Add Batch 7ba69d39-e709-4a74-911f-fdfbbd5025ae data

.gitattributes

@@ -6745,3 +6745,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Byte[[:space:]]Pair[[:space:]]Encoding[[:space:]]for[[:space:]]Symbolic[[:space:]]Music/58bf9db4-7e37-4f11-b09d-7484d505eee2_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/ByteSized32_[[:space:]]A[[:space:]]Corpus[[:space:]]and[[:space:]]Challenge[[:space:]]Task[[:space:]]for[[:space:]]Generating[[:space:]]Task-Specific[[:space:]]World[[:space:]]Models[[:space:]]Expressed[[:space:]]as[[:space:]]Text[[:space:]]Games/a4e7e94c-ef95-400a-8a5d-d1d6e5cef91d_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/C-STS_[[:space:]]Conditional[[:space:]]Semantic[[:space:]]Textual[[:space:]]Similarity/0b267bb1-c5ff-4364-9e31-34b13a7a31a9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CAPSTONE_[[:space:]]Curriculum[[:space:]]Sampling[[:space:]]for[[:space:]]Dense[[:space:]]Retrieval[[:space:]]with[[:space:]]Document[[:space:]]Expansion/d66707f9-dcd1-4cb5-a6c2-30915753860b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CESAR_[[:space:]]Automatic[[:space:]]Induction[[:space:]]of[[:space:]]Compositional[[:space:]]Instructions[[:space:]]for[[:space:]]Multi-turn[[:space:]]Dialogs/54a33beb-057a-4282-9d89-7a5dd37b8152_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CHEF[[:space:]]in[[:space:]]the[[:space:]]Language[[:space:]]Kitchen_[[:space:]]A[[:space:]]Generative[[:space:]]Data[[:space:]]Augmentation[[:space:]]Leveraging[[:space:]]Korean[[:space:]]Morpheme[[:space:]]Ingredients/8a3db38c-4d3c-4329-8082-2813529f2076_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CLAD-ST_[[:space:]]Contrastive[[:space:]]Learning[[:space:]]with[[:space:]]Adversarial[[:space:]]Data[[:space:]]for[[:space:]]Robust[[:space:]]Speech[[:space:]]Translation/3c4dc6ba-99d3-4dfa-ade5-337f84d977db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CLAIR_[[:space:]]Evaluating[[:space:]]Image[[:space:]]Captions[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/ed931ad6-a4d3-4ac0-93d7-f7030245fd98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CLEME_[[:space:]]Debiasing[[:space:]]Multi-reference[[:space:]]Evaluation[[:space:]]for[[:space:]]Grammatical[[:space:]]Error[[:space:]]Correction/ef2f484a-30da-421b-8e1e-02febc3b3494_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CLEVR-Implicit_[[:space:]]A[[:space:]]Diagnostic[[:space:]]Dataset[[:space:]]for[[:space:]]Implicit[[:space:]]Reasoning[[:space:]]in[[:space:]]Referring[[:space:]]Expression[[:space:]]Comprehension/7a840248-0cf8-41af-aee2-5d7039744b4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/COFFEE_[[:space:]]Counterfactual[[:space:]]Fairness[[:space:]]for[[:space:]]Personalized[[:space:]]Text[[:space:]]Generation[[:space:]]in[[:space:]]Explainable[[:space:]]Recommendation/495d6ac8-cce4-4b58-8d12-6baca0a4fa1d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/COHESENTIA_[[:space:]]A[[:space:]]Novel[[:space:]]Benchmark[[:space:]]of[[:space:]]Incremental[[:space:]]versus[[:space:]]Holistic[[:space:]]Assessment[[:space:]]of[[:space:]]Coherence[[:space:]]in[[:space:]]Generated[[:space:]]Texts/c71a49be-8040-4fab-8e36-9e7ed39b9d60_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CORE_[[:space:]]A[[:space:]]Few-Shot[[:space:]]Company[[:space:]]Relation[[:space:]]Classification[[:space:]]Dataset[[:space:]]for[[:space:]]Robust[[:space:]]Domain[[:space:]]Adaptation./1cbf97d6-fb6a-43c5-abad-995e9092f433_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/COVID-19[[:space:]]Vaccine[[:space:]]Misinformation[[:space:]]in[[:space:]]Middle[[:space:]]Income[[:space:]]Countries/98b30198-34a4-46bd-9d4b-062f536aa150_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CP-BCS_[[:space:]]Binary[[:space:]]Code[[:space:]]Summarization[[:space:]]Guided[[:space:]]by[[:space:]]Control[[:space:]]Flow[[:space:]]Graph[[:space:]]and[[:space:]]Pseudo[[:space:]]Code/44299b87-3a5c-4f9a-a3cf-312f7224d327_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CQE_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Quantity[[:space:]]Extractor/6181a88e-acfc-40ef-b30f-9cf1b53ba9cf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CRAB_[[:space:]]Assessing[[:space:]]the[[:space:]]Strength[[:space:]]of[[:space:]]Causal[[:space:]]Relationships[[:space:]]Between[[:space:]]Real-world[[:space:]]Events/a0e72cc1-3b2b-4d97-8b8e-6f7b1e950a96_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CRT-QA_[[:space:]]A[[:space:]]Dataset[[:space:]]of[[:space:]]Complex[[:space:]]Reasoning[[:space:]]Question[[:space:]]Answering[[:space:]]over[[:space:]]Tabular[[:space:]]Data/347d884d-a55c-486b-9ae2-6df172305f13_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CRUSH4SQL_[[:space:]]Collective[[:space:]]Retrieval[[:space:]]Using[[:space:]]Schema[[:space:]]Hallucination[[:space:]]For[[:space:]]Text2SQL/2dd0a715-9b9e-4af5-af4a-7a5bf2329397_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CRaSh_[[:space:]]Clustering,[[:space:]]Removing,[[:space:]]and[[:space:]]Sharing[[:space:]]Enhance[[:space:]]Fine-tuning[[:space:]]without[[:space:]]Full[[:space:]]Large[[:space:]]Language[[:space:]]Model/e14386f4-849f-4812-aa5e-c6198a4ccf70_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CRoW_[[:space:]]Benchmarking[[:space:]]Commonsense[[:space:]]Reasoning[[:space:]]in[[:space:]]Real-World[[:space:]]Tasks/3d52bb9f-0520-4c02-9249-06daf9755136_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CS2W_[[:space:]]A[[:space:]]Chinese[[:space:]]Spoken-to-Written[[:space:]]Style[[:space:]]Conversion[[:space:]]Dataset[[:space:]]with[[:space:]]Multiple[[:space:]]Conversion[[:space:]]Types/c387e0a9-9ee6-4761-973a-5635df455450_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CT-GAT_[[:space:]]Cross-Task[[:space:]]Generative[[:space:]]Adversarial[[:space:]]Attack[[:space:]]based[[:space:]]on[[:space:]]Transferability/938204ab-3680-4afd-b900-774c246d9bea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cabbage[[:space:]]Sweeter[[:space:]]than[[:space:]]Cake_[[:space:]]Analysing[[:space:]]the[[:space:]]Potential[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Learning[[:space:]]Conceptual[[:space:]]Spaces/e227df07-382c-4617-a120-73c877f6b88a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Calc-X[[:space:]]and[[:space:]]Calcformers_[[:space:]]Empowering[[:space:]]Arithmetical[[:space:]]Chain-of-Thought[[:space:]]through[[:space:]]Interaction[[:space:]]with[[:space:]]Symbolic[[:space:]]Systems/6c48cc57-1e61-4519-9062-f87857f49fb4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]LLMs[[:space:]]Facilitate[[:space:]]Interpretation[[:space:]]of[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models_/418b25ff-2a56-4f2c-8938-751a7a5c0fcc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]LMs[[:space:]]Generalize[[:space:]]to[[:space:]]Future[[:space:]]Data_[[:space:]]An[[:space:]]Empirical[[:space:]]Analysis[[:space:]]on[[:space:]]Text[[:space:]]Summarization/8abfa476-8edd-457f-b89d-1b3f7d7ebc64_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Language[[:space:]]Models[[:space:]]Laugh[[:space:]]at[[:space:]]YouTube[[:space:]]Short-form[[:space:]]Videos_/a6451894-c68b-430f-87c6-767c5e37c3bd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Language[[:space:]]Models[[:space:]]Understand[[:space:]]Physical[[:space:]]Concepts_/4a4447ff-e579-4554-b3bd-b9982f3abf2c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Capture[[:space:]]Dissenting[[:space:]]Human[[:space:]]Voices_/30ed6f41-d955-4c57-a438-ddbd35345db7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]Pre-trained[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]Models[[:space:]]Answer[[:space:]]Visual[[:space:]]Information-Seeking[[:space:]]Questions_/2b89895c-0dd3-4c3c-a08a-0fd7aef7adbe_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]We[[:space:]]Edit[[:space:]]Factual[[:space:]]Knowledge[[:space:]]by[[:space:]]In-Context[[:space:]]Learning_/f85804e5-0a03-4123-a25a-e7f8ade04abd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]We[[:space:]]Edit[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models_/0a84dc45-7a65-46e0-98a2-3c815afa61e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]You[[:space:]]Follow[[:space:]]Me_[[:space:]]Testing[[:space:]]Situational[[:space:]]Understanding[[:space:]]for[[:space:]]ChatGPT/90ab7090-e589-4ebb-8df2-3c2e6368f974_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Can[[:space:]]language[[:space:]]models[[:space:]]learn[[:space:]]analogical[[:space:]]reasoning_[[:space:]]Investigating[[:space:]]training[[:space:]]objectives[[:space:]]and[[:space:]]comparisons[[:space:]]to[[:space:]]human[[:space:]]performance/0a695125-e670-4132-b376-9ffea3d70587_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CaseEncoder_[[:space:]]A[[:space:]]Knowledge-enhanced[[:space:]]Pre-trained[[:space:]]Model[[:space:]]for[[:space:]]Legal[[:space:]]Case[[:space:]]Encoding/ff07a9aa-896b-4a7a-a75e-aece91860295_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Causal[[:space:]]Document-Grounded[[:space:]]Dialogue[[:space:]]Pre-training/5f8208f0-5950-4e8b-9d1c-025b2121b7a9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Causal[[:space:]]Reasoning[[:space:]]through[[:space:]]Two[[:space:]]Cognition[[:space:]]Layers[[:space:]]for[[:space:]]Improving[[:space:]]Generalization[[:space:]]in[[:space:]]Visual[[:space:]]Question[[:space:]]Answering/1115d50f-9bcb-460c-9253-af4c231d0385_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Centering[[:space:]]the[[:space:]]Margins_[[:space:]]Outlier-Based[[:space:]]Identification[[:space:]]of[[:space:]]Harmed[[:space:]]Populations[[:space:]]in[[:space:]]Toxicity[[:space:]]Detection/3e62efce-99c8-4636-891a-1fd4111eee34_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Chain-of-Questions[[:space:]]Training[[:space:]]with[[:space:]]Latent[[:space:]]Answers[[:space:]]for[[:space:]]Robust[[:space:]]Multistep[[:space:]]Question[[:space:]]Answering/543a3826-98a8-42f3-b635-53570128a924_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Chain-of-Thought[[:space:]]Tuning_[[:space:]]Masked[[:space:]]Language[[:space:]]Models[[:space:]]can[[:space:]]also[[:space:]]Think[[:space:]]Step[[:space:]]By[[:space:]]Step[[:space:]]in[[:space:]]Natural[[:space:]]Language[[:space:]]Understanding/f39719a5-38d7-412d-aa0c-275385f223dc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Challenges[[:space:]]in[[:space:]]Context-Aware[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/9323dec0-c769-43ab-ba96-8936c62a04c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Character-LLM_[[:space:]]A[[:space:]]Trainable[[:space:]]Agent[[:space:]]for[[:space:]]Role-Playing/69310f58-c867-4070-a4f0-495ebe8d7e24_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Characterizing[[:space:]]Mechanisms[[:space:]]for[[:space:]]Factual[[:space:]]Recall[[:space:]]in[[:space:]]Language[[:space:]]Models/af167b33-7912-41ea-b8fb-c5720fc1f775_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Characterizing[[:space:]]and[[:space:]]Verifying[[:space:]]Scientific[[:space:]]Claims_[[:space:]]Qualitative[[:space:]]Causal[[:space:]]Structure[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need/9a12de47-c5e9-4ba3-8324-d8f09608dbaa_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ChatEdit_[[:space:]]Towards[[:space:]]Multi-turn[[:space:]]Interactive[[:space:]]Facial[[:space:]]Image[[:space:]]Editing[[:space:]]via[[:space:]]Dialogue/8b2d5345-c61f-429f-8f74-448264c711ae_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ChatGPT[[:space:]]to[[:space:]]Replace[[:space:]]Crowdsourcing[[:space:]]of[[:space:]]Paraphrases[[:space:]]for[[:space:]]Intent[[:space:]]Classification_[[:space:]]Higher[[:space:]]Diversity[[:space:]]and[[:space:]]Comparable[[:space:]]Model[[:space:]]Robustness/4f40ae78-1d75-4789-a9c2-d9f8dc223a42_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Chinese[[:space:]]Lexical[[:space:]]Substitution_[[:space:]]Dataset[[:space:]]and[[:space:]]Method/591f8c86-814d-4da5-8fa9-eb5ce6205fa5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CiteBench_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Scientific[[:space:]]Citation[[:space:]]Text[[:space:]]Generation/b110e6cf-1be6-4ef0-af56-91fec678b615_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CleanCoNLL_[[:space:]]A[[:space:]]Nearly[[:space:]]Noise-Free[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition[[:space:]]Dataset/a3107bc4-775b-4757-8f6c-55b9500f038f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ClimateBERT-NetZero_[[:space:]]Detecting[[:space:]]and[[:space:]]Assessing[[:space:]]Net[[:space:]]Zero[[:space:]]and[[:space:]]Reduction[[:space:]]Targets/519c5905-1393-44e9-b1b5-ed3548f9ac74_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Clinical[[:space:]]Contradiction[[:space:]]Detection/f816f086-94a0-4a4b-9182-6743a53d48ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ClusterLLM_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]as[[:space:]]a[[:space:]]Guide[[:space:]]for[[:space:]]Text[[:space:]]Clustering/37122fb8-9420-4c59-bd88-516dc5a8d0a7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Clustering[[:space:]]Pseudo[[:space:]]Language[[:space:]]Family[[:space:]]in[[:space:]]Multilingual[[:space:]]Translation[[:space:]]Models[[:space:]]with[[:space:]]Fisher[[:space:]]Information[[:space:]]Matrix/940b30c2-af69-4a20-a86e-0ef4f8810037_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoAnnotating_[[:space:]]Uncertainty-Guided[[:space:]]Work[[:space:]]Allocation[[:space:]]between[[:space:]]Human[[:space:]]and[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Data[[:space:]]Annotation/7d6a7f79-47cb-4985-ad7f-de9d3a2b2c9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoCo_[[:space:]]Coherence-Enhanced[[:space:]]Machine-Generated[[:space:]]Text[[:space:]]Detection[[:space:]]Under[[:space:]]Low[[:space:]]Resource[[:space:]]With[[:space:]]Contrastive[[:space:]]Learning/336d76c0-8801-43e7-affc-0835ce81177f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoF-CoT_[[:space:]]Enhancing[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Coarse-to-Fine[[:space:]]Chain-of-Thought[[:space:]]Prompting[[:space:]]for[[:space:]]Multi-domain[[:space:]]NLU[[:space:]]Tasks/c481b2fa-03a3-4cec-910d-7463e2665aee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoLT5_[[:space:]]Faster[[:space:]]Long-Range[[:space:]]Transformers[[:space:]]with[[:space:]]Conditional[[:space:]]Computation/42da2671-592c-47e5-b883-7422e483134b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoMPosT_[[:space:]]Characterizing[[:space:]]and[[:space:]]Evaluating[[:space:]]Caricature[[:space:]]in[[:space:]]LLM[[:space:]]Simulations/c2f38d30-01cf-48e7-960e-348eb8cbfc30_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoRec_[[:space:]]An[[:space:]]Easy[[:space:]]Approach[[:space:]]for[[:space:]]Coordination[[:space:]]Recognition/bde8d71c-8bdc-412a-af90-ae0d1b0f10d2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CoSyn_[[:space:]]Detecting[[:space:]]Implicit[[:space:]]Hate[[:space:]]Speech[[:space:]]in[[:space:]]Online[[:space:]]Conversations[[:space:]]Using[[:space:]]a[[:space:]]Context[[:space:]]Synergized[[:space:]]Hyperbolic[[:space:]]Network/abc9f66b-ee17-4832-a682-a4344a30d816_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Coarse-to-Fine[[:space:]]Contrastive[[:space:]]Learning[[:space:]]in[[:space:]]Image-Text-Graph[[:space:]]Space[[:space:]]for[[:space:]]Improved[[:space:]]Vision-Language[[:space:]]Compositionality/2e238f87-67d3-4105-aa9e-9dade8193f77_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Code-Switching[[:space:]]Metrics[[:space:]]Using[[:space:]]Intonation[[:space:]]Units/16e021c9-fc3b-4c90-b4e3-096a36ce23db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CodeBERTScore_[[:space:]]Evaluating[[:space:]]Code[[:space:]]Generation[[:space:]]with[[:space:]]Pretrained[[:space:]]Models[[:space:]]of[[:space:]]Code/00d04abb-ead0-4cae-bd93-7a9b68c8a502_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CodeFusion_[[:space:]]A[[:space:]]Pre-trained[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Code[[:space:]]Generation/90c04282-a3d1-41e6-b01f-95a594903e8e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/CodeT5+_[[:space:]]Open[[:space:]]Code[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Code[[:space:]]Understanding[[:space:]]and[[:space:]]Generation/a09efd13-e521-41b9-ba56-49bc61a8cb36_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cognate[[:space:]]Transformer[[:space:]]for[[:space:]]Automated[[:space:]]Phonological[[:space:]]Reconstruction[[:space:]]and[[:space:]]Cognate[[:space:]]Reflex[[:space:]]Prediction/1503bb9a-8e29-46b0-a137-adc2464016aa_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/CAPSTONE_ Curriculum Sampling for Dense Retrieval with Document Expansion/d66707f9-dcd1-4cb5-a6c2-30915753860b_content_list.json

