Add Batch 52d0d5c5-557f-47bc-85f0-05418df7164c data

.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

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

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.
+
+<table><tr><td>Task</td><td>Dataset</td><td>Train</td><td>Valid</td></tr><tr><td rowspan="2">Query</td><td>WizInt (Komeili et al., 2022)</td><td>351,375</td><td>2,467</td></tr><tr><td>Fits (xu2)</td><td>3,587</td><td>392</td></tr><tr><td rowspan="4">KG Response</td><td>WizInt (Komeili et al., 2022)</td><td>22,488</td><td>1,687</td></tr><tr><td>Fits (xu2)</td><td>6,279</td><td>656</td></tr><tr><td>Ms Marco (Nguyen et al., 2017)</td><td>281,636</td><td>36,859</td></tr><tr><td>NQ Open (Adolphs et al., 2021b)</td><td>79,168</td><td>8.757</td></tr><tr><td rowspan="5">Dialogue Response</td><td>WizInt (Komeili et al., 2022)</td><td>8,335</td><td>587</td></tr><tr><td>MSC (Xu et al., 2022a)</td><td>105,549</td><td>17,691</td></tr><tr><td>SaferDialogues (Ung et al., 2022)</td><td>6,306</td><td>7,88</td></tr><tr><td>PersonaChat (Zhang et al., 2018)</td><td>131,438</td><td>7.801</td></tr><tr><td>EmpatheicDialogues (Rashkin et al., 2019)</td><td>64,636</td><td>5,738</td></tr></table>
+
+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,
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Seen</td><td colspan="5">Unseen</td></tr><tr><td>PPL ↓</td><td>B3 ↑</td><td>B4 ↑</td><td>R1 ↑</td><td>R2 ↑</td><td>PPL ↓</td><td>B3 ↑</td><td>B4 ↑</td><td>R1 ↑</td><td>R2 ↑</td></tr><tr><td>PostKS (Lian et al., 2019)</td><td>79.1</td><td>-</td><td>-</td><td>13.0</td><td>1.0</td><td>193.8</td><td>-</td><td>-</td><td>13.1</td><td>1.0</td></tr><tr><td>ITDD (Li et al., 2019)</td><td>17.8</td><td>4.0</td><td>2.5</td><td>16.2</td><td>-</td><td>44.8</td><td>2.1</td><td>1.1</td><td>11.4</td><td>-</td></tr><tr><td>SKT (Kim et al., 2020)</td><td>52.0</td><td>-</td><td>-</td><td>19.3</td><td>6.8</td><td>81.4</td><td>-</td><td>-</td><td>16.1</td><td>4.2</td></tr><tr><td>PIPM (Chen et al., 2020)</td><td>42.7</td><td>-</td><td>3.3</td><td>19.9</td><td>7.3</td><td>65.7</td><td>-</td><td>2.5</td><td>17.6</td><td>5.4</td></tr><tr><td>ZRKGK (Li et al., 2021)</td><td>40.4</td><td>-</td><td>-</td><td>-</td><td>-</td><td>41.5</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>CoLV (Zhan et al., 2021)</td><td>39.6</td><td>-</td><td>2.9</td><td>20.6</td><td>7.9</td><td>54.3</td><td>-</td><td>2.1</td><td>19.7</td><td>6.3</td></tr><tr><td>KnowledGPT (Zhao et al., 2020)</td><td>19.2</td><td>9.5</td><td>7.2</td><td>22.0</td><td>7.9</td><td>22.3</td><td>8.3</td><td>6.0</td><td>20.5</td><td>6.7</td></tr><tr><td>SPI (Xu et al., 2023)</td><td>17.1</td><td>10.2</td><td>7.7</td><td>22.7</td><td>8.8</td><td>19.1</td><td>9.6</td><td>7.3</td><td>22.0</td><td>8.5</td></tr><tr><td>R2C2</td><td>40.7</td><td>1.7</td><td>0.9</td><td>12.3</td><td>1.9</td><td>46.6</td><td>1.6</td><td>0.9</td><td>11.9</td><td>5.4</td></tr><tr><td>R2C2-PT</td><td>24.7</td><td>6.0</td><td>4.4</td><td>17.3</td><td>5.2</td><td>30.4</td><td>6.2</td><td>4.6</td><td>17.7</td><td>5.4</td></tr><tr><td>ELVM-OK</td><td>19.8</td><td>12.0</td><td>9.5</td><td>25.3</td><td>10.7</td><td>25.8</td><td>12.1</td><td>9.7</td><td>25.2</td><td>10.5</td></tr><tr><td>ELVM-EM-Like</td><td>16.2</td><td>11.6</td><td>9.1</td><td>25.7</td><td>10.7</td><td>19.4</td><td>12.0</td><td>9.5</td><td>25.7</td><td>10.8</td></tr><tr><td>ELVM-from-R2C2</td><td>15.8</td><td>14.5</td><td>11.8</td><td>28.8</td><td>13.3</td><td>18.9</td><td>14.3</td><td>11.5</td><td>28.8</td><td>13.3</td></tr><tr><td>ELVM</td><td>14.6</td><td>14.7</td><td>11.9</td><td>29.3</td><td>13.9</td><td>18.0</td><td>14.5</td><td>11.8</td><td>29.2</td><td>13.7</td></tr></table>
+
+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><tr><td rowspan="2">Model</td><td colspan="3">Seen</td><td colspan="3">Unseen</td></tr><tr><td>R@1</td><td>R@5</td><td>R@10</td><td>R@1</td><td>R@5</td><td>R@10</td></tr><tr><td>R2C2-PT</td><td>25.7</td><td>63.8</td><td>74.3</td><td>29.2</td><td>64.8</td><td>76.7</td></tr><tr><td>ELVM-EM-Like</td><td>26.4</td><td>63.9</td><td>74.7</td><td>28.9</td><td>64.5</td><td>75.8</td></tr><tr><td>ELVM-β=0</td><td>26.1</td><td>64.6</td><td>75.7</td><td>28.9</td><td>65.3</td><td>76.8</td></tr><tr><td>ELVM</td><td>29.1</td><td>68.4</td><td>80.3</td><td>30.4</td><td>68.3</td><td>80.3</td></tr></table>
+
+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.  
+
+<table><tr><td>Model</td><td>F1 ↑</td><td>EM ↑</td><td>R1 ↑</td><td>R2 ↑</td><td>PPL ↓</td></tr><tr><td>Q.Rewriting (Raposo et al., 2022)</td><td>18.9</td><td>1.0</td><td>-</td><td>-</td><td>-</td></tr><tr><td>DPR-IHN (Kim and Kim, 2022)</td><td>30.4</td><td>4.7</td><td>-</td><td>-</td><td>-</td></tr><tr><td>QKConv (Cai et al., 2023)</td><td>33.5</td><td>5.9</td><td>-</td><td>-</td><td>-</td></tr><tr><td>R2C2-PT</td><td>22.8</td><td>1.3</td><td>25.7</td><td>11.5</td><td>9.4</td></tr><tr><td>ELVM-OK</td><td>33.0</td><td>4.5</td><td>35.6</td><td>21.9</td><td>5.6</td></tr><tr><td>ELVM</td><td>36.5</td><td>6.2</td><td>39.0</td><td>25.8</td><td>4.8</td></tr></table>
+
+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><tr><td rowspan="2">Model</td><td colspan="2">Fluency (%)</td><td colspan="2">Relevance (%)</td></tr><tr><td>Seen</td><td>Unseen</td><td>Seen</td><td>Unseen</td></tr><tr><td>KnowledGPT</td><td>69.8</td><td>62.9</td><td>41.5</td><td>41.1</td></tr><tr><td>ELVM</td><td>73.5</td><td>77.5*</td><td>49.5</td><td>63.6**</td></tr></table>
+
+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  $p_{\phi}$
+
+<table><tr><td rowspan="2">m</td><td colspan="2">Seen</td><td colspan="2">Unseen</td></tr><tr><td>B4↑</td><td>R2↑</td><td>B4↑</td><td>R2↑</td></tr><tr><td>1</td><td>10.8</td><td>12.4</td><td>10.5</td><td>12.2</td></tr><tr><td>2</td><td>11.1</td><td>13.0</td><td>11.1</td><td>12.8</td></tr><tr><td>4*</td><td>11.9</td><td>13.9</td><td>11.8</td><td>13.7</td></tr><tr><td>8</td><td>11.8</td><td>13.7</td><td>11.7</td><td>13.6</td></tr></table>
+
+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><tr><td rowspan="2">β</td><td colspan="3">Wow Unseen</td><td colspan="3">QReCC</td></tr><tr><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td><td>F1 ↑</td><td>EM ↑</td><td>R@5 ↑</td></tr><tr><td>0</td><td>9.3</td><td>10.6</td><td>65.3</td><td>34.5</td><td>5.1</td><td>55.8</td></tr><tr><td>1*</td><td>11.8</td><td>13.7</td><td>68.3</td><td>36.5</td><td>6.9</td><td>65.6</td></tr><tr><td>5</td><td>10.8</td><td>12.4</td><td>66.7</td><td>31.3</td><td>3.5</td><td>31.9</td></tr></table>
+
+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.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">WoW Unseen</td><td colspan="3">QReCC</td></tr><tr><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td><td>F1 ↑</td><td>EM ↑</td><td>R@5 ↑</td></tr><tr><td>ELVM</td><td>11.8</td><td>13.7</td><td>68.3</td><td>36.5</td><td>6.2</td><td>59.9</td></tr><tr><td>ELVMDR</td><td>12.8</td><td>15.5</td><td>81.2</td><td>37.9</td><td>6.9</td><td>65.6</td></tr></table>
+
+Table 8: Performance results achieved by employing the dense retriever instead of the default BM25 retriever for WoW unseen and QReCC evaluation.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Seen</td><td colspan="3">Unseen</td></tr><tr><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td></tr><tr><td>R2C2-PT</td><td>4.4</td><td>5.2</td><td>63.8</td><td>4.6</td><td>5.4</td><td>64.8</td></tr><tr><td>ELVM-OK</td><td>9.5</td><td>10.7</td><td>-</td><td>9.7</td><td>10.5</td><td>-</td></tr><tr><td>ELVM</td><td>11.9</td><td>13.9</td><td>68.4</td><td>11.8</td><td>13.7</td><td>68.3</td></tr><tr><td>R2C2-PTLarge</td><td>8.7</td><td>10.2</td><td>62.9</td><td>8.9</td><td>10.3</td><td>64.4</td></tr><tr><td>ELVM-OKLarge</td><td>9.6</td><td>10.6</td><td>-</td><td>10.2</td><td>11.2</td><td>-</td></tr><tr><td>ELVMLarge</td><td>12.6</td><td>14.3</td><td>72.6</td><td>12.4</td><td>14.1</td><td>74.3</td></tr></table>
+
+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
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Seen</td><td colspan="3">Unseen</td></tr><tr><td>PPL ↓</td><td>B4 ↑</td><td>R2 ↑</td><td>PPL ↓</td><td>B4 ↑</td><td>R2 ↑</td></tr><tr><td>CoLV</td><td>39.6</td><td>2.9</td><td>7.9</td><td>54.3</td><td>2.1</td><td>6.3</td></tr><tr><td>KnowledGPT</td><td>19.2</td><td>7.2</td><td>7.9</td><td>22.3</td><td>6.0</td><td>6.7</td></tr><tr><td>SPI</td><td>17.1</td><td>7.7</td><td>8.8</td><td>19.1</td><td>7.3</td><td>8.5</td></tr><tr><td>BART</td><td>974</td><td>1.9</td><td>2.6</td><td>973</td><td>1.3</td><td>2.6</td></tr><tr><td>ELVM-from-BART</td><td>15.3</td><td>8.4</td><td>11.1</td><td>16.4</td><td>8.2</td><td>10.9</td></tr><tr><td>ELVM</td><td>14.6</td><td>11.9</td><td>13.9</td><td>18.0</td><td>11.8</td><td>13.7</td></tr></table>
+
+Table 10: Performance comparison with BART as our initial model.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Train Ret</td><td rowspan="2">Test Ret</td><td colspan="4">WoW Unseen</td></tr><tr><td>PPL ↓</td><td>B4 ↑</td><td>R2 ↑</td><td>F1 ↑</td></tr><tr><td>CoLV</td><td>-</td><td>-</td><td>54.3</td><td>2.1</td><td>6.3</td><td>18.5</td></tr><tr><td>KnowledgeGPT</td><td>-</td><td>-</td><td>22.3</td><td>6.0</td><td>6.7</td><td>20.5</td></tr><tr><td>SPI</td><td>-</td><td>-</td><td>19.1</td><td>7.3</td><td>8.5</td><td>-</td></tr><tr><td>ELVM</td><td>BM25</td><td>BM25</td><td>18.0</td><td>11.8</td><td>13.7</td><td>27.0</td></tr><tr><td>ELVM</td><td>BM25</td><td>Dense</td><td>14.8</td><td>12.6</td><td>15.2</td><td>29.3</td></tr><tr><td>ELVM</td><td>Dense</td><td>Dense</td><td>12.9</td><td>12.8</td><td>15.5</td><td>29.0</td></tr><tr><td>ELVM</td><td>Dense</td><td>BM25</td><td>21.6</td><td>11.8</td><td>13.1</td><td>28.2</td></tr></table>
+
+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
+
+<table><tr><td>Dialogue Context</td><td rowspan="2" colspan="2">I want to teach my family about investing by buying them a few shares of stock, but I really don&#x27;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&#x27;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. [...]</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td><td colspan="2">Yes, it&#x27;s like the shares of a company are sold to investors through equity crowdfunding platforms.</td></tr><tr><td rowspan="2">KnownGPT</td><td>Query: n/a</td><td>GT knowledge selected: False</td></tr><tr><td colspan="2">Response: yes, you can get shares in a company like apple or xbox. you can also get shares in company like stock market.</td></tr><tr><td rowspan="2">R2C2-PT</td><td>Query: How to teach kids about investing</td><td>GT knowledge selected: False</td></tr><tr><td colspan="2">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.</td></tr><tr><td rowspan="2">ELVM-β = 0</td><td>Query: Investing tips for beginners</td><td>GT knowledge selected: False</td></tr><tr><td colspan="2">Response: That&#x27;s a good idea. The stock market is a place where you can buy and sell stocks.</td></tr><tr><td rowspan="2">ELVM w/ BM25</td><td>Query: Apple xbox snapshot stock share buy apple stock instagram stock investing</td><td>GT knowledge selected: True</td></tr><tr><td colspan="2">Response: That is a good idea. You can also buy shares of private companies which are sold to investors through equity crowdfunding platforms.</td></tr><tr><td rowspan="2">ELVM w/ DR</td><td>Query: Best stocks for teaching kids about investing</td><td>GT knowledge selected: True</td></tr><tr><td colspan="2">Response: That is a good idea. The latter include shares of private companies which are sold to investors through equity crowdfunding platforms.</td></tr></table>
+
+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  
+
+<table><tr><td>Parameter</td><td>Value</td></tr><tr><td>Training Steps</td><td>10,000</td></tr><tr><td>Batch Size</td><td>128</td></tr><tr><td>Optimizer</td><td>AdamW</td></tr><tr><td>Learning Rate (LR)</td><td>5e-5</td></tr><tr><td>LR Scheduler</td><td>Cosine</td></tr><tr><td>Adam β1</td><td>0.9</td></tr><tr><td>Adam β2</td><td>0.999</td></tr><tr><td>Adam ε</td><td>1e-8</td></tr><tr><td>Warmup Steps</td><td>1000</td></tr></table>
+
+Table 13: Hyperparameters used for stage-1 training.  $k$  indicates number of retrieved documents.  
+
