (token-AL: 2.38, word-AL: 1.0)." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15628" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.125, + 0.365, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.262, + 0.348, + 0.275 + ], + "angle": 0, + "content": "(a) De-En Wait-k Transformer base" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.125, + 0.621, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.262, + 0.602, + 0.275 + ], + "angle": 0, + "content": "(b) De-En MoE Transformer base" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.125, + 0.88, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.262, + 0.862, + 0.275 + ], + "angle": 0, + "content": "(c) De-En ITST Transformer base" + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.287, + 0.796, + 0.302 + ], + "angle": 0, + "content": "Figure 10: Results of translation quality v.s. word-level AL, De-En transformer-base." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.397, + 0.363, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.534, + 0.349, + 0.546 + ], + "angle": 0, + "content": "(a) En-Fr Wait-k Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.397, + 0.622, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.534, + 0.603, + 0.546 + ], + "angle": 0, + "content": "(b) En-Fr MoE Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.397, + 0.88, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.534, + 0.862, + 0.546 + ], + "angle": 0, + "content": "(c) En-Fr ITST Transformer small" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.547, + 0.362, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.684, + 0.349, + 0.695 + ], + "angle": 0, + "content": "(d) De-En Wait-k Transformer base" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.547, + 0.622, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.684, + 0.602, + 0.695 + ], + "angle": 0, + "content": "(e) De-En MoE Transformer base" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.547, + 0.88, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.684, + 0.861, + 0.695 + ], + "angle": 0, + "content": "(f) De-En ITST Transformer base" + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.696, + 0.362, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.833, + 0.345, + 0.846 + ], + "angle": 0, + "content": "(g) De-En Wait-k Transformer big" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.696, + 0.622, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.833, + 0.598, + 0.846 + ], + "angle": 0, + "content": "(h) De-En MoE Transformer big" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.696, + 0.88, + 0.828 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.833, + 0.857, + 0.846 + ], + "angle": 0, + "content": "(i) De-En ITST Transformer big" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.857, + 0.703, + 0.872 + ], + "angle": 0, + "content": "Figure 11: Results of translation quality v.s. token-level AL." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15629" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.274, + 0.446, + 0.684 + ], + "angle": 0, + "content": "
Token-level Wait-k
ktoken ALword ALBLEU
11.951.9132.32
33.833.4338.90
55.674.8240.68
77.516.2340.95
99.287.5841.06
1110.918.8241.14
1312.399.9541.12
Word-level Wait-k
ktoken ALword ALBLEU
12.241.7432.43
34.183.3239.36
56.555.1541.09
78.856.9441.48
910.928.5741.57
1112.759.9841.58
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.181.7433.47
34.233.3740.83
56.585.1942.78
78.866.9743.00
910.948.5843.08
1112.7610.0043.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.693, + 0.487, + 0.721 + ], + "angle": 0, + "content": "Table 4: Main results of transformer-small Waik-k models on IWLST17 En.Fr dataset." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.274, + 0.84, + 0.683 + ], + "angle": 0, + "content": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.901.8832.20
33.833.4238.92
55.694.8440.51
77.536.2540.83
99.317.6040.92
1110.948.8540.95
1312.429.9740.96
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.011.6032.01
34.183.3239.28
56.575.1740.77
78.866.9541.03
910.948.5841.11
1112.769.9941.10
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.171.7233.01
34.223.3740.46
56.565.1742.35
78.846.9542.61
910.918.5742.66
1112.749.9942.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.693, + 0.882, + 0.721 + ], + "angle": 0, + "content": "Table 5: Main results of transformer-small MoE Waik-k models on IWSLT17 En.Fr dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.527, + 0.941 + ], + "angle": 0, + "content": "15630" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.185, + 0.451, + 0.771 + ], + "angle": 0, + "content": "
Token-level ITST
δtoken ALword ALBLEU
0.23.333.0537.81
0.33.593.2538.64
0.44.143.7039.40
0.55.184.5940.19
0.66.715.9240.84
0.78.627.6341.27
0.759.568.4541.34
0.810.729.4141.42
0.8512.1110.5341.42
0.913.8411.8041.42
Word-level ITST
δtoken ALword ALBLEU
0.23.272.6437.25
0.33.462.8037.88
0.43.913.2138.69
0.54.673.9139.57
0.65.814.9640.44
0.77.296.2840.94
0.758.217.0541.13
0.89.418.0241.36
0.8511.199.3741.50
0.913.3611.0041.53
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.252.6138.30
0.33.302.6538.75
0.43.482.8139.61
0.53.943.2340.50
0.64.804.0441.35
0.76.025.1941.88
0.756.765.8642.09
0.87.726.6742.30
0.858.997.6842.56
0.910.909.1642.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.782, + 0.486, + 0.81 + ], + "angle": 0, + "content": "Table 6: Main results of transformer-small ITST models on IWSLT17 En.Fr dataset." + }, + { + "type": "table", + "bbox": [ + 0.553, + 0.258, + 0.84, + 0.7 + ], + "angle": 0, + "content": "
Token-level Wait-k
ktoken ALword ALBLEU
10.520.7517.63
32.161.9123.24
53.923.0526.14
75.774.2327.46
97.765.5028.77
119.646.6929.37
1311.497.8629.88
1513.188.9229.89
1714.799.9330.19
Word-level Wait-k
ktoken ALword ALBLEU
12.961.8424.02
35.333.4328.08
58.085.2329.35
710.827.0129.87
913.478.7230.17
1115.8810.2530.38
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.961.8424.71
35.273.3828.49
58.055.2029.95
710.817.0030.56
913.448.6930.74
1115.8510.2430.95
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.709, + 0.882, + 0.737 + ], + "angle": 0, + "content": "Table 7: Main results of transformer-base Waik-k models on WMT15 De.En dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.524, + 0.941 + ], + "angle": 0, + "content": "15631" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.258, + 0.445, + 0.437 + ], + "angle": 0, + "content": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.171.1117.28
32.502.0923.29
54.173.1825.95
76.004.3627.86
97.815.5128.81
119.626.6629.53
1311.427.8029.66
1513.158.8929.81
1714.809.9129.79
" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.438, + 0.445, + 0.568 + ], + "angle": 0, + "content": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
13.121.9123.59
35.293.3927.75
58.035.1929.29
710.796.9830.14
913.388.6530.39
1115.8110.230.6
" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.569, + 0.444, + 0.699 + ], + "angle": 0, + "content": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
13.081.9224.97
35.323.4428.66
58.085.2330.21
710.877.0430.63
913.478.7330.9
1115.8810.2631.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.709, + 0.487, + 0.737 + ], + "angle": 0, + "content": "Table 8: Main results of transformer-base MoE Waik-k models on WMT15 De.En dataset." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.186, + 0.847, + 0.382 + ], + "angle": 0, + "content": "
Token-level ITST
δtoken ALword ALBLEU
0.22.452.0822.19
0.32.892.3824.15
0.43.822.9626.21
0.55.233.9228.19
0.67.285.5029.25
0.79.877.4829.47
0.7511.668.8129.76
0.813.7910.3829.79
0.8516.6012.1129.90
0.919.8113.8529.71
" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.383, + 0.846, + 0.575 + ], + "angle": 0, + "content": "
Word-level ITST
δtoken ALword ALBLEU
0.24.132.6426.81
0.34.192.6927.01
0.44.392.8127.47
0.55.093.3028.04
0.66.434.2828.87
0.78.605.9329.54
0.759.916.9229.54
0.811.548.0929.67
0.8513.839.6129.82
0.916.7511.3829.93
" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.577, + 0.846, + 0.771 + ], + "angle": 0, + "content": "
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.24.152.6427.74
0.34.242.7027.99
0.44.552.9228.44
0.55.333.4929.15
0.66.774.6329.97
0.78.576.0830.49
0.759.516.8030.53
0.810.717.6730.60
0.8512.328.7230.68
0.914.289.9330.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.782, + 0.884, + 0.81 + ], + "angle": 0, + "content": "Table 9: Main results of transformer-base ITST models on WMT15 De.En dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15632" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.258, + 0.445, + 0.438 + ], + "angle": 0, + "content": "
Token-level Wait-k
ktoken ALword ALBLEU
1-0.130.3919.20
31.801.7124.01
53.742.9526.86
75.814.2628.86
97.805.5329.96
119.726.7430.62
1311.517.8730.91
1513.208.9531.16
1714.849.9731.31
" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.438, + 0.445, + 0.568 + ], + "angle": 0, + "content": "
Word-level Wait-k
ktoken ALword ALBLEU
12.551.5924.63
35.113.3028.82
57.955.1630.43
710.776.9931.35
913.378.67831.86
1115.8010.2232.03
" + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.568, + 0.445, + 0.699 + ], + "angle": 0, + "content": "
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.891.7926.21
35.213.3630.16
57.985.1731.36
710.827.0132.29
913.428.7132.58
1115.7410.1632.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.709, + 0.487, + 0.737 + ], + "angle": 0, + "content": "Table 10: Main results of transformer-big Waik-k models on WMT15 De.En dataset." + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.258, + 0.84, + 0.438 + ], + "angle": 0, + "content": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
10.860.9418.44
32.191.9224.07
53.903.0426.70
75.704.1928.47
97.695.4729.59
119.606.6730.36
1311.447.8330.51
1513.148.9130.56
1714.759.9130.65
" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.438, + 0.84, + 0.568 + ], + "angle": 0, + "content": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.861.7624.79
35.243.3728.95
57.985.1530.26
710.776.9731.01
913.378.6631.47
1115.8010.2031.55
" + }, + { + "type": "table", + "bbox": [ + 0.554, + 0.568, + 0.84, + 0.699 + ], + "angle": 0, + "content": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.601.6025.26
35.193.3129.03
58.015.1630.79
710.836.9931.66
913.438.6731.82
1115.8310.2132.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.509, + 0.709, + 0.884, + 0.737 + ], + "angle": 0, + "content": "Table 11: Main results of transformer-big MoE Waik-k models on WMT15 De.En dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.928, + 0.525, + 0.941 + ], + "angle": 0, + "content": "15633" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.151, + 0.185, + 0.451, + 0.771 + ], + "angle": 0, + "content": "
Token-level ITST
δtoken ALword ALBLEU
0.21.491.5523.22
0.32.412.1224.75
0.43.933.0626.67
0.55.994.3928.23
0.69.126.5729.85
0.712.699.3231.02
0.7514.7210.7531.28
0.816.9512.2331.54
0.8519.0613.4931.70
0.921.2614.6931.84
Word-level ITST
δtoken ALword ALBLEU
0.23.692.3526.33
0.33.792.4226.62
0.43.942.5426.62
0.54.693.0227.55
0.66.604.2728.99
0.710.286.7730.41
0.7512.508.3430.95
0.814.229.6531.54
0.8515.9010.8131.56
0.918.1312.2331.61
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.972.5328.12
0.34.002.5528.13
0.44.102.6228.29
0.54.392.8228.92
0.65.373.4629.66
0.77.855.1331.18
0.759.736.4831.77
0.811.908.0432.10
0.8514.049.5732.34
0.916.1410.9732.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.782, + 0.49, + 0.81 + ], + "angle": 0, + "content": "Table 12: Main results of transformer-big ITST models on WMT15 De.En dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.478, + 0.929, + 0.526, + 0.941 + ], + "angle": 0, + "content": "15634" + } + ] +] \ No newline at end of file diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_origin.pdf b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4e7912653b3d8ebbecc69d3439fcc13abbb2e4cf --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/6976f9e4-3ce3-45d6-a81a-1951bc56ac12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e40646e37aff02ccd122b9a8c6bdf4a7e6e8ae3ca4a2a9fe9be7b05b20cb61db +size 1026354 diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/full.md b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b2fab9d169d826ea62a0411a549cc790128947c4 --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/full.md @@ -0,0 +1,469 @@ +# Enhanced Simultaneous Machine Translation with Word-level Policies + +Kang Kim*, Hankyu Cho* + +XL8 Inc. + +{kai, matthew}@x18.ai + +# Abstract + +Recent years have seen remarkable advances in the field of Simultaneous Machine Translation (SiMT) due to the introduction of innovative policies that dictate whether to READ or WRITE at each step of the translation process. However, a common assumption in many existing studies is that operations are carried out at the subword level, even though the standard unit for input and output in most practical scenarios is typically at the word level. This paper demonstrates that policies devised and validated at the subword level are surpassed by those operating at the word level, which process multiple subwords to form a complete word in a single step. Additionally, we suggest a method to boost SiMT models using language models (LMs), wherein the proposed word-level policy plays a vital role in addressing the subword disparity between LMs and SiMT models. Code is available at https://github.com/x18-ai/WordSiMT. + +# 1 Introduction + +Simultaneous Machine Translation (SiMT) commences the translation process while simultaneously receiving the input, making it an effective approach for applications that require minimal latency such as simultaneous interpretation or live broadcast. The development of a novel policy is central to research efforts in SiMT. This policy dictates the translation process by determining whether to execute a READ or WRITE action at each step of the process. + +Neural SiMT models, like offline Neural Machine Translation (NMT) models, commonly employ Byte Pair Encoding (BPE) (Sennrich et al., 2016) or similar techniques to encode an input sentence into a sequence of tokens. Typically, a single READ or WRITE action of a SiMT policy is responsible for handling an encoded token, which may sometimes be a word but often a subword. + +![](images/698d58ed3af2627eea3939c9cb293ed07e64a927e1c670a7484c4fa413657720.jpg) +Figure 1: Illustration of both token-level Wait-1 (red arrow lines) and word-level Wait-1 (green arrow lines) policies. The symbol " $\_$ " at the end of a token indicates a word boundary. + +The development of BPE-based SiMT models and their policies has resulted in researchers focusing on working at the subword level. The performance analysis and implementation of many SiMT systems, to the best of our knowledge, have been carried out on encoded sequences of source and target subwords, rather than on the original source and target sentences1. This has led to two critical issues that need to be addressed. + +The first issue is the lack of a standardized tokenization and encoding scheme, meaning that different implementations may employ varying token sequences to encode identical text. This variability can impact latency evaluation results and complicate score comparisons across different systems. + +The second issue is the missed opportunity to process more source tokens before writing each target token without added latency. For a BPE-based SiMT model, the input must be received on a word-by-word basis to ensure proper encoding of each word into a sequence of subwords. Consequently, when the model encodes a word and performs a READ to process only a subword, it delays the reading of the remaining subwords without any benefit in actual latency², and may adversely impact + +![](images/231157632db099fa480ad5e3f153a799abe77907d7c96e89f011863bd87a35c4.jpg) +Figure 2: An exemplary case depicting the difference between token-level and word-level Wait-k policies and their Average Lagging (AL) scores. The token-level model begins translating in the middle of reading "Tremendously" and fails to recover from the incorrectly translated target prefix. On the other hand, the word-level model processes "Tremendously" in a single READ action and produces a correct translation. + +translation quality by causing the model to rely on incomplete representations extracted from partially-read words. Similarly, performing a WRITE to generate a subword earlier than the remaining subwords does not necessarily reduce latency, as the subword must wait for a complete word to be generated before it can be displayed. + +In this paper, we show that establishing the unit of measuring and operating SiMT systems at the word level, rather than the subword level, offers a viable solution to these issues. Specifically, to tackle the first issue, we propose word-level latency metric calculation for measuring the latency of SiMT systems. This not only enables consistent comparisons between different systems but also provides a more accurate reflection of the actual latency experienced in SiMT applications that display translation results word by word. + +To address the second issue, we illustrate that an existing token-level policy can be transformed into a word-level policy that inherently overcomes the second issue, resulting in improved performance. Word-level policies take into consideration word boundaries and perform a READ or WRITE action to sequentially process a sequence of tokens that form a complete word. This conversion process can be applied to any token-level policies, and our experiments reveal that state-of-the-art fixed and adaptive policies exhibit significantly better performance when transformed into their word-level counterparts. Notably, these word-level policies often outperform token-level policies, even when evaluated using a token-level latency metric, due to their enhanced utilization of input and output tokens. + +Additionally, to boost translation accuracy, we suggest incorporating a pre-trained language model (LM) into a SiMT model, where the word-level policy plays a crucial role as a pivotal component. One of the major hurdles in utilizing an LM for SiMT is the vocabulary mismatch between the SiMT model + +and the LM. The difficulty of handling subword disparities when utilizing an LM for a downstream task is widely acknowledged (Liu et al., 2021a; Wang et al., 2022), and it becomes particularly problematic in SiMT, as inconsistent subwords between the LM and SiMT model make processing the same source or target prefix challenging. Our study demonstrates that our proposed word-level policy effectively tackles this challenge, enabling a successful integration of LMs into SiMT systems. + +# 2 Related Work + +# 2.1 Simultaneous Machine Translation + +SiMT systems that employ a fixed policy utilize a pre-defined sequence of READ and WRITE operations for each source sentence. STATIC-RW (Dalvi et al., 2018) and Wait-k (Ma et al., 2019a) policies first read k source tokens, then alternate between reading and writing a single token. Elbayad et al. (2020) propose the multi-path training of a Wait-k model to train a single model that supports different k values at test time. Zhang et al. (2021) improve Wait-k policy using knowledge distillation from an offline MT model, while Zhang and Feng (2021) suggest a Mixture-of-Experts Wait-k Policy where predictions from multiple k values are combined inside a single model. + +In contrast, research efforts on adaptive policies focus on the development of dynamic decision-making processes for READ/WRITE actions. Cho and Esipova (2016) firstly introduce model-based adaptive criteria for Neural SiMT. Gu et al. (2017) propose to learn a policy by using reinforcement learning. Raffel et al. (2017) introduce Monotonic Attention that ensures monotonic alignment in the attention mechanism. Succeeding works improve it by extending the alignment window (Chiu and Raffel, 2017; Arivazhagan et al., 2019), extending it as monotonic multi-head attention (MMA) (Ma et al., 2019b) or learning transposed policies between the forward and backward models. (Zhang and Feng, + +2022c). Zheng et al. (2020) derive a policy by composing Wait-k models trained with different values of k. Zhang and Feng (2022a) model to predict the alignment between each target token and the source token. Zhang and Feng (2022b) measure accumulated information from source tokens to decide whether to write the current target token. + +Despite significant advancements, the impact of operating policies at the word level has not been thoroughly explored in existing works, which have mainly focused on developing and evaluating systems at the token level. In this paper, we address this gap by demonstrating that implementing various types of policies at the word level consistently outperforms their token-level counterparts. + +# 2.2 Utilizing pre-trained LM for MT + +Since the successful utilization of Transformer-based LMs pre-trained on large text corpora for downstream NLP tasks (Devlin et al., 2019; Liu et al., 2019; Lample and Conneau, 2019), the utilization of these models for MT has become a significant research area. Several studies have demonstrated the effectiveness of incorporating encoder-only LMs into NMT models. Weng et al. (2020); Yang et al. (2020) combine the LM representations with the encoder's representation using a gating mechanism. Zhu et al. (2020) propose attention between BERT and both the encoder and decoder. Weng et al. (2022) leverage mBERT as an encoder and introduce a decoder that attends to grouped representations of the encoder output. + +Another research direction focuses on developing LMs with the encoder-decoder architecture designed for NMT as the target downstream task (Lewis et al., 2020; Liu et al., 2020). These models show improvements particularly for low-resource language pairs. To enhance their adaptation for MT, various methods have been proposed, including fine-tuning specific parts of the LMs (Cooper Stickland et al., 2021), reducing domain mismatch and overestimation (Wang et al., 2022) and mitigating the copying behavior (Liu et al., 2021b). + +The integration of pre-trained LMs into SiMT remains an underexplored area of research. To date, Indurthi et al. (2022) is the only related study we are aware of. It improves MMA by integrating the LM's prediction of the next target token, which is encoded using their model's vocabulary before being inputted into the model. However, this approach sacrifices the semantic coherence of the + +original tokens due to token fragmentation. Additionally, their approach falls under target-side LM integration, overlooking the potential advantages of source-side LM integration. + +In this paper, we demonstrate an effective way of integrating a source-side LM into SiMT systems, offering a more versatile solution that can be integrated into most existing neural SiMT models. Building upon previous research conducted in offline MT (Zhu et al., 2020), we introduce essential modifications, with a particular focus on word-level policies as a pivotal component. The effective management of vocabulary mismatches between the LM and the SiMT model is contingent upon the successful implementation of a word-level SiMT policy, a key aspect that we address in our study. + +# 3 Proposed Methods + +In this section, we propose the concept of employing a word-level latency metric and outline our conversion process for translating token-level policies into their word-level equivalents. Additionally, we present an integration of LM into SiMT, highlighting the advantages of utilizing word-level policies. + +# 3.1 Preliminaries + +Given a source sentence $\mathbf{x} = (x_{1}, x_{2}, \ldots x_{n})$ , the goal of a SiMT model is to generate a target sentence of $\mathbf{y} = (y_{1}, y_{2}, \ldots y_{m})$ while minimizing latency metrics. A SiMT model's policy, represented by the variable $g_{i}$ , determines the number of source tokens to process before predicting target token $y_{i}$ . Then the probability of generating $\mathbf{y}$ given $\mathbf{x}$ is formulated as follows: + +$$ +p (\mathbf {y} | \mathbf {x}) = \prod_ {i} ^ {| \mathbf {y} |} p \left(y _ {i} \mid \mathbf {x} _ {\leq g _ {i}}, \mathbf {y} _ {< i}; \theta\right) \tag {1} +$$ + +where $\theta$ is the model's parameters which are commonly optimized with a cross-entropy loss. + +Transformer encoder-decoder model (Vaswani et al., 2017) is currently the most widely used architecture for SiMT. To avoid redundant encoding of the input sequence after each READ operation, the encoder is typically modified to encode the source tokens unidirectionally (Elbayad et al., 2020). Alternatively, more advanced techniques like the recurrent Linear Transformer (Kahardipraja et al., 2021) or Partial Bidirectional Encoding (Iranzo Sanchez et al., 2022) can be adopted to enhance the encoding capabilities further. + +During the evaluation of SiMT systems, translation quality is commonly assessed in conjunction with the latency required for generating translations. Various metrics have been proposed to calculate latency scores, with Average Lagging (AL) (Ma et al., 2019a) being the most commonly used metric. + +# 3.2 Measuring latency based on the word level + +As detailed in A.1, a substantial body of prior research work assesses the performance of SiMT systems by utilizing a latency metric on encoded source tokens under different tokenization and encoding schemes. This practice results in each system being evaluated on non-identical token sequences for the same dataset, thereby making it challenging to accurately compare scores across different systems. + +To address this, we propose word-level latency score calculation by considering the word boundaries in token-level sequences. Specifically, when the first token of a source word is processed through a READ operation, we consider it as reading the corresponding word. Similarly, when the last token of a target word is written via a WRITE operation, we consider it as writing that word. By doing so, the latency scores are calculated consistently, regardless of the tokenization and encoding of the input. This ensures that results from different systems can be compared fairly. + +# 3.3 Word-level SiMT policies + +The proposed word-level policy restricts a SiMT policy's transition from READ to WRITE or vice versa to occur exclusively at the boundaries of words. Any token-level policy can be transformed to operate at the word-level by following the conversion process we outline below. + +Concretely, we ensure that a word-level policy does not write a target token in the middle of reading a sequence of source tokens that make up a word. To accomplish this word-level READ, we delay $g_{i}$ until it reaches the nearest source word boundary. We formally define $r_{i}$ that has a refined value of $g_{i}$ based on the word boundaries in $\mathbf{x}$ as follows: + +$$ +r _ {i} := \min \{j | j \geq g _ {i} \wedge j \in B _ {S} \} \tag {2} +$$ + +Here, $B_{S}$ denotes the indices of the source words' last tokens. Substituting $r_i$ for $g_i$ as a policy transforms it into another policy that upholds the same + +![](images/c9c360f553d97fd7c1279b3790fec24e9433423a6cc35154833a43eb39f6596e.jpg) +Token-level unidirectional encoding +Figure 3: Comparison of masking in the token-level unidirectional attention (left) and the intra-word bidirectional encoding (right). Word boundaries are represented by vertical/horizontal bars on each axis. + +![](images/eeb2ecb69355b1b06599dbab51b7beb5aab89743c0897a11dbf4adddc39a2098.jpg) +Intra-word bidirectional encoding + +decision-making criterion while ensuring uninterrupted reading of an entire word when initiating the reading of a token. + +Similarly to the word-level READ, we design a word-level WRITE to balance READ and WRITE actions throughout the translation. To achieve this, we can modify $r_i$ such that it writes until it produces a token that ends with an end-of-word symbol. We define $w_i$ that satisfies this as follows: + +$$ +b _ {i} := \min \{j | j \geq i \wedge j \in B _ {T} \} \tag {3} +$$ + +$$ +w _ {i} := \left\{ \begin{array}{l l} r _ {i}, & \text {i f} i = 1 \vee b _ {i - 1} \neq b _ {i} \\ w _ {i - 1}, & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +where $B_{T}$ denotes the indices of the target words' last tokens and $b_{i}$ represents the index of the final token in the word that includes $y_{i}$ . By employing $w_{i}$ in place of $r_{i}$ (or $g_{i}$ ), we ensure that the policy consistently composes entire words without interruptions from any READ actions. This approach effectively reduces latency by facilitating faster writing of certain tokens compared to the original policy, thereby compensating for the increased latency resulting from word-level READ operations. Figure 1 provides a visual comparison between word-level and token-level policies in the context of Wait-1, with the word-level policy encompassing both word-level READ and WRITE operations. + +# 3.4 Intra-word bidirectional encoding + +Unidirectional encoding in SiMT is vital for managing computational complexity and training efficiency. However, it has an inevitable consequence of weakening the source sequence representations compared to bidirectional encoding. This is an additional factor contributing to the lower translation + +![](images/a727ea4b5c0a6c130677d30ed24fcb7c203ad12fa20e003632cbfa99f9e280a6.jpg) +Figure 4: Illustration of LM-fused attention with the word-level Wait-1. Word and model activations processed at a specific stage are highlighted in red. When a source word is received, it is independently encoded by the LM's vocabulary and the in-domain vocabulary. The hidden activations of the word from the LM is then utilized in the encoder. The decoder generates a sequence of tokens for a word by using both the LM and encoder activations. + +![](images/ff2cb4a8e818630be28676722bcc1420696bbff0092e9c83b3b5ecdea1197ad2.jpg) + +quality of SiMT compared to offline models, along with the early translation from partial inputs. + +To mitigate this issue, we utilize a technique called intra-word bidirectional encoding. At the word level, this approach involves unidirectional encoding for each word in the input sentence, meaning past words cannot attend to future words. However, at the subword level, subwords within the same word can attend to each other, allowing past subwords to attend to future subwords within the same word. Since READ operates at the word level in word-level policies, this encoding does not require recomputation during each WRITE operation. It only necessitates a single forward pass, similar to token-level unidirectional encoding. However, it can produce a better encoded representation by enabling attention to more tokens. An example masking to enable intra-word bidirectional encoding is depicted in Figure 3. + +# 3.5 Integration of LM into SiMT through word-level policies + +In this subsection, we showcase an additional benefit of word-level policies when integrating an LM into a SiMT system. One of the key challenges in this integration is the vocabulary mismatch between the LM and the SiMT model, which hinders ensuring that both models process an equal amount of input prefix at each translation step. + +One possible solution is to use the LM's vocabulary for the SiMT model. However, the LM's training data may not align well with the specific domain targeted by the SiMT system (Wang et al., 2022). This can result in suboptimal vocabulary for the SiMT model compared to a vocabulary obtained from in-domain data (Liu et al., 2021a). An + +other option is to explore methods to bridge vocabulary gaps (Kim et al., 2019; Sato et al., 2020; Liu et al., 2021a), but they are either validated only in certain transfer learning scenarios or require an additional training phase to train adapters or finetuning the entire LM using pre-training data. + +In this paper, we introduce a method for leveraging an LM in a manner that facilitates the integration of an off-the-shelf LM into a SiMT model, utilizing a word-level policy, regardless of vocabulary mismatches and the internal structure of the SiMT model. Specifically, we employ an LM fused attention for both the encoder and decoder, following the approach outlined in (Zhu et al., 2020), but with two notable modifications. + +Firstly, we replace BERT with a decoder-only auto-regressive LM (Radford et al., 2019; Lin et al., 2022) for unidirectional encoding of the input, aligning with SiMT models for efficient training and inference. Secondly, the attention between the SiMT model and the LM occurs when both models execute a word-level READ for an input word. This ensures they interact only when they process an equal amount of input prefix, naturally resolving the synchronization issue. Additionally, as they align at every word boundary, the SiMT model can operate independently with a vocabulary derived from in-domain data, while the LM continues to use its original vocabulary. Unlike methods targeting specific SiMT models (Indurthi et al., 2022), our approach can benefit any Neural SiMT model with any decoder-only LM. Figure 4 illustrates the proposed integration of the LM with word-level Wait-1. + +![](images/d5bcec75ccb8102db8d55b6645ec1b3a09e7c3c9ad3e216c6f2cdca793b7754b.jpg) + +![](images/24e7a2f13f43e53f3d020453632e5e0a7c5f48531eac80496f7b5711a1de1be3.jpg) + +![](images/0743529918e4cac20bc0bb494314a3051c8894e58014c09ac047a402ac000152.jpg) + +![](images/bcf1943b0977721a88373dcdf9a230257aa37eb2d928590877358cfcfef7a357.jpg) +(a) En-Fr Wait-k Transformer small +(d) De-En Wait-k Transformer big +Figure 5: Results of translation quality v.s. word-level AL. + +![](images/97ec0498c4715076be48915533852f0c18de28f3c468d5e5f253a81e911273a5.jpg) +(b) En-Fr MoE Transformer small +(e) De-En MoE Transformer big + +![](images/33c00968ad7a1b50905ae7a5d41482ddc496a57786dd6a9eba79a39c63e3dd2c.jpg) +(c) En-Fr ITST Transformer small +(f) De-En ITST Transformer big + +# 4 Experiments + +# 4.1 Datasets + +We use the following two datasets for experiments: IWSLT 17 English to French (En-Fr) (Cettolo et al., 2017) consists of $242\mathrm{k}$ pairs split into $233\mathrm{k}$ training, 890 validation, and 8,597 test pairs. + +WMT 15 German to English (De-En) comprises 4.5 million pairs. We use newstest2013 (3,000 pairs) for validation and newstest2015 (2,169 pairs) for testing. + +# 4.2 Experimental settings + +Our methods are evaluated on the following systems, all of which are based on Transformer (Vaswani et al., 2017) with unidirectional encoding. + +Wait-k (Ma et al., 2019a): A model operating under Wait-k policy. A single model is trained for all k values by randomly sampling k during training (Elbayad et al., 2020). For the word-level Wait-k policy, we define it as reading the first k words and then alternating between reading one source word and writing one target word. + +MoE Wait-k (Zhang and Feng, 2021): A Mixture-of-Experts model initially trained with + +fixed Experts weights and then fine-tuned with dynamic Experts weights. The word-level Wait-k policy of different k is applied to each expert. + +ITST (Zhang and Feng, 2022b): A SoTA SiMT model equipped with Information-Transport-based policy that quantifies information weights from each source to the current target token. To implement word-level ITST, we convert the number of source tokens required for the first target token of each target word into the corresponding number of source words using Equation 2. Once the required number of source words is read, we complete the translation of the word. Additionally, we calculate the latency cost at the word level. + +We compare each system by training both token-level and word-level models, with and without an LM. For models with an LM, we use XGLM-564M (Lin et al., 2022) and employ the two-stage training in which we first train the SiMT model without the LM and initialize the encoder and decoder from the LM-fused model with the pre-trained weights (Zhu et al., 2020). We also tested the single stage training where all trainable parameters are trained jointly with the LM from scratch. The difference of these strategies are discussed in Section 5.3. We tokenized and encoded the input using sentencepiece (Kudo and Richardson, 2018) and applied BPE with a vocabulary size of $32k$ . We use sacreBLEU for BLEU calculation (Post, 2018). For models with token-level policies, we trained + +models with the official implementations $^{4 5 6}$ , and implemented word-level policies based on these implementations. More training details are described in A.2. + +# 4.3 Main results + +The performance of each system is compared in Figure 5. Notably, when measuring latency at the word level, the word-level policy proves to be highly effective for all three systems across different latency levels and datasets, resulting in superior performance. The only exception is observed in the En-Fr ITST models that demonstrate similar levels of performance. The incorporation of an LM using the proposed LM-fused attention further enhances the performance for all word-level configurations. This observation highlights the suitability of word-level policies for the LM-fused attention approach and underscores the effectiveness of leveraging an LM to enhance SiMT systems. + +Notably, as depicted in Figure 11, the word-level policies also outperform or compete with the token-level policies in token-level latency. This can be attributed to the enhanced token representations under the word-level policy, thanks to the contextual information provided by all other tokens belonging to the same word for each token. + +# 5 Analysis + +To validate the effectiveness of word-level policies from multiple angles, we conduct several analyses on various settings. All the experiments were conducted on WMT De.En with transformer-big unless specified otherwise. + +# 5.1 Ablation study + +# 5.1.1 Effects of Word-level READ and WRITE + +To gain insights into the functionality of word-level READ and WRITE actions, we trained Wait-k models with various policy settings and conducted a performance comparison. Specifically, we examined models with the following policy settings: + +WW: word-level READ and WRITE. + +TW: token-level READ and word-level WRITE. + +WT: word-level READ and token-level WRITE. + +TkTk: a simpler baseline policy which involves alternating reading k source tokens and writing k + +![](images/1617a6da950878273b5b9d7426a25edf9b884d66ba1fa0358f278f70b88bb03a.jpg) +(a) +Figure 6: Ablation studies for word-level policies. (a): Comparison of word-level Wait-k policies with different policies. (b): Comparison of word-level Wait-k models with and without the intra-word bidirectional encoding. + +![](images/9c1e514dc8f8250f15e8c991565d6d7356e7701d4a87a47975224a0d4f40d773.jpg) +(b) + +target tokens without considering word boundaries. + +The results are presented in Figure 6 (a). The word-level policy (WW) consistently outperforms TW across all latency settings. This is attributed to its imbalance between the number of source and target prefixes processed in each step. Additionally, WT achieves a minimal AL of approximately 3.6, indicating that it is not well-suited for scenarios that require low latency. Lastly, TkTk shows significantly worse performance than WW, suggesting that reading or writing a few consecutive tokens without considering semantic boundaries offers no benefits, unlike word-level policies. + +# 5.1.2 Effects of intra-word bidirectional encoding + +In order to assess the impact of the proposed intraword bidirectional encoding, we trained word-level Wait-K models with and without it and compared the accuracy of the two models across different AL settings. The results are presented in Figure 6 (b). + +Remarkably, the model equipped with the intraword bidirectional encoding consistently achieved higher BLEU scores compared to the model without it, across all tested latency settings. This provides strong evidence of the effectiveness of the intra-word bidirectional encoding in enhancing SiMT performance. + +# 5.2 Effectiveness of word-level policies for LM + +In this subsection, we aim to explore the significance of word-level policies in leveraging LM for SiMT. We compare different configurations based on three factors: + +Word vs. Token: The policy type that the model operates with. + +In-domain vocab vs. LM vocab: Whether the + +![](images/779b878c35c67f128b72cafc92caeeafcaf256e3dd6b25c7f9734de8b0e5742f.jpg) +Figure 7: Comparison of models with different LM integration. + +model uses an in-domain vocabulary obtained from the in-domain training data or uses the LM's vocabulary for the source language. Note that the use of "in-domain vocab" is specific to the "Word" configuration due to the vocabulary mismatch. + +LM-fused attention vs. LM embed: Whether the model incorporates an LM using the LM-fused attention or replacing the embedding layer of the encoder with the LM's embedding (Xu et al., 2021). The latter approach uses "LM vocab" by design. + +Figure 7 showcases the results. The models with word-level policies consistently outperform those with token-level policies by a significant margin in both LM-fused attention and LM embedding settings, underscoring the importance of word-level policy adoption for effective LM integration. The top-performing configuration is the proposed LMAtn-In-domain vocab-Word, demonstrating that the highest translation accuracy is achieved when the SiMT model operates with an in-domain vocabulary. Additionally, it is evident that the "LM embed" approach performs notably worse than the proposed LM-fused attention, further affirming the latter's superiority. + +# 5.3 Effects of LM-fused attention with various LMs and training configurations + +To assess the effectiveness and broad applicability of our proposed LM integration, we conducted experiments on the IWSLT17 En-Fr dataset with two decoder-only LMs of different sizes: the 137M parameter GPT-2 model (Radford et al., 2018) and the XGLM-564M model. Additionally, we explore the option of training models in the single training stage instead of the two-stage training. The results, presented in Figure 8, demonstrate that GPT-2 model also exhibits improved performance with LM-fused attention, although their impact is + +![](images/c99091fa070de565656017ad1894b99576ec056ee67b7d47ac0af000302a2a50.jpg) +(a) +Figure 8: Translation accuracy comparison of LM-fused attention models with different training configurations. (a): En-Fr Transformer small (b) De-En Transformer big + +![](images/31ef32ad43fd802b73ec692f820117850df0cf34c92ee5f6260366c2e282f676.jpg) +(b) + +naturally less pronounced compared to XGLM due to the difference in model size. Moreover, although models trained using the single-stage training generally exhibit lower performance compared to those trained using the two-stage training, they still outperform models without the LM for most configurations. This indicates that the LM-fused attention is applicable to various types of LMs and remains effective even when using the single-stage training strategy. This flexibility allows users to choose a training approach and model configuration that aligns best with their desired accuracy goals and computational constraints. + +# 5.4 Policy quality comparison + +To assess the accuracy of policies in determining when to read or write, we adopt the methodology of estimating the quality of a policy in prior research (Zhang and Feng, 2022b,c; Guo et al., 2022). We measure the quality of a policy by analyzing the proportion of aligned source words received before translating on RWTH De-En alignment dataset7. + +To ensure accurate word-level alignment calculation, we consider an aligned source word is read before writing a ground truth (GT) target word if the last token of the source word is read before the first token of the target word is written. The results of this analysis are presented in Figure 9. It is observed that word-level policies, both for ITST and Wait-k, exhibit better alignments across most latency settings. This suggests that word-level policies contribute to revising premature WRITE actions by guiding the model to read the remaining tokens of the aligned word, without negatively impacting the model's latency. + +![](images/96dac2a388686619654ce0211c754f994a329e79389415a4a32950cde31ea028.jpg) +Figure 9: The percentage of aligned source words read before the translation of the target words started. + +# 6 Conclusions + +This paper explores the potential benefits of word-level operations in SiMT systems. We propose word-level latency calculation to ensure fair and accurate latency comparisons. We introduce a conversion process that transforms token-level policies into word-level policies, enabling the processing of multiple subwords that form a word within a single READ or WRITE action. Additionally, we propose the integration of LM-fused attention, which combines an autoregressive LM model into SiMT models with word-level policies. Experimental results demonstrate the superiority of word-level policies compared to token-level policies, as well as the effectiveness of the LM integration. Our findings highlight the crucial role of word-level policies in the integration process. + +# 7 Limitations + +While the proposed word-level policy implementation is widely applicable to most existing SiMT systems, it is important to note that systems utilizing languages with a writing style that lacks spaces or other delimiters between words or sentences (e.g., Chinese) are unable to derive benefits from this approach. Furthermore, it is important to consider that while the proposed LM-fused attention proves effective in enhancing translation quality across all latency levels, integrating a large LM may necessitate a faster compute capability to fulfill the low-latency demands of the SiMT task. + +# References + +Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. 