@@ -0,0 +1,1591 @@
+[
+    {
+        "type": "text",
+        "text": "CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion",
+        "text_level": 1,
+        "bbox": [
+            208,
+            87,
+            789,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xingwei He $^{1*}$ , Yeyun Gong $^{2,\\dagger}$ , A-Long Jin $^{1}$ , Hang Zhang $^{2}$ , Anlei Dong $^{3}$ , Jian Jiao $^{3}$ , Siu Ming Yiu $^{1,\\dagger}$ , Nan Duan $^{2}$",
+        "bbox": [
+            225,
+            141,
+            776,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>The University of Hong Kong, <sup>2</sup>Microsoft Research Asia, <sup>3</sup>Microsoft hexingwei15@gmail.com, ajin@eee.hku.hk, smyiu@cs.hku.hk, {yegong, v-zhang, anlei.dong, jian.jiao, NANDuan}@microsoft.com",
+        "bbox": [
+            176,
+            175,
+            825,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query. In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality query-informed document representations. Experimental results on both in-domain and out-of-domain datasets demonstrate that our approach outperforms previous dense retrieval models.",
+        "bbox": [
+            144,
+            279,
+            460,
+            705
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            715,
+            258,
+            731
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Text retrieval aims to find the relevant documents for a given query from a large collection of documents, playing an indispensable role in open-domain question answering (Chen et al., 2017), fact verification (Thorne et al., 2018) and retrieval-augmented generation (Lewis et al., 2020; He et al., 2022). At the early stage, sparse retrieval methods such as TF-IDF or BM25 dominated passage retrieval by relying mainly on lexical term matching",
+        "bbox": [
+            112,
+            741,
+            489,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "to compute relevance between the query and document. Recently, there has been a surge of research interest in neural network-based dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021). Different from sparse retrieval, dense retrieval resorts to neural encoders to compute the dense representations of the query and document. This enables dense retrieval to infer the relevance between them at the semantic level rather than the surface level, thus circumventing the term mismatch problem suffered by the sparse retrieval models.",
+        "bbox": [
+            507,
+            253,
+            884,
+            429
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, the dual-encoder architecture has been a standard workhorse for dense retrieval. One major disadvantage of this architecture is that it can only partially extract the interactions between the query and document, since it encodes them separately. By comparison, the cross-encoder architecture can effectively capture the deep correlation between them by taking the concatenation of the query and document as input. By directly concatenating the query and document, the cross-encoder gains an advantage in capturing interactions, but also loses the advantage of pre-computing document representations during inference. Therefore, cross-encoder cannot wholly replace dual-encoder.",
+        "bbox": [
+            507,
+            432,
+            884,
+            657
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To enhance the retrieval models' ability to capture interactions between queries and documents while maintaining retrieval efficiency, previous work mainly focuses on generating query-informed document representations. One approach, known as late interaction (Khattab and Zaharia, 2020), involves encoding the query and document independently in the early layers, while the later layers model their interactions. Late interaction combines dual-encoder and cross-encoder, making a trade-off between retrieval efficiency and performance. On the other hand, Li et al. (2022) proposed a promising retrieval architecture, dual-cross-encoder. As shown in Figure 1 (c), this architecture computes the query-related document representation by expanding the document with a real or pseudo query.",
+        "bbox": [
+            507,
+            661,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done during internship at Microsoft Research Asia. †Corresponding authors.",
+        "bbox": [
+            139,
+            892,
+            485,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10531",
+        "bbox": [
+            475,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10531-10541 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            842,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Compared with late interaction, dual-cross-encoder (i.e., dense retrieval with document expansion) gets the query-related document representation without sacrificing the retrieval efficiency at inference. However, there exists a discrepancy between training and inference in the current dual-cross-encoder retriever. Specifically, during training, the document is expanded using a real query, whereas during inference, the document is enriched with a generated query. This discrepancy causes the learned retriever overly focus on the query, yet neglect the document, when computing the document representation. During inference, if the generated query  $q'$  significantly differs from the user-input query  $q$ , the query-related document representation will be misled by  $q'$ , thus degrading the performance. That is why the dual-cross-encoder even underperforms the vanilla dual-encoder. To address this issue, Li et al. (2022) proposed a solution by computing multiview document representations using different generated queries for each document. While multiview document representations improve retrieval performance, they come at the cost of significantly increased retrieval latency, which scales linearly with the number of views.",
+        "bbox": [
+            112,
+            84,
+            492,
+            487
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose CAPSTONE, a curriculum sampling for dense retrieval with document expansion, to bridge the gap between training and inference for dual-cross-encoder. Our motivation is to expect the dual-cross-encoder retrieval model can utilize both the document  $d$  and pseudo query  $q^{\\prime}$  to compute the query-informed document representation. To achieve this, we train the dual-cross-encoder retriever by gradually increasing the relevance of the pseudo query  $q^{\\prime}$  to the gold query  $q$ . Specifically, at the early training stage, a pseudo query  $q^{\\prime}$  irrelevant to  $q$  is selected, causing the retriever to solely rely on the document. As we progress to the late training stage, a highly related pseudo query  $q^{\\prime}$  is chosen, allowing the retriever to learn to augment the document representation with the pseudo query. By doing so, we alleviate the discrepancy between training and inference. During inference, if the user-input query  $q$  is similar to the pseudo query  $q^{\\prime}$ , then  $q^{\\prime}$  will contribute more to making the target document  $d$  be retrieved. Otherwise, the retrieval model will mainly rely on the relevance between  $q$  and  $d$ .",
+        "bbox": [
+            112,
+            494,
+            490,
+            865
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To summarize, the main contributions of this paper are as follows: (1) We propose a curriculum learning approach to bridge the gap between train",
+        "bbox": [
+            112,
+            870,
+            492,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ing and inference for dense retrieval with document expansion, further improving the query-informed document representation<sup>1</sup>. (2) We propose to compute the typical document representation rather than using multiview document representations at inference, which balances the retrieval efficiency and performance. (3) To verify the effectiveness, we apply our proposed approach to two different dense retrieval models, DPR (Karpukhin et al., 2020) and coCondenser, and conduct extensive experiments on three in-domain retrieval datasets, and the zero-shot BEIR benchmark. Experimental results show our proposed approach brings substantial gains over competitive baselines. (4) To the best of our knowledge, this is the first time that document expansion is successfully applied to dense retrieval without incurring extra retrieval latency.",
+        "bbox": [
+            507,
+            84,
+            885,
+            361
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            507,
+            370,
+            650,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce the definition of text retrieval and three architectures for dense retrieval.",
+        "bbox": [
+            507,
+            397,
+            885,
+            430
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Task Description. Text retrieval is meant to find the most relevant  $M$  documents  $D^{+} = \\{d_{1}^{+}, d_{2}^{+}, \\ldots, d_{M}^{+}\\}$  for the given query  $q$  from a large corpus  $D = \\{d_{1}, d_{2}, \\ldots, d_{N}\\}$  with  $N$  documents ( $M \\ll N$ ).",
+        "bbox": [
+            507,
+            456,
+            885,
+            536
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Dual-Encoder Architecture. Dual-encoder (DE) is the typical dense retrieval architecture. As shown in Figure 1 (a), it consists of a query encoder  $E_{q}$  and a document encoder  $E_{d}$ , which encode the query  $q$  and document  $d$  into dense vectors  $E_{q}(q)$  and  $E_{d}(d)$ , respectively. Previous works usually initialize the encoders with BERT and use the [CLS] vector at the last layer as the dense representation. The similarity score between  $q$  and  $d$  is measured with the inner product of their vectors:",
+        "bbox": [
+            507,
+            546,
+            885,
+            722
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d). \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            589,
+            736,
+            882,
+            756
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Cross-Encoder Architecture. Since DE models  $q$  and  $d$  separately, it is not good at capturing the relevance between them. To capture the interactions between them, cross-encoder (CE) directly takes the concatenation of  $q$  and  $d$  as input, as shown in Figure 1 (b). It first computes the dense representation  $E(q + d)$  for them and then uses the fully",
+        "bbox": [
+            507,
+            769,
+            885,
+            882
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "1Our code is available at: https://github.com/microsoft/SimXNS/CAPSTONE.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10532",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/012cff165e4a6b68de3be894d4ddfbd24d766bd2c0923035a2c1221467023e8a.jpg",
+        "image_caption": [
+            "(a) Dual-encoder"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            124,
+            85,
+            344,
+            211
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/f68ed6d10d993773052bbf470053a2fb6732b7da7fbe9c3c3eca6eec1f0184ae.jpg",
+        "image_caption": [
+            "(b) Cross-encoder"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            351,
+            86,
+            563,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/4f3f43b05c0c590da89241fc943a88d725bbed1a25aef9272e6198f7136d442f.jpg",
+        "image_caption": [
+            "(c) Dual-cross-encoder"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            569,
+            86,
+            875,
+            212
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "connected layers (FCL) to compute the similarity score:",
+        "bbox": [
+            112,
+            298,
+            485,
+            330
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s i m} (q, d) = F C L (E (d + q)), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            184,
+            342,
+            485,
+            360
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $^{\\prime} + ^{\\prime}$  is the concatenation operation.",
+        "bbox": [
+            112,
+            370,
+            418,
+            386
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Dual-Cross-Encoder Architecture. Although CE can extract more fine-grained relevance between  $q$  and  $d$ , it is not suitable for retrieval. To combine the advantages of DE and CE, Li et al. (2022) proposed dual-cross-encoder (DCE). As shown in Figure 1 (c), DCE also consists of a query encoder and a document encoder. The only difference between DE and DCE is that the document encoder of DCE takes the document and the generated query  $q'$  as input during inference. Therefore, DCE can be regarded as DE with document expansion. Similar to DE, DCE computes the similarity score between  $q$  and  $d$  with the inner product:",
+        "bbox": [
+            112,
+            394,
+            489,
+            602
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d + q ^ {\\prime}). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            613,
+            487,
+            633
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            642,
+            263,
+            659
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In Section 3.1, we first identify the discrepancy between training and inference of DCE. Then, we introduce how to bridge the gap with curriculum learning in Section 3.2. Finally, we will show our proposed inference method in Section 3.3.",
+        "bbox": [
+            112,
+            668,
+            487,
+            747
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Discrepancy in Training and Inference",
+        "text_level": 1,
+        "bbox": [
+            112,
+            758,
+            463,
+            774
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The training objective of DE is to learn dense representations of queries and documents to maximize the similarity score between the query and positive document. The training loss is defined as follows:",
+        "bbox": [
+            112,
+            778,
+            487,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L (q, d ^ {+}, D ^ {-}) \\\\ = - l o g \\frac {e ^ {s i m (q , d ^ {+})}}{e ^ {s i m (q , d ^ {+})} + \\sum_ {d ^ {-} \\in D ^ {-}} e ^ {s i m (q , d ^ {-})}}, \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            149,
+            852,
+            448,
+            921
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/30085f324ac07f9808fc296835fc77d1212dbb3a0d64a3525ec0e1a691b89ebc.jpg",
+        "image_caption": [
+            "Figure 1: Illustration of the dual-encoder, cross-encoder and dual-cross-encoder architectures.  $q$  and  $q'$  denote the gold query and generated query for the document  $d$ . '+ is the concatenation operation.",
+            "Figure 2: Comparison of different document expansion methods when training DCE, where the left is the vanilla method and the right is our proposed method.  $d$  is the positive document for  $q$ .  $d_{1}, \\ldots, d_{k}$  are the hard negatives.  $d_{k+1}, \\ldots, d_{n}$  are the in-batch negatives.  $q'$  and  $q_{i}'$  are the generated query for  $d$  and  $d_{i}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            294,
+            884,
+            443
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $D^{-}$  is the set of negative documents, containing hard negatives and in-batch negatives.",
+        "bbox": [
+            507,
+            564,
+            882,
+            596
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "During training, DCE expands the positive document  $d$  and hard negatives with the gold query  $q$  (e.g., replacing  $d$  with  $d + q$ ). For ease of understanding, we show how to construct positive, hard negatives and in-batch negatives in the left of Figure 2. As shown in this figure, DCE can filter out all in-batch negatives by solely relying on the query information from the expanded document, since it only needs to observe whether the input query appears in the expanded document. The utilization of document information is necessary only for distinguishing the positive from the hard negatives. Consequently, the learned representation of the expanded document has a strong bias toward the query, while almost neglecting the document.",
+        "bbox": [
+            507,
+            596,
+            884,
+            838
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "At inference, DCE enriches the document with the generated query  $q'$  rather than the user-input query  $q$ . When  $q$  and  $q'$  are different types of queries for the target document  $d$ ,  $q$  or  $E_q(q)$  will be far from  $q'$  or  $E_d(q')$ . As the retrieval model",
+        "bbox": [
+            507,
+            839,
+            882,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10533",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/26d553333d43693bc1b567699099fd6b33e03e09c0626836a25e3ac15d685877.jpg",
+        "image_caption": [
+            "Figure 3: The query selection process based on curriculum learning in each training iteration. We first compute the relevance score  $M(q, q_i')$  between the generated query  $q_i'$  and the gold query  $q$ , and then sort the generated queries in ascending order according to their relevance scores to  $q$ . Next, we divide them into K groups. At the  $i$ -th training iteration, we randomly sample one generated query from the  $i$ -th group, and then use the sampled query to update the dual-cross encoder."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            80,
+            487,
+            244
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "overly relies on the query part when computing the document representation,  $E_{d}(d + q^{\\prime})$  will be close to  $E_{d}(q^{\\prime})$ . Based on these, we can easily infer that  $E_{d}(d + q^{\\prime})$  is far away from  $E_{q}(q)$ , making it more difficult to retrieve the target document  $d$ . As a result, the dense retriever with document expansion even underperforms its counterpart without using document expansion. We attribute the performance degradation to the discrepancy between training and inference.",
+        "bbox": [
+            112,
+            412,
+            487,
+            571
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Bridging the Gap with Curriculum Sampling",
+        "text_level": 1,
+        "bbox": [
+            112,
+            588,
+            435,
+            620
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To bridge the gap between training and inference, we propose to expand the document with its generated query during training. We show the training process of our proposed method in the right of Figure 2. Although this idea is simple and intuitive, it brings a significant advantage: the retriever cannot take a shortcut during training, where the retriever cannot exclude in-batch negatives by simply checking whether the user-input query appears in the expanded document.",
+        "bbox": [
+            112,
+            627,
+            487,
+            788
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "However, this is not the end of the story. If the generated query  $q'$  used to expand  $d$  is irrelevant to the gold query  $q$ , the document retriever of DCE will ignore  $q'$ , which means DCE will degenerate into DE. On the other hand, if  $q'$  and  $q$  are similar, we will encounter the abovementioned problem (DCE will overly depend on  $q'$ ). We expect the dense retriever with document expansion can use",
+        "bbox": [
+            112,
+            790,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/147ac2e7998f584b10a083627fb114a5d33e5ae3739c44fc46fe7da2421b93be.jpg",
+        "image_caption": [
+            "Figure 4: The overview of the retrieval stage.  $d_{i}$  is the  $i - th$  document, and  $N$  refers to the size of the external corpus.  $q_{i}^{j}$  means the  $j$ -th generated query for  $d_{i}$ ."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            510,
+            80,
+            885,
+            190
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "both the document and the generated query during training, so that the learned document representation contains the information of both parts.",
+        "bbox": [
+            507,
+            268,
+            882,
+            315
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To fulfill this goal, we further propose the curriculum sampling. To be concrete, at the early training stage, the selected  $q'$  has low correlations with  $q$ , forcing the retriever to use the document. As the training goes on, we select  $q'$  with gradually increased relevance to  $q$ , encouraging the retriever to use the query. Driven by this motivation, we first generate some queries for each document  $d$ . Then, we compute the relevance score between the generated query  $q_i'$  and the gold query  $q$  with an automatic evaluation metric  $M$  (i.e., ROUGE- $L^2$ ). After that, we sort the generated queries in ascending order according to their relevance scores to  $q$  and divide them into  $K$  groups. We summarize the curriculum sampling in Figure 3.",
+        "bbox": [
+            507,
+            316,
+            884,
+            557
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Inference",
+        "text_level": 1,
+        "bbox": [
+            507,
+            569,
+            630,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Before retrieving, we first generate  $S$  queries for each document with the generator. Next, we concatenate a document with one generated query and compute its latent representation with the document encoder. Then, we will get  $S$  different representations for each document.",
+        "bbox": [
+            507,
+            590,
+            882,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Corpus Expansion. Following Li et al. (2022), we keep  $S$  representations for each document, so the original corpus will be expanded to  $S$  times. When a query comes, we need to retrieve the relevant documents from the expanded corpus, which increases the retrieval latency by  $S$  times.",
+        "bbox": [
+            507,
+            696,
+            882,
+            791
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Computing the Typical Representation of Different Views. To avoid increasing the retrieval latency, we can retain a single representation for each document. The easiest way is to set  $S$  to 1. However, we found that the retrieval performance",
+        "bbox": [
+            507,
+            803,
+            882,
+            883
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2The evaluation code for ROUGE-L is available at: https://huggingface.co/spaces/evaluate-metric/rouge.",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10534",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/4f4525089551f2e02835e9c9c5fe7f3a4a894e0be54334334a3cf5acba68c302.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td><td>#Passage</td></tr><tr><td>MS-MARCO</td><td>502,939</td><td>6,980</td><td>-</td><td>8,841,823</td></tr><tr><td>TREC-19</td><td>-</td><td>-</td><td>43</td><td>8,841,823</td></tr><tr><td>TREC-20</td><td>-</td><td>-</td><td>54</td><td>8,841,823</td></tr></table>",
+        "bbox": [
+            115,
+            82,
+            487,
+            140
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Basic statistics of retrieval datasets. #Passage is the number of unique passages in the retrieval corpus.",
+        "bbox": [
+            112,
+            149,
+            485,
+            180
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "is positively correlated to the number of views,  $S$ . This is because a larger value of  $S$  increases the probability of the user-input query being relevant to one of the generated queries. Therefore, setting  $S$  to 1 will degrade the retrieval performance compared with expanding the corpus.",
+        "bbox": [
+            112,
+            206,
+            487,
+            303
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Given this, we propose to compute the typical representation for a document by averaging all  $S$  different views. This approach allows us to maintain only one representation for each document. As shown in Figure 4, our proposed method does not bring additional retrieval latency $^3$ , yet it performs much better than simply setting  $S$  to 1 (please refer to Section 4.3 for more details).",
+        "bbox": [
+            112,
+            305,
+            487,
+            434
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            451,
+            260,
+            467
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Experimental Setups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            480,
+            324,
+            495
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Datasets. We conduct experiments on three passage retrieval datasets: MS-MARCO passage ranking (Bajaj et al., 2016), TREC Deep Learning (DL) Track 2019 (Craswell et al., 2020) and 2020 (Craswell et al., 2021). MS-MARCO collects real user queries from Bing search and passages from web collection. TREC DL tracks share the same retrieval corpus with MS-MARCO and provide another two annotated test sets. We train our models on the MS-MARCO train set, and then evaluate them on the MS-MARCO development set $^4$ , TREC DL 19 and 20 test sets. Table 1 shows the statistics of these datasets. We evaluate the zero-shot retrieval performance on BEIR benchmark (Thakur et al., 2021), which contains 18 datasets across different domains.",
+        "bbox": [
+            112,
+            504,
+            489,
+            760
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Metrics. Following previous work, we use MRR@10, Recall@50, and Recall@1000 to evaluate the retrieval performance on MS-MARCO, where MRR@10 is the most important metric. We resort to nDCG@10 for TREC DL",
+        "bbox": [
+            112,
+            776,
+            489,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "tracks and BEIR benchmark<sup>5</sup>.",
+        "bbox": [
+            507,
+            83,
+            732,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details. We set the maximum document length to be 144 tokens, and the maximum query length to be 32 tokens. All models are optimized using the AdamW optimizer (Loshchilov and Hutter, 2019) with the learning rate of  $5 \\times 10^{-6}$ , and a linear learning rate schedule with  $10\\%$  warm-up steps for around three epochs. Every training batch contains 64 queries, and each query is accompanied by one positive and 31 hard negative documents. Following Gao and Callan (2022), we train our proposed model for two stages and initialize the retriever with coCondenser at each stage. At the first training stage, the hard negatives are sampled from the official BM25 hard negatives, but at the second training stage, the hard negatives are sampled from the mined hard negatives. For each training stage, we evaluate the last model training checkpoint on the retrieval datasets.",
+        "bbox": [
+            507,
+            111,
+            884,
+            400
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The query generator is built upon the seq2seq architecture (Sutskever et al., 2014; He and Yiu, 2022), designed to take a passage as its input and has been specifically trained to create synthetic queries that correspond to the provided passage. Nogueira and Lin (2019) initialized the query generator with the T5-base model. They trained it on the MS-MARCO dataset and generated 80 queries for each passage within MS-MARCO. Following Li et al. (2022), we use the queries released by Nogueira and Lin (2019) to expand the passages in the MS-MARCO dataset. In parallel, for BEIR, we use the query generator fine-tuned on MS-MARCO to generate 5 queries for each document within the BEIR benchmark. During the query generation process, we employ the top- $k$  sampling approach with  $k$  set to 10, and we limit the maximum length of the synthetic queries to 64 tokens.",
+        "bbox": [
+            507,
+            401,
+            884,
+            690
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During training, we implement our proposed curriculum sampling with all 80 queries and divide them into  $K$  groups, but at inference, we resort to the first 10 (i.e.,  $S = 10$ ) and first 5 (i.e.,  $S = 5$ ) queries to compute the typical document representation for MS-MARCO and BEIR benchmark, respectively. At the first training stage, we set  $K$  to 3, and at the second training stage, we set  $K$  to 4.",
+        "bbox": [
+            507,
+            692,
+            884,
+            820
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "5The official evaluation toolkit is available at https:// github.com/beir-cellar/beir.",
+        "bbox": [
+            507,
+            831,
+            882,
+            854
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{6}$ https://git.uwaterloo.ca/jimmylin/doc2query-data/raw/master/T5-passage/predicted Queries_topk_sampleing.zip",
+        "bbox": [
+            509,
+            856,
+            870,
+            892
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>7</sup>https://huggingface.co/castorini/doc2query-t5-base-msmarco",
+        "bbox": [
+            510,
+            892,
+            791,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>We can pre-generate queries and pre-compute document representations. Therefore, the retrieval latency mentioned in this paper does not encompass these pre-processing steps.",
+        "bbox": [
+            112,
+            866,
+            487,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "4The MS-MARCO test set is not publicly available.",
+        "bbox": [
+            137,
+            904,
+            450,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10535",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/fbb9e76c0a6e74d1ab47fc2f3c6f2ba3c292a0683dab218ad41dcfb627b0c963.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"3\">MS-MARCO</td><td>TREC DL 19</td><td>TREC DL 20</td></tr><tr><td>MRR@10</td><td>R@50</td><td>R@1000</td><td>nDCG@10</td><td>nDCG@10</td></tr><tr><td>Sparse retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>BM25 (Yang et al., 2017)</td><td>18.5</td><td>58.5</td><td>85.7</td><td>51.2</td><td>47.7</td></tr><tr><td>DeepCT (Dai and Callan, 2019)</td><td>24.3</td><td>69.0</td><td>91.0</td><td>57.2</td><td>-</td></tr><tr><td>DocT5Query (Nogueira and Lin, 2019)</td><td>27.7</td><td>75.6</td><td>94.7</td><td>64.2</td><td>-</td></tr><tr><td>Dense retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>DPR (Karpukhin et al., 2020)</td><td>31.4</td><td>-</td><td>95.3</td><td>59.0</td><td>62.1</td></tr><tr><td>ANCE (Xiong et al., 2021)</td><td>33.0</td><td>-</td><td>95.9</td><td>64.5</td><td>64.6</td></tr><tr><td>SEED (Lu et al., 2021)</td><td>33.9</td><td>-</td><td>96.1</td><td>-</td><td>-</td></tr><tr><td>STAR (Zhan et al., 2021)</td><td>34.7</td><td>-</td><td>-</td><td>68.3</td><td>-</td></tr><tr><td>TAS-B (Hofstätter et al., 2021)</td><td>34.0</td><td>-</td><td>97.5</td><td>71.2</td><td>69.3</td></tr><tr><td>RocketQA (Qu et al., 2021)</td><td>37.0</td><td>85.5</td><td>97.9</td><td>-</td><td>-</td></tr><tr><td>COIL (Gao et al., 2021)</td><td>35.5</td><td>-</td><td>96.3</td><td>70.4</td><td>-</td></tr><tr><td>ColBERT (Khattab and Zaharia, 2020)</td><td>36.0</td><td>82.9</td><td>96.8</td><td>-</td><td>-</td></tr><tr><td>DCE (Li et al., 2022)</td><td>36.0</td><td>-</td><td>96.4</td><td>68.3</td><td>68.9</td></tr><tr><td>RetroMAE (Xiao et al., 2022)</td><td>35.0</td><td>-</td><td>97.6</td><td>-</td><td>-</td></tr><tr><td>Condenser (Gao and Callan, 2021)</td><td>36.6</td><td>-</td><td>97.4</td><td>69.8</td><td>-</td></tr><tr><td>coCondenser (Gao and Callan, 2022)*</td><td>37.9</td><td>86.3</td><td>98.4</td><td>70.7</td><td>69.8</td></tr><tr><td>CAPSTONE</td><td>38.6</td><td>86.6</td><td>98.6</td><td>71.1</td><td>70.3</td></tr></table>",
+        "bbox": [
+            141,
+            80,
+            857,
+            331
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Passage retrieval results on MS-Marco Dev, and TREC datasets. Results with * are from our reproduction.",
+        "bbox": [
+            112,
+            340,
+            880,
+            356
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "During inference, we pre-compute representations for all documents in the retrieval corpus, and build IndexFlatIP indexes for document representations with the FAISS library (Johnson et al., 2019) to accelerate retrieval.",
+        "bbox": [
+            112,
+            380,
+            487,
+            460
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We implement all models with the HuggingFace Transformers library (Wolf et al., 2019) and conduct all experiments on 8 NVIDIA Tesla V100 GPUs with 32 GB memory. We also use the gradient checkpointing technology (Chen et al., 2016) to reduce the memory of deep neural models, and mixed precision training $^{8}$  to speed up training.",
+        "bbox": [
+            112,
+            461,
+            489,
+            575
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            587,
+            331,
+            602
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In-domain Performance. Table 2 shows the performance of our proposed model and baselines on MS-MARCO, TREC-2019 and TREC-2020. Since we initialize our model, CAPSTONE, with coCondenser, coCondenser is the main comparison model. From Table 2, we can see our proposed model improves coCondenser by a large margin, increasing by 0.7 points in MRR@10 on MS-MARCO, 0.4 points on TREC DL 19 and 0.5 points on TREC DL 20 datasets. These improvements verify the effectiveness of our proposed curriculum learning and typical representation strategy.",
+        "bbox": [
+            112,
+            609,
+            489,
+            803
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zero-shot Performance. To test the out-of-domain generalization capabilities, we first use the T5-base model (Nogueira and Lin, 2019) trained on MS-MARCO to generate 5 queries for each document of BEIR benchmark, and then evaluate our",
+        "bbox": [
+            112,
+            814,
+            489,
+            894
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "model, CAPSTONE, fine-tuned with MS-MARCO on the BEIR benchmark. During evaluation, we compute the typical representation for each document with  $S = 5$ . As shown in Table 3, CAPSTONE performs much better than all baselines in the average performance. Compared with its counterpart without document expansion (i.e., coCondenser), our model achieves improvements on 11 out of 18 datasets and increases nDCG@10 by 0.6 points in the average performance. These improvements demonstrate that our approach generalizes well on diverse domains under zero-shot settings.",
+        "bbox": [
+            507,
+            380,
+            884,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Ablation Study and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            583,
+            781,
+            599
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To analyze our method more directly, experiments in this section are based on the vanilla DPR initialized with the ERNIE-2.0-base model (Sun et al., 2020) without using dense retrieval pre-training strategies, if there is no particular statement.",
+        "bbox": [
+            507,
+            604,
+            882,
+            684
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Corpus Expansion vs. Document Expansion. In this paper, we append only one query to a document both at training and inference. As stated in 3.3, corpus expansion will enlarge the original corpus to  $S$  times during inference. In contrast, document expansion for sparse retrieval (Nogueira and Lin, 2019) appends multiple generated queries to a document at inference, thus without bringing extra retrieval latency. Since document expansion has shown effectiveness in sparse retrieval, it is very intuitive to consider whether it is effective for dense retrieval. To answer this question, we",
+        "bbox": [
+            507,
+            692,
+            884,
+            885
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "In this section, document expansion means expanding a document with multiple queries.",
+        "bbox": [
+            507,
+            891,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{8}$ https://github.com/NVIDIA/apex",
+        "bbox": [
+            134,
+            903,
+            347,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10536",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/d693fc8a55519183081a5d4acd062bf829de0cb81a6b6c852f25630d6eb9e22f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>BERT</td><td>LaPraDoR</td><td>SimCSE</td><td>DiffCSE</td><td>SEED</td><td>Condenser</td><td>coCondenser*</td><td>CAPSTONE</td></tr><tr><td>TREC-COVID</td><td>64.9</td><td>49.5</td><td>52.4</td><td>49.2</td><td>61.2</td><td>75.4</td><td>74.0</td><td>77.9</td></tr><tr><td>BioASQ</td><td>26.2</td><td>23.9</td><td>26.4</td><td>25.8</td><td>29.7</td><td>31.7</td><td>34.1</td><td>34.3</td></tr><tr><td>NFCorpus</td><td>25.7</td><td>28.3</td><td>25.0</td><td>25.9</td><td>25.6</td><td>27.8</td><td>32.4</td><td>33.0</td></tr><tr><td>NQ</td><td>43.8</td><td>41.5</td><td>41.2</td><td>41.2</td><td>42.5</td><td>45.9</td><td>50.5</td><td>50.5</td></tr><tr><td>HotpotQA</td><td>47.8</td><td>48.8</td><td>50.2</td><td>49.9</td><td>52.8</td><td>53.7</td><td>56.4</td><td>56.7</td></tr><tr><td>FiQA-2018</td><td>23.7</td><td>26.6</td><td>24.0</td><td>22.9</td><td>24.4</td><td>26.1</td><td>30.0</td><td>30.4</td></tr><tr><td>Signal-1M (RT)</td><td>21.6</td><td>24.5</td><td>26.4</td><td>26.0</td><td>24.6</td><td>25.8</td><td>24.7</td><td>23.1</td></tr><tr><td>TREC-NEWS</td><td>36.2</td><td>20.6</td><td>36.8</td><td>36.3</td><td>33.5</td><td>35.3</td><td>39.1</td><td>40.3</td></tr><tr><td>Robust04</td><td>36.4</td><td>31.0</td><td>35.3</td><td>34.3</td><td>34.8</td><td>35.2</td><td>40.3</td><td>40.7</td></tr><tr><td>ArguAna</td><td>35.7</td><td>50.3</td><td>43.6</td><td>46.8</td><td>34.7</td><td>37.5</td><td>40.9</td><td>39.2</td></tr><tr><td>Touche-2020</td><td>27.0</td><td>17.8</td><td>17.8</td><td>16.8</td><td>18.0</td><td>22.3</td><td>27.0</td><td>31.0</td></tr><tr><td>CQADupStack</td><td>28.4</td><td>32.6</td><td>29.5</td><td>30.5</td><td>28.5</td><td>31.6</td><td>30.0</td><td>30.0</td></tr><tr><td>Quora</td><td>78.2</td><td>84.3</td><td>84.8</td><td>85.0</td><td>84.9</td><td>85.5</td><td>84.3</td><td>83.8</td></tr><tr><td>DBPedia</td><td>29.8</td><td>32.8</td><td>30.4</td><td>30.3</td><td>32.4</td><td>33.1</td><td>37.2</td><td>38.0</td></tr><tr><td>SCIDOCS</td><td>11.5</td><td>14.5</td><td>12.5</td><td>12.5</td><td>11.7</td><td>13.6</td><td>14.3</td><td>14.3</td></tr><tr><td>FEVER</td><td>68.4</td><td>51.8</td><td>65.1</td><td>64.1</td><td>65.3</td><td>68.2</td><td>72.4</td><td>72.7</td></tr><tr><td>Climate-FEVER</td><td>20.5</td><td>17.2</td><td>22.2</td><td>20.0</td><td>17.6</td><td>19.9</td><td>19.4</td><td>19.3</td></tr><tr><td>SciFact</td><td>50.4</td><td>48.3</td><td>54.5</td><td>52.3</td><td>55.6</td><td>57.0</td><td>58.3</td><td>60.5</td></tr><tr><td>Avg. Performance</td><td>37.6</td><td>35.8</td><td>37.7</td><td>37.2</td><td>37.7</td><td>40.3</td><td>42.7</td><td>43.3</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            878,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Zero-shot dense retrieval performances on BEIR benchmark (measured with nDCG@10). Results with * are from our reproduction.",
+        "bbox": [
+            112,
+            344,
+            882,
+            373
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/882d0aea860342f7dd897bcdd58dc088de610c50976e45de633b778d5854bfc1.jpg",
+        "image_caption": [
+            "Figure 5: Results of DPR with corpus, document, and asymmetric expansion on the dev set of MS-Marco."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            394,
+            487,
+            536
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "conduct experiments on DPR with three different settings: (1) corpus expansion; (2) document expansion; (3) asymmetric expansion, which appends one query to a document at training, but appends  $S$  queries to a document at inference. As shown in Figure 5, unlike corpus expansion, using more queries does not bring clear improvements for document and asymmetric expansion, but sometimes damages their performances. To explain this, we illustrate document expansion and corpus expansion in Figure 6. Document expansion models the target document and all queries together, where the irrelevant queries acting like noise will corrupt the document representation. By comparison, corpus expansion nicely bypasses this problem by modeling different queries independently.",
+        "bbox": [
+            112,
+            600,
+            489,
+            858
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Effect of Query Selection Strategies. To demonstrate the advantage of our proposed curriculum sampling, we expand the positive and negative doc",
+        "bbox": [
+            112,
+            870,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/5b11124203b9a5fa5f683d449d7256f0610e3e3ad7b9f7d3fb0a3a2410ac575f.jpg",
+        "image_caption": [
+            "Figure 6: Illustration of document expansion and corpus expansion at inference. For simplicity, only the positive document  $d_{i}$  of the query  $q$  is shown."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            394,
+            882,
+            554
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "uments with different query selection strategies: (1) gold, meaning using the gold query (i.e., strategy used by the vanilla DCE (Li et al., 2022)); (2) random, meaning randomly sampling a query from the generated queries; (3) top- $k$  and (4) bottom- $k$  meaning sampling a query from the top- $k$  and bottom- $k$  sorted queries, respectively.",
+        "bbox": [
+            507,
+            640,
+            882,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From Figure 7, we observe that: (1) DPR trained with the gold query performs worst among all strategies, even worse than the vanilla DPR (DPR w/o query) when  $S < 8$ . As we have stated above, expanding the document with the gold query during training will make the retriever pay more attention to the query while ignoring the document. At inference, when the provided queries are limited (i.e.,  $S$  is small), it is unlikely to find a query similar to the user-input query. However, as  $S$  increases, the",
+        "bbox": [
+            507,
+            758,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10537",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/2ed9c2c8b6f790185e2947d69e1a1d6d81c37a46d58e97bac77d333f4736637b.jpg",
+        "image_caption": [
+            "Figure 7: Comparison of different query selection strategies. During training, we first expand the document with a query selected from the generated queries using a specific strategy, and then train DPR with the expanded document. During inference, we evaluate the well-trained DPR retrievers with corpus expansion (solid lines) and the typical representation (dotted lines) on the dev set of MS-Marco, where the solid and dotted lines with the same color are results for the same DPR."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            487,
+            278
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "probability of finding a query similar to the user-input query also increases. (2) DPR trained with the top-1 query performs much better than DPR trained with the gold query, since it slightly shifts its attention from the query to the document. However, it still cannot exceed DPR w/o query when  $S = 1$ , indicating that it fails to fully utilize the document like DPR. (3) If the document is expanded with a weakly related (bottom-1) or random query at training, DPR with both settings will outperform DPR w/o query when  $S = 1$ . These two strategies enable DPR to mine query information without sacrificing document information. However, their performances are not strongly positively correlated with  $S$  like the top-1 strategies, indicating it does not fully use the query information. (4) DPR with the curriculum strategy outperforms DPR w/o query when  $S = 1$ , and its performance linearly increases with  $S$ , verifying this strategy enables DPR to utilize both the document and query to model the query-informed document representation.",
+        "bbox": [
+            110,
+            441,
+            489,
+            781
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Effect of the Typical Representation. Although corpus expansion makes the retrieval performance of our proposed curriculum sampling improve with the increase of  $S$  (see the solid red line in Figure 7), the retrieval latency also increases linearly. To alleviate this, we propose to use the average of different document views as the typical representation. Among all strategies (see the dotted lines in the",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8c1c131d46514858221bcd604a2735b389dfef6d02b743c91136d5e94625d42e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Variants</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td></tr><tr><td>CAPSTONE+DPR w/ S = 1</td><td>35.15</td><td>97.19</td></tr><tr><td>CAPSTONE+DPR w/ average</td><td>35.66</td><td>97.28</td></tr><tr><td>CAPSTONE+DPR w/ max</td><td>35.45</td><td>97.25</td></tr><tr><td>CAPSTONE+DPR w/ median</td><td>35.64</td><td>97.20</td></tr></table>",
+        "bbox": [
+            510,
+            82,
+            882,
+            149
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/54def682231ed546680a63aea158c3ce80e0930c1644d4a1b4575d56caade3a6.jpg",
+        "table_caption": [
+            "Table 4: Results of CAPSTONE initialized with DPR using different pooling methods to compute the typical representation on the dev set of MS-Marco."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">BM25 Negatives</td><td colspan=\"2\">Mined Negatives</td></tr><tr><td>MRR@10</td><td>R@1000</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td><td>36.44</td><td>97.65</td></tr><tr><td>CAPSTONE+DPR</td><td>35.66</td><td>97.28</td><td>37.28</td><td>97.82</td></tr><tr><td>coCondenser</td><td>35.91</td><td>98.21</td><td>37.94</td><td>98.41</td></tr><tr><td>CAPSTONE+coCondenser</td><td>36.75</td><td>98.21</td><td>38.65</td><td>98.60</td></tr></table>",
+        "bbox": [
+            510,
+            218,
+            884,
+            286
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Performance of CAPSTONE initialized with DPR and coCondenser at the two training stages on the dev set of MS-Marco.",
+        "bbox": [
+            507,
+            294,
+            882,
+            337
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 7), our proposed curriculum sampling performs best, which again verifies the effectiveness of curriculum sampling. In addition, the retrieval performance of typical representations outperforms its counterpart using corpus expansion with  $S = 1$  for all query selection strategies. More importantly, for the curriculum learning strategy, the typical representation enhances the vanilla DPR by 1.2 on MRR10, and is even on par with its counterpart using corpus expansion with  $S = 3$ , but there still exists a clear performance gap between the typical representation and corpus expansion with  $S = 10$ . Therefore, the typical representation balances the retrieval efficiency and performance.",
+        "bbox": [
+            505,
+            365,
+            884,
+            592
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Comparison of Methods for Computing the Typical Representation. We consider three different pooling methods to compute the typical document representation: taking the average/max/median pooling of different document views. From Table 4, we observe that (1) DPR with  $S = 1$  improves DPR by 0.9 on MRR@10; (2) DPR with the typical representation brings an extra 0.5 improvements on MRR@10; (3) There is no significant difference in the results of the three pooling methods. Given the simplicity, we use the average pooling to compute the typical representation.",
+        "bbox": [
+            507,
+            602,
+            882,
+            795
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-stage Retrieval Performance. To further verify the effectiveness of our proposed approach, we apply our proposed approach to DPR and coCondenser. From Table 5, we observe that our approach brings clear improvements to DPR with 1.4 and 0.8 increases on MRR@10 at the two training stages, respectively. In addition, we witness similar",
+        "bbox": [
+            507,
+            806,
+            884,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10538",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "improvements on coCondenser, proving that our approach does not depend on retrieval models and is applicable to different retrieval models.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            145,
+            270,
+            160
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.1 Dense Retrieval",
+        "text_level": 1,
+        "bbox": [
+            114,
+            171,
+            284,
+            186
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In recent years, with the development of large-scale pre-trained language models, such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), we witness the research interest in information retrieval shifting from the traditional sparse retrieval to neural dense retrieval. Karpukhin et al. (2020); Xiong et al. (2021) proposed dense passage retriever (DPR), which uses two neural encoders initialized with BERT to model the query and document, independently. The subsequent works follow this dual-encoder framework. One line of work improves dense retrieval by mining hard negatives, where they select the top-ranked documents retrieved by the recent retriever as hard negatives and then re-train the retriever with the newly mined hard negatives (Xiong et al., 2021; Qu et al., 2021). However, the mined hard negatives are highly likely to contain false negatives, harming the performance. To mitigate this, following studies denoise the mined hard negatives with re-rankers (Qu et al., 2021; Ren et al., 2021; Zhang et al., 2022). Another kind of work focuses on pre-training to make the pre-trained models more suitable for dense retrieval, such as Condenser (Gao and Callan, 2021), coCondenser (Gao and Callan, 2022) and SIMLM (Wang et al., 2022).",
+        "bbox": [
+            115,
+            193,
+            489,
+            611
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Dense retrieval models typically depend on extensive supervised data, comprising pairs of queries and positive documents. To address the challenge of limited training data, Ma et al. (2021); Sun et al. (2021) proposed to train a query generator on high-resource information retrieval data, and then used the query generator to generate synthetic queries for low-source target domains. Additionally, Dai et al. (2023) harnessed large language models in zero/few settings to produce synthetic queries for documents in target domains, eliminating the need for training a general query generator. In contrast to these approaches, our work leverages synthetic queries for document expansion.",
+        "bbox": [
+            112,
+            612,
+            489,
+            838
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.2 Query Expansion",
+        "text_level": 1,
+        "bbox": [
+            114,
+            848,
+            297,
+            865
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Query expansion enriches the query with various heuristically discovered relevant contexts. For example, GAR (Mao et al., 2021) expands the query",
+        "bbox": [
+            112,
+            871,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "by adding relevant contexts, including the answer, the sentence containing the answer, and the title of a passage where the answer belongs. However, query expansion will increase the retrieval latency during inference, since the generation model is required to generate these relevant contexts for query expansion. Therefore, we mainly focus on document expansion in this work.",
+        "bbox": [
+            507,
+            84,
+            885,
+            212
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.3 Document Expansion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            228,
+            724,
+            243
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Document expansion augments the document with generated queries, which the document might answer. Compared with query expansion, document expansion can be conducted prior to indexing without incurring extra retrieval latency. Document expansion has shown its effectiveness on sparse retrieval models (Nogueira et al., 2019; Nogueira and Lin, 2019), yet how to apply it to dense retrieval models is still under-explored. Li et al. (2022) first applied document expansion to dense retrieval models. However, their models suffer from the discrepancy between training and inference, thus even degrading the vanilla dense retrieval models under the single-view document representation settings. To mitigate this discrepancy, we propose a curriculum learning approach, which successfully applies document expansion to dense retrieval models.",
+        "bbox": [
+            507,
+            250,
+            885,
+            525
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            541,
+            640,
+            556
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work proposes the curriculum sampling for dense retrieval with document expansion, which enables dense retrieval models to learn much better query-related document representations. In addition, we propose to compute the typical representation of different document views, which balances inference efficiency and effectiveness. Our experimental results on the in- and out-of-domain datasets verify the effectiveness of the curriculum sampling and typical representation.",
+        "bbox": [
+            507,
+            569,
+            885,
+            731
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            507,
+            746,
+            645,
+            760
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "There are two possible limitations of this work. The first limitation is that we need to generate synthetic queries for each document in the retrieval corpus, which is very time-consuming. Luckily, this process does not bring extra delay to retrieval. In addition, limited by sufficient computational resources, we only verify the effectiveness of our method on vanilla DPR and coCondenser. In the future, we plan to apply our approach to other dense retrieval",
+        "bbox": [
+            507,
+            774,
+            885,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10539",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "models and verify the effectiveness of our method on these models.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            128,
+            285,
+            143
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This project is supported by HKU-SCF FinTech Academy, Shenzhen-Hong Kong-Macao Science and Technology Plan Project (Category C Project: SGDX20210823103537030), and Theme-based Research Scheme of RGC, Hong Kong (T35-710/20-R). We would like to thank the anonymous reviewers for their constructive and informative feedback on this work.",
+        "bbox": [
+            112,
+            152,
+            489,
+            282
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            307,
+            213,
+            322
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268.",
+            "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer open-domain questions. In Proceedings of ACL, pages 1870-1879, Vancouver, Canada. Association for Computational Linguistics.",
+            "Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. 2016. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174.",
+            "Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the trec 2020 deep learning track. arXiv preprint arXiv:2102.07662.",
+            "Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the trec 2019 deep learning track. arXiv preprint arXiv:2003.07820.",
+            "Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. arXiv preprint arXiv:1910.10687.",
+            "Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In ICLR.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.",
+            "Luyu Gao and Jamie Callan. 2021. Condenser: a pretraining architecture for dense retrieval. In Proceedings of EMNLP, pages 981-993, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            330,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Luyu Gao and Jamie Callan. 2022. Unsupervised corpus aware language model pre-training for dense passage retrieval. In Proceedings of ACL, pages 2843-2853, Dublin, Ireland. Association for Computational Linguistics.",
+            "Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. COIL: Revisit exact lexical match in information retrieval with contextualized inverted list. In Proceedings of NAACL-HLT, pages 3030-3042, Online. Association for Computational Linguistics.",
+            "Xingwei He, Yeyun Gong, A-Long Jin, Weizhen Qi, Hang Zhang, Jian Jiao, Bartuer Zhou, Biao Cheng, Sm Yiu, and Nan Duan. 2022. Metric-guided distillation: Distilling knowledge from the metric to ranker and retriever for generative commonsense reasoning. In Proceedings of EMNLP, pages 839-852, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Xingwei He and Siu Ming Yiu. 2022. Controllable dictionary example generation: Generating example sentences for specific targeted audiences. In Proceedings of ACL, pages 610-627, Dublin, Ireland. Association for Computational Linguistics.",
+            "Sebastian Hofstätter, Sheng-Chieh Lin, Zheng-Hong Yang, Jimmy Lin, and Allan Hanbury. 2021. Efficiently teaching an effective dense retriever with balanced topic aware sampling. In Proceedings of SIGIR, page 113-122, New York, NY, USA. Association for Computing Machinery.",
+            "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547.",
+            "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 EMNLP, pages 6769-6781, Online. Association for Computational Linguistics.",
+            "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of SIGIR, page 39-48, New York, NY, USA. Association for Computing Machinery.",
+            "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of NIPS, Red Hook, NY, USA. Curran Associates Inc.",
+            "Zehan Li, Nan Yang, Liang Wang, and Furu Wei. 2022. Learning diverse document representations with deep query interactions for dense retrieval. arXiv preprint arXiv:2208.04232."
+        ],
+        "bbox": [
+            510,
+            85,
+            885,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10540",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "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.",
+            "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR.",
+            "Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is more: Pretrain a strong Siamese encoder for dense text retrieval using a weak decoder. In Proceedings of EMNLP, pages 2780-2791, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot neural passage retrieval via domain-targeted synthetic question generation. In Proceedings of EACL, pages 1075-1088, Online. Association for Computational Linguistics.",
+            "Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for opendomain question answering. In Proceedings of ACL, pages 4089-4100, Online. Association for Computational Linguistics.",
+            "Rodrigo Nogueira and Jimmy Lin. 2019. From doc2query to docctttttquery.",
+            "Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction. arXiv preprint arXiv:1904.08375.",
+            "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 NAACL-HLT, pages 5835-5847, Online. Association for Computational Linguistics.",
+            "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of EMNLP, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Si Sun, Yingzhuo Qian, Zhenghao Liu, Chenyan Xiong, Kaitao Zhang, Jie Bao, Zhiyuan Liu, and Paul Bennett. 2021. Few-shot text ranking with meta adapted synthetic weak supervision. In Proceedings of ACL, pages 5030-5043, Online. Association for Computational Linguistics.",
+            "Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Ernie 2.0: A continual pre-training framework for language understanding. In Proceedings of AAAI, volume 34, pages 8968-8975."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In NIPS, volume 27. Curran Associates, Inc.",
+            "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).",
+            "James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopoulos, and Arpit Mittal. 2018. The fact extraction and VERIFICATION (FEVER) shared task. In Proceedings of the First Workshop on Fact Extraction and VERIFICATION (FEVER), pages 1-9, Brussels, Belgium. Association for Computational Linguistics.",
+            "Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. SimLM: Pre-training with representation bottleneck for dense passage retrieval. arXiv preprint arXiv:2207.02578.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, R'emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.",
+            "Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. RetroMAE: Pre-training retrieval-oriented language models via masked auto-encoder. In Proceedings of EMNLP, pages 538-548, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.",
+            "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR.",
+            "Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of SIGIR, page 1253-1256, New York, NY, USA. Association for Computing Machinery.",
+            "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of SIGIR, page 1503-1512, New York, NY, USA. Association for Computing Machinery.",
+            "Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022. Adversarial retriever-ranker for dense text retrieval. In ICLR."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            845
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10541",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/CAPSTONE_ Curriculum Sampling for Dense Retrieval with Document Expansion/d66707f9-dcd1-4cb5-a6c2-30915753860b_model.json