+<table><tr><td></td><td>WoW</td><td>QReCC</td></tr><tr><td>Train Epochs</td><td>3</td><td>1</td></tr><tr><td>Batch Size</td><td>64</td><td>64</td></tr><tr><td>Optimizer</td><td>AdamW</td><td>AdamW</td></tr><tr><td>Learning Rate (LR)</td><td>1e-4</td><td>5e-4</td></tr><tr><td>LR Scheduler</td><td>Cosine</td><td>Cosine</td></tr><tr><td>Adam β1</td><td>0.9</td><td>0.9</td></tr><tr><td>Adam β2</td><td>0.999</td><td>0.999</td></tr><tr><td>Adam ε</td><td>1e-8</td><td>1e-8</td></tr><tr><td>Warmup Steps</td><td>100</td><td>0</td></tr><tr><td>Min Query Length</td><td>1</td><td>1</td></tr><tr><td>Max Query Length</td><td>32</td><td>32</td></tr><tr><td>Top-k</td><td>100</td><td>100</td></tr><tr><td>Top-p</td><td>0.95</td><td>0.5</td></tr><tr><td>Temperature</td><td>1.0</td><td>2.0</td></tr></table>
+
+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
+
+<table><tr><td rowspan="2">Model</td><td colspan="5">Seen</td><td colspan="5">Unseen</td></tr><tr><td>PPL ↓</td><td>B3 ↑</td><td>B4 ↑</td><td>R1 ↑</td><td>R2 ↑</td><td>PPL ↓</td><td>B3 ↑</td><td>B4 ↑</td><td>R1 ↑</td><td>R2 ↑</td></tr><tr><td>ELVM</td><td>14.6</td><td>14.7</td><td>11.9</td><td>29.3</td><td>13.9</td><td>18.0</td><td>14.5</td><td>11.8</td><td>29.2</td><td>13.7</td></tr><tr><td>ELVM-w-GT-Doc</td><td>9.7</td><td>22.1</td><td>18.5</td><td>40.9</td><td>23.9</td><td>10.5</td><td>21.5</td><td>17.9</td><td>40.7</td><td>23.3</td></tr></table>
+
+Table 15: Performance comparison in WoW in the presence of ground-truth document during evaluation (ELVM-w-GT-Doc).  
+
+<table><tr><td>Model</td><td>F1 ↑</td><td>EM ↑</td></tr><tr><td>Q Rewriting (Raposo et al., 2022)</td><td>18.9</td><td>1.0</td></tr><tr><td>DPR-IHN (Kim and Kim, 2022)</td><td>30.4</td><td>4.7</td></tr><tr><td>QKConv (Cai et al., 2023)</td><td>33.5</td><td>5.9</td></tr><tr><td>ELVM</td><td>36.5</td><td>6.2</td></tr><tr><td>ELVM-w-GT-Doc</td><td>55.4</td><td>16.2</td></tr></table>
+
+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).  
+
+<table><tr><td>Query from</td><td>R@1</td><td>R@5</td><td>R@10</td></tr><tr><td>Dialogue History</td><td>33.4</td><td>58.6</td><td>68.7</td></tr><tr><td>Last Utterance</td><td>25.5</td><td>50.8</td><td>62.6</td></tr><tr><td>R2C2-PT</td><td>17.4</td><td>33.0</td><td>39.5</td></tr><tr><td>ELVM-β=0</td><td>30.6</td><td>55.8</td><td>65.7</td></tr><tr><td>ELVM</td><td>35.2</td><td>59.9</td><td>68.6</td></tr><tr><td>ELVMDR</td><td>37.5</td><td>65.6</td><td>75.1</td></tr></table>
+
+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.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="4">Seen</td><td colspan="4">Unseen</td></tr><tr><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td><td>R@10 ↑</td><td>B4 ↑</td><td>R2 ↑</td><td>R@5 ↑</td><td>R@10 ↑</td></tr><tr><td>k=5</td><td>11.9</td><td>13.9</td><td>68.4</td><td>80.3</td><td>11.8</td><td>13.7</td><td>68.3</td><td>80.3</td></tr><tr><td>k=5Bulk</td><td>9.4</td><td>10.9</td><td>62.7</td><td>73.9</td><td>8.2</td><td>9.5</td><td>47.2</td><td>61.9</td></tr><tr><td>k=10</td><td>10.3</td><td>12.2</td><td>64.8</td><td>76.4</td><td>10.6</td><td>12.7</td><td>64.8</td><td>76.9</td></tr><tr><td>k=10Bulk</td><td>10.2</td><td>12.2</td><td>62.1</td><td>74.1</td><td>9.0</td><td>10.5</td><td>47.2</td><td>61.6</td></tr></table>
+
+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-
+
+<table><tr><td rowspan="2">Model</td><td colspan="3">Seen</td><td colspan="3">Unseen</td></tr><tr><td>PPL ↓</td><td>B4 ↑</td><td>R2 ↑</td><td>PPL ↓</td><td>B4 ↑</td><td>R2 ↑</td></tr><tr><td>ELVM</td><td>14.6</td><td>11.9</td><td>13.9</td><td>18.0</td><td>11.8</td><td>13.7</td></tr><tr><td>ELVM-Decouple</td><td>16.2</td><td>12.6</td><td>14.5</td><td>21.6</td><td>12.2</td><td>14.0</td></tr></table>
+
+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><tr><td rowspan="2">Model</td><td colspan="2">Seen</td><td colspan="2">Unseen</td></tr><tr><td>F1 ↑</td><td>KF1 ↑</td><td>F1 ↑</td><td>KF1 ↑</td></tr><tr><td>CoLV</td><td>20.3</td><td>18.2</td><td>18.5</td><td>17.5</td></tr><tr><td>KnownGPT</td><td>22.0</td><td>23.8</td><td>20.5</td><td>22.1</td></tr><tr><td>SPI</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>R2C2-PT</td><td>15.4</td><td>20.5</td><td>15.7</td><td>22.3</td></tr><tr><td>ELVM-OK</td><td>23.0</td><td>33.0</td><td>22.9</td><td>33.0</td></tr><tr><td>ELVM-EM-Like</td><td>24.5</td><td>33.2</td><td>23.5</td><td>33.5</td></tr><tr><td>ELVM</td><td>27.2</td><td>40.8</td><td>27.0</td><td>40.3</td></tr></table>
+
+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.
+
+<table><tr><td>Dialogue Context</td><td rowspan="3">Hey, 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&#x27;s going out of fashion now I think [...] Cardigan (sweater) By contrast, a pullover does not open in front but must be &quot;pulled over&quot; the head to be worn. [...] But the pullover does not open in front but must be &quot;pulled over&quot; the head to be worn.</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td></tr><tr><td>KnownGPT</td><td>Query: n/a GT knowledge selected: False Response: i like the hand-knitted version. it&#x27;s more comfortable. i like mine made of a cardigan.</td></tr><tr><td>R2C2-PT</td><td>Query: fallen fashion GT knowledge selected: False Response: I&#x27;m sorry, I didn&#x27;t mean to offend you.</td></tr><tr><td>ELVM-β = 0</td><td>Query: fallen fashion cardigans GT knowledge selected: True Response: I like the one that is tied is instead considered a robe.</td></tr><tr><td>ELVM w/ BM25</td><td>Query: 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 &quot;pulled over&quot; the head to be worn.</td></tr><tr><td>ELVM w/ DR</td><td>Query: fallen waterfall cardigan GT knowledge selected: True Response: I like the pullover that does not open in front but must be &quot;pulled over&quot; the head to be worn.</td></tr><tr><td>Dialogue Context</td><td>[...] Was she diagnosed as a child? It is much easier to diagnose children, since the symptoms begin occurring early in life</td></tr><tr><td>GT knowledge</td><td rowspan="2">Yes 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.</td></tr><tr><td>GT Response</td></tr><tr><td>KnownGPT</td><td>Query: 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.</td></tr><tr><td>R2C2-PT</td><td>Query: types of epilepsy GT knowledge selected: False Response: I am glad to hear she is receiving treatment. Dementia is such a devastating disease.</td></tr><tr><td>ELVM-β = 0</td><td>Query: types of epilepsy GT knowledge selected: False Response: I am glad she is receiving treatment. Dementia has a significant effect on a person&#x27;s caregivers.</td></tr><tr><td>ELVM w/ BM25</td><td>Query: 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.</td></tr><tr><td>ELVM w/ DR</td><td>Query: 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.</td></tr></table>
+
+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.
+
+<table><tr><td>Dialogue Context</td><td rowspan="3">Do 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&#x27;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 &quot;Big Money Hustlas&quot; and &quot;Big Money Rustlas&quot;. [...] 
+The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films &quot;Big Money Hustlas</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td></tr><tr><td rowspan="2">KnownedGPT</td><td>Response: Query: n/a 
+GT knowledge selected: False</td></tr><tr><td>Response: the duo has earned two platinum and five gold albums. they have also earned two platinum and five gold albums</td></tr><tr><td rowspan="2">R2C2-PT</td><td>Query: ICP wiki 
+GT knowledge selected: False</td></tr><tr><td>Response: The duo has earned two platinum ,with 5 golden album</td></tr><tr><td rowspan="2">ELVM-β = 0</td><td>Query: ICP facts 
+GT knowledge selected: False</td></tr><tr><td>Response: The group has been around for a long time .</td></tr><tr><td rowspan="2">ELVM w/ BM25</td><td>Query: Insane clown posse ICP facts 
+GT knowledge selected: True</td></tr><tr><td>Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films &quot;Big Money Hustlas&quot;</td></tr><tr><td rowspan="2">ELVM w/ DR</td><td>Query: Insane clown posse facts 
+GT knowledge selected: True</td></tr><tr><td>Response: The duo founded the independent record label Psychopathic Records with Alex Abbiss as manager, and produced and starred in the feature films &quot;Big Money Hustlas&quot;</td></tr><tr><td>Dialogue Context</td><td>[...] 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</td></tr><tr><td>GT knowledge</td><td>[...] 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. [...]</td></tr><tr><td>GT Response</td><td>Red meat is divined as any meat that has myoglobin than white</td></tr><tr><td rowspan="2">KnownedGPT</td><td>Query: n/a 
+GT knowledge selected: False</td></tr><tr><td>Response: red meat is classified as white meat. i think it&#x27;s interesting that only flesh from mammals or fowl is classified as red.</td></tr><tr><td rowspan="2">R2C2-PT</td><td>Query: protein powder 
+GT knowledge selected: False</td></tr><tr><td>Response: Athletes normally eat meat in large quantities.</td></tr><tr><td rowspan="2">ELVM-β = 0</td><td>Query: red meat protein 
+GT knowledge selected: True</td></tr><tr><td>Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.</td></tr><tr><td rowspan="2">ELVM w/ BM25</td><td>Query: red meat protein 
+GT knowledge selected: True</td></tr><tr><td>Response: In nutritional science, red meat is defined as any meat that has more myoglobin than white meat.</td></tr><tr><td rowspan="2">ELVM w/ DR</td><td>Query: red meat protein 
+GT knowledge selected: True</td></tr><tr><td>Response: Red meat is defined as any meat that has more myoglobin than white meat.</td></tr></table>
+
+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.
+
+<table><tr><td>Dialogue Context</td><td rowspan="2">When did Mark Spitz&#x27; 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 [...]</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td><td>Mark Spitz started a successful real-estate company in Beverly Hills</td></tr><tr><td>R2C2-PT</td><td>Query: 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.</td></tr><tr><td>ELVM-β = 0</td><td>Query: 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</td></tr><tr><td>ELVM w/ BM25</td><td>Query: Spitz endorsements after swimming career ends 
+GT knowledge selected: True 
+Response: Mark Spitz started a successful real-estate company in Beverly Hills.</td></tr><tr><td>Dialogue Context</td><td>What was the New Scheme? 
+The &#x27;new scheme&#x27; and the &#x27;new dedication&#x27; 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?</td></tr><tr><td>GT knowledge</td><td>[...] 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 [...]</td></tr><tr><td>GT Response</td><td>The funds were also spent on the training and dispatching of Ahmadiyya missionaries outside the Indian sub-continent and their sustenance.</td></tr><tr><td>R2C2-PT</td><td>Query: 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.</td></tr><tr><td>ELVM-β = 0</td><td>Query: 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.</td></tr><tr><td>ELVM w/ BM25</td><td>Query: 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.</td></tr></table>
+
+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.
+
+<table><tr><td>Dialogue Context</td><td rowspan="3">How 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&#x27;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&#x27;s last work a success?
+Sondheim and Lapine&#x27;s last work together was the rhapsodic Passion (1994), adapted from Ettore Scola&#x27;s Italian film Passione D&#x27;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.</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td></tr><tr><td>R2C2-PT</td><td>Query: Was the play &quot;rhapsodic passion&quot; a success or a failure?
+GT knowledge selected: TrueResponse: Sondheim and Lapine&#x27;s last work together was the rhapsodic Passion (1994).</td></tr><tr><td>ELVM-β = 0</td><td>Query: 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&#x27;s last work together was the rhapsodic Passion (1994).</td></tr><tr><td>ELVM w/ BM25</td><td>Query: 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.</td></tr><tr><td>Dialogue Context</td><td rowspan="3">Did 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.</td></tr><tr><td>GT knowledge</td></tr><tr><td>GT Response</td></tr><tr><td>R2C2-PT</td><td>Query: Frank Hague prosecutionGT knowledge selected: FalseResponse: No, no one tried to prosecute Frank Hague.</td></tr><tr><td>ELVM-β = 0</td><td>Query: 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.</td></tr><tr><td>ELVM w/ BM25</td><td>Query: 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.</td></tr></table>
+
+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.