2019. Monotonic infinite lookback attention for simultaneous machine + +translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1313-1323, Florence, Italy. Association for Computational Linguistics. +Ozan Caglayan, Julia Ive, Veneta Haralampieva, Pranava Madhyastha, Loic Barrault, and Lucia Specia. 2020. Simultaneous machine translation with visual context. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2350-2361, Online. Association for Computational Linguistics. +Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuhito Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the IWSLT 2017 evaluation campaign. In Proceedings of the 14th International Conference on Spoken Language Translation, pages 2-14, Tokyo, Japan. International Workshop on Spoken Language Translation. +Chung-Cheng Chiu and Colin Raffel. 2017. Monotonic chunkwise attention. arXiv preprint arXiv:1712.05382. +Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? arXiv preprint arXiv:1606.02012. +Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3440-3453, Online. Association for Computational Linguistics. +Fahim Dalvi, Nadir Durrani, Hassan Sajjad, and Stephan Vogel. 2018. Incremental decoding and training methods for simultaneous translation in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 493-499, New Orleans, Louisiana. Association for Computational Linguistics. +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics. +Maha Elbayad, Laurent Besacier, and Jakob Verbeek. 2020. Efficient Wait-k Models for Simultaneous Machine Translation. In Proc. Interspeech 2020, pages 1461-1465. +Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor O.K. Li. 2017. Learning to translate in real-time + +with neural machine translation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1053-1062, Valencia, Spain. Association for Computational Linguistics. +Shoutao Guo, Shaolei Zhang, and Yang Feng. 2022. Turning fixed to adaptive: Integrating post-evaluation into simultaneous machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2264–2278, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Shoutao Guo, Shaolei Zhang, and Yang Feng. 2023. Learning optimal policy for simultaneous machine translation via binary search. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2318-2333, Toronto, Canada. Association for Computational Linguistics. +Sathish Reddy Indurthi, Mohd Abbas Zaidi, Beomseok Lee, Nikhil Kumar Lakumarapu, and Sangha Kim. 2022. Language model augmented monotonic attention for simultaneous translation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 38-45, Seattle, United States. Association for Computational Linguistics. +Javier Iranzo Sanchez, Jorge Civera, and Alfons Juan-Ciscar. 2022. From simultaneous to streaming machine translation by leveraging streaming history. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6972-6985, Dublin, Ireland. Association for Computational Linguistics. +Patrick Kahardipraja, Brielen Madureira, and David Schlangen. 2021. Towards incremental transformers: An empirical analysis of transformer models for incremental NLU. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1178-1189, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1246–1257, Florence, Italy. Association for Computational Linguistics. +Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics. + +Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint arXiv:1901.07291. +Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics. +Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettle-moyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, and Xian Li. 2022. Few-shot learning with multilingual generative language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9019-9052, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Xin Liu, Baosong Yang, Dayiheng Liu, Haibo Zhang, Weihua Luo, Min Zhang, Haiying Zhang, and Jinsong Su. 2021a. Bridging subword gaps in pretrainfinetune paradigm for natural language generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6001-6011, Online. Association for Computational Linguistics. +Xuebo Liu, Longyue Wang, Derek F. Wong, Liang Ding, Lidia S. Chao, Shuming Shi, and Zhaopeng Tu. 2021b. On the copying behaviors of pre-training for neural machine translation. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 4265-4275, Online. Association for Computational Linguistics. +Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742. +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. +Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, Hua Wu, and Haifeng Wang. 2019a. STACL: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework. In Proceedings of the 57th Annual Meeting of the Association for + +Computational Linguistics, pages 3025-3036, Florence, Italy. Association for Computational Linguistics. +Xutai Ma, Juan Pino, James Cross, Liezl Puzon, and Jiatao Gu. 2019b. Monotonic multihead attention. arXiv preprint arXiv:1909.12406. +Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics. +Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training. +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9. +Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, and Douglas Eck. 2017. Online and linear-time attention by enforcing monotonic alignments. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 2837-2846. PMLR. +Shoetsu Sato, Jin Sakuma, Naoki Yoshinaga, Masashi Toyoda, and Masaru Kitsuregawa. 2020. Vocabulary adaptation for domain adaptation in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4269-4279, Online. Association for Computational Linguistics. +Felix Schneider and Alexander Waibel. 2020. Towards stream translation: Adaptive computation time for simultaneous machine translation. In Proceedings of the 17th International Conference on Spoken Language Translation, pages 228-236, Online. Association for Computational Linguistics. +Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics. +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. +Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. 2023. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2334-2349, Toronto, Canada. Association for Computational Linguistics. + +Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2591-2600, Dublin, Ireland. Association for Computational Linguistics. +Rongxiang Weng, Heng Yu, Shujian Huang, Shanbo Cheng, and Weihua Luo. 2020. Acquiring knowledge from pre-trained model to neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9266-9273. +Rongxiang Weng, Heng Yu, Weihua Luo, and Min Zhang. 2022. Deep fusing pre-trained models into neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11468-11476. +Haoran Xu, Benjamin Van Durme, and Kenton Murray. 2021. BERT, mBERT, or BiBERT? a study on contextualized embeddings for neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6663-6675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Jiacheng Yang, Mingxuan Wang, Hao Zhou, Chengqi Zhao, Weinan Zhang, Yong Yu, and Lei Li. 2020. Towards making the most of bert in neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9378-9385. +Shaolei Zhang and Yang Feng. 2021. Universal simultaneous machine translation with mixture-of-experts wait-k policy. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7306-7317, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. +Shaolei Zhang and Yang Feng. 2022a. Gaussian multihead attention for simultaneous machine translation. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 3019–3030, Dublin, Ireland. Association for Computational Linguistics. +Shaolei Zhang and Yang Feng. 2022b. Information-transport-based policy for simultaneous translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 992-1013, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. +Shaolei Zhang and Yang Feng. 2022c. Modeling dual read/write paths for simultaneous machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2461-2477, Dublin, Ireland. Association for Computational Linguistics. + +Shaolei Zhang and Yang Feng. 2023. Hidden markov transformer for simultaneous machine translation. arXiv preprint arXiv:2303.00257. + +Shaolei Zhang, Yang Feng, and Liangyou Li. 2021. Future-guided incremental transformer for simultaneous translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14428-14436. + +Shaolei Zhang, Shoutao Guo, and Yang Feng. 2022. Wait-info policy: Balancing source and target at information level for simultaneous machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2249-2263, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics. + +Baigong Zheng, Kaibo Liu, Renjie Zheng, Mingbo Ma, Hairong Liu, and Liang Huang. 2020. Simultaneous translation policies: From fixed to adaptive. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2847-2853, Online. Association for Computational Linguistics. + +Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. 2020. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823. + +# A Appendix + +# A.1 Token-level latency evaluation in previous works + +Many previous works in the field of SiMT have employed token-level latency calculation as a fundamental metric for evaluating system performance. In the first group of papers, which includes (Arivazhagan et al., 2019; Ma et al., 2019b; Schneider and Waibel, 2020; Zheng et al., 2020; Zhang and Feng, 2022b), researchers explicitly stated their use of token-level latency calculation or incorporated token-level scores in their analyses. On the other hand, the second group, comprising (Zhang and Feng, 2022b; Elbayad et al., 2020; Zhang and Feng, 2022c, 2021, 2022a; Zhang et al., 2022; Guo et al., 2022; Zhang and Feng, 2023; Guo et al., 2023; Caglayan et al., 2020; Wang et al., 2023), can be identified by their publicly available official implementations, which incorporate token-level latency evaluation code. This underscores the prevalence of token-level latency calculation in assessing the effectiveness of SiMT systems in the cutting-edge SiMT research works. + +# A.2 More training details + +For Transformer model configurations, we follow the Efficient Wait-k (Elbayad et al., 2020)'s set + +tings. $^{8}$ The details of hyperparameter settings for each model is in Table 1. + +To ensure optimal performance of the ITST models, we conducted multiple training runs for each model configuration and dataset. We observed performance fluctuations across the training runs. To select the best-performing models for both token-level and word-level policies, we repeated the training process three times and selected the checkpoint with the lowest validation loss for each configuration. + +For LM integrated models, the intra-word bidirectional encoding was not applied to the LM to prevent the need for fine-tuning. + +# A.3 Examples and discussions + +In Table 2, an illustrative case is presented to demonstrate the distinction between a token-level policy (token-level Wait-1) and a word-level policy (word-level Wait-1). At Step 2, the token-level model predicts the target token “B” after processing the subword “B” for the word “Beine” (which means “leg” in German). Subsequently, it fails to recover from this incorrect prediction and continues by predicting “ody_” as the next token, even after processing the remaining token “eine_”. In contrast, the word-level model accurately predicts “legs” (encoded as “leg s_”) after processing the complete word ‘Beine’ (encoded as ‘B eine_’). Furthermore, the token model also makes erroneous predictions for all steps when processing the word ‘blutüberströmmt.’ (encoded as ‘bl ut über ström t_’, meaning “covered in blood” in German), while the word-level model accurately predicts a correct word in a single step. + +Another case can be observed in Table 3, where the token-level model initiates the translation before fully reading the word "Irgendetwas" (encoded as "Irgen det was"). As a consequence, it produces an incorrect translation. On the other hand, the word-level model accurately translates the sentence by processing the complete word before making any predictions. + +# A.4 Table of main results + +# A.5 Graphs of main results + +
Hyperparametertransformer-smalltransformer-basetransformer-big
attention heads4816
embedding dime2565121024
ffn embedding dim102420484096
dropout0.30.30.3
encoder layers666
decoder layers666
lr5e-45e-45e-4
lr schedulerinverse sqrtinverse sqrtinverse sqrt
optimizerAdam (0.9, 0.98)Adam (0.9, 0.98)Adam (0.9, 0.98)
clip-norm000
warmup-updates400040004000
warmup-init-lr1e-71e-71e-7
weight decay000
label smoothing0.10.10.1
max tokens327685242881048576
+ +Table 1: Hyperparameters of each Transformer model. + +
Src:Meine_B eine_waren_bl ut über ström t.
Ref:My leg s_were Covered_in_blood.
StepStreaminig InputToken Wait-1 OutputWord Wait-1 Output
1MeineMyMy
2Meine_BB
3Meine_B eineodyleg s
4Meine_B eine_warenwaswere
5Meine_B eine_waren_blblue
6Meine_B eine_waren_bl utand
7Meine_B eine_waren_bl ut übermy
8Meine_B eine_waren_bl ut über strömleg
9Meine_B eine_waren_bl ut über ström ts
10Meine_B eine_waren_bl ut über ström t.were_blood - ri dden.blood dy.
+ +Table 2: An example case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.56, word-AL: 1.69) and Word-level Wait-1 (token-AL: 1.44, word-AL: 1.0). + +
Src:Ir gen det was lag in der Luft .
Ref:Some thing was up .
StepStreaming InputToken Wait-1 OutputWord Wait-1 Output
1IrIr
2Ir gengen
3Ir gen detgen
4Ir gen det wasgenSome thing_
5Ir gen det was laggenlay_
6Ir gen det was lag insin_
7Ir gen det was lag in der,the_
8Ir gen det was lag in der Luftin
9Ir gen det was lag in der Luft .fact , was in the air .air .
+ +Table 3: An Eexample case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.69, word-AL: 4.25) and Word-level Wait-1 (token-AL: 2.38, word-AL: 1.0). + +![](images/a7c2f1f08eb9e88dd9ef08d785d6c4d82ff39d39211e8d4a400c58abdefa40f8.jpg) +(a) De-En Wait-k Transformer base + +![](images/84c13e68369add4451739ed4985cf2fb0a7369e75084a4fb33c26288da6bc039.jpg) +(b) De-En MoE Transformer base +Figure 10: Results of translation quality v.s. word-level AL, De-En transformer-base. + +![](images/cd652d3f4e088a493be32d365fa0f641cc4dc5ab82a6105e750d9829cf91bf6b.jpg) +(c) De-En ITST Transformer base + +![](images/0e4a8516a1f8797398d67d483ca155986b84acc7cc5ec468050494c9897920fc.jpg) + +![](images/0ada3edc3344d0d06ec3dbf9bd85e2fb68caf26a3742abbb66fb233aa6b1dbe3.jpg) + +![](images/3f5c144dc9fe3d0c40307c39c86dab14519469c5de082683a35f9b04a5c995d2.jpg) + +![](images/323f89b5334a770fa7032530970985fae90748ef9da6d7a287e0d0b646cc03df.jpg) +(a) En-Fr Wait-k Transformer small + +![](images/aafe8835c201e63e967ad66c630b512e34ecebeabb9c769a7b3549bc2c37a95f.jpg) +(b) En-Fr MoE Transformer small + +![](images/eee39750e797615ae83ba316087c8f4da0123f2576d70ffd41df469544822c36.jpg) +(c) En-Fr ITST Transformer small + +![](images/cbf2f32a207e488330d3e6161e6f6e583f398e48e59c8ea6d91e5c204171cc12.jpg) +(d) De-En Wait-k Transformer base +(g) De-En Wait-k Transformer big +Figure 11: Results of translation quality v.s. token-level AL. + +![](images/21ab652590c28f0f7a7cebd049f50ff4ba38d82570f5d7c63b42d018b98f631c.jpg) +(e) De-En MoE Transformer base +(h) De-En MoE Transformer big + +![](images/177f24907f5fd2547d3881b704fdf7ff430fdffd2d5f4ce05a256029d5feb0dd.jpg) +(f) De-En ITST Transformer base +(i) De-En ITST Transformer big + +
Token-level Wait-k
ktoken ALword ALBLEU
11.951.9132.32
33.833.4338.90
55.674.8240.68
77.516.2340.95
99.287.5841.06
1110.918.8241.14
1312.399.9541.12
Word-level Wait-k
ktoken ALword ALBLEU
12.241.7432.43
34.183.3239.36
56.555.1541.09
78.856.9441.48
910.928.5741.57
1112.759.9841.58
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.181.7433.47
34.233.3740.83
56.585.1942.78
78.866.9743.00
910.948.5843.08
1112.7610.0043.09
+ +Table 4: Main results of transformer-small Waik-k models on IWLST17 En.Fr dataset. + +
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.901.8832.20
33.833.4238.92
55.694.8440.51
77.536.2540.83
99.317.6040.92
1110.948.8540.95
1312.429.9740.96
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.011.6032.01
34.183.3239.28
56.575.1740.77
78.866.9541.03
910.948.5841.11
1112.769.9941.10
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.171.7233.01
34.223.3740.46
56.565.1742.35
78.846.9542.61
910.918.5742.66
1112.749.9942.67
+ +Table 5: Main results of transformer-small MoE Waik-k models on IWSLT17 En.Fr dataset. + +
Token-level ITST
δtoken ALword ALBLEU
0.23.333.0537.81
0.33.593.2538.64
0.44.143.7039.40
0.55.184.5940.19
0.66.715.9240.84
0.78.627.6341.27
0.759.568.4541.34
0.810.729.4141.42
0.8512.1110.5341.42
0.913.8411.8041.42
Word-level ITST
δtoken ALword ALBLEU
0.23.272.6437.25
0.33.462.8037.88
0.43.913.2138.69
0.54.673.9139.57
0.65.814.9640.44
0.77.296.2840.94
0.758.217.0541.13
0.89.418.0241.36
0.8511.199.3741.50
0.913.3611.0041.53
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.252.6138.30
0.33.302.6538.75
0.43.482.8139.61
0.53.943.2340.50
0.64.804.0441.35
0.76.025.1941.88
0.756.765.8642.09
0.87.726.6742.30
0.858.997.6842.56
0.910.909.1642.63
+ +Table 6: Main results of transformer-small ITST models on IWSLT17 En.Fr dataset. + +
Token-level Wait-k
ktoken ALword ALBLEU
10.520.7517.63
32.161.9123.24
53.923.0526.14
75.774.2327.46
97.765.5028.77
119.646.6929.37
1311.497.8629.88
1513.188.9229.89
1714.799.9330.19
Word-level Wait-k
ktoken ALword ALBLEU
12.961.8424.02
35.333.4328.08
58.085.2329.35
710.827.0129.87
913.478.7230.17
1115.8810.2530.38
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.961.8424.71
35.273.3828.49
58.055.2029.95
710.817.0030.56
913.448.6930.74
1115.8510.2430.95
+ +Table 7: Main results of transformer-base Waik-k models on WMT15 De.En dataset. + +
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.171.1117.28
32.502.0923.29
54.173.1825.95
76.004.3627.86
97.815.5128.81
119.626.6629.53
1311.427.8029.66
1513.158.8929.81
1714.809.9129.79
+ +
Word-level MoE Wait-k
ktoken ALword ALBLEU
13.121.9123.59
35.293.3927.75
58.035.1929.29
710.796.9830.14
913.388.6530.39
1115.8110.230.6
+ +
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
13.081.9224.97
35.323.4428.66
58.085.2330.21
710.877.0430.63
913.478.7330.9
1115.8810.2631.0
+ +Table 8: Main results of transformer-base MoE Waik-k models on WMT15 De.En dataset. + +
Token-level ITST
δtoken ALword ALBLEU
0.22.452.0822.19
0.32.892.3824.15
0.43.822.9626.21
0.55.233.9228.19
0.67.285.5029.25
0.79.877.4829.47
0.7511.668.8129.76
0.813.7910.3829.79
0.8516.6012.1129.90
0.919.8113.8529.71
+ +
Word-level ITST
δtoken ALword ALBLEU
0.24.132.6426.81
0.34.192.6927.01
0.44.392.8127.47
0.55.093.3028.04
0.66.434.2828.87
0.78.605.9329.54
0.759.916.9229.54
0.811.548.0929.67
0.8513.839.6129.82
0.916.7511.3829.93
+ +
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.24.152.6427.74
0.34.242.7027.99
0.44.552.9228.44
0.55.333.4929.15
0.66.774.6329.97
0.78.576.0830.49
0.759.516.8030.53
0.810.717.6730.60
0.8512.328.7230.68
0.914.289.9330.71
+ +Table 9: Main results of transformer-base ITST models on WMT15 De.En dataset. + +
Token-level Wait-k
ktoken ALword ALBLEU
1-0.130.3919.20
31.801.7124.01
53.742.9526.86
75.814.2628.86
97.805.5329.96
119.726.7430.62
1311.517.8730.91
1513.208.9531.16
1714.849.9731.31
+ +
Word-level Wait-k
ktoken ALword ALBLEU
12.551.5924.63
35.113.3028.82
57.955.1630.43
710.776.9931.35
913.378.67831.86
1115.8010.2232.03
+ +
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.891.7926.21
35.213.3630.16
57.985.1731.36
710.827.0132.29
913.428.7132.58
1115.7410.1632.45
+ +Table 10: Main results of transformer-big Waik-k models on WMT15 De.En dataset. + +
Token-level MoE Wait-k
ktoken ALword ALBLEU
10.860.9418.44
32.191.9224.07
53.903.0426.70
75.704.1928.47
97.695.4729.59
119.606.6730.36
1311.447.8330.51
1513.148.9130.56
1714.759.9130.65
+ +
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.861.7624.79
35.243.3728.95
57.985.1530.26
710.776.9731.01
913.378.6631.47
1115.8010.2031.55
+ +
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.601.6025.26
35.193.3129.03
58.015.1630.79
710.836.9931.66
913.438.6731.82
1115.8310.2132.17
+ +Table 11: Main results of transformer-big MoE Waik-k models on WMT15 De.En dataset. + +
Token-level ITST
δtoken ALword ALBLEU
0.21.491.5523.22
0.32.412.1224.75
0.43.933.0626.67
0.55.994.3928.23
0.69.126.5729.85
0.712.699.3231.02
0.7514.7210.7531.28
0.816.9512.2331.54
0.8519.0613.4931.70
0.921.2614.6931.84
Word-level ITST
δtoken ALword ALBLEU
0.23.692.3526.33
0.33.792.4226.62
0.43.942.5426.62
0.54.693.0227.55
0.66.604.2728.99
0.710.286.7730.41
0.7512.508.3430.95
0.814.229.6531.54
0.8515.9010.8131.56
0.918.1312.2331.61
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.972.5328.12
0.34.002.5528.13
0.44.102.6228.29
0.54.392.8228.92
0.65.373.4629.66
0.77.855.1331.18
0.759.736.4831.77
0.811.908.0432.10
0.8514.049.5732.34
0.916.1410.9732.62
+ +Table 12: Main results of transformer-big ITST models on WMT15 De.En dataset. \ No newline at end of file diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/images.zip b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf01aeac67ccf2ef25322f2cc3244468eb67ad21 --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b3508ff77778ceeac49f4d61dcef766ebaa4e772c12561d749cb17a6b03ab8b +size 1405954 diff --git a/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/layout.json b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..55ca0f377a905d8030097ef35e15c67880658b89 --- /dev/null +++ b/2023/Enhanced Simultaneous Machine Translation with Word-level Policies/layout.json @@ -0,0 +1,11406 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 75, + 513, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 75, + 513, + 92 + ], + "spans": [ + { + "bbox": [ + 80, + 75, + 513, + 92 + ], + "type": "text", + "content": "Enhanced Simultaneous Machine Translation with Word-level Policies" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 229, + 120, + 366, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 120, + 366, + 133 + ], + "spans": [ + { + "bbox": [ + 229, + 120, + 366, + 133 + ], + "type": "text", + "content": "Kang Kim*, Hankyu Cho*" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 275, + 135, + 320, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 275, + 135, + 320, + 146 + ], + "spans": [ + { + "bbox": [ + 275, + 135, + 320, + 146 + ], + "type": "text", + "content": "XL8 Inc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 235, + 149, + 361, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 149, + 361, + 161 + ], + "spans": [ + { + "bbox": [ + 235, + 149, + 361, + 161 + ], + "type": "text", + "content": "{kai, matthew}@x18.ai" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "spans": [ + { + "bbox": [ + 155, + 212, + 202, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 232, + 274, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 232, + 274, + 483 + ], + "spans": [ + { + "bbox": [ + 84, + 232, + 274, + 483 + ], + "type": "text", + "content": "Recent years have seen remarkable advances in the field of Simultaneous Machine Translation (SiMT) due to the introduction of innovative policies that dictate whether to READ or WRITE at each step of the translation process. However, a common assumption in many existing studies is that operations are carried out at the subword level, even though the standard unit for input and output in most practical scenarios is typically at the word level. This paper demonstrates that policies devised and validated at the subword level are surpassed by those operating at the word level, which process multiple subwords to form a complete word in a single step. Additionally, we suggest a method to boost SiMT models using language models (LMs), wherein the proposed word-level policy plays a vital role in addressing the subword disparity between LMs and SiMT models. Code is available at https://github.com/x18-ai/WordSiMT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 491, + 154, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 491, + 154, + 503 + ], + "spans": [ + { + "bbox": [ + 68, + 491, + 154, + 503 + ], + "type": "text", + "content": "1 Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 511, + 291, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 511, + 291, + 645 + ], + "spans": [ + { + "bbox": [ + 67, + 511, + 291, + 645 + ], + "type": "text", + "content": "Simultaneous Machine Translation (SiMT) commences the translation process while simultaneously receiving the input, making it an effective approach for applications that require minimal latency such as simultaneous interpretation or live broadcast. The development of a novel policy is central to research efforts in SiMT. This policy dictates the translation process by determining whether to execute a READ or WRITE action at each step of the process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 647, + 291, + 754 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 647, + 291, + 754 + ], + "spans": [ + { + "bbox": [ + 67, + 647, + 291, + 754 + ], + "type": "text", + "content": "Neural SiMT models, like offline Neural Machine Translation (NMT) models, commonly employ Byte Pair Encoding (BPE) (Sennrich et al., 2016) or similar techniques to encode an input sentence into a sequence of tokens. Typically, a single READ or WRITE action of a SiMT policy is responsible for handling an encoded token, which may sometimes be a word but often a subword." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 210, + 526, + 306 + ], + "blocks": [ + { + "bbox": [ + 304, + 210, + 526, + 306 + ], + "lines": [ + { + "bbox": [ + 304, + 210, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 526, + 306 + ], + "type": "image", + "image_path": "698d58ed3af2627eea3939c9cb293ed07e64a927e1c670a7484c4fa413657720.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 314, + 526, + 363 + ], + "lines": [ + { + "bbox": [ + 302, + 314, + 526, + 363 + ], + "spans": [ + { + "bbox": [ + 302, + 314, + 526, + 363 + ], + "type": "text", + "content": "Figure 1: Illustration of both token-level Wait-1 (red arrow lines) and word-level Wait-1 (green arrow lines) policies. The symbol \" " + }, + { + "bbox": [ + 302, + 314, + 526, + 363 + ], + "type": "inline_equation", + "content": "\\_" + }, + { + "bbox": [ + 302, + 314, + 526, + 363 + ], + "type": "text", + "content": " \" at the end of a token indicates a word boundary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 384, + 525, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 384, + 525, + 504 + ], + "spans": [ + { + "bbox": [ + 302, + 384, + 525, + 504 + ], + "type": "text", + "content": "The development of BPE-based SiMT models and their policies has resulted in researchers focusing on working at the subword level. The performance analysis and implementation of many SiMT systems, to the best of our knowledge, have been carried out on encoded sequences of source and target subwords, rather than on the original source and target sentences1. This has led to two critical issues that need to be addressed." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 506, + 525, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 506, + 525, + 587 + ], + "spans": [ + { + "bbox": [ + 302, + 506, + 525, + 587 + ], + "type": "text", + "content": "The first issue is the lack of a standardized tokenization and encoding scheme, meaning that different implementations may employ varying token sequences to encode identical text. This variability can impact latency evaluation results and complicate score comparisons across different systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 588, + 526, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 588, + 526, + 723 + ], + "spans": [ + { + "bbox": [ + 302, + 588, + 526, + 723 + ], + "type": "text", + "content": "The second issue is the missed opportunity to process more source tokens before writing each target token without added latency. For a BPE-based SiMT model, the input must be received on a word-by-word basis to ensure proper encoding of each word into a sequence of subwords. Consequently, when the model encodes a word and performs a READ to process only a subword, it delays the reading of the remaining subwords without any benefit in actual latency², and may adversely impact" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 302, + 730, + 525, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 730, + 525, + 750 + ], + "spans": [ + { + "bbox": [ + 302, + 730, + 525, + 750 + ], + "type": "text", + "content": "1We provide a list of reference works pertaining to this case in A.1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 751, + 525, + 772 + ], + "type": "text", + "content": "Here we assume that the processing time of a READ is negligible, compared to the interval between receiving words." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 81, + 761, + 156, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 156, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 156, + 772 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 283, + 780, + 312, + 791 + ], + "type": "text", + "content": "15616" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 124, + 795, + 468, + 806 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 795, + 468, + 806 + ], + "spans": [ + { + "bbox": [ + 124, + 795, + 468, + 806 + ], + "type": "text", + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15616-15634" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 165, + 807, + 428, + 817 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 807, + 428, + 817 + ], + "spans": [ + { + "bbox": [ + 165, + 807, + 428, + 817 + ], + "type": "text", + "content": "December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 526, + 131 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 526, + 131 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 526, + 131 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 526, + 131 + ], + "type": "image", + "image_path": "231157632db099fa480ad5e3f153a799abe77907d7c96e89f011863bd87a35c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 140, + 526, + 188 + ], + "lines": [ + { + "bbox": [ + 66, + 140, + 526, + 188 + ], + "spans": [ + { + "bbox": [ + 66, + 140, + 526, + 188 + ], + "type": "text", + "content": "Figure 2: An exemplary case depicting the difference between token-level and word-level Wait-k policies and their Average Lagging (AL) scores. The token-level model begins translating in the middle of reading \"Tremendously\" and fails to recover from the incorrectly translated target prefix. On the other hand, the word-level model processes \"Tremendously\" in a single READ action and produces a correct translation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 210, + 290, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 210, + 290, + 304 + ], + "spans": [ + { + "bbox": [ + 66, + 210, + 290, + 304 + ], + "type": "text", + "content": "translation quality by causing the model to rely on incomplete representations extracted from partially-read words. Similarly, performing a WRITE to generate a subword earlier than the remaining subwords does not necessarily reduce latency, as the subword must wait for a complete word to be generated before it can be displayed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 307, + 290, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 307, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 66, + 307, + 290, + 455 + ], + "type": "text", + "content": "In this paper, we show that establishing the unit of measuring and operating SiMT systems at the word level, rather than the subword level, offers a viable solution to these issues. Specifically, to tackle the first issue, we propose word-level latency metric calculation for measuring the latency of SiMT systems. This not only enables consistent comparisons between different systems but also provides a more accurate reflection of the actual latency experienced in SiMT applications that display translation results word by word." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 459, + 290, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 459, + 290, + 688 + ], + "spans": [ + { + "bbox": [ + 66, + 459, + 290, + 688 + ], + "type": "text", + "content": "To address the second issue, we illustrate that an existing token-level policy can be transformed into a word-level policy that inherently overcomes the second issue, resulting in improved performance. Word-level policies take into consideration word boundaries and perform a READ or WRITE action to sequentially process a sequence of tokens that form a complete word. This conversion process can be applied to any token-level policies, and our experiments reveal that state-of-the-art fixed and adaptive policies exhibit significantly better performance when transformed into their word-level counterparts. Notably, these word-level policies often outperform token-level policies, even when evaluated using a token-level latency metric, due to their enhanced utilization of input and output tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 692, + 291, + 772 + ], + "type": "text", + "content": "Additionally, to boost translation accuracy, we suggest incorporating a pre-trained language model (LM) into a SiMT model, where the word-level policy plays a crucial role as a pivotal component. One of the major hurdles in utilizing an LM for SiMT is the vocabulary mismatch between the SiMT model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 210, + 526, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 210, + 526, + 344 + ], + "spans": [ + { + "bbox": [ + 302, + 210, + 526, + 344 + ], + "type": "text", + "content": "and the LM. The difficulty of handling subword disparities when utilizing an LM for a downstream task is widely acknowledged (Liu et al., 2021a; Wang et al., 2022), and it becomes particularly problematic in SiMT, as inconsistent subwords between the LM and SiMT model make processing the same source or target prefix challenging. Our study demonstrates that our proposed word-level policy effectively tackles this challenge, enabling a successful integration of LMs into SiMT systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 355, + 395, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 355, + 395, + 367 + ], + "spans": [ + { + "bbox": [ + 302, + 355, + 395, + 367 + ], + "type": "text", + "content": "2 Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 376, + 494, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 376, + 494, + 389 + ], + "spans": [ + { + "bbox": [ + 302, + 376, + 494, + 389 + ], + "type": "text", + "content": "2.1 Simultaneous Machine Translation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 394, + 525, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 394, + 525, + 582 + ], + "spans": [ + { + "bbox": [ + 302, + 394, + 525, + 582 + ], + "type": "text", + "content": "SiMT systems that employ a fixed policy utilize a pre-defined sequence of READ and WRITE operations for each source sentence. STATIC-RW (Dalvi et al., 2018) and Wait-k (Ma et al., 2019a) policies first read k source tokens, then alternate between reading and writing a single token. Elbayad et al. (2020) propose the multi-path training of a Wait-k model to train a single model that supports different k values at test time. Zhang et al. (2021) improve Wait-k policy using knowledge distillation from an offline MT model, while Zhang and Feng (2021) suggest a Mixture-of-Experts Wait-k Policy where predictions from multiple k values are combined inside a single model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 584, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 584, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 584, + 525, + 772 + ], + "type": "text", + "content": "In contrast, research efforts on adaptive policies focus on the development of dynamic decision-making processes for READ/WRITE actions. Cho and Esipova (2016) firstly introduce model-based adaptive criteria for Neural SiMT. Gu et al. (2017) propose to learn a policy by using reinforcement learning. Raffel et al. (2017) introduce Monotonic Attention that ensures monotonic alignment in the attention mechanism. Succeeding works improve it by extending the alignment window (Chiu and Raffel, 2017; Arivazhagan et al., 2019), extending it as monotonic multi-head attention (MMA) (Ma et al., 2019b) or learning transposed policies between the forward and backward models. (Zhang and Feng," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15617" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 290, + 166 + ], + "type": "text", + "content": "2022c). Zheng et al. (2020) derive a policy by composing Wait-k models trained with different values of k. Zhang and Feng (2022a) model to predict the alignment between each target token and the source token. Zhang and Feng (2022b) measure accumulated information from source tokens to decide whether to write the current target token." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 168, + 291, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 168, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 168, + 291, + 275 + ], + "type": "text", + "content": "Despite significant advancements, the impact of operating policies at the word level has not been thoroughly explored in existing works, which have mainly focused on developing and evaluating systems at the token level. In this paper, we address this gap by demonstrating that implementing various types of policies at the word level consistently outperforms their token-level counterparts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 290, + 248, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 290, + 248, + 303 + ], + "spans": [ + { + "bbox": [ + 67, + 290, + 248, + 303 + ], + "type": "text", + "content": "2.2 Utilizing pre-trained LM for MT" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 310, + 291, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 310, + 291, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 310, + 291, + 513 + ], + "type": "text", + "content": "Since the successful utilization of Transformer-based LMs pre-trained on large text corpora for downstream NLP tasks (Devlin et al., 2019; Liu et al., 2019; Lample and Conneau, 2019), the utilization of these models for MT has become a significant research area. Several studies have demonstrated the effectiveness of incorporating encoder-only LMs into NMT models. Weng et al. (2020); Yang et al. (2020) combine the LM representations with the encoder's representation using a gating mechanism. Zhu et al. (2020) propose attention between BERT and both the encoder and decoder. Weng et al. (2022) leverage mBERT as an encoder and introduce a decoder that attends to grouped representations of the encoder output." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 514, + 291, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 514, + 291, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 514, + 291, + 663 + ], + "type": "text", + "content": "Another research direction focuses on developing LMs with the encoder-decoder architecture designed for NMT as the target downstream task (Lewis et al., 2020; Liu et al., 2020). These models show improvements particularly for low-resource language pairs. To enhance their adaptation for MT, various methods have been proposed, including fine-tuning specific parts of the LMs (Cooper Stickland et al., 2021), reducing domain mismatch and overestimation (Wang et al., 2022) and mitigating the copying behavior (Liu et al., 2021b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 291, + 772 + ], + "type": "text", + "content": "The integration of pre-trained LMs into SiMT remains an underexplored area of research. To date, Indurthi et al. (2022) is the only related study we are aware of. It improves MMA by integrating the LM's prediction of the next target token, which is encoded using their model's vocabulary before being inputted into the model. However, this approach sacrifices the semantic coherence of the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "spans": [ + { + "bbox": [ + 302, + 71, + 526, + 125 + ], + "type": "text", + "content": "original tokens due to token fragmentation. Additionally, their approach falls under target-side LM integration, overlooking the potential advantages of source-side LM integration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 126, + 527, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 126, + 527, + 288 + ], + "spans": [ + { + "bbox": [ + 302, + 126, + 527, + 288 + ], + "type": "text", + "content": "In this paper, we demonstrate an effective way of integrating a source-side LM into SiMT systems, offering a more versatile solution that can be integrated into most existing neural SiMT models. Building upon previous research conducted in offline MT (Zhu et al., 2020), we introduce essential modifications, with a particular focus on word-level policies as a pivotal component. The effective management of vocabulary mismatches between the LM and the SiMT model is contingent upon the successful implementation of a word-level SiMT policy, a key aspect that we address in our study." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 300, + 420, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 300, + 420, + 313 + ], + "spans": [ + { + "bbox": [ + 302, + 300, + 420, + 313 + ], + "type": "text", + "content": "3 Proposed Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 322, + 527, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 322, + 527, + 403 + ], + "spans": [ + { + "bbox": [ + 302, + 322, + 527, + 403 + ], + "type": "text", + "content": "In this section, we propose the concept of employing a word-level latency metric and outline our conversion process for translating token-level policies into their word-level equivalents. Additionally, we present an integration of LM into SiMT, highlighting the advantages of utilizing word-level policies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 414, + 394, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 414, + 394, + 426 + ], + "spans": [ + { + "bbox": [ + 302, + 414, + 394, + 426 + ], + "type": "text", + "content": "3.1 Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "spans": [ + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": "Given a source sentence " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x_{1}, x_{2}, \\ldots x_{n})" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": ", the goal of a SiMT model is to generate a target sentence of " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (y_{1}, y_{2}, \\ldots y_{m})" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": " while minimizing latency metrics. A SiMT model's policy, represented by the variable " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": ", determines the number of source tokens to process before predicting target token " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": ". Then the probability of generating " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 302, + 433, + 525, + 540 + ], + "type": "text", + "content": " is formulated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 553, + 525, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 553, + 525, + 590 + ], + "spans": [ + { + "bbox": [ + 343, + 553, + 525, + 590 + ], + "type": "interline_equation", + "content": "p (\\mathbf {y} | \\mathbf {x}) = \\prod_ {i} ^ {| \\mathbf {y} |} p \\left(y _ {i} \\mid \\mathbf {x} _ {\\leq g _ {i}}, \\mathbf {y} _ {< i}; \\theta\\right) \\tag {1}", + "image_path": "a62911540cea2abcee78145e7bb9fc053080690b46f8d85af8bac45de2feb177.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 597, + 526, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 597, + 526, + 624 + ], + "spans": [ + { + "bbox": [ + 302, + 597, + 526, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 597, + 526, + 624 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 302, + 597, + 526, + 624 + ], + "type": "text", + "content": " is the model's parameters which are commonly optimized with a cross-entropy loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 624, + 527, + 772 + ], + "type": "text", + "content": "Transformer encoder-decoder model (Vaswani et al., 2017) is currently the most widely used architecture for SiMT. To avoid redundant encoding of the input sequence after each READ operation, the encoder is typically modified to encode the source tokens unidirectionally (Elbayad et al., 2020). Alternatively, more advanced techniques like the recurrent Linear Transformer (Kahardipraja et al., 2021) or Partial Bidirectional Encoding (Iranzo Sanchez et al., 2022) can be adopted to enhance the encoding capabilities further." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15618" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 291, + 153 + ], + "type": "text", + "content": "During the evaluation of SiMT systems, translation quality is commonly assessed in conjunction with the latency required for generating translations. Various metrics have been proposed to calculate latency scores, with Average Lagging (AL) (Ma et al., 2019a) being the most commonly used metric." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 164, + 289, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 164, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 67, + 164, + 289, + 178 + ], + "type": "text", + "content": "3.2 Measuring latency based on the word level" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 184, + 290, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 184, + 290, + 305 + ], + "spans": [ + { + "bbox": [ + 67, + 184, + 290, + 305 + ], + "type": "text", + "content": "As detailed in A.1, a substantial body of prior research work assesses the performance of SiMT systems by utilizing a latency metric on encoded source tokens under different tokenization and encoding schemes. This practice results in each system being evaluated on non-identical token sequences for the same dataset, thereby making it challenging to accurately compare scores across different systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 306, + 291, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 306, + 291, + 469 + ], + "spans": [ + { + "bbox": [ + 67, + 306, + 291, + 469 + ], + "type": "text", + "content": "To address this, we propose word-level latency score calculation by considering the word boundaries in token-level sequences. Specifically, when the first token of a source word is processed through a READ operation, we consider it as reading the corresponding word. Similarly, when the last token of a target word is written via a WRITE operation, we consider it as writing that word. By doing so, the latency scores are calculated consistently, regardless of the tokenization and encoding of the input. This ensures that results from different systems can be compared fairly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 481, + 214, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 481, + 214, + 495 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 214, + 495 + ], + "type": "text", + "content": "3.3 Word-level SiMT policies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 501, + 290, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 501, + 290, + 581 + ], + "spans": [ + { + "bbox": [ + 67, + 501, + 290, + 581 + ], + "type": "text", + "content": "The proposed word-level policy restricts a SiMT policy's transition from READ to WRITE or vice versa to occur exclusively at the boundaries of words. Any token-level policy can be transformed to operate at the word-level by following the conversion process we outline below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "spans": [ + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "content": "Concretely, we ensure that a word-level policy does not write a target token in the middle of reading a sequence of source tokens that make up a word. To accomplish this word-level READ, we delay " + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "content": " until it reaches the nearest source word boundary. We formally define " + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "content": " that has a refined value of " + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "content": " based on the word boundaries in " + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 67, + 583, + 290, + 690 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 705, + 290, + 719 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 705, + 290, + 719 + ], + "spans": [ + { + "bbox": [ + 108, + 705, + 290, + 719 + ], + "type": "interline_equation", + "content": "r _ {i} := \\min \\{j | j \\geq g _ {i} \\wedge j \\in B _ {S} \\} \\tag {2}", + "image_path": "af48315684ee251688177b956273bfb5dfb95ebce4571001d15f1f265264c866.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "inline_equation", + "content": "B_{S}" + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": " denotes the indices of the source words' last tokens. Substituting " + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 67, + 732, + 291, + 772 + ], + "type": "text", + "content": " as a policy transforms it into another policy that upholds the same" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 304, + 69, + 414, + 170 + ], + "blocks": [ + { + "bbox": [ + 304, + 69, + 414, + 170 + ], + "lines": [ + { + "bbox": [ + 304, + 69, + 414, + 170 + ], + "spans": [ + { + "bbox": [ + 304, + 69, + 414, + 170 + ], + "type": "image", + "image_path": "c9c360f553d97fd7c1279b3790fec24e9433423a6cc35154833a43eb39f6596e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 172, + 414, + 182 + ], + "lines": [ + { + "bbox": [ + 304, + 172, + 414, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 414, + 182 + ], + "type": "text", + "content": "Token-level unidirectional encoding" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 190, + 526, + 238 + ], + "lines": [ + { + "bbox": [ + 302, + 190, + 526, + 238 + ], + "spans": [ + { + "bbox": [ + 302, + 190, + 526, + 238 + ], + "type": "text", + "content": "Figure 3: Comparison of masking in the token-level unidirectional attention (left) and the intra-word bidirectional encoding (right). Word boundaries are represented by vertical/horizontal bars on each axis." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 418, + 68, + 524, + 170 + ], + "blocks": [ + { + "bbox": [ + 418, + 68, + 524, + 170 + ], + "lines": [ + { + "bbox": [ + 418, + 68, + 524, + 170 + ], + "spans": [ + { + "bbox": [ + 418, + 68, + 524, + 170 + ], + "type": "image", + "image_path": "eeb2ecb69355b1b06599dbab51b7beb5aab89743c0897a11dbf4adddc39a2098.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 172, + 523, + 182 + ], + "lines": [ + { + "bbox": [ + 421, + 172, + 523, + 182 + ], + "spans": [ + { + "bbox": [ + 421, + 172, + 523, + 182 + ], + "type": "text", + "content": "Intra-word bidirectional encoding" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 261, + 525, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 261, + 525, + 301 + ], + "spans": [ + { + "bbox": [ + 302, + 261, + 525, + 301 + ], + "type": "text", + "content": "decision-making criterion while ensuring uninterrupted reading of an entire word when initiating the reading of a token." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "spans": [ + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "text", + "content": "Similarly to the word-level READ, we design a word-level WRITE to balance READ and WRITE actions throughout the translation. To achieve this, we can modify " + }, + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "text", + "content": " such that it writes until it produces a token that ends with an end-of-word symbol. We define " + }, + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "inline_equation", + "content": "w_i" + }, + { + "bbox": [ + 302, + 302, + 525, + 383 + ], + "type": "text", + "content": " that satisfies this as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 398, + 525, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 398, + 525, + 412 + ], + "spans": [ + { + "bbox": [ + 345, + 398, + 525, + 412 + ], + "type": "interline_equation", + "content": "b _ {i} := \\min \\{j | j \\geq i \\wedge j \\in B _ {T} \\} \\tag {3}", + "image_path": "e8f46cf71a152fb0f2ce73b5b4e8360f9e67e12035f2bee00e9cc478f0e12c8e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 332, + 419, + 525, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 419, + 525, + 455 + ], + "spans": [ + { + "bbox": [ + 332, + 419, + 525, + 455 + ], + "type": "interline_equation", + "content": "w _ {i} := \\left\\{ \\begin{array}{l l} r _ {i}, & \\text {i f} i = 1 \\vee b _ {i - 1} \\neq b _ {i} \\\\ w _ {i - 1}, & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "3d060bc27d0d8628a1bd468afcd0776d0e232b9960b28a2b2f704c5e9b61d430.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "spans": [ + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "B_{T}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": " denotes the indices of the target words' last tokens and " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": " represents the index of the final token in the word that includes " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": ". By employing " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": " in place of " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "r_{i}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 302, + 460, + 525, + 663 + ], + "type": "text", + "content": "), we ensure that the policy consistently composes entire words without interruptions from any READ actions. This approach effectively reduces latency by facilitating faster writing of certain tokens compared to the original policy, thereby compensating for the increased latency resulting from word-level READ operations. Figure 1 provides a visual comparison between word-level and token-level policies in the context of Wait-1, with the word-level policy encompassing both word-level READ and WRITE operations." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 674, + 490, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 674, + 490, + 687 + ], + "spans": [ + { + "bbox": [ + 302, + 674, + 490, + 687 + ], + "type": "text", + "content": "3.4 Intra-word bidirectional encoding" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 692, + 525, + 772 + ], + "type": "text", + "content": "Unidirectional encoding in SiMT is vital for managing computational complexity and training efficiency. However, it has an inevitable consequence of weakening the source sequence representations compared to bidirectional encoding. This is an additional factor contributing to the lower translation" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15619" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 68, + 261, + 200 + ], + "blocks": [ + { + "bbox": [ + 68, + 68, + 261, + 200 + ], + "lines": [ + { + "bbox": [ + 68, + 68, + 261, + 200 + ], + "spans": [ + { + "bbox": [ + 68, + 68, + 261, + 200 + ], + "type": "image", + "image_path": "a727ea4b5c0a6c130677d30ed24fcb7c203ad12fa20e003632cbfa99f9e280a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 207, + 526, + 255 + ], + "lines": [ + { + "bbox": [ + 67, + 207, + 526, + 255 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 526, + 255 + ], + "type": "text", + "content": "Figure 4: Illustration of LM-fused attention with the word-level Wait-1. Word and model activations processed at a specific stage are highlighted in red. When a source word is received, it is independently encoded by the LM's vocabulary and the in-domain vocabulary. The hidden activations of the word from the LM is then utilized in the encoder. The decoder generates a sequence of tokens for a word by using both the LM and encoder activations." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 263, + 68, + 526, + 200 + ], + "blocks": [ + { + "bbox": [ + 263, + 68, + 526, + 200 + ], + "lines": [ + { + "bbox": [ + 263, + 68, + 526, + 200 + ], + "spans": [ + { + "bbox": [ + 263, + 68, + 526, + 200 + ], + "type": "image", + "image_path": "ff2cb4a8e818630be28676722bcc1420696bbff0092e9c83b3b5ecdea1197ad2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 276, + 289, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 276, + 289, + 303 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 289, + 303 + ], + "type": "text", + "content": "quality of SiMT compared to offline models, along with the early translation from partial inputs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 305, + 291, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 305, + 291, + 534 + ], + "spans": [ + { + "bbox": [ + 67, + 305, + 291, + 534 + ], + "type": "text", + "content": "To mitigate this issue, we utilize a technique called intra-word bidirectional encoding. At the word level, this approach involves unidirectional encoding for each word in the input sentence, meaning past words cannot attend to future words. However, at the subword level, subwords within the same word can attend to each other, allowing past subwords to attend to future subwords within the same word. Since READ operates at the word level in word-level policies, this encoding does not require recomputation during each WRITE operation. It only necessitates a single forward pass, similar to token-level unidirectional encoding. However, it can produce a better encoded representation by enabling attention to more tokens. An example masking to enable intra-word bidirectional encoding is depicted in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 549, + 271, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 549, + 271, + 575 + ], + "spans": [ + { + "bbox": [ + 67, + 549, + 271, + 575 + ], + "type": "text", + "content": "3.5 Integration of LM into SiMT through word-level policies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 582, + 290, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 582, + 290, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 290, + 677 + ], + "type": "text", + "content": "In this subsection, we showcase an additional benefit of word-level policies when integrating an LM into a SiMT system. One of the key challenges in this integration is the vocabulary mismatch between the LM and the SiMT model, which hinders ensuring that both models process an equal amount of input prefix at each translation step." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 678, + 291, + 772 + ], + "type": "text", + "content": "One possible solution is to use the LM's vocabulary for the SiMT model. However, the LM's training data may not align well with the specific domain targeted by the SiMT system (Wang et al., 2022). This can result in suboptimal vocabulary for the SiMT model compared to a vocabulary obtained from in-domain data (Liu et al., 2021a). An" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 302, + 276, + 526, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 276, + 526, + 358 + ], + "spans": [ + { + "bbox": [ + 302, + 276, + 526, + 358 + ], + "type": "text", + "content": "other option is to explore methods to bridge vocabulary gaps (Kim et al., 2019; Sato et al., 2020; Liu et al., 2021a), but they are either validated only in certain transfer learning scenarios or require an additional training phase to train adapters or finetuning the entire LM using pre-training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 302, + 376, + 526, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 376, + 526, + 497 + ], + "spans": [ + { + "bbox": [ + 302, + 376, + 526, + 497 + ], + "type": "text", + "content": "In this paper, we introduce a method for leveraging an LM in a manner that facilitates the integration of an off-the-shelf LM into a SiMT model, utilizing a word-level policy, regardless of vocabulary mismatches and the internal structure of the SiMT model. Specifically, we employ an LM fused attention for both the encoder and decoder, following the approach outlined in (Zhu et al., 2020), but with two notable modifications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 516, + 526, + 772 + ], + "type": "text", + "content": "Firstly, we replace BERT with a decoder-only auto-regressive LM (Radford et al., 2019; Lin et al., 2022) for unidirectional encoding of the input, aligning with SiMT models for efficient training and inference. Secondly, the attention between the SiMT model and the LM occurs when both models execute a word-level READ for an input word. This ensures they interact only when they process an equal amount of input prefix, naturally resolving the synchronization issue. Additionally, as they align at every word boundary, the SiMT model can operate independently with a vocabulary derived from in-domain data, while the LM continues to use its original vocabulary. Unlike methods targeting specific SiMT models (Indurthi et al., 2022), our approach can benefit any Neural SiMT model with any decoder-only LM. Figure 4 illustrates the proposed integration of the LM with word-level Wait-1." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15620" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 217, + 180 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 217, + 180 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 217, + 180 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 217, + 180 + ], + "type": "image", + "image_path": "d5bcec75ccb8102db8d55b6645ec1b3a09e7c3c9ad3e216c6f2cdca793b7754b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 68, + 369, + 180 + ], + "blocks": [ + { + "bbox": [ + 224, + 68, + 369, + 180 + ], + "lines": [ + { + "bbox": [ + 224, + 68, + 369, + 180 + ], + "spans": [ + { + "bbox": [ + 224, + 68, + 369, + 180 + ], + "type": "image", + "image_path": "24e7a2f13f43e53f3d020453632e5e0a7c5f48531eac80496f7b5711a1de1be3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 68, + 524, + 180 + ], + "blocks": [ + { + "bbox": [ + 378, + 68, + 524, + 180 + ], + "lines": [ + { + "bbox": [ + 378, + 68, + 524, + 180 + ], + "spans": [ + { + "bbox": [ + 378, + 68, + 524, + 180 + ], + "type": "image", + "image_path": "0743529918e4cac20bc0bb494314a3051c8894e58014c09ac047a402ac000152.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 69, + 195, + 215, + 306 + ], + "blocks": [ + { + "bbox": [ + 77, + 184, + 208, + 194 + ], + "lines": [ + { + "bbox": [ + 77, + 184, + 208, + 194 + ], + "spans": [ + { + "bbox": [ + 77, + 184, + 208, + 194 + ], + "type": "text", + "content": "(a) En-Fr Wait-k Transformer small" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 195, + 215, + 306 + ], + "lines": [ + { + "bbox": [ + 69, + 195, + 215, + 306 + ], + "spans": [ + { + "bbox": [ + 69, + 195, + 215, + 306 + ], + "type": "image", + "image_path": "bcf1943b0977721a88373dcdf9a230257aa37eb2d928590877358cfcfef7a357.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 309, + 205, + 321 + ], + "lines": [ + { + "bbox": [ + 80, + 309, + 205, + 321 + ], + "spans": [ + { + "bbox": [ + 80, + 309, + 205, + 321 + ], + "type": "text", + "content": "(d) De-En Wait-k Transformer big" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 329, + 415, + 342 + ], + "lines": [ + { + "bbox": [ + 177, + 329, + 415, + 342 + ], + "spans": [ + { + "bbox": [ + 177, + 329, + 415, + 342 + ], + "type": "text", + "content": "Figure 5: Results of translation quality v.s. word-level AL." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 195, + 370, + 306 + ], + "blocks": [ + { + "bbox": [ + 235, + 184, + 359, + 194 + ], + "lines": [ + { + "bbox": [ + 235, + 184, + 359, + 194 + ], + "spans": [ + { + "bbox": [ + 235, + 184, + 359, + 194 + ], + "type": "text", + "content": "(b) En-Fr MoE Transformer small" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 195, + 370, + 306 + ], + "lines": [ + { + "bbox": [ + 223, + 195, + 370, + 306 + ], + "spans": [ + { + "bbox": [ + 223, + 195, + 370, + 306 + ], + "type": "image", + "image_path": "97ec0498c4715076be48915533852f0c18de28f3c468d5e5f253a81e911273a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 309, + 355, + 321 + ], + "lines": [ + { + "bbox": [ + 238, + 309, + 355, + 321 + ], + "spans": [ + { + "bbox": [ + 238, + 309, + 355, + 321 + ], + "type": "text", + "content": "(e) De-En MoE Transformer big" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 377, + 195, + 524, + 306 + ], + "blocks": [ + { + "bbox": [ + 388, + 184, + 513, + 194 + ], + "lines": [ + { + "bbox": [ + 388, + 184, + 513, + 194 + ], + "spans": [ + { + "bbox": [ + 388, + 184, + 513, + 194 + ], + "type": "text", + "content": "(c) En-Fr ITST Transformer small" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 195, + 524, + 306 + ], + "lines": [ + { + "bbox": [ + 377, + 195, + 524, + 306 + ], + "spans": [ + { + "bbox": [ + 377, + 195, + 524, + 306 + ], + "type": "image", + "image_path": "33c00968ad7a1b50905ae7a5d41482ddc496a57786dd6a9eba79a39c63e3dd2c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 309, + 510, + 321 + ], + "lines": [ + { + "bbox": [ + 392, + 309, + 510, + 321 + ], + "spans": [ + { + "bbox": [ + 392, + 309, + 510, + 321 + ], + "type": "text", + "content": "(f) De-En ITST Transformer big" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 363, + 154, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 363, + 154, + 376 + ], + "spans": [ + { + "bbox": [ + 67, + 363, + 154, + 376 + ], + "type": "text", + "content": "4 Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 386, + 135, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 386, + 135, + 397 + ], + "spans": [ + { + "bbox": [ + 67, + 386, + 135, + 397 + ], + "type": "text", + "content": "4.1 Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "spans": [ + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "text", + "content": "We use the following two datasets for experiments: IWSLT 17 English to French (En-Fr) (Cettolo et al., 2017) consists of " + }, + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "inline_equation", + "content": "242\\mathrm{k}" + }, + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "text", + "content": " pairs split into " + }, + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "inline_equation", + "content": "233\\mathrm{k}" + }, + { + "bbox": [ + 67, + 405, + 290, + 460 + ], + "type": "text", + "content": " training, 890 validation, and 8,597 test pairs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 461, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 461, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 461, + 290, + 515 + ], + "type": "text", + "content": "WMT 15 German to English (De-En) comprises 4.5 million pairs. We use newstest2013 (3,000 pairs) for validation and newstest2015 (2,169 pairs) for testing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 528, + 197, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 528, + 197, + 540 + ], + "spans": [ + { + "bbox": [ + 67, + 528, + 197, + 540 + ], + "type": "text", + "content": "4.2 Experimental settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 547, + 290, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 547, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 547, + 290, + 587 + ], + "type": "text", + "content": "Our methods are evaluated on the following systems, all of which are based on Transformer (Vaswani et al., 2017) with unidirectional encoding." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 67, + 588, + 290, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 588, + 290, + 682 + ], + "spans": [ + { + "bbox": [ + 67, + 588, + 290, + 682 + ], + "type": "text", + "content": "Wait-k (Ma et al., 2019a): A model operating under Wait-k policy. A single model is trained for all k values by randomly sampling k during training (Elbayad et al., 2020). For the word-level Wait-k policy, we define it as reading the first k words and then alternating between reading one source word and writing one target word." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 683, + 290, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 290, + 710 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 290, + 710 + ], + "type": "text", + "content": "MoE Wait-k (Zhang and Feng, 2021): A Mixture-of-Experts model initially trained with" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 363, + 525, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 363, + 525, + 404 + ], + "spans": [ + { + "bbox": [ + 302, + 363, + 525, + 404 + ], + "type": "text", + "content": "fixed Experts weights and then fine-tuned with dynamic Experts weights. The word-level Wait-k policy of different k is applied to each expert." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 406, + 525, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 406, + 525, + 554 + ], + "spans": [ + { + "bbox": [ + 302, + 406, + 525, + 554 + ], + "type": "text", + "content": "ITST (Zhang and Feng, 2022b): A SoTA SiMT model equipped with Information-Transport-based policy that quantifies information weights from each source to the current target token. To implement word-level ITST, we convert the number of source tokens required for the first target token of each target word into the corresponding number of source words using Equation 2. Once the required number of source words is read, we complete the translation of the word. Additionally, we calculate the latency cost at the word level." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "content": "We compare each system by training both token-level and word-level models, with and without an LM. For models with an LM, we use XGLM-564M (Lin et al., 2022) and employ the two-stage training in which we first train the SiMT model without the LM and initialize the encoder and decoder from the LM-fused model with the pre-trained weights (Zhu et al., 2020). We also tested the single stage training where all trainable parameters are trained jointly with the LM from scratch. The difference of these strategies are discussed in Section 5.3. We tokenized and encoded the input using sentencepiece (Kudo and Richardson, 2018) and applied BPE with a vocabulary size of " + }, + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "inline_equation", + "content": "32k" + }, + { + "bbox": [ + 302, + 556, + 526, + 772 + ], + "type": "text", + "content": ". We use sacreBLEU for BLEU calculation (Post, 2018). For models with token-level policies, we trained" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "text", + "content": "3Technically, the word-level policy derived from the token-level Wait-k through the conversion process in Section 3.3 can wait between 1 and " + }, + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "text", + "content": " tokens, depending on the input encoding. Therefore, it is not equivalent to the word-level Wait-k policy we define here, which always waits for " + }, + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 67, + 720, + 290, + 772 + ], + "type": "text", + "content": " words." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "15621" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 70, + 291, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 70, + 291, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 291, + 125 + ], + "type": "text", + "content": "models with the official implementations" + }, + { + "bbox": [ + 67, + 70, + 291, + 125 + ], + "type": "inline_equation", + "content": "^{4 5 6}" + }, + { + "bbox": [ + 67, + 70, + 291, + 125 + ], + "type": "text", + "content": ", and implemented word-level policies based on these implementations. More training details are described in A.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 135, + 155, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 135, + 155, + 147 + ], + "spans": [ + { + "bbox": [ + 67, + 135, + 155, + 147 + ], + "type": "text", + "content": "4.3 Main results" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 153, + 291, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 291, + 341 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 291, + 341 + ], + "type": "text", + "content": "The performance of each system is compared in Figure 5. Notably, when measuring latency at the word level, the word-level policy proves to be highly effective for all three systems across different latency levels and datasets, resulting in superior performance. The only exception is observed in the En-Fr ITST models that demonstrate similar levels of performance. The incorporation of an LM using the proposed LM-fused attention further enhances the performance for all word-level configurations. This observation highlights the suitability of word-level policies for the LM-fused attention approach and underscores the effectiveness of leveraging an LM to enhance SiMT systems." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 343, + 291, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 343, + 291, + 436 + ], + "spans": [ + { + "bbox": [ + 67, + 343, + 291, + 436 + ], + "type": "text", + "content": "Notably, as depicted in Figure 11, the word-level policies also outperform or compete with the token-level policies in token-level latency. This can be attributed to the enhanced token representations under the word-level policy, thanks to the contextual information provided by all other tokens belonging to the same word for each token." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 448, + 133, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 448, + 133, + 461 + ], + "spans": [ + { + "bbox": [ + 67, + 448, + 133, + 461 + ], + "type": "text", + "content": "5 Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 470, + 290, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 470, + 290, + 538 + ], + "spans": [ + { + "bbox": [ + 67, + 470, + 290, + 538 + ], + "type": "text", + "content": "To validate the effectiveness of word-level policies from multiple angles, we conduct several analyses on various settings. All the experiments were conducted on WMT De.En with transformer-big unless specified otherwise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 547, + 164, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 547, + 164, + 560 + ], + "spans": [ + { + "bbox": [ + 67, + 547, + 164, + 560 + ], + "type": "text", + "content": "5.1 Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 565, + 255, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 565, + 255, + 591 + ], + "spans": [ + { + "bbox": [ + 67, + 565, + 255, + 591 + ], + "type": "text", + "content": "5.1.1 Effects of Word-level READ and WRITE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 595, + 291, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 595, + 291, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 595, + 291, + 662 + ], + "type": "text", + "content": "To gain insights into the functionality of word-level READ and WRITE actions, we trained Wait-k models with various policy settings and conducted a performance comparison. Specifically, we examined models with the following policy settings:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 663, + 236, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 663, + 236, + 675 + ], + "spans": [ + { + "bbox": [ + 67, + 663, + 236, + 675 + ], + "type": "text", + "content": "WW: word-level READ and WRITE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 677, + 284, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 677, + 284, + 688 + ], + "spans": [ + { + "bbox": [ + 67, + 677, + 284, + 688 + ], + "type": "text", + "content": "TW: token-level READ and word-level WRITE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 691, + 282, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 691, + 282, + 702 + ], + "spans": [ + { + "bbox": [ + 67, + 691, + 282, + 702 + ], + "type": "text", + "content": "WT: word-level READ and token-level WRITE." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 704, + 289, + 731 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 289, + 731 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 289, + 731 + ], + "type": "text", + "content": "TkTk: a simpler baseline policy which involves alternating reading k source tokens and writing k" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 409, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 409, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 409, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 409, + 174 + ], + "type": "image", + "image_path": "1617a6da950878273b5b9d7426a25edf9b884d66ba1fa0358f278f70b88bb03a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "lines": [ + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "spans": [ + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 201, + 525, + 251 + ], + "lines": [ + { + "bbox": [ + 302, + 201, + 525, + 251 + ], + "spans": [ + { + "bbox": [ + 302, + 201, + 525, + 251 + ], + "type": "text", + "content": "Figure 6: Ablation studies for word-level policies. (a): Comparison of word-level Wait-k policies with different policies. (b): Comparison of word-level Wait-k models with and without the intra-word bidirectional encoding." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 417, + 73, + 520, + 174 + ], + "blocks": [ + { + "bbox": [ + 417, + 73, + 520, + 174 + ], + "lines": [ + { + "bbox": [ + 417, + 73, + 520, + 174 + ], + "spans": [ + { + "bbox": [ + 417, + 73, + 520, + 174 + ], + "type": "image", + "image_path": "9c1e514dc8f8250f15e8c991565d6d7356e7701d4a87a47975224a0d4f40d773.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "lines": [ + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "spans": [ + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 272, + 525, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 272, + 525, + 285 + ], + "spans": [ + { + "bbox": [ + 302, + 272, + 525, + 285 + ], + "type": "text", + "content": "target tokens without considering word boundaries." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 302, + 286, + 525, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 286, + 525, + 448 + ], + "spans": [ + { + "bbox": [ + 302, + 286, + 525, + 448 + ], + "type": "text", + "content": "The results are presented in Figure 6 (a). The word-level policy (WW) consistently outperforms TW across all latency settings. This is attributed to its imbalance between the number of source and target prefixes processed in each step. Additionally, WT achieves a minimal AL of approximately 3.6, indicating that it is not well-suited for scenarios that require low latency. Lastly, TkTk shows significantly worse performance than WW, suggesting that reading or writing a few consecutive tokens without considering semantic boundaries offers no benefits, unlike word-level policies." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 457, + 498, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 457, + 498, + 484 + ], + "spans": [ + { + "bbox": [ + 302, + 457, + 498, + 484 + ], + "type": "text", + "content": "5.1.2 Effects of intra-word bidirectional encoding" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 487, + 525, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 487, + 525, + 554 + ], + "spans": [ + { + "bbox": [ + 302, + 487, + 525, + 554 + ], + "type": "text", + "content": "In order to assess the impact of the proposed intraword bidirectional encoding, we trained word-level Wait-K models with and without it and compared the accuracy of the two models across different AL settings. The results are presented in Figure 6 (b)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 302, + 555, + 525, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 555, + 525, + 650 + ], + "spans": [ + { + "bbox": [ + 302, + 555, + 525, + 650 + ], + "type": "text", + "content": "Remarkably, the model equipped with the intraword bidirectional encoding consistently achieved higher BLEU scores compared to the model without it, across all tested latency settings. This provides strong evidence of the effectiveness of the intra-word bidirectional encoding in enhancing SiMT performance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 302, + 660, + 524, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 660, + 524, + 672 + ], + "spans": [ + { + "bbox": [ + 302, + 660, + 524, + 672 + ], + "type": "text", + "content": "5.2 Effectiveness of word-level policies for LM" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 302, + 678, + 525, + 730 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 678, + 525, + 730 + ], + "spans": [ + { + "bbox": [ + 302, + 678, + 525, + 730 + ], + "type": "text", + "content": "In this subsection, we aim to explore the significance of word-level policies in leveraging LM for SiMT. We compare different configurations based on three factors:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 302, + 732, + 524, + 759 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 732, + 524, + 759 + ], + "spans": [ + { + "bbox": [ + 302, + 732, + 524, + 759 + ], + "type": "text", + "content": "Word vs. Token: The policy type that the model operates with." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 302, + 760, + 524, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 760, + 524, + 772 + ], + "spans": [ + { + "bbox": [ + 302, + 760, + 524, + 772 + ], + "type": "text", + "content": "In-domain vocab vs. LM vocab: Whether the" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 80, + 739, + 277, + 750 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 739, + 277, + 750 + ], + "spans": [ + { + "bbox": [ + 80, + 739, + 277, + 750 + ], + "type": "text", + "content": "4Efficient Wait-k: https://github.com/elbayadm/attn2d" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 81, + 750, + 269, + 761 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 750, + 269, + 761 + ], + "spans": [ + { + "bbox": [ + 81, + 750, + 269, + 761 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 81, + 750, + 269, + 761 + ], + "type": "text", + "content": " MoE Wait-k: https://github.com/ictnlp/MoE-Waitk" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 81, + 761, + 220, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 761, + 220, + 772 + ], + "spans": [ + { + "bbox": [ + 81, + 761, + 220, + 772 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 81, + 761, + 220, + 772 + ], + "type": "text", + "content": "ITST: https://github.com/ictnlp/ITST" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15622" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 74, + 271, + 199 + ], + "blocks": [ + { + "bbox": [ + 73, + 74, + 271, + 199 + ], + "lines": [ + { + "bbox": [ + 73, + 74, + 271, + 199 + ], + "spans": [ + { + "bbox": [ + 73, + 74, + 271, + 199 + ], + "type": "image", + "image_path": "779b878c35c67f128b72cafc92caeeafcaf256e3dd6b25c7f9734de8b0e5742f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 210, + 291, + 236 + ], + "lines": [ + { + "bbox": [ + 67, + 210, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 67, + 210, + 291, + 236 + ], + "type": "text", + "content": "Figure 7: Comparison of models with different LM integration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 257, + 291, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 257, + 291, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 291, + 323 + ], + "type": "text", + "content": "model uses an in-domain vocabulary obtained from the in-domain training data or uses the LM's vocabulary for the source language. Note that the use of \"in-domain vocab\" is specific to the \"Word\" configuration due to the vocabulary mismatch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 324, + 291, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 324, + 291, + 391 + ], + "spans": [ + { + "bbox": [ + 67, + 324, + 291, + 391 + ], + "type": "text", + "content": "LM-fused attention vs. LM embed: Whether the model incorporates an LM using the LM-fused attention or replacing the embedding layer of the encoder with the LM's embedding (Xu et al., 2021). The latter approach uses \"LM vocab\" by design." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 393, + 291, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 393, + 291, + 582 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 291, + 582 + ], + "type": "text", + "content": "Figure 7 showcases the results. The models with word-level policies consistently outperform those with token-level policies by a significant margin in both LM-fused attention and LM embedding settings, underscoring the importance of word-level policy adoption for effective LM integration. The top-performing configuration is the proposed LMAtn-In-domain vocab-Word, demonstrating that the highest translation accuracy is achieved when the SiMT model operates with an in-domain vocabulary. Additionally, it is evident that the \"LM embed\" approach performs notably worse than the proposed LM-fused attention, further affirming the latter's superiority." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 592, + 290, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 592, + 290, + 620 + ], + "spans": [ + { + "bbox": [ + 67, + 592, + 290, + 620 + ], + "type": "text", + "content": "5.3 Effects of LM-fused attention with various LMs and training configurations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 624, + 291, + 772 + ], + "type": "text", + "content": "To assess the effectiveness and broad applicability of our proposed LM integration, we conducted experiments on the IWSLT17 En-Fr dataset with two decoder-only LMs of different sizes: the 137M parameter GPT-2 model (Radford et al., 2018) and the XGLM-564M model. Additionally, we explore the option of training models in the single training stage instead of the two-stage training. The results, presented in Figure 8, demonstrate that GPT-2 model also exhibits improved performance with LM-fused attention, although their impact is" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 411, + 174 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 411, + 174 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 411, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 411, + 174 + ], + "type": "image", + "image_path": "c99091fa070de565656017ad1894b99576ec056ee67b7d47ac0af000302a2a50.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "lines": [ + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "spans": [ + { + "bbox": [ + 352, + 182, + 364, + 192 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 202, + 526, + 250 + ], + "lines": [ + { + "bbox": [ + 302, + 202, + 526, + 250 + ], + "spans": [ + { + "bbox": [ + 302, + 202, + 526, + 250 + ], + "type": "text", + "content": "Figure 8: Translation accuracy comparison of LM-fused attention models with different training configurations. (a): En-Fr Transformer small (b) De-En Transformer big" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 418, + 73, + 521, + 173 + ], + "blocks": [ + { + "bbox": [ + 418, + 73, + 521, + 173 + ], + "lines": [ + { + "bbox": [ + 418, + 73, + 521, + 173 + ], + "spans": [ + { + "bbox": [ + 418, + 73, + 521, + 173 + ], + "type": "image", + "image_path": "31ef32ad43fd802b73ec692f820117850df0cf34c92ee5f6260366c2e282f676.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "lines": [ + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "spans": [ + { + "bbox": [ + 464, + 182, + 476, + 192 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 271, + 526, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 271, + 526, + 447 + ], + "spans": [ + { + "bbox": [ + 302, + 271, + 526, + 447 + ], + "type": "text", + "content": "naturally less pronounced compared to XGLM due to the difference in model size. Moreover, although models trained using the single-stage training generally exhibit lower performance compared to those trained using the two-stage training, they still outperform models without the LM for most configurations. This indicates that the LM-fused attention is applicable to various types of LMs and remains effective even when using the single-stage training strategy. This flexibility allows users to choose a training approach and model configuration that aligns best with their desired accuracy goals and computational constraints." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 456, + 453, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 456, + 453, + 470 + ], + "spans": [ + { + "bbox": [ + 302, + 456, + 453, + 470 + ], + "type": "text", + "content": "5.4 Policy quality comparison" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 302, + 474, + 525, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 474, + 525, + 567 + ], + "spans": [ + { + "bbox": [ + 302, + 474, + 525, + 567 + ], + "type": "text", + "content": "To assess the accuracy of policies in determining when to read or write, we adopt the methodology of estimating the quality of a policy in prior research (Zhang and Feng, 2022b,c; Guo et al., 2022). We measure the quality of a policy by analyzing the proportion of aligned source words received before translating on RWTH De-En alignment dataset7." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 569, + 525, + 745 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 569, + 525, + 745 + ], + "spans": [ + { + "bbox": [ + 302, + 569, + 525, + 745 + ], + "type": "text", + "content": "To ensure accurate word-level alignment calculation, we consider an aligned source word is read before writing a ground truth (GT) target word if the last token of the source word is read before the first token of the target word is written. The results of this analysis are presented in Figure 9. It is observed that word-level policies, both for ITST and Wait-k, exhibit better alignments across most latency settings. This suggests that word-level policies contribute to revising premature WRITE actions by guiding the model to read the remaining tokens of the aligned word, without negatively impacting the model's latency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 303, + 750, + 440, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 750, + 440, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 750, + 440, + 772 + ], + "type": "text", + "content": "7https://www-i6.informatik.rwth-aachen.de/goldAlignment/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15623" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 73, + 267, + 195 + ], + "blocks": [ + { + "bbox": [ + 70, + 73, + 267, + 195 + ], + "lines": [ + { + "bbox": [ + 70, + 73, + 267, + 195 + ], + "spans": [ + { + "bbox": [ + 70, + 73, + 267, + 195 + ], + "type": "image", + "image_path": "96dac2a388686619654ce0211c754f994a329e79389415a4a32950cde31ea028.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 210, + 289, + 234 + ], + "lines": [ + { + "bbox": [ + 67, + 210, + 289, + 234 + ], + "spans": [ + { + "bbox": [ + 67, + 210, + 289, + 234 + ], + "type": "text", + "content": "Figure 9: The percentage of aligned source words read before the translation of the target words started." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 255, + 151, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 255, + 151, + 267 + ], + "spans": [ + { + "bbox": [ + 67, + 255, + 151, + 267 + ], + "type": "text", + "content": "6 Conclusions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 276, + 291, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 276, + 291, + 493 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 291, + 493 + ], + "type": "text", + "content": "This paper explores the potential benefits of word-level operations in SiMT systems. We propose word-level latency calculation to ensure fair and accurate latency comparisons. We introduce a conversion process that transforms token-level policies into word-level policies, enabling the processing of multiple subwords that form a word within a single READ or WRITE action. Additionally, we propose the integration of LM-fused attention, which combines an autoregressive LM model into SiMT models with word-level policies. Experimental results demonstrate the superiority of word-level policies compared to token-level policies, as well as the effectiveness of the LM integration. Our findings highlight the crucial role of word-level policies in the integration process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 502, + 148, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 502, + 148, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 502, + 148, + 515 + ], + "type": "text", + "content": "7 Limitations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 523, + 291, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 523, + 291, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 523, + 291, + 686 + ], + "type": "text", + "content": "While the proposed word-level policy implementation is widely applicable to most existing SiMT systems, it is important to note that systems utilizing languages with a writing style that lacks spaces or other delimiters between words or sentences (e.g., Chinese) are unable to derive benefits from this approach. Furthermore, it is important to consider that while the proposed LM-fused attention proves effective in enhancing translation quality across all latency levels, integrating a large LM may necessitate a faster compute capability to fulfill the low-latency demands of the SiMT task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 709, + 127, + 721 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 709, + 127, + 721 + ], + "spans": [ + { + "bbox": [ + 68, + 709, + 127, + 721 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 728, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 728, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 68, + 728, + 290, + 772 + ], + "type": "text", + "content": "Naveen Arivazhagan, Colin Cherry, Wolfgang Macherey, Chung-Cheng Chiu, Semih Yavuz, Ruoming Pang, Wei Li, and Colin Raffel. 2019. Monotonic infinite lookback attention for simultaneous machine" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 303, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 72, + 525, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 525, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 525, + 117 + ], + "type": "text", + "content": "translation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1313-1323, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 125, + 526, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 526, + 202 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 526, + 202 + ], + "type": "text", + "content": "Ozan Caglayan, Julia Ive, Veneta Haralampieva, Pranava Madhyastha, Loic Barrault, and Lucia Specia. 2020. Simultaneous machine translation with visual context. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2350-2361, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 211, + 526, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 526, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 526, + 299 + ], + "type": "text", + "content": "Mauro Cettolo, Marcello Federico, Luisa Bentivogli, Jan Niehues, Sebastian Stüker, Katsuhito Sudoh, Koichiro Yoshino, and Christian Federmann. 2017. Overview of the IWSLT 2017 evaluation campaign. In Proceedings of the 14th International Conference on Spoken Language Translation, pages 2-14, Tokyo, Japan. International Workshop on Spoken Language Translation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 308, + 525, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 525, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 525, + 342 + ], + "type": "text", + "content": "Chung-Cheng Chiu and Colin Raffel. 2017. Monotonic chunkwise attention. arXiv preprint arXiv:1712.05382." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 351, + 525, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 525, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 525, + 386 + ], + "type": "text", + "content": "Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? arXiv preprint arXiv:1606.02012." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 394, + 525, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 525, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 525, + 471 + ], + "type": "text", + "content": "Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3440-3453, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 480, + 525, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 525, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 525, + 579 + ], + "type": "text", + "content": "Fahim Dalvi, Nadir Durrani, Hassan Sajjad, and Stephan Vogel. 2018. Incremental decoding and training methods for simultaneous translation in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pages 493-499, New Orleans, Louisiana. Association for Computational Linguistics." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 303, + 587, + 525, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 587, + 525, + 687 + ], + "spans": [ + { + "bbox": [ + 303, + 587, + 525, + 687 + ], + "type": "text", + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 303, + 696, + 525, + 740 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 696, + 525, + 740 + ], + "spans": [ + { + "bbox": [ + 303, + 696, + 525, + 740 + ], + "type": "text", + "content": "Maha Elbayad, Laurent Besacier, and Jakob Verbeek. 2020. Efficient Wait-k Models for Simultaneous Machine Translation. In Proc. Interspeech 2020, pages 1461-1465." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 303, + 749, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 303, + 749, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 303, + 749, + 525, + 772 + ], + "type": "text", + "content": "Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor O.K. Li. 2017. Learning to translate in real-time" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15624" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 127 + ], + "type": "text", + "content": "with neural machine translation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pages 1053-1062, Valencia, Spain. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 137, + 291, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 291, + 215 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 291, + 215 + ], + "type": "text", + "content": "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2022. Turning fixed to adaptive: Integrating post-evaluation into simultaneous machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2264–2278, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 225, + 291, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 225, + 291, + 302 + ], + "spans": [ + { + "bbox": [ + 69, + 225, + 291, + 302 + ], + "type": "text", + "content": "Shoutao Guo, Shaolei Zhang, and Yang Feng. 2023. Learning optimal policy for simultaneous machine translation via binary search. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2318-2333, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 312, + 291, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 312, + 291, + 412 + ], + "spans": [ + { + "bbox": [ + 69, + 312, + 291, + 412 + ], + "type": "text", + "content": "Sathish Reddy Indurthi, Mohd Abbas Zaidi, Beomseok Lee, Nikhil Kumar Lakumarapu, and Sangha Kim. 2022. Language model augmented monotonic attention for simultaneous translation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 38-45, Seattle, United States. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 422, + 291, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 422, + 291, + 499 + ], + "spans": [ + { + "bbox": [ + 69, + 422, + 291, + 499 + ], + "type": "text", + "content": "Javier Iranzo Sanchez, Jorge Civera, and Alfons Juan-Ciscar. 2022. From simultaneous to streaming machine translation by leveraging streaming history. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6972-6985, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 509, + 291, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 509, + 291, + 597 + ], + "spans": [ + { + "bbox": [ + 69, + 509, + 291, + 597 + ], + "type": "text", + "content": "Patrick Kahardipraja, Brielen Madureira, and David Schlangen. 2021. Towards incremental transformers: An empirical analysis of transformer models for incremental NLU. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1178-1189, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 607, + 291, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 607, + 291, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 607, + 291, + 685 + ], + "type": "text", + "content": "Yunsu Kim, Yingbo Gao, and Hermann Ney. 2019. Effective cross-lingual transfer of neural machine translation models without shared vocabularies. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1246–1257, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 291, + 772 + ], + "type": "text", + "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 66-71, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 526, + 772 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 525, + 105 + ], + "type": "text", + "content": "Guillaume Lample and Alexis Conneau. 2019. Crosslingual language model pretraining. arXiv preprint arXiv:1901.07291." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 112, + 526, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 112, + 526, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 112, + 526, + 211 + ], + "type": "text", + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 218, + 526, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 526, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 526, + 351 + ], + "type": "text", + "content": "Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettle-moyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, and Xian Li. 2022. Few-shot learning with multilingual generative language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9019-9052, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 358, + 526, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 526, + 468 + ], + "type": "text", + "content": "Xin Liu, Baosong Yang, Dayiheng Liu, Haibo Zhang, Weihua Luo, Min Zhang, Haiying Zhang, and Jinsong Su. 2021a. Bridging subword gaps in pretrainfinetune paradigm for natural language generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 6001-6011, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 475, + 525, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 525, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 525, + 552 + ], + "type": "text", + "content": "Xuebo Liu, Longyue Wang, Derek F. Wong, Liang Ding, Lidia S. Chao, Shuming Shi, and Zhaopeng Tu. 2021b. On the copying behaviors of pre-training for neural machine translation. In Findings of the Association for Computational Linguistics: ACLIJCNLP 2021, pages 4265-4275, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 559, + 525, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 525, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 525, + 624 + ], + "type": "text", + "content": "Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 632, + 525, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 632, + 525, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 632, + 525, + 688 + ], + "type": "text", + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 694, + 525, + 772 + ], + "type": "text", + "content": "Mingbo Ma, Liang Huang, Hao Xiong, Renjie Zheng, Kaibo Liu, Baigong Zheng, Chuanqiang Zhang, Zhongjun He, Hairong Liu, Xing Li, Hua Wu, and Haifeng Wang. 2019a. STACL: Simultaneous translation with implicit anticipation and controllable latency using prefix-to-prefix framework. In Proceedings of the 57th Annual Meeting of the Association for" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "15625" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 291, + 772 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 80, + 72, + 291, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 291, + 105 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 291, + 105 + ], + "type": "text", + "content": "Computational Linguistics, pages 3025-3036, Florence, Italy. Association for Computational Linguistics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 113, + 290, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 113, + 290, + 146 + ], + "spans": [ + { + "bbox": [ + 69, + 113, + 290, + 146 + ], + "type": "text", + "content": "Xutai Ma, Juan Pino, James Cross, Liezl Puzon, and Jiatao Gu. 2019b. Monotonic multihead attention. arXiv preprint arXiv:1909.12406." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 153, + 290, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 153, + 290, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 153, + 290, + 208 + ], + "type": "text", + "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 215, + 290, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 215, + 290, + 249 + ], + "spans": [ + { + "bbox": [ + 69, + 215, + 290, + 249 + ], + "type": "text", + "content": "Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. 2018. Improving language understanding by generative pre-training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 255, + 290, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 255, + 290, + 300 + ], + "spans": [ + { + "bbox": [ + 69, + 255, + 290, + 300 + ], + "type": "text", + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 306, + 290, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 306, + 290, + 383 + ], + "spans": [ + { + "bbox": [ + 69, + 306, + 290, + 383 + ], + "type": "text", + "content": "Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss, and Douglas Eck. 2017. Online and linear-time attention by enforcing monotonic alignments. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 2837-2846. PMLR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 391, + 290, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 391, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 69, + 391, + 290, + 469 + ], + "type": "text", + "content": "Shoetsu Sato, Jin Sakuma, Naoki Yoshinaga, Masashi Toyoda, and Masaru Kitsuregawa. 2020. Vocabulary adaptation for domain adaptation in neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 4269-4279, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 475, + 290, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 475, + 290, + 542 + ], + "spans": [ + { + "bbox": [ + 69, + 475, + 290, + 542 + ], + "type": "text", + "content": "Felix Schneider and Alexander Waibel. 2020. Towards stream translation: Adaptive computation time for simultaneous machine translation. In Proceedings of the 17th International Conference on Spoken Language Translation, pages 228-236, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 548, + 290, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 548, + 290, + 626 + ], + "spans": [ + { + "bbox": [ + 69, + 548, + 290, + 626 + ], + "type": "text", + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715-1725, Berlin, Germany. Association for Computational Linguistics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 632, + 290, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 632, + 290, + 688 + ], + "spans": [ + { + "bbox": [ + 69, + 632, + 290, + 688 + ], + "type": "text", + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "spans": [ + { + "bbox": [ + 69, + 694, + 290, + 772 + ], + "type": "text", + "content": "Shushu Wang, Jing Wu, Kai Fan, Wei Luo, Jun Xiao, and Zhongqiang Huang. 2023. Better simultaneous translation with monotonic knowledge distillation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2334-2349, Toronto, Canada. Association for Computational Linguistics." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 72, + 525, + 772 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 305, + 72, + 525, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 525, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 525, + 160 + ], + "type": "text", + "content": "Wenxuan Wang, Wenxiang Jiao, Yongchang Hao, Xing Wang, Shuming Shi, Zhaopeng Tu, and Michael Lyu. 2022. Understanding and improving sequence-to-sequence pretraining for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2591-2600, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 170, + 525, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 525, + 226 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 525, + 226 + ], + "type": "text", + "content": "Rongxiang Weng, Heng Yu, Shujian Huang, Shanbo Cheng, and Weihua Luo. 2020. Acquiring knowledge from pre-trained model to neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9266-9273." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 236, + 525, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 525, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 525, + 291 + ], + "type": "text", + "content": "Rongxiang Weng, Heng Yu, Weihua Luo, and Min Zhang. 2022. Deep fusing pre-trained models into neural machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11468-11476." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 301, + 525, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 525, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 525, + 390 + ], + "type": "text", + "content": "Haoran Xu, Benjamin Van Durme, and Kenton Murray. 2021. BERT, mBERT, or BiBERT? a study on contextualized embeddings for neural machine translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6663-6675, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 400, + 525, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 525, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 525, + 465 + ], + "type": "text", + "content": "Jiacheng Yang, Mingxuan Wang, Hao Zhou, Chengqi Zhao, Weinan Zhang, Yong Yu, and Lei Li. 2020. Towards making the most of bert in neural machine translation. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9378-9385." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 476, + 525, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 525, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 525, + 553 + ], + "type": "text", + "content": "Shaolei Zhang and Yang Feng. 2021. Universal simultaneous machine translation with mixture-of-experts wait-k policy. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7306-7317, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 564, + 525, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 525, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 525, + 619 + ], + "type": "text", + "content": "Shaolei Zhang and Yang Feng. 2022a. Gaussian multihead attention for simultaneous machine translation. In *Findings of the Association for Computational Linguistics: ACL 2022*, pages 3019–3030, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 525, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 525, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 525, + 695 + ], + "type": "text", + "content": "Shaolei Zhang and Yang Feng. 2022b. Information-transport-based policy for simultaneous translation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 992-1013, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "spans": [ + { + "bbox": [ + 304, + 706, + 525, + 772 + ], + "type": "text", + "content": "Shaolei Zhang and Yang Feng. 2022c. Modeling dual read/write paths for simultaneous machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2461-2477, Dublin, Ireland. Association for Computational Linguistics." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "15626" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 290, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 290, + 105 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 290, + 105 + ], + "type": "text", + "content": "Shaolei Zhang and Yang Feng. 2023. Hidden markov transformer for simultaneous machine translation. arXiv preprint arXiv:2303.00257." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 112, + 291, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 112, + 291, + 168 + ], + "spans": [ + { + "bbox": [ + 68, + 112, + 291, + 168 + ], + "type": "text", + "content": "Shaolei Zhang, Yang Feng, and Liangyou Li. 2021. Future-guided incremental transformer for simultaneous translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14428-14436." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 175, + 291, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 175, + 291, + 253 + ], + "spans": [ + { + "bbox": [ + 68, + 175, + 291, + 253 + ], + "type": "text", + "content": "Shaolei Zhang, Shoutao Guo, and Yang Feng. 2022. Wait-info policy: Balancing source and target at information level for simultaneous machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 2249-2263, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 260, + 291, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 260, + 291, + 338 + ], + "spans": [ + { + "bbox": [ + 68, + 260, + 291, + 338 + ], + "type": "text", + "content": "Baigong Zheng, Kaibo Liu, Renjie Zheng, Mingbo Ma, Hairong Liu, and Liang Huang. 2020. Simultaneous translation policies: From fixed to adaptive. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 2847-2853, Online. Association for Computational Linguistics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 344, + 291, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 344, + 291, + 390 + ], + "spans": [ + { + "bbox": [ + 68, + 344, + 291, + 390 + ], + "type": "text", + "content": "Jinhua Zhu, Yingce Xia, Lijun Wu, Di He, Tao Qin, Wengang Zhou, Houqiang Li, and Tie-Yan Liu. 2020. Incorporating bert into neural machine translation. arXiv preprint arXiv:2002.06823." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 397, + 142, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 397, + 142, + 411 + ], + "spans": [ + { + "bbox": [ + 68, + 397, + 142, + 411 + ], + "type": "text", + "content": "A Appendix" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 418, + 289, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 418, + 289, + 444 + ], + "spans": [ + { + "bbox": [ + 68, + 418, + 289, + 444 + ], + "type": "text", + "content": "A.1 Token-level latency evaluation in previous works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 449, + 291, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 449, + 291, + 719 + ], + "spans": [ + { + "bbox": [ + 67, + 449, + 291, + 719 + ], + "type": "text", + "content": "Many previous works in the field of SiMT have employed token-level latency calculation as a fundamental metric for evaluating system performance. In the first group of papers, which includes (Arivazhagan et al., 2019; Ma et al., 2019b; Schneider and Waibel, 2020; Zheng et al., 2020; Zhang and Feng, 2022b), researchers explicitly stated their use of token-level latency calculation or incorporated token-level scores in their analyses. On the other hand, the second group, comprising (Zhang and Feng, 2022b; Elbayad et al., 2020; Zhang and Feng, 2022c, 2021, 2022a; Zhang et al., 2022; Guo et al., 2022; Zhang and Feng, 2023; Guo et al., 2023; Caglayan et al., 2020; Wang et al., 2023), can be identified by their publicly available official implementations, which incorporate token-level latency evaluation code. This underscores the prevalence of token-level latency calculation in assessing the effectiveness of SiMT systems in the cutting-edge SiMT research works." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 729, + 197, + 741 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 729, + 197, + 741 + ], + "spans": [ + { + "bbox": [ + 68, + 729, + 197, + 741 + ], + "type": "text", + "content": "A.2 More training details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 746, + 291, + 772 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 746, + 291, + 772 + ], + "spans": [ + { + "bbox": [ + 67, + 746, + 291, + 772 + ], + "type": "text", + "content": "For Transformer model configurations, we follow the Efficient Wait-k (Elbayad et al., 2020)'s set" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 302, + 70, + 524, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 70, + 524, + 96 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 524, + 96 + ], + "type": "text", + "content": "tings." + }, + { + "bbox": [ + 302, + 70, + 524, + 96 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 302, + 70, + 524, + 96 + ], + "type": "text", + "content": " The details of hyperparameter settings for each model is in Table 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 302, + 98, + 526, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 98, + 526, + 218 + ], + "spans": [ + { + "bbox": [ + 302, + 98, + 526, + 218 + ], + "type": "text", + "content": "To ensure optimal performance of the ITST models, we conducted multiple training runs for each model configuration and dataset. We observed performance fluctuations across the training runs. To select the best-performing models for both token-level and word-level policies, we repeated the training process three times and selected the checkpoint with the lowest validation loss for each configuration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 302, + 220, + 525, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 220, + 525, + 260 + ], + "spans": [ + { + "bbox": [ + 302, + 220, + 525, + 260 + ], + "type": "text", + "content": "For LM integrated models, the intra-word bidirectional encoding was not applied to the LM to prevent the need for fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 302, + 269, + 453, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 269, + 453, + 282 + ], + "spans": [ + { + "bbox": [ + 302, + 269, + 453, + 282 + ], + "type": "text", + "content": "A.3 Examples and discussions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 301, + 286, + 526, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 286, + 526, + 544 + ], + "spans": [ + { + "bbox": [ + 301, + 286, + 526, + 544 + ], + "type": "text", + "content": "In Table 2, an illustrative case is presented to demonstrate the distinction between a token-level policy (token-level Wait-1) and a word-level policy (word-level Wait-1). At Step 2, the token-level model predicts the target token “B” after processing the subword “B” for the word “Beine” (which means “leg” in German). Subsequently, it fails to recover from this incorrect prediction and continues by predicting “ody_” as the next token, even after processing the remaining token “eine_”. In contrast, the word-level model accurately predicts “legs” (encoded as “leg s_”) after processing the complete word ‘Beine’ (encoded as ‘B eine_’). Furthermore, the token model also makes erroneous predictions for all steps when processing the word ‘blutüberströmmt.’ (encoded as ‘bl ut über ström t_’, meaning “covered in blood” in German), while the word-level model accurately predicts a correct word in a single step." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 302, + 544, + 525, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 544, + 525, + 652 + ], + "spans": [ + { + "bbox": [ + 302, + 544, + 525, + 652 + ], + "type": "text", + "content": "Another case can be observed in Table 3, where the token-level model initiates the translation before fully reading the word \"Irgendetwas\" (encoded as \"Irgen det was\"). As a consequence, it produces an incorrect translation. On the other hand, the word-level model accurately translates the sentence by processing the complete word before making any predictions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 302, + 661, + 430, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 661, + 430, + 673 + ], + "spans": [ + { + "bbox": [ + 302, + 661, + 430, + 673 + ], + "type": "text", + "content": "A.4 Table of main results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 302, + 679, + 440, + 692 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 679, + 440, + 692 + ], + "spans": [ + { + "bbox": [ + 302, + 679, + 440, + 692 + ], + "type": "text", + "content": "A.5 Graphs of main results" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 760, + 450, + 772 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 760, + 450, + 772 + ], + "spans": [ + { + "bbox": [ + 316, + 760, + 450, + 772 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 316, + 760, + 450, + 772 + ], + "type": "text", + "content": "https://github.com/elbayadm/attn2d" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15627" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 89, + 492, + 309 + ], + "blocks": [ + { + "bbox": [ + 102, + 89, + 492, + 309 + ], + "lines": [ + { + "bbox": [ + 102, + 89, + 492, + 309 + ], + "spans": [ + { + "bbox": [ + 102, + 89, + 492, + 309 + ], + "type": "table", + "html": "
Hyperparametertransformer-smalltransformer-basetransformer-big
attention heads4816
embedding dime2565121024
ffn embedding dim102420484096
dropout0.30.30.3
encoder layers666
decoder layers666
lr5e-45e-45e-4
lr schedulerinverse sqrtinverse sqrtinverse sqrt
optimizerAdam (0.9, 0.98)Adam (0.9, 0.98)Adam (0.9, 0.98)
clip-norm000
warmup-updates400040004000
warmup-init-lr1e-71e-71e-7
weight decay000
label smoothing0.10.10.1
max tokens327685242881048576
", + "image_path": "6f02bb8f188a8cb248e2712614dfedcb4bf155f86a411b5a423c84eac5ad4cd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 377, + 511, + 513 + ], + "blocks": [ + { + "bbox": [ + 185, + 317, + 406, + 331 + ], + "lines": [ + { + "bbox": [ + 185, + 317, + 406, + 331 + ], + "spans": [ + { + "bbox": [ + 185, + 317, + 406, + 331 + ], + "type": "text", + "content": "Table 1: Hyperparameters of each Transformer model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 377, + 511, + 513 + ], + "lines": [ + { + "bbox": [ + 83, + 377, + 511, + 513 + ], + "spans": [ + { + "bbox": [ + 83, + 377, + 511, + 513 + ], + "type": "table", + "html": "
Src:Meine_B eine_waren_bl ut über ström t.
Ref:My leg s_were Covered_in_blood.
StepStreaminig InputToken Wait-1 OutputWord Wait-1 Output
1MeineMyMy
2Meine_BB
3Meine_B eineodyleg s
4Meine_B eine_warenwaswere
5Meine_B eine_waren_blblue
6Meine_B eine_waren_bl utand
7Meine_B eine_waren_bl ut übermy
8Meine_B eine_waren_bl ut über strömleg
9Meine_B eine_waren_bl ut über ström ts
10Meine_B eine_waren_bl ut über ström t.were_blood - ri dden.blood dy.
", + "image_path": "8868b7b9646288d5fad9842cddb3ff3a7faf811aaff0945f1e1cafc0592865e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 592, + 511, + 717 + ], + "blocks": [ + { + "bbox": [ + 67, + 520, + 525, + 544 + ], + "lines": [ + { + "bbox": [ + 67, + 520, + 525, + 544 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 525, + 544 + ], + "type": "text", + "content": "Table 2: An example case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.56, word-AL: 1.69) and Word-level Wait-1 (token-AL: 1.44, word-AL: 1.0)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 592, + 511, + 717 + ], + "lines": [ + { + "bbox": [ + 83, + 592, + 511, + 717 + ], + "spans": [ + { + "bbox": [ + 83, + 592, + 511, + 717 + ], + "type": "table", + "html": "
Src:Ir gen det was lag in der Luft .
Ref:Some thing was up .
StepStreaming InputToken Wait-1 OutputWord Wait-1 Output
1IrIr
2Ir gengen
3Ir gen detgen
4Ir gen det wasgenSome thing_
5Ir gen det was laggenlay_
6Ir gen det was lag insin_
7Ir gen det was lag in der,the_
8Ir gen det was lag in der Luftin
9Ir gen det was lag in der Luft .fact , was in the air .air .