@@ -0,0 +1,2070 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.209,
+                0.089,
+                0.791,
+                0.13
+            ],
+            "angle": 0,
+            "content": "CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.226,
+                0.142,
+                0.777,
+                0.176
+            ],
+            "angle": 0,
+            "content": "Xingwei He\\(^{1*}\\), Yeyun Gong\\(^{2,\\dagger}\\), A-Long Jin\\(^{1}\\), Hang Zhang\\(^{2}\\), Anlei Dong\\(^{3}\\), Jian Jiao\\(^{3}\\), Siu Ming Yiu\\(^{1,\\dagger}\\), Nan Duan\\(^{2}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.178,
+                0.177,
+                0.826,
+                0.227
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>The University of Hong Kong, <sup>2</sup>Microsoft Research Asia, <sup>3</sup>Microsoft hexingwei15@gmail.com, ajin@eee.hku.hk, smyiu@cs.hku.hk, {yegong, v-zhang, anlei.dong, jian.jiao, NANDuan}@microsoft.com"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.145,
+                0.28,
+                0.461,
+                0.706
+            ],
+            "angle": 0,
+            "content": "The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query. In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality query-informed document representations. Experimental results on both in-domain and out-of-domain datasets demonstrate that our approach outperforms previous dense retrieval models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.717,
+                0.26,
+                0.732
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.887
+            ],
+            "angle": 0,
+            "content": "Text retrieval aims to find the relevant documents for a given query from a large collection of documents, playing an indispensable role in open-domain question answering (Chen et al., 2017), fact verification (Thorne et al., 2018) and retrieval-augmented generation (Lewis et al., 2020; He et al., 2022). At the early stage, sparse retrieval methods such as TF-IDF or BM25 dominated passage retrieval by relying mainly on lexical term matching"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.254,
+                0.885,
+                0.43
+            ],
+            "angle": 0,
+            "content": "to compute relevance between the query and document. Recently, there has been a surge of research interest in neural network-based dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021). Different from sparse retrieval, dense retrieval resorts to neural encoders to compute the dense representations of the query and document. This enables dense retrieval to infer the relevance between them at the semantic level rather than the surface level, thus circumventing the term mismatch problem suffered by the sparse retrieval models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.434,
+                0.885,
+                0.658
+            ],
+            "angle": 0,
+            "content": "In recent years, the dual-encoder architecture has been a standard workhorse for dense retrieval. One major disadvantage of this architecture is that it can only partially extract the interactions between the query and document, since it encodes them separately. By comparison, the cross-encoder architecture can effectively capture the deep correlation between them by taking the concatenation of the query and document as input. By directly concatenating the query and document, the cross-encoder gains an advantage in capturing interactions, but also loses the advantage of pre-computing document representations during inference. Therefore, cross-encoder cannot wholly replace dual-encoder."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.662,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "To enhance the retrieval models' ability to capture interactions between queries and documents while maintaining retrieval efficiency, previous work mainly focuses on generating query-informed document representations. One approach, known as late interaction (Khattab and Zaharia, 2020), involves encoding the query and document independently in the early layers, while the later layers model their interactions. Late interaction combines dual-encoder and cross-encoder, making a trade-off between retrieval efficiency and performance. On the other hand, Li et al. (2022) proposed a promising retrieval architecture, dual-cross-encoder. As shown in Figure 1 (c), this architecture computes the query-related document representation by expanding the document with a real or pseudo query."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.141,
+                0.894,
+                0.486,
+                0.919
+            ],
+            "angle": 0,
+            "content": "*Work done during internship at Microsoft Research Asia. †Corresponding authors."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10531"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.153,
+                0.946,
+                0.843,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10531-10541 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.488
+            ],
+            "angle": 0,
+            "content": "Compared with late interaction, dual-cross-encoder (i.e., dense retrieval with document expansion) gets the query-related document representation without sacrificing the retrieval efficiency at inference. However, there exists a discrepancy between training and inference in the current dual-cross-encoder retriever. Specifically, during training, the document is expanded using a real query, whereas during inference, the document is enriched with a generated query. This discrepancy causes the learned retriever overly focus on the query, yet neglect the document, when computing the document representation. During inference, if the generated query \\( q' \\) significantly differs from the user-input query \\( q \\), the query-related document representation will be misled by \\( q' \\), thus degrading the performance. That is why the dual-cross-encoder even underperforms the vanilla dual-encoder. To address this issue, Li et al. (2022) proposed a solution by computing multiview document representations using different generated queries for each document. While multiview document representations improve retrieval performance, they come at the cost of significantly increased retrieval latency, which scales linearly with the number of views."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.495,
+                0.492,
+                0.866
+            ],
+            "angle": 0,
+            "content": "In this paper, we propose CAPSTONE, a curriculum sampling for dense retrieval with document expansion, to bridge the gap between training and inference for dual-cross-encoder. Our motivation is to expect the dual-cross-encoder retrieval model can utilize both the document \\(d\\) and pseudo query \\(q^{\\prime}\\) to compute the query-informed document representation. To achieve this, we train the dual-cross-encoder retriever by gradually increasing the relevance of the pseudo query \\(q^{\\prime}\\) to the gold query \\(q\\). Specifically, at the early training stage, a pseudo query \\(q^{\\prime}\\) irrelevant to \\(q\\) is selected, causing the retriever to solely rely on the document. As we progress to the late training stage, a highly related pseudo query \\(q^{\\prime}\\) is chosen, allowing the retriever to learn to augment the document representation with the pseudo query. By doing so, we alleviate the discrepancy between training and inference. During inference, if the user-input query \\(q\\) is similar to the pseudo query \\(q^{\\prime}\\), then \\(q^{\\prime}\\) will contribute more to making the target document \\(d\\) be retrieved. Otherwise, the retrieval model will mainly rely on the relevance between \\(q\\) and \\(d\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.871,
+                0.493,
+                0.922
+            ],
+            "angle": 0,
+            "content": "To summarize, the main contributions of this paper are as follows: (1) We propose a curriculum learning approach to bridge the gap between train"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.887,
+                0.362
+            ],
+            "angle": 0,
+            "content": "ing and inference for dense retrieval with document expansion, further improving the query-informed document representation<sup>1</sup>. (2) We propose to compute the typical document representation rather than using multiview document representations at inference, which balances the retrieval efficiency and performance. (3) To verify the effectiveness, we apply our proposed approach to two different dense retrieval models, DPR (Karpukhin et al., 2020) and coCondenser, and conduct extensive experiments on three in-domain retrieval datasets, and the zero-shot BEIR benchmark. Experimental results show our proposed approach brings substantial gains over competitive baselines. (4) To the best of our knowledge, this is the first time that document expansion is successfully applied to dense retrieval without incurring extra retrieval latency."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.372,
+                0.651,
+                0.389
+            ],
+            "angle": 0,
+            "content": "2 Preliminary"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.398,
+                0.887,
+                0.431
+            ],
+            "angle": 0,
+            "content": "In this section, we introduce the definition of text retrieval and three architectures for dense retrieval."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.457,
+                0.886,
+                0.537
+            ],
+            "angle": 0,
+            "content": "Task Description. Text retrieval is meant to find the most relevant \\( M \\) documents \\( D^{+} = \\{d_{1}^{+}, d_{2}^{+}, \\ldots, d_{M}^{+}\\} \\) for the given query \\( q \\) from a large corpus \\( D = \\{d_{1}, d_{2}, \\ldots, d_{N}\\} \\) with \\( N \\) documents (\\( M \\ll N \\))."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.548,
+                0.886,
+                0.724
+            ],
+            "angle": 0,
+            "content": "Dual-Encoder Architecture. Dual-encoder (DE) is the typical dense retrieval architecture. As shown in Figure 1 (a), it consists of a query encoder \\( E_{q} \\) and a document encoder \\( E_{d} \\), which encode the query \\( q \\) and document \\( d \\) into dense vectors \\( E_{q}(q) \\) and \\( E_{d}(d) \\), respectively. Previous works usually initialize the encoders with BERT and use the [CLS] vector at the last layer as the dense representation. The similarity score between \\( q \\) and \\( d \\) is measured with the inner product of their vectors:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.591,
+                0.737,
+                0.884,
+                0.757
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d). \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.77,
+                0.886,
+                0.883
+            ],
+            "angle": 0,
+            "content": "Cross-Encoder Architecture. Since DE models \\(q\\) and \\(d\\) separately, it is not good at capturing the relevance between them. To capture the interactions between them, cross-encoder (CE) directly takes the concatenation of \\(q\\) and \\(d\\) as input, as shown in Figure 1 (b). It first computes the dense representation \\(E(q + d)\\) for them and then uses the fully"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": "1Our code is available at: https://github.com/microsoft/SimXNS/CAPSTONE."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10532"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.125,
+                0.086,
+                0.345,
+                0.212
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.18,
+                0.217,
+                0.29,
+                0.23
+            ],
+            "angle": 0,
+            "content": "(a) Dual-encoder"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.352,
+                0.087,
+                0.564,
+                0.213
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.4,
+                0.217,
+                0.515,
+                0.23
+            ],
+            "angle": 0,
+            "content": "(b) Cross-encoder"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.571,
+                0.087,
+                0.877,
+                0.213
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.651,
+                0.217,
+                0.796,
+                0.23
+            ],
+            "angle": 0,
+            "content": "(c) Dual-cross-encoder"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.245,
+                0.882,
+                0.275
+            ],
+            "angle": 0,
+            "content": "Figure 1: Illustration of the dual-encoder, cross-encoder and dual-cross-encoder architectures. \\( q \\) and \\( q' \\) denote the gold query and generated query for the document \\( d \\). '+ is the concatenation operation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.299,
+                0.487,
+                0.331
+            ],
+            "angle": 0,
+            "content": "connected layers (FCL) to compute the similarity score:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.185,
+                0.343,
+                0.487,
+                0.361
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {s i m} (q, d) = F C L (E (d + q)), \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.371,
+                0.42,
+                0.387
+            ],
+            "angle": 0,
+            "content": "where \\(^{\\prime} + ^{\\prime}\\) is the concatenation operation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.395,
+                0.49,
+                0.604
+            ],
+            "angle": 0,
+            "content": "Dual-Cross-Encoder Architecture. Although CE can extract more fine-grained relevance between \\( q \\) and \\( d \\), it is not suitable for retrieval. To combine the advantages of DE and CE, Li et al. (2022) proposed dual-cross-encoder (DCE). As shown in Figure 1 (c), DCE also consists of a query encoder and a document encoder. The only difference between DE and DCE is that the document encoder of DCE takes the document and the generated query \\( q' \\) as input during inference. Therefore, DCE can be regarded as DE with document expansion. Similar to DE, DCE computes the similarity score between \\( q \\) and \\( d \\) with the inner product:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.178,
+                0.614,
+                0.488,
+                0.634
+            ],
+            "angle": 0,
+            "content": "\\[\n\\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d + q ^ {\\prime}). \\tag {3}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.643,
+                0.264,
+                0.66
+            ],
+            "angle": 0,
+            "content": "3 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.669,
+                0.489,
+                0.749
+            ],
+            "angle": 0,
+            "content": "In Section 3.1, we first identify the discrepancy between training and inference of DCE. Then, we introduce how to bridge the gap with curriculum learning in Section 3.2. Finally, we will show our proposed inference method in Section 3.3."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.759,
+                0.465,
+                0.775
+            ],
+            "angle": 0,
+            "content": "3.1 Discrepancy in Training and Inference"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.78,
+                0.489,
+                0.844
+            ],
+            "angle": 0,
+            "content": "The training objective of DE is to learn dense representations of queries and documents to maximize the similarity score between the query and positive document. The training loss is defined as follows:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.151,
+                0.853,
+                0.45,
+                0.922
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} L (q, d ^ {+}, D ^ {-}) \\\\ = - l o g \\frac {e ^ {s i m (q , d ^ {+})}}{e ^ {s i m (q , d ^ {+})} + \\sum_ {d ^ {-} \\in D ^ {-}} e ^ {s i m (q , d ^ {-})}}, \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.295,
+                0.885,
+                0.444
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.453,
+                0.884,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Figure 2: Comparison of different document expansion methods when training DCE, where the left is the vanilla method and the right is our proposed method. \\(d\\) is the positive document for \\(q\\). \\(d_{1}, \\ldots, d_{k}\\) are the hard negatives. \\(d_{k+1}, \\ldots, d_{n}\\) are the in-batch negatives. \\(q'\\) and \\(q_{i}'\\) are the generated query for \\(d\\) and \\(d_{i}\\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.565,
+                0.884,
+                0.597
+            ],
+            "angle": 0,
+            "content": "where \\(D^{-}\\) is the set of negative documents, containing hard negatives and in-batch negatives."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.598,
+                0.885,
+                0.839
+            ],
+            "angle": 0,
+            "content": "During training, DCE expands the positive document \\(d\\) and hard negatives with the gold query \\(q\\) (e.g., replacing \\(d\\) with \\(d + q\\)). For ease of understanding, we show how to construct positive, hard negatives and in-batch negatives in the left of Figure 2. As shown in this figure, DCE can filter out all in-batch negatives by solely relying on the query information from the expanded document, since it only needs to observe whether the input query appears in the expanded document. The utilization of document information is necessary only for distinguishing the positive from the hard negatives. Consequently, the learned representation of the expanded document has a strong bias toward the query, while almost neglecting the document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.884,
+                0.92
+            ],
+            "angle": 0,
+            "content": "At inference, DCE enriches the document with the generated query \\( q' \\) rather than the user-input query \\( q \\). When \\( q \\) and \\( q' \\) are different types of queries for the target document \\( d \\), \\( q \\) or \\( E_q(q) \\) will be far from \\( q' \\) or \\( E_d(q') \\). As the retrieval model"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10533"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.081,
+                0.488,
+                0.246
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.256,
+                0.49,
+                0.384
+            ],
+            "angle": 0,
+            "content": "Figure 3: The query selection process based on curriculum learning in each training iteration. We first compute the relevance score \\( M(q, q_i') \\) between the generated query \\( q_i' \\) and the gold query \\( q \\), and then sort the generated queries in ascending order according to their relevance scores to \\( q \\). Next, we divide them into K groups. At the \\( i \\)-th training iteration, we randomly sample one generated query from the \\( i \\)-th group, and then use the sampled query to update the dual-cross encoder."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.413,
+                0.489,
+                0.572
+            ],
+            "angle": 0,
+            "content": "overly relies on the query part when computing the document representation, \\( E_{d}(d + q^{\\prime}) \\) will be close to \\( E_{d}(q^{\\prime}) \\). Based on these, we can easily infer that \\( E_{d}(d + q^{\\prime}) \\) is far away from \\( E_{q}(q) \\), making it more difficult to retrieve the target document \\( d \\). As a result, the dense retriever with document expansion even underperforms its counterpart without using document expansion. We attribute the performance degradation to the discrepancy between training and inference."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.589,
+                0.436,
+                0.621
+            ],
+            "angle": 0,
+            "content": "3.2 Bridging the Gap with Curriculum Sampling"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.629,
+                0.489,
+                0.789
+            ],
+            "angle": 0,
+            "content": "To bridge the gap between training and inference, we propose to expand the document with its generated query during training. We show the training process of our proposed method in the right of Figure 2. Although this idea is simple and intuitive, it brings a significant advantage: the retriever cannot take a shortcut during training, where the retriever cannot exclude in-batch negatives by simply checking whether the user-input query appears in the expanded document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "However, this is not the end of the story. If the generated query \\( q' \\) used to expand \\( d \\) is irrelevant to the gold query \\( q \\), the document retriever of DCE will ignore \\( q' \\), which means DCE will degenerate into DE. On the other hand, if \\( q' \\) and \\( q \\) are similar, we will encounter the abovementioned problem (DCE will overly depend on \\( q' \\)). We expect the dense retriever with document expansion can use"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.512,
+                0.082,
+                0.886,
+                0.191
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.199,
+                0.883,
+                0.243
+            ],
+            "angle": 0,
+            "content": "Figure 4: The overview of the retrieval stage. \\( d_{i} \\) is the \\( i - th \\) document, and \\( N \\) refers to the size of the external corpus. \\( q_{i}^{j} \\) means the \\( j \\)-th generated query for \\( d_{i} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.269,
+                0.884,
+                0.316
+            ],
+            "angle": 0,
+            "content": "both the document and the generated query during training, so that the learned document representation contains the information of both parts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.317,
+                0.885,
+                0.558
+            ],
+            "angle": 0,
+            "content": "To fulfill this goal, we further propose the curriculum sampling. To be concrete, at the early training stage, the selected \\( q' \\) has low correlations with \\( q \\), forcing the retriever to use the document. As the training goes on, we select \\( q' \\) with gradually increased relevance to \\( q \\), encouraging the retriever to use the query. Driven by this motivation, we first generate some queries for each document \\( d \\). Then, we compute the relevance score between the generated query \\( q_i' \\) and the gold query \\( q \\) with an automatic evaluation metric \\( M \\) (i.e., ROUGE-\\( L^2 \\)). After that, we sort the generated queries in ascending order according to their relevance scores to \\( q \\) and divide them into \\( K \\) groups. We summarize the curriculum sampling in Figure 3."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.57,
+                0.631,
+                0.584
+            ],
+            "angle": 0,
+            "content": "3.3 Inference"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.592,
+                0.884,
+                0.686
+            ],
+            "angle": 0,
+            "content": "Before retrieving, we first generate \\( S \\) queries for each document with the generator. Next, we concatenate a document with one generated query and compute its latent representation with the document encoder. Then, we will get \\( S \\) different representations for each document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.697,
+                0.884,
+                0.793
+            ],
+            "angle": 0,
+            "content": "Corpus Expansion. Following Li et al. (2022), we keep \\(S\\) representations for each document, so the original corpus will be expanded to \\(S\\) times. When a query comes, we need to retrieve the relevant documents from the expanded corpus, which increases the retrieval latency by \\(S\\) times."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.804,
+                0.884,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Computing the Typical Representation of Different Views. To avoid increasing the retrieval latency, we can retain a single representation for each document. The easiest way is to set \\( S \\) to 1. However, we found that the retrieval performance"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "2The evaluation code for ROUGE-L is available at: https://huggingface.co/spaces/evaluate-metric/rouge."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10534"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.083,
+                0.488,
+                0.141
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td><td>#Passage</td></tr><tr><td>MS-MARCO</td><td>502,939</td><td>6,980</td><td>-</td><td>8,841,823</td></tr><tr><td>TREC-19</td><td>-</td><td>-</td><td>43</td><td>8,841,823</td></tr><tr><td>TREC-20</td><td>-</td><td>-</td><td>54</td><td>8,841,823</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.15,
+                0.486,
+                0.181
+            ],
+            "angle": 0,
+            "content": "Table 1: Basic statistics of retrieval datasets. #Passage is the number of unique passages in the retrieval corpus."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.208,
+                0.489,
+                0.304
+            ],
+            "angle": 0,
+            "content": "is positively correlated to the number of views, \\( S \\). This is because a larger value of \\( S \\) increases the probability of the user-input query being relevant to one of the generated queries. Therefore, setting \\( S \\) to 1 will degrade the retrieval performance compared with expanding the corpus."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.306,
+                0.489,
+                0.435
+            ],
+            "angle": 0,
+            "content": "Given this, we propose to compute the typical representation for a document by averaging all \\( S \\) different views. This approach allows us to maintain only one representation for each document. As shown in Figure 4, our proposed method does not bring additional retrieval latency\\(^3\\), yet it performs much better than simply setting \\( S \\) to 1 (please refer to Section 4.3 for more details)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.452,
+                0.262,
+                0.468
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.481,
+                0.326,
+                0.497
+            ],
+            "angle": 0,
+            "content": "4.1 Experimental Setups"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.505,
+                0.49,
+                0.761
+            ],
+            "angle": 0,
+            "content": "Datasets. We conduct experiments on three passage retrieval datasets: MS-MARCO passage ranking (Bajaj et al., 2016), TREC Deep Learning (DL) Track 2019 (Craswell et al., 2020) and 2020 (Craswell et al., 2021). MS-MARCO collects real user queries from Bing search and passages from web collection. TREC DL tracks share the same retrieval corpus with MS-MARCO and provide another two annotated test sets. We train our models on the MS-MARCO train set, and then evaluate them on the MS-MARCO development set\\(^4\\), TREC DL 19 and 20 test sets. Table 1 shows the statistics of these datasets. We evaluate the zero-shot retrieval performance on BEIR benchmark (Thakur et al., 2021), which contains 18 datasets across different domains."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.777,
+                0.49,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Evaluation Metrics. Following previous work, we use MRR@10, Recall@50, and Recall@1000 to evaluate the retrieval performance on MS-MARCO, where MRR@10 is the most important metric. We resort to nDCG@10 for TREC DL"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.084,
+                0.734,
+                0.099
+            ],
+            "angle": 0,
+            "content": "tracks and BEIR benchmark<sup>5</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.112,
+                0.885,
+                0.401
+            ],
+            "angle": 0,
+            "content": "Implementation Details. We set the maximum document length to be 144 tokens, and the maximum query length to be 32 tokens. All models are optimized using the AdamW optimizer (Loshchilov and Hutter, 2019) with the learning rate of \\(5 \\times 10^{-6}\\), and a linear learning rate schedule with \\(10\\%\\) warm-up steps for around three epochs. Every training batch contains 64 queries, and each query is accompanied by one positive and 31 hard negative documents. Following Gao and Callan (2022), we train our proposed model for two stages and initialize the retriever with coCondenser at each stage. At the first training stage, the hard negatives are sampled from the official BM25 hard negatives, but at the second training stage, the hard negatives are sampled from the mined hard negatives. For each training stage, we evaluate the last model training checkpoint on the retrieval datasets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.403,
+                0.885,
+                0.692
+            ],
+            "angle": 0,
+            "content": "The query generator is built upon the seq2seq architecture (Sutskever et al., 2014; He and Yiu, 2022), designed to take a passage as its input and has been specifically trained to create synthetic queries that correspond to the provided passage. Nogueira and Lin (2019) initialized the query generator with the T5-base model. They trained it on the MS-MARCO dataset and generated 80 queries for each passage within MS-MARCO. Following Li et al. (2022), we use the queries released by Nogueira and Lin (2019) to expand the passages in the MS-MARCO dataset. In parallel, for BEIR, we use the query generator fine-tuned on MS-MARCO to generate 5 queries for each document within the BEIR benchmark. During the query generation process, we employ the top-\\(k\\) sampling approach with \\(k\\) set to 10, and we limit the maximum length of the synthetic queries to 64 tokens."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.693,
+                0.885,
+                0.821
+            ],
+            "angle": 0,
+            "content": "During training, we implement our proposed curriculum sampling with all 80 queries and divide them into \\(K\\) groups, but at inference, we resort to the first 10 (i.e., \\(S = 10\\)) and first 5 (i.e., \\(S = 5\\)) queries to compute the typical document representation for MS-MARCO and BEIR benchmark, respectively. At the first training stage, we set \\(K\\) to 3, and at the second training stage, we set \\(K\\) to 4."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.832,
+                0.884,
+                0.856
+            ],
+            "angle": 0,
+            "content": "5The official evaluation toolkit is available at https:// github.com/beir-cellar/beir."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.51,
+                0.857,
+                0.871,
+                0.894
+            ],
+            "angle": 0,
+            "content": "\\(^{6}\\)https://git.uwaterloo.ca/jimmylin/doc2query-data/raw/master/T5-passage/predicted Queries_topk_sampleing.zip"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.511,
+                0.894,
+                0.792,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>7</sup>https://huggingface.co/castorini/doc2query-t5-base-msmarco"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.832,
+                0.884,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.868,
+                0.488,
+                0.906
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>We can pre-generate queries and pre-compute document representations. Therefore, the retrieval latency mentioned in this paper does not encompass these pre-processing steps."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.138,
+                0.906,
+                0.452,
+                0.918
+            ],
+            "angle": 0,
+            "content": "4The MS-MARCO test set is not publicly available."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.868,
+                0.488,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10535"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.142,
+                0.082,
+                0.858,
+                0.332
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"3\">MS-MARCO</td><td>TREC DL 19</td><td>TREC DL 20</td></tr><tr><td>MRR@10</td><td>R@50</td><td>R@1000</td><td>nDCG@10</td><td>nDCG@10</td></tr><tr><td>Sparse retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>BM25 (Yang et al., 2017)</td><td>18.5</td><td>58.5</td><td>85.7</td><td>51.2</td><td>47.7</td></tr><tr><td>DeepCT (Dai and Callan, 2019)</td><td>24.3</td><td>69.0</td><td>91.0</td><td>57.2</td><td>-</td></tr><tr><td>DocT5Query (Nogueira and Lin, 2019)</td><td>27.7</td><td>75.6</td><td>94.7</td><td>64.2</td><td>-</td></tr><tr><td>Dense retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>DPR (Karpukhin et al., 2020)</td><td>31.4</td><td>-</td><td>95.3</td><td>59.0</td><td>62.1</td></tr><tr><td>ANCE (Xiong et al., 2021)</td><td>33.0</td><td>-</td><td>95.9</td><td>64.5</td><td>64.6</td></tr><tr><td>SEED (Lu et al., 2021)</td><td>33.9</td><td>-</td><td>96.1</td><td>-</td><td>-</td></tr><tr><td>STAR (Zhan et al., 2021)</td><td>34.7</td><td>-</td><td>-</td><td>68.3</td><td>-</td></tr><tr><td>TAS-B (Hofstätter et al., 2021)</td><td>34.0</td><td>-</td><td>97.5</td><td>71.2</td><td>69.3</td></tr><tr><td>RocketQA (Qu et al., 2021)</td><td>37.0</td><td>85.5</td><td>97.9</td><td>-</td><td>-</td></tr><tr><td>COIL (Gao et al., 2021)</td><td>35.5</td><td>-</td><td>96.3</td><td>70.4</td><td>-</td></tr><tr><td>ColBERT (Khattab and Zaharia, 2020)</td><td>36.0</td><td>82.9</td><td>96.8</td><td>-</td><td>-</td></tr><tr><td>DCE (Li et al., 2022)</td><td>36.0</td><td>-</td><td>96.4</td><td>68.3</td><td>68.9</td></tr><tr><td>RetroMAE (Xiao et al., 2022)</td><td>35.0</td><td>-</td><td>97.6</td><td>-</td><td>-</td></tr><tr><td>Condenser (Gao and Callan, 2021)</td><td>36.6</td><td>-</td><td>97.4</td><td>69.8</td><td>-</td></tr><tr><td>coCondenser (Gao and Callan, 2022)*</td><td>37.9</td><td>86.3</td><td>98.4</td><td>70.7</td><td>69.8</td></tr><tr><td>CAPSTONE</td><td>38.6</td><td>86.6</td><td>98.6</td><td>71.1</td><td>70.3</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.341,
+                0.882,
+                0.357
+            ],
+            "angle": 0,
+            "content": "Table 2: Passage retrieval results on MS-Marco Dev, and TREC datasets. Results with * are from our reproduction."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.381,
+                0.489,
+                0.461
+            ],
+            "angle": 0,
+            "content": "During inference, we pre-compute representations for all documents in the retrieval corpus, and build IndexFlatIP indexes for document representations with the FAISS library (Johnson et al., 2019) to accelerate retrieval."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.462,