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

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 4.3<sup>1</sup>. 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": "<table><tr><td>Methods</td><td>C</td><td>ArXiv</td><td>PubMed</td></tr><tr><td>Full attention</td><td>1,610M</td><td>8.63</td><td>7.63</td></tr><tr><td>Block (b=1024)</td><td>201M</td><td>13.19</td><td>12.19</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>10.75</td><td>10.13</td></tr><tr><td>MASFormer (l=2)</td><td>318M</td><td>10.25</td><td>8.75</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>9.31</td><td>8.25</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>9.63</td><td>8.25</td></tr></table>",
+        "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": "<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>8.00</td><td>19.32</td><td>28.83</td></tr><tr><td>Window(w=1024)</td><td>402M</td><td>4.32</td><td>13.51</td><td>17.03</td></tr><tr><td>Block(b=2048)</td><td>402M</td><td>5.03</td><td>9.61</td><td>12.31</td></tr><tr><td>MASFormer(l=4)</td><td>436M</td><td>6.59</td><td>14.91</td><td>18.82</td></tr><tr><td>Window(w=2048)</td><td>805M</td><td>5.05</td><td>15.21</td><td>22.79</td></tr><tr><td>Block(b=4096)</td><td>805M</td><td>5.15</td><td>14.50</td><td>23.64</td></tr><tr><td>MASFormer(l=6)</td><td>553M</td><td>7.15</td><td>15.72</td><td>21.20</td></tr><tr><td>MASFormer(l=8)</td><td>671M</td><td>7.46</td><td>17.00</td><td>24.42</td></tr><tr><td>MASFormer(l=12)</td><td>906M</td><td>8.70</td><td>18.58</td><td>26.26</td></tr></table>",
+        "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": "<table><tr><td>l=8</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>w.o. C.T.</td><td>29.33/6.43/25.71</td><td>53.28/23.61/51.74</td></tr><tr><td>C.T. (n=2048)</td><td>29.87/7.16/26.15</td><td>52.28/23.01/49.83</td></tr><tr><td>C.T. (n=8192)</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr></table>",
+        "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": "<table><tr><td>Position</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>Every 3</td><td>28.26/6.94/25.03</td><td>26.16/12.37/24.82</td></tr><tr><td>Top 8</td><td>20.89/4.52/18.37</td><td>-/-/-</td></tr><tr><td>Middle 8</td><td>27.27/5.99/24.06</td><td>20.80/9.01/19.52</td></tr><tr><td>Bottom 8</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr><tr><td>Every 2</td><td>31.27/8.19/27.41</td><td>35.34/16.04/33.68</td></tr><tr><td>Bottom 12</td><td>32.53/8.70/28.75</td><td>56.98/26.26/54.46</td></tr></table>",
+        "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": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\">Example Count</td><td colspan=\"3\">Input Length</td></tr><tr><td>Train</td><td>Valid</td><td>Test</td><td>Average</td><td>Median</td><td>90th percentile</td></tr><tr><td>ArXiv</td><td>203,037</td><td>6,436</td><td>6,440</td><td>10,720</td><td>8,519</td><td>20,170</td></tr><tr><td>PubMed</td><td>119,924</td><td>6,633</td><td>6,658</td><td>4,748</td><td>3,883</td><td>8,883</td></tr><tr><td>QMSUM</td><td>1,257</td><td>272</td><td>281</td><td>9,497</td><td>14,197</td><td>27,761</td></tr><tr><td>GovReport</td><td>17,457</td><td>972</td><td>973</td><td>7,886</td><td>8,841</td><td>18,835</td></tr></table>",
+        "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": "<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>31.50 / 8.00 / 27.81</td><td>46.13 / 19.32 / 41.89</td><td>60.53 / 28.83 / 57.88</td></tr><tr><td>Window (w=1024)</td><td>402M</td><td>23.31 / 4.32 / 20.62</td><td>35.90 / 13.51 / 32.19</td><td>49.82 / 17.03 / 47.42</td></tr><tr><td>Window (w=2048)</td><td>805M</td><td>26.73 / 5.05 / 23.40</td><td>38.74 / 15.21 / 34.87</td><td>56.14 / 22.79 / 53.50</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>26.24 / 5.03 / 23.13</td><td>21.85 / 9.61 / 19.86</td><td>26.37 / 12.31 / 25.18</td></tr><tr><td>Block (b=4096)</td><td>805M</td><td>26.96 / 5.15 / 23.85</td><td>35.95 / 14.50 / 32.37</td><td>49.83 / 23.64 / 47.50</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>29.86 / 6.59 / 25.87</td><td>38.85 / 14.91 / 34.98</td><td>46.67 / 18.82 / 44.39</td></tr><tr><td>MASFormer (l=6)</td><td>553M</td><td>30.83 / 7.15 / 27.12</td><td>36.29 / 15.72 / 32.96</td><td>49.26 / 21.20 / 46.89</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>30.91 / 8.00 / 27.81</td><td>43.31 / 17.00 / 39.12</td><td>54.37 / 24.42 / 51.87</td></tr><tr><td>MASFormer (l=12)</td><td>906M</td><td>32.53 / 8.70 / 28.75</td><td>45.19 / 18.58 / 40.72</td><td>56.98 / 26.26 / 54.46</td></tr></table>",
+        "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": "<table><tr><td>Methods</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr><tr><td>Window attention (w=1024)</td><td>5 × 10-4</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>Window attention (w=2048)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>1 × 10-4</td></tr><tr><td>Block attention (w=2048)</td><td>1 × 10-4</td><td>1 × 10-5</td><td>1 × 10-5</td></tr><tr><td>Block attention (w=4096)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>1 × 10-5</td></tr><tr><td>MASFormer (l=4)</td><td>5 × 10-5</td><td>1 × 10-3</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=6)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=8)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=12)</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr></table>",
+        "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": "<table><tr><td>Hyperparameter</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Training steps</td><td>3000</td><td>12000</td><td>8000</td></tr><tr><td>Batch size</td><td>32</td><td>32</td><td>64</td></tr><tr><td>Input length</td><td>8192</td><td>8192</td><td>8192</td></tr><tr><td>Maximum generation length</td><td>256</td><td>512</td><td>1024</td></tr><tr><td>Weight decay</td><td>0.001</td><td>0.001</td><td>0.001</td></tr></table>",
+        "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
+    }
+]

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 4.3<sup>1</sup>. 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": "<table><tr><td>Methods</td><td>C</td><td>ArXiv</td><td>PubMed</td></tr><tr><td>Full attention</td><td>1,610M</td><td>8.63</td><td>7.63</td></tr><tr><td>Block (b=1024)</td><td>201M</td><td>13.19</td><td>12.19</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>10.75</td><td>10.13</td></tr><tr><td>MASFormer (l=2)</td><td>318M</td><td>10.25</td><td>8.75</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>9.31</td><td>8.25</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>9.63</td><td>8.25</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>8.00</td><td>19.32</td><td>28.83</td></tr><tr><td>Window(w=1024)</td><td>402M</td><td>4.32</td><td>13.51</td><td>17.03</td></tr><tr><td>Block(b=2048)</td><td>402M</td><td>5.03</td><td>9.61</td><td>12.31</td></tr><tr><td>MASFormer(l=4)</td><td>436M</td><td>6.59</td><td>14.91</td><td>18.82</td></tr><tr><td>Window(w=2048)</td><td>805M</td><td>5.05</td><td>15.21</td><td>22.79</td></tr><tr><td>Block(b=4096)</td><td>805M</td><td>5.15</td><td>14.50</td><td>23.64</td></tr><tr><td>MASFormer(l=6)</td><td>553M</td><td>7.15</td><td>15.72</td><td>21.20</td></tr><tr><td>MASFormer(l=8)</td><td>671M</td><td>7.46</td><td>17.00</td><td>24.42</td></tr><tr><td>MASFormer(l=12)</td><td>906M</td><td>8.70</td><td>18.58</td><td>26.26</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>l=8</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>w.o. C.T.</td><td>29.33/6.43/25.71</td><td>53.28/23.61/51.74</td></tr><tr><td>C.T. (n=2048)</td><td>29.87/7.16/26.15</td><td>52.28/23.01/49.83</td></tr><tr><td>C.T. (n=8192)</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Position</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>Every 3</td><td>28.26/6.94/25.03</td><td>26.16/12.37/24.82</td></tr><tr><td>Top 8</td><td>20.89/4.52/18.37</td><td>-/-/-</td></tr><tr><td>Middle 8</td><td>27.27/5.99/24.06</td><td>20.80/9.01/19.52</td></tr><tr><td>Bottom 8</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr><tr><td>Every 2</td><td>31.27/8.19/27.41</td><td>35.34/16.04/33.68</td></tr><tr><td>Bottom 12</td><td>32.53/8.70/28.75</td><td>56.98/26.26/54.46</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td rowspan=\"2\">Dataset</td><td colspan=\"3\">Example Count</td><td colspan=\"3\">Input Length</td></tr><tr><td>Train</td><td>Valid</td><td>Test</td><td>Average</td><td>Median</td><td>90th percentile</td></tr><tr><td>ArXiv</td><td>203,037</td><td>6,436</td><td>6,440</td><td>10,720</td><td>8,519</td><td>20,170</td></tr><tr><td>PubMed</td><td>119,924</td><td>6,633</td><td>6,658</td><td>4,748</td><td>3,883</td><td>8,883</td></tr><tr><td>QMSUM</td><td>1,257</td><td>272</td><td>281</td><td>9,497</td><td>14,197</td><td>27,761</td></tr><tr><td>GovReport</td><td>17,457</td><td>972</td><td>973</td><td>7,886</td><td>8,841</td><td>18,835</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>31.50 / 8.00 / 27.81</td><td>46.13 / 19.32 / 41.89</td><td>60.53 / 28.83 / 57.88</td></tr><tr><td>Window (w=1024)</td><td>402M</td><td>23.31 / 4.32 / 20.62</td><td>35.90 / 13.51 / 32.19</td><td>49.82 / 17.03 / 47.42</td></tr><tr><td>Window (w=2048)</td><td>805M</td><td>26.73 / 5.05 / 23.40</td><td>38.74 / 15.21 / 34.87</td><td>56.14 / 22.79 / 53.50</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>26.24 / 5.03 / 23.13</td><td>21.85 / 9.61 / 19.86</td><td>26.37 / 12.31 / 25.18</td></tr><tr><td>Block (b=4096)</td><td>805M</td><td>26.96 / 5.15 / 23.85</td><td>35.95 / 14.50 / 32.37</td><td>49.83 / 23.64 / 47.50</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>29.86 / 6.59 / 25.87</td><td>38.85 / 14.91 / 34.98</td><td>46.67 / 18.82 / 44.39</td></tr><tr><td>MASFormer (l=6)</td><td>553M</td><td>30.83 / 7.15 / 27.12</td><td>36.29 / 15.72 / 32.96</td><td>49.26 / 21.20 / 46.89</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>30.91 / 8.00 / 27.81</td><td>43.31 / 17.00 / 39.12</td><td>54.37 / 24.42 / 51.87</td></tr><tr><td>MASFormer (l=12)</td><td>906M</td><td>32.53 / 8.70 / 28.75</td><td>45.19 / 18.58 / 40.72</td><td>56.98 / 26.26 / 54.46</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Methods</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr><tr><td>Window attention (w=1024)</td><td>5 × 10-4</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>Window attention (w=2048)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>1 × 10-4</td></tr><tr><td>Block attention (w=2048)</td><td>1 × 10-4</td><td>1 × 10-5</td><td>1 × 10-5</td></tr><tr><td>Block attention (w=4096)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>1 × 10-5</td></tr><tr><td>MASFormer (l=4)</td><td>5 × 10-5</td><td>1 × 10-3</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=6)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=8)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=12)</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Hyperparameter</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Training steps</td><td>3000</td><td>12000</td><td>8000</td></tr><tr><td>Batch size</td><td>32</td><td>32</td><td>64</td></tr><tr><td>Input length</td><td>8192</td><td>8192</td><td>8192</td></tr><tr><td>Maximum generation length</td><td>256</td><td>512</td><td>1024</td></tr><tr><td>Weight decay</td><td>0.001</td><td>0.001</td><td>0.001</td></tr></table>"
+        },
+        {
+            "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"
+        }
+    ]
+]