", + "image_path": "83390317b18088f0dcb2c19432051eafb4888477fbd11093c859a791f7a32385.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 724, + 525, + 749 + ], + "lines": [ + { + "bbox": [ + 67, + 724, + 525, + 749 + ], + "spans": [ + { + "bbox": [ + 67, + 724, + 525, + 749 + ], + "type": "text", + "content": "Table 3: An Eexample case from WMT De.En test set (transformer-big) for Token-level Wait-1 (token-AL: 2.69, word-AL: 4.25) and Word-level Wait-1 (token-AL: 2.38, word-AL: 1.0)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "15628" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 105, + 217, + 217 + ], + "blocks": [ + { + "bbox": [ + 70, + 105, + 217, + 217 + ], + "lines": [ + { + "bbox": [ + 70, + 105, + 217, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 105, + 217, + 217 + ], + "type": "image", + "image_path": "a7c2f1f08eb9e88dd9ef08d785d6c4d82ff39d39211e8d4a400c58abdefa40f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 220, + 207, + 231 + ], + "lines": [ + { + "bbox": [ + 78, + 220, + 207, + 231 + ], + "spans": [ + { + "bbox": [ + 78, + 220, + 207, + 231 + ], + "type": "text", + "content": "(a) De-En Wait-k Transformer base" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 105, + 369, + 217 + ], + "blocks": [ + { + "bbox": [ + 224, + 105, + 369, + 217 + ], + "lines": [ + { + "bbox": [ + 224, + 105, + 369, + 217 + ], + "spans": [ + { + "bbox": [ + 224, + 105, + 369, + 217 + ], + "type": "image", + "image_path": "84c13e68369add4451739ed4985cf2fb0a7369e75084a4fb33c26288da6bc039.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 220, + 358, + 231 + ], + "lines": [ + { + "bbox": [ + 235, + 220, + 358, + 231 + ], + "spans": [ + { + "bbox": [ + 235, + 220, + 358, + 231 + ], + "type": "text", + "content": "(b) De-En MoE Transformer base" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 241, + 473, + 253 + ], + "lines": [ + { + "bbox": [ + 119, + 241, + 473, + 253 + ], + "spans": [ + { + "bbox": [ + 119, + 241, + 473, + 253 + ], + "type": "text", + "content": "Figure 10: Results of translation quality v.s. word-level AL, De-En transformer-base." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 105, + 523, + 217 + ], + "blocks": [ + { + "bbox": [ + 378, + 105, + 523, + 217 + ], + "lines": [ + { + "bbox": [ + 378, + 105, + 523, + 217 + ], + "spans": [ + { + "bbox": [ + 378, + 105, + 523, + 217 + ], + "type": "image", + "image_path": "cd652d3f4e088a493be32d365fa0f641cc4dc5ab82a6105e750d9829cf91bf6b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 220, + 512, + 231 + ], + "lines": [ + { + "bbox": [ + 389, + 220, + 512, + 231 + ], + "spans": [ + { + "bbox": [ + 389, + 220, + 512, + 231 + ], + "type": "text", + "content": "(c) De-En ITST Transformer base" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 70, + 333, + 215, + 444 + ], + "blocks": [ + { + "bbox": [ + 70, + 333, + 215, + 444 + ], + "lines": [ + { + "bbox": [ + 70, + 333, + 215, + 444 + ], + "spans": [ + { + "bbox": [ + 70, + 333, + 215, + 444 + ], + "type": "image", + "image_path": "0e4a8516a1f8797398d67d483ca155986b84acc7cc5ec468050494c9897920fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 224, + 333, + 370, + 444 + ], + "blocks": [ + { + "bbox": [ + 224, + 333, + 370, + 444 + ], + "lines": [ + { + "bbox": [ + 224, + 333, + 370, + 444 + ], + "spans": [ + { + "bbox": [ + 224, + 333, + 370, + 444 + ], + "type": "image", + "image_path": "0ada3edc3344d0d06ec3dbf9bd85e2fb68caf26a3742abbb66fb233aa6b1dbe3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 378, + 333, + 523, + 444 + ], + "blocks": [ + { + "bbox": [ + 378, + 333, + 523, + 444 + ], + "lines": [ + { + "bbox": [ + 378, + 333, + 523, + 444 + ], + "spans": [ + { + "bbox": [ + 378, + 333, + 523, + 444 + ], + "type": "image", + "image_path": "3f5c144dc9fe3d0c40307c39c86dab14519469c5de082683a35f9b04a5c995d2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 70, + 460, + 215, + 571 + ], + "blocks": [ + { + "bbox": [ + 77, + 449, + 207, + 459 + ], + "lines": [ + { + "bbox": [ + 77, + 449, + 207, + 459 + ], + "spans": [ + { + "bbox": [ + 77, + 449, + 207, + 459 + ], + "type": "text", + "content": "(a) En-Fr Wait-k Transformer small" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 460, + 215, + 571 + ], + "lines": [ + { + "bbox": [ + 70, + 460, + 215, + 571 + ], + "spans": [ + { + "bbox": [ + 70, + 460, + 215, + 571 + ], + "type": "image", + "image_path": "323f89b5334a770fa7032530970985fae90748ef9da6d7a287e0d0b646cc03df.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 224, + 460, + 370, + 571 + ], + "blocks": [ + { + "bbox": [ + 235, + 449, + 358, + 459 + ], + "lines": [ + { + "bbox": [ + 235, + 449, + 358, + 459 + ], + "spans": [ + { + "bbox": [ + 235, + 449, + 358, + 459 + ], + "type": "text", + "content": "(b) En-Fr MoE Transformer small" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 460, + 370, + 571 + ], + "lines": [ + { + "bbox": [ + 224, + 460, + 370, + 571 + ], + "spans": [ + { + "bbox": [ + 224, + 460, + 370, + 571 + ], + "type": "image", + "image_path": "aafe8835c201e63e967ad66c630b512e34ecebeabb9c769a7b3549bc2c37a95f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 378, + 460, + 523, + 571 + ], + "blocks": [ + { + "bbox": [ + 389, + 449, + 512, + 459 + ], + "lines": [ + { + "bbox": [ + 389, + 449, + 512, + 459 + ], + "spans": [ + { + "bbox": [ + 389, + 449, + 512, + 459 + ], + "type": "text", + "content": "(c) En-Fr ITST Transformer small" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 460, + 523, + 571 + ], + "lines": [ + { + "bbox": [ + 378, + 460, + 523, + 571 + ], + "spans": [ + { + "bbox": [ + 378, + 460, + 523, + 571 + ], + "type": "image", + "image_path": "eee39750e797615ae83ba316087c8f4da0123f2576d70ffd41df469544822c36.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 70, + 585, + 215, + 696 + ], + "blocks": [ + { + "bbox": [ + 78, + 575, + 207, + 584 + ], + "lines": [ + { + "bbox": [ + 78, + 575, + 207, + 584 + ], + "spans": [ + { + "bbox": [ + 78, + 575, + 207, + 584 + ], + "type": "text", + "content": "(d) De-En Wait-k Transformer base" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 585, + 215, + 696 + ], + "lines": [ + { + "bbox": [ + 70, + 585, + 215, + 696 + ], + "spans": [ + { + "bbox": [ + 70, + 585, + 215, + 696 + ], + "type": "image", + "image_path": "cbf2f32a207e488330d3e6161e6f6e583f398e48e59c8ea6d91e5c204171cc12.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 700, + 205, + 711 + ], + "lines": [ + { + "bbox": [ + 80, + 700, + 205, + 711 + ], + "spans": [ + { + "bbox": [ + 80, + 700, + 205, + 711 + ], + "type": "text", + "content": "(g) De-En Wait-k Transformer big" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 174, + 720, + 418, + 733 + ], + "lines": [ + { + "bbox": [ + 174, + 720, + 418, + 733 + ], + "spans": [ + { + "bbox": [ + 174, + 720, + 418, + 733 + ], + "type": "text", + "content": "Figure 11: Results of translation quality v.s. token-level AL." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 224, + 585, + 370, + 696 + ], + "blocks": [ + { + "bbox": [ + 236, + 575, + 358, + 584 + ], + "lines": [ + { + "bbox": [ + 236, + 575, + 358, + 584 + ], + "spans": [ + { + "bbox": [ + 236, + 575, + 358, + 584 + ], + "type": "text", + "content": "(e) De-En MoE Transformer base" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 585, + 370, + 696 + ], + "lines": [ + { + "bbox": [ + 224, + 585, + 370, + 696 + ], + "spans": [ + { + "bbox": [ + 224, + 585, + 370, + 696 + ], + "type": "image", + "image_path": "21ab652590c28f0f7a7cebd049f50ff4ba38d82570f5d7c63b42d018b98f631c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 700, + 355, + 711 + ], + "lines": [ + { + "bbox": [ + 238, + 700, + 355, + 711 + ], + "spans": [ + { + "bbox": [ + 238, + 700, + 355, + 711 + ], + "type": "text", + "content": "(h) De-En MoE Transformer big" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 378, + 585, + 523, + 696 + ], + "blocks": [ + { + "bbox": [ + 390, + 575, + 512, + 584 + ], + "lines": [ + { + "bbox": [ + 390, + 575, + 512, + 584 + ], + "spans": [ + { + "bbox": [ + 390, + 575, + 512, + 584 + ], + "type": "text", + "content": "(f) De-En ITST Transformer base" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 378, + 585, + 523, + 696 + ], + "lines": [ + { + "bbox": [ + 378, + 585, + 523, + 696 + ], + "spans": [ + { + "bbox": [ + 378, + 585, + 523, + 696 + ], + "type": "image", + "image_path": "177f24907f5fd2547d3881b704fdf7ff430fdffd2d5f4ce05a256029d5feb0dd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 700, + 509, + 711 + ], + "lines": [ + { + "bbox": [ + 392, + 700, + 509, + 711 + ], + "spans": [ + { + "bbox": [ + 392, + 700, + 509, + 711 + ], + "type": "text", + "content": "(i) De-En ITST Transformer big" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "15629" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 230, + 265, + 575 + ], + "blocks": [ + { + "bbox": [ + 94, + 230, + 265, + 575 + ], + "lines": [ + { + "bbox": [ + 94, + 230, + 265, + 575 + ], + "spans": [ + { + "bbox": [ + 94, + 230, + 265, + 575 + ], + "type": "table", + "html": "
Token-level Wait-k
ktoken ALword ALBLEU
11.951.9132.32
33.833.4338.90
55.674.8240.68
77.516.2340.95
99.287.5841.06
1110.918.8241.14
1312.399.9541.12
Word-level Wait-k
ktoken ALword ALBLEU
12.241.7432.43
34.183.3239.36
56.555.1541.09
78.856.9441.48
910.928.5741.57
1112.759.9841.58
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.181.7433.47
34.233.3740.83
56.585.1942.78
78.866.9743.00
910.948.5843.08
1112.7610.0043.09
", + "image_path": "f45195fa04605265623e7451aab3c7af80b050b8d9992cc96066d26e898704ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 329, + 230, + 499, + 574 + ], + "blocks": [ + { + "bbox": [ + 67, + 582, + 289, + 606 + ], + "lines": [ + { + "bbox": [ + 67, + 582, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 582, + 289, + 606 + ], + "type": "text", + "content": "Table 4: Main results of transformer-small Waik-k models on IWLST17 En.Fr dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 230, + 499, + 574 + ], + "lines": [ + { + "bbox": [ + 329, + 230, + 499, + 574 + ], + "spans": [ + { + "bbox": [ + 329, + 230, + 499, + 574 + ], + "type": "table", + "html": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.901.8832.20
33.833.4238.92
55.694.8440.51
77.536.2540.83
99.317.6040.92
1110.948.8540.95
1312.429.9740.96
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.011.6032.01
34.183.3239.28
56.575.1740.77
78.866.9541.03
910.948.5841.11
1112.769.9941.10
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.171.7233.01
34.223.3740.46
56.565.1742.35
78.846.9542.61
910.918.5742.66
1112.749.9942.67
", + "image_path": "5f1e40807cbfee63a2bc86b703237032e584713f25b2da20279d617de819fd42.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 302, + 582, + 524, + 606 + ], + "lines": [ + { + "bbox": [ + 302, + 582, + 524, + 606 + ], + "spans": [ + { + "bbox": [ + 302, + 582, + 524, + 606 + ], + "type": "text", + "content": "Table 5: Main results of transformer-small MoE Waik-k models on IWSLT17 En.Fr dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 313, + 791 + ], + "type": "text", + "content": "15630" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 155, + 268, + 648 + ], + "blocks": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "lines": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "spans": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "type": "table", + "html": "
Token-level ITST
δtoken ALword ALBLEU
0.23.333.0537.81
0.33.593.2538.64
0.44.143.7039.40
0.55.184.5940.19
0.66.715.9240.84
0.78.627.6341.27
0.759.568.4541.34
0.810.729.4141.42
0.8512.1110.5341.42
0.913.8411.8041.42
Word-level ITST
δtoken ALword ALBLEU
0.23.272.6437.25
0.33.462.8037.88
0.43.913.2138.69
0.54.673.9139.57
0.65.814.9640.44
0.77.296.2840.94
0.758.217.0541.13
0.89.418.0241.36
0.8511.199.3741.50
0.913.3611.0041.53
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.252.6138.30
0.33.302.6538.75
0.43.482.8139.61
0.53.943.2340.50
0.64.804.0441.35
0.76.025.1941.88
0.756.765.8642.09
0.87.726.6742.30
0.858.997.6842.56
0.910.909.1642.63
", + "image_path": "4cf5870a862751b5063aeb22225ccc9ed6ed18e72a9934510e3b4f2484cde731.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 329, + 216, + 499, + 588 + ], + "blocks": [ + { + "bbox": [ + 67, + 657, + 289, + 681 + ], + "lines": [ + { + "bbox": [ + 67, + 657, + 289, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 657, + 289, + 681 + ], + "type": "text", + "content": "Table 6: Main results of transformer-small ITST models on IWSLT17 En.Fr dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 216, + 499, + 588 + ], + "lines": [ + { + "bbox": [ + 329, + 216, + 499, + 588 + ], + "spans": [ + { + "bbox": [ + 329, + 216, + 499, + 588 + ], + "type": "table", + "html": "
Token-level Wait-k
ktoken ALword ALBLEU
10.520.7517.63
32.161.9123.24
53.923.0526.14
75.774.2327.46
97.765.5028.77
119.646.6929.37
1311.497.8629.88
1513.188.9229.89
1714.799.9330.19
Word-level Wait-k
ktoken ALword ALBLEU
12.961.8424.02
35.333.4328.08
58.085.2329.35
710.827.0129.87
913.478.7230.17
1115.8810.2530.38
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.961.8424.71
35.273.3828.49
58.055.2029.95
710.817.0030.56
913.448.6930.74
1115.8510.2430.95
", + "image_path": "de2dd378cf60b689e0f8a55f8a4a3cc518a59e6b6fffed493b522eb8737e3a14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 302, + 596, + 524, + 619 + ], + "lines": [ + { + "bbox": [ + 302, + 596, + 524, + 619 + ], + "spans": [ + { + "bbox": [ + 302, + 596, + 524, + 619 + ], + "type": "text", + "content": "Table 7: Main results of transformer-base Waik-k models on WMT15 De.En dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 311, + 791 + ], + "type": "text", + "content": "15631" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 216, + 264, + 367 + ], + "blocks": [ + { + "bbox": [ + 94, + 216, + 264, + 367 + ], + "lines": [ + { + "bbox": [ + 94, + 216, + 264, + 367 + ], + "spans": [ + { + "bbox": [ + 94, + 216, + 264, + 367 + ], + "type": "table", + "html": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
11.171.1117.28
32.502.0923.29
54.173.1825.95
76.004.3627.86
97.815.5128.81
119.626.6629.53
1311.427.8029.66
1513.158.8929.81
1714.809.9129.79
", + "image_path": "a0a4c3007bbf594f661970f89be975077d28c64a72f69d6260fc80e77028437c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 94, + 368, + 264, + 477 + ], + "blocks": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "lines": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "spans": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "type": "table", + "html": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
13.121.9123.59
35.293.3927.75
58.035.1929.29
710.796.9830.14
913.388.6530.39
1115.8110.230.6
", + "image_path": "80a3f5ca7a97da89b89c9198454ce5e1a8d09b5641e6ca5b3d633faa4e98611a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 94, + 478, + 264, + 587 + ], + "blocks": [ + { + "bbox": [ + 94, + 478, + 264, + 587 + ], + "lines": [ + { + "bbox": [ + 94, + 478, + 264, + 587 + ], + "spans": [ + { + "bbox": [ + 94, + 478, + 264, + 587 + ], + "type": "table", + "html": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
13.081.9224.97
35.323.4428.66
58.085.2330.21
710.877.0430.63
913.478.7330.9
1115.8810.2631.0
", + "image_path": "c20796ce6611dc73a8c4ffa97d571256d860ad733c34cfe6aca3c164e60769d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 324, + 156, + 503, + 321 + ], + "blocks": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "lines": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "type": "text", + "content": "Table 8: Main results of transformer-base MoE Waik-k models on WMT15 De.En dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 156, + 503, + 321 + ], + "lines": [ + { + "bbox": [ + 324, + 156, + 503, + 321 + ], + "spans": [ + { + "bbox": [ + 324, + 156, + 503, + 321 + ], + "type": "table", + "html": "
Token-level ITST
δtoken ALword ALBLEU
0.22.452.0822.19
0.32.892.3824.15
0.43.822.9626.21
0.55.233.9228.19
0.67.285.5029.25
0.79.877.4829.47
0.7511.668.8129.76
0.813.7910.3829.79
0.8516.6012.1129.90
0.919.8113.8529.71
", + "image_path": "2fe04fde2cfb6d0d0088f385c3c2e60a7a4009023f7f335f64a116078aa5b797.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 324, + 322, + 503, + 483 + ], + "blocks": [ + { + "bbox": [ + 324, + 322, + 503, + 483 + ], + "lines": [ + { + "bbox": [ + 324, + 322, + 503, + 483 + ], + "spans": [ + { + "bbox": [ + 324, + 322, + 503, + 483 + ], + "type": "table", + "html": "
Word-level ITST
δtoken ALword ALBLEU
0.24.132.6426.81
0.34.192.6927.01
0.44.392.8127.47
0.55.093.3028.04
0.66.434.2828.87
0.78.605.9329.54
0.759.916.9229.54
0.811.548.0929.67
0.8513.839.6129.82
0.916.7511.3829.93
", + "image_path": "45b1f2f452d95bc1ae4cfd6baf626f680655c71f178cce3c6bf282c890649ca6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 485, + 503, + 648 + ], + "blocks": [ + { + "bbox": [ + 324, + 485, + 503, + 648 + ], + "lines": [ + { + "bbox": [ + 324, + 485, + 503, + 648 + ], + "spans": [ + { + "bbox": [ + 324, + 485, + 503, + 648 + ], + "type": "table", + "html": "
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.24.152.6427.74
0.34.242.7027.99
0.44.552.9228.44
0.55.333.4929.15
0.66.774.6329.97
0.78.576.0830.49
0.759.516.8030.53
0.810.717.6730.60
0.8512.328.7230.68
0.914.289.9330.71
", + "image_path": "e6a677d3aa2464e7579de5bfac355e06e292e3c7e02df482165b6c9f5575b58c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 657, + 525, + 681 + ], + "lines": [ + { + "bbox": [ + 302, + 657, + 525, + 681 + ], + "spans": [ + { + "bbox": [ + 302, + 657, + 525, + 681 + ], + "type": "text", + "content": "Table 9: Main results of transformer-base ITST models on WMT15 De.En dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15632" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 216, + 264, + 368 + ], + "blocks": [ + { + "bbox": [ + 94, + 216, + 264, + 368 + ], + "lines": [ + { + "bbox": [ + 94, + 216, + 264, + 368 + ], + "spans": [ + { + "bbox": [ + 94, + 216, + 264, + 368 + ], + "type": "table", + "html": "
Token-level Wait-k
ktoken ALword ALBLEU
1-0.130.3919.20
31.801.7124.01
53.742.9526.86
75.814.2628.86
97.805.5329.96
119.726.7430.62
1311.517.8730.91
1513.208.9531.16
1714.849.9731.31
", + "image_path": "8341018f2ceaecb0a4c02bccbb7234fa7303c99ba3930b43e3175d70f36e4351.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 94, + 368, + 264, + 477 + ], + "blocks": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "lines": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "spans": [ + { + "bbox": [ + 94, + 368, + 264, + 477 + ], + "type": "table", + "html": "
Word-level Wait-k
ktoken ALword ALBLEU
12.551.5924.63
35.113.3028.82
57.955.1630.43
710.776.9931.35
913.378.67831.86
1115.8010.2232.03
", + "image_path": "576895b19625c6abe42f685905293d4a8237e8b631dbd4a18baae5f5c7d0174e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 94, + 477, + 264, + 587 + ], + "blocks": [ + { + "bbox": [ + 94, + 477, + 264, + 587 + ], + "lines": [ + { + "bbox": [ + 94, + 477, + 264, + 587 + ], + "spans": [ + { + "bbox": [ + 94, + 477, + 264, + 587 + ], + "type": "table", + "html": "
Word-level Wait-k w/ LM
ktoken ALword ALBLEU
12.891.7926.21
35.213.3630.16
57.985.1731.36
710.827.0132.29
913.428.7132.58
1115.7410.1632.45
", + "image_path": "2dd2afc0415868650d3ced5ce1e32e9880bf3bb5124c1e7efb60e19bc69fb1f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 329, + 216, + 499, + 368 + ], + "blocks": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "lines": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 67, + 596, + 289, + 619 + ], + "type": "text", + "content": "Table 10: Main results of transformer-big Waik-k models on WMT15 De.En dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 216, + 499, + 368 + ], + "lines": [ + { + "bbox": [ + 329, + 216, + 499, + 368 + ], + "spans": [ + { + "bbox": [ + 329, + 216, + 499, + 368 + ], + "type": "table", + "html": "
Token-level MoE Wait-k
ktoken ALword ALBLEU
10.860.9418.44
32.191.9224.07
53.903.0426.70
75.704.1928.47
97.695.4729.59
119.606.6730.36
1311.447.8330.51
1513.148.9130.56
1714.759.9130.65
", + "image_path": "6488edd0bf33944d926ae4289a4578216ef5571f7ffa6461bca910f2a26efde6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 329, + 368, + 499, + 477 + ], + "blocks": [ + { + "bbox": [ + 329, + 368, + 499, + 477 + ], + "lines": [ + { + "bbox": [ + 329, + 368, + 499, + 477 + ], + "spans": [ + { + "bbox": [ + 329, + 368, + 499, + 477 + ], + "type": "table", + "html": "
Word-level MoE Wait-k
ktoken ALword ALBLEU
12.861.7624.79
35.243.3728.95
57.985.1530.26
710.776.9731.01
913.378.6631.47
1115.8010.2031.55
", + "image_path": "d94c887699214c539e6755c3467e04adcbfb0332e0484828d815cff89dbc626b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 329, + 477, + 499, + 587 + ], + "blocks": [ + { + "bbox": [ + 329, + 477, + 499, + 587 + ], + "lines": [ + { + "bbox": [ + 329, + 477, + 499, + 587 + ], + "spans": [ + { + "bbox": [ + 329, + 477, + 499, + 587 + ], + "type": "table", + "html": "
Word-level MoE Wait-k w/ LM
ktoken ALword ALBLEU
12.601.6025.26
35.193.3129.03
58.015.1630.79
710.836.9931.66
913.438.6731.82
1115.8310.2132.17
", + "image_path": "5281560146f781c0093c98ae66e11f7eb04243b8577c17b9caac85c3cf526158.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 302, + 596, + 525, + 619 + ], + "lines": [ + { + "bbox": [ + 302, + 596, + 525, + 619 + ], + "spans": [ + { + "bbox": [ + 302, + 596, + 525, + 619 + ], + "type": "text", + "content": "Table 11: Main results of transformer-big MoE Waik-k models on WMT15 De.En dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 780, + 312, + 791 + ], + "type": "text", + "content": "15633" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 155, + 268, + 648 + ], + "blocks": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "lines": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "spans": [ + { + "bbox": [ + 89, + 155, + 268, + 648 + ], + "type": "table", + "html": "
Token-level ITST
δtoken ALword ALBLEU
0.21.491.5523.22
0.32.412.1224.75
0.43.933.0626.67
0.55.994.3928.23
0.69.126.5729.85
0.712.699.3231.02
0.7514.7210.7531.28
0.816.9512.2331.54
0.8519.0613.4931.70
0.921.2614.6931.84
Word-level ITST
δtoken ALword ALBLEU
0.23.692.3526.33
0.33.792.4226.62
0.43.942.5426.62
0.54.693.0227.55
0.66.604.2728.99
0.710.286.7730.41
0.7512.508.3430.95
0.814.229.6531.54
0.8515.9010.8131.56
0.918.1312.2331.61
Word-level ITST w/ LM
δtoken ALword ALBLEU
0.23.972.5328.12
0.34.002.5528.13
0.44.102.6228.29
0.54.392.8228.92
0.65.373.4629.66
0.77.855.1331.18
0.759.736.4831.77
0.811.908.0432.10
0.8514.049.5732.34
0.916.1410.9732.62
", + "image_path": "570f67e802876b0c8be79a328058a509cecb9bdb8fe48832da96736f62ee0cdb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 657, + 291, + 681 + ], + "lines": [ + { + "bbox": [ + 67, + 657, + 291, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 657, + 291, + 681 + ], + "type": "text", + "content": "Table 12: Main results of transformer-big ITST models on WMT15 De.En dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "spans": [ + { + "bbox": [ + 284, + 781, + 312, + 791 + ], + "type": "text", + "content": "15634" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 595, + 841 + ], + "page_idx": 18 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_content_list.json b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f50e971d930259a213d5d3555f1b9326074e310 --- /dev/null +++ b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_content_list.json @@ -0,0 +1,1764 @@ +[ + { + "type": "text", + "text": "Enhancing Abstractiveness of Summarization Models through Calibrated Distillation", + "text_level": 1, + "bbox": [ + 220, + 90, + 779, + 130 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hwanjun Song, Igor Shalyminov, Hang Su, Siffi Singh, Kaiheng Yao, Saab Mansour \nAWS AI Labs", + "bbox": [ + 137, + 154, + 870, + 187 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hwanjuns, shalymin, shawnsu, siffis, kaishey, saabm} @amazon.com", + "bbox": [ + 220, + 189, + 781, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 260, + 252, + 339, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sequence-level knowledge distillation reduces the size of Seq2Seq models for more efficient abstractive summarization. However, it often leads to a loss of abstractiveness in summarization. In this paper, we propose a novel approach named DisCal to enhance the level of abstractiveness (measured by $n$ -gram overlap) without sacrificing the informativeness (measured by ROUGE) of generated summaries. DisCal exposes diverse pseudo summaries with two supervision to the student model. Firstly, the best pseudo summary is identified in terms of abstractiveness and informativeness and used for sequence-level distillation. Secondly, their ranks are used to ensure the student model to assign higher prediction scores to summaries with higher ranks. Our experiments show that DisCal outperforms prior methods in abstractive summarization distillation, producing highly abstractive and informative summaries. Code is publicly available at https://c1kj.short.gy/discal.", + "bbox": [ + 141, + 279, + 460, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction", + "text_level": 1, + "bbox": [ + 114, + 604, + 260, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text summarization is the task of generating a concise and condensed summary of a source document while preserving its most important information (Gupta and Gupta, 2019). Unlike extractive summarization, which involves selecting and concatenating sentences from the original document (Nallapati et al., 2017), abstractive summarization is a sequence-to-sequence (Seq2Seq) problem that can generate novel phrases and sentences that were not present in the original document (Nallapati et al., 2016; Paulus et al.; Fan et al., 2018; Gupta and Gupta, 2019). Recent advances in large pre-trained language models have greatly accelerated summarization modeling progress (Lewis et al., 2020; Zhong et al., 2022), but there are still significant concerns for the large language model's real use cases due to the slow inference speed under a production-level environment.", + "bbox": [ + 112, + 629, + 489, + 917 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Input: Jose Mourinho has lauded Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. (...) Chelsea have topped the Premier League since the opening day but Jose Mourinho (left) will remain focused. Blues captain John Terry has been a pivotal (...)", + "bbox": [ + 512, + 248, + 885, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicted Summaries", + "text_level": 1, + "bbox": [ + 636, + 288, + 742, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "w.o. Distil: Chelsea are top the Premier League since opening day. Chelsea have led or shared the lead since opening round of fixtures. The Blues have been a key figure in keeping the side in consistent form. Jose Mourinho says his side will remain focused. Manchester City slip up at Crystal Palace last week to end their title hopes. (ROUGE-1: 30.8 / Novel 5-Gram: 82.8)", + "bbox": [ + 512, + 298, + 885, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seq-level Distil: Chelsea have topped the Premier League since the opening day. Jose Mourinho has praised Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. The Blues have led or shared the lead since the Opening round of fixtures and entered this weekend's matches seven points clear with eight matches remaining.", + "bbox": [ + 512, + 350, + 885, + 400 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(ROUGE-1: 39.2 / Novel 5-Gram: 27.3)", + "bbox": [ + 615, + 401, + 788, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DisCal: Chelsea won the Premier League since the opening day. The Blues have been sevenhave points clear with eight matches remaining. Jose Mourinho has praised the consistency and confidence. Chelsea face QPR at Loftus Road. Mourinho says Manchester City lost 2-1 at Crystal Palace.", + "bbox": [ + 512, + 413, + 882, + 453 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(ROUGE-1: 42.0 / Novel 5-Gram: 90.1)", + "bbox": [ + 613, + 454, + 788, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1: Summaries generated from the models without using knowledge distillation, \"w.o. Distil\"; using sequence-level distillation, \"Seq-level Distil\" (Zhang et al., 2022a); and using Calibrated Distillation (\"Dis-Cal\", ours) on CNNDM data. Fragments from the input are color-coded to indicate overlap: green, yellow, and red for over three, five, and ten tokens, respectively.", + "bbox": [ + 507, + 468, + 885, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Knowledge distillation is a widely used technique for compressing a large model into a smaller one for faster inference with minimal performance loss (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). A prominent direction for abstractive summarization is known as sequence-level distillation (Kim and Rush, 2016; Zhang et al., 2022a). This method involves generating a pseudo summary for each training document using a teacher model, and training a student model on pairs of training documents and their corresponding pseudo summaries. Compared to methods that rely on word-level information, such as minimizing the cross-entropy loss between teacher and student prediction distributions (Gou et al., 2021), this approach enables the student to better mimic the teacher model's generation at the sequence level.", + "bbox": [ + 507, + 577, + 884, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the high ROUGE(Lin, 2004) score achieved through sequence-level distillation, we argue that the pseudo summary generated by the teacher model exacerbates the student model's ten", + "bbox": [ + 507, + 854, + 884, + 917 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7026", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7026-7036 December 6-10, 2023 ©2023 Association for Computational Linguistics", + "bbox": [ + 216, + 945, + 779, + 973 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c0aefa51247fef47b7480ee89dfb10e0ec30788d1b3c52b409ce32fa497d9dd.jpg", + "image_caption": [ + "Figure 2: Overview of DisCal: The teacher efficiently transfers its knowledge through two approaches: firstly, employing sequence-level distillation, utilizing the best pseudo summary in terms of abstractiveness and informativeness, and secondly, applying output calibration, making higher-ranked summaries receive correspondingly higher predicted scores." + ], + "image_footnote": [], + "bbox": [ + 117, + 82, + 489, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dency to copy continuous text segments from the source documents, thus intensifying the problem of copy bias during summary generation. As seen in Figure 1, relying solely on the teacher's pseudo summaries for distilling knowledge, without utilizing gold summaries, compels the student model to generate extractive-like summaries due to the inherent copy bias (see the Seq-level Distil). Thus, the level of abstractiveness remains limited, hindering the student model's capacity to produce truly informative and coherent abstractive summaries.", + "bbox": [ + 112, + 399, + 489, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This trade-off between informativeness and abstractiveness is a significant challenge in abstractive summarization (Zhang et al., 2018; Lin and Ng, 2019), yet it has not been addressed within the context of knowledge distillation.", + "bbox": [ + 112, + 577, + 489, + 658 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present the notion of calibrated distillation, which entails distilling knowledge by precisely calibrating the pseudo summaries provided by the teacher. Our proposed method, referred to as DisCal, as illustrated in Figure 2, leverages the teacher model as a dynamic summary generator that generates diverse pseudo summaries for each input text. To enhance the diversity, we dynamically manipulate the attention temperature of the teacher model throughout the distillation process, mitigating copy-bias by exposing numerous summaries to the student model.", + "bbox": [ + 112, + 659, + 489, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the quality of the pseudo summaries, we employ a ranking system based on two factors: informativeness, which is assessed using the ROUGE score, and abstractiveness, which is mea", + "bbox": [ + 112, + 854, + 489, + 917 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sured by the ratio of novel $n$ -grams in a summary that are not present in the input text. For knowledge distillation, we select the best pseudo summary in terms of the two factors to supervise the student model through sequence-level distillation. Additionally, the ranking of the summaries is used to calibrate the student model, ensuring that it assigns higher prediction scores to summaries with higher ranks. By doing these, DisCal enhances the level of abstractiveness and improves the ROUGE score, showing promising potential even when the gold summaries in training data are less abstractive.", + "bbox": [ + 507, + 84, + 884, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are threefold: (1) we unveil the issue of reduced abstractiveness in current sequence-level distillation, (2) we introduce a novel approach called calibrated distillation to achieve better informativeness-abstractiveness trade-off, and (3) the proposed DisCal method surpasses existing state-of-the-art approaches in abstractive summarization distillation on three news and dialogue summarization datasets, namely CNNDM (Hermann et al., 2015), XSUM (Narayan et al., 2018), and SAMSum (Gliwa et al., 2019).", + "bbox": [ + 507, + 277, + 884, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Related Work", + "text_level": 1, + "bbox": [ + 507, + 463, + 665, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large pre-trained Seq2Seq models have emerged as the de facto standard for abstractive summarization due to their exceptional performance and versatility. They excel in capturing the salient information from documents through various techniques. For instance, T5 (Raffel et al., 2020) predicts corrupted text spans, BART (Lewis et al., 2020) employs denoising auto-encoding, PEGASUS (Zhang et al., 2020) identifies the most summary-worthy sentences, and DialogLED (Zhong et al., 2022) employs window-based denoising.", + "bbox": [ + 507, + 483, + 884, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the high computational cost associated with large models, there has been a surge of research focused on compressing these models (Gou et al., 2021; Frantar et al., 2023). One prominent approach in this field is known as knowledge distillation, which involves training a smaller student model to mimic the predictions of a larger teacher model by minimizing the difference between the teacher and student predictions (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). Particularly, in the context of abstractive summarization distillation with Seq2Seq models, Kim and Rush (2016) proposed the sequence-level knowledge distillation approach which involves training a student model using pseudo summaries generated by the teacher model using beam search decoding. On", + "bbox": [ + 507, + 661, + 884, + 917 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7027", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the other hand, Shleifer and Rush (2020) proposed the shrink and fine-tune (SFT) framework. This approach involves removing certain layers from the teacher model to create a smaller student model, which is then fine-tuned using gold summaries. In a recent study, Zhang et al. (2022a) introduced the method PLATE, which aims to smooth attention distributions of teacher models during pseudo summary generation and then fine-tune the shrunken student model with them.", + "bbox": [ + 112, + 84, + 489, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, there is an interesting line of work called model calibration (Liu and Liu, 2021; Zhang et al., 2022b; Liu et al., 2022), where they leverage different candidate summaries to calibrate the model's predictions to overcome the problem of exposure bias (Bengio et al., 2015). In contrast to prior research, our work focuses on the previously overlooked problem of decreased abstractiveness when distilling summarization models. We propose a solution called Calibrated Distillation, which achieves a high level of informativeness and abstractiveness using a smaller model.", + "bbox": [ + 115, + 244, + 489, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 Preliminary", + "text_level": 1, + "bbox": [ + 112, + 449, + 253, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 Seq2Seq Abstractive Summarization", + "text_level": 1, + "bbox": [ + 112, + 473, + 448, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Abstractive summarization aims at generating a concise summary of a given document or text using novel phrases and sentences. The objective of abstractive summarization is to learn a neural Transformer model $\\Theta^1$ that receives a source document $X = \\{x_{1},x_{2},\\ldots ,x_{|X|}\\}$ and generates its appropriate summary $Y = \\{y_{1},y_{2},\\ldots ,y_{|Y|}\\}$ , where $x_{t}$ and $y_{t}$ are the word token in the document and its summary at time $t$ , respectively.", + "bbox": [ + 112, + 495, + 489, + 639 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For this objective, the Seq2Seq Transformer can be trained to maximize the conditional probability:", + "bbox": [ + 112, + 640, + 487, + 671 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (Y | X; \\Theta) = \\prod_ {t = 1} ^ {| Y |} p \\left(y _ {t} \\mid Y _ {< t}, X; \\Theta\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 686, + 487, + 730 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the notation $Y_{1We mainly focus on Seq2Seq Transformer models (Vaswani et al., 2017) for abstractive summarization.", + "bbox": [ + 112, + 892, + 487, + 917 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7028", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fore, we randomly re-scale the attention temperature of the Transformer model,", + "bbox": [ + 112, + 84, + 489, + 116 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{k \\sqrt {d}}\\right) V, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 121, + 487, + 158 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q, K, V$ are linear projections of hidden states of each Transformer layer; and $k$ is a randomly drawn re-scaling factor from the uniform distribution $\\mathrm{U}(1, \\gamma)$ ; and $\\gamma$ is the maximum value for re-scaling. Thus, the teacher model generates a list of $n$ pseudo summaries via diverse beam search, $\\tilde{\\mathcal{Y}} = \\{\\tilde{Y}_1, \\tilde{Y}_2, \\dots, \\tilde{Y}_n\\}$ , and these summaries differ even for the same document according to which re-scaling factor is chosen. Examples of pseudo summaries can be found in Appendix A.", + "bbox": [ + 112, + 162, + 489, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2 Calibrated Distillation", + "text_level": 1, + "bbox": [ + 112, + 332, + 337, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a new concept of distillation named calibrated distillation, which is built on the standard sequence-level distillation pipeline but differs in terms of considering two more aspects: (1) we utilize the gold summary to identify the most reliable pseudo summary from the summary list $\\tilde{\\mathcal{V}}$ and (2) we calibrate the student model's output such that it can generate the summary with high informativeness and abstractiveness.", + "bbox": [ + 112, + 353, + 487, + 495 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, given the list of pseudo summaries $\\tilde{\\mathcal{V}}$ , DisCal evaluates and ranks the $n$ summaries in the list in terms of informativeness and abstractiveness. We define the calibration score to evaluate the ranks by employing ROUGE and novel n-gram scores, which are respectively for informativeness and abstractiveness², as in Definition 4.1.", + "bbox": [ + 112, + 499, + 489, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4.1. Let $\\tilde{\\mathcal{V}} = \\{\\tilde{Y}_1,\\tilde{Y}_2,\\dots ,\\tilde{Y}_n\\}$ be the list of $n$ pseudo summaries for an input document. Then, the informativeness score $s_{\\mathrm{info}}(\\tilde{Y}_i)$ for the $i$ -th pseudo summary is the average of ROUGE-1, ROUGE-2, and ROUGE-L F1 scores on the gold summary $Y^{*}$ , and the abstractiveness score $s_{\\mathrm{abs}}(\\tilde{Y}_i)$ for the $i$ -th pseudo summary is the average of novel 1-gram, 3-gram, and 5-gram scores with respect to the input document $X$ . Hence, the calibration score is formulated by the weighted sum of the two scores normalized over $n$ pseudo summaries in the list $\\tilde{\\mathcal{V}}$ as:", + "bbox": [ + 112, + 613, + 489, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\mathrm {c a l i b}} (\\tilde {Y} _ {i}) = (1 - \\lambda) \\bar {s} _ {\\mathrm {i n f o}} (\\tilde {Y} _ {i}) + \\lambda \\bar {s} _ {\\mathrm {a b s}} (\\tilde {Y} _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 812, + 448, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\bar {s} _ {\\text {i n f o}} (\\tilde {Y} _ {i}) = s _ {\\text {i n f o}} (\\tilde {Y} _ {i}) / \\sum_ {j = 1} ^ {n} s _ {\\text {i n f o}} (\\tilde {Y} _ {j}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 139, + 835, + 487, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {a n d} \\bar {s} _ {\\mathrm {a b s}} (\\tilde {Y _ {i}}) = s _ {\\mathrm {a b s}} (\\tilde {Y _ {i}}) / \\sum_ {j = 1} ^ {n} s _ {\\mathrm {a b s}} (\\tilde {Y _ {j}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 139, + 879, + 440, + 921 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is the balancing term of $s_{\\mathrm{info}}$ and $s_{\\mathrm{abs}}$ , adjusting the importance of the two factors.", + "bbox": [ + 507, + 84, + 882, + 116 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we now obtain a list of ranked pseudo summaries $\\tilde{\\mathcal{Y}}' = \\{Y_1', Y_2', \\ldots, Y_n'\\}$ such that $\\forall_{i < j} s_{\\mathrm{calib}}(Y_i') < s_{\\mathrm{calib}}(Y_j')$ . We use this updated list for calibrated knowledge distillation.", + "bbox": [ + 507, + 123, + 882, + 187 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, the summary $Y_{n}^{\\prime}$ is selected as the best summary among all pseudo summaries in $\\tilde{\\mathcal{V}}^{\\prime}$ since it exhibits the highest calibration score $s_{\\mathrm{calib}}$ . Hence, we employ $Y_{n}^{\\prime}$ as the target summary for guiding the student model through sequence-level knowledge distillation. The student model learns from the teacher's knowledge by minimizing a modified NLL loss. In this case, the loss equation remains identical to Eq. (2), but the target is substituted with the rank 1 pseudo summary, denoted as $\\ell_{\\mathrm{NLL}}(X,Y_n')$ . By incorporating the ROUGE score into the assessment process, we ensure that the selected summary has a high level of informativeness.", + "bbox": [ + 507, + 187, + 884, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Secondly, motivated by the work that leverages the order of candidate summaries (Zhang et al., 2022b; Liu et al., 2022), we encourage the student model's prediction output such that it assigns higher estimated probabilities to high ranked summaries. For a given pseudo summary $\\hat{Y}$ , the length-normalized estimated log-probability (Liu et al., 2022) by the student model is formulated as:", + "bbox": [ + 507, + 398, + 882, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (\\tilde {Y}) = \\frac {1}{| \\tilde {Y} | ^ {\\alpha}} \\sum_ {t = 1} ^ {| \\tilde {Y} |} \\log p (\\tilde {y} _ {t} | X, \\tilde {Y} _ {< t}; \\Theta_ {s}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 542, + 882, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{Y} = \\{\\tilde{y}_1,\\tilde{y}_2,\\dots ,\\tilde{y}_{|\\tilde{Y} |}\\}$ and $\\alpha$ is a length penalty hyperparameter similarly used for beam search. Then, our calibration loss is formulated by using the margin based pairwise ranking loss (Hopkins and May, 2011) as:", + "bbox": [ + 507, + 602, + 882, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {C a l i b}} (X, \\tilde {\\mathcal {Y} ^ {\\prime}}) = \\sum_ {i < j} \\max (0, f (\\tilde {Y} _ {j}) - f (\\tilde {Y} _ {i}) + m _ {i j}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 505, + 699, + 880, + 745 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_{ij} = (j - i)*m$ represents the margin multiplied by the difference in rank between two pseudo summaries. Intuitively, this encourages the student model's log-probability $f(\\tilde{Y}_j)$ to be greater than $f(\\tilde{Y}_i)$ since $s_{\\mathrm{calib}}(\\tilde{Y}_j) > s_{\\mathrm{calib}}(\\tilde{Y}_i)$ , thereby generating summaries with high levels of informativeness and abstractiveness.", + "bbox": [ + 507, + 747, + 882, + 858 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{2}$ Novel $n$ -gram score is the ratio of $n$ -grams in the summary that do not appear in the input document,which is widely used to measure the abstractiveness in the literature (Liu and Lapata, 2019; Zhang et al., 2022a; Dreyer et al., 2023)", + "bbox": [ + 507, + 869, + 882, + 917 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7029", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/672d0e52eccb9fdeec623a2d177204ddfc4f14263da1bfe26014238bbbf63af4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Training# Validation# Testingnovel 1-gramnovel 3-gramnovel 5-gramTask