+                0.49,
+                0.576
+            ],
+            "angle": 0,
+            "content": "We implement all models with the HuggingFace Transformers library (Wolf et al., 2019) and conduct all experiments on 8 NVIDIA Tesla V100 GPUs with 32 GB memory. We also use the gradient checkpointing technology (Chen et al., 2016) to reduce the memory of deep neural models, and mixed precision training\\(^{8}\\) to speed up training."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.588,
+                0.332,
+                0.604
+            ],
+            "angle": 0,
+            "content": "4.2 Experimental Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.61,
+                0.49,
+                0.804
+            ],
+            "angle": 0,
+            "content": "In-domain Performance. Table 2 shows the performance of our proposed model and baselines on MS-MARCO, TREC-2019 and TREC-2020. Since we initialize our model, CAPSTONE, with coCondenser, coCondenser is the main comparison model. From Table 2, we can see our proposed model improves coCondenser by a large margin, increasing by 0.7 points in MRR@10 on MS-MARCO, 0.4 points on TREC DL 19 and 0.5 points on TREC DL 20 datasets. These improvements verify the effectiveness of our proposed curriculum learning and typical representation strategy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.815,
+                0.49,
+                0.895
+            ],
+            "angle": 0,
+            "content": "Zero-shot Performance. To test the out-of-domain generalization capabilities, we first use the T5-base model (Nogueira and Lin, 2019) trained on MS-MARCO to generate 5 queries for each document of BEIR benchmark, and then evaluate our"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.381,
+                0.885,
+                0.575
+            ],
+            "angle": 0,
+            "content": "model, CAPSTONE, fine-tuned with MS-MARCO on the BEIR benchmark. During evaluation, we compute the typical representation for each document with \\( S = 5 \\). As shown in Table 3, CAPSTONE performs much better than all baselines in the average performance. Compared with its counterpart without document expansion (i.e., coCondenser), our model achieves improvements on 11 out of 18 datasets and increases nDCG@10 by 0.6 points in the average performance. These improvements demonstrate that our approach generalizes well on diverse domains under zero-shot settings."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.584,
+                0.783,
+                0.6
+            ],
+            "angle": 0,
+            "content": "4.3 Ablation Study and Analysis"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.605,
+                0.884,
+                0.686
+            ],
+            "angle": 0,
+            "content": "To analyze our method more directly, experiments in this section are based on the vanilla DPR initialized with the ERNIE-2.0-base model (Sun et al., 2020) without using dense retrieval pre-training strategies, if there is no particular statement."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.693,
+                0.885,
+                0.887
+            ],
+            "angle": 0,
+            "content": "Corpus Expansion vs. Document Expansion. In this paper, we append only one query to a document both at training and inference. As stated in 3.3, corpus expansion will enlarge the original corpus to \\( S \\) times during inference. In contrast, document expansion for sparse retrieval (Nogueira and Lin, 2019) appends multiple generated queries to a document at inference, thus without bringing extra retrieval latency. Since document expansion has shown effectiveness in sparse retrieval, it is very intuitive to consider whether it is effective for dense retrieval. To answer this question, we"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.892,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this section, document expansion means expanding a document with multiple queries."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.904,
+                0.348,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{8}\\)https://github.com/NVIDIA/apex"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10536"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.12,
+                0.082,
+                0.88,
+                0.334
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td>BERT</td><td>LaPraDoR</td><td>SimCSE</td><td>DiffCSE</td><td>SEED</td><td>Condenser</td><td>coCondenser*</td><td>CAPSTONE</td></tr><tr><td>TREC-COVID</td><td>64.9</td><td>49.5</td><td>52.4</td><td>49.2</td><td>61.2</td><td>75.4</td><td>74.0</td><td>77.9</td></tr><tr><td>BioASQ</td><td>26.2</td><td>23.9</td><td>26.4</td><td>25.8</td><td>29.7</td><td>31.7</td><td>34.1</td><td>34.3</td></tr><tr><td>NFCorpus</td><td>25.7</td><td>28.3</td><td>25.0</td><td>25.9</td><td>25.6</td><td>27.8</td><td>32.4</td><td>33.0</td></tr><tr><td>NQ</td><td>43.8</td><td>41.5</td><td>41.2</td><td>41.2</td><td>42.5</td><td>45.9</td><td>50.5</td><td>50.5</td></tr><tr><td>HotpotQA</td><td>47.8</td><td>48.8</td><td>50.2</td><td>49.9</td><td>52.8</td><td>53.7</td><td>56.4</td><td>56.7</td></tr><tr><td>FiQA-2018</td><td>23.7</td><td>26.6</td><td>24.0</td><td>22.9</td><td>24.4</td><td>26.1</td><td>30.0</td><td>30.4</td></tr><tr><td>Signal-1M (RT)</td><td>21.6</td><td>24.5</td><td>26.4</td><td>26.0</td><td>24.6</td><td>25.8</td><td>24.7</td><td>23.1</td></tr><tr><td>TREC-NEWS</td><td>36.2</td><td>20.6</td><td>36.8</td><td>36.3</td><td>33.5</td><td>35.3</td><td>39.1</td><td>40.3</td></tr><tr><td>Robust04</td><td>36.4</td><td>31.0</td><td>35.3</td><td>34.3</td><td>34.8</td><td>35.2</td><td>40.3</td><td>40.7</td></tr><tr><td>ArguAna</td><td>35.7</td><td>50.3</td><td>43.6</td><td>46.8</td><td>34.7</td><td>37.5</td><td>40.9</td><td>39.2</td></tr><tr><td>Touche-2020</td><td>27.0</td><td>17.8</td><td>17.8</td><td>16.8</td><td>18.0</td><td>22.3</td><td>27.0</td><td>31.0</td></tr><tr><td>CQADupStack</td><td>28.4</td><td>32.6</td><td>29.5</td><td>30.5</td><td>28.5</td><td>31.6</td><td>30.0</td><td>30.0</td></tr><tr><td>Quora</td><td>78.2</td><td>84.3</td><td>84.8</td><td>85.0</td><td>84.9</td><td>85.5</td><td>84.3</td><td>83.8</td></tr><tr><td>DBPedia</td><td>29.8</td><td>32.8</td><td>30.4</td><td>30.3</td><td>32.4</td><td>33.1</td><td>37.2</td><td>38.0</td></tr><tr><td>SCIDOCS</td><td>11.5</td><td>14.5</td><td>12.5</td><td>12.5</td><td>11.7</td><td>13.6</td><td>14.3</td><td>14.3</td></tr><tr><td>FEVER</td><td>68.4</td><td>51.8</td><td>65.1</td><td>64.1</td><td>65.3</td><td>68.2</td><td>72.4</td><td>72.7</td></tr><tr><td>Climate-FEVER</td><td>20.5</td><td>17.2</td><td>22.2</td><td>20.0</td><td>17.6</td><td>19.9</td><td>19.4</td><td>19.3</td></tr><tr><td>SciFact</td><td>50.4</td><td>48.3</td><td>54.5</td><td>52.3</td><td>55.6</td><td>57.0</td><td>58.3</td><td>60.5</td></tr><tr><td>Avg. Performance</td><td>37.6</td><td>35.8</td><td>37.7</td><td>37.2</td><td>37.7</td><td>40.3</td><td>42.7</td><td>43.3</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.345,
+                0.884,
+                0.374
+            ],
+            "angle": 0,
+            "content": "Table 3: Zero-shot dense retrieval performances on BEIR benchmark (measured with nDCG@10). Results with * are from our reproduction."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.395,
+                0.489,
+                0.537
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.114,
+                0.544,
+                0.487,
+                0.573
+            ],
+            "angle": 0,
+            "content": "Figure 5: Results of DPR with corpus, document, and asymmetric expansion on the dev set of MS-Marco."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.601,
+                0.49,
+                0.859
+            ],
+            "angle": 0,
+            "content": "conduct experiments on DPR with three different settings: (1) corpus expansion; (2) document expansion; (3) asymmetric expansion, which appends one query to a document at training, but appends \\( S \\) queries to a document at inference. As shown in Figure 5, unlike corpus expansion, using more queries does not bring clear improvements for document and asymmetric expansion, but sometimes damages their performances. To explain this, we illustrate document expansion and corpus expansion in Figure 6. Document expansion models the target document and all queries together, where the irrelevant queries acting like noise will corrupt the document representation. By comparison, corpus expansion nicely bypasses this problem by modeling different queries independently."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.871,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Effect of Query Selection Strategies. To demonstrate the advantage of our proposed curriculum sampling, we expand the positive and negative doc"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.51,
+                0.395,
+                0.883,
+                0.555
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.563,
+                0.884,
+                0.606
+            ],
+            "angle": 0,
+            "content": "Figure 6: Illustration of document expansion and corpus expansion at inference. For simplicity, only the positive document \\( d_{i} \\) of the query \\( q \\) is shown."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.641,
+                0.884,
+                0.754
+            ],
+            "angle": 0,
+            "content": "uments with different query selection strategies: (1) gold, meaning using the gold query (i.e., strategy used by the vanilla DCE (Li et al., 2022)); (2) random, meaning randomly sampling a query from the generated queries; (3) top-\\(k\\) and (4) bottom-\\(k\\) meaning sampling a query from the top-\\(k\\) and bottom-\\(k\\) sorted queries, respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "From Figure 7, we observe that: (1) DPR trained with the gold query performs worst among all strategies, even worse than the vanilla DPR (DPR w/o query) when \\( S < 8 \\). As we have stated above, expanding the document with the gold query during training will make the retriever pay more attention to the query while ignoring the document. At inference, when the provided queries are limited (i.e., \\( S \\) is small), it is unlikely to find a query similar to the user-input query. However, as \\( S \\) increases, the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10537"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.116,
+                0.082,
+                0.488,
+                0.279
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.288,
+                0.49,
+                0.416
+            ],
+            "angle": 0,
+            "content": "Figure 7: Comparison of different query selection strategies. During training, we first expand the document with a query selected from the generated queries using a specific strategy, and then train DPR with the expanded document. During inference, we evaluate the well-trained DPR retrievers with corpus expansion (solid lines) and the typical representation (dotted lines) on the dev set of MS-Marco, where the solid and dotted lines with the same color are results for the same DPR."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.442,
+                0.49,
+                0.782
+            ],
+            "angle": 0,
+            "content": "probability of finding a query similar to the user-input query also increases. (2) DPR trained with the top-1 query performs much better than DPR trained with the gold query, since it slightly shifts its attention from the query to the document. However, it still cannot exceed DPR w/o query when \\( S = 1 \\), indicating that it fails to fully utilize the document like DPR. (3) If the document is expanded with a weakly related (bottom-1) or random query at training, DPR with both settings will outperform DPR w/o query when \\( S = 1 \\). These two strategies enable DPR to mine query information without sacrificing document information. However, their performances are not strongly positively correlated with \\( S \\) like the top-1 strategies, indicating it does not fully use the query information. (4) DPR with the curriculum strategy outperforms DPR w/o query when \\( S = 1 \\), and its performance linearly increases with \\( S \\), verifying this strategy enables DPR to utilize both the document and query to model the query-informed document representation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.791,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Effect of the Typical Representation. Although corpus expansion makes the retrieval performance of our proposed curriculum sampling improve with the increase of \\( S \\) (see the solid red line in Figure 7), the retrieval latency also increases linearly. To alleviate this, we propose to use the average of different document views as the typical representation. Among all strategies (see the dotted lines in the"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.511,
+                0.083,
+                0.883,
+                0.151
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Variants</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td></tr><tr><td>CAPSTONE+DPR w/ S = 1</td><td>35.15</td><td>97.19</td></tr><tr><td>CAPSTONE+DPR w/ average</td><td>35.66</td><td>97.28</td></tr><tr><td>CAPSTONE+DPR w/ max</td><td>35.45</td><td>97.25</td></tr><tr><td>CAPSTONE+DPR w/ median</td><td>35.64</td><td>97.20</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.16,
+                0.883,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Table 4: Results of CAPSTONE initialized with DPR using different pooling methods to compute the typical representation on the dev set of MS-Marco."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.511,
+                0.219,
+                0.885,
+                0.287
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Models</td><td colspan=\"2\">BM25 Negatives</td><td colspan=\"2\">Mined Negatives</td></tr><tr><td>MRR@10</td><td>R@1000</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td><td>36.44</td><td>97.65</td></tr><tr><td>CAPSTONE+DPR</td><td>35.66</td><td>97.28</td><td>37.28</td><td>97.82</td></tr><tr><td>coCondenser</td><td>35.91</td><td>98.21</td><td>37.94</td><td>98.41</td></tr><tr><td>CAPSTONE+coCondenser</td><td>36.75</td><td>98.21</td><td>38.65</td><td>98.60</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.296,
+                0.883,
+                0.338
+            ],
+            "angle": 0,
+            "content": "Table 5: Performance of CAPSTONE initialized with DPR and coCondenser at the two training stages on the dev set of MS-Marco."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.366,
+                0.885,
+                0.593
+            ],
+            "angle": 0,
+            "content": "Figure 7), our proposed curriculum sampling performs best, which again verifies the effectiveness of curriculum sampling. In addition, the retrieval performance of typical representations outperforms its counterpart using corpus expansion with \\( S = 1 \\) for all query selection strategies. More importantly, for the curriculum learning strategy, the typical representation enhances the vanilla DPR by 1.2 on MRR10, and is even on par with its counterpart using corpus expansion with \\( S = 3 \\), but there still exists a clear performance gap between the typical representation and corpus expansion with \\( S = 10 \\). Therefore, the typical representation balances the retrieval efficiency and performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.603,
+                0.884,
+                0.796
+            ],
+            "angle": 0,
+            "content": "Comparison of Methods for Computing the Typical Representation. We consider three different pooling methods to compute the typical document representation: taking the average/max/median pooling of different document views. From Table 4, we observe that (1) DPR with \\( S = 1 \\) improves DPR by 0.9 on MRR@10; (2) DPR with the typical representation brings an extra 0.5 improvements on MRR@10; (3) There is no significant difference in the results of the three pooling methods. Given the simplicity, we use the average pooling to compute the typical representation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Multi-stage Retrieval Performance. To further verify the effectiveness of our proposed approach, we apply our proposed approach to DPR and coCondenser. From Table 5, we observe that our approach brings clear improvements to DPR with 1.4 and 0.8 increases on MRR@10 at the two training stages, respectively. In addition, we witness similar"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10538"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.488,
+                0.133
+            ],
+            "angle": 0,
+            "content": "improvements on coCondenser, proving that our approach does not depend on retrieval models and is applicable to different retrieval models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.146,
+                0.271,
+                0.161
+            ],
+            "angle": 0,
+            "content": "5 Related Work"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.172,
+                0.285,
+                0.187
+            ],
+            "angle": 0,
+            "content": "5.1 Dense Retrieval"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.194,
+                0.49,
+                0.612
+            ],
+            "angle": 0,
+            "content": "In recent years, with the development of large-scale pre-trained language models, such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), we witness the research interest in information retrieval shifting from the traditional sparse retrieval to neural dense retrieval. Karpukhin et al. (2020); Xiong et al. (2021) proposed dense passage retriever (DPR), which uses two neural encoders initialized with BERT to model the query and document, independently. The subsequent works follow this dual-encoder framework. One line of work improves dense retrieval by mining hard negatives, where they select the top-ranked documents retrieved by the recent retriever as hard negatives and then re-train the retriever with the newly mined hard negatives (Xiong et al., 2021; Qu et al., 2021). However, the mined hard negatives are highly likely to contain false negatives, harming the performance. To mitigate this, following studies denoise the mined hard negatives with re-rankers (Qu et al., 2021; Ren et al., 2021; Zhang et al., 2022). Another kind of work focuses on pre-training to make the pre-trained models more suitable for dense retrieval, such as Condenser (Gao and Callan, 2021), coCondenser (Gao and Callan, 2022) and SIMLM (Wang et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.613,
+                0.49,
+                0.839
+            ],
+            "angle": 0,
+            "content": "Dense retrieval models typically depend on extensive supervised data, comprising pairs of queries and positive documents. To address the challenge of limited training data, Ma et al. (2021); Sun et al. (2021) proposed to train a query generator on high-resource information retrieval data, and then used the query generator to generate synthetic queries for low-source target domains. Additionally, Dai et al. (2023) harnessed large language models in zero/few settings to produce synthetic queries for documents in target domains, eliminating the need for training a general query generator. In contrast to these approaches, our work leverages synthetic queries for document expansion."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.85,
+                0.299,
+                0.866
+            ],
+            "angle": 0,
+            "content": "5.2 Query Expansion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.872,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Query expansion enriches the query with various heuristically discovered relevant contexts. For example, GAR (Mao et al., 2021) expands the query"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.886,
+                0.214
+            ],
+            "angle": 0,
+            "content": "by adding relevant contexts, including the answer, the sentence containing the answer, and the title of a passage where the answer belongs. However, query expansion will increase the retrieval latency during inference, since the generation model is required to generate these relevant contexts for query expansion. Therefore, we mainly focus on document expansion in this work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.229,
+                0.725,
+                0.244
+            ],
+            "angle": 0,
+            "content": "5.3 Document Expansion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.252,
+                0.886,
+                0.526
+            ],
+            "angle": 0,
+            "content": "Document expansion augments the document with generated queries, which the document might answer. Compared with query expansion, document expansion can be conducted prior to indexing without incurring extra retrieval latency. Document expansion has shown its effectiveness on sparse retrieval models (Nogueira et al., 2019; Nogueira and Lin, 2019), yet how to apply it to dense retrieval models is still under-explored. Li et al. (2022) first applied document expansion to dense retrieval models. However, their models suffer from the discrepancy between training and inference, thus even degrading the vanilla dense retrieval models under the single-view document representation settings. To mitigate this discrepancy, we propose a curriculum learning approach, which successfully applies document expansion to dense retrieval models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.542,
+                0.642,
+                0.557
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.57,
+                0.886,
+                0.732
+            ],
+            "angle": 0,
+            "content": "This work proposes the curriculum sampling for dense retrieval with document expansion, which enables dense retrieval models to learn much better query-related document representations. In addition, we propose to compute the typical representation of different document views, which balances inference efficiency and effectiveness. Our experimental results on the in- and out-of-domain datasets verify the effectiveness of the curriculum sampling and typical representation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.747,
+                0.646,
+                0.762
+            ],
+            "angle": 0,
+            "content": "7 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.775,
+                0.886,
+                0.92
+            ],
+            "angle": 0,
+            "content": "There are two possible limitations of this work. The first limitation is that we need to generate synthetic queries for each document in the retrieval corpus, which is very time-consuming. Luckily, this process does not bring extra delay to retrieval. In addition, limited by sufficient computational resources, we only verify the effectiveness of our method on vanilla DPR and coCondenser. In the future, we plan to apply our approach to other dense retrieval"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10539"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.489,
+                0.117
+            ],
+            "angle": 0,
+            "content": "models and verify the effectiveness of our method on these models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.129,
+                0.287,
+                0.145
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.153,
+                0.49,
+                0.283
+            ],
+            "angle": 0,
+            "content": "This project is supported by HKU-SCF FinTech Academy, Shenzhen-Hong Kong-Macao Science and Technology Plan Project (Category C Project: SGDX20210823103537030), and Theme-based Research Scheme of RGC, Hong Kong (T35-710/20-R). We would like to thank the anonymous reviewers for their constructive and informative feedback on this work."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.308,
+                0.214,
+                0.323
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.331,
+                0.49,
+                0.41
+            ],
+            "angle": 0,
+            "content": "Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, et al. 2016. Ms marco: A human generated machine reading comprehension dataset. arXiv preprint arXiv:1611.09268."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.419,
+                0.49,
+                0.485
+            ],
+            "angle": 0,
+            "content": "Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer open-domain questions. In Proceedings of ACL, pages 1870-1879, Vancouver, Canada. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.494,
+                0.489,
+                0.534
+            ],
+            "angle": 0,
+            "content": "Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. 2016. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.543,
+                0.489,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the trec 2020 deep learning track. arXiv preprint arXiv:2102.07662."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.605,
+                0.489,
+                0.657
+            ],
+            "angle": 0,
+            "content": "Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the trec 2019 deep learning track. arXiv preprint arXiv:2003.07820."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.667,
+                0.488,
+                0.707
+            ],
+            "angle": 0,
+            "content": "Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. arXiv preprint arXiv:1910.10687."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.716,
+                0.49,
+                0.769
+            ],
+            "angle": 0,
+            "content": "Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In ICLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.778,
+                0.49,
+                0.843
+            ],
+            "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 NAACL-HLT, pages 4171-4186."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.853,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Luyu Gao and Jamie Callan. 2021. Condenser: a pretraining architecture for dense retrieval. In Proceedings of EMNLP, pages 981-993, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.331,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.886,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Luyu Gao and Jamie Callan. 2022. Unsupervised corpus aware language model pre-training for dense passage retrieval. In Proceedings of ACL, pages 2843-2853, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.165,
+                0.885,
+                0.232
+            ],
+            "angle": 0,
+            "content": "Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. COIL: Revisit exact lexical match in information retrieval with contextualized inverted list. In Proceedings of NAACL-HLT, pages 3030-3042, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.245,
+                0.885,
+                0.35
+            ],
+            "angle": 0,
+            "content": "Xingwei He, Yeyun Gong, A-Long Jin, Weizhen Qi, Hang Zhang, Jian Jiao, Bartuer Zhou, Biao Cheng, Sm Yiu, and Nan Duan. 2022. Metric-guided distillation: Distilling knowledge from the metric to ranker and retriever for generative commonsense reasoning. In Proceedings of EMNLP, pages 839-852, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.363,
+                0.885,
+                0.43
+            ],
+            "angle": 0,
+            "content": "Xingwei He and Siu Ming Yiu. 2022. Controllable dictionary example generation: Generating example sentences for specific targeted audiences. In Proceedings of ACL, pages 610-627, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.442,
+                0.885,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Sebastian Hofstätter, Sheng-Chieh Lin, Zheng-Hong Yang, Jimmy Lin, and Allan Hanbury. 2021. Efficiently teaching an effective dense retriever with balanced topic aware sampling. In Proceedings of SIGIR, page 113-122, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.535,
+                0.885,
+                0.575
+            ],
+            "angle": 0,
+            "content": "Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.588,
+                0.885,
+                0.668
+            ],
+            "angle": 0,
+            "content": "Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense passage retrieval for open-domain question answering. In Proceedings of EMNLP, pages 6769-6781, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.681,
+                0.885,
+                0.748
+            ],
+            "angle": 0,
+            "content": "Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of SIGIR, page 39-48, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.76,
+                0.885,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of NIPS, Red Hook, NY, USA. Curran Associates Inc."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.866,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Zehan Li, Nan Yang, Liang Wang, and Furu Wei. 2022. Learning diverse document representations with deep query interactions for dense retrieval. arXiv preprint arXiv:2208.04232."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.886,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10540"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                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,
+                0.163,
+                0.486,
+                0.19
+            ],
+            "angle": 0,
+            "content": "Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.2,
+                0.487,
+                0.292
+            ],
+            "angle": 0,
+            "content": "Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is more: Pretrain a strong Siamese encoder for dense text retrieval using a weak decoder. In Proceedings of EMNLP, pages 2780-2791, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.303,
+                0.487,
+                0.369
+            ],
+            "angle": 0,
+            "content": "Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot neural passage retrieval via domain-targeted synthetic question generation. In Proceedings of EACL, pages 1075-1088, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.38,
+                0.487,
+                0.458
+            ],
+            "angle": 0,
+            "content": "Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for opendomain question answering. In Proceedings of ACL, pages 4089-4100, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.469,
+                0.486,
+                0.497
+            ],
+            "angle": 0,
+            "content": "Rodrigo Nogueira and Jimmy Lin. 2019. From doc2query to docctttttquery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.507,
+                0.486,
+                0.547
+            ],
+            "angle": 0,
+            "content": "Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction. arXiv preprint arXiv:1904.08375."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.557,
+                0.487,
+                0.65
+            ],
+            "angle": 0,
+            "content": "Yingqi Qu, Yuchen Ding, Jing Liu, Kai Liu, Ruiyang Ren, Wayne Xin Zhao, Daxiang Dong, Hua Wu, and Haifeng Wang. 2021. RocketQA: An optimized training approach to dense passage retrieval for open-domain question answering. In Proceedings of NAACL-HLT, pages 5835-5847, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.66,
+                0.487,
+                0.752
+            ],
+            "angle": 0,
+            "content": "Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of EMNLP, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.763,
+                0.487,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Si Sun, Yingzhuo Qian, Zhenghao Liu, Chenyan Xiong, Kaitao Zhang, Jie Bao, Zhiyuan Liu, and Paul Bennett. 2021. Few-shot text ranking with meta adapted synthetic weak supervision. In Proceedings of ACL, pages 5030-5043, Online. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.118,
+                0.852,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Ernie 2.0: A continual pre-training framework for language understanding. In Proceedings of AAAI, volume 34, pages 8968-8975."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In NIPS, volume 27. Curran Associates, Inc."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.883,
+                0.214
+            ],
+            "angle": 0,
+            "content": "Nandan Thakur, Nils Reimers, Andreas Rückle, Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogeneous benchmark for zero-shot evaluation of information retrieval models. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.225,
+                0.883,
+                0.316
+            ],
+            "angle": 0,
+            "content": "James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopoulos, and Arpit Mittal. 2018. The fact extraction and VERIFICATION (FEVER) shared task. In Proceedings of the First Workshop on Fact Extraction and VERIFICATION (FEVER), pages 1-9, Brussels, Belgium. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.327,
+                0.883,
+                0.392
+            ],
+            "angle": 0,
+            "content": "Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. SimLM: Pre-training with representation bottleneck for dense passage retrieval. arXiv preprint arXiv:2207.02578."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.402,
+                0.883,
+                0.481
+            ],
+            "angle": 0,
+            "content": "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, R'emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.491,
+                0.883,
+                0.569
+            ],
+            "angle": 0,
+            "content": "Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. RetroMAE: Pre-training retrieval-oriented language models via masked auto-encoder. In Proceedings of EMNLP, pages 538-548, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.58,
+                0.883,
+                0.644
+            ],
+            "angle": 0,
+            "content": "Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.655,
+                0.883,
+                0.72
+            ],
+            "angle": 0,
+            "content": "Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of SIGIR, page 1253-1256, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.731,
+                0.883,
+                0.797
+            ],
+            "angle": 0,
+            "content": "Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of SIGIR, page 1503-1512, New York, NY, USA. Association for Computing Machinery."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.807,
+                0.883,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022. Adversarial retriever-ranker for dense text retrieval. In ICLR."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.846
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10541"
+        }
+    ]
+]