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

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 4.3<sup>1</sup>. 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  $b = 2048$ .
+
+<table><tr><td>Methods</td><td>C</td><td>ArXiv</td><td>PubMed</td></tr><tr><td>Full attention</td><td>1,610M</td><td>8.63</td><td>7.63</td></tr><tr><td>Block (b=1024)</td><td>201M</td><td>13.19</td><td>12.19</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>10.75</td><td>10.13</td></tr><tr><td>MASFormer (l=2)</td><td>318M</td><td>10.25</td><td>8.75</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>9.31</td><td>8.25</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>9.63</td><td>8.25</td></tr></table>
+
+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.
+
+<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>8.00</td><td>19.32</td><td>28.83</td></tr><tr><td>Window(w=1024)</td><td>402M</td><td>4.32</td><td>13.51</td><td>17.03</td></tr><tr><td>Block(b=2048)</td><td>402M</td><td>5.03</td><td>9.61</td><td>12.31</td></tr><tr><td>MASFormer(l=4)</td><td>436M</td><td>6.59</td><td>14.91</td><td>18.82</td></tr><tr><td>Window(w=2048)</td><td>805M</td><td>5.05</td><td>15.21</td><td>22.79</td></tr><tr><td>Block(b=4096)</td><td>805M</td><td>5.15</td><td>14.50</td><td>23.64</td></tr><tr><td>MASFormer(l=6)</td><td>553M</td><td>7.15</td><td>15.72</td><td>21.20</td></tr><tr><td>MASFormer(l=8)</td><td>671M</td><td>7.46</td><td>17.00</td><td>24.42</td></tr><tr><td>MASFormer(l=12)</td><td>906M</td><td>8.70</td><td>18.58</td><td>26.26</td></tr></table>
+
+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.
+
+<table><tr><td>l=8</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>w.o. C.T.</td><td>29.33/6.43/25.71</td><td>53.28/23.61/51.74</td></tr><tr><td>C.T. (n=2048)</td><td>29.87/7.16/26.15</td><td>52.28/23.01/49.83</td></tr><tr><td>C.T. (n=8192)</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr></table>
+
+# 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.  
+
+<table><tr><td>Position</td><td>QMSUM</td><td>GovReport</td></tr><tr><td>Every 3</td><td>28.26/6.94/25.03</td><td>26.16/12.37/24.82</td></tr><tr><td>Top 8</td><td>20.89/4.52/18.37</td><td>-/-/-</td></tr><tr><td>Middle 8</td><td>27.27/5.99/24.06</td><td>20.80/9.01/19.52</td></tr><tr><td>Bottom 8</td><td>30.91/7.45/27.02</td><td>54.37/24.42/51.87</td></tr><tr><td>Every 2</td><td>31.27/8.19/27.41</td><td>35.34/16.04/33.68</td></tr><tr><td>Bottom 12</td><td>32.53/8.70/28.75</td><td>56.98/26.26/54.46</td></tr></table>
+
+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.
+
+<table><tr><td rowspan="2">Dataset</td><td colspan="3">Example Count</td><td colspan="3">Input Length</td></tr><tr><td>Train</td><td>Valid</td><td>Test</td><td>Average</td><td>Median</td><td>90th percentile</td></tr><tr><td>ArXiv</td><td>203,037</td><td>6,436</td><td>6,440</td><td>10,720</td><td>8,519</td><td>20,170</td></tr><tr><td>PubMed</td><td>119,924</td><td>6,633</td><td>6,658</td><td>4,748</td><td>3,883</td><td>8,883</td></tr><tr><td>QMSUM</td><td>1,257</td><td>272</td><td>281</td><td>9,497</td><td>14,197</td><td>27,761</td></tr><tr><td>GovReport</td><td>17,457</td><td>972</td><td>973</td><td>7,886</td><td>8,841</td><td>18,835</td></tr></table>
+
+# 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.  
+
+<table><tr><td>Methods</td><td>C</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1610M</td><td>31.50 / 8.00 / 27.81</td><td>46.13 / 19.32 / 41.89</td><td>60.53 / 28.83 / 57.88</td></tr><tr><td>Window (w=1024)</td><td>402M</td><td>23.31 / 4.32 / 20.62</td><td>35.90 / 13.51 / 32.19</td><td>49.82 / 17.03 / 47.42</td></tr><tr><td>Window (w=2048)</td><td>805M</td><td>26.73 / 5.05 / 23.40</td><td>38.74 / 15.21 / 34.87</td><td>56.14 / 22.79 / 53.50</td></tr><tr><td>Block (b=2048)</td><td>402M</td><td>26.24 / 5.03 / 23.13</td><td>21.85 / 9.61 / 19.86</td><td>26.37 / 12.31 / 25.18</td></tr><tr><td>Block (b=4096)</td><td>805M</td><td>26.96 / 5.15 / 23.85</td><td>35.95 / 14.50 / 32.37</td><td>49.83 / 23.64 / 47.50</td></tr><tr><td>MASFormer (l=4)</td><td>436M</td><td>29.86 / 6.59 / 25.87</td><td>38.85 / 14.91 / 34.98</td><td>46.67 / 18.82 / 44.39</td></tr><tr><td>MASFormer (l=6)</td><td>553M</td><td>30.83 / 7.15 / 27.12</td><td>36.29 / 15.72 / 32.96</td><td>49.26 / 21.20 / 46.89</td></tr><tr><td>MASFormer (l=8)</td><td>671M</td><td>30.91 / 8.00 / 27.81</td><td>43.31 / 17.00 / 39.12</td><td>54.37 / 24.42 / 51.87</td></tr><tr><td>MASFormer (l=12)</td><td>906M</td><td>32.53 / 8.70 / 28.75</td><td>45.19 / 18.58 / 40.72</td><td>56.98 / 26.26 / 54.46</td></tr></table>
+
+# B.2 Training Details
+
+Table 6: Finetuning performance of different attention methods.  
+
+<table><tr><td>Methods</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Full attention</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr><tr><td>Window attention (w=1024)</td><td>5 × 10-4</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>Window attention (w=2048)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>1 × 10-4</td></tr><tr><td>Block attention (w=2048)</td><td>1 × 10-4</td><td>1 × 10-5</td><td>1 × 10-5</td></tr><tr><td>Block attention (w=4096)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>1 × 10-5</td></tr><tr><td>MASFormer (l=4)</td><td>5 × 10-5</td><td>1 × 10-3</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=6)</td><td>5 × 10-5</td><td>5 × 10-5</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=8)</td><td>5 × 10-5</td><td>5 × 10-4</td><td>5 × 10-4</td></tr><tr><td>MASFormer (l=12)</td><td>1 × 10-5</td><td>1 × 10-4</td><td>1 × 10-4</td></tr></table>
+
+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.
+
+<table><tr><td>Hyperparameter</td><td>QMSUM</td><td>ArXiv</td><td>GovReport</td></tr><tr><td>Training steps</td><td>3000</td><td>12000</td><td>8000</td></tr><tr><td>Batch size</td><td>32</td><td>32</td><td>64</td></tr><tr><td>Input length</td><td>8192</td><td>8192</td><td>8192</td></tr><tr><td>Maximum generation length</td><td>256</td><td>512</td><td>1024</td></tr><tr><td>Weight decay</td><td>0.001</td><td>0.001</td><td>0.001</td></tr></table>
+
+Table 8: The other fine-tuning parameters for each dataset, which remain the same for every method.