CNNDM287,11313,36811,49018.0576.0588.87News Summarization
XSUM204,04511,33211,33485.4099.7899.98News Summarization
SAMSum14,73281881959.0793.9398.84Dialogue Summarization
", + "bbox": [ + 115, + 83, + 884, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1: Summary of datasets. Novel $n$ -gram scores are computed on pairs of input documents and their gold summaries. A higher $n$ -gram score indicates that the gold summaries in the test set are more abstractive.", + "bbox": [ + 112, + 148, + 882, + 177 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, the student model is trained by combining the two loss objectives for sequence-level knowledge distillation and model output calibration, respectively, as:", + "bbox": [ + 112, + 185, + 489, + 249 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {D i s C a l}} = \\eta * \\underbrace {\\ell_ {\\text {N L L}} \\left(X , Y _ {n} ^ {\\prime}\\right)} _ {\\text {S e q - l e v e l K D}} + \\underbrace {\\ell_ {\\text {c a l i b}} \\left(X , \\tilde {\\mathcal {Y}}\\right)} _ {\\text {O u t p u t C a l i b .}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 254, + 487, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ is the weight for the NLL loss.", + "bbox": [ + 112, + 304, + 406, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 Evaluation", + "text_level": 1, + "bbox": [ + 112, + 332, + 243, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate DisCal on three widely-used abstractive summarization datasets: two news summarization datasets of CNN/DailyMail (Hermann et al., 2015) and XSUM (Narayan et al., 2018); and a dialogue summarization dataset of SAMSum (Gliwa et al., 2019).", + "bbox": [ + 112, + 357, + 487, + 455 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The CNNDM dataset comprises online news articles sourced from the CNN and DailyMail websites, each accompanied by corresponding highlight summaries for reference.", + "- The XSUM dataset contains online articles from BBC News with single sentence summaries, which are more abstractive than those in CNNDM (Dreyer et al., 2023).", + "- The SAMSum dataset contains messenger-like conversations with summaries written by linguists. Unlike CNNDM and XSUM, SAMSum involves dialogue data that includes more than two participants." + ], + "bbox": [ + 117, + 460, + 489, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The detailed statistics of the datasets can be found in Table 1. It is important to note that each dataset exhibits varying levels of abstractiveness in its gold summaries. XSUM and SAMSum exhibit a very high level of abstractiveness compared to CNNDM, probably because their summary length is very short; mostly a single sentence.", + "bbox": [ + 112, + 684, + 487, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Teacher and Student Models. Following the literature (Zhang et al., 2022a), we consider BART Large (Lewis et al., 2020), one of the widely used Seq2Seq Transformer architectures for abstractive summarization. The BART Large model is trained on the entire dataset with gold summaries as a teacher model. Then, we configure two student", + "bbox": [ + 112, + 806, + 489, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "models with identical Transformer encoder layers to the teacher, but they differ in the number of decoder layers: BART 12-6 and BART 12-3, with six and three decoding layers, respectively. Referring to the SFT pipeline (Shleifer and Rush, 2020), the student models are initialized from the 12-encoder-layer/12-decoder-layer teacher. The two student models copy the full encoder from the teacher model. But, the decoder of BART 12-6 is copied from the $\\{0,2,4,6,8,10\\}$ decoder layers of the teacher, while the decoder of BART 12-3 from the $\\{0,5,11\\}$ decoder layers. This initialization is simple but effective since it eliminates the need for separately pre-training the two student models.", + "bbox": [ + 507, + 185, + 884, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Appendix D, we validate the generalizability of DisCal on a different state-of-the-art Seq2Seq model, DialogLED (Zhong et al., 2022).", + "bbox": [ + 507, + 410, + 882, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithms. We compare DisCal with the three prior knowledge distillation approaches, namely shrink and then fine-tune (SFT) (Shleifer and Rush, 2020) and two sequence-level knowledge distillation methods, Seq-Distil (Kim and Rush, 2016) and PLATE (Zhang et al., 2022a). The SFT method trains the student model on pairs of documents and their corresponding gold summaries without using pseudo summaries. On the other hand, the other two methods only rely on the pseudo summary generated by the teacher model using beam search decoding; PLATE is different from Seq-Distil in terms of scaling up the teacher's attention temperature in pseudo summary generation. We re-implement all compared methods and train them in the same environment using eight NVIDIA V100 GPUs and Pytorch 1.13.1 (Paszke et al., 2019).", + "bbox": [ + 507, + 463, + 882, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Similar to recent studies (Rohde et al., 2021; Zhang et al., 2022a), we train BART Large (teacher model) using the Adam optimizer (Kingma and Ba, 2014) with an initial learning rate of $5e-5$ and a label smoothing of 0.1. The teacher model is trained for 20,000 steps on CNNDM and XSUM with a weight decay of 0.001 and a batch size of 64, while 5,000 steps on SAM-Sum with a weight decay of 0.1 and a batch size of 16. We use the same training configuration on the two student models for Seq-distil and PLATE.", + "bbox": [ + 507, + 741, + 884, + 917 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7030", + "bbox": [ + 480, + 927, + 521, + 940 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bb05b0e8b50416c3d01f0dbc0f0ead8261012117920c5a2943a5925141c81dcb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
Teacher Model: BART Large with 406M parameters
BART Large (Lewis et al., 2020)44.8021.4741.807.4636.7452.27
Student Model: BART 12-6 with 306M parameters
SFT (Shleifer and Rush, 2020)44.7321.3741.766.8635.6551.78
Seq-Distil (Kim and Rush, 2016)44.1421.3341.165.5228.2642.24
PLATE (Zhang et al., 2022a)45.3322.1342.526.8735.2650.92
DisCal (ours)46.7622.5844.0710.7756.7676.62
Student Model: BART 12-3 with 255M parameters
SFT (Shleifer and Rush, 2020)44.4721.3141.717.3337.3955.14
Seq-Distil (Kim and Rush, 2016)44.2321.2241.714.6924.9838.40
PLATE (Zhang et al., 2022a)44.7821.6542.036.5432.7248.45
DisCal (ours)46.1621.9243.6210.9156.9976.88
", + "bbox": [ + 114, + 83, + 885, + 278 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cbe089e77746df243585ddbde0d4614483fe137e6dc898d6d5581d673d1b88df.jpg", + "table_caption": [ + "Table 2: Comparison on CNNDM data for news summarization. We reproduced all the methods. The reproduced BART Large shows a better ROUGE-1 score than the original implementation performance of 44.16." + ], + "table_footnote": [], + "table_body": "
MethodROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
BART-Large (Lewis et al., 2020)Teacher Model: BART Large with 406M parameters
45.3522.5037.5037.7393.0898.34
SFT (Shleifer and Rush, 2020)Student Model: BART 12-6 with 306M parameters
44.8421.4236.3536.7692.9498.49
Seq-Distil (Kim and Rush, 2016)44.2020.8635.6735.4490.6497.11
PLATE (Zhang et al., 2022a)44.7121.4336.5335.4291.3297.67
DisCal (ours)45.2421.9137.2536.8892.6898.30
SFT (Shleifer and Rush, 2020)Student Model: BART 12-3 with 255M parameters
43.6820.9136.2637.4593.7298.92
Seq-Distil (Kim and Rush, 2016)43.4820.3635.4335.5690.5697.18
PLATE (Zhang et al., 2022a)43.7820.8636.1135.6191.4197.81
DisCal (ours)44.3021.1436.7337.3592.9898.43
", + "bbox": [ + 114, + 318, + 885, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3: Comparison on XSUM data for news summarization. We reproduced all the methods. The reproduced BART Large shows a better ROUGE-1 score than the original implementation performance of 45.14.", + "bbox": [ + 112, + 502, + 882, + 533 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for our hyperparameters, we tune them on validation sets. The maximum value $\\gamma$ for re-scaling in Eq. (3), the balancing term $\\lambda$ for the calibration score in Eq. (4), and the weight for NNL loss in Eq. (7) are respectively set at 2.0, 0.2, and 0.01 on CNNDM; 1.5, 0.2, and 1.0 on XSUM; and 1.5, 0.2, and 0.1 on SAMSum. The number of pseudo summaries $n$ per document is set at 6. The detailed implementation including hyperparameter settings for all methods are provided in Appendix B.1.", + "bbox": [ + 112, + 542, + 487, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Regarding inference, we apply beam search following the convention (Lewis et al., 2020; Zhang et al., 2022a). We set the beam size, length penalty, minimum length, and maximum length to 4, 2.0, 55, and 142 on CNNDM; 6, 2.0, 10, and 62 on XSUM; 6, 2.0, 11, and 62 on SAMSum. For evaluation, we use ROUGE F1 and novel $n$ -gram scores as the informativeness and abstractiveness metrics. Refer to Appendix B.2 for details.", + "bbox": [ + 112, + 703, + 489, + 850 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 Results on News Summarization", + "text_level": 1, + "bbox": [ + 112, + 864, + 415, + 879 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 2 and 3 summarize the results obtained from two news summarization datasets. The first", + "bbox": [ + 112, + 887, + 487, + 917 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "block shows the performance of the teacher model (BART Large), while the second and third blocks include the results achieved by the two student models (BART 12-6 and BART 12-3) trained using four different knowledge distillation methods.", + "bbox": [ + 507, + 542, + 882, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In general, DisCal exhibits the best performance in terms of informativeness and abstractiveness in both datasets. Particularly, DisCal shows significant performance improvements on the CNNDM dataset. This dataset, as indicated in Table 1, exhibits a low level of abstractiveness in its gold summaries, leaving ample room for improvement. The two students models with DisCal even surpass the performance of their teacher model with large margin. On the other hand, the two existing sequence-level distillation methods, Seq-Distil and PLATE sacrifice the level of abstractiveness compared to the teacher model and SFT. For XSUM, we observe the similar trend of exhibiting the highest ROUGE while maintaining better novel $n$ -gram scores than the two sequence-level distillation methods. A less significant improvement to abstractiveness comes from short length summaries of XSUM.", + "bbox": [ + 507, + 627, + 884, + 917 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7031", + "bbox": [ + 480, + 928, + 517, + 940 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bab39393a13e9afaceeeffa5c5ab3a32a7598c10da066b58e74f0340d30758a9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
Teacher Model: BART Large with 406M parameters
BART Large (Lewis et al., 2020)53.2428.6549.2346.6083.5294.03
Student Model: BART 12-6 with 306M parameters
SFT (Shleifer and Rush, 2020)52.5227.4948.2147.8885.1295.12
Seq-Distil (Kim and Rush, 2016)52.1627.2247.9444.4080.3192.22
PLATE (Zhang et al., 2022a)53.1128.3449.0345.2581.1992.62
DisCal (ours)53.6628.9649.9747.8584.7294.65
Student Model: BART 12-3 with 255M parameters
SFT (Shleifer and Rush, 2020)47.3823.6343.7252.7489.0197.44
Seq-Distil (Kim and Rush, 2016)50.6826.0846.8446.2881.5393.10
PLATE (Zhang et al., 2022a)50.7326.3446.9047.9682.9893.67
DisCal (ours)51.6526.7248.0848.7484.7994.92
", + "bbox": [ + 114, + 83, + 885, + 278 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8b61fc91bcb923e86a504b689d9dcc760db393825c89cc98114fd900e6ff902.jpg", + "table_caption": [ + "Table 4: Comparison on SAMSum data for dialogue summarization. We reproduced all the methods." + ], + "table_footnote": [], + "table_body": "
Model#ParamInference Latency
CNNDMXSUMSAMSum
BART Large406M700ms304ms385ms
BART 12-6306M377ms212ms190ms
BART 12-3255M275ms129ms125ms
", + "bbox": [ + 115, + 300, + 487, + 376 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/33d69957b8a21bd09ebe7b3a5d12d099b7c15ac682f4c9240769189e5bd27b14.jpg", + "table_caption": [ + "Table 5: Number of parameters and latency (milliseconds per document) on V100 GPU with batch size 1." + ], + "table_footnote": [], + "table_body": "
ComponentInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
SFT (Default)44.4721.3155.14
\\( \\ell_{\\text{NLL}} (\\lambda = 0.0) \\)45.3221.9952.83
\\( \\ell_{\\text{NLL}} (\\lambda = 0.2) \\)45.3121.9555.24
\\( \\ell_{\\text{Calib}} (\\lambda = 0.2) \\)2.520.000.0
\\( \\ell_{\\text{NLL}} + \\ell_{\\text{Calib}} \\)46.1621.9276.88
", + "bbox": [ + 115, + 414, + 489, + 514 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Results on Dialogue Summarization", + "text_level": 1, + "bbox": [ + 112, + 541, + 442, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 shows the results on the dialogue SAMSum dataset. DisCal maintains its performance dominance compared to the other distillation approaches. Similar to the CNNDM dataset, BART 12-6 with DisCal exhibits ROUGE and novel $n$ -gram scores higher than its teacher model.", + "bbox": [ + 112, + 561, + 487, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 Inference Latency", + "text_level": 1, + "bbox": [ + 112, + 671, + 302, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5 summarizes the number of trainable parameters and inference latency of BART models we used. By reducing the number of decoder layers, the parameter size decreases from 406M to 255M. Notably, as the decoder is the most computationally intensive component during inference due to auto-regressive decoding, the student models demonstrate a significantly lower inference latency compared to the teacher model. Specifically, BART 12-6 and BART 12-3 achieve inference speed improvements of $1.43 - 2.03$ and $2.36 - 3.08$ times faster than BART Large, respectively. Despite faster inference speed, the student models enhanced with DisCal exhibit comparable or superior perfor", + "bbox": [ + 112, + 694, + 489, + 919 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69e31fb1c7ef9dc3002926a67eeb71b2087bf4d373b857c40944d5371c98de32.jpg", + "table_caption": [ + "Table 6: Ablation study by adding each loss component." + ], + "table_footnote": [], + "table_body": "
CoefficientInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
λ = 0.246.1621.9276.88
λ = 0.444.8219.9689.56
λ = 0.643.7518.7192.78
", + "bbox": [ + 510, + 299, + 882, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d4bfcb0917a5986744e07c2d97ee89f17f66cec281b5b77fee8b96e013ac5f7f.jpg", + "table_caption": [ + "Table 7: Varying the terms $\\lambda$ which balances between abstractiveness and informativeness according to Eq. (4)." + ], + "table_footnote": [], + "table_body": "
NumberInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
n = 345.4921.6072.66
n = 646.1621.9276.88
n = 946.2921.7583.56
n = 1246.4221.3086.67
", + "bbox": [ + 510, + 411, + 880, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8: Varying the number of pseudo summaries.", + "bbox": [ + 519, + 499, + 867, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance to BART Large in generating informative and highly abstractive summaries, as demonstrated from the results presented in Tables 2, 3, and 4.", + "bbox": [ + 507, + 521, + 880, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4 Detailed Analysis on Main Component", + "text_level": 1, + "bbox": [ + 507, + 580, + 858, + 596 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform a detailed analysis of DisCal on CNNDM data using BART 12-3.", + "bbox": [ + 507, + 601, + 882, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4.1 Loss Component Ablation Study", + "text_level": 1, + "bbox": [ + 507, + 642, + 823, + 658 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform an ablation study on DisCal by gradually adding each loss component on top of the SFT method. The results are shown in Table 6. Firstly, we use the NLL loss in Eq. (2) with $\\lambda = 0.0$ , where we consider only the informativeness score $s_{\\mathrm{info}}$ for selecting the best summary without utilizing the calibration loss from Eq. (6). In this setting, although the ROUGE score exhibits a considerable improvement, the novel 5-gram score drops. Secondly, when increasing the $\\lambda$ value to 0.2, where the best pseudo summary is selected by considering both informativeness and abstractiveness, the ROUGE and novel 5-gram scores are both improved. Next, by utilizing both the NLL loss and the calibration loss, we observe further enhancements in the ROUGE and novel 5-gram scores due to their synergistic", + "bbox": [ + 505, + 661, + 882, + 919 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7032", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/92fa06cd20dd2dff9b1cfb48983b784601c078fd4dd8f747a1a535152c842088.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSummaryDocument
GoldThe wildfire started in Miami-Dade County on Sunday. By Monday night, it had grown to nearly 2,000 acres. The fire was 50% contained, officials said.Parts of Miami-Dade County's skyline was hidden from view Monday as smoke from a growing 1,850-acre wildfire loomed over portions of the Florida county. What started as a nonthreatening and seemingly shrinking grass fire on Sunday, consuming fewer than 100 acres according to Miami-Dade Fire Rescue Battalion Chief Al Cruz, grew to be more than 10 times that within the next 24 hours. By Monday night, the fire had burned nearly 2,000 acres and was 50% contained, high temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire units and a helicopter are battling the blaze.
STF (21.1 / 63.6)The fire has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, state forester says. One school is evacuated as a precaution. The fire is 50 percent contained. Florida agriculture commissioner: The fire is around the clock
Seq-Distil (33.6 / 45.9)A grass fire in Miami-Dade County has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire units and a helicopter are battling the blaze.Temperatures and gusty winds helped the fire spread, State Forester Jim Karels said. Several fire units and a helicopter with the capacity to drop 400 gallons of water at a time were battling the blaze, Cruz said. "The Florida Forest Service and Miami-Dade Fire Rescue have worked around the clock to protect Southwest Miami-Dade County," Florida Agriculture Commissioner Adam H. Putnam said in a statement. Early Monday night, officials were considering road closures, and one school, Lincoln Marti, was evacuated as a precaution, according to the Fire Department.
PLATE (22.3 / 36.1)The fire has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire crews and a helicopter with the capacity to drop 400 gallons of water at a time are battling the blaze.
DisCal (42.4 / 71.4)The fire has burned nearly 2,000 acres and is 50% contained, fire officials say. The fire started as a nonthreatening grass fire on Sunday. A school is evacuated as a precaution. Firefighters are battling the blaze in Miami-Dade County.
", + "bbox": [ + 117, + 84, + 878, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dca6c49d63e59d89de14a1731cc6fe5600d04c23ff7b15b4de19c87388015e26.jpg", + "table_caption": [ + "Table 9: Example of summaries generated from four different methods including gold summary on CNNDM using BART 12-3. Fragments that overlap the input document by five or more words are marked in red. The values in parenthesis under the method name are the informativeness and abstractiveness scores, $(s_{\\mathrm{info}} / s_{\\mathrm{abs}})$ ." + ], + "table_footnote": [], + "table_body": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
DisCal46.1621.9276.88
+ Self Distil47.5023.5867.57
", + "bbox": [ + 115, + 331, + 485, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "impact. However, using only the calibration loss does not work as there is no supervision from the NLL loss for summary generation.", + "bbox": [ + 112, + 423, + 485, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2 Balancing Term for Calibration Score", + "text_level": 1, + "bbox": [ + 112, + 481, + 470, + 495 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The hyperparameter $\\lambda$ in Eq. (4) balances the importance between informativeness and abstractiveness in evaluating pseudo summaries. The higher the $\\lambda$ value, the greater the importance of abstractiveness in the score. Table 7 demonstrates how the ROUGE and novel 5-gram scores are affected by adjusting the $\\lambda$ value. With increasing $\\lambda$ values, we observe a trade-off between the levels of informativeness and abstractiveness in the summaries. Placing excessive weight on abstractiveness compromises the level of informativeness in the summary; the abstractiveness improves while the informativeness drops considerably.", + "bbox": [ + 112, + 502, + 489, + 711 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.3 Number of Pseudo Summaries", + "text_level": 1, + "bbox": [ + 112, + 722, + 413, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Intuitively, increasing the number $n$ of pseudo summaries from the dynamic teacher model provide more performance gain with DisCal. Table 8 shows how increasing $n$ affects the performance of DisCal. As the number increases, DisCal generates better summaries in terms of the ROUGE-1 and novel 5-gram scores, while the ROUGE-2 score begins to drop slightly when $n$ is greater than 6. Therefore, in general, having more pseudo summaries helps generate highly abstract summaries without sacrificing much informativeness.", + "bbox": [ + 112, + 741, + 489, + 917 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f7202f6410c5bc59a68b942a5b3e05bcd7a13c275be4d883f8db933a6e9bb5f3.jpg", + "table_caption": [ + "Table 10: Improvement with self-calibrated distillation." + ], + "table_footnote": [], + "table_body": "
MethodConCohRelFlu
BART Large4.914.674.062.99
STF4.904.573.882.97
Seq-Distil4.794.553.992.97
PLATE4.824.553.922.98
DisCal (ours)4.824.534.042.96
", + "bbox": [ + 510, + 331, + 880, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 11: Human-like evaluation using G-EVAL on consistency (Con), coherence (Coh), relevance (Rel), and fluency (Flu), where BART Large is the teacher model of the four distilled models.", + "bbox": [ + 507, + 423, + 882, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.4 Qualitative Analysis", + "text_level": 1, + "bbox": [ + 507, + 493, + 727, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9 presents an example of generated summaries on the test data from CNNDM. The two approaches, Seq-Distil and PLATE, generate more informative summaries compared to SFT, as they exhibit high informativeness score $s_{\\mathrm{info}}$ , However, they sacrifice the abstractiveness score $s_{\\mathrm{abs}}$ , which is lower than that of SFT. This indicates that they copy a large portion of the summaries from the input document (see the red fragments shown in table 9). In contrast, DisCal is not only robust against copy bias but also achieves a very high informativeness score compared to other methods. We provide additional examples in Appendix C.", + "bbox": [ + 505, + 514, + 882, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.5 Self-Calibrated Distillation", + "text_level": 1, + "bbox": [ + 507, + 736, + 781, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method has potential in leveraging enhanced student model as self-teacher for subsequent training. Here, we set $\\lambda = 0.0$ in this bootstrapping experiment as the student model has already been trained with DisCal ( $\\lambda = 0.2$ ). Table 10 presents the results before and after self-calibration using BART 12-3 trained with DisCal on CNNDM. While ROUGE scores got improved, novel 5-gram scores decrease. Thus, a $\\lambda$ value greater than 0.0 is still necessary to maintain high abstractiveness.", + "bbox": [ + 505, + 757, + 884, + 917 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7033", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2457e6033b9896c07ec361d3052769a2930bcd2c1c051b0ec7d194dbf11bfa77.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
SFT w. Back-translation41.8217.2938.5617.6567.2784.44
Seq-Distil w. Back-translation41.1617.1637.9317.1264.8982.29
PLATE w. Back-translation42.3617.9039.2717.5366.8384.22
DisCal wo. Back-translation46.7622.5844.0710.7756.7676.62
", + "bbox": [ + 114, + 82, + 885, + 177 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 12: Impact of using back-translation to existing distillation methods using BART 12-6 on CNNDM.", + "bbox": [ + 139, + 179, + 855, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.5 Human-like Evaluation using GPT-4", + "text_level": 1, + "bbox": [ + 112, + 206, + 448, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct human-like evaluation using G-EVAL (Liu et al., 2023). This is a novel LLM-based evaluation approach employing GPT-4, outperforming all prior automated methods and also displaying a substantial Spearman correlation of 0.513 with human scores in summarization tasks. We use exactly the same prompt suggested from the authors, employing a scale of 1 (worst) to 5 (best) for consistency, coherence, and relevance, and 1 (worst) to 3 (best) for fluency. Table 11 shows the results of four distillation models using BART 12-6, including their teacher BART Large, on CNNDM.", + "bbox": [ + 112, + 228, + 489, + 420 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our analysis yields the two insights. Firstly, all distillation methods have slight impact on consistency, coherence, relevance, and fluency; up to 0.18 difference compared to the teacher. This likely stems from the use of teacher-generated pseudo summaries, which effectively prevents performance divergence in student models. Secondly, DisCal enhances abstractiveness while maintaining high consistency. This is achieved through the integration of ROUGE (between pseudo and gold summary) in summary selection and output calibration, ensuring the student model to retain crucial contents from the gold summary during training.", + "bbox": [ + 112, + 423, + 489, + 632 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5.6 Comparison with Paraphrasing", + "text_level": 1, + "bbox": [ + 112, + 645, + 410, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Paraphrasing is a simple method for enhancing the abstractiveness of provided summaries (Li et al., 2022; Zhou and Bhat, 2021). Hence, we evaluate one of the paraphrasing techniques known as back-translation. We utilize Amazon Translate $^{2}$ , a fluent and accurate machine translation service and explore a form of back translation: English $\\rightarrow$ German $\\rightarrow$ English. Table 12 summarizes the impact of back-translation on CNNDM when it is applied to the model output generated by SFT, Seq-Distil, and PLATE using BART 12-6.", + "bbox": [ + 112, + 667, + 489, + 843 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The results demonstrate that the back-translation effectively enhances abstractiveness of existing methods, yet it noticeably reduces informativeness", + "bbox": [ + 112, + 845, + 487, + 892 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "(i.e., ROUGE) compared to not using it. In contrast, our approach, Discal, strikes a more favorable balance between informativeness and abstractiveness by the proposed calibrated distillation, resulting in improvements in both aspects.", + "bbox": [ + 507, + 206, + 884, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Conclusion", + "text_level": 1, + "bbox": [ + 507, + 294, + 640, + 309 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We propose DisCal, an improved knowledge distillation method that leverages diverse candidate summaries generated by teacher model. By evaluating and ranking pseudo summaries during distillation, DisCal chooses the best summaries in terms of informativeness and abstractiveness, and this enhances model predictions based on output calibration. Experiments on three summarization datasets demonstrate that DisCal produces summaries with a higher level of abstractiveness as well as informativeness.", + "bbox": [ + 507, + 313, + 884, + 491 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 507, + 499, + 616, + 514 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DisCal introduces additional training overhead. Generating pseudo summaries from the teacher model involves beam search decoding, which is computationally intensive compared to simple teacher forcing. However, this computational overhead in training phase does not affect inference in testing, i.e. DisCal does not require any changes in inference.", + "bbox": [ + 507, + 520, + 882, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Regarding the training overhead, some recent studies show that beam decoding can be expedited using techniques such as early exiting (Liu et al., 2021b; Schuster et al., 2022) and parallel decoding (Santilli et al., 2023). These research shows great potential on alleviating the burden associated with beam decoding during training.", + "bbox": [ + 507, + 649, + 882, + 762 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ethics Statement", + "text_level": 1, + "bbox": [ + 507, + 769, + 662, + 784 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper focuses on general abstractive summarization and knowledge distillation, introducing a novel calibrated distillation method that produces summaries with high levels of abstractiveness and informativeness. To evaluate our method, we use public benchmark datasets, i.e. CNNDM, XSUM, and SAMSum. Therefore, we do not anticipate any negative ethical and social impact.", + "bbox": [ + 507, + 790, + 884, + 919 + ], + "page_idx": 8 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://aws.amazon.com/translate/", + "bbox": [ + 134, + 903, + 352, + 917 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7034", + "bbox": [ + 480, + 927, + 519, + 940 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 115, + 84, + 213, + 98 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jimmy Ba and Rich Caruana. 2014. Do deep nets really need to be deep? In NeurIPS.", + "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. 2015. Scheduled sampling for sequence prediction with recurrent neural networks. In NeurIPS.", + "Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2020. Distilling knowledge learned in bert for text generation. In ACL.", + "Markus Dreyer, Mengwen Liu, Feng Nan, Sandeep Atluri, and Sujith Ravi. 2023. Evaluating the tradeoff between abstractiveness and factuality in abstractive summarization. In ECAL.", + "Angela Fan, David Grangier, and Michael Auli. 2018. Controllable abstractive summarization. In ACL.", + "Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. Optq: Accurate quantization for generative pre-trained transformers. In ICLR.", + "Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMsum corpus: A human-annotated dialogue dataset for abstractive summarization. In New Frontiers in Summarization.", + "Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. 2021. Knowledge distillation: A survey. IJCV, 129.", + "Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. 2020. Reducing the teacher-student gap via spherical knowledge disitllation. arXiv preprint arXiv:2010.07485.", + "Som Gupta and Sanjai Kumar Gupta. 2019. Abstractive summarization: An overview of the state of the art. Expert Systems with Applications, 121.", + "Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In NeurIPS.", + "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.", + "Mark Hopkins and Jonathan May. 2011. Tuning as ranking. In EMNLP.", + "Yoon Kim and Alexander M Rush. 2016. Sequence-level knowledge distillation. In ACL.", + "Diederik P Kingma and Jimmy Ba. 2014. ADAM: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.", + "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL." + ], + "bbox": [ + 115, + 105, + 485, + 917 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. *Ai Open*, 3:71-90.", + "Xiaobo Liang, Lijun Wu, Juntao Li, Tao Qin, Min Zhang, and Tie-Yan Liu. 2022. Multi-teacher distillation with single model for neural machine translation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30.", + "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out.", + "Hui Lin and Vincent Ng. 2019. Abstractive summarization: A survey of the state of the art. In AAAI.", + "Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv e-prints, pages arXiv-2303.", + "Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In EMNLP.", + "Yang Liu, Sheng Shen, and Mirella Lapata. 2021a. Noisy self-knowledge distillation for text summarization. In NAACL.", + "Yijin Liu, Fandong Meng, Jie Zhou, Yufeng Chen, and Jinan Xu. 2021b. Faster depth-adaptive transformers. In AAAI.", + "Yixin Liu and Pengfei Liu. 2021. Simcls: A simple framework for contrastive learning of abstractive summarization. In ACL.", + "Yixin Liu, Pengfei Liu, Dragomir Radev, and Graham Neubig. 2022. Brio: Bringing order to abstractive summarization. In ACL.", + "Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. Summarrunner: A recurrent neural network based sequence model for extractive summarization of documents. In AAAI.", + "Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In SIGNLL.", + "Shashi Narayan, Shay B Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In EMNLP.", + "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS.", + "Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. In ICML." + ], + "bbox": [ + 510, + 85, + 880, + 917 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7035", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1).", + "Tobias Rohde, Xiaoxia Wu, and Yinhan Liu. 2021. Hierarchical learning for generation with long source sequences. arXiv preprint arXiv:2104.07545.", + "Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. 2023. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427.", + "Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. 2022. Confident adaptive language modeling. In NeurIPS.", + "Sam Shleifer and Alexander M Rush. 2020. Pre-trained summarization distillation. arXiv e-prints.", + "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS.", + "Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2018. Diverse beam search for improved description of complex scenes. In AAAI.", + "Fang-Fang Zhang, Jin-ge Yao, and Rui Yan. 2018. On the abstractiveness of neural document summarization. In EMNLP.", + "Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In ICML.", + "Shengqiang Zhang, Xingxing Zhang, Hangbo Bao, and Furu Wei. 2022a. Attention temperature matters in abstractive summarization distillation. In ACL.", + "Xingxing Zhang, Yiran Liu, Xun Wang, Pengcheng He, Yang Yu, Si-Qing Chen, Wayne Xiong, and Furu Wei. 2022b. Momentum calibration for text generation. arXiv preprint arXiv:2212.04257.", + "Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. Dialoglm: Pre-trained model for long dialogue understanding and summarization. In AAAI.", + "Jianing Zhou and Suma Bhat. 2021. Paraphrase generation: A survey of the state of the art. In EMNLP, pages 5075-5086." + ], + "bbox": [ + 115, + 85, + 489, + 839 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7036", + "bbox": [ + 480, + 928, + 519, + 940 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_model.json b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c9e027bb6dc005b4b32a1628d08ea27a563b8b --- /dev/null +++ b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "title", + "bbox": [ + 0.221, + 0.091, + 0.78, + 0.131 + ], + "angle": 0, + "content": "Enhancing Abstractiveness of Summarization Models through Calibrated Distillation" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.155, + 0.871, + 0.188 + ], + "angle": 0, + "content": "Hwanjun Song, Igor Shalyminov, Hang Su, Siffi Singh, Kaiheng Yao, Saab Mansour \nAWS AI Labs" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.19, + 0.783, + 0.206 + ], + "angle": 0, + "content": "{hwanjuns, shalymin, shawnsu, siffis, kaishey, saabm} @amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.261, + 0.253, + 0.341, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.28, + 0.461, + 0.594 + ], + "angle": 0, + "content": "Sequence-level knowledge distillation reduces the size of Seq2Seq models for more efficient abstractive summarization. However, it often leads to a loss of abstractiveness in summarization. In this paper, we propose a novel approach named DisCal to enhance the level of abstractiveness (measured by \\( n \\)-gram overlap) without sacrificing the informativeness (measured by ROUGE) of generated summaries. DisCal exposes diverse pseudo summaries with two supervision to the student model. Firstly, the best pseudo summary is identified in terms of abstractiveness and informativeness and used for sequence-level distillation. Secondly, their ranks are used to ensure the student model to assign higher prediction scores to summaries with higher ranks. Our experiments show that DisCal outperforms prior methods in abstractive summarization distillation, producing highly abstractive and informative summaries. Code is publicly available at https://c1kj.short.gy/discal." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.605, + 0.262, + 0.62 + ], + "angle": 0, + "content": "1 Introduction" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.63, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Text summarization is the task of generating a concise and condensed summary of a source document while preserving its most important information (Gupta and Gupta, 2019). Unlike extractive summarization, which involves selecting and concatenating sentences from the original document (Nallapati et al., 2017), abstractive summarization is a sequence-to-sequence (Seq2Seq) problem that can generate novel phrases and sentences that were not present in the original document (Nallapati et al., 2016; Paulus et al.; Fan et al., 2018; Gupta and Gupta, 2019). Recent advances in large pre-trained language models have greatly accelerated summarization modeling progress (Lewis et al., 2020; Zhong et al., 2022), but there are still significant concerns for the large language model's real use cases due to the slow inference speed under a production-level environment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.249, + 0.886, + 0.288 + ], + "angle": 0, + "content": "Input: Jose Mourinho has lauded Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. (...) Chelsea have topped the Premier League since the opening day but Jose Mourinho (left) will remain focused. Blues captain John Terry has been a pivotal (...)" + }, + { + "type": "title", + "bbox": [ + 0.638, + 0.289, + 0.744, + 0.298 + ], + "angle": 0, + "content": "Predicted Summaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.886, + 0.35 + ], + "angle": 0, + "content": "w.o. Distil: Chelsea are top the Premier League since opening day. Chelsea have led or shared the lead since opening round of fixtures. The Blues have been a key figure in keeping the side in consistent form. Jose Mourinho says his side will remain focused. Manchester City slip up at Crystal Palace last week to end their title hopes. (ROUGE-1: 30.8 / Novel 5-Gram: 82.8)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.351, + 0.886, + 0.401 + ], + "angle": 0, + "content": "Seq-level Distil: Chelsea have topped the Premier League since the opening day. Jose Mourinho has praised Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. The Blues have led or shared the lead since the Opening round of fixtures and entered this weekend's matches seven points clear with eight matches remaining." + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.402, + 0.789, + 0.412 + ], + "angle": 0, + "content": "(ROUGE-1: 39.2 / Novel 5-Gram: 27.3)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.884, + 0.454 + ], + "angle": 0, + "content": "DisCal: Chelsea won the Premier League since the opening day. The Blues have been sevenhave points clear with eight matches remaining. Jose Mourinho has praised the consistency and confidence. Chelsea face QPR at Loftus Road. Mourinho says Manchester City lost 2-1 at Crystal Palace." + }, + { + "type": "text", + "bbox": [ + 0.615, + 0.455, + 0.789, + 0.465 + ], + "angle": 0, + "content": "(ROUGE-1: 42.0 / Novel 5-Gram: 90.1)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.469, + 0.886, + 0.569 + ], + "angle": 0, + "content": "Figure 1: Summaries generated from the models without using knowledge distillation, \"w.o. Distil\"; using sequence-level distillation, \"Seq-level Distil\" (Zhang et al., 2022a); and using Calibrated Distillation (\"Dis-Cal\", ours) on CNNDM data. Fragments from the input are color-coded to indicate overlap: green, yellow, and red for over three, five, and ten tokens, respectively." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.579, + 0.885, + 0.852 + ], + "angle": 0, + "content": "Knowledge distillation is a widely used technique for compressing a large model into a smaller one for faster inference with minimal performance loss (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). A prominent direction for abstractive summarization is known as sequence-level distillation (Kim and Rush, 2016; Zhang et al., 2022a). This method involves generating a pseudo summary for each training document using a teacher model, and training a student model on pairs of training documents and their corresponding pseudo summaries. Compared to methods that rely on word-level information, such as minimizing the cross-entropy loss between teacher and student prediction distributions (Gou et al., 2021), this approach enables the student to better mimic the teacher model's generation at the sequence level." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.855, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Despite the high ROUGE(Lin, 2004) score achieved through sequence-level distillation, we argue that the pseudo summary generated by the teacher model exacerbates the student model's ten" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7026" + }, + { + "type": "footer", + "bbox": [ + 0.218, + 0.946, + 0.781, + 0.974 + ], + "angle": 0, + "content": "Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7026-7036 December 6-10, 2023 ©2023 Association for Computational Linguistics" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.083, + 0.49, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.291, + 0.49, + 0.391 + ], + "angle": 0, + "content": "Figure 2: Overview of DisCal: The teacher efficiently transfers its knowledge through two approaches: firstly, employing sequence-level distillation, utilizing the best pseudo summary in terms of abstractiveness and informativeness, and secondly, applying output calibration, making higher-ranked summaries receive correspondingly higher predicted scores." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.4, + 0.49, + 0.576 + ], + "angle": 0, + "content": "dency to copy continuous text segments from the source documents, thus intensifying the problem of copy bias during summary generation. As seen in Figure 1, relying solely on the teacher's pseudo summaries for distilling knowledge, without utilizing gold summaries, compels the student model to generate extractive-like summaries due to the inherent copy bias (see the Seq-level Distil). Thus, the level of abstractiveness remains limited, hindering the student model's capacity to produce truly informative and coherent abstractive summaries." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.579, + 0.49, + 0.659 + ], + "angle": 0, + "content": "This trade-off between informativeness and abstractiveness is a significant challenge in abstractive summarization (Zhang et al., 2018; Lin and Ng, 2019), yet it has not been addressed within the context of knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.661, + 0.49, + 0.852 + ], + "angle": 0, + "content": "In this paper, we present the notion of calibrated distillation, which entails distilling knowledge by precisely calibrating the pseudo summaries provided by the teacher. Our proposed method, referred to as DisCal, as illustrated in Figure 2, leverages the teacher model as a dynamic summary generator that generates diverse pseudo summaries for each input text. To enhance the diversity, we dynamically manipulate the attention temperature of the teacher model throughout the distillation process, mitigating copy-bias by exposing numerous summaries to the student model." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.856, + 0.49, + 0.919 + ], + "angle": 0, + "content": "To evaluate the quality of the pseudo summaries, we employ a ranking system based on two factors: informativeness, which is assessed using the ROUGE score, and abstractiveness, which is mea" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.885, + 0.277 + ], + "angle": 0, + "content": "sured by the ratio of novel \\( n \\)-grams in a summary that are not present in the input text. For knowledge distillation, we select the best pseudo summary in terms of the two factors to supervise the student model through sequence-level distillation. Additionally, the ranking of the summaries is used to calibrate the student model, ensuring that it assigns higher prediction scores to summaries with higher ranks. By doing these, DisCal enhances the level of abstractiveness and improves the ROUGE score, showing promising potential even when the gold summaries in training data are less abstractive." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.278, + 0.885, + 0.455 + ], + "angle": 0, + "content": "Our contributions are threefold: (1) we unveil the issue of reduced abstractiveness in current sequence-level distillation, (2) we introduce a novel approach called calibrated distillation to achieve better informativeness-abstractiveness trade-off, and (3) the proposed DisCal method surpasses existing state-of-the-art approaches in abstractive summarization distillation on three news and dialogue summarization datasets, namely CNNDM (Hermann et al., 2015), XSUM (Narayan et al., 2018), and SAMSum (Gliwa et al., 2019)." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.464, + 0.666, + 0.479 + ], + "angle": 0, + "content": "2 Related Work" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.485, + 0.885, + 0.661 + ], + "angle": 0, + "content": "Large pre-trained Seq2Seq models have emerged as the de facto standard for abstractive summarization due to their exceptional performance and versatility. They excel in capturing the salient information from documents through various techniques. For instance, T5 (Raffel et al., 2020) predicts corrupted text spans, BART (Lewis et al., 2020) employs denoising auto-encoding, PEGASUS (Zhang et al., 2020) identifies the most summary-worthy sentences, and DialogLED (Zhong et al., 2022) employs window-based denoising." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.662, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Due to the high computational cost associated with large models, there has been a surge of research focused on compressing these models (Gou et al., 2021; Frantar et al., 2023). One prominent approach in this field is known as knowledge distillation, which involves training a smaller student model to mimic the predictions of a larger teacher model by minimizing the difference between the teacher and student predictions (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). Particularly, in the context of abstractive summarization distillation with Seq2Seq models, Kim and Rush (2016) proposed the sequence-level knowledge distillation approach which involves training a student model using pseudo summaries generated by the teacher model using beam search decoding. On" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7027" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.113, + 0.085, + 0.49, + 0.244 + ], + "angle": 0, + "content": "the other hand, Shleifer and Rush (2020) proposed the shrink and fine-tune (SFT) framework. This approach involves removing certain layers from the teacher model to create a smaller student model, which is then fine-tuned using gold summaries. In a recent study, Zhang et al. (2022a) introduced the method PLATE, which aims to smooth attention distributions of teacher models during pseudo summary generation and then fine-tune the shrunken student model with them." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.246, + 0.49, + 0.44 + ], + "angle": 0, + "content": "In addition, there is an interesting line of work called model calibration (Liu and Liu, 2021; Zhang et al., 2022b; Liu et al., 2022), where they leverage different candidate summaries to calibrate the model's predictions to overcome the problem of exposure bias (Bengio et al., 2015). In contrast to prior research, our work focuses on the previously overlooked problem of decreased abstractiveness when distilling summarization models. We propose a solution called Calibrated Distillation, which achieves a high level of informativeness and abstractiveness using a smaller model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.45, + 0.254, + 0.466 + ], + "angle": 0, + "content": "3 Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.474, + 0.449, + 0.489 + ], + "angle": 0, + "content": "3.1 Seq2Seq Abstractive Summarization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.496, + 0.49, + 0.64 + ], + "angle": 0, + "content": "Abstractive summarization aims at generating a concise summary of a given document or text using novel phrases and sentences. The objective of abstractive summarization is to learn a neural Transformer model \\(\\Theta^1\\) that receives a source document \\(X = \\{x_{1},x_{2},\\ldots ,x_{|X|}\\}\\) and generates its appropriate summary \\(Y = \\{y_{1},y_{2},\\ldots ,y_{|Y|}\\}\\), where \\(x_{t}\\) and \\(y_{t}\\) are the word token in the document and its summary at time \\(t\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.642, + 0.489, + 0.673 + ], + "angle": 0, + "content": "For this objective, the Seq2Seq Transformer can be trained to maximize the conditional probability:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.687, + 0.488, + 0.731 + ], + "angle": 0, + "content": "\\[\np (Y | X; \\Theta) = \\prod_ {t = 1} ^ {| Y |} p \\left(y _ {t} \\mid Y _ {< t}, X; \\Theta\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.745, + 0.488, + 0.825 + ], + "angle": 0, + "content": "where the notation \\( Y_{1We mainly focus on Seq2Seq Transformer models (Vaswani et al., 2017) for abstractive summarization." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.085, + 0.855, + 0.1 + ], + "angle": 0, + "content": "3.2 Sequence-level Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.107, + 0.885, + 0.349 + ], + "angle": 0, + "content": "Let \\(\\Theta_t\\) and \\(\\Theta_s\\) be the teacher and student models, where the student must be smaller in size compared to the teacher. Given the teacher model \\(\\Theta_t^*\\) trained by Eq. (2), the teacher's output distribution for the document \\(D\\) is approximated by the pseudo summary \\(\\tilde{Y} = \\{\\tilde{y}_1,\\tilde{y}_2,\\dots ,\\tilde{y}_{|\\tilde{Y} |}\\}\\), which is the output from running beam search with the teacher model (Kim and Rush, 2016), where the summary with the highest beam score is selected. Therefore, the student model is trained to mimic the teacher's summary generation process by minimizing the NLL loss on the teacher-generated summary \\(\\tilde{Y}\\), i.e., \\(\\ell_{\\mathrm{NLL}}(X,\\tilde{Y})\\). The gold summary \\(Y^{*}\\) in training data is no longer used in sequence-level knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.363, + 0.659, + 0.38 + ], + "angle": 0, + "content": "4 Methodology" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.387, + 0.885, + 0.516 + ], + "angle": 0, + "content": "We introduce a new distillation method DisCal (Abstractive Summarization Distillation with Calibration) in this section. Briefly speaking, the dynamic summary generator (in Section 4.1) provides a list of feasible pseudo summaries for each document and the calibrated distillation (in Section 4.2) tunes the student model to output the summary with high informativeness and abstractiveness." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.528, + 0.796, + 0.543 + ], + "angle": 0, + "content": "4.1 Dynamic Summary Generator" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.549, + 0.884, + 0.74 + ], + "angle": 0, + "content": "In sequence-level knowledge distillation, using a single deterministic pseudo-summary generated by the trained teacher model is sub-optimal. This approach limits the exposure of the model to diverse valid summaries for a given input document (Liu et al., 2021a), leading to reduced abstractiveness. Additionally, it easily propagates incorrect predictions from the teacher model to the student model due to overly confident predictions (Guo et al., 2020; Liang et al., 2022). Consequently, this can lead to poor performance in generating accurate abstractive summaries." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "To address the issues, we utilize the teacher model as a dynamic summary generator (in Figure 2), enabling it to generate diverse pseudo summaries in real-time during the distillation process. This is achieved by employing the diverse beam search technique (Vijayakumar et al., 2018) and randomly re-scaling its attention temperature within a predefined range. Manipulating attention weights in all attention modules is recognized for its effectiveness in mitigating the copy bias in the generated summary (Zhang et al., 2022a). There-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7028" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.114, + 0.085, + 0.49, + 0.117 + ], + "angle": 0, + "content": "fore, we randomly re-scale the attention temperature of the Transformer model," + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.122, + 0.488, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{k \\sqrt {d}}\\right) V, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.163, + 0.49, + 0.325 + ], + "angle": 0, + "content": "where \\(Q, K, V\\) are linear projections of hidden states of each Transformer layer; and \\(k\\) is a randomly drawn re-scaling factor from the uniform distribution \\(\\mathrm{U}(1, \\gamma)\\); and \\(\\gamma\\) is the maximum value for re-scaling. Thus, the teacher model generates a list of \\(n\\) pseudo summaries via diverse beam search, \\(\\tilde{\\mathcal{Y}} = \\{\\tilde{Y}_1, \\tilde{Y}_2, \\dots, \\tilde{Y}_n\\}\\), and these summaries differ even for the same document according to which re-scaling factor is chosen. Examples of pseudo summaries can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.334, + 0.338, + 0.349 + ], + "angle": 0, + "content": "4.2 Calibrated Distillation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.354, + 0.489, + 0.497 + ], + "angle": 0, + "content": "We introduce a new concept of distillation named calibrated distillation, which is built on the standard sequence-level distillation pipeline but differs in terms of considering two more aspects: (1) we utilize the gold summary to identify the most reliable pseudo summary from the summary list \\(\\tilde{\\mathcal{V}}\\) and (2) we calibrate the student model's output such that it can generate the summary with high informativeness and abstractiveness." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.5, + 0.49, + 0.611 + ], + "angle": 0, + "content": "Specifically, given the list of pseudo summaries \\(\\tilde{\\mathcal{V}}\\), DisCal evaluates and ranks the \\(n\\) summaries in the list in terms of informativeness and abstractiveness. We define the calibration score to evaluate the ranks by employing ROUGE and novel n-gram scores, which are respectively for informativeness and abstractiveness², as in Definition 4.1." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.614, + 0.49, + 0.807 + ], + "angle": 0, + "content": "Definition 4.1. Let \\(\\tilde{\\mathcal{V}} = \\{\\tilde{Y}_1,\\tilde{Y}_2,\\dots ,\\tilde{Y}_n\\}\\) be the list of \\(n\\) pseudo summaries for an input document. Then, the informativeness score \\(s_{\\mathrm{info}}(\\tilde{Y}_i)\\) for the \\(i\\)-th pseudo summary is the average of ROUGE-1, ROUGE-2, and ROUGE-L F1 scores on the gold summary \\(Y^{*}\\), and the abstractiveness score \\(s_{\\mathrm{abs}}(\\tilde{Y}_i)\\) for the \\(i\\)-th pseudo summary is the average of novel 1-gram, 3-gram, and 5-gram scores with respect to the input document \\(X\\). Hence, the calibration score is formulated by the weighted sum of the two scores normalized over \\(n\\) pseudo summaries in the list \\(\\tilde{\\mathcal{V}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.813, + 0.45, + 0.833 + ], + "angle": 0, + "content": "\\[\ns _ {\\mathrm {c a l i b}} (\\tilde {Y} _ {i}) = (1 - \\lambda) \\bar {s} _ {\\mathrm {i n f o}} (\\tilde {Y} _ {i}) + \\lambda \\bar {s} _ {\\mathrm {a b s}} (\\tilde {Y} _ {i}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.836, + 0.488, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\bar {s} _ {\\text {i n f o}} (\\tilde {Y} _ {i}) = s _ {\\text {i n f o}} (\\tilde {Y} _ {i}) / \\sum_ {j = 1} ^ {n} s _ {\\text {i n f o}} (\\tilde {Y} _ {j}) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.881, + 0.441, + 0.922 + ], + "angle": 0, + "content": "\\[\n\\mathrm {a n d} \\bar {s} _ {\\mathrm {a b s}} (\\tilde {Y _ {i}}) = s _ {\\mathrm {a b s}} (\\tilde {Y _ {i}}) / \\sum_ {j = 1} ^ {n} s _ {\\mathrm {a b s}} (\\tilde {Y _ {j}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.085, + 0.884, + 0.117 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the balancing term of \\(s_{\\mathrm{info}}\\) and \\(s_{\\mathrm{abs}}\\), adjusting the importance of the two factors." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.124, + 0.884, + 0.188 + ], + "angle": 0, + "content": "Accordingly, we now obtain a list of ranked pseudo summaries \\(\\tilde{\\mathcal{Y}}' = \\{Y_1', Y_2', \\ldots, Y_n'\\}\\) such that \\(\\forall_{i < j} s_{\\mathrm{calib}}(Y_i') < s_{\\mathrm{calib}}(Y_j')\\). We use this updated list for calibrated knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.189, + 0.885, + 0.398 + ], + "angle": 0, + "content": "Firstly, the summary \\(Y_{n}^{\\prime}\\) is selected as the best summary among all pseudo summaries in \\(\\tilde{\\mathcal{V}}^{\\prime}\\) since it exhibits the highest calibration score \\(s_{\\mathrm{calib}}\\). Hence, we employ \\(Y_{n}^{\\prime}\\) as the target summary for guiding the student model through sequence-level knowledge distillation. The student model learns from the teacher's knowledge by minimizing a modified NLL loss. In this case, the loss equation remains identical to Eq. (2), but the target is substituted with the rank 1 pseudo summary, denoted as \\(\\ell_{\\mathrm{NLL}}(X,Y_n')\\). By incorporating the ROUGE score into the assessment process, we ensure that the selected summary has a high level of informativeness." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.399, + 0.884, + 0.527 + ], + "angle": 0, + "content": "Secondly, motivated by the work that leverages the order of candidate summaries (Zhang et al., 2022b; Liu et al., 2022), we encourage the student model's prediction output such that it assigns higher estimated probabilities to high ranked summaries. For a given pseudo summary \\(\\hat{Y}\\), the length-normalized estimated log-probability (Liu et al., 2022) by the student model is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.543, + 0.884, + 0.589 + ], + "angle": 0, + "content": "\\[\nf (\\tilde {Y}) = \\frac {1}{| \\tilde {Y} | ^ {\\alpha}} \\sum_ {t = 1} ^ {| \\tilde {Y} |} \\log p (\\tilde {y} _ {t} | X, \\tilde {Y} _ {< t}; \\Theta_ {s}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.604, + 0.883, + 0.685 + ], + "angle": 0, + "content": "where \\(\\tilde{Y} = \\{\\tilde{y}_1,\\tilde{y}_2,\\dots ,\\tilde{y}_{|\\tilde{Y} |}\\}\\) and \\(\\alpha\\) is a length penalty hyperparameter similarly used for beam search. Then, our calibration loss is formulated by using the margin based pairwise ranking loss (Hopkins and May, 2011) as:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.7, + 0.882, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {C a l i b}} (X, \\tilde {\\mathcal {Y} ^ {\\prime}}) = \\sum_ {i < j} \\max (0, f (\\tilde {Y} _ {j}) - f (\\tilde {Y} _ {i}) + m _ {i j}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.748, + 0.884, + 0.859 + ], + "angle": 0, + "content": "where \\(m_{ij} = (j - i)*m\\) represents the margin multiplied by the difference in rank between two pseudo summaries. Intuitively, this encourages the student model's log-probability \\(f(\\tilde{Y}_j)\\) to be greater than \\(f(\\tilde{Y}_i)\\) since \\(s_{\\mathrm{calib}}(\\tilde{Y}_j) > s_{\\mathrm{calib}}(\\tilde{Y}_i)\\), thereby generating summaries with high levels of informativeness and abstractiveness." + }, + { + "type": "page_footnote", + "bbox": [ + 0.508, + 0.87, + 0.884, + 0.919 + ], + "angle": 0, + "content": "\\( {}^{2} \\) Novel \\( n \\) -gram score is the ratio of \\( n \\) -grams in the summary that do not appear in the input document,which is widely used to measure the abstractiveness in the literature (Liu and Lapata, 2019; Zhang et al., 2022a; Dreyer et al., 2023)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7029" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.084, + 0.885, + 0.147 + ], + "angle": 0, + "content": "
Dataset# Training# Validation# Testingnovel 1-gramnovel 3-gramnovel 5-gramTask
CNNDM287,11313,36811,49018.0576.0588.87News Summarization
XSUM204,04511,33211,33485.4099.7899.98News Summarization
SAMSum14,73281881959.0793.9398.84Dialogue Summarization
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.149, + 0.883, + 0.178 + ], + "angle": 0, + "content": "Table 1: Summary of datasets. Novel \\( n \\)-gram scores are computed on pairs of input documents and their gold summaries. A higher \\( n \\)-gram score indicates that the gold summaries in the test set are more abstractive." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.186, + 0.49, + 0.25 + ], + "angle": 0, + "content": "As a result, the student model is trained by combining the two loss objectives for sequence-level knowledge distillation and model output calibration, respectively, as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.255, + 0.489, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {D i s C a l}} = \\eta * \\underbrace {\\ell_ {\\text {N L L}} \\left(X , Y _ {n} ^ {\\prime}\\right)} _ {\\text {S e q - l e v e l K D}} + \\underbrace {\\ell_ {\\text {c a l i b}} \\left(X , \\tilde {\\mathcal {Y}}\\right)} _ {\\text {O u t p u t C a l i b .}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.305, + 0.408, + 0.32 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the weight for the NLL loss." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.333, + 0.245, + 0.348 + ], + "angle": 0, + "content": "5 Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.359, + 0.489, + 0.456 + ], + "angle": 0, + "content": "Datasets. We evaluate DisCal on three widely-used abstractive summarization datasets: two news summarization datasets of CNN/DailyMail (Hermann et al., 2015) and XSUM (Narayan et al., 2018); and a dialogue summarization dataset of SAMSum (Gliwa et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.461, + 0.49, + 0.525 + ], + "angle": 0, + "content": "- The CNNDM dataset comprises online news articles sourced from the CNN and DailyMail websites, each accompanied by corresponding highlight summaries for reference." + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.531, + 0.49, + 0.595 + ], + "angle": 0, + "content": "- The XSUM dataset contains online articles from BBC News with single sentence summaries, which are more abstractive than those in CNNDM (Dreyer et al., 2023)." + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.601, + 0.49, + 0.681 + ], + "angle": 0, + "content": "- The SAMSum dataset contains messenger-like conversations with summaries written by linguists. Unlike CNNDM and XSUM, SAMSum involves dialogue data that includes more than two participants." + }, + { + "type": "list", + "bbox": [ + 0.118, + 0.461, + 0.49, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.686, + 0.489, + 0.799 + ], + "angle": 0, + "content": "The detailed statistics of the datasets can be found in Table 1. It is important to note that each dataset exhibits varying levels of abstractiveness in its gold summaries. XSUM and SAMSum exhibit a very high level of abstractiveness compared to CNNDM, probably because their summary length is very short; mostly a single sentence." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.807, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Teacher and Student Models. Following the literature (Zhang et al., 2022a), we consider BART Large (Lewis et al., 2020), one of the widely used Seq2Seq Transformer architectures for abstractive summarization. The BART Large model is trained on the entire dataset with gold summaries as a teacher model. Then, we configure two student" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.186, + 0.885, + 0.411 + ], + "angle": 0, + "content": "models with identical Transformer encoder layers to the teacher, but they differ in the number of decoder layers: BART 12-6 and BART 12-3, with six and three decoding layers, respectively. Referring to the SFT pipeline (Shleifer and Rush, 2020), the student models are initialized from the 12-encoder-layer/12-decoder-layer teacher. The two student models copy the full encoder from the teacher model. But, the decoder of BART 12-6 is copied from the \\(\\{0,2,4,6,8,10\\}\\) decoder layers of the teacher, while the decoder of BART 12-3 from the \\(\\{0,5,11\\}\\) decoder layers. This initialization is simple but effective since it eliminates the need for separately pre-training the two student models." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.411, + 0.884, + 0.459 + ], + "angle": 0, + "content": "In Appendix D, we validate the generalizability of DisCal on a different state-of-the-art Seq2Seq model, DialogLED (Zhong et al., 2022)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.464, + 0.884, + 0.737 + ], + "angle": 0, + "content": "Algorithms. We compare DisCal with the three prior knowledge distillation approaches, namely shrink and then fine-tune (SFT) (Shleifer and Rush, 2020) and two sequence-level knowledge distillation methods, Seq-Distil (Kim and Rush, 2016) and PLATE (Zhang et al., 2022a). The SFT method trains the student model on pairs of documents and their corresponding gold summaries without using pseudo summaries. On the other hand, the other two methods only rely on the pseudo summary generated by the teacher model using beam search decoding; PLATE is different from Seq-Distil in terms of scaling up the teacher's attention temperature in pseudo summary generation. We re-implement all compared methods and train them in the same environment using eight NVIDIA V100 GPUs and Pytorch 1.13.1 (Paszke et al., 2019)." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.743, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Implementation Details. Similar to recent studies (Rohde et al., 2021; Zhang et al., 2022a), we train BART Large (teacher model) using the Adam optimizer (Kingma and Ba, 2014) with an initial learning rate of \\(5e-5\\) and a label smoothing of 0.1. The teacher model is trained for 20,000 steps on CNNDM and XSUM with a weight decay of 0.001 and a batch size of 64, while 5,000 steps on SAM-Sum with a weight decay of 0.1 and a batch size of 16. We use the same training configuration on the two student models for Seq-distil and PLATE." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.522, + 0.941 + ], + "angle": 0, + "content": "7030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.084, + 0.887, + 0.279 + ], + "angle": 0, + "content": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
Teacher Model: BART Large with 406M parameters
BART Large (Lewis et al., 2020)44.8021.4741.807.4636.7452.27
Student Model: BART 12-6 with 306M parameters
SFT (Shleifer and Rush, 2020)44.7321.3741.766.8635.6551.78
Seq-Distil (Kim and Rush, 2016)44.1421.3341.165.5228.2642.24
PLATE (Zhang et al., 2022a)45.3322.1342.526.8735.2650.92
DisCal (ours)46.7622.5844.0710.7756.7676.62
Student Model: BART 12-3 with 255M parameters
SFT (Shleifer and Rush, 2020)44.4721.3141.717.3337.3955.14
Seq-Distil (Kim and Rush, 2016)44.2321.2241.714.6924.9838.40
PLATE (Zhang et al., 2022a)44.7821.6542.036.5432.7248.45
DisCal (ours)46.1621.9243.6210.9156.9976.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.28, + 0.884, + 0.309 + ], + "angle": 0, + "content": "Table 2: Comparison on CNNDM data for news summarization. We reproduced all the methods. The reproduced BART Large shows a better ROUGE-1 score than the original implementation performance of 44.16." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.319, + 0.887, + 0.502 + ], + "angle": 0, + "content": "
MethodROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
BART-Large (Lewis et al., 2020)Teacher Model: BART Large with 406M parameters
45.3522.5037.5037.7393.0898.34
SFT (Shleifer and Rush, 2020)Student Model: BART 12-6 with 306M parameters
44.8421.4236.3536.7692.9498.49
Seq-Distil (Kim and Rush, 2016)44.2020.8635.6735.4490.6497.11
PLATE (Zhang et al., 2022a)44.7121.4336.5335.4291.3297.67
DisCal (ours)45.2421.9137.2536.8892.6898.30
SFT (Shleifer and Rush, 2020)Student Model: BART 12-3 with 255M parameters
43.6820.9136.2637.4593.7298.92
Seq-Distil (Kim and Rush, 2016)43.4820.3635.4335.5690.5697.18
PLATE (Zhang et al., 2022a)43.7820.8636.1135.6191.4197.81
DisCal (ours)44.3021.1436.7337.3592.9898.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.504, + 0.884, + 0.534 + ], + "angle": 0, + "content": "Table 3: Comparison on XSUM data for news summarization. We reproduced all the methods. The reproduced BART Large shows a better ROUGE-1 score than the original implementation performance of 45.14." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.543, + 0.489, + 0.703 + ], + "angle": 0, + "content": "As for our hyperparameters, we tune them on validation sets. The maximum value \\(\\gamma\\) for re-scaling in Eq. (3), the balancing term \\(\\lambda\\) for the calibration score in Eq. (4), and the weight for NNL loss in Eq. (7) are respectively set at 2.0, 0.2, and 0.01 on CNNDM; 1.5, 0.2, and 1.0 on XSUM; and 1.5, 0.2, and 0.1 on SAMSum. The number of pseudo summaries \\(n\\) per document is set at 6. The detailed implementation including hyperparameter settings for all methods are provided in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.705, + 0.49, + 0.851 + ], + "angle": 0, + "content": "Regarding inference, we apply beam search following the convention (Lewis et al., 2020; Zhang et al., 2022a). We set the beam size, length penalty, minimum length, and maximum length to 4, 2.0, 55, and 142 on CNNDM; 6, 2.0, 10, and 62 on XSUM; 6, 2.0, 11, and 62 on SAMSum. For evaluation, we use ROUGE F1 and novel \\( n \\)-gram scores as the informativeness and abstractiveness metrics. Refer to Appendix B.2 for details." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.865, + 0.416, + 0.88 + ], + "angle": 0, + "content": "5.1 Results on News Summarization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.888, + 0.489, + 0.919 + ], + "angle": 0, + "content": "Tables 2 and 3 summarize the results obtained from two news summarization datasets. The first" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.543, + 0.883, + 0.624 + ], + "angle": 0, + "content": "block shows the performance of the teacher model (BART Large), while the second and third blocks include the results achieved by the two student models (BART 12-6 and BART 12-3) trained using four different knowledge distillation methods." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.629, + 0.885, + 0.919 + ], + "angle": 0, + "content": "In general, DisCal exhibits the best performance in terms of informativeness and abstractiveness in both datasets. Particularly, DisCal shows significant performance improvements on the CNNDM dataset. This dataset, as indicated in Table 1, exhibits a low level of abstractiveness in its gold summaries, leaving ample room for improvement. The two students models with DisCal even surpass the performance of their teacher model with large margin. On the other hand, the two existing sequence-level distillation methods, Seq-Distil and PLATE sacrifice the level of abstractiveness compared to the teacher model and SFT. For XSUM, we observe the similar trend of exhibiting the highest ROUGE while maintaining better novel \\( n \\)-gram scores than the two sequence-level distillation methods. A less significant improvement to abstractiveness comes from short length summaries of XSUM." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.519, + 0.941 + ], + "angle": 0, + "content": "7031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.084, + 0.887, + 0.279 + ], + "angle": 0, + "content": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
Teacher Model: BART Large with 406M parameters
BART Large (Lewis et al., 2020)53.2428.6549.2346.6083.5294.03
Student Model: BART 12-6 with 306M parameters
SFT (Shleifer and Rush, 2020)52.5227.4948.2147.8885.1295.12
Seq-Distil (Kim and Rush, 2016)52.1627.2247.9444.4080.3192.22
PLATE (Zhang et al., 2022a)53.1128.3449.0345.2581.1992.62
DisCal (ours)53.6628.9649.9747.8584.7294.65
Student Model: BART 12-3 with 255M parameters
SFT (Shleifer and Rush, 2020)47.3823.6343.7252.7489.0197.44
Seq-Distil (Kim and Rush, 2016)50.6826.0846.8446.2881.5393.10
PLATE (Zhang et al., 2022a)50.7326.3446.9047.9682.9893.67
DisCal (ours)51.6526.7248.0848.7484.7994.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.28, + 0.838, + 0.295 + ], + "angle": 0, + "content": "Table 4: Comparison on SAMSum data for dialogue summarization. We reproduced all the methods." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.302, + 0.489, + 0.378 + ], + "angle": 0, + "content": "
Model#ParamInference Latency
CNNDMXSUMSAMSum
BART Large406M700ms304ms385ms
BART 12-6306M377ms212ms190ms
BART 12-3255M275ms129ms125ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.379, + 0.489, + 0.407 + ], + "angle": 0, + "content": "Table 5: Number of parameters and latency (milliseconds per document) on V100 GPU with batch size 1." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.415, + 0.49, + 0.515 + ], + "angle": 0, + "content": "
ComponentInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
SFT (Default)44.4721.3155.14
\\( \\ell_{\\text{NLL}} (\\lambda = 0.0) \\)45.3221.9952.83
\\( \\ell_{\\text{NLL}} (\\lambda = 0.2) \\)45.3121.9555.24
\\( \\ell_{\\text{Calib}} (\\lambda = 0.2) \\)2.520.000.0
\\( \\ell_{\\text{NLL}} + \\ell_{\\text{Calib}} \\)46.1621.9276.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.515, + 0.486, + 0.53 + ], + "angle": 0, + "content": "Table 6: Ablation study by adding each loss component." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.542, + 0.443, + 0.557 + ], + "angle": 0, + "content": "5.2 Results on Dialogue Summarization" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.562, + 0.489, + 0.658 + ], + "angle": 0, + "content": "Table 4 shows the results on the dialogue SAMSum dataset. DisCal maintains its performance dominance compared to the other distillation approaches. Similar to the CNNDM dataset, BART 12-6 with DisCal exhibits ROUGE and novel \\( n \\)-gram scores higher than its teacher model." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.672, + 0.304, + 0.687 + ], + "angle": 0, + "content": "5.3 Inference Latency" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.695, + 0.49, + 0.92 + ], + "angle": 0, + "content": "Table 5 summarizes the number of trainable parameters and inference latency of BART models we used. By reducing the number of decoder layers, the parameter size decreases from 406M to 255M. Notably, as the decoder is the most computationally intensive component during inference due to auto-regressive decoding, the student models demonstrate a significantly lower inference latency compared to the teacher model. Specifically, BART 12-6 and BART 12-3 achieve inference speed improvements of \\(1.43 - 2.03\\) and \\(2.36 - 3.08\\) times faster than BART Large, respectively. Despite faster inference speed, the student models enhanced with DisCal exhibit comparable or superior perfor" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.3, + 0.883, + 0.376 + ], + "angle": 0, + "content": "
CoefficientInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
λ = 0.246.1621.9276.88
λ = 0.444.8219.9689.56
λ = 0.643.7518.7192.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.377, + 0.884, + 0.407 + ], + "angle": 0, + "content": "Table 7: Varying the terms \\( \\lambda \\) which balances between abstractiveness and informativeness according to Eq. (4)." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.412, + 0.882, + 0.499 + ], + "angle": 0, + "content": "
NumberInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
n = 345.4921.6072.66
n = 646.1621.9276.88
n = 946.2921.7583.56
n = 1246.4221.3086.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.5, + 0.868, + 0.514 + ], + "angle": 0, + "content": "Table 8: Varying the number of pseudo summaries." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.523, + 0.882, + 0.571 + ], + "angle": 0, + "content": "mance to BART Large in generating informative and highly abstractive summaries, as demonstrated from the results presented in Tables 2, 3, and 4." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.581, + 0.86, + 0.597 + ], + "angle": 0, + "content": "5.4 Detailed Analysis on Main Component" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.602, + 0.884, + 0.634 + ], + "angle": 0, + "content": "We perform a detailed analysis of DisCal on CNNDM data using BART 12-3." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.643, + 0.825, + 0.659 + ], + "angle": 0, + "content": "5.4.1 Loss Component Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.662, + 0.884, + 0.92 + ], + "angle": 0, + "content": "We perform an ablation study on DisCal by gradually adding each loss component on top of the SFT method. The results are shown in Table 6. Firstly, we use the NLL loss in Eq. (2) with \\(\\lambda = 0.0\\), where we consider only the informativeness score \\(s_{\\mathrm{info}}\\) for selecting the best summary without utilizing the calibration loss from Eq. (6). In this setting, although the ROUGE score exhibits a considerable improvement, the novel 5-gram score drops. Secondly, when increasing the \\(\\lambda\\) value to 0.2, where the best pseudo summary is selected by considering both informativeness and abstractiveness, the ROUGE and novel 5-gram scores are both improved. Next, by utilizing both the NLL loss and the calibration loss, we observe further enhancements in the ROUGE and novel 5-gram scores due to their synergistic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7032" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.085, + 0.88, + 0.276 + ], + "angle": 0, + "content": "
MethodSummaryDocument
GoldThe wildfire started in Miami-Dade County on Sunday. By Monday night, it had grown to nearly 2,000 acres. The fire was 50% contained, officials said.Parts of Miami-Dade County's skyline was hidden from view Monday as smoke from a growing 1,850-acre wildfire loomed over portions of the Florida county. What started as a nonthreatening and seemingly shrinking grass fire on Sunday, consuming fewer than 100 acres according to Miami-Dade Fire Rescue Battalion Chief Al Cruz, grew to be more than 10 times that within the next 24 hours. By Monday night, the fire had burned nearly 2,000 acres and was 50% contained, high temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire units and a helicopter are battling the blaze.