2023/CAPSTONE_ Curriculum Sampling for Dense Retrieval with Document Expansion/d66707f9-dcd1-4cb5-a6c2-30915753860b_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:819c887dd35743eaa5dda8d34635df2eff96e3d2e74d6a382600c7001b25fe3e
+size 756548

2023/CAPSTONE_ Curriculum Sampling for Dense Retrieval with Document Expansion/full.md

@@ -0,0 +1,277 @@
+# CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion
+
+Xingwei He $^{1*}$ , Yeyun Gong $^{2,\dagger}$ , A-Long Jin $^{1}$ , Hang Zhang $^{2}$ , Anlei Dong $^{3}$ , Jian Jiao $^{3}$ , Siu Ming Yiu $^{1,\dagger}$ , Nan Duan $^{2}$
+
+<sup>1</sup>The University of Hong Kong, <sup>2</sup>Microsoft Research Asia, <sup>3</sup>Microsoft hexingwei15@gmail.com, ajin@eee.hku.hk, smyiu@cs.hku.hk, {yegong, v-zhang, anlei.dong, jian.jiao, NANDuan}@microsoft.com
+
+# Abstract
+
+The dual-encoder has become the de facto architecture for dense retrieval. Typically, it computes the latent representations of the query and document independently, thus failing to fully capture the interactions between the query and document. To alleviate this, recent research has focused on obtaining query-informed document representations. During training, it expands the document with a real query, but during inference, it replaces the real query with a generated one. This inconsistency between training and inference causes the dense retrieval model to prioritize query information while disregarding the document when computing the document representation. Consequently, it performs even worse than the vanilla dense retrieval model because its performance heavily relies on the relevance between the generated queries and the real query. In this paper, we propose a curriculum sampling strategy that utilizes pseudo queries during training and progressively enhances the relevance between the generated query and the real query. By doing so, the retrieval model learns to extend its attention from the document alone to both the document and query, resulting in high-quality query-informed document representations. Experimental results on both in-domain and out-of-domain datasets demonstrate that our approach outperforms previous dense retrieval models.
+
+# 1 Introduction
+
+Text retrieval aims to find the relevant documents for a given query from a large collection of documents, playing an indispensable role in open-domain question answering (Chen et al., 2017), fact verification (Thorne et al., 2018) and retrieval-augmented generation (Lewis et al., 2020; He et al., 2022). At the early stage, sparse retrieval methods such as TF-IDF or BM25 dominated passage retrieval by relying mainly on lexical term matching
+
+to compute relevance between the query and document. Recently, there has been a surge of research interest in neural network-based dense retrieval (Karpukhin et al., 2020; Xiong et al., 2021). Different from sparse retrieval, dense retrieval resorts to neural encoders to compute the dense representations of the query and document. This enables dense retrieval to infer the relevance between them at the semantic level rather than the surface level, thus circumventing the term mismatch problem suffered by the sparse retrieval models.
+
+In recent years, the dual-encoder architecture has been a standard workhorse for dense retrieval. One major disadvantage of this architecture is that it can only partially extract the interactions between the query and document, since it encodes them separately. By comparison, the cross-encoder architecture can effectively capture the deep correlation between them by taking the concatenation of the query and document as input. By directly concatenating the query and document, the cross-encoder gains an advantage in capturing interactions, but also loses the advantage of pre-computing document representations during inference. Therefore, cross-encoder cannot wholly replace dual-encoder.
+
+To enhance the retrieval models' ability to capture interactions between queries and documents while maintaining retrieval efficiency, previous work mainly focuses on generating query-informed document representations. One approach, known as late interaction (Khattab and Zaharia, 2020), involves encoding the query and document independently in the early layers, while the later layers model their interactions. Late interaction combines dual-encoder and cross-encoder, making a trade-off between retrieval efficiency and performance. On the other hand, Li et al. (2022) proposed a promising retrieval architecture, dual-cross-encoder. As shown in Figure 1 (c), this architecture computes the query-related document representation by expanding the document with a real or pseudo query.
+
+Compared with late interaction, dual-cross-encoder (i.e., dense retrieval with document expansion) gets the query-related document representation without sacrificing the retrieval efficiency at inference. However, there exists a discrepancy between training and inference in the current dual-cross-encoder retriever. Specifically, during training, the document is expanded using a real query, whereas during inference, the document is enriched with a generated query. This discrepancy causes the learned retriever overly focus on the query, yet neglect the document, when computing the document representation. During inference, if the generated query  $q'$  significantly differs from the user-input query  $q$ , the query-related document representation will be misled by  $q'$ , thus degrading the performance. That is why the dual-cross-encoder even underperforms the vanilla dual-encoder. To address this issue, Li et al. (2022) proposed a solution by computing multiview document representations using different generated queries for each document. While multiview document representations improve retrieval performance, they come at the cost of significantly increased retrieval latency, which scales linearly with the number of views.
+
+In this paper, we propose CAPSTONE, a curriculum sampling for dense retrieval with document expansion, to bridge the gap between training and inference for dual-cross-encoder. Our motivation is to expect the dual-cross-encoder retrieval model can utilize both the document  $d$  and pseudo query  $q^{\prime}$  to compute the query-informed document representation. To achieve this, we train the dual-cross-encoder retriever by gradually increasing the relevance of the pseudo query  $q^{\prime}$  to the gold query  $q$ . Specifically, at the early training stage, a pseudo query  $q^{\prime}$  irrelevant to  $q$  is selected, causing the retriever to solely rely on the document. As we progress to the late training stage, a highly related pseudo query  $q^{\prime}$  is chosen, allowing the retriever to learn to augment the document representation with the pseudo query. By doing so, we alleviate the discrepancy between training and inference. During inference, if the user-input query  $q$  is similar to the pseudo query  $q^{\prime}$ , then  $q^{\prime}$  will contribute more to making the target document  $d$  be retrieved. Otherwise, the retrieval model will mainly rely on the relevance between  $q$  and  $d$ .
+
+To summarize, the main contributions of this paper are as follows: (1) We propose a curriculum learning approach to bridge the gap between train
+
+ing and inference for dense retrieval with document expansion, further improving the query-informed document representation<sup>1</sup>. (2) We propose to compute the typical document representation rather than using multiview document representations at inference, which balances the retrieval efficiency and performance. (3) To verify the effectiveness, we apply our proposed approach to two different dense retrieval models, DPR (Karpukhin et al., 2020) and coCondenser, and conduct extensive experiments on three in-domain retrieval datasets, and the zero-shot BEIR benchmark. Experimental results show our proposed approach brings substantial gains over competitive baselines. (4) To the best of our knowledge, this is the first time that document expansion is successfully applied to dense retrieval without incurring extra retrieval latency.
+
+# 2 Preliminary
+
+In this section, we introduce the definition of text retrieval and three architectures for dense retrieval.
+
+Task Description. Text retrieval is meant to find the most relevant  $M$  documents  $D^{+} = \{d_{1}^{+}, d_{2}^{+}, \ldots, d_{M}^{+}\}$  for the given query  $q$  from a large corpus  $D = \{d_{1}, d_{2}, \ldots, d_{N}\}$  with  $N$  documents ( $M \ll N$ ).
+
+Dual-Encoder Architecture. Dual-encoder (DE) is the typical dense retrieval architecture. As shown in Figure 1 (a), it consists of a query encoder  $E_{q}$  and a document encoder  $E_{d}$ , which encode the query  $q$  and document  $d$  into dense vectors  $E_{q}(q)$  and  $E_{d}(d)$ , respectively. Previous works usually initialize the encoders with BERT and use the [CLS] vector at the last layer as the dense representation. The similarity score between  $q$  and  $d$  is measured with the inner product of their vectors:
+
+$$
+\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d). \tag {1}
+$$
+
+Cross-Encoder Architecture. Since DE models  $q$  and  $d$  separately, it is not good at capturing the relevance between them. To capture the interactions between them, cross-encoder (CE) directly takes the concatenation of  $q$  and  $d$  as input, as shown in Figure 1 (b). It first computes the dense representation  $E(q + d)$  for them and then uses the fully
+
+![](images/012cff165e4a6b68de3be894d4ddfbd24d766bd2c0923035a2c1221467023e8a.jpg)  
+(a) Dual-encoder
+
+![](images/f68ed6d10d993773052bbf470053a2fb6732b7da7fbe9c3c3eca6eec1f0184ae.jpg)  
+(b) Cross-encoder
+
+![](images/4f3f43b05c0c590da89241fc943a88d725bbed1a25aef9272e6198f7136d442f.jpg)  
+(c) Dual-cross-encoder
+
+connected layers (FCL) to compute the similarity score:
+
+$$
+\operatorname {s i m} (q, d) = F C L (E (d + q)), \tag {2}
+$$
+
+where  $^{\prime} + ^{\prime}$  is the concatenation operation.
+
+Dual-Cross-Encoder Architecture. Although CE can extract more fine-grained relevance between  $q$  and  $d$ , it is not suitable for retrieval. To combine the advantages of DE and CE, Li et al. (2022) proposed dual-cross-encoder (DCE). As shown in Figure 1 (c), DCE also consists of a query encoder and a document encoder. The only difference between DE and DCE is that the document encoder of DCE takes the document and the generated query  $q'$  as input during inference. Therefore, DCE can be regarded as DE with document expansion. Similar to DE, DCE computes the similarity score between  $q$  and  $d$  with the inner product:
+
+$$
+\operatorname {s i m} (q, d) = E _ {q} (q) ^ {T} E _ {d} (d + q ^ {\prime}). \tag {3}
+$$
+
+# 3 Methodology
+
+In Section 3.1, we first identify the discrepancy between training and inference of DCE. Then, we introduce how to bridge the gap with curriculum learning in Section 3.2. Finally, we will show our proposed inference method in Section 3.3.
+
+# 3.1 Discrepancy in Training and Inference
+
+The training objective of DE is to learn dense representations of queries and documents to maximize the similarity score between the query and positive document. The training loss is defined as follows:
+
+$$
+\begin{array}{l} L (q, d ^ {+}, D ^ {-}) \\ = - l o g \frac {e ^ {s i m (q , d ^ {+})}}{e ^ {s i m (q , d ^ {+})} + \sum_ {d ^ {-} \in D ^ {-}} e ^ {s i m (q , d ^ {-})}}, \\ \end{array}
+$$
+
+![](images/30085f324ac07f9808fc296835fc77d1212dbb3a0d64a3525ec0e1a691b89ebc.jpg)  
+Figure 1: Illustration of the dual-encoder, cross-encoder and dual-cross-encoder architectures.  $q$  and  $q'$  denote the gold query and generated query for the document  $d$ . '+ is the concatenation operation.  
+Figure 2: Comparison of different document expansion methods when training DCE, where the left is the vanilla method and the right is our proposed method.  $d$  is the positive document for  $q$ .  $d_{1}, \ldots, d_{k}$  are the hard negatives.  $d_{k+1}, \ldots, d_{n}$  are the in-batch negatives.  $q'$  and  $q_{i}'$  are the generated query for  $d$  and  $d_{i}$ .
+
+where  $D^{-}$  is the set of negative documents, containing hard negatives and in-batch negatives.
+
+During training, DCE expands the positive document  $d$  and hard negatives with the gold query  $q$  (e.g., replacing  $d$  with  $d + q$ ). For ease of understanding, we show how to construct positive, hard negatives and in-batch negatives in the left of Figure 2. As shown in this figure, DCE can filter out all in-batch negatives by solely relying on the query information from the expanded document, since it only needs to observe whether the input query appears in the expanded document. The utilization of document information is necessary only for distinguishing the positive from the hard negatives. Consequently, the learned representation of the expanded document has a strong bias toward the query, while almost neglecting the document.
+
+At inference, DCE enriches the document with the generated query  $q'$  rather than the user-input query  $q$ . When  $q$  and  $q'$  are different types of queries for the target document  $d$ ,  $q$  or  $E_q(q)$  will be far from  $q'$  or  $E_d(q')$ . As the retrieval model
+
+![](images/26d553333d43693bc1b567699099fd6b33e03e09c0626836a25e3ac15d685877.jpg)  
+Figure 3: The query selection process based on curriculum learning in each training iteration. We first compute the relevance score  $M(q, q_i')$  between the generated query  $q_i'$  and the gold query  $q$ , and then sort the generated queries in ascending order according to their relevance scores to  $q$ . Next, we divide them into K groups. At the  $i$ -th training iteration, we randomly sample one generated query from the  $i$ -th group, and then use the sampled query to update the dual-cross encoder.
+
+overly relies on the query part when computing the document representation,  $E_{d}(d + q^{\prime})$  will be close to  $E_{d}(q^{\prime})$ . Based on these, we can easily infer that  $E_{d}(d + q^{\prime})$  is far away from  $E_{q}(q)$ , making it more difficult to retrieve the target document  $d$ . As a result, the dense retriever with document expansion even underperforms its counterpart without using document expansion. We attribute the performance degradation to the discrepancy between training and inference.
+
+# 3.2 Bridging the Gap with Curriculum Sampling
+
+To bridge the gap between training and inference, we propose to expand the document with its generated query during training. We show the training process of our proposed method in the right of Figure 2. Although this idea is simple and intuitive, it brings a significant advantage: the retriever cannot take a shortcut during training, where the retriever cannot exclude in-batch negatives by simply checking whether the user-input query appears in the expanded document.
+
+However, this is not the end of the story. If the generated query  $q'$  used to expand  $d$  is irrelevant to the gold query  $q$ , the document retriever of DCE will ignore  $q'$ , which means DCE will degenerate into DE. On the other hand, if  $q'$  and  $q$  are similar, we will encounter the abovementioned problem (DCE will overly depend on  $q'$ ). We expect the dense retriever with document expansion can use
+
+![](images/147ac2e7998f584b10a083627fb114a5d33e5ae3739c44fc46fe7da2421b93be.jpg)  
+Figure 4: The overview of the retrieval stage.  $d_{i}$  is the  $i - th$  document, and  $N$  refers to the size of the external corpus.  $q_{i}^{j}$  means the  $j$ -th generated query for  $d_{i}$ .
+
+both the document and the generated query during training, so that the learned document representation contains the information of both parts.
+
+To fulfill this goal, we further propose the curriculum sampling. To be concrete, at the early training stage, the selected  $q'$  has low correlations with  $q$ , forcing the retriever to use the document. As the training goes on, we select  $q'$  with gradually increased relevance to  $q$ , encouraging the retriever to use the query. Driven by this motivation, we first generate some queries for each document  $d$ . Then, we compute the relevance score between the generated query  $q_i'$  and the gold query  $q$  with an automatic evaluation metric  $M$  (i.e., ROUGE- $L^2$ ). After that, we sort the generated queries in ascending order according to their relevance scores to  $q$  and divide them into  $K$  groups. We summarize the curriculum sampling in Figure 3.
+
+# 3.3 Inference
+
+Before retrieving, we first generate  $S$  queries for each document with the generator. Next, we concatenate a document with one generated query and compute its latent representation with the document encoder. Then, we will get  $S$  different representations for each document.
+
+Corpus Expansion. Following Li et al. (2022), we keep  $S$  representations for each document, so the original corpus will be expanded to  $S$  times. When a query comes, we need to retrieve the relevant documents from the expanded corpus, which increases the retrieval latency by  $S$  times.
+
+Computing the Typical Representation of Different Views. To avoid increasing the retrieval latency, we can retain a single representation for each document. The easiest way is to set  $S$  to 1. However, we found that the retrieval performance
+
+<table><tr><td>Datasets</td><td>Train</td><td>Dev</td><td>Test</td><td>#Passage</td></tr><tr><td>MS-MARCO</td><td>502,939</td><td>6,980</td><td>-</td><td>8,841,823</td></tr><tr><td>TREC-19</td><td>-</td><td>-</td><td>43</td><td>8,841,823</td></tr><tr><td>TREC-20</td><td>-</td><td>-</td><td>54</td><td>8,841,823</td></tr></table>
+
+Table 1: Basic statistics of retrieval datasets. #Passage is the number of unique passages in the retrieval corpus.
+
+is positively correlated to the number of views,  $S$ . This is because a larger value of  $S$  increases the probability of the user-input query being relevant to one of the generated queries. Therefore, setting  $S$  to 1 will degrade the retrieval performance compared with expanding the corpus.
+
+Given this, we propose to compute the typical representation for a document by averaging all  $S$  different views. This approach allows us to maintain only one representation for each document. As shown in Figure 4, our proposed method does not bring additional retrieval latency $^3$ , yet it performs much better than simply setting  $S$  to 1 (please refer to Section 4.3 for more details).
+
+# 4 Experiments
+
+# 4.1 Experimental Setups
+
+Datasets. We conduct experiments on three passage retrieval datasets: MS-MARCO passage ranking (Bajaj et al., 2016), TREC Deep Learning (DL) Track 2019 (Craswell et al., 2020) and 2020 (Craswell et al., 2021). MS-MARCO collects real user queries from Bing search and passages from web collection. TREC DL tracks share the same retrieval corpus with MS-MARCO and provide another two annotated test sets. We train our models on the MS-MARCO train set, and then evaluate them on the MS-MARCO development set $^4$ , TREC DL 19 and 20 test sets. Table 1 shows the statistics of these datasets. We evaluate the zero-shot retrieval performance on BEIR benchmark (Thakur et al., 2021), which contains 18 datasets across different domains.
+
+Evaluation Metrics. Following previous work, we use MRR@10, Recall@50, and Recall@1000 to evaluate the retrieval performance on MS-MARCO, where MRR@10 is the most important metric. We resort to nDCG@10 for TREC DL
+
+tracks and BEIR benchmark<sup>5</sup>.
+
+Implementation Details. We set the maximum document length to be 144 tokens, and the maximum query length to be 32 tokens. All models are optimized using the AdamW optimizer (Loshchilov and Hutter, 2019) with the learning rate of  $5 \times 10^{-6}$ , and a linear learning rate schedule with  $10\%$  warm-up steps for around three epochs. Every training batch contains 64 queries, and each query is accompanied by one positive and 31 hard negative documents. Following Gao and Callan (2022), we train our proposed model for two stages and initialize the retriever with coCondenser at each stage. At the first training stage, the hard negatives are sampled from the official BM25 hard negatives, but at the second training stage, the hard negatives are sampled from the mined hard negatives. For each training stage, we evaluate the last model training checkpoint on the retrieval datasets.
+
+The query generator is built upon the seq2seq architecture (Sutskever et al., 2014; He and Yiu, 2022), designed to take a passage as its input and has been specifically trained to create synthetic queries that correspond to the provided passage. Nogueira and Lin (2019) initialized the query generator with the T5-base model. They trained it on the MS-MARCO dataset and generated 80 queries for each passage within MS-MARCO. Following Li et al. (2022), we use the queries released by Nogueira and Lin (2019) to expand the passages in the MS-MARCO dataset. In parallel, for BEIR, we use the query generator fine-tuned on MS-MARCO to generate 5 queries for each document within the BEIR benchmark. During the query generation process, we employ the top- $k$  sampling approach with  $k$  set to 10, and we limit the maximum length of the synthetic queries to 64 tokens.
+
+During training, we implement our proposed curriculum sampling with all 80 queries and divide them into  $K$  groups, but at inference, we resort to the first 10 (i.e.,  $S = 10$ ) and first 5 (i.e.,  $S = 5$ ) queries to compute the typical document representation for MS-MARCO and BEIR benchmark, respectively. At the first training stage, we set  $K$  to 3, and at the second training stage, we set  $K$  to 4.
+
+<table><tr><td rowspan="2">Models</td><td colspan="3">MS-MARCO</td><td>TREC DL 19</td><td>TREC DL 20</td></tr><tr><td>MRR@10</td><td>R@50</td><td>R@1000</td><td>nDCG@10</td><td>nDCG@10</td></tr><tr><td>Sparse retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>BM25 (Yang et al., 2017)</td><td>18.5</td><td>58.5</td><td>85.7</td><td>51.2</td><td>47.7</td></tr><tr><td>DeepCT (Dai and Callan, 2019)</td><td>24.3</td><td>69.0</td><td>91.0</td><td>57.2</td><td>-</td></tr><tr><td>DocT5Query (Nogueira and Lin, 2019)</td><td>27.7</td><td>75.6</td><td>94.7</td><td>64.2</td><td>-</td></tr><tr><td>Dense retrieval</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>DPR (Karpukhin et al., 2020)</td><td>31.4</td><td>-</td><td>95.3</td><td>59.0</td><td>62.1</td></tr><tr><td>ANCE (Xiong et al., 2021)</td><td>33.0</td><td>-</td><td>95.9</td><td>64.5</td><td>64.6</td></tr><tr><td>SEED (Lu et al., 2021)</td><td>33.9</td><td>-</td><td>96.1</td><td>-</td><td>-</td></tr><tr><td>STAR (Zhan et al., 2021)</td><td>34.7</td><td>-</td><td>-</td><td>68.3</td><td>-</td></tr><tr><td>TAS-B (Hofstätter et al., 2021)</td><td>34.0</td><td>-</td><td>97.5</td><td>71.2</td><td>69.3</td></tr><tr><td>RocketQA (Qu et al., 2021)</td><td>37.0</td><td>85.5</td><td>97.9</td><td>-</td><td>-</td></tr><tr><td>COIL (Gao et al., 2021)</td><td>35.5</td><td>-</td><td>96.3</td><td>70.4</td><td>-</td></tr><tr><td>ColBERT (Khattab and Zaharia, 2020)</td><td>36.0</td><td>82.9</td><td>96.8</td><td>-</td><td>-</td></tr><tr><td>DCE (Li et al., 2022)</td><td>36.0</td><td>-</td><td>96.4</td><td>68.3</td><td>68.9</td></tr><tr><td>RetroMAE (Xiao et al., 2022)</td><td>35.0</td><td>-</td><td>97.6</td><td>-</td><td>-</td></tr><tr><td>Condenser (Gao and Callan, 2021)</td><td>36.6</td><td>-</td><td>97.4</td><td>69.8</td><td>-</td></tr><tr><td>coCondenser (Gao and Callan, 2022)*</td><td>37.9</td><td>86.3</td><td>98.4</td><td>70.7</td><td>69.8</td></tr><tr><td>CAPSTONE</td><td>38.6</td><td>86.6</td><td>98.6</td><td>71.1</td><td>70.3</td></tr></table>
+
+Table 2: Passage retrieval results on MS-Marco Dev, and TREC datasets. Results with * are from our reproduction.
+
+During inference, we pre-compute representations for all documents in the retrieval corpus, and build IndexFlatIP indexes for document representations with the FAISS library (Johnson et al., 2019) to accelerate retrieval.
+
+We implement all models with the HuggingFace Transformers library (Wolf et al., 2019) and conduct all experiments on 8 NVIDIA Tesla V100 GPUs with 32 GB memory. We also use the gradient checkpointing technology (Chen et al., 2016) to reduce the memory of deep neural models, and mixed precision training $^{8}$  to speed up training.
+
+# 4.2 Experimental Results
+
+In-domain Performance. Table 2 shows the performance of our proposed model and baselines on MS-MARCO, TREC-2019 and TREC-2020. Since we initialize our model, CAPSTONE, with coCondenser, coCondenser is the main comparison model. From Table 2, we can see our proposed model improves coCondenser by a large margin, increasing by 0.7 points in MRR@10 on MS-MARCO, 0.4 points on TREC DL 19 and 0.5 points on TREC DL 20 datasets. These improvements verify the effectiveness of our proposed curriculum learning and typical representation strategy.
+
+Zero-shot Performance. To test the out-of-domain generalization capabilities, we first use the T5-base model (Nogueira and Lin, 2019) trained on MS-MARCO to generate 5 queries for each document of BEIR benchmark, and then evaluate our
+
+model, CAPSTONE, fine-tuned with MS-MARCO on the BEIR benchmark. During evaluation, we compute the typical representation for each document with  $S = 5$ . As shown in Table 3, CAPSTONE performs much better than all baselines in the average performance. Compared with its counterpart without document expansion (i.e., coCondenser), our model achieves improvements on 11 out of 18 datasets and increases nDCG@10 by 0.6 points in the average performance. These improvements demonstrate that our approach generalizes well on diverse domains under zero-shot settings.
+
+# 4.3 Ablation Study and Analysis
+
+To analyze our method more directly, experiments in this section are based on the vanilla DPR initialized with the ERNIE-2.0-base model (Sun et al., 2020) without using dense retrieval pre-training strategies, if there is no particular statement.
+
+Corpus Expansion vs. Document Expansion. In this paper, we append only one query to a document both at training and inference. As stated in 3.3, corpus expansion will enlarge the original corpus to  $S$  times during inference. In contrast, document expansion for sparse retrieval (Nogueira and Lin, 2019) appends multiple generated queries to a document at inference, thus without bringing extra retrieval latency. Since document expansion has shown effectiveness in sparse retrieval, it is very intuitive to consider whether it is effective for dense retrieval. To answer this question, we
+
+<table><tr><td>Dataset</td><td>BERT</td><td>LaPraDoR</td><td>SimCSE</td><td>DiffCSE</td><td>SEED</td><td>Condenser</td><td>coCondenser*</td><td>CAPSTONE</td></tr><tr><td>TREC-COVID</td><td>64.9</td><td>49.5</td><td>52.4</td><td>49.2</td><td>61.2</td><td>75.4</td><td>74.0</td><td>77.9</td></tr><tr><td>BioASQ</td><td>26.2</td><td>23.9</td><td>26.4</td><td>25.8</td><td>29.7</td><td>31.7</td><td>34.1</td><td>34.3</td></tr><tr><td>NFCorpus</td><td>25.7</td><td>28.3</td><td>25.0</td><td>25.9</td><td>25.6</td><td>27.8</td><td>32.4</td><td>33.0</td></tr><tr><td>NQ</td><td>43.8</td><td>41.5</td><td>41.2</td><td>41.2</td><td>42.5</td><td>45.9</td><td>50.5</td><td>50.5</td></tr><tr><td>HotpotQA</td><td>47.8</td><td>48.8</td><td>50.2</td><td>49.9</td><td>52.8</td><td>53.7</td><td>56.4</td><td>56.7</td></tr><tr><td>FiQA-2018</td><td>23.7</td><td>26.6</td><td>24.0</td><td>22.9</td><td>24.4</td><td>26.1</td><td>30.0</td><td>30.4</td></tr><tr><td>Signal-1M (RT)</td><td>21.6</td><td>24.5</td><td>26.4</td><td>26.0</td><td>24.6</td><td>25.8</td><td>24.7</td><td>23.1</td></tr><tr><td>TREC-NEWS</td><td>36.2</td><td>20.6</td><td>36.8</td><td>36.3</td><td>33.5</td><td>35.3</td><td>39.1</td><td>40.3</td></tr><tr><td>Robust04</td><td>36.4</td><td>31.0</td><td>35.3</td><td>34.3</td><td>34.8</td><td>35.2</td><td>40.3</td><td>40.7</td></tr><tr><td>ArguAna</td><td>35.7</td><td>50.3</td><td>43.6</td><td>46.8</td><td>34.7</td><td>37.5</td><td>40.9</td><td>39.2</td></tr><tr><td>Touche-2020</td><td>27.0</td><td>17.8</td><td>17.8</td><td>16.8</td><td>18.0</td><td>22.3</td><td>27.0</td><td>31.0</td></tr><tr><td>CQADupStack</td><td>28.4</td><td>32.6</td><td>29.5</td><td>30.5</td><td>28.5</td><td>31.6</td><td>30.0</td><td>30.0</td></tr><tr><td>Quora</td><td>78.2</td><td>84.3</td><td>84.8</td><td>85.0</td><td>84.9</td><td>85.5</td><td>84.3</td><td>83.8</td></tr><tr><td>DBPedia</td><td>29.8</td><td>32.8</td><td>30.4</td><td>30.3</td><td>32.4</td><td>33.1</td><td>37.2</td><td>38.0</td></tr><tr><td>SCIDOCS</td><td>11.5</td><td>14.5</td><td>12.5</td><td>12.5</td><td>11.7</td><td>13.6</td><td>14.3</td><td>14.3</td></tr><tr><td>FEVER</td><td>68.4</td><td>51.8</td><td>65.1</td><td>64.1</td><td>65.3</td><td>68.2</td><td>72.4</td><td>72.7</td></tr><tr><td>Climate-FEVER</td><td>20.5</td><td>17.2</td><td>22.2</td><td>20.0</td><td>17.6</td><td>19.9</td><td>19.4</td><td>19.3</td></tr><tr><td>SciFact</td><td>50.4</td><td>48.3</td><td>54.5</td><td>52.3</td><td>55.6</td><td>57.0</td><td>58.3</td><td>60.5</td></tr><tr><td>Avg. Performance</td><td>37.6</td><td>35.8</td><td>37.7</td><td>37.2</td><td>37.7</td><td>40.3</td><td>42.7</td><td>43.3</td></tr></table>
+
+Table 3: Zero-shot dense retrieval performances on BEIR benchmark (measured with nDCG@10). Results with * are from our reproduction.
+
+![](images/882d0aea860342f7dd897bcdd58dc088de610c50976e45de633b778d5854bfc1.jpg)  
+Figure 5: Results of DPR with corpus, document, and asymmetric expansion on the dev set of MS-Marco.
+
+conduct experiments on DPR with three different settings: (1) corpus expansion; (2) document expansion; (3) asymmetric expansion, which appends one query to a document at training, but appends  $S$  queries to a document at inference. As shown in Figure 5, unlike corpus expansion, using more queries does not bring clear improvements for document and asymmetric expansion, but sometimes damages their performances. To explain this, we illustrate document expansion and corpus expansion in Figure 6. Document expansion models the target document and all queries together, where the irrelevant queries acting like noise will corrupt the document representation. By comparison, corpus expansion nicely bypasses this problem by modeling different queries independently.
+
+Effect of Query Selection Strategies. To demonstrate the advantage of our proposed curriculum sampling, we expand the positive and negative doc
+
+![](images/5b11124203b9a5fa5f683d449d7256f0610e3e3ad7b9f7d3fb0a3a2410ac575f.jpg)  
+Figure 6: Illustration of document expansion and corpus expansion at inference. For simplicity, only the positive document  $d_{i}$  of the query  $q$  is shown.
+
+uments with different query selection strategies: (1) gold, meaning using the gold query (i.e., strategy used by the vanilla DCE (Li et al., 2022)); (2) random, meaning randomly sampling a query from the generated queries; (3) top- $k$  and (4) bottom- $k$  meaning sampling a query from the top- $k$  and bottom- $k$  sorted queries, respectively.
+
+From Figure 7, we observe that: (1) DPR trained with the gold query performs worst among all strategies, even worse than the vanilla DPR (DPR w/o query) when  $S < 8$ . As we have stated above, expanding the document with the gold query during training will make the retriever pay more attention to the query while ignoring the document. At inference, when the provided queries are limited (i.e.,  $S$  is small), it is unlikely to find a query similar to the user-input query. However, as  $S$  increases, the
+
+![](images/2ed9c2c8b6f790185e2947d69e1a1d6d81c37a46d58e97bac77d333f4736637b.jpg)  
+Figure 7: Comparison of different query selection strategies. During training, we first expand the document with a query selected from the generated queries using a specific strategy, and then train DPR with the expanded document. During inference, we evaluate the well-trained DPR retrievers with corpus expansion (solid lines) and the typical representation (dotted lines) on the dev set of MS-Marco, where the solid and dotted lines with the same color are results for the same DPR.
+
+probability of finding a query similar to the user-input query also increases. (2) DPR trained with the top-1 query performs much better than DPR trained with the gold query, since it slightly shifts its attention from the query to the document. However, it still cannot exceed DPR w/o query when  $S = 1$ , indicating that it fails to fully utilize the document like DPR. (3) If the document is expanded with a weakly related (bottom-1) or random query at training, DPR with both settings will outperform DPR w/o query when  $S = 1$ . These two strategies enable DPR to mine query information without sacrificing document information. However, their performances are not strongly positively correlated with  $S$  like the top-1 strategies, indicating it does not fully use the query information. (4) DPR with the curriculum strategy outperforms DPR w/o query when  $S = 1$ , and its performance linearly increases with  $S$ , verifying this strategy enables DPR to utilize both the document and query to model the query-informed document representation.
+
+Effect of the Typical Representation. Although corpus expansion makes the retrieval performance of our proposed curriculum sampling improve with the increase of  $S$  (see the solid red line in Figure 7), the retrieval latency also increases linearly. To alleviate this, we propose to use the average of different document views as the typical representation. Among all strategies (see the dotted lines in the
+
+<table><tr><td>Variants</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td></tr><tr><td>CAPSTONE+DPR w/ S = 1</td><td>35.15</td><td>97.19</td></tr><tr><td>CAPSTONE+DPR w/ average</td><td>35.66</td><td>97.28</td></tr><tr><td>CAPSTONE+DPR w/ max</td><td>35.45</td><td>97.25</td></tr><tr><td>CAPSTONE+DPR w/ median</td><td>35.64</td><td>97.20</td></tr></table>
+
+Table 4: Results of CAPSTONE initialized with DPR using different pooling methods to compute the typical representation on the dev set of MS-Marco.  
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">BM25 Negatives</td><td colspan="2">Mined Negatives</td></tr><tr><td>MRR@10</td><td>R@1000</td><td>MRR@10</td><td>R@1000</td></tr><tr><td>DPR</td><td>34.26</td><td>97.02</td><td>36.44</td><td>97.65</td></tr><tr><td>CAPSTONE+DPR</td><td>35.66</td><td>97.28</td><td>37.28</td><td>97.82</td></tr><tr><td>coCondenser</td><td>35.91</td><td>98.21</td><td>37.94</td><td>98.41</td></tr><tr><td>CAPSTONE+coCondenser</td><td>36.75</td><td>98.21</td><td>38.65</td><td>98.60</td></tr></table>
+
+Table 5: Performance of CAPSTONE initialized with DPR and coCondenser at the two training stages on the dev set of MS-Marco.
+
+Figure 7), our proposed curriculum sampling performs best, which again verifies the effectiveness of curriculum sampling. In addition, the retrieval performance of typical representations outperforms its counterpart using corpus expansion with  $S = 1$  for all query selection strategies. More importantly, for the curriculum learning strategy, the typical representation enhances the vanilla DPR by 1.2 on MRR10, and is even on par with its counterpart using corpus expansion with  $S = 3$ , but there still exists a clear performance gap between the typical representation and corpus expansion with  $S = 10$ . Therefore, the typical representation balances the retrieval efficiency and performance.
+
+Comparison of Methods for Computing the Typical Representation. We consider three different pooling methods to compute the typical document representation: taking the average/max/median pooling of different document views. From Table 4, we observe that (1) DPR with  $S = 1$  improves DPR by 0.9 on MRR@10; (2) DPR with the typical representation brings an extra 0.5 improvements on MRR@10; (3) There is no significant difference in the results of the three pooling methods. Given the simplicity, we use the average pooling to compute the typical representation.
+
+Multi-stage Retrieval Performance. To further verify the effectiveness of our proposed approach, we apply our proposed approach to DPR and coCondenser. From Table 5, we observe that our approach brings clear improvements to DPR with 1.4 and 0.8 increases on MRR@10 at the two training stages, respectively. In addition, we witness similar
+
+improvements on coCondenser, proving that our approach does not depend on retrieval models and is applicable to different retrieval models.
+
+# 5 Related Work
+
+# 5.1 Dense Retrieval
+
+In recent years, with the development of large-scale pre-trained language models, such as BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019), we witness the research interest in information retrieval shifting from the traditional sparse retrieval to neural dense retrieval. Karpukhin et al. (2020); Xiong et al. (2021) proposed dense passage retriever (DPR), which uses two neural encoders initialized with BERT to model the query and document, independently. The subsequent works follow this dual-encoder framework. One line of work improves dense retrieval by mining hard negatives, where they select the top-ranked documents retrieved by the recent retriever as hard negatives and then re-train the retriever with the newly mined hard negatives (Xiong et al., 2021; Qu et al., 2021). However, the mined hard negatives are highly likely to contain false negatives, harming the performance. To mitigate this, following studies denoise the mined hard negatives with re-rankers (Qu et al., 2021; Ren et al., 2021; Zhang et al., 2022). Another kind of work focuses on pre-training to make the pre-trained models more suitable for dense retrieval, such as Condenser (Gao and Callan, 2021), coCondenser (Gao and Callan, 2022) and SIMLM (Wang et al., 2022).
+
+Dense retrieval models typically depend on extensive supervised data, comprising pairs of queries and positive documents. To address the challenge of limited training data, Ma et al. (2021); Sun et al. (2021) proposed to train a query generator on high-resource information retrieval data, and then used the query generator to generate synthetic queries for low-source target domains. Additionally, Dai et al. (2023) harnessed large language models in zero/few settings to produce synthetic queries for documents in target domains, eliminating the need for training a general query generator. In contrast to these approaches, our work leverages synthetic queries for document expansion.
+
+# 5.2 Query Expansion
+
+Query expansion enriches the query with various heuristically discovered relevant contexts. For example, GAR (Mao et al., 2021) expands the query
+
+by adding relevant contexts, including the answer, the sentence containing the answer, and the title of a passage where the answer belongs. However, query expansion will increase the retrieval latency during inference, since the generation model is required to generate these relevant contexts for query expansion. Therefore, we mainly focus on document expansion in this work.
+
+# 5.3 Document Expansion
+
+Document expansion augments the document with generated queries, which the document might answer. Compared with query expansion, document expansion can be conducted prior to indexing without incurring extra retrieval latency. Document expansion has shown its effectiveness on sparse retrieval models (Nogueira et al., 2019; Nogueira and Lin, 2019), yet how to apply it to dense retrieval models is still under-explored. Li et al. (2022) first applied document expansion to dense retrieval models. However, their models suffer from the discrepancy between training and inference, thus even degrading the vanilla dense retrieval models under the single-view document representation settings. To mitigate this discrepancy, we propose a curriculum learning approach, which successfully applies document expansion to dense retrieval models.
+
+# 6 Conclusion
+
+This work proposes the curriculum sampling for dense retrieval with document expansion, which enables dense retrieval models to learn much better query-related document representations. In addition, we propose to compute the typical representation of different document views, which balances inference efficiency and effectiveness. Our experimental results on the in- and out-of-domain datasets verify the effectiveness of the curriculum sampling and typical representation.
+
+# 7 Limitations
+
+There are two possible limitations of this work. The first limitation is that we need to generate synthetic queries for each document in the retrieval corpus, which is very time-consuming. Luckily, this process does not bring extra delay to retrieval. In addition, limited by sufficient computational resources, we only verify the effectiveness of our method on vanilla DPR and coCondenser. In the future, we plan to apply our approach to other dense retrieval
+
+models and verify the effectiveness of our method on these models.
+
+# Acknowledgements
+
+This project is supported by HKU-SCF FinTech Academy, Shenzhen-Hong Kong-Macao Science and Technology Plan Project (Category C Project: SGDX20210823103537030), and Theme-based Research Scheme of RGC, Hong Kong (T35-710/20-R). We would like to thank the anonymous reviewers for their constructive and informative feedback on this work.
+
+# 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.  
+Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. 2017. Reading Wikipedia to answer open-domain questions. In Proceedings of ACL, pages 1870-1879, Vancouver, Canada. Association for Computational Linguistics.  
+Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. 2016. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174.  
+Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the trec 2020 deep learning track. arXiv preprint arXiv:2102.07662.  
+Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the trec 2019 deep learning track. arXiv preprint arXiv:2003.07820.  
+Zhuyun Dai and Jamie Callan. 2019. Context-aware sentence/passage term importance estimation for first stage retrieval. arXiv preprint arXiv:1910.10687.  
+Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2023. Promptagator: Few-shot dense retrieval from 8 examples. In ICLR.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186.  
+Luyu Gao and Jamie Callan. 2021. Condenser: a pretraining architecture for dense retrieval. In Proceedings of EMNLP, pages 981-993, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+Luyu Gao and Jamie Callan. 2022. Unsupervised corpus aware language model pre-training for dense passage retrieval. In Proceedings of ACL, pages 2843-2853, Dublin, Ireland. Association for Computational Linguistics.  
+Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. COIL: Revisit exact lexical match in information retrieval with contextualized inverted list. In Proceedings of NAACL-HLT, pages 3030-3042, Online. Association for Computational Linguistics.  
+Xingwei He, Yeyun Gong, A-Long Jin, Weizhen Qi, Hang Zhang, Jian Jiao, Bartuer Zhou, Biao Cheng, Sm Yiu, and Nan Duan. 2022. Metric-guided distillation: Distilling knowledge from the metric to ranker and retriever for generative commonsense reasoning. In Proceedings of EMNLP, pages 839-852, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Xingwei He and Siu Ming Yiu. 2022. Controllable dictionary example generation: Generating example sentences for specific targeted audiences. In Proceedings of ACL, pages 610-627, Dublin, Ireland. Association for Computational Linguistics.  
+Sebastian Hofstätter, Sheng-Chieh Lin, Zheng-Hong Yang, Jimmy Lin, and Allan Hanbury. 2021. Efficiently teaching an effective dense retriever with balanced topic aware sampling. In Proceedings of SIGIR, page 113-122, New York, NY, USA. Association for Computing Machinery.  
+Jeff Johnson, Matthijs Douze, and Herve Jégou. 2019. Billion-scale similarity search with GPUs. IEEE Transactions on Big Data, 7(3):535-547.  
+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 EMNLP, pages 6769-6781, Online. Association for Computational Linguistics.  
+Omar Khattab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of SIGIR, page 39-48, New York, NY, USA. Association for Computing Machinery.  
+Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. In Proceedings of NIPS, Red Hook, NY, USA. Curran Associates Inc.  
+Zehan Li, Nan Yang, Liang Wang, and Furu Wei. 2022. Learning diverse document representations with deep query interactions for dense retrieval. arXiv preprint arXiv:2208.04232.
+
+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.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In ICLR.  
+Shuqi Lu, Di He, Chenyan Xiong, Guolin Ke, Waleed Malik, Zhicheng Dou, Paul Bennett, Tie-Yan Liu, and Arnold Overwijk. 2021. Less is more: Pretrain a strong Siamese encoder for dense text retrieval using a weak decoder. In Proceedings of EMNLP, pages 2780-2791, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Ji Ma, Ivan Korotkov, Yinfei Yang, Keith Hall, and Ryan McDonald. 2021. Zero-shot neural passage retrieval via domain-targeted synthetic question generation. In Proceedings of EACL, pages 1075-1088, Online. Association for Computational Linguistics.  
+Yuning Mao, Pengcheng He, Xiaodong Liu, Yelong Shen, Jianfeng Gao, Jiawei Han, and Weizhu Chen. 2021. Generation-augmented retrieval for opendomain question answering. In Proceedings of ACL, pages 4089-4100, Online. Association for Computational Linguistics.  
+Rodrigo Nogueira and Jimmy Lin. 2019. From doc2query to docctttttquery.  
+Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 2019. Document expansion by query prediction. arXiv preprint arXiv:1904.08375.  
+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 NAACL-HLT, pages 5835-5847, Online. Association for Computational Linguistics.  
+Ruiyang Ren, Yingqi Qu, Jing Liu, Wayne Xin Zhao, QiaoQiao She, Hua Wu, Haifeng Wang, and Ji-Rong Wen. 2021. RocketQAv2: A joint training method for dense passage retrieval and passage re-ranking. In Proceedings of EMNLP, pages 2825-2835, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Si Sun, Yingzhuo Qian, Zhenghao Liu, Chenyan Xiong, Kaitao Zhang, Jie Bao, Zhiyuan Liu, and Paul Bennett. 2021. Few-shot text ranking with meta adapted synthetic weak supervision. In Proceedings of ACL, pages 5030-5043, Online. Association for Computational Linguistics.  
+Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Hao Tian, Hua Wu, and Haifeng Wang. 2020. Ernie 2.0: A continual pre-training framework for language understanding. In Proceedings of AAAI, volume 34, pages 8968-8975.
+
+Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In NIPS, volume 27. Curran Associates, Inc.  
+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).  
+James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopoulos, and Arpit Mittal. 2018. The fact extraction and VERIFICATION (FEVER) shared task. In Proceedings of the First Workshop on Fact Extraction and VERIFICATION (FEVER), pages 1-9, Brussels, Belgium. Association for Computational Linguistics.  
+Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2022. SimLM: Pre-training with representation bottleneck for dense passage retrieval. arXiv preprint arXiv:2207.02578.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, R'emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771.  
+Shitao Xiao, Zheng Liu, Yingxia Shao, and Zhao Cao. 2022. RetroMAE: Pre-training retrieval-oriented language models via masked auto-encoder. In Proceedings of EMNLP, pages 538-548, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk. 2021. Approximate nearest neighbor negative contrastive learning for dense text retrieval. In ICLR.  
+Peilin Yang, Hui Fang, and Jimmy Lin. 2017. Anserini: Enabling the use of lucene for information retrieval research. In Proceedings of SIGIR, page 1253-1256, New York, NY, USA. Association for Computing Machinery.  
+Jingtao Zhan, Jiaxin Mao, Yiqun Liu, Jiafeng Guo, Min Zhang, and Shaoping Ma. 2021. Optimizing dense retrieval model training with hard negatives. In Proceedings of SIGIR, page 1503-1512, New York, NY, USA. Association for Computing Machinery.  
+Hang Zhang, Yeyun Gong, Yelong Shen, Jiancheng Lv, Nan Duan, and Weizhu Chen. 2022. Adversarial retriever-ranker for dense text retrieval. In ICLR.