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

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

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 T5<sub>SMALL</sub> (Raffel et al., 2020) and mT5<sub>SMALL</sub> (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.
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Lang.</td><td rowspan="2">Vocabulary</td><td rowspan="2">Parameters</td><td colspan="4">QA</td><td colspan="3">QG</td></tr><tr><td>No-Trim</td><td>Post-FT</td><td>Pre-FT</td><td>No-Trim</td><td>Post-FT</td><td>Pre-FT</td><td></td></tr><tr><td rowspan="7">mT5</td><td>EN</td><td>209K (83.6%)</td><td>258M (86.1%)</td><td>70.1 / 55.5</td><td>70.2 / 55.5</td><td>70.1 / 56.4</td><td>23.8 / 90.0</td><td>23.8 / 90.0</td><td>24.0 / 90.1</td><td></td></tr><tr><td>ES</td><td>131K (52.4%)</td><td>178M (59.4%)</td><td>55.9 / 34.7</td><td>55.9 / 34.7</td><td>57.8 / 37.5</td><td>22.7 / 84.1</td><td>22.7 / 84.1</td><td>22.3 / 84.2</td><td></td></tr><tr><td>FR</td><td>131K (52.4%)</td><td>178M (59.4%)</td><td>50.0 / 30.9</td><td>50.0 / 30.9</td><td>48.6 / 29.4</td><td>17.5 / 80.7</td><td>17.5 / 80.7</td><td>16.1 / 79.2</td><td></td></tr><tr><td>IT</td><td>111K (44.4%)</td><td>157M (52.6%)</td><td>53.2 / 37.6</td><td>53.4 / 37.8</td><td>51.5 / 36.0</td><td>17.6 / 80.8</td><td>17.6 / 80.8</td><td>17.5 / 80.6</td><td></td></tr><tr><td>JA</td><td>125K (50.0%)</td><td>172M (57.6%)</td><td>65.7 / 65.7</td><td>65.7 / 65.7</td><td>63.0 / 63.0</td><td>29.0 / 80.9</td><td>29.0 / 80.9</td><td>28.6 / 81.0</td><td></td></tr><tr><td>KO</td><td>73K (29.2%)</td><td>119M (39.7%)</td><td>77.1 / 70.6</td><td>77.1 / 70.5</td><td>74.5 / 67.3</td><td>27.5 / 82.9</td><td>27.5 / 83.0</td><td>28.0 / 83.7</td><td></td></tr><tr><td>RU</td><td>147K (58.8%)</td><td>195M (65.1%)</td><td>73.7 / 51.4</td><td>73.8 / 51.4</td><td>74.8 / 53.4</td><td>26.4 / 84.3</td><td>26.4 / 84.3</td><td>28.9 / 86.4</td><td></td></tr><tr><td rowspan="7">mBART</td><td>EN</td><td>173K (69.2%)</td><td>532M (87.1%)</td><td>76.9 / 62.6</td><td>77.0 / 62.7</td><td>78.4 / 65.7</td><td>25.1 / 90.4</td><td>25.1 / 90.4</td><td>24.7 / 90.1</td><td></td></tr><tr><td>ES</td><td>87K (34.8%)</td><td>443M (72.7%)</td><td>64.1 / 42.2</td><td>64.5 / 42.8</td><td>63.7 / 43.9</td><td>22.9 / 83.6</td><td>22.8 / 83.6</td><td>22.8 / 84.0</td><td></td></tr><tr><td>FR</td><td>85K (34.0%)</td><td>442M (72.5%)</td><td>60.4 / 39.3</td><td>61.0 / 39.8</td><td>66.4 / 45.1</td><td>19.8 / 81.7</td><td>19.8 / 81.7</td><td>18.4 / 79.7</td><td></td></tr><tr><td>IT</td><td>67K (26.8%)</td><td>424M (69.5%)</td><td>64.7 / 50.0</td><td>64.9 / 50.2</td><td>65.8 / 49.8</td><td>18.0 / 80.6</td><td>17.9 / 80.7</td><td>18.9 / 81.1</td><td></td></tr><tr><td>JA</td><td>77K (30.8%)</td><td>434M (71.1%)</td><td>68.2 / 68.2</td><td>68.2 / 68.2</td><td>70.6 / 70.6</td><td>30.0 / 82.3</td><td>29.7 / 82.1</td><td>29.1 / 80.8</td><td></td></tr><tr><td>KO</td><td>46K (18.4%)</td><td>402M (65.9%)</td><td>79.3 / 72.3</td><td>79.2 / 72.1</td><td>83.2 / 77.3</td><td>30.2 / 83.9</td><td>30.3 / 84.0</td><td>30.2 / 83.8</td><td></td></tr><tr><td>RU</td><td>99K (39.6%)</td><td>456M (74.8%)</td><td>78.7 / 58.0</td><td>79.0 / 58.2</td><td>75.5 / 49.9</td><td>29.3 / 87.2</td><td>28.7 / 87.0</td><td>28.3 / 86.7</td><td></td></tr></table>
+
+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 Tune<sup>3</sup> 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
+
+<table><tr><td rowspan="2"></td><td rowspan="2">Lang.</td><td rowspan="2">Vocabulary</td><td rowspan="2">Parameter</td><td colspan="3">Sentiment</td><td colspan="3">NLI</td></tr><tr><td>No-Trim</td><td>Post-FT</td><td>Pre-FT</td><td>No-Trim</td><td>Post-FT</td><td>Pre-FT</td></tr><tr><td rowspan="7">XLM-R</td><td>AR</td><td>49K (19.6%)</td><td>124M (44.7%)</td><td>66.3</td><td>66.3</td><td>60.9</td><td>75.7</td><td>75.7</td><td>73.8</td></tr><tr><td>DE</td><td>91K (36.4%)</td><td>156M (56.3%)</td><td>73.2</td><td>73.3</td><td>73.5</td><td>79.9</td><td>79.9</td><td>78.3</td></tr><tr><td>EN</td><td>173K (69.2%)</td><td>219M (78.7%)</td><td>68.4</td><td>68.5</td><td>67.9</td><td>84.6</td><td>84.6</td><td>70.6</td></tr><tr><td>ES</td><td>87K (34.8%)</td><td>153M (55.0%)</td><td>69.0</td><td>69.0</td><td>65.0</td><td>79.8</td><td>79.8</td><td>67.2</td></tr><tr><td>FR</td><td>85K (34.0%)</td><td>151M (54.6%)</td><td>71.8</td><td>71.8</td><td>72.1</td><td>80.1</td><td>80.1</td><td>79.6</td></tr><tr><td>IT</td><td>67K (26.8%)</td><td>138M (49.7%)</td><td>62.9</td><td>62.9</td><td>70.8</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PT</td><td>66K (26.4%)</td><td>137M (49.3%)</td><td>70.7</td><td>70.8</td><td>70.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td rowspan="7">XLM-V</td><td>AR</td><td>92K (11.8%)</td><td>157M (20.2%)</td><td>59.8</td><td>59.8</td><td>64.7</td><td>75.5</td><td>75.6</td><td>76.1</td></tr><tr><td>DE</td><td>239K (30.7%)</td><td>269M (34.7%)</td><td>73.5</td><td>73.5</td><td>73.0</td><td>78.9</td><td>78.9</td><td>79.0</td></tr><tr><td>EN</td><td>484K (62.2%)</td><td>458M (58.9%)</td><td>63.9</td><td>63.9</td><td>61.3</td><td>84.4</td><td>84.4</td><td>84.5</td></tr><tr><td>ES</td><td>243K (31.2%)</td><td>279M (35.1%)</td><td>60.7</td><td>60.7</td><td>66.6</td><td>80.7</td><td>80.7</td><td>80.6</td></tr><tr><td>FR</td><td>218K (28.0%)</td><td>253M (32.6%)</td><td>68.8</td><td>68.8</td><td>59.5</td><td>78.6</td><td>78.6</td><td>79.0</td></tr><tr><td>IT</td><td>184K (23.7%)</td><td>227M (29.3%)</td><td>70.2</td><td>70.2</td><td>74.2</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PT</td><td>181K (23.3%)</td><td>225M (28.9%)</td><td>66.6</td><td>66.5</td><td>52.8</td><td>-</td><td>-</td><td>-</td></tr></table>
+
+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 mT5<sub>SMALL</sub> 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-R<sub>BASE</sub> with  $n$  from [5K, 10K, 15K, 30K, 60K], which correspond to [89M, 93M, 97M, 109M, 132M] of parameter size, respectively<sup>4</sup>.
+
+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).<sup>8</sup> 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)
+
+<table><tr><td rowspan="3">Base Model</td><td rowspan="3">Vocab Trimming</td><td colspan="2">Vocabulary</td><td colspan="4">Social Bias</td><td>General Performance</td></tr><tr><td rowspan="2">Pre-FT</td><td rowspan="2">Post-FT</td><td colspan="2">AULA</td><td colspan="2">EEC-AULA</td><td rowspan="2">GLUE</td></tr><tr><td>CP</td><td>SS</td><td>Gender</td><td>Race</td></tr><tr><td>Monolingual (RoBERTa)</td><td>-</td><td>50K</td><td>50K</td><td>58.1</td><td>58.8</td><td>64.8</td><td>85.7</td><td>78.0</td></tr><tr><td rowspan="5">Multilingual (XLM-R)</td><td>-</td><td>250K</td><td>250K</td><td>49.5</td><td>54.9</td><td>44.3</td><td>62.0</td><td>77.9</td></tr><tr><td>Pre-FT VT (EN)</td><td>173K</td><td>173K</td><td>49.5</td><td>54.9</td><td>42.5</td><td>56.9</td><td>78.0</td></tr><tr><td>Post-FT VT (EN)</td><td>250K</td><td>173k</td><td>49.5</td><td>54.9</td><td>44.3</td><td>62.0</td><td>77.9</td></tr><tr><td>Pre-FT VT (50K)</td><td>50K</td><td>50K</td><td>49.3</td><td>55.0</td><td>41.0</td><td>62.4</td><td>76.9</td></tr><tr><td>Post-FT VT (50K)</td><td>250K</td><td>50K</td><td>49.5</td><td>54.9</td><td>44.3</td><td>62.0</td><td>77.9</td></tr></table>
+
+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
+
+<table><tr><td>Vocab.</td><td>No-Trim (250K)</td><td>5K</td><td>10K</td><td>15K</td><td>30K</td><td>60K</td></tr><tr><td>Param.</td><td>278M</td><td>89M</td><td>93M</td><td>97M</td><td>109M</td><td>132M</td></tr><tr><td rowspan="7">Post-FT (Sentiment)</td><td>AR</td><td>66.3</td><td>64.5</td><td>64.5</td><td>65.9</td><td>65.9</td></tr><tr><td>DE</td><td>73.2</td><td>70.4</td><td>72.4</td><td>72.1</td><td>73.7</td></tr><tr><td>EN</td><td>68.4</td><td>64.0</td><td>66.5</td><td>67.4</td><td>68.6</td></tr><tr><td>ES</td><td>69.0</td><td>66.2</td><td>67.2</td><td>67.8</td><td>68.4</td></tr><tr><td>FR</td><td>71.8</td><td>68.6</td><td>71.4</td><td>71.7</td><td>71.6</td></tr><tr><td>IT</td><td>62.9</td><td>60.8</td><td>61.9</td><td>63.5</td><td>62.3</td></tr><tr><td>PT</td><td>70.7</td><td>63.6</td><td>65.9</td><td>68.2</td><td>69.4</td></tr><tr><td rowspan="7">Pre-FT (Sentiment)</td><td>AR</td><td>66.3</td><td>58.6</td><td>61.9</td><td>63.1</td><td>63.2</td></tr><tr><td>DE</td><td>73.2</td><td>70.6</td><td>71.8</td><td>73.1</td><td>71.5</td></tr><tr><td>EN</td><td>68.4</td><td>64.6</td><td>66.6</td><td>67.7</td><td>66.0</td></tr><tr><td>ES</td><td>69.0</td><td>60.4</td><td>65.6</td><td>66.3</td><td>65.9</td></tr><tr><td>FR</td><td>71.8</td><td>70.3</td><td>74.3</td><td>73.8</td><td>72.2</td></tr><tr><td>IT</td><td>62.9</td><td>66.2</td><td>67.1</td><td>68.1</td><td>68.2</td></tr><tr><td>PT</td><td>70.7</td><td>65.5</td><td>67.5</td><td>69.6</td><td>67.9</td></tr><tr><td rowspan="5">Post-FT (NLI)</td><td>AR</td><td>75.7</td><td>75.0</td><td>75.8</td><td>75.8</td><td>75.7</td></tr><tr><td>DE</td><td>79.9</td><td>76.6</td><td>78.9</td><td>79.6</td><td>79.9</td></tr><tr><td>EN</td><td>84.6</td><td>80.5</td><td>83.6</td><td>84.3</td><td>84.6</td></tr><tr><td>ES</td><td>79.8</td><td>75.3</td><td>78.4</td><td>79.4</td><td>79.8</td></tr><tr><td>FR</td><td>80.1</td><td>77.2</td><td>80.0</td><td>80.1</td><td>80.2</td></tr><tr><td rowspan="5">Pre-FT (NLI)</td><td>AR</td><td>75.7</td><td>73.0</td><td>75.0</td><td>75.4</td><td>75.7</td></tr><tr><td>DE</td><td>79.9</td><td>77.4</td><td>79.0</td><td>79.0</td><td>79.6</td></tr><tr><td>EN</td><td>84.6</td><td>83.3</td><td>84.7</td><td>84.4</td><td>85.1</td></tr><tr><td>ES</td><td>79.8</td><td>78.9</td><td>79.1</td><td>79.2</td><td>81.0</td></tr><tr><td>FR</td><td>80.1</td><td>77.3</td><td>78.9</td><td>78.2</td><td>80.1</td></tr></table>
+
+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.
+
+<table><tr><td>Vocab. 
+Param.</td><td>No-Trim (250K) 
+300M</td><td>5K 
+49M</td><td>10K 
+54M</td><td>15K 
+59M</td><td>30K 
+74M</td><td>60K 
+105M</td><td>90K 
+136M</td><td>120K 
+166M</td></tr><tr><td>EN</td><td>70.1 / 55.5</td><td>72.4 / 59.7</td><td>73.2 / 60.5</td><td>73.7 / 60.9</td><td>72.7 / 58.8</td><td>71.0 / 56.4</td><td>70.5 / 55.8</td><td>70.3 / 55.6</td></tr><tr><td>ES</td><td>55.9 / 34.7</td><td>53.5 / 32.5</td><td>54.4 / 33.1</td><td>54.5 / 33.2</td><td>55.2 / 33.8</td><td>56.0 / 34.8</td><td>55.9 / 34.8</td><td>55.9 / 34.7</td></tr><tr><td>FR</td><td>50.0 / 30.9</td><td>50.0 / 31.5</td><td>48.9 / 29.9</td><td>48.6 / 29.3</td><td>49.9 / 30.1</td><td>49.8 / 30.5</td><td>50.2 / 30.9</td><td>50.0 / 30.8</td></tr><tr><td>IT</td><td>53.2 / 37.6</td><td>54.3 / 39.6</td><td>53.8 / 38.4</td><td>54.3 / 38.8</td><td>54.2 / 38.5</td><td>53.7 / 38.1</td><td>53.4 / 37.8</td><td>-</td></tr><tr><td>JA</td><td>65.7 / 65.7</td><td>54.4 / 54.4</td><td>60.3 / 60.3</td><td>62.6 / 62.6</td><td>60.4 / 60.4</td><td>64.9 / 64.9</td><td>65.6 / 65.6</td><td>65.7 / 65.7</td></tr><tr><td>KO</td><td>77.1 / 70.6</td><td>75.0 / 67.8</td><td>75.8 / 68.7</td><td>76.2 / 69.3</td><td>76.9 / 70.2</td><td>77.3 / 70.7</td><td>-</td><td>-</td></tr><tr><td>RU</td><td>73.7 / 51.4</td><td>70.1 / 48.8</td><td>71.4 / 49.5</td><td>70.3 / 47.6</td><td>70.3 / 47.9</td><td>73.8 / 51.1</td><td>73.9 / 51.4</td><td>73.8 / 51.4</td></tr><tr><td>EN</td><td>23.8 / 90.0</td><td>23.6 / 89.8</td><td>24.1 / 89.9</td><td>24.2 / 90.0</td><td>24.2 / 90.0</td><td>23.9 / 90.0</td><td>23.9 / 90.0</td><td>23.9 / 90.0</td></tr><tr><td>ES</td><td>22.7 / 84.1</td><td>21.4 / 83.6</td><td>22.0 / 83.8</td><td>22.2 / 83.9</td><td>22.4 / 84.0</td><td>22.7 / 84.1</td><td>22.7 / 84.1</td><td>22.7 / 84.1</td></tr><tr><td>FR</td><td>17.5 / 80.7</td><td>17.1 / 80.3</td><td>17.1 / 80.4</td><td>17.3 / 80.5</td><td>17.3 / 80.5</td><td>17.4 / 80.6</td><td>17.5 / 80.7</td><td>17.5 / 80.7</td></tr><tr><td>IT</td><td>17.6 / 80.8</td><td>17.3 / 80.7</td><td>17.5 / 80.8</td><td>17.5 / 80.8</td><td>17.6 / 80.7</td><td>17.6 / 80.8</td><td>17.6 / 80.8</td><td>-</td></tr><tr><td>JA</td><td>29.0 / 80.9</td><td>25.7 / 79.2</td><td>27.4 / 80.1</td><td>28.2 / 80.5</td><td>28.8 / 80.9</td><td>29.0 / 80.9</td><td>29.0 / 80.9</td><td>29.0 / 80.9</td></tr><tr><td>KO</td><td>27.5 / 82.9</td><td>27.3 / 82.8</td><td>27.4 / 82.9</td><td>27.4 / 82.9</td><td>27.4 / 82.9</td><td>27.5 / 83.0</td><td>-</td><td>-</td></tr><tr><td>RU</td><td>26.4 / 84.3</td><td>26.0 / 84.0</td><td>26.2 / 84.2</td><td>26.3 / 84.2</td><td>26.3 / 84.2</td><td>26.3 / 84.3</td><td>26.4 / 84.3</td><td>26.4 / 84.3</td></tr><tr><td>EN</td><td>70.1 / 55.5</td><td>74.0 / 61.3</td><td>71.8 / 59.1</td><td>73.4 / 60.7</td><td>74.3 / 61.3</td><td>71.3 / 56.5</td><td>68.6 / 54.3</td><td>66.2 / 52.1</td></tr><tr><td>ES</td><td>55.9 / 34.7</td><td>56.7 / 36.7</td><td>58.1 / 37.5</td><td>62.2 / 41.4</td><td>59.8 / 40.1</td><td>58.0 / 37.0</td><td>58.9 / 38.1</td><td>52.0 / 32.7</td></tr><tr><td>FR</td><td>50.0 / 30.9</td><td>49.0 / 32.7</td><td>47.6 / 30.2</td><td>43.3 / 26.9</td><td>44.9 / 27.4</td><td>34.3 / 20.0</td><td>43.1 / 24.1</td><td>40.0 / 23.7</td></tr><tr><td>IT</td><td>53.2 / 37.6</td><td>57.6 / 43.7</td><td>61.5 / 46.8</td><td>61.2 / 45.6</td><td>57.8 / 42.1</td><td>56.5 / 40.5</td><td>55.4 / 39.2</td><td>-</td></tr><tr><td>JA</td><td>65.7 / 65.7</td><td>55.3 / 55.3</td><td>52.2 / 52.2</td><td>61.3 / 61.3</td><td>60.4 / 60.4</td><td>64.6 / 64.6</td><td>66.9 / 66.9</td><td>67.2 / 67.2</td></tr><tr><td>KO</td><td>77.1 / 70.6</td><td>79.5 / 72.6</td><td>83.7 / 77.7</td><td>82.7 / 76.4</td><td>80.4 / 73.7</td><td>81.6 / 75.1</td><td>-</td><td>-</td></tr><tr><td>RU</td><td>73.7 / 51.4</td><td>73.5 / 51.7</td><td>76.9 / 56.4</td><td>77.4 / 56.9</td><td>75.9 / 54.2</td><td>75.1 / 53.0</td><td>75.0 / 53.0</td><td>73.3 / 51.4</td></tr><tr><td>EN</td><td>23.8 / 90.0</td><td>24.8 / 90.0</td><td>24.4 / 89.9</td><td>24.3 / 89.9</td><td>24.2 / 90.0</td><td>24.2 / 90.0</td><td>23.3 / 89.9</td><td>23.6 / 89.8</td></tr><tr><td>ES</td><td>22.7 / 84.1</td><td>21.9 / 84.1</td><td>21.9 / 83.9</td><td>22.7 / 83.7</td><td>22.0 / 84.3</td><td>22.6 / 84.4</td><td>20.4 / 79.6</td><td>22.6 / 84.1</td></tr><tr><td>FR</td><td>17.5 / 80.7</td><td>16.1 / 79.0</td><td>16.9 / 79.9</td><td>16.9 / 79.5</td><td>15.8 / 78.8</td><td>14.6 / 78.0</td><td>15.7 / 79.0</td><td>17.0 / 79.2</td></tr><tr><td>IT</td><td>17.6 / 80.8</td><td>17.4 / 80.4</td><td>17.8 / 81.1</td><td>18.0 / 80.8</td><td>17.1 / 80.6</td><td>17.4 / 80.9</td><td>17.4 / 80.8</td><td>-</td></tr><tr><td>JA</td><td>29.0 / 80.9</td><td>26.6 / 79.8</td><td>27.0 / 79.2</td><td>27.5 / 79.7</td><td>27.3 / 79.9</td><td>27.7 / 80.3</td><td>28.3 / 80.8</td><td>28.2 / 80.2</td></tr><tr><td>KO</td><td>27.5 / 82.9</td><td>28.4 / 83.4</td><td>27.8 / 83.4</td><td>28.2 / 84.1</td><td>28.8 / 83.4</td><td>28.4 / 83.4</td><td>-</td><td>-</td></tr><tr><td>RU</td><td>26.4 / 84.3</td><td>27.7 / 85.9</td><td>30.0 / 86.6</td><td>29.0 / 86.5</td><td>29.2 / 86.3</td><td>29.0 / 86.6</td><td>29.2 / 86.7</td><td>28.9 / 86.0</td></tr></table>
+
+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><tr><td></td><td>Model</td><td>AULA-EEC</td><td>Anger</td><td>Sadness</td><td>Fear</td><td>Joy</td><td>No Emotion Type</td></tr><tr><td rowspan="6">Gender Bias</td><td>FT RoBERTa</td><td>64.8</td><td>50.0</td><td>67.8</td><td>71.5</td><td>64.3</td><td>60.8</td></tr><tr><td>FT XLM-R</td><td>44.3</td><td>46.9</td><td>39.4</td><td>51.4</td><td>34.4</td><td>39.7</td></tr><tr><td>Pre-FT XLM-R (EN)</td><td>42.5</td><td>33.8</td><td>55.8</td><td>45.1</td><td>40.7</td><td>42.7</td></tr><tr><td>Post-FT XLM-R (EN)</td><td>44.3</td><td>46.9</td><td>39.4</td><td>51.4</td><td>34.4</td><td>39.7</td></tr><tr><td>Pre-FT XLM-R (50k)</td><td>44.3</td><td>46.9</td><td>39.4</td><td>51.4</td><td>34.4</td><td>39.7</td></tr><tr><td>Post-FT XLM-R (50k)</td><td>41.0</td><td>41.2</td><td>45.8</td><td>47.2</td><td>36.9</td><td>37.2</td></tr><tr><td rowspan="6">Race Bias</td><td>FT RoBERTa</td><td>85.7</td><td>50.0</td><td>88.0</td><td>65.4</td><td>50.0</td><td>100.0</td></tr><tr><td>FT XLM-R</td><td>62.0</td><td>74.5</td><td>32.1</td><td>47.3</td><td>74.6</td><td>17.4</td></tr><tr><td>Pre-FT VT XLM-R (EN)</td><td>56.9</td><td>33.7</td><td>58.8</td><td>69.2</td><td>60.1</td><td>50.0</td></tr><tr><td>Pre-FT VT XLM-R (EN)</td><td>62.0</td><td>74.5</td><td>32.1</td><td>47.3</td><td>74.6</td><td>17.4</td></tr><tr><td>Pre-FT VT XLM-R (50k)</td><td>62.0</td><td>74.5</td><td>32.1</td><td>47.3</td><td>74.6</td><td>17.4</td></tr><tr><td>Post-FT VT XLM-R (50k)</td><td>62.4</td><td>55.7</td><td>70.3</td><td>70.9</td><td>59.1</td><td>59.0</td></tr></table>
+
+Table 6: Social bias scores of LMs fine-tuned on sentiment analysis on each emotion type w/wo VT on the EEC dataset.
+
+<table><tr><td></td><td></td><td>RoBERTa</td><td>XLM-R</td><td>VT XLM-R (EN)</td><td>VT XLM-R (50K)</td></tr><tr><td rowspan="10">CP Dataset</td><td>AULA</td><td>58.1</td><td>49.5</td><td>49.5</td><td>49.3</td></tr><tr><td>Age</td><td>59.8</td><td>49.4</td><td>49.4</td><td>49.4</td></tr><tr><td>Disability</td><td>68.3</td><td>66.7</td><td>66.7</td><td>68.3</td></tr><tr><td>Gender</td><td>53.4</td><td>48.5</td><td>48.5</td><td>48.1</td></tr><tr><td>Nationality</td><td>56.6</td><td>44.0</td><td>44.0</td><td>44.7</td></tr><tr><td>Physical-Appearance</td><td>52.4</td><td>58.7</td><td>58.7</td><td>55.6</td></tr><tr><td>Race-Color</td><td>56.8</td><td>43.8</td><td>43.8</td><td>44.0</td></tr><tr><td>Religion</td><td>53.3</td><td>54.3</td><td>54.3</td><td>53.3</td></tr><tr><td>Sexual-Orientation</td><td>67.9</td><td>54.8</td><td>54.8</td><td>54.8</td></tr><tr><td>Socioeconomic</td><td>66.3</td><td>58.1</td><td>58.1</td><td>57.6</td></tr><tr><td rowspan="5">SS Dataset</td><td>AULA</td><td>58.8</td><td>54.9</td><td>54.9</td><td>55.0</td></tr><tr><td>Gender</td><td>62.4</td><td>55.3</td><td>55.3</td><td>54.9</td></tr><tr><td>Profession</td><td>60.4</td><td>54.3</td><td>54.3</td><td>54.6</td></tr><tr><td>Race</td><td>57.0</td><td>56.0</td><td>56.0</td><td>56.0</td></tr><tr><td>Religion</td><td>54.4</td><td>46.8</td><td>46.8</td><td>46.8</td></tr></table>
+
+Table 7: Social bias scores of VT LMs compared to their original ones on both CP and SS datasets for each bias type.