STF (21.1 / 63.6)The fire has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, state forester says. One school is evacuated as a precaution. The fire is 50 percent contained. Florida agriculture commissioner: The fire is around the clock
Seq-Distil (33.6 / 45.9)A grass fire in Miami-Dade County has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire units and a helicopter are battling the blaze.Temperatures and gusty winds helped the fire spread, State Forester Jim Karels said. Several fire units and a helicopter with the capacity to drop 400 gallons of water at a time were battling the blaze, Cruz said. "The Florida Forest Service and Miami-Dade Fire Rescue have worked around the clock to protect Southwest Miami-Dade County," Florida Agriculture Commissioner Adam H. Putnam said in a statement. Early Monday night, officials were considering road closures, and one school, Lincoln Marti, was evacuated as a precaution, according to the Fire Department.
PLATE (22.3 / 36.1)The fire has burned nearly 2,000 acres and is 50% contained. High temperatures and gusty winds helped the fire spread, State Forester Jim Karels says. Several fire crews and a helicopter with the capacity to drop 400 gallons of water at a time are battling the blaze.
DisCal (42.4 / 71.4)The fire has burned nearly 2,000 acres and is 50% contained, fire officials say. The fire started as a nonthreatening grass fire on Sunday. A school is evacuated as a precaution. Firefighters are battling the blaze in Miami-Dade County.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.279, + 0.882, + 0.322 + ], + "angle": 0, + "content": "Table 9: Example of summaries generated from four different methods including gold summary on CNNDM using BART 12-3. Fragments that overlap the input document by five or more words are marked in red. The values in parenthesis under the method name are the informativeness and abstractiveness scores, \\((s_{\\mathrm{info}} / s_{\\mathrm{abs}})\\)." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.332, + 0.486, + 0.397 + ], + "angle": 0, + "content": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2Novel 5-gram
DisCal46.1621.9276.88
+ Self Distil47.5023.5867.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.399, + 0.485, + 0.413 + ], + "angle": 0, + "content": "Table 10: Improvement with self-calibrated distillation." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.424, + 0.486, + 0.471 + ], + "angle": 0, + "content": "impact. However, using only the calibration loss does not work as there is no supervision from the NLL loss for summary generation." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.482, + 0.472, + 0.497 + ], + "angle": 0, + "content": "5.4.2 Balancing Term for Calibration Score" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.503, + 0.49, + 0.712 + ], + "angle": 0, + "content": "The hyperparameter \\(\\lambda\\) in Eq. (4) balances the importance between informativeness and abstractiveness in evaluating pseudo summaries. The higher the \\(\\lambda\\) value, the greater the importance of abstractiveness in the score. Table 7 demonstrates how the ROUGE and novel 5-gram scores are affected by adjusting the \\(\\lambda\\) value. With increasing \\(\\lambda\\) values, we observe a trade-off between the levels of informativeness and abstractiveness in the summaries. Placing excessive weight on abstractiveness compromises the level of informativeness in the summary; the abstractiveness improves while the informativeness drops considerably." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.723, + 0.414, + 0.737 + ], + "angle": 0, + "content": "5.4.3 Number of Pseudo Summaries" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.743, + 0.49, + 0.919 + ], + "angle": 0, + "content": "Intuitively, increasing the number \\( n \\) of pseudo summaries from the dynamic teacher model provide more performance gain with DisCal. Table 8 shows how increasing \\( n \\) affects the performance of DisCal. As the number increases, DisCal generates better summaries in terms of the ROUGE-1 and novel 5-gram scores, while the ROUGE-2 score begins to drop slightly when \\( n \\) is greater than 6. Therefore, in general, having more pseudo summaries helps generate highly abstract summaries without sacrificing much informativeness." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.332, + 0.882, + 0.42 + ], + "angle": 0, + "content": "
MethodConCohRelFlu
BART Large4.914.674.062.99
STF4.904.573.882.97
Seq-Distil4.794.553.992.97
PLATE4.824.553.922.98
DisCal (ours)4.824.534.042.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.424, + 0.884, + 0.48 + ], + "angle": 0, + "content": "Table 11: Human-like evaluation using G-EVAL on consistency (Con), coherence (Coh), relevance (Rel), and fluency (Flu), where BART Large is the teacher model of the four distilled models." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.494, + 0.729, + 0.51 + ], + "angle": 0, + "content": "5.4.4 Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.515, + 0.884, + 0.724 + ], + "angle": 0, + "content": "Table 9 presents an example of generated summaries on the test data from CNNDM. The two approaches, Seq-Distil and PLATE, generate more informative summaries compared to SFT, as they exhibit high informativeness score \\( s_{\\mathrm{info}} \\), However, they sacrifice the abstractiveness score \\( s_{\\mathrm{abs}} \\), which is lower than that of SFT. This indicates that they copy a large portion of the summaries from the input document (see the red fragments shown in table 9). In contrast, DisCal is not only robust against copy bias but also achieves a very high informativeness score compared to other methods. We provide additional examples in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.737, + 0.782, + 0.752 + ], + "angle": 0, + "content": "5.4.5 Self-Calibrated Distillation" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.758, + 0.885, + 0.919 + ], + "angle": 0, + "content": "Our method has potential in leveraging enhanced student model as self-teacher for subsequent training. Here, we set \\(\\lambda = 0.0\\) in this bootstrapping experiment as the student model has already been trained with DisCal (\\(\\lambda = 0.2\\)). Table 10 presents the results before and after self-calibration using BART 12-3 trained with DisCal on CNNDM. While ROUGE scores got improved, novel 5-gram scores decrease. Thus, a \\(\\lambda\\) value greater than 0.0 is still necessary to maintain high abstractiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7033" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.083, + 0.887, + 0.178 + ], + "angle": 0, + "content": "
MethodInformativenessAbstractiveness
ROUGE-1ROUGE-2ROUGE-LNovel 1-gramNovel 3-gramNovel 5-gram
SFT w. Back-translation41.8217.2938.5617.6567.2784.44
Seq-Distil w. Back-translation41.1617.1637.9317.1264.8982.29
PLATE w. Back-translation42.3617.9039.2717.5366.8384.22
DisCal wo. Back-translation46.7622.5844.0710.7756.7676.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.18, + 0.856, + 0.195 + ], + "angle": 0, + "content": "Table 12: Impact of using back-translation to existing distillation methods using BART 12-6 on CNNDM." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.207, + 0.449, + 0.223 + ], + "angle": 0, + "content": "5.5 Human-like Evaluation using GPT-4" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.229, + 0.49, + 0.421 + ], + "angle": 0, + "content": "We conduct human-like evaluation using G-EVAL (Liu et al., 2023). This is a novel LLM-based evaluation approach employing GPT-4, outperforming all prior automated methods and also displaying a substantial Spearman correlation of 0.513 with human scores in summarization tasks. We use exactly the same prompt suggested from the authors, employing a scale of 1 (worst) to 5 (best) for consistency, coherence, and relevance, and 1 (worst) to 3 (best) for fluency. Table 11 shows the results of four distillation models using BART 12-6, including their teacher BART Large, on CNNDM." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.424, + 0.49, + 0.633 + ], + "angle": 0, + "content": "Our analysis yields the two insights. Firstly, all distillation methods have slight impact on consistency, coherence, relevance, and fluency; up to 0.18 difference compared to the teacher. This likely stems from the use of teacher-generated pseudo summaries, which effectively prevents performance divergence in student models. Secondly, DisCal enhances abstractiveness while maintaining high consistency. This is achieved through the integration of ROUGE (between pseudo and gold summary) in summary selection and output calibration, ensuring the student model to retain crucial contents from the gold summary during training." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.646, + 0.411, + 0.661 + ], + "angle": 0, + "content": "5.6 Comparison with Paraphrasing" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.668, + 0.49, + 0.844 + ], + "angle": 0, + "content": "Paraphrasing is a simple method for enhancing the abstractiveness of provided summaries (Li et al., 2022; Zhou and Bhat, 2021). Hence, we evaluate one of the paraphrasing techniques known as back-translation. We utilize Amazon Translate\\(^{2}\\), a fluent and accurate machine translation service and explore a form of back translation: English \\(\\rightarrow\\) German \\(\\rightarrow\\) English. Table 12 summarizes the impact of back-translation on CNNDM when it is applied to the model output generated by SFT, Seq-Distil, and PLATE using BART 12-6." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.846, + 0.489, + 0.893 + ], + "angle": 0, + "content": "The results demonstrate that the back-translation effectively enhances abstractiveness of existing methods, yet it noticeably reduces informativeness" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.207, + 0.885, + 0.288 + ], + "angle": 0, + "content": "(i.e., ROUGE) compared to not using it. In contrast, our approach, Discal, strikes a more favorable balance between informativeness and abstractiveness by the proposed calibrated distillation, resulting in improvements in both aspects." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.295, + 0.642, + 0.31 + ], + "angle": 0, + "content": "6 Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.315, + 0.885, + 0.492 + ], + "angle": 0, + "content": "We propose DisCal, an improved knowledge distillation method that leverages diverse candidate summaries generated by teacher model. By evaluating and ranking pseudo summaries during distillation, DisCal chooses the best summaries in terms of informativeness and abstractiveness, and this enhances model predictions based on output calibration. Experiments on three summarization datasets demonstrate that DisCal produces summaries with a higher level of abstractiveness as well as informativeness." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.5, + 0.617, + 0.515 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.521, + 0.884, + 0.648 + ], + "angle": 0, + "content": "DisCal introduces additional training overhead. Generating pseudo summaries from the teacher model involves beam search decoding, which is computationally intensive compared to simple teacher forcing. However, this computational overhead in training phase does not affect inference in testing, i.e. DisCal does not require any changes in inference." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.65, + 0.884, + 0.763 + ], + "angle": 0, + "content": "Regarding the training overhead, some recent studies show that beam decoding can be expedited using techniques such as early exiting (Liu et al., 2021b; Schuster et al., 2022) and parallel decoding (Santilli et al., 2023). These research shows great potential on alleviating the burden associated with beam decoding during training." + }, + { + "type": "title", + "bbox": [ + 0.509, + 0.77, + 0.663, + 0.785 + ], + "angle": 0, + "content": "Ethics Statement" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.791, + 0.885, + 0.92 + ], + "angle": 0, + "content": "This paper focuses on general abstractive summarization and knowledge distillation, introducing a novel calibrated distillation method that produces summaries with high levels of abstractiveness and informativeness. To evaluate our method, we use public benchmark datasets, i.e. CNNDM, XSUM, and SAMSum. Therefore, we do not anticipate any negative ethical and social impact." + }, + { + "type": "page_footnote", + "bbox": [ + 0.136, + 0.904, + 0.353, + 0.919 + ], + "angle": 0, + "content": "\\(^{2}\\)https://aws.amazon.com/translate/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7034" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.116, + 0.085, + 0.214, + 0.099 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.107, + 0.486, + 0.134 + ], + "angle": 0, + "content": "Jimmy Ba and Rich Caruana. 2014. Do deep nets really need to be deep? In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.143, + 0.487, + 0.182 + ], + "angle": 0, + "content": "Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. 2015. Scheduled sampling for sequence prediction with recurrent neural networks. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.192, + 0.486, + 0.231 + ], + "angle": 0, + "content": "Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2020. Distilling knowledge learned in bert for text generation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.24, + 0.487, + 0.292 + ], + "angle": 0, + "content": "Markus Dreyer, Mengwen Liu, Feng Nan, Sandeep Atluri, and Sujith Ravi. 2023. Evaluating the tradeoff between abstractiveness and factuality in abstractive summarization. In ECAL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.302, + 0.487, + 0.328 + ], + "angle": 0, + "content": "Angela Fan, David Grangier, and Michael Auli. 2018. Controllable abstractive summarization. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.338, + 0.487, + 0.378 + ], + "angle": 0, + "content": "Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2023. Optq: Accurate quantization for generative pre-trained transformers. In ICLR." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.387, + 0.487, + 0.439 + ], + "angle": 0, + "content": "Bogdan Gliwa, Iwona Mochol, Maciej Biesek, and Aleksander Wawer. 2019. SAMsum corpus: A human-annotated dialogue dataset for abstractive summarization. In New Frontiers in Summarization." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.449, + 0.487, + 0.488 + ], + "angle": 0, + "content": "Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. 2021. Knowledge distillation: A survey. IJCV, 129." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.497, + 0.487, + 0.551 + ], + "angle": 0, + "content": "Jia Guo, Minghao Chen, Yao Hu, Chen Zhu, Xiaofei He, and Deng Cai. 2020. Reducing the teacher-student gap via spherical knowledge disitllation. arXiv preprint arXiv:2010.07485." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.56, + 0.487, + 0.6 + ], + "angle": 0, + "content": "Som Gupta and Sanjai Kumar Gupta. 2019. Abstractive summarization: An overview of the state of the art. Expert Systems with Applications, 121." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.608, + 0.487, + 0.661 + ], + "angle": 0, + "content": "Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.67, + 0.487, + 0.711 + ], + "angle": 0, + "content": "Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.719, + 0.487, + 0.746 + ], + "angle": 0, + "content": "Mark Hopkins and Jonathan May. 2011. Tuning as ranking. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.755, + 0.487, + 0.782 + ], + "angle": 0, + "content": "Yoon Kim and Alexander M Rush. 2016. Sequence-level knowledge distillation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.791, + 0.487, + 0.83 + ], + "angle": 0, + "content": "Diederik P Kingma and Jimmy Ba. 2014. ADAM: A method for stochastic optimization. arXiv preprint arXiv:1412.6980." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.84, + 0.487, + 0.919 + ], + "angle": 0, + "content": "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In ACL." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.107, + 0.487, + 0.919 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.126 + ], + "angle": 0, + "content": "Bohan Li, Yutai Hou, and Wanxiang Che. 2022. Data augmentation approaches in natural language processing: A survey. *Ai Open*, 3:71-90." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.137, + 0.882, + 0.203 + ], + "angle": 0, + "content": "Xiaobo Liang, Lijun Wu, Juntao Li, Tao Qin, Min Zhang, and Tie-Yan Liu. 2022. Multi-teacher distillation with single model for neural machine translation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.214, + 0.882, + 0.253 + ], + "angle": 0, + "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.265, + 0.882, + 0.292 + ], + "angle": 0, + "content": "Hui Lin and Vincent Ng. 2019. Abstractive summarization: A survey of the state of the art. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.303, + 0.882, + 0.357 + ], + "angle": 0, + "content": "Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv e-prints, pages arXiv-2303." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.367, + 0.882, + 0.394 + ], + "angle": 0, + "content": "Yang Liu and Mirella Lapata. 2019. Text summarization with pretrained encoders. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.405, + 0.882, + 0.444 + ], + "angle": 0, + "content": "Yang Liu, Sheng Shen, and Mirella Lapata. 2021a. Noisy self-knowledge distillation for text summarization. In NAACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.456, + 0.882, + 0.495 + ], + "angle": 0, + "content": "Yijin Liu, Fandong Meng, Jie Zhou, Yufeng Chen, and Jinan Xu. 2021b. Faster depth-adaptive transformers. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.507, + 0.882, + 0.547 + ], + "angle": 0, + "content": "Yixin Liu and Pengfei Liu. 2021. Simcls: A simple framework for contrastive learning of abstractive summarization. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.558, + 0.882, + 0.598 + ], + "angle": 0, + "content": "Yixin Liu, Pengfei Liu, Dragomir Radev, and Graham Neubig. 2022. Brio: Bringing order to abstractive summarization. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.609, + 0.882, + 0.661 + ], + "angle": 0, + "content": "Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. Summarrunner: A recurrent neural network based sequence model for extractive summarization of documents. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.673, + 0.882, + 0.726 + ], + "angle": 0, + "content": "Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In SIGNLL." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.737, + 0.882, + 0.79 + ], + "angle": 0, + "content": "Shashi Narayan, Shay B Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.801, + 0.882, + 0.868 + ], + "angle": 0, + "content": "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.512, + 0.878, + 0.882, + 0.918 + ], + "angle": 0, + "content": "Romain Paulus, Caiming Xiong, and Richard Socher. A deep reinforced model for abstractive summarization. In ICML." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.086, + 0.882, + 0.918 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7035" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.165 + ], + "angle": 0, + "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1)." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.175, + 0.489, + 0.215 + ], + "angle": 0, + "content": "Tobias Rohde, Xiaoxia Wu, and Yinhan Liu. 2021. Hierarchical learning for generation with long source sequences. arXiv preprint arXiv:2104.07545." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.225, + 0.489, + 0.291 + ], + "angle": 0, + "content": "Andrea Santilli, Silvio Severino, Emilian Postolache, Valentino Maiorca, Michele Mancusi, Riccardo Marin, and Emanuele Rodola. 2023. Accelerating transformer inference for translation via parallel decoding. arXiv preprint arXiv:2305.10427." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.3, + 0.489, + 0.352 + ], + "angle": 0, + "content": "Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. 2022. Confident adaptive language modeling. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.363, + 0.486, + 0.39 + ], + "angle": 0, + "content": "Sam Shleifer and Alexander M Rush. 2020. Pre-trained summarization distillation. arXiv e-prints." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.4, + 0.486, + 0.453 + ], + "angle": 0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.463, + 0.489, + 0.515 + ], + "angle": 0, + "content": "Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2018. Diverse beam search for improved description of complex scenes. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.525, + 0.489, + 0.564 + ], + "angle": 0, + "content": "Fang-Fang Zhang, Jin-ge Yao, and Rui Yan. 2018. On the abstractiveness of neural document summarization. In EMNLP." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.575, + 0.489, + 0.614 + ], + "angle": 0, + "content": "Jingqing Zhang, Yao Zhao, Mohammad Saleh, and Peter Liu. 2020. Pegasus: Pre-training with extracted gap-sentences for abstractive summarization. In ICML." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.625, + 0.486, + 0.663 + ], + "angle": 0, + "content": "Shengqiang Zhang, Xingxing Zhang, Hangbo Bao, and Furu Wei. 2022a. Attention temperature matters in abstractive summarization distillation. In ACL." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.674, + 0.489, + 0.727 + ], + "angle": 0, + "content": "Xingxing Zhang, Yiran Liu, Xun Wang, Pengcheng He, Yang Yu, Si-Qing Chen, Wayne Xiong, and Furu Wei. 2022b. Momentum calibration for text generation. arXiv preprint arXiv:2212.04257." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.737, + 0.489, + 0.789 + ], + "angle": 0, + "content": "Ming Zhong, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. 2022. Dialoglm: Pre-trained model for long dialogue understanding and summarization. In AAAI." + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.8, + 0.489, + 0.84 + ], + "angle": 0, + "content": "Jianing Zhou and Suma Bhat. 2021. Paraphrase generation: A survey of the state of the art. In EMNLP, pages 5075-5086." + }, + { + "type": "list", + "bbox": [ + 0.117, + 0.086, + 0.49, + 0.84 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.929, + 0.521, + 0.941 + ], + "angle": 0, + "content": "7036" + } + ] +] \ No newline at end of file diff --git a/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_origin.pdf b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b55630be4f0527a01ec50d093d907b9310acaa30 --- /dev/null +++ b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/fcd475e0-81b7-494f-bfbc-0be5e7b94a8d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:167b47aaba66b9b1923a193d495529e6cb1e92ef13536274d7353a3c5ff64d13 +size 929386 diff --git a/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/full.md b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fb41580b55bfb72c45962877e4c48936c11197f0 --- /dev/null +++ b/2023/Enhancing Abstractiveness of Summarization Models through Calibrated Distillation/full.md @@ -0,0 +1,340 @@ +# Enhancing Abstractiveness of Summarization Models through Calibrated Distillation + +Hwanjun Song, Igor Shalyminov, Hang Su, Siffi Singh, Kaiheng Yao, Saab Mansour +AWS AI Labs + +{hwanjuns, shalymin, shawnsu, siffis, kaishey, saabm} @amazon.com + +# Abstract + +Sequence-level knowledge distillation reduces the size of Seq2Seq models for more efficient abstractive summarization. However, it often leads to a loss of abstractiveness in summarization. In this paper, we propose a novel approach named DisCal to enhance the level of abstractiveness (measured by $n$ -gram overlap) without sacrificing the informativeness (measured by ROUGE) of generated summaries. DisCal exposes diverse pseudo summaries with two supervision to the student model. Firstly, the best pseudo summary is identified in terms of abstractiveness and informativeness and used for sequence-level distillation. Secondly, their ranks are used to ensure the student model to assign higher prediction scores to summaries with higher ranks. Our experiments show that DisCal outperforms prior methods in abstractive summarization distillation, producing highly abstractive and informative summaries. Code is publicly available at https://c1kj.short.gy/discal. + +# 1 Introduction + +Text summarization is the task of generating a concise and condensed summary of a source document while preserving its most important information (Gupta and Gupta, 2019). Unlike extractive summarization, which involves selecting and concatenating sentences from the original document (Nallapati et al., 2017), abstractive summarization is a sequence-to-sequence (Seq2Seq) problem that can generate novel phrases and sentences that were not present in the original document (Nallapati et al., 2016; Paulus et al.; Fan et al., 2018; Gupta and Gupta, 2019). Recent advances in large pre-trained language models have greatly accelerated summarization modeling progress (Lewis et al., 2020; Zhong et al., 2022), but there are still significant concerns for the large language model's real use cases due to the slow inference speed under a production-level environment. + +Input: Jose Mourinho has lauded Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. (...) Chelsea have topped the Premier League since the opening day but Jose Mourinho (left) will remain focused. Blues captain John Terry has been a pivotal (...) + +# Predicted Summaries + +w.o. Distil: Chelsea are top the Premier League since opening day. Chelsea have led or shared the lead since opening round of fixtures. The Blues have been a key figure in keeping the side in consistent form. Jose Mourinho says his side will remain focused. Manchester City slip up at Crystal Palace last week to end their title hopes. (ROUGE-1: 30.8 / Novel 5-Gram: 82.8) + +Seq-level Distil: Chelsea have topped the Premier League since the opening day. Jose Mourinho has praised Chelsea's consistency, with a hint of caution, as his side bid to wrap up a wire-to-wire Premier League victory. The Blues have led or shared the lead since the Opening round of fixtures and entered this weekend's matches seven points clear with eight matches remaining. + +(ROUGE-1: 39.2 / Novel 5-Gram: 27.3) + +DisCal: Chelsea won the Premier League since the opening day. The Blues have been sevenhave points clear with eight matches remaining. Jose Mourinho has praised the consistency and confidence. Chelsea face QPR at Loftus Road. Mourinho says Manchester City lost 2-1 at Crystal Palace. + +(ROUGE-1: 42.0 / Novel 5-Gram: 90.1) + +Figure 1: Summaries generated from the models without using knowledge distillation, "w.o. Distil"; using sequence-level distillation, "Seq-level Distil" (Zhang et al., 2022a); and using Calibrated Distillation ("Dis-Cal", ours) on CNNDM data. Fragments from the input are color-coded to indicate overlap: green, yellow, and red for over three, five, and ten tokens, respectively. + +Knowledge distillation is a widely used technique for compressing a large model into a smaller one for faster inference with minimal performance loss (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). A prominent direction for abstractive summarization is known as sequence-level distillation (Kim and Rush, 2016; Zhang et al., 2022a). This method involves generating a pseudo summary for each training document using a teacher model, and training a student model on pairs of training documents and their corresponding pseudo summaries. Compared to methods that rely on word-level information, such as minimizing the cross-entropy loss between teacher and student prediction distributions (Gou et al., 2021), this approach enables the student to better mimic the teacher model's generation at the sequence level. + +Despite the high ROUGE(Lin, 2004) score achieved through sequence-level distillation, we argue that the pseudo summary generated by the teacher model exacerbates the student model's ten + +![](images/7c0aefa51247fef47b7480ee89dfb10e0ec30788d1b3c52b409ce32fa497d9dd.jpg) +Figure 2: Overview of DisCal: The teacher efficiently transfers its knowledge through two approaches: firstly, employing sequence-level distillation, utilizing the best pseudo summary in terms of abstractiveness and informativeness, and secondly, applying output calibration, making higher-ranked summaries receive correspondingly higher predicted scores. + +dency to copy continuous text segments from the source documents, thus intensifying the problem of copy bias during summary generation. As seen in Figure 1, relying solely on the teacher's pseudo summaries for distilling knowledge, without utilizing gold summaries, compels the student model to generate extractive-like summaries due to the inherent copy bias (see the Seq-level Distil). Thus, the level of abstractiveness remains limited, hindering the student model's capacity to produce truly informative and coherent abstractive summaries. + +This trade-off between informativeness and abstractiveness is a significant challenge in abstractive summarization (Zhang et al., 2018; Lin and Ng, 2019), yet it has not been addressed within the context of knowledge distillation. + +In this paper, we present the notion of calibrated distillation, which entails distilling knowledge by precisely calibrating the pseudo summaries provided by the teacher. Our proposed method, referred to as DisCal, as illustrated in Figure 2, leverages the teacher model as a dynamic summary generator that generates diverse pseudo summaries for each input text. To enhance the diversity, we dynamically manipulate the attention temperature of the teacher model throughout the distillation process, mitigating copy-bias by exposing numerous summaries to the student model. + +To evaluate the quality of the pseudo summaries, we employ a ranking system based on two factors: informativeness, which is assessed using the ROUGE score, and abstractiveness, which is mea + +sured by the ratio of novel $n$ -grams in a summary that are not present in the input text. For knowledge distillation, we select the best pseudo summary in terms of the two factors to supervise the student model through sequence-level distillation. Additionally, the ranking of the summaries is used to calibrate the student model, ensuring that it assigns higher prediction scores to summaries with higher ranks. By doing these, DisCal enhances the level of abstractiveness and improves the ROUGE score, showing promising potential even when the gold summaries in training data are less abstractive. + +Our contributions are threefold: (1) we unveil the issue of reduced abstractiveness in current sequence-level distillation, (2) we introduce a novel approach called calibrated distillation to achieve better informativeness-abstractiveness trade-off, and (3) the proposed DisCal method surpasses existing state-of-the-art approaches in abstractive summarization distillation on three news and dialogue summarization datasets, namely CNNDM (Hermann et al., 2015), XSUM (Narayan et al., 2018), and SAMSum (Gliwa et al., 2019). + +# 2 Related Work + +Large pre-trained Seq2Seq models have emerged as the de facto standard for abstractive summarization due to their exceptional performance and versatility. They excel in capturing the salient information from documents through various techniques. For instance, T5 (Raffel et al., 2020) predicts corrupted text spans, BART (Lewis et al., 2020) employs denoising auto-encoding, PEGASUS (Zhang et al., 2020) identifies the most summary-worthy sentences, and DialogLED (Zhong et al., 2022) employs window-based denoising. + +Due to the high computational cost associated with large models, there has been a surge of research focused on compressing these models (Gou et al., 2021; Frantar et al., 2023). One prominent approach in this field is known as knowledge distillation, which involves training a smaller student model to mimic the predictions of a larger teacher model by minimizing the difference between the teacher and student predictions (Ba and Caruana, 2014; Hinton et al., 2015; Chen et al., 2020). Particularly, in the context of abstractive summarization distillation with Seq2Seq models, Kim and Rush (2016) proposed the sequence-level knowledge distillation approach which involves training a student model using pseudo summaries generated by the teacher model using beam search decoding. On + +the other hand, Shleifer and Rush (2020) proposed the shrink and fine-tune (SFT) framework. This approach involves removing certain layers from the teacher model to create a smaller student model, which is then fine-tuned using gold summaries. In a recent study, Zhang et al. (2022a) introduced the method PLATE, which aims to smooth attention distributions of teacher models during pseudo summary generation and then fine-tune the shrunken student model with them. + +In addition, there is an interesting line of work called model calibration (Liu and Liu, 2021; Zhang et al., 2022b; Liu et al., 2022), where they leverage different candidate summaries to calibrate the model's predictions to overcome the problem of exposure bias (Bengio et al., 2015). In contrast to prior research, our work focuses on the previously overlooked problem of decreased abstractiveness when distilling summarization models. We propose a solution called Calibrated Distillation, which achieves a high level of informativeness and abstractiveness using a smaller model. + +# 3 Preliminary + +# 3.1 Seq2Seq Abstractive Summarization + +Abstractive summarization aims at generating a concise summary of a given document or text using novel phrases and sentences. The objective of abstractive summarization is to learn a neural Transformer model $\Theta^1$ that receives a source document $X = \{x_{1},x_{2},\ldots ,x_{|X|}\}$ and generates its appropriate summary $Y = \{y_{1},y_{2},\ldots ,y_{|Y|}\}$ , where $x_{t}$ and $y_{t}$ are the word token in the document and its summary at time $t$ , respectively. + +For this objective, the Seq2Seq Transformer can be trained to maximize the conditional probability: + +$$ +p (Y | X; \Theta) = \prod_ {t = 1} ^ {| Y |} p \left(y _ {t} \mid Y _ {< t}, X; \Theta\right), \tag {1} +$$ + +where the notation $Y_{