2023/CAPSTONE_ Curriculum Sampling for Dense Retrieval with Document Expansion/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3604d17429f2fe6299a22d4957ff607c0f9d4478a23d5b0d2f75f4e1f6f130ba
+size 492340

2023/CESAR_ Automatic Induction of Compositional Instructions for Multi-turn Dialogs/54a33beb-057a-4282-9d89-7a5dd37b8152_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0074d621948f19b1be7c1265a5c6c5fb4222bea0bf8adb187c78dcb11b8def0
+size 1208836

2023/CESAR_ Automatic Induction of Compositional Instructions for Multi-turn Dialogs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b0d1d5423b45796a86383eb9e06dabdfba93737df780ba1e65f5a4fb12876b2
+size 3297524

2023/CHEF in the Language Kitchen_ A Generative Data Augmentation Leveraging Korean Morpheme Ingredients/8a3db38c-4d3c-4329-8082-2813529f2076_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31ace1d8149c8db6926197f0cfba59ff0c5f8746354962d12f47b8fa664a357c
+size 1459721

2023/CHEF in the Language Kitchen_ A Generative Data Augmentation Leveraging Korean Morpheme Ingredients/full.md

@@ -0,0 +1,414 @@
+# CHEF in the Language Kitchen: A Generative Data Augmentation Leveraging Korean Morpheme Ingredients
+
+Jaehyung Seo $^{1}$ , Hyeonseok Moon $^{1}$ , Jaewook Lee $^{1}$ , Sugyeong Eo $^{1}$ , Chanjun Park $^{2}$  and Heuseok Lim $^{1\dagger}$
+
+$^{1}$ Department of Computer Science and Engineering, Korea University
+
+2Upstage
+
+$^{1}$  [seojae777, glee889, jaewook133, djtnrud, limhseok]@korea.ac.kr
+
+2chanjun.park@upstage.ai
+
+# Abstract
+
+Korean morphological variations present unique opportunities and challenges in natural language processing (NLP), necessitating an advanced understanding of morpheme-based sentence construction. The complexity of morphological variations allows for diverse sentence forms based on the syntactic-semantic integration of functional morphemes (i.e., affixes) to lexical morphemes (i.e., roots). With this in mind, we propose a method - CHEF, replicating the morphological transformations inherent in sentences based on lexical and functional morpheme combinations through generative data augmentation. CHEF operates using a morpheme blender and a label discriminator, thereby enhancing the diversity of Korean sentence forms by capturing the properties of agglutination while maintaining label consistency. We conduct experiments on Korean multiple classification datasets, improving model performance in full- and few-shot settings. Our proposed method boosts performance beyond the preceding data augmentation methods without incurring external data usage. We demonstrate that our approach achieves comparable results yielded by augmentation techniques that use large language models (LLMs).
+
+# 1 Introduction
+
+As an agglutinative language, Korean encompasses a rich array of functional morphemes (Song, 2006; Park et al., 2018). Deep learning-based research in Korean NLP aims to enhance linguistic construction efficiency through morpheme segmentation. Previous studies have explored morphological analysis to improve model performance (Song and Park, 2019; Lee et al., 2020; Kim and Colineau, 2020; Kim et al., 2022). The Korean language presents distinct challenges and opportunities with respect to transformations contingent upon the combinatory patterns between lexical and functional morphemes (Matteson et al., 2018; Seo et al., 2022).
+
+![](images/0561fa6077ab8ff5e09ae59ef339a73121b63f7047f73b68673f2bd2fd83a17c.jpg)  
+Figure 1: CHEF for data augmentation. The Korean verbs ‘默默地’, ‘默默地’, ‘默默地’ and ‘默默地’ have equivalent meaning of ‘eat’. The color variations in the shapes indicate morphological changes resulting from combinations with functional morphemes. The ‘Recipe N’ denotes a morpheme set that can compose various sentences. ‘Positive’ represents the example of a label attached to the original sentence.
+
+Consequently, the incorporation of morphological alterations following combinatory rules enables a significant diversity of linguistic forms to be generated from given lexical morphemes.
+
+In this sense, Korean datasets are constrained in capturing the intricate rules governing the interactions between lexical and functional morphemes. As a result, the representations derived from these datasets only contain a small fraction of the potential linguistic forms. For instance, Figure 1 demonstrates multiple approaches to represent the concept of 'eat' by employing different ending forms. Additionally, previous data augmentation approaches are rarely designed to generate synthetic data that considers the morphological characteristic of the Korean language, resulting in limited effectiveness in data augmentation. To address these limitations,
+
+it is crucial to develop data augmentation methods that explicitly consider morphological diversity. By incorporating such properties into the augmentation process, we can contribute to the robustness of models and facilitate performance enhancements.
+
+In this paper, we propose CHEF, a data augmentation method designed to construct new synthetic data in alignment with the combinations of lexical and functional morphemes. CHEF is composed of a morpheme blender<sup>1</sup> and label discriminator. For the initial phase, we prepare Korean lexical morphemes from the training dataset as ingredients. The morpheme blender employs the lexical morphemes and generates the new synthetic sentence by leveraging the knowledge of derivational and inflectional rules in conjunction with functional morphemes obtained from a pre-trained generative model. Morpheme blender can generate diverse sentence expressions by incorporating lexical and functional morphemes, yet it does not guarantee the conservation of label information. To adjust the unintended blending, the label discriminator controls the synthetic data generation process of the morpheme blender. Its primary objective is to prevent substantial shifts in meaning and ensure alignment with the original labels. The teacher forcing of the morpheme blender incorporates global information about synthetic data completed by combinations of lexical and functional morphemes. This is achieved through contrastive learning with the label discriminator. Integrating these two modules, CHEF ensures that synthetic data maintains label consistency while blending new morpheme combinations.
+
+As illustrated in Figure 1, CHEF augmentation enriches text data by attaching various functional morphemes to a given set of lexical morphemes from the original sentence while preserving the label information. The main contributions of our method are summarized as follows:
+
+- Despite difficulties enhancing performance in full-shot settings through data augmentation, CHEF demonstrates effectiveness and robustness through in-depth analysis.  
+- We observe that employing contrastive learning between a label discriminator and a morpheme blender is a suitable data augmentation approach for maintaining label consistency.
+
+- CHEF unlocks the morphological diversity in the training data without using any additional external data and approximates or even outperforms the LLM-based data augmentations.  
+- CHEF exhibits effectiveness even with small amounts of data augmentation.
+
+# 2 Related Work
+
+Data augmentation is widely recognized in deep learning research as a valuable approach to addressing the scarcity of annotated data. It aids in ensuring that the distribution of the training data maintains robustness even in unseen tests. Over time, several methods have been developed to augment textual content, expanding the diversity and quantity of available training data.
+
+Zhang et al. (2015) proposed a method of word substitution using a synonym thesaurus based on rules, which is followed by augmentation methods utilizing a thesaurus in the studies of (Dai and Adel, 2020; Daval-Frerot and Weis, 2020). Wei and Zou (2019) suggested an easy data augmentation (EDA) method using word-level replacement, random insertion, random deletion, and random swap. Karimi et al. (2021) introduced an easier data augmentation (AEDA) technique, improving text classification performance by randomly inserting six pre-defined punctuation marks. Another strategy involves Mixup-based data augmentation (Zhang et al., 2018), combining two sample data to create new training data. Mixup augmentation was initially applied to image data and extended to text-based deep neural networks (Guo, 2020). Subsequently, it has also been applied to textual data augmentation using Transformer-based methods, serving as a technique to mitigate model overfitting and enhance generalization capabilities (Sun et al., 2020; Yoon et al., 2021; Kong et al., 2022).
+
+Sennrich et al. (2016) proposed a back translation method for augmenting data by performing a round-trip translation on data written in the original language, using a neural machine translation system trained on human-annotated data. Data augmentation based on back translation has evolved as one of the crucial techniques since it allows for rewriting the entire sentence, rather than just word-level alterations (Fabbri et al., 2021; Lowell et al., 2021; Park et al., 2021a).
+
+Recently, research on data augmentation using pre-trained language models has been actively pursued (Kumar et al., 2020; Du et al., 2021; Schick
+
+![](images/b9b47806489e2668d5a52eb1dc96b89aee0928fb3e14cb7d85b2bb318d36c8a2.jpg)  
+Figure 2: Training process of CHEF with Morpheme Blender  $\mathbf{M}_{\mathrm{G}}$  and Label Discriminator  $\mathbf{M}_{\mathrm{D}}$
+
+and Schütze, 2021; Zhou et al., 2022; Mekala et al., 2022). Among them, the data augmentation technique through counterfactual is particularly noteworthy, demonstrating improvements to the state-of-the-art performance levels across various benchmark datasets (Liu et al., 2021; Joshi and He, 2022; Ou et al., 2022; Wen et al., 2022). With the advent of LLMs, considerable progress has been made in overcoming augmentation constraints through generative AI (Yoo et al., 2021). Thus, we proceed assuming that it is feasible to secure textual content with LLMs. Nevertheless, it is still necessary to expend financial and temporal resources to identify the optimal direction for augmentation. As a result, we focus on determining which augmentation method could benefit model learning from Korean data. Given the diverse derivational and inflectional forms inherent in the agglutinative Korean language, we propose CHEF - an augmentation method that considers these linguistic properties.
+
+# 3 CHEF
+
+Taking inspiration from CommonGen (Lin et al., 2020) and Korean CommonGen (Seo et al., 2022), we focus on the data-to-text generation capabilities of sequence-to-sequence models. CommonGen and Korean CommonGen necessitate generative commonsense reasoning based on given concept sets, where given concept sets are used to determine the feasibility of constructing plausible sentences. Both conditional generation tasks are to make generative language models learn a function  $f: \mathcal{C} \to \mathcal{T}$  that maps a set of input concept
+
+set  $\mathcal{C} = \{c_1, \dots, c_n\}$  to make a target sentence  $\mathcal{T}$  based on the relation within  $\mathcal{C}$ . We perceive a parallel between this process and combining ingredients to create a complete dish. In particular, Seo et al. (2022) have shown that combining lexical morphemes yields optimal parsing for generating Korean sentences. Consequently, we concentrate on harnessing the sentence augmentation effect of sequence-to-sequence models by assembling ingredients of morpheme sets.
+
+# 3.1 Preliminary
+
+The primary objective of CHEF is to augment dataset  $\mathcal{D} = \{(\mathrm{inp}^i,l^i)\}_{i = 1}^N$ , where  $\mathrm{inp}^i = (h^i,p^i)$  denotes textual input and  $l^i$  is its corresponding label. Considering multiple sentence settings such as NLI and STS, we regard  $h^i$  as a hypothesis and define auxiliary input  $p^i$  as a premise. If the input comprises a single sentence (e.g., YNAT), we regard  $h^i$  as the same text as  $\mathrm{inp}^i$ , and  $p^i$  is conceived as an empty string.
+
+The structure of CHEF is illustrated in Figure 2. It consists of two components: the label discriminator module  $(\mathbf{M}_{\mathrm{D}})$  that helps maintain label consistency and the morpheme blender  $(\mathbf{M}_{\mathrm{G}})$  that augments data by referring each morpheme set. In this framework, the label discriminator module is presented as an encoder architecture, such as BERT (Devlin et al., 2018), and the morpheme blender module possesses an encoder-decoder architecture, such as BART (Lewis et al.,  $2020)^{2}$ .
+
+Both modules are trained to generate synthetic data according to the task and label to be augmented. In the following sections, we describe each module in detail.
+
+# 3.2 Label Discriminator
+
+Label discriminator module  $(\mathbf{M}_{\mathbb{D}})$  helps retain the intended label consistency during the data augmentation process.  $\mathbf{M}_{\mathbb{D}}$  is designed to take  $\mathrm{inp}^i$  and offer probabilities to be classified into each label.  $\mathbf{M}_{\mathbb{D}}$  encode inp into the vector space by applying a linear pooling layer followed by the softmax, to the [BOS] position of the last hidden state. We denote the processed output as  $\mathbf{M}_{\mathbb{D}}(\mathrm{inp}^i)\in \mathbb{R}^{1\times n_c}$ , where  $n_c$  indicates the number of classes for the task dealt with  $\mathcal{D}$ .  $\mathbf{M}_{\mathbb{D}}$  is supervised to maximize the probability of each  $\mathrm{inp}^i$  to be classified into  $l^i$ .
+
+Concisely,  $\mathbf{M}_{\mathbb{D}}$  is a task module trained with  $\mathcal{D}$  that provides the label expectation probability obtained by the full contextualized sentence  $\mathrm{inp}^i$ , and is utilized as a label instructor of the morpheme blender, described in the latter sections.
+
+# 3.3 Morpheme Blender
+
+Ingredients Preparation The main objective of the morpheme blender  $(\mathbf{M}_{\mathrm{G}})$  is to synthesize the sentence by given morpheme ingredients and its targeting label. For training  $\mathbf{M}_{\mathrm{G}}$ , we extract lexical morpheme ingredients  $m^{i} = \{m_{1}^{i},\dots ,m_{n_{i}}^{i}\}$  from the  $h^i$ , where  $n^i$  denotes the number of lexical morphemes in  $h^i$ . In preparing  $m^i$ , we use a Korean morphological analyzer, mecab-ko<sup>3</sup>, to make the lexical morpheme set for  $h^i$ . We leverage that the diversity of Korean sentences that can be combined for the same lexical morpheme ingredients is potentially high. Therefore, we construct the morpheme set based on substantial lexical morphemes, allowing the morphemes within the model-aware recipe to be combined with various endings, functional morphemes, and particles.
+
+We also define a label mapping function  $\psi : l \to v$  that maps label  $l^i$  into the label word  $v^i$  that has a form of natural language. For instance, in the case of NLI, we adopt "contradiction" for the label "-1", "neutral" for "0", and "entailment" for "1".
+
+In gathering these, we establish the blender that takes morpheme ingredients  $m^i$  (optionally with  $p^i$ ) and label word  $v^i$  as inputs, and returns newly synthesized sentence  $h'^i$ . Detailed training procedures are described in the following section.
+
+Teacher Forcing For attaining ingredient synthesizing capacity,  $\mathbf{M}_{\mathrm{G}}$  is trained to generate  $h^i$  by referring to the concatenated sequence of  $m^i$ ,  $p^i$ , and  $v^i$ . The encoder framework of  $\mathbf{M}_{\mathrm{G}}$ , denoted as  $\mathbf{Enc}_{\mathrm{G}}$ , receives a concatenated sequence and generates a contextualized representation by capturing the bidirectional interactions of the words in the input. Subsequently,  $\mathbf{M}_{\mathrm{G}}$  produces sentences by utilizing the contextual representation acquired by the encoder. We can define the loss function  $\mathcal{L}_{ce}$  for training sequence-to-sequence generation of  $\mathbf{M}_{\mathrm{G}}$  as in equation (1). For clarity in expression, we denote  $[\dots]$  as a sequentialized concatenation of all the elements in it, and define the concatenated input  $seq^i = [m^i, p^i, v^i]$ .
+
+$$
+\mathcal {L} _ {c e} = - \frac {1}{| \mathcal {D} |} \sum_ {i = 1} ^ {N} \sum_ {j} \log P _ {\mathrm {M} _ {6}} \left(h _ {j} ^ {i} \mid h _ {<   j} ^ {i}; s e q ^ {i}\right) \tag {1}
+$$
+
+Through the sequence-to-sequence training,  $\mathbf{M}_{\mathrm{G}}$  can generate a whole sentence that covers a given morpheme set with weakly reflecting the label information granted by  $v^{i}$  in  $\text{seq}^i$ .
+
+Label Consistency Supervision Even if label information is reflected in the generation process by feeding it as an input, we find that generation outputs from  $\mathbf{M}_{\mathrm{G}}$  still suffer from label inconsistency. To alleviate this, we exploit the auxiliary training objective utilizing pre-trained discriminator  $\mathbf{M}_{\mathrm{G}}$  to consider label consistency.
+
+We argue that by aligning the encoded outputs of  $\mathbf{M}_{\mathrm{D}}$  and  $\mathbf{Enc}_{\mathrm{G}}$ , we can make  $\mathbf{Enc}_{\mathrm{G}}$  better embed contextual relations with the corresponding morpheme sets and task-specific labels and promote label maintenance. This can be distilled to the training objective that maximizes the similarity  $\sin^i$  defined as the following equations:
+
+$$
+\mathbf {r} _ {\mathrm {G}} ^ {i} = \operatorname {S o f t M a x} \left(\operatorname {E n c} _ {\mathrm {G}} (s e q ^ {i}) \cdot \mathrm {W} _ {\mathrm {G}}\right) \tag {2}
+$$
+
+$$
+\operatorname {s i m} ^ {i} (\mathrm {i n p}) = \frac {\mathbf {r} _ {\mathrm {G}} ^ {i} \cdot \mathbf {M} _ {\mathrm {D}} (\mathrm {i n p}))}{\| \mathbf {r} _ {\mathrm {G}} ^ {i} \| \| \mathbf {M} _ {\mathrm {D}} (\mathrm {i n p})) \|} \tag {3}
+$$
+
+Note that through the pre-training of  $\mathbf{M}_{\mathrm{D}}$ , the encoded output of  $\mathbf{M}_{\mathrm{D}}$  represents the label information of the whole sentence  $h^i$  that  $\mathbf{M}_{\mathrm{G}}$  should generate. The auxiliary training objective aims to make the label representation encoded through the set of ingredients  $seq^i$ , be aligned with the label representation yielded by  $h^i$ .
+
+For the direct supervision, we define contrastive sample set  $\mathcal{C}^i\subset \mathcal{D}$  for each  $(\mathrm{inp}^i,l^i)\in \mathcal{D}$ , and
+
+contrastive label set  $\mathbf{l}^i = \mathbf{L}\backslash \{l^i\}$  for each label  $l^i$  , where  $\mathbf{L}$  denotes the set of all possible labels considered in  $\mathcal{D}$  . In comprising  $\mathcal{C}^i$  , we randomly extract a single sample from  $\mathcal{D}$  for each label in  $\mathbf{l}^i$  . Then the loss function  $\mathcal{L}_{cont}$  for learning label consistency is defined as equation (4)
+
+$$
+\mathcal {L} _ {\text {c o n t}} ^ {i} = - \log \frac {\exp \left(\sin^ {i} \left(\operatorname {i n p} ^ {i}\right) / \tau\right)}{\sum_ {\left(\operatorname {i n p} ^ {j} , l ^ {j}\right) \in \mathcal {C} ^ {i}} \exp \left(\sin^ {i} \left(\operatorname {i n p} ^ {j}\right) / \tau\right)} \tag {4}
+$$
+
+$\mathrm{W_G}$  denotes the linear pooling layer that maps the encoded representation on the [BOS] position into the label classification probability. The temperature parameter  $\tau$  controls the sharpness of the softmax distribution with larger values leading to a smoother distribution and smaller values leading to a more peaked distribution.
+
+In summing these,  $\mathbf{M}_{\mathrm{G}}$  is trained with  $\mathcal{L}_{total}$  defined as the following equation:
+
+$$
+\mathcal {L} _ {\text {c o n t}} = - \frac {1}{| \mathcal {D} |} \sum_ {i = 1} ^ {N} \mathcal {L} _ {\text {c o n t}} ^ {i} \tag {5}
+$$
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = (1 - \lambda) \mathcal {L} _ {c e} + \lambda \mathcal {L} _ {\text {c o n t}} \tag {6}
+$$
+
+In equation (5) and (6),  $\lambda$  is the balance parameter between the two losses.
+
+# 3.4 Augmentation Pipeline
+
+In utilizing  $\mathbf{M}_{\mathbb{G}}$  we generate a single data for each  $(\mathrm{inp}^i,l^i)$  in  $\mathcal{D}$ . The pipeline is as follows:
+
+1. Extract morpheme set  $m^i$  of  $h^i$  in  $\mathsf{inp}^i$  
+2. Generate  $h_{aug}^{i} = \mathbf{M}_{\mathbb{G}}([m^{i},p^{i},v^{i}])$ , where  $v^{i}$  is the label word of  $l^i$ .  
+3.  $h_{\text{aug}}^i$  is regarded as an augmented data for  $\mathcal{D}$ , which label is  $l^i$ .
+
+Considering that excessive augmentation may lead to error accumulation and the following label confusion, we apply CHEF to the small fraction of  $\mathcal{D}$  in implementing full-shot learning.
+
+# 4 Experimental Settings
+
+We introduce experimental settings used for the experiments. More details are in Appendix A
+
+# 4.1 Datasets
+
+We adopt Korean multiple classification benchmark datasets. Each dataset is used for training and evaluation according to the proposed task format.
+
+KLUE-NLI The KLUE-NLI dataset (Park et al., 2021b) has explicitly been curated for the natural language inference (NLI) task (Bowman et al., 2015). The training, validation, and test data comprises 24,998, 3,000, and 3,000 sentence pairs. In this task, models are required to process pairs of sentences, referred to as the premise and hypothesis, and deduce the underlying relationship, which could be entailment, contradiction, or neutral.
+
+KorNLI The KorNLI dataset (Ham et al., 2020) is also designed for Korean natural language inference. It is generated by translating the English Standard NLI (SNLI) and Multi-Genre NLI (MNLI) datasets, as well as the Cross-lingual NLI (XNLI) dataset into Korean. The training data of the KorNLI dataset consists of 942,854 sentence pairs, machine-translated from the SNLI and MNLI datasets, while the evaluation data comprise 7,500 translated sentence pairs from the XNLI dataset.
+
+KLUE-STS The KLUE-STS dataset (Park et al., 2021b) has been meticulously assembled for the semantic textual similarity (STS) task, encompassing 11,668 sentence pairs for training, 519 for validation, and 1,037 for testing. In this task, models assess pairs of sentences and determine their degree of semantic similarity.
+
+KLUE-YNAT KLUE-YNAT dataset (Park et al., 2021b) has been designed for the topic classification task. The dataset includes training, validation, and test data composed of 45,678, 9,107, and 9,107 samples. In this task, models are tasked with processing sentences and assigning them to predefined news categories based on the underlying topic.
+
+NSMC The NSMC dataset<sup>4</sup> has been constructed for the sentiment analysis task in Korean. It is derived from movie reviews and their respective ratings from the Naver platform. The training data of the NSMC dataset comprises 150,000 reviews, and the test set consists of 50,000 reviews. In this task, models are required to process individual sentences and determine their underlying sentiment, which can be positive or negative.
+
+# 4.2 Models
+
+The morpheme blender employs an encoder-decoder structure and uses KoBART (Lewis et al., 2020), a pre-trained generative language model for
+
+<table><tr><td>Model</td><td>KLUE-NLI (Acc.)</td><td>KorNLI-MNLI (Acc.)</td><td>KorNLI-SNLI (Acc.)</td><td>KLUE-STS (Pearsonr)</td><td>KLUE-YNAT (F1)</td><td>NSMC (Acc.)</td></tr><tr><td>BERTKBase</td><td>81.53</td><td>80.63</td><td>71.56</td><td>90.85</td><td>85.73</td><td>90.30</td></tr><tr><td>+KoEDA</td><td>81.50</td><td>80.12</td><td>69.78</td><td>90.80</td><td>86.01</td><td>90.42</td></tr><tr><td>+AEDA</td><td>81.47</td><td>80.76</td><td>69.52</td><td>90.86</td><td>86.35</td><td>90.35</td></tr><tr><td>+BT</td><td>81.30</td><td>80.08</td><td>70.71</td><td>90.65</td><td>85.15</td><td>90.35</td></tr><tr><td>+Mixup</td><td>82.10</td><td>80.14</td><td>71.74</td><td>90.97</td><td>86.28</td><td>90.40</td></tr><tr><td>+GPT-3.5+CA</td><td>78.90</td><td>79.87</td><td>68.52</td><td>89.49</td><td>84.36</td><td>89.98</td></tr><tr><td>+GPT-3.5+PARA</td><td>80.40</td><td>80.00</td><td>69.42</td><td>91.24</td><td>86.32</td><td>90.45</td></tr><tr><td>+GPT-4+CA</td><td>80.73</td><td>80.60</td><td>65.26</td><td>90.92</td><td>85.50</td><td>90.45</td></tr><tr><td>+GPT-4+PARA</td><td>80.90</td><td>80.18</td><td>68.59</td><td>91.44</td><td>86.14</td><td>90.43</td></tr><tr><td>+CHEF</td><td>82.63</td><td>81.07</td><td>72.06</td><td>91.26</td><td>86.68</td><td>90.52</td></tr><tr><td>RoBERTaKBase</td><td>84.83</td><td>80.84</td><td>71.84</td><td>92.50</td><td>85.07</td><td>90.08</td></tr><tr><td>+KoEDA</td><td>85.23</td><td>81.27</td><td>70.49</td><td>92.94</td><td>85.18</td><td>90.71</td></tr><tr><td>+AEDA</td><td>85.57</td><td>80.56</td><td>71.70</td><td>92.80</td><td>85.19</td><td>90.63</td></tr><tr><td>+BT</td><td>84.53</td><td>80.78</td><td>69.78</td><td>92.68</td><td>84.47</td><td>90.65</td></tr><tr><td>+Mixup</td><td>85.83</td><td>80.60</td><td>70.85</td><td>92.77</td><td>85.55</td><td>90.00</td></tr><tr><td>+GPT-3.5+CA</td><td>84.90</td><td>81.11</td><td>69.72</td><td>91.16</td><td>83.93</td><td>90.46</td></tr><tr><td>+GPT-3.5+PARA</td><td>84.60</td><td>80.58</td><td>70.12</td><td>92.95</td><td>85.11</td><td>90.87</td></tr><tr><td>+GPT-4+CA</td><td>84.83</td><td>81.85</td><td>62.13</td><td>92.91</td><td>84.73</td><td>90.83</td></tr><tr><td>+GPT-4+PARA</td><td>85.03</td><td>81.23</td><td>70.08</td><td>92.99</td><td>85.47</td><td>90.94</td></tr><tr><td>+CHEF</td><td>86.00</td><td>81.77</td><td>72.22</td><td>93.05</td><td>86.18</td><td>90.74</td></tr><tr><td>RoBERTaLARGE</td><td>89.17</td><td>81.73</td><td>73.26</td><td>93.35</td><td>85.69</td><td>91.28</td></tr><tr><td>+KoEDA</td><td>89.63</td><td>82.24</td><td>72.52</td><td>93.24</td><td>85.32</td><td>91.12</td></tr><tr><td>+AEDA</td><td>89.17</td><td>81.45</td><td>72.26</td><td>93.47</td><td>85.56</td><td>91.28</td></tr><tr><td>+BT</td><td>33.33</td><td>80.36</td><td>72.64</td><td>92.89</td><td>85.24</td><td>91.05</td></tr><tr><td>+Mixup</td><td>90.83</td><td>81.95</td><td>72.39</td><td>93.43</td><td>85.99</td><td>90.33</td></tr><tr><td>+GPT-3.5+CA</td><td>88.03</td><td>82.96</td><td>33.33</td><td>90.73</td><td>84.23</td><td>33.72</td></tr><tr><td>+GPT-3.5+PARA</td><td>88.33</td><td>83.12</td><td>71.86</td><td>93.28</td><td>85.88</td><td>91.31</td></tr><tr><td>+GPT-4+CA</td><td>89.10</td><td>83.48</td><td>68.94</td><td>93.20</td><td>85.46</td><td>67.97</td></tr><tr><td>+GPT-4+PARA</td><td>90.13</td><td>82.88</td><td>72.87</td><td>93.51</td><td>86.19</td><td>91.21</td></tr><tr><td>+CHEF</td><td>90.47</td><td>82.76</td><td>73.56</td><td>93.61</td><td>86.12</td><td>91.38</td></tr></table>
+
+Table 1: Full-shot evaluation results on multiple classification datasets. The best performance in each dataset is bolded. Underlines indicate that CHEF has outperformed the baselines. Scores in red denote a training failure due to poor quality. Accuracy (Acc.), F1-score (F1), and Pearson correlation (Pearsonr) are used as evaluation metrics.
+
+Korean, as its backbone. We choose KoBART because it exhibits acceptable Korean text generation abilities even in its small model parameters (124M) (Seo et al., 2022). KoBART takes an input sequence consisting of a lexical morpheme set, a label word, and an optional premise to generate synthetic data. The label discriminator features an encoder architecture and utilizes KLUE-BERT-base (Devlin et al., 2018), a pre-trained language model for Korean, as its backbone. KLUE-BERT-base conveys the original sentence's label to KoBART through a contrastive loss. We opt for the KLUE-BERT-base as it is the most compact model among the evaluated alternatives, mitigating the possible distillation effects that could emerge during the discrimination process due to model sizes (Park et al., 2021b). To validate the efficacy of CHEF, we select the following Korean language understanding models:  $\mathrm{BERT^{K}_{Base}}$  (KLUE-BERT-base), RoBERTa $^{\mathrm{K}}_{\mathrm{Base}}$  (KLUE-RoBERTa-base), and RoBERTa $^{\mathrm{K}}_{\mathrm{Large}}$  (KLUE-RoBERTa-large).
+
+# 4.3 Compared Methods
+
+The experiments are conducted based on full- and few-shot learning. We conduct comparative ex
+
+periments using BackTranslation (BT) (Sennrich et al., 2016), Korean-EDA (KoEDA) $^5$  (Wei and Zou, 2019), AEDA (Karimi et al., 2021), and Mixup (Zhang et al., 2018; Sun et al., 2020). We also employ GPT-3.5 (Ouyang et al., 2022; Brown et al., 2020) and GPT-4 (OpenAI, 2023) as backbone augmentation models. Our LLM-based approaches included semantic-based paraphrasing (PARA) (Fadaee et al., 2017; Kobayashi, 2018) and counterfactual augmentation (CA) (Liu et al., 2021; Ou et al., 2022).
+
+# 5 Experimental Results
+
+In this section, we evaluate a set of multiple classification datasets with our proposed method.
+
+# 5.1 Full-shot Learning
+
+As shown in Table 1, we compare the effectiveness and extent of performance improvement of CHEF against other data augmentation methods in the full-shot settings. We proportionally augmented the data with respect to the size of the training dataset, ensuring that an amount corresponding to  $1\%$  of the training data is added to each label.
+
+![](images/e41b4df5ed1de15663a0aa61641cac890b109a7da9dd8e89eb2acbfa7b0a77fa.jpg)  
+(a) KLUE-NLI
+
+![](images/b11d942aab19fd08e7e3629aaa06ff277810806e7243d3729885ba4d05398948.jpg)  
+(b) KorNLI-MNLI
+
+![](images/d0d0cebfca8a8a3761038d16eb4ff1e4235b6ae2056f723f5777e20886282be4.jpg)  
+(c) KorNLI-SNLI  
+Figure 3: Performance of the language models trained on multiple classification datasets by adjusting the augmentation ratio of CHEF. The first column of each plot is the baseline performance without data augmentation.
+
+![](images/f83f8c7e82dd8b6a0544cb560cb3fd0c3d8103ab6a9d6fab026002e282a3f107.jpg)  
+(d) KLUE-STS
+
+![](images/1291caf2ff380400209ca42c20a402001dd781be244d289e75124aa3807cce7c.jpg)  
+(e) KLUE-YNAT
+
+![](images/c9938274c1fc8a3f218590327533102818c4f4cb2043a4d99f6225682a53d6bf.jpg)  
+(f) NSMC
+
+Augmentation methods employing LLMs exhibit the generation of sentences with high diversity and superior qualitative properties. However, they occasionally demonstrate decreases in performance of up to  $2\%$ . These declines can be attributed to the significant alteration of the overall data distribution, which makes it challenging to fully account for the underlying labeling scheme intrinsic to the task. In the case of counterfactual augmentation, the instability in label transformations is amplified, even when provided with chain-of-thought prompts (Wei et al., 2022). Original sentences and labels undergo counterfactual changes without any review for conformance, resulting in increased variability that hinders the effectiveness of the augmentation process.
+
+BT cannot guarantee the preservation of the same labels, resulting in performance decrease cases. AEDA does not show the same level of improvements for all tasks as reported in (Karimi et al., 2021) when applied to pre-trained Korean language models. Furthermore, Mixup exhibits notable performance improvements in KLUE-NLI; however, these enhancements do not consistently ensure improved results across other tasks. KoEDA utilizes a morphological analyzer to segment the data and leverages a thesaurus for Korean. This approach shows modest performance improvements, validating the effectiveness of morpheme-based data augmentation in line with the characteristics of the Korean language. However, its performance enhancements are limited due to challenges in ensuring label consistency and the absence of contextual understanding. We observe that CHEF exhibits the most significant performance improvements. By leveraging a comprehensive understanding of derivation and inflectional properties in conjunction with functional morphemes, CHEF effectively enhances models across various scenarios. As a result, CHEF overcomes these challenges and demon