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

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

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 score<sup>2</sup>. 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
+
+<table><tr><td>Dataset</td><td>Domain</td><td>|I|</td><td>(|Qtrain|, |Qtest|) Splits</td><td>Train Query (Qtrain) Type</td></tr><tr><td>ZESHEL</td><td>YuGiOh</td><td>10,031</td><td>(100/3274), (500/2874), (2000/1374)</td><td>Real Queries</td></tr><tr><td>ZESHEL</td><td>StarTrek</td><td>34,430</td><td>(100/4127), (500/3727), (2000/2227)</td><td>Real Queries</td></tr><tr><td>ZESHEL</td><td>Military</td><td>104,520</td><td>(100/2300), (500/1900), (2000/0400)</td><td>Real Queries</td></tr><tr><td>BEIR</td><td>SciDocs</td><td>25,657</td><td>(1000/1000)</td><td>Pseudo-Queries</td></tr><tr><td>BEIR</td><td>HotpotQA</td><td>5,233,329</td><td>(1000/1000)</td><td>Pseudo-Queries</td></tr></table>
+
+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) SoftMax<sub>0,ε</sub>, 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)

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

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

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

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": "<table><tr><td>Task</td><td>Parameters</td><td>Tokens</td><td>Interaction</td></tr><tr><td>Hindsight Neglect</td><td>t(60)=-4.22, p&lt;0.001</td><td>t(60)=-4.69, p&lt;0.001</td><td>t(60)=-2.88, p=0.012</td></tr><tr><td>Into the Unknown</td><td>t(60)=-0.31, p=0.824</td><td>t(60)=2.04, p=0.079</td><td>t(60)=0.02, p=0.986</td></tr><tr><td>Memo Trap</td><td>t(60)=-10.05, p&lt;0.001</td><td>t(60)=-11.34, p&lt;0.001</td><td>t(60)=-9.71, p&lt;0.001</td></tr><tr><td>Modus Tollens</td><td>t(60)=-0.16, p=0.927</td><td>t(60)=-2.13, p=0.071</td><td>t(60)=-1.50, p=0.208</td></tr><tr><td>NeQA</td><td>t(60)=0.79, p=0.559</td><td>t(60)=-0.08, p=0.963</td><td>t(60)=2.45, p=0.034</td></tr><tr><td>Pattern Match Supp.</td><td>t(60)=-3.20, p=0.005</td><td>t(60)=-9.58, p&lt;0.001</td><td>t(60)=-6.37, p&lt;0.001</td></tr><tr><td>Redefine</td><td>t(60)=0.60, p=0.645</td><td>t(60)=-0.79, p=0.559</td><td>t(60)=-1.53, p=0.205</td></tr><tr><td>Repetitive Algebra</td><td>t(60)=-0.49, p=0.706</td><td>t(60)=-2.11, p=0.071</td><td>t(60)=-1.00, p=0.443</td></tr><tr><td>Resisting Correction</td><td>t(60)=5.13, p&lt;0.001</td><td>t(60)=5.63, p&lt;0.001</td><td>t(60)=-1.89, p=0.104</td></tr><tr><td>Sig Figs</td><td>t(60)=-0.59, p=0.645</td><td>t(60)=-0.74, p=0.574</td><td>t(60)=-1.46, p=0.215</td></tr><tr><td>TruthfulQA-MC1</td><td>t(60)=-10.90, p&lt;0.001</td><td>t(60)=-11.45, p&lt;0.001</td><td>t(60)=-2.97, p=0.010</td></tr><tr><td>TruthfulQA-MC2</td><td>t(60)=-24.72, p&lt;0.001</td><td>t(60)=-23.89, p&lt;0.001</td><td>t(60)=-12.02, p&lt;0.001</td></tr></table>",
+        "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
+    }
+]

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": "<table><tr><td>Task</td><td>Parameters</td><td>Tokens</td><td>Interaction</td></tr><tr><td>Hindsight Neglect</td><td>t(60)=-4.22, p&lt;0.001</td><td>t(60)=-4.69, p&lt;0.001</td><td>t(60)=-2.88, p=0.012</td></tr><tr><td>Into the Unknown</td><td>t(60)=-0.31, p=0.824</td><td>t(60)=2.04, p=0.079</td><td>t(60)=0.02, p=0.986</td></tr><tr><td>Memo Trap</td><td>t(60)=-10.05, p&lt;0.001</td><td>t(60)=-11.34, p&lt;0.001</td><td>t(60)=-9.71, p&lt;0.001</td></tr><tr><td>Modus Tollens</td><td>t(60)=-0.16, p=0.927</td><td>t(60)=-2.13, p=0.071</td><td>t(60)=-1.50, p=0.208</td></tr><tr><td>NeQA</td><td>t(60)=0.79, p=0.559</td><td>t(60)=-0.08, p=0.963</td><td>t(60)=2.45, p=0.034</td></tr><tr><td>Pattern Match Supp.</td><td>t(60)=-3.20, p=0.005</td><td>t(60)=-9.58, p&lt;0.001</td><td>t(60)=-6.37, p&lt;0.001</td></tr><tr><td>Redefine</td><td>t(60)=0.60, p=0.645</td><td>t(60)=-0.79, p=0.559</td><td>t(60)=-1.53, p=0.205</td></tr><tr><td>Repetitive Algebra</td><td>t(60)=-0.49, p=0.706</td><td>t(60)=-2.11, p=0.071</td><td>t(60)=-1.00, p=0.443</td></tr><tr><td>Resisting Correction</td><td>t(60)=5.13, p&lt;0.001</td><td>t(60)=5.63, p&lt;0.001</td><td>t(60)=-1.89, p=0.104</td></tr><tr><td>Sig Figs</td><td>t(60)=-0.59, p=0.645</td><td>t(60)=-0.74, p=0.574</td><td>t(60)=-1.46, p=0.215</td></tr><tr><td>TruthfulQA-MC1</td><td>t(60)=-10.90, p&lt;0.001</td><td>t(60)=-11.45, p&lt;0.001</td><td>t(60)=-2.97, p=0.010</td></tr><tr><td>TruthfulQA-MC2</td><td>t(60)=-24.72, p&lt;0.001</td><td>t(60)=-23.89, p&lt;0.001</td><td>t(60)=-12.02, p&lt;0.001</td></tr></table>"
+        },
+        {
+            "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"
+        }
+    ]
+]

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

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).
+
+<table><tr><td>Task</td><td>Parameters</td><td>Tokens</td><td>Interaction</td></tr><tr><td>Hindsight Neglect</td><td>t(60)=-4.22, p&lt;0.001</td><td>t(60)=-4.69, p&lt;0.001</td><td>t(60)=-2.88, p=0.012</td></tr><tr><td>Into the Unknown</td><td>t(60)=-0.31, p=0.824</td><td>t(60)=2.04, p=0.079</td><td>t(60)=0.02, p=0.986</td></tr><tr><td>Memo Trap</td><td>t(60)=-10.05, p&lt;0.001</td><td>t(60)=-11.34, p&lt;0.001</td><td>t(60)=-9.71, p&lt;0.001</td></tr><tr><td>Modus Tollens</td><td>t(60)=-0.16, p=0.927</td><td>t(60)=-2.13, p=0.071</td><td>t(60)=-1.50, p=0.208</td></tr><tr><td>NeQA</td><td>t(60)=0.79, p=0.559</td><td>t(60)=-0.08, p=0.963</td><td>t(60)=2.45, p=0.034</td></tr><tr><td>Pattern Match Supp.</td><td>t(60)=-3.20, p=0.005</td><td>t(60)=-9.58, p&lt;0.001</td><td>t(60)=-6.37, p&lt;0.001</td></tr><tr><td>Redefine</td><td>t(60)=0.60, p=0.645</td><td>t(60)=-0.79, p=0.559</td><td>t(60)=-1.53, p=0.205</td></tr><tr><td>Repetitive Algebra</td><td>t(60)=-0.49, p=0.706</td><td>t(60)=-2.11, p=0.071</td><td>t(60)=-1.00, p=0.443</td></tr><tr><td>Resisting Correction</td><td>t(60)=5.13, p&lt;0.001</td><td>t(60)=5.63, p&lt;0.001</td><td>t(60)=-1.89, p=0.104</td></tr><tr><td>Sig Figs</td><td>t(60)=-0.59, p=0.645</td><td>t(60)=-0.74, p=0.574</td><td>t(60)=-1.46, p=0.215</td></tr><tr><td>TruthfulQA-MC1</td><td>t(60)=-10.90, p&lt;0.001</td><td>t(60)=-11.45, p&lt;0.001</td><td>t(60)=-2.97, p=0.010</td></tr><tr><td>TruthfulQA-MC2</td><td>t(60)=-24.72, p&lt;0.001</td><td>t(60)=-23.89, p&lt;0.001</td><td>t(60)=-12.02, p&lt;0.001</td></tr></table>
+
+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.