+
+strates the efficacy of data augmentation for full-shot learning.
+
+# 5.2 Changes in Augmentation Size
+
+Figure 3 presents the performance changes according to each dataset's augmentation ratio of CHEF. As the amount of synthetic data increases, the probability of the involvement of data that renders negative noise accumulations. In the full-shot setting, simply expanding the number of synthetic data samples does not guarantee a performance improvement. However, CHEF maintains the effectiveness of data augmentation and shows superior enhancements compared to the baseline in most cases where data augmentation is limited to within  $10\%$  (More details in Appendix 7).
+
+# 5.3 Few-shot and Synthetic-only
+
+We further probe the efficacy of CHEF in the NLI task, which is relatively capricious data augmentation effects in Section §5.2. We conduct comparative experiments in a few-shot setting with 32 samples and add 32 augmented sentences for each label. Table 2 shows the results averaged and maximum value over three different seeds. The data augmentation methods using the LLMs are more effective in low-resource regimes than in full-shot settings. CHEF significantly boosts the model's performance, even in situations with limited data. Furthermore, we evaluate the effectiveness of solely using the synthetic data generated by CHEF (i.e., CHEF-SynOnly) to train the models. The outcomes reveal a spectrum of enhancements in performance and demonstrate the quality of the CHEF augmentation.
+
+# 5.4 Ablation Study
+
+To precisely evaluate the importance of labeling consistency, we conduct ablation studies by systematically altering the label discriminator com
+
+<table><tr><td>Model</td><td>KLUE-NLI
+(Acc.)</td><td>KorNLI-MNLI
+(Acc.)</td><td>KorNLI-SNLI
+(Acc.)</td></tr><tr><td>BERTKBase</td><td>37.6/40.1</td><td>40.5/41.7</td><td>37.5/40.0</td></tr><tr><td>+KoEDA</td><td>43.0/45.0</td><td>43.5/44.9</td><td>36.8/38.6</td></tr><tr><td>+AEDA</td><td>43.1/45.5</td><td>42.1/43.9</td><td>38.0/39.8</td></tr><tr><td>+BT</td><td>41.1/43.0</td><td>42.9/44.8</td><td>38.0/40.2</td></tr><tr><td>+Mixup</td><td>40.6/41.5</td><td>42.8/44.2</td><td>38.9/39.2</td></tr><tr><td>+GPT-3.5+CA</td><td>37.1/37.7</td><td>41.6/44.6</td><td>38.2/39.5</td></tr><tr><td>+GPT-3.5+PARA</td><td>41.5/43.4</td><td>42.9/44.3</td><td>38.7/39.7</td></tr><tr><td>+GPT-4+CA</td><td>41.7/43.0</td><td>46.3/48.5</td><td>40.5/41.8</td></tr><tr><td>+GPT-4+PARA</td><td>44.7/46.1</td><td>43.5/45.7</td><td>37.7/40.7</td></tr><tr><td>+CHEF</td><td>46.5/47.6</td><td>47.7/49.2</td><td>41.1/42.3</td></tr><tr><td>+CHEF-SynOnly</td><td>41.7/42.5</td><td>46.8/47.4</td><td>35.1/37.1</td></tr><tr><td>RoBERTaKBase</td><td>37.6/40.1</td><td>36.5/36.7</td><td>36.1/36.7</td></tr><tr><td>+KoEDA</td><td>41.2/44.4</td><td>38.4/39.9</td><td>36.3/36.5</td></tr><tr><td>+AEDA</td><td>42.4/48.0</td><td>39.5/42.9</td><td>36.8/37.2</td></tr><tr><td>+BT</td><td>39.9/42.2</td><td>38.9/41.3</td><td>36.2/37.1</td></tr><tr><td>+Mixup</td><td>39.4/41.4</td><td>38.4/39.5</td><td>36.0/36.4</td></tr><tr><td>+GPT-3.5+CA</td><td>36.1/36.9</td><td>39.1/42.6</td><td>36.1/37.1</td></tr><tr><td>+GPT-3.5+PARA</td><td>39.8/41.8</td><td>39.4/40.7</td><td>36.1/36.7</td></tr><tr><td>+GPT-4+CA</td><td>40.3/42.3</td><td>41.0/43.3</td><td>39.6/41.4</td></tr><tr><td>+GPT-4+PARA</td><td>43.3/45.8</td><td>39.4/41.7</td><td>39.4/41.6</td></tr><tr><td>+CHEF</td><td>41.2/43.3</td><td>43.9/45.8</td><td>40.1/44.7</td></tr><tr><td>+CHEF-SynOnly</td><td>38.7/39.5</td><td>43.0/45.1</td><td>33.9/34.8</td></tr><tr><td>RoBERTaKLarge</td><td>47.7/51.7</td><td>40.8/44.4</td><td>37.5/39.4</td></tr><tr><td>+KoEDA</td><td>51.2/52.6</td><td>43.6/47.0</td><td>38.9/40.9</td></tr><tr><td>+AEDA</td><td>54.6/56.1</td><td>45.0/48.6</td><td>38.7/40.3</td></tr><tr><td>+BT</td><td>54.4/56.3</td><td>43.5/48.7</td><td>38.0/38.8</td></tr><tr><td>+Mixup</td><td>57.1/57.7</td><td>42.2/44.5</td><td>40.2/40.5</td></tr><tr><td>+GPT-3.5+CA</td><td>41.4/43.9</td><td>41.7/45.9</td><td>38.5/40.6</td></tr><tr><td>+GPT-3.5+PARA</td><td>52.6/53.7</td><td>43.8/47.1</td><td>38.5/39.8</td></tr><tr><td>+GPT-4+CA</td><td>45.6/51.0</td><td>46.1/48.8</td><td>41.4/44.3</td></tr><tr><td>+GPT-4+PARA</td><td>56.9/59.5</td><td>45.3/48.2</td><td>38.6/41.2</td></tr><tr><td>+CHEF</td><td>56.5/57.2</td><td>51.5/54.5</td><td>44.9/47.6</td></tr><tr><td>+CHEF-SynOnly</td><td>50.0/57.1</td><td>49.8/54.5</td><td>33.8/34.2</td></tr></table>
+
+ponent within the CHEF. The variations in performance resulting from the exclusion or absence of pre-training in the discriminator are presented in Table 3. Upon removal of the label discriminator, we observe a decline in the augmentation effectiveness in the full-shot setting. Notably, even without the pre-training phase of the discriminator, CHEF maintains its capacity for enhancing performance. By integrating the morpheme blender and the pretrained label discriminator using contrastive loss, CHEF achieves the highest level of performance. These empirical findings provide compelling evidence for the effectiveness of incorporating the discriminator to ensure label consistency.
+
+# 5.5 Cross-domain Analysis
+
+We conduct cross-domain experiments in the NLI task. As described in §4.1, KLUE-NLI and KormNLI/SNLI are datasets derived from different sources. To assess performance, we employ cross
+
+Table 2: Few-shot evaluation results. The scores on the left side of the  $\prime$  represent the averages, and those on the right indicate the maximums. The best performance in each dataset is bolded. Underlines indicate that CHEF has outperformed the baselines.  
+
+<table><tr><td>Model</td><td>KLUE-NLI (Acc.)</td><td>KLUE-STS (Pearsonr)</td><td>KLUE-YNAT (F1)</td></tr><tr><td>BERTKBase</td><td>81.53</td><td>90.85</td><td>85.73</td></tr><tr><td>+MG</td><td>81.83</td><td>90.62</td><td>86.02</td></tr><tr><td>+MG+MD</td><td>82.54</td><td>91.05</td><td>86.34</td></tr><tr><td>+MG+MD+PT</td><td>82.63</td><td>91.26</td><td>86.68</td></tr></table>
+
+Table 3: Results of the ablation study on the CHEF.  $\mathbf{M}_{\mathrm{G}}$  and  $\mathbf{M}_{\mathrm{D}}$  respectively denote the morpheme blender and discriminator. PT refers to the pre-training of  $\mathbf{M}_{\mathrm{D}}$ . The values in bold represent the highest performance.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">TRAIN → TEST</td></tr><tr><td>KLUE-NLI → Kor-NLI</td><td>Kor-NLI → KLUE-NLI</td></tr><tr><td>BERTKBase(Full)</td><td>61.14</td><td>71.37</td></tr><tr><td>+CHEF</td><td>60.09</td><td>72.43</td></tr><tr><td>BERTKBase(Few)</td><td>41.01</td><td>36.06</td></tr><tr><td>-CROSS†</td><td>39.04</td><td>38.93</td></tr><tr><td>+CHEF</td><td>49.66</td><td>40.93</td></tr></table>
+
+Table 4: Results of cross-domain analysis.  $\rightarrow$  indicates that the training and test data belong to different domains. -CROSS† denotes the results without cross-domain settings. The best performance is bolded.
+
+evaluation of the model trained on data augmented using CHEF. Table 4 presents the observed performance improvements in three out of four cases, compared to the baselines without augmentation. The cross-domain effectiveness of CHEF is more pronounced in the few-shot setting. Furthermore, few-shot CHEF outperforms models trained and evaluated solely on a single domain. These results alleviate concerns regarding the degradation of the model's generalization ability due to the augmentation effects, reaffirming its robust performance across different domains.
+
+# 5.6 Larger and Multilingual Blender
+
+To comprehend the implications of employing a larger generative model for CHEF, we evaluate the performance leveraging the mBART-large (Liu et al., 2020). To avoid unintended knowledge distillation effects resulting from using larger discriminator modules, we maintain the same discriminator. As presented in Table 5, the results present that using a larger generative model as the blender does not necessarily ensure higher performance, but comparable effects can be achieved as well. We further find that employing a multilingual model as the blender yields similar augmentation effects.
+
+# 5.7 Morpheme Ingredients Variants
+
+Table 6 shows the CHEF's performance employing different morpheme ingredients. We modify
+
+<table><tr><td>Model</td><td>KLUE-NLI (Acc.)</td><td>KLUE-STS (Pearsonr)</td><td>KLUE-YNAT (F1)</td></tr><tr><td>BERTKBase</td><td>81.53</td><td>90.85</td><td>85.73</td></tr><tr><td>+CHEF(BART)</td><td>82.63</td><td>91.26</td><td>86.68</td></tr><tr><td>+CHEF(mBART)</td><td>82.00</td><td>91.26</td><td>87.11</td></tr></table>
+
+Table 5: Evaluation results on larger morpheme blender. The best performance in each dataset is bolded.  
+
+<table><tr><td>Model</td><td>KLUE-NLI</td></tr><tr><td>BERTKBase</td><td>81.53</td></tr><tr><td>+ CHEFSynonym</td><td>80.22</td></tr><tr><td>+ CHEFAntonym</td><td>79.84</td></tr></table>
+
+Table 6: Evaluation results on morpheme ingredients variants.  $\mathrm{CHEF_{Synonym}}$  and  $\mathrm{CHEF_{Antonym}}$  denote the methods of replacing morphemes and labels.
+
+the lexical morphemes by substituting them with synonyms or altering them to antonyms considering labels. Replacing lexical morphemes based on antonyms or synonyms is not guaranteed to preserve the compositionality across other lexical morphemes. The involvement of non-contextualized morphemes leads to diminished generative capabilities and the generation of low-quality synthetic data, resulting in a decline in performance.
+
+# 6 Conclusion
+
+We introduce CHEF, a novel data augmentation method designed for the Korean language, which is inherently agglutinative and rich in morphological variations. CHEF leverages the combinatory properties of lexical and functional morphemes in Korean to construct linguistically diverse and label-consistent synthetic data. By incorporating a morpheme blender and a label discriminator module, CHEF ensures that the generated synthetic data preserves the label information of the original dataset while introducing new linguistic forms through morphological alterations. Our experiments demonstrate the effectiveness and robustness of CHEF across various scenarios. In future work, we plan to explore the adaptability of the CHEF architecture to other morphologically rich languages and further optimize the interaction between the morpheme blender and label discriminator.
+
+# Limitations
+
+This study proposes an effective augmentation method suitable for Korean datasets, considering the unique linguistic characteristics of the Korean
+
+language. However, the method of this study primarily focuses on Korean and does not sufficiently consider other languages. This area can be further explored and improved in future research. Also, due to the performance limitations of the off-the-shelf pre-trained generative model, unnecessary word duplication occurs in sentences augmented using CHEF. This issue can be addressed by introducing a more optimized decoding strategy or employing a more advanced generative model, expected to produce higher-quality results. Moreover, The mecab-ko analyzer, which we used for constructing the set of morphemes, can have an error rate depending on the domain and data it is applied to. The proportion of morpheme combination rules in CHEF may differ based on which morpheme analyzer is used to construct the morpheme set. This presents a potential risk: if a less proficient morpheme analyzer is used, it may fail to recover the original morphemes accurately and be susceptible to errors when processing data from untrained domains. Therefore, leveraging a more advanced morpheme analyzer could enable data augmentation that more accurately reflects the linguistic characteristics of the Korean language.
+
+# Ethics Statement
+
+We employed Korean multiple classification benchmark datasets in our experiments. Data augmentation was conducted by altering the morphological composition of sentences present in the training datasets. Excluding the NSMC dataset, each benchmark dataset has been officially released and has undergone validation to ensure ethical considerations using human annotators. The NSMC dataset may contain some unethical expressions among negative reviews. Furthermore, we acknowledge that the pre-trained language models (PLMs), used as the backbone for the morpheme blender, could have been exposed to toxic data during the pretraining process, thereby possessing the potential to generate inappropriate synthetic data.
+
+# Acknowledgments
+
+This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2023-2018-0-01405) supervised by the IITP (Institute for Information & Communications Technology Planning & Evaluation). This work was supported by Institute of Information &
+
+communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2020-0-00368, A Neural-Symbolic Model for Knowledge Acquisition and Inference Techniques). This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program (IITP-2023-2020-0-01819) supervised by the IITP(Institute for Information & communications Technology Planning & Evaluation).
+
+# References
+
+Daniel M Bikel. 2000. Automatic wordnet mapping using word sense disambiguation. In 2000 Joint SIG-DAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, pages 142-147.  
+Samuel Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632-642.  
+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.  
+Xiang Dai and Heike Adel. 2020. An analysis of simple data augmentation for named entity recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3861-3867, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Guillaume Daval-Frerot and Yannick Weis. 2020. WMD at SemEval-2020 tasks 7 and 11: Assessing humor and propaganda using unsupervised data augmentation. In Proceedings of the Fourteenth Workshop on Semantic Evaluation, pages 1865-1874, Barcelona (online). International Committee for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Jingfei Du, Edouard Grave, Belize Gunel, Vishrav Chaudhary, Onur Celebi, Michael Auli, Veselin Stoyanov, and Alexis Conneau. 2021. Self-training improves pre-training for natural language understanding. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5408-5418, Online. Association for Computational Linguistics.
+
+Alexander Fabbri, Simeng Han, Haoyuan Li, Haoran Li, Marjan Ghazvininejad, Shafiq Joty, Dragomir Radev, and Yashar Mehdad. 2021. Improving zero and few-shot abstractive summarization with intermediate fine-tuning and data augmentation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 704-717, Online. Association for Computational Linguistics.  
+Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2017. Data augmentation for low-resource neural machine translation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 567-573.  
+Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, et al. 2021. Beyond english-centric multilingual machine translation. The Journal of Machine Learning Research, 22(1):4839-4886.  
+Hongyu Guo. 2020. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4044-4051.  
+Jiyeon Ham, Yo Joong Choe, Kyubyong Park, Ilji Choi, and Hyungjoon Soh. 2020. Kornli and korsts: New benchmark datasets for korean natural language understanding. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 422-430.  
+Nitish Joshi and He He. 2022. An investigation of the (in) effectiveness of counterfactually augmented data. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3668-3681.  
+Akbar Karimi, Leonardo Rossi, and Andrea Prati. 2021. Aeda: An easier data augmentation technique for text classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 2748-2754.  
+Myung Hee Kim and Nathalie Colineau. 2020. An enhanced mapping scheme of the universal part-of-speech for korean. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 3826-3833.  
+Nayeon Kim, Jun-Hyung Park, Joon-Young Choi, Eojin Jeon, Youjin Kang, and SangKeun Lee. 2022. Break it down into bts: Basic, tiniest subword units for korean. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7007-7024.  
+Sosuke Kobayashi. 2018. Contextual augmentation: Data augmentation by words with paradigmatic relations. In Proceedings of NAACL-HLT, pages 452-457.
+
+Fanshuang Kong, Richong Zhang, Xiaohui Guo, Samuel Mensah, and Yongyi Mao. 2022. Dropmix: A textual data augmentation combining dropout with mixup. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 890-899.  
+Varun Kumar, Ashutosh Choudhary, and Eunah Cho. 2020. Data augmentation using pre-trained transformer models. In Proceedings of the 2nd Workshop on Life-long Learning for Spoken Language Systems, pages 18-26.  
+Young-Jun Lee, Chae-Gyun Lim, and Ho-Jin Choi. 2020. Korean-specific emotion annotation procedure using n-gram-based distant supervision and koreanspecific-feature-based distant supervision. In Proceedings of the Twelfth Language Resources and Evaluation Conference, pages 1603-1610.  
+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.  
+Bill Yuchen Lin, Wangchunshu Zhou, Ming Shen, Pei Zhou, Chandra Bhagavatula, Yejin Choi, and Xiang Ren. 2020. Commongen: A constrained text generation challenge for generative commonsense reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1823-1840.  
+Qi Liu, Matt Kusner, and Phil Blunsom. 2021. Counterfactual data augmentation for neural machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 187-197.  
+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.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+David Lowell, Brian Howard, Zachary C. Lipton, and Byron Wallace. 2021. Unsupervised data augmentation with naive augmentation and without unlabeled data. In Proceedings of the 2021 Conference on
+
+Empirical Methods in Natural Language Processing, pages 4992-5001, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Andrew Matteson, Chanhee Lee, Youngbum Kim, and Heuseok Lim. 2018. Rich character-level information for Korean morphological analysis and part-of-speech tagging. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2482-2492, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Dheeraj Mekala, Tu Vu, Timo Schick, and Jingbo Shang. 2022. Leveraging qa datasets to improve generative data augmentation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9737-9750.  
+OpenAI. 2023. Gpt-4 technical report.  
+Jiao Ou, Jinchao Zhang, Yang Feng, and Jie Zhou. 2022. Counterfactual data augmentation via perspective transition for open-domain dialogues. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1635-1648, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+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.  
+Chanjun Park, Jaehyung Seo, Seolhwa Lee, Chanhee Lee, Hyeonseok Moon, Sugyeong Eo, and Heuiseok Lim. 2021a. BTS: Back TranScription for speech-to-text post-processor using text-to-speech-to-text. In Proceedings of the 8th Workshop on Asian Translation (WAT2021), pages 106–116, Online. Association for Computational Linguistics.  
+Sungjoon Park, Jeongmin Byun, Sion Baek, Yongseok Cho, and Alice Oh. 2018. Subword-level word vector representations for korean. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2429-2438.  
+Sungjoon Park, Jihyung Moon, Sungdong Kim, Won Ik Cho, Jiyoon Han, Jangwon Park, Chisung Song, Jun-seong Kim, Yongsook Song, Taehwan Oh, et al. 2021b. Klue: Korean language understanding evaluation. arXiv preprint arXiv:2105.09680.  
+Timo Schick and Hinrich Schütze. 2021. Generating datasets with pretrained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6943-6951.  
+Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th
+
+Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86-96.  
+Jaehyung Seo, Seounghoon Lee, Chanjun Park, Yoonna Jang, Hyeonseok Moon, Sugyeong Eo, Seonmin Koo, and Heui-Seok Lim. 2022. A dog is passing over the jet? a text-generation dataset for korean commonsense reasoning and evaluation. In *Findings of the Association for Computational Linguistics: NAACL* 2022, pages 2233-2249.  
+Hyun-Je Song and Seong-Bae Park. 2019. Korean morphological analysis with tied sequence-to-sequence multi-task model. 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 1436-1441.  
+Jae Jung Song. 2006. The Korean language: Structure, use and context. Routledge.  
+Lichao Sun, Congying Xia, Wenpeng Yin, Tingting Liang, S Yu Philip, and Lifang He. 2020. Mixup-transformer: Dynamic data augmentation for nlp tasks. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3436-3440.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed H Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems.  
+Jason Wei and Kai Zou. 2019. Eda: Easy data augmentation techniques for boosting performance on text classification tasks. 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 6382-6388.  
+Jiaxin Wen, Yeshuang Zhu, Jinchao Zhang, Jie Zhou, and Minlie Huang. 2022. AutoCAD: Automatically generate counterfactuals for mitigating shortcut learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 2302-2317, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Kang Min Yoo, Dongju Park, Jaewook Kang, Sang-Woo Lee, and Woomyoung Park. 2021. GPT3Mix: Leveraging large-scale language models for text augmentation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2225-2239, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Soyoung Yoon, Gyuwan Kim, and Kyumin Park. 2021. Ssmix: Saliency-based span mixup for text classification. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3225-3234.
+
+Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc.  
+Jing Zhou, Yanan Zheng, Jie Tang, Li Jian, and Zhilin Yang. 2022. Flipda: Effective and robust data augmentation for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8646-8665.
+
+# A Experimental Details
+
+We trained our models on a single NVIDIA A6000 GPU (48GB) and AMD EPYC 7513 32-Core Processor CPUs.
+
+CHEF Modules As a discriminator, CHEF uses KLUE-BERT-base $^6$  (Devlin et al., 2018; Park et al., 2021a), which has 768 embedding sizes, 768 hidden sizes, 12 layers, and 12 attention heads. As a morpheme blender, CHEF employs KoBART-base $^7$  (Lewis et al., 2020), where each encoder and decoder have 768 hidden sizes, 6 layers, and 16 attention heads, respectively.
+
+Korean Language Understanding Models We used three pre-trained language models for Korean. KLUE-BERT-base (Devlin et al., 2018; Park et al., 2021a) has 768 embedding sizes, 768 hidden sizes, 12 layers, and 12 attention heads. KLUE-RoBERTa-base (Liu et al., 2019; Park et al., 2021a) also has 768 embedding sizes, 768 hidden sizes, 12 layers, and 12 attention heads. KLUE-RoBERTa-large (Liu et al., 2019; Park et al., 2021a) has 1024 embedding sizes, 1024 hidden sizes, 24 layers, and 16 attention heads.
+
+EDA & BackTranslation We applied Back-Translation (BT) (Sennrich et al., 2016) and easy data augmentation (EDA) (Wei and Zou, 2019) to the data to generate synthetic data. For BT, we used M2M100 (Fan et al., 2021) to translate Korean text into Japanese and back into Korean. The model we used for BT is facebook-M2M100 (418M), which has a 1024 embedding size, 12 layers, and 16 attention heads. We used KoEDA  $^{8}$ , a library that utilizes the Korean WordNet (Bikel, 2000) for EDA.
+
+AEDA & Mixup We used data augmentation with the default settings presented in the AEDA paper (Karimi et al., 2021) $^9$ . Mixup (Zhang et al., 2018; Sun et al., 2020), which is suitable for Transformer-based models, was implemented by randomly shuffling sample indices within a batch. Subsequently, the hidden state of each sample and the label scalar were mixed at a ratio of 0.2, following the default  $\lambda$  value, to create synthetic data.
+
+Large Language Models We included GPT-3.5 (gpt-3.5-turbo-0301) (Brown et al., 2020; Ouyang et al., 2022) and GPT-4 (gpt-4-0314) (OpenAI, 2023) as the large language models. We applied augmentation to the given training data using the OpenAI API<sup>10</sup> and LangChain<sup>11</sup>. As depicted in Figure 4 and 5, the prompt consists of an example in a one-shot template for the given task and instructions for the augmentation method. The cost incurred due to the OpenAI API calls amounted to $245.17, and the data augmentation approach utilizing LLMs did not show a significant improvement in model performance relative to the cost incurred.
+
+Hyperparameters Hyperparameters to train the CHEF are batch size 8, learning rate  $1 \times 10^{-5}$ , AdamW optimizer (Loshchilov and Hutter, 2019)  $(\beta_{1} = 0.9, \beta_{2} = 0.999, \epsilon = 1e - 8)$ , lambda 0.2, max source length 200, max target length 168, and 5 epochs. To train Korean language understanding models is batch size 32, learning rate  $2 \times 10^{-5}$ , AdamW optimizer  $(\beta_{1} = 0.9, \beta_{2} = 0.999, \epsilon = 1e - 8)$ , and 20 epochs. In the case of few-shot learning, all other hyperparameter settings remained unchanged except for the epoch adjusted to 30, considering the model's overfitting point.
+
+Decoding Strategy Within the framework of CHEF augmentation, we imposed specific constraints on the decoding strategy. We established the following settings: a beam size of 10, a maximum sequence length of 168, a minimum sequence length of 5, a repetition penalty of 2, and a no-repeat n-gram size of 3 to penalize the generation of duplicate tokens.
+
+# B Qualitative Analysis
+
+We applied CHEF to several benchmark datasets and conducted qualitative analyses of the aug-
+
+mented output. CHEF augmentations typically take the following form of three types of variations.
+
+As described in Figure 6, we observe that the outputs generated by CHEF primarily involve modifications of particles and determiners that are closely associated with nouns. However, no significant changes are observed in the verbs of the sentences. The combination of lexical morphemes introduces slight variations in the meaning of the generated sentences but does not lead to changes in the labels assigned to them.
+
+As shown in Figure 7, the augmented output primarily consisted of conjugation variations, leading to a diverse range of sentences with different ending conjunction rules. However, there is minimal impact on the overall meaning of the sentences themselves. Notably, the augmented hypotheses are labeled as having significantly lower similarity to the corresponding premises regarding semantic similarity evaluation.
+
+As illustrated in Figure 8, the augmentation process involves single sentences resembling news headlines fitting into the "Life & Culture" section. Unlike the other benchmark datasets, KLUEYNAT is characterized by the fact that the original sentences contain relatively fewer functional morphemes. Based on this characteristic of the dataset, CHEF was applied to augment the data by changing the order or role of the lexical morphemes in the sentence.
+
+<table><tr><td>Model</td><td>KLUE-NLI</td><td>KorNLI-MNLI</td><td>KorNLI-SNLI</td><td>KLUE-STS</td><td>KLUE-YNAT</td><td>NSMC</td></tr><tr><td>BERTKBase</td><td>81.53</td><td>80.63</td><td>71.56</td><td>90.85</td><td>85.73</td><td>90.30</td></tr><tr><td>+CHEF 1%</td><td>82.63</td><td>81.07</td><td>72.06</td><td>91.26</td><td>86.68</td><td>90.52</td></tr><tr><td>+CHEF 2%</td><td>81.47</td><td>79.83</td><td>71.60</td><td>91.12</td><td>86.76</td><td>90.61</td></tr><tr><td>+CHEF 3%</td><td>81.57</td><td>79.85</td><td>70.61</td><td>91.36</td><td>86.63</td><td>90.39</td></tr><tr><td>+CHEF 5%</td><td>82.30</td><td>79.81</td><td>70.63</td><td>91.52</td><td>86.51</td><td>90.50</td></tr><tr><td>+CHEF 10%</td><td>81.43</td><td>79.85</td><td>70.33</td><td>91.05</td><td>86.90</td><td>90.41</td></tr><tr><td>RoBERTaKBase</td><td>84.83</td><td>80.84</td><td>71.84</td><td>92.50</td><td>85.07</td><td>90.08</td></tr><tr><td>+CHEF 1%</td><td>86.00</td><td>81.77</td><td>72.22</td><td>93.05</td><td>86.18</td><td>90.74</td></tr><tr><td>+CHEF 2%</td><td>85.47</td><td>81.63</td><td>71.50</td><td>93.16</td><td>86.28</td><td>90.70</td></tr><tr><td>+CHEF 3%</td><td>86.20</td><td>81.11</td><td>71.58</td><td>92.71</td><td>86.64</td><td>90.76</td></tr><tr><td>+CHEF 5%</td><td>85.77</td><td>80.70</td><td>71.19</td><td>92.35</td><td>86.23</td><td>90.80</td></tr><tr><td>+CHEF 10%</td><td>85.53</td><td>81.17</td><td>69.94</td><td>92.83</td><td>86.58</td><td>90.69</td></tr><tr><td>RoBERTaKLarge</td><td>89.17</td><td>81.73</td><td>73.26</td><td>93.35</td><td>85.69</td><td>91.28</td></tr><tr><td>+CHEF 1%</td><td>90.47</td><td>82.76</td><td>73.56</td><td>93.61</td><td>86.12</td><td>91.38</td></tr><tr><td>+CHEF 2%</td><td>90.30</td><td>82.98</td><td>73.40</td><td>93.21</td><td>86.63</td><td>91.35</td></tr><tr><td>+CHEF 3%</td><td>90.00</td><td>82.44</td><td>74.14</td><td>93.40</td><td>86.30</td><td>91.33</td></tr><tr><td>+CHEF 5%</td><td>90.07</td><td>83.08</td><td>73.17</td><td>93.29</td><td>87.02</td><td>91.28</td></tr><tr><td>+CHEF 10%</td><td>89.53</td><td>83.12</td><td>73.23</td><td>92.69</td><td>86.64</td><td>91.34</td></tr></table>
+
+Table 7: Performance of language models on multiple classification datasets by adjusting CHEF's augmentation rate. The best-performing method on each dataset is shown in **bold**. Underlines indicate that the method outperformed the baselines.
+
+![](images/2b8783de22b5d777e859c2845a263b037ff172210287a19020949fa8193756ff.jpg)  
+Figure 4: An example of a prompt template for augmenting data with given original data. (a) is a human instruction to paraphrase the hypothesis of the given data. (b) is an additional directive to limit the output to JSON format. (c) is an example of paraphrasing. (d) is the format of the given input sequence and the generated LLM output.
+
+(a)
+
+Considering the "premise" and "labels" of the given example, make the counterfactual hypothesis so that the data can be augmented.
+
+A SET OF "labels": ["entailment", "contradiction", "neutral"].
+
+You should randomly select one of the labels in A SET OF "labels" and change the "labels" in YOUR ANSWER to be different from the "labels" in the ORIGINAL DATA and paraphrase with an appropriate "hypothesis".
+
+(b)
+
+Use the Following JSON format!
+
+(c)
+
+ORIGINAL DATA: {"hypothesis": "groenstang is het huischijingschag en die hecht uusuusgeurogelreonded."  
+"premise": "huiuustiigchng is iierhan stgca gaa jiscnogacnncn hongjohcng zengrsga  
+guiinlde'soetdhgsmtoeiergetdgoanhaoqsgvngu  
+"labels": "entailment"}
+
+YOUR ANSWER: {{“hypothesis”：“상품진화성은 희성군의 죽과가부棄고 기재선 힍류시렸다.”,
+
+"premise": "동류진화종이 이려한 썼과가지수고能够在 펋일화성조성과군만들어 so목thy상에이어렸다고 기재고 썸성고을 춤우수상에로结局렸다.",
+
+"labels": "contradiction"}}
+
+(d)
+
+Begin!
+
+ORIGINAL DATA: {input sequence}
+
+YOUR ANSWER: {generated LLM output}
+
+![](images/dc7b341b475358ae04dadb04b5d80289c9c9e35325dbac0ff37f69fea62e7432.jpg)  
+Figure 5: An example of prompt template for augmenting data with given original data. (a) is a human instruction for generating a counterfactual sequence for a hypothesis on a given input and instructions on how to set the label of the counterfactual sequence. (b) is an additional directive to limit the output to JSON format. (c) is an example of counterfactual sequence generation. (d) is the format of the given input sequence and the generated LLM output.  
+Figure 6: An example of augmentation result from CHEF in KLUE-NLI
+
+![](images/84563db6130583dedcbe209b60c514a297ebaa5a98630b2393fbd27fd7137ee2.jpg)  
+Figure 7: An example of augmentation result from CHEF in KLUE-STS
+
+![](images/89bee0ee169333ecb4511c24d2cc3e697f9eb0fba8aa371709fb276fa05ca332.jpg)  
+Figure 8: An example of augmentation result from CHEF in KLUE-YNAT

2023/CHEF in the Language Kitchen_ A Generative Data Augmentation Leveraging Korean Morpheme Ingredients/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53f04215223b1d52f588460a3b62e7057538b91daa151f430c2228cf04f3c099
+size 1016450

2023/CLAD-ST_ Contrastive Learning with Adversarial Data for Robust Speech Translation/3c4dc6ba-99d3-4dfa-ade5-337f84d977db_content_list.json

@@ -0,0 +1,1211 @@
+[
+    {
+        "type": "text",
+        "text": "CLAD-ST: Contrastive Learning with Adversarial Data for Robust Speech Translation",
+        "text_level": 1,
+        "bbox": [
+            208,
+            87,
+            789,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Sathish Reddy Indurthi, Shamil Chollampatt, Ravi Agrawal, Marco Turchi",
+        "bbox": [
+            300,
+            142,
+            699,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zoom Video Communications",
+        "bbox": [
+            376,
+            177,
+            623,
+            190
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sathishreddy.indurthi@zoom.us, shamil.chollampatt@zoom.us,",
+        "bbox": [
+            215,
+            192,
+            786,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ravi.agrawal@zoom.us,marco.turchi@zoom.us",
+        "bbox": [
+            294,
+            210,
+            707,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The cascaded approach continues to be the most popular choice for speech translation (ST). This approach consists of an automatic speech recognition (ASR) model and a machine translation (MT) model that are used in a pipeline to translate speech in one language to text in another language. MT models are often trained on well-formed text and therefore lack robustness while translating noisy ASR outputs in the cascaded approach, degrading the overall translation quality significantly. We address this robustness problem in downstream MT models by forcing the MT encoder to bring the representations of a noisy input closer to its clean version in the semantic space. This is achieved by introducing a contrastive learning method that leverages adversarial examples in the form of ASR outputs paired with their corresponding human transcripts to optimize the network parameters. In addition, a curriculum learning strategy is then used to stabilize the training by alternating the standard MT log-likelihood loss and the contrastive losses. Our approach achieves significant gains of up to 3 BLEU scores in English-German and English-French speech translation without hurting the translation quality on clean text.",
+        "bbox": [
+            141,
+            275,
+            460,
+            659
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            668,
+            258,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Neural machine translation (NMT) has made significant advancements over the past several years with claims of achieving 'human parity' (Hassan et al., 2018), and 'super-human performance' (Barrault et al., 2019). However, despite making tremendous progress in quality and coverage (Costa-jussa et al., 2022), NMT models have been identified to lack robustness in dealing with noisy inputs (Belinkov and Bisk, 2018).",
+        "bbox": [
+            112,
+            693,
+            487,
+            837
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Robustness is especially important in cascaded speech translation (ST) systems, where an NMT model works on the output of the upstream automatic speech recognition (ASR) system. In this scenario, significant MT performance degradation has",
+        "bbox": [
+            112,
+            839,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Human Transcript: I'm not sure that's wise, given the importance of the problem, but there's now the geoengineering discussion about: Should that be in the back pocket in case things happen faster, or this innovation goes a lot slower than we expect?",
+        "bbox": [
+            519,
+            252,
+            873,
+            313
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ASR Output: I'm not sure that's why it's given the importance of the problem. But now that the geoengineering discussion about should that be in the back pocket in case things happen faster or this innovation goes a lot slower than we expect.",
+        "bbox": [
+            517,
+            317,
+            873,
+            378
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Example human transcript and OpenAI Whisper ASR output from the MuST-C En-De devset. The underlined text indicates erroneous phrases and italicized indicates punctuation errors.",
+        "bbox": [
+            507,
+            390,
+            885,
+            447
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "been measured due to  $i$  error propagation from the ASR and ii) the mismatch between training-testing condition as the NMT model is typically trained on well-formed text making it weak in dealing with noisy inputs. For these reasons, there has been significant effort towards building end-to-end ST models (Bérard et al., 2016; Weiss et al., 2017). However, due to the lack of sufficient speech translation datasets in many languages and the flexibility to independently optimize ASR and MT systems, cascaded systems continue to be a dominant approach (Anastasopoulos et al., 2022).",
+        "bbox": [
+            507,
+            481,
+            884,
+            674
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Prior research has tried to tackle the robustness problem in NMT models independently by (1) synthetic noise injection (Belinkov and Bisk, 2018), including mimicking ASR errors for the ST task (Sperber et al., 2017; Li et al., 2018; Martucci et al., 2021; Wang et al., 2022) or (2) by adding noise to the continuous-space representations learned by the NMT model (Sato et al., 2019; Cheng et al., 2020; Wei et al., 2022). Some of the synthetic noise injection methods replace words with probable words from ASR confusion lattices (Martucci et al., 2021; Wang et al., 2022) but fail to consider the full context and to realistically replicate ASR errors, like the phrase-level errors and segmentation errors that ASR models often make (see Table 1). On the other",
+        "bbox": [
+            507,
+            677,
+            885,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "9049",
+        "bbox": [
+            478,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9049-9056 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            836,
+            973
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/c1dcfeffafb9c5b1034ea7a4c031f9c6f73334a996f6901650f569b6520acb62.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            186,
+            84,
+            416,
+            181
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/929bdd23670515232314d61d779cee3a1e7e69db6801a0f2fa31b9f60adbbbe1.jpg",
+        "image_caption": [
+            "Figure 1: Proposed model and the two alternating training objectives.  $\\mathcal{L}_{\\mathrm{NMT}}$  is the standard NMT negative log-likelihood loss and  $\\mathcal{L}_{\\mathrm{CTL}}$  is the contrastive loss for improving robustness to ASR outputs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            194,
+            186,
+            389,
+            280
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "hand, the latter approaches predominantly generate arbitrary or random error distributions and overlook the real-world error signals from the actual upstream ASR model.",
+        "bbox": [
+            112,
+            379,
+            487,
+            441
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the robustness problem, particularly in the context of cascaded ST, we propose to combine the best of the two approaches. This is obtained by training the NMT model with adversarial examples generated from the ASR outputs and encouraging the encoder representations of both the ASR outputs and their corresponding human transcripts to be closer to each other. This is done via contrastive learning (Saunshi et al., 2019), in addition to the standard log-likelihood NMT training, and through curriculum learning (Bengio et al., 2009) by pretraining on log-likelihood loss followed by iteratively alternating between the two objectives. Our method leads to significant improvements in the ST task on two language pairs (En-De, En-Fr) without hurting the performance on MT of well-formed texts. Our approach also follows the standard cascade approach that does not rely on any end-to-end ST data (speech aligned with their transcription and translation) unlike that of Bahar et al. (2021) and Lam et al. (2021), making it easier to apply to a wider range of language pairs.",
+        "bbox": [
+            112,
+            444,
+            489,
+            813
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            828,
+            216,
+            841
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our NMT model is a Transformer model (Vaswani et al., 2017) with several encoder and decoder layers, with self-attention modules within each block and a cross-attention module to link the",
+        "bbox": [
+            112,
+            854,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "encoder and decoder components. Training the network parameters  $\\Theta$  is done by minimizing the mean negative log-likelihood over a batch of  $n$  source-target text translation pairs  $(X,Y) = (\\mathbf{x}^{(1)},\\mathbf{y}^{(1)}),\\ldots ,(\\mathbf{x}^{(n)},\\mathbf{y}^{(n)})$ . This objective encourages producing a target sentence  $\\mathbf{y}$  with a sequence of tokens  $y_{1},y_{2},\\ldots ,y_{|\\mathbf{y}|}$  given the source sentence  $\\mathbf{x}$  with sequence of tokens  $x_{1},x_{2},\\ldots ,x_{|\\mathbf{x}|}$ .",
+        "bbox": [
+            507,
+            84,
+            882,
+            231
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N M T}} = - \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {| \\mathbf {y} ^ {(i)} |} \\log P (y _ {j} ^ {(i)} | \\mathbf {x} ^ {(i)}, \\mathbf {y} _ {<   j} ^ {(i)}, \\Theta)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            250,
+            865,
+            299
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To improve the robustness of the NMT model on noisy ASR outputs for cascaded speech translation, we use a contrastive learning method (Wei et al., 2021) that is aimed to bring the representations of a noisy sentence generated by the upstream ASR closer to its clean version (the human transcript for the same audio) in the semantic space modeled by the NMT encoder.",
+        "bbox": [
+            507,
+            306,
+            882,
+            432
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To get the encoder sentence representations efficiently for contrastive learning, a [CLS] token is prepended to the input sentences similar to the BERT model (Devlin et al., 2019). We found that using encoder output corresponding to the [CLS] token is slightly better than using the mean of encoder outputs as the sentence representation.  $\\mathbf{s}_{\\mathbf{x}}$  which also gets added to each of the decoder outputs before softmax (see Figure 1).",
+        "bbox": [
+            507,
+            435,
+            882,
+            580
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Contrastive learning uses speech transcription corpora (i.e., speech paired with human transcripts). The speech input is passed through the upstream ASR model to obtain the ASR outputs. Given the noisy ASR output  $\\bar{\\mathbf{x}}$  paired with its corresponding clean human transcript  $\\mathbf{x}$ , we minimize the contrastive objective  $\\mathcal{L}_{\\mathrm{CTL}}$ , which is an average of two symmetric sentence-level contrastive loss functions.",
+        "bbox": [
+            507,
+            580,
+            882,
+            722
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\text {G i v e n a b a t c h o f} n \\text {e x a m p l e s} (X, \\bar {X}) = \\\\ \\left\\{\\left(\\mathbf {x} ^ {(1)}, \\bar {\\mathbf {x}} ^ {(1)}\\right), \\dots , \\left(\\mathbf {x} ^ {(n)}, \\bar {\\mathbf {x}} ^ {(n)}\\right) \\right\\} \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            510,
+            724,
+            880,
+            758
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C T L}} = \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\mathcal {L} _ {c} \\left(\\mathbf {x} ^ {(i)}\\right) + \\mathcal {L} _ {c} \\left(\\bar {\\mathbf {x}} ^ {(i)}\\right)\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            534,
+            766,
+            882,
+            806
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "where",
+        "bbox": [
+            509,
+            815,
+            559,
+            827
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {c} (\\mathbf {x}) = \\frac {- \\exp \\left(D (\\mathbf {s} _ {\\mathbf {x