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

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": "<sup>2</sup>University 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": "<sup>1</sup>https://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": "<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr></table>",
+        "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)<sup>3</sup>, Vicuna<sup>4</sup>, 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": "<sup>4</sup>https://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": "<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>Full training*</td><td>75.0</td><td>24.0</td></tr><tr><td>Vicuna</td><td>73.810.95</td><td>11.230.78</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + ZCoT</td><td>77.101.21</td><td>20.211.02</td></tr><tr><td>ChatGPT + DoT</td><td>81.190.11</td><td>22.250.70</td></tr><tr><td>GPT-4</td><td>83.040.51</td><td>33.860.83</td></tr><tr><td>GPT-4 + ZCoT</td><td>81.971.21</td><td>33.221.36</td></tr><tr><td>GPT-4 + DoT</td><td>82.770.81</td><td>34.641.40</td></tr></table>",
+        "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": "<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + S1</td><td>79.621.12</td><td>18.721.95</td></tr><tr><td>ChatGPT + S1 + S2</td><td>80.700.48</td><td>20.111.02</td></tr><tr><td>ChatGPT + S1 + S2 + S3</td><td>81.190.11</td><td>22.250.70</td></tr></table>",
+        "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": "<table><tr><td>Methods</td><td>Quality</td><td>Subjectivity Assessment</td><td>Contrastive Reasoning</td><td>Schema Analysis</td></tr><tr><td rowspan=\"3\">ChatGPT +DoT</td><td>Comprehensive</td><td>68.5%</td><td>45.0%</td><td>50.5%</td></tr><tr><td>Partially good</td><td>24.5%</td><td>30.5%</td><td>27.0%</td></tr><tr><td>Invalid</td><td>7.0%</td><td>24.5%</td><td>22.5%</td></tr><tr><td rowspan=\"3\">GPT4 +DoT</td><td>Comprehensive</td><td>84.5%</td><td>69.5%</td><td>76.5%</td></tr><tr><td>Partially good</td><td>13.5%</td><td>25.5%</td><td>18.5%</td></tr><tr><td>Invalid</td><td>2.0%</td><td>5.0%</td><td>5.0%</td></tr></table>",
+        "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 UpWork<sup>5</sup>, 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 Eliza<sup>6</sup> 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 UpWork<sup>8</sup>, 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": "<sup>7</sup>https://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": "<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>Overgeneralization</td><td>Major conclusions are drawn based on limited information.</td><td>Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr><tr><td>Emotional reasoning</td><td>Letting one&#x27;s feeling about something overrule facts to the contrary.</td><td>Even though Steve is here at work late every day, I know I work harder than anyone else at my job.</td></tr><tr><td>Labeling</td><td>Giving someone or something a label without finding out more about it/them.</td><td>My daughter would never do anything I disapproved of.</td></tr><tr><td>Magnification</td><td>Emphasizing the negative or playing down the positive of a situation.</td><td>My professor said he made some corrections on my paper, so I know I&#x27;ll probably fail the class.</td></tr><tr><td>Mental filter</td><td>Placing all one&#x27;s attention o, or seeing only, the negatives of a situation.</td><td>My husband says he wishes I was better at housekeeping, so I must be a lousy wife.</td></tr><tr><td>Should statements</td><td>Should 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 &#x27;should&#x27; or it&#x27;s synonyms (ought to, must etc.) be present in the statements containing this distortion.</td><td>I should get all A&#x27;s to be a good student.</td></tr><tr><td>Fortune-telling</td><td>As 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.</td><td>I was afraid of job interviews so I decided to start my own thing.</td></tr></table>",
+        "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": "<table><tr><td>Objective Fact:</td><td>The patient has been feeling extremely unhappy and empty lately.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>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.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>",
+        "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": "<table><tr><td>Objective Fact:</td><td>The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td>Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.</td></tr><tr><td>Reasoning Process</td><td rowspan=\"2\">Emotions 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&#x27;s future. positive events and achievements are also likely to occur.</td></tr><tr><td>Contradicting the Thoughts:</td></tr><tr><td>Cognition Schema:</td><td>Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life&#x27;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.</td></tr></table>",
+        "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": "<table><tr><td>Objective Fact:</td><td>The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.</td></tr><tr><td>Subjective Thought:</td><td>The patient feels that they are the reason for their friends leaving and wonders if it&#x27;s their personality or behavior causing the issue. they cannot stand being alone.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>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.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>There may be other reasons for the loss of these friendships that the patient is not considering.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>",
+        "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": "<table><tr><td>Objective Fact:</td><td>The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.</td></tr><tr><td>Subjective Thought:</td><td>The patient believes that their personality or behavior might be the reason their friendships don&#x27;t last, and they are limited in opportunities to make new friends.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td></td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Contradicting the Thoughts:</td><td></td></tr><tr><td>Cognition Schema:</td><td>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&#x27;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.</td></tr></table>",
+        "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
+    }
+]

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": "<sup>2</sup>University 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": "<sup>1</sup>https://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": "<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr></table>"
+        },
+        {
+            "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)<sup>3</sup>, Vicuna<sup>4</sup>, 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": "<sup>4</sup>https://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": "<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>Full training*</td><td>75.0</td><td>24.0</td></tr><tr><td>Vicuna</td><td>73.810.95</td><td>11.230.78</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + ZCoT</td><td>77.101.21</td><td>20.211.02</td></tr><tr><td>ChatGPT + DoT</td><td>81.190.11</td><td>22.250.70</td></tr><tr><td>GPT-4</td><td>83.040.51</td><td>33.860.83</td></tr><tr><td>GPT-4 + ZCoT</td><td>81.971.21</td><td>33.221.36</td></tr><tr><td>GPT-4 + DoT</td><td>82.770.81</td><td>34.641.40</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + S1</td><td>79.621.12</td><td>18.721.95</td></tr><tr><td>ChatGPT + S1 + S2</td><td>80.700.48</td><td>20.111.02</td></tr><tr><td>ChatGPT + S1 + S2 + S3</td><td>81.190.11</td><td>22.250.70</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Methods</td><td>Quality</td><td>Subjectivity Assessment</td><td>Contrastive Reasoning</td><td>Schema Analysis</td></tr><tr><td rowspan=\"3\">ChatGPT +DoT</td><td>Comprehensive</td><td>68.5%</td><td>45.0%</td><td>50.5%</td></tr><tr><td>Partially good</td><td>24.5%</td><td>30.5%</td><td>27.0%</td></tr><tr><td>Invalid</td><td>7.0%</td><td>24.5%</td><td>22.5%</td></tr><tr><td rowspan=\"3\">GPT4 +DoT</td><td>Comprehensive</td><td>84.5%</td><td>69.5%</td><td>76.5%</td></tr><tr><td>Partially good</td><td>13.5%</td><td>25.5%</td><td>18.5%</td></tr><tr><td>Invalid</td><td>2.0%</td><td>5.0%</td><td>5.0%</td></tr></table>"
+        },
+        {
+            "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 UpWork<sup>5</sup>, 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 Eliza<sup>6</sup> 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 UpWork<sup>8</sup>, 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": "<sup>7</sup>https://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": "<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>Overgeneralization</td><td>Major conclusions are drawn based on limited information.</td><td>Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr><tr><td>Emotional reasoning</td><td>Letting one&#x27;s feeling about something overrule facts to the contrary.</td><td>Even though Steve is here at work late every day, I know I work harder than anyone else at my job.</td></tr><tr><td>Labeling</td><td>Giving someone or something a label without finding out more about it/them.</td><td>My daughter would never do anything I disapproved of.</td></tr><tr><td>Magnification</td><td>Emphasizing the negative or playing down the positive of a situation.</td><td>My professor said he made some corrections on my paper, so I know I&#x27;ll probably fail the class.</td></tr><tr><td>Mental filter</td><td>Placing all one&#x27;s attention o, or seeing only, the negatives of a situation.</td><td>My husband says he wishes I was better at housekeeping, so I must be a lousy wife.</td></tr><tr><td>Should statements</td><td>Should 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 &#x27;should&#x27; or it&#x27;s synonyms (ought to, must etc.) be present in the statements containing this distortion.</td><td>I should get all A&#x27;s to be a good student.</td></tr><tr><td>Fortune-telling</td><td>As 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.</td><td>I was afraid of job interviews so I decided to start my own thing.</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Objective Fact:</td><td>The patient has been feeling extremely unhappy and empty lately.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>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.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Objective Fact:</td><td>The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td>Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.</td></tr><tr><td>Reasoning Process</td><td rowspan=\"2\">Emotions 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&#x27;s future. positive events and achievements are also likely to occur.</td></tr><tr><td>Contradicting the Thoughts:</td></tr><tr><td>Cognition Schema:</td><td>Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life&#x27;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.</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Objective Fact:</td><td>The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.</td></tr><tr><td>Subjective Thought:</td><td>The patient feels that they are the reason for their friends leaving and wonders if it&#x27;s their personality or behavior causing the issue. they cannot stand being alone.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>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.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>There may be other reasons for the loss of these friendships that the patient is not considering.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Objective Fact:</td><td>The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.</td></tr><tr><td>Subjective Thought:</td><td>The patient believes that their personality or behavior might be the reason their friendships don&#x27;t last, and they are limited in opportunities to make new friends.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td></td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Contradicting the Thoughts:</td><td></td></tr><tr><td>Cognition Schema:</td><td>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&#x27;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.</td></tr></table>"
+        },
+        {
+            "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"
+        }
+    ]
+]

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

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
+
+<sup>2</sup>University 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
+
+<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr></table>
+
+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)<sup>3</sup>, Vicuna<sup>4</sup>, 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
+
+<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>Full training*</td><td>75.0</td><td>24.0</td></tr><tr><td>Vicuna</td><td>73.810.95</td><td>11.230.78</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + ZCoT</td><td>77.101.21</td><td>20.211.02</td></tr><tr><td>ChatGPT + DoT</td><td>81.190.11</td><td>22.250.70</td></tr><tr><td>GPT-4</td><td>83.040.51</td><td>33.860.83</td></tr><tr><td>GPT-4 + ZCoT</td><td>81.971.21</td><td>33.221.36</td></tr><tr><td>GPT-4 + DoT</td><td>82.770.81</td><td>34.641.40</td></tr></table>
+
+Table 2: Main Results (Standard deviation in subscript).  $(^{\star})$  results copied from (Shreevastava and Foltz, 2021).  
+
+<table><tr><td>Methods</td><td>Distortion Assessment (F-1)</td><td>Distortion Classification (Weighted F-1)</td></tr><tr><td>ChatGPT</td><td>73.470.58</td><td>19.241.00</td></tr><tr><td>ChatGPT + S1</td><td>79.621.12</td><td>18.721.95</td></tr><tr><td>ChatGPT + S1 + S2</td><td>80.700.48</td><td>20.111.02</td></tr><tr><td>ChatGPT + S1 + S2 + S3</td><td>81.190.11</td><td>22.250.70</td></tr></table>
+
+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.  
+
+<table><tr><td>Methods</td><td>Quality</td><td>Subjectivity Assessment</td><td>Contrastive Reasoning</td><td>Schema Analysis</td></tr><tr><td rowspan="3">ChatGPT +DoT</td><td>Comprehensive</td><td>68.5%</td><td>45.0%</td><td>50.5%</td></tr><tr><td>Partially good</td><td>24.5%</td><td>30.5%</td><td>27.0%</td></tr><tr><td>Invalid</td><td>7.0%</td><td>24.5%</td><td>22.5%</td></tr><tr><td rowspan="3">GPT4 +DoT</td><td>Comprehensive</td><td>84.5%</td><td>69.5%</td><td>76.5%</td></tr><tr><td>Partially good</td><td>13.5%</td><td>25.5%</td><td>18.5%</td></tr><tr><td>Invalid</td><td>2.0%</td><td>5.0%</td><td>5.0%</td></tr></table>
+
+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 UpWork<sup>5</sup>, 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 Eliza<sup>6</sup> 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 UpWork<sup>8</sup>, 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
+
+<table><tr><td>Cognitive Distortion Type</td><td>Interpretation</td><td>Example Distorted Speech</td></tr><tr><td>Personalization</td><td>Personalizing or taking up the blame for a situation, that in reality involved many factors and was out of the person&#x27;s control.</td><td>My son is pretty quiet today. I wonder what I did to upset him.</td></tr><tr><td>Mind Reading</td><td>Suspecting what others are thinking or what are the motivations behind their actions.</td><td>My house was dirty when my friends came over, they must think I&#x27;m a slob!</td></tr><tr><td>Overgeneralization</td><td>Major conclusions are drawn based on limited information.</td><td>Last time I was in the pool I almost drowned, I am a terrible swimmer and should not go into the water again.</td></tr><tr><td>All-or-nothing thinking</td><td>Looking at a situation as either black or white or thinking that there are only two possible outcomes to a situation.</td><td>If I cannot get my Ph.D., then I am a total failure.</td></tr><tr><td>Emotional reasoning</td><td>Letting one&#x27;s feeling about something overrule facts to the contrary.</td><td>Even though Steve is here at work late every day, I know I work harder than anyone else at my job.</td></tr><tr><td>Labeling</td><td>Giving someone or something a label without finding out more about it/them.</td><td>My daughter would never do anything I disapproved of.</td></tr><tr><td>Magnification</td><td>Emphasizing the negative or playing down the positive of a situation.</td><td>My professor said he made some corrections on my paper, so I know I&#x27;ll probably fail the class.</td></tr><tr><td>Mental filter</td><td>Placing all one&#x27;s attention o, or seeing only, the negatives of a situation.</td><td>My husband says he wishes I was better at housekeeping, so I must be a lousy wife.</td></tr><tr><td>Should statements</td><td>Should 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 &#x27;should&#x27; or it&#x27;s synonyms (ought to, must etc.) be present in the statements containing this distortion.</td><td>I should get all A&#x27;s to be a good student.</td></tr><tr><td>Fortune-telling</td><td>As 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.</td><td>I was afraid of job interviews so I decided to start my own thing.</td></tr></table>
+
+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):  
+
+<table><tr><td>Objective Fact:</td><td>The patient has been feeling extremely unhappy and empty lately.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>The patient may believe that their thoughts are true because they are experiencing these feelings and emotions firsthand.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>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.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>
+
+Diagnosis of Thought (GPT4):  
+
+<table><tr><td>Objective Fact:</td><td>The patient has been extremely unhappy recently. The patient feels empty and alone most of the time. he patient has friends and close friends.</td></tr><tr><td>Subjective Thought:</td><td>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.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td>Feeling alone even when surrounded by friends could cause them to think that their friendships are not fulfilling or meaningful.</td></tr><tr><td>Reasoning Process</td><td rowspan="2">Emotions 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&#x27;s future. positive events and achievements are also likely to occur.</td></tr><tr><td>Contradicting the Thoughts:</td></tr><tr><td>Cognition Schema:</td><td>Emotional reasoning: they believe that their feelings of unhappiness, emptiness, and loneliness are true reflections of their life&#x27;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.</td></tr></table>
+
+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):  
+
+<table><tr><td>Objective Fact:</td><td>The patient has no close friends at the moment and feels lonely. The patient has made acquaintances but no close friends.</td></tr><tr><td>Subjective Thought:</td><td>The patient feels that they are the reason for their friends leaving and wonders if it&#x27;s their personality or behavior causing the issue. they cannot stand being alone.</td></tr><tr><td>Reasoning Process Supporting the Thoughts:</td><td>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.</td></tr><tr><td>Reasoning Process Contradicting the Thoughts:</td><td>There may be other reasons for the loss of these friendships that the patient is not considering.</td></tr><tr><td>Cognition Schema:</td><td>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.</td></tr></table>
+
+Diagnosis of Thought (GPT4):  
+
+<table><tr><td>Objective Fact:</td><td>The patient currently does not have close friends and mainly interacts with school friends. they feel lonely and seek help.</td></tr><tr><td>Subjective Thought:</td><td>The patient believes that their personality or behavior might be the reason their friendships don&#x27;t last, and they are limited in opportunities to make new friends.</td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Supporting the Thoughts:</td><td></td></tr><tr><td>Reasoning Process</td><td>The 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.</td></tr><tr><td>Contradicting the Thoughts:</td><td></td></tr><tr><td>Cognition Schema:</td><td>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&#x27;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.</td></tr></table>
+
+Figure 3: Showcases of Diagnosis of Thought using ChatGPT and GPT4 given the Patient Message.

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

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.<sup>1</sup> 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,<sup>2</sup> 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.<sup>3</sup> 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": "<sup>1</sup>Called \"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
+    }
+]

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.<sup>1</sup> 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,<sup>2</sup> 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.<sup>3</sup> 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": "<sup>1</sup>Called \"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
+        }
+    ]
+]

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

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.<sup>1</sup> 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,<sup>2</sup> 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.<sup>3</sup> 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.

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": "<sup>1</sup>Harbin 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": "<sup>1</sup>The 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": "<table><tr><td>Language</td><td>#Word</td><td>#Image</td><td>Average</td></tr><tr><td>En</td><td>4,975</td><td>45,228</td><td>9.09</td></tr><tr><td>Fr</td><td>4,646</td><td>36,927</td><td>7.95</td></tr><tr><td>De</td><td>10,384</td><td>85,419</td><td>8.23</td></tr><tr><td>Cs</td><td>10,706</td><td>72,942</td><td>6.81</td></tr><tr><td>Zh</td><td>10,805</td><td>85,144</td><td>7.88</td></tr><tr><td>Total</td><td>41,516</td><td>325,660</td><td>7.84</td></tr></table>",
+        "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 Spacy<sup>3</sup> and Stanford CoreNLP<sup>4</sup>. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images<sup>5</sup> 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 codebase<sup>6</sup>. 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": "<sup>3</sup>https://github.com/explosion/spaCy",
+        "bbox": [
+            134,
+            877,
+            361,
+            891
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://github.com/stanfordnlp/CoreNLP",
+        "bbox": [
+            136,
+            891,
+            391,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>https://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": "<table><tr><td></td><td colspan=\"4\">En-Fr</td><td colspan=\"4\">En-De</td><td colspan=\"2\">En-Cs</td></tr><tr><td>Approach</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F18</td></tr><tr><td>UMMT</td><td>39.80*</td><td>-</td><td>-</td><td>-</td><td>23.50*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PVP</td><td>52.30*</td><td>-</td><td>-</td><td>-</td><td>33.90*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>XLM</td><td>53.08</td><td>46.82</td><td>35.31</td><td>43.09</td><td>31.81</td><td>26.83</td><td>25.62</td><td>22.95</td><td>25.14</td><td>20.15</td></tr><tr><td>UNMT-CS</td><td>54.24</td><td>46.30</td><td>36.06</td><td>43.49</td><td>33.43</td><td>27.51</td><td>26.67</td><td>24.27</td><td>26.46</td><td>22.71</td></tr><tr><td>Ours</td><td>54.07</td><td>46.95</td><td>36.21</td><td>44.44</td><td>33.06</td><td>29.12</td><td>27.37</td><td>24.75</td><td>26.34</td><td>22.88</td></tr><tr><td>Δ</td><td>+0.99</td><td>+0.13</td><td>+0.90</td><td>+1.35</td><td>+1.25</td><td>+2.29</td><td>+1.75</td><td>+1.80</td><td>+1.20</td><td>+2.73</td></tr><tr><td></td><td colspan=\"4\">Fr-En</td><td colspan=\"4\">De-En</td><td colspan=\"2\">Cs-En</td></tr><tr><td>Approach</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F18</td></tr><tr><td>UMMT</td><td>40.50*</td><td>-</td><td>-</td><td>-</td><td>26.40*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PVP</td><td>46.00*</td><td>-</td><td>-</td><td>-</td><td>31.60*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>XLM</td><td>48.04</td><td>42.14</td><td>37.03</td><td>43.58</td><td>37.51</td><td>33.21</td><td>30.73</td><td>27.98</td><td>33.13</td><td>30.29</td></tr><tr><td>UNMT-CS</td><td>49.28</td><td>42.86</td><td>37.19</td><td>44.39</td><td>38.83</td><td>34.48</td><td>31.44</td><td>28.07</td><td>33.91</td><td>31.33</td></tr><tr><td>Ours</td><td>49.10</td><td>43.01</td><td>37.41</td><td>44.57</td><td>40.32</td><td>35.20</td><td>32.70</td><td>29.81</td><td>34.29</td><td>31.56</td></tr><tr><td>Δ</td><td>+1.06</td><td>+0.87</td><td>+0.38</td><td>+0.99</td><td>+2.81</td><td>+1.99</td><td>+1.97</td><td>+1.83</td><td>+1.16</td><td>+1.27</td></tr></table>",
+        "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 Moses<sup>7</sup>. 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": "<sup>7</sup>https://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": "<table><tr><td>GT-DE</td><td>Menschen an der Seitenlinie bei einem Fußballspiel.</td></tr><tr><td>GT-EN</td><td>People on the sideline of a soccer match.</td></tr><tr><td>XLM</td><td>People at the plate in a baseball game.</td></tr><tr><td>Ours</td><td>People at the stands in a soccer game.</td></tr><tr><td>GT-DE</td><td>Ein kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.</td></tr><tr><td>GT-EN</td><td>A young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.</td></tr><tr><td>XLM</td><td>A small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.</td></tr><tr><td>Ours</td><td>A small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.</td></tr><tr><td>GT-FR</td><td>Un 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.</td></tr><tr><td>GT-EN</td><td>A 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.</td></tr><tr><td>XLM</td><td>A 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.</td></tr><tr><td>Ours</td><td>A 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.</td></tr></table>",
+        "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": "<table><tr><td>Approach</td><td>Zh-En</td><td>En-Zh</td><td>Cs-En</td><td>En-Cs</td><td>Fr-En</td><td>En-Fr</td></tr><tr><td>XLM</td><td>8.43</td><td>7.91</td><td>25.14</td><td>17.79</td><td>33.73</td><td>35.77</td></tr><tr><td>Our Method</td><td>9.87</td><td>9.36</td><td>25.51</td><td>18.24</td><td>34.14</td><td>36.46</td></tr><tr><td>Δ</td><td>+1.44</td><td>+1.45</td><td>+0.37</td><td>+0.45</td><td>+0.41</td><td>+0.69</td></tr></table>",
+        "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
+    }
+]

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": "<sup>1</sup>Harbin 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": "<sup>1</sup>The 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": "<table><tr><td>Language</td><td>#Word</td><td>#Image</td><td>Average</td></tr><tr><td>En</td><td>4,975</td><td>45,228</td><td>9.09</td></tr><tr><td>Fr</td><td>4,646</td><td>36,927</td><td>7.95</td></tr><tr><td>De</td><td>10,384</td><td>85,419</td><td>8.23</td></tr><tr><td>Cs</td><td>10,706</td><td>72,942</td><td>6.81</td></tr><tr><td>Zh</td><td>10,805</td><td>85,144</td><td>7.88</td></tr><tr><td>Total</td><td>41,516</td><td>325,660</td><td>7.84</td></tr></table>"
+        },
+        {
+            "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 Spacy<sup>3</sup> and Stanford CoreNLP<sup>4</sup>. We choose nouns because they typically correspond to real-world entities and are more likely to yield high-quality images. Next, use Google Images<sup>5</sup> 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 codebase<sup>6</sup>. 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": "<sup>3</sup>https://github.com/explosion/spaCy"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.892,
+                0.392,
+                0.905
+            ],
+            "angle": 0,
+            "content": "<sup>4</sup>https://github.com/stanfordnlp/CoreNLP"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.905,
+                0.306,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>5</sup>https://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": "<table><tr><td></td><td colspan=\"4\">En-Fr</td><td colspan=\"4\">En-De</td><td colspan=\"2\">En-Cs</td></tr><tr><td>Approach</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F18</td></tr><tr><td>UMMT</td><td>39.80*</td><td>-</td><td>-</td><td>-</td><td>23.50*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PVP</td><td>52.30*</td><td>-</td><td>-</td><td>-</td><td>33.90*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>XLM</td><td>53.08</td><td>46.82</td><td>35.31</td><td>43.09</td><td>31.81</td><td>26.83</td><td>25.62</td><td>22.95</td><td>25.14</td><td>20.15</td></tr><tr><td>UNMT-CS</td><td>54.24</td><td>46.30</td><td>36.06</td><td>43.49</td><td>33.43</td><td>27.51</td><td>26.67</td><td>24.27</td><td>26.46</td><td>22.71</td></tr><tr><td>Ours</td><td>54.07</td><td>46.95</td><td>36.21</td><td>44.44</td><td>33.06</td><td>29.12</td><td>27.37</td><td>24.75</td><td>26.34</td><td>22.88</td></tr><tr><td>Δ</td><td>+0.99</td><td>+0.13</td><td>+0.90</td><td>+1.35</td><td>+1.25</td><td>+2.29</td><td>+1.75</td><td>+1.80</td><td>+1.20</td><td>+2.73</td></tr><tr><td></td><td colspan=\"4\">Fr-En</td><td colspan=\"4\">De-En</td><td colspan=\"2\">Cs-En</td></tr><tr><td>Approach</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F17</td><td>F18</td><td>C17</td><td>F16</td><td>F18</td></tr><tr><td>UMMT</td><td>40.50*</td><td>-</td><td>-</td><td>-</td><td>26.40*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>PVP</td><td>46.00*</td><td>-</td><td>-</td><td>-</td><td>31.60*</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>XLM</td><td>48.04</td><td>42.14</td><td>37.03</td><td>43.58</td><td>37.51</td><td>33.21</td><td>30.73</td><td>27.98</td><td>33.13</td><td>30.29</td></tr><tr><td>UNMT-CS</td><td>49.28</td><td>42.86</td><td>37.19</td><td>44.39</td><td>38.83</td><td>34.48</td><td>31.44</td><td>28.07</td><td>33.91</td><td>31.33</td></tr><tr><td>Ours</td><td>49.10</td><td>43.01</td><td>37.41</td><td>44.57</td><td>40.32</td><td>35.20</td><td>32.70</td><td>29.81</td><td>34.29</td><td>31.56</td></tr><tr><td>Δ</td><td>+1.06</td><td>+0.87</td><td>+0.38</td><td>+0.99</td><td>+2.81</td><td>+1.99</td><td>+1.97</td><td>+1.83</td><td>+1.16</td><td>+1.27</td></tr></table>"
+        },
+        {
+            "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 Moses<sup>7</sup>. 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": "<sup>7</sup>https://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": "<table><tr><td>GT-DE</td><td>Menschen an der Seitenlinie bei einem Fußballspiel.</td></tr><tr><td>GT-EN</td><td>People on the sideline of a soccer match.</td></tr><tr><td>XLM</td><td>People at the plate in a baseball game.</td></tr><tr><td>Ours</td><td>People at the stands in a soccer game.</td></tr><tr><td>GT-DE</td><td>Ein kleines Kind in einer orangefarbenen Rettungswestehalt ein Padel und paddelt in einem blauen Kajak auf einem Gewässer.</td></tr><tr><td>GT-EN</td><td>A young child wearing an orange life vest holding an oar paddling a blue kayak in a body of water.</td></tr><tr><td>XLM</td><td>A small child in an orange-red life vest holds a Paddy Power and paddles in a blue cadet on a body of water.</td></tr><tr><td>Ours</td><td>A small child in an orange life vest is holding a paddle and paddling in a blue kayak on a body of water.</td></tr><tr><td>GT-FR</td><td>Un 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.</td></tr><tr><td>GT-EN</td><td>A 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.</td></tr><tr><td>XLM</td><td>A 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.</td></tr><tr><td>Ours</td><td>A 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.</td></tr></table>"
+        },
+        {
+            "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": "<table><tr><td>Approach</td><td>Zh-En</td><td>En-Zh</td><td>Cs-En</td><td>En-Cs</td><td>Fr-En</td><td>En-Fr</td></tr><tr><td>XLM</td><td>8.43</td><td>7.91</td><td>25.14</td><td>17.79</td><td>33.73</td><td>35.77</td></tr><tr><td>Our Method</td><td>9.87</td><td>9.36</td><td>25.51</td><td>18.24</td><td>34.14</td><td>36.46</td></tr><tr><td>Δ</td><td>+1.44</td><td>+1.45</td><td>+0.37</td><td>+0.45</td><td>+0.41</td><td>+0.69</td></tr></table>"
+        },
+        {
+            "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"
+        }
+    ]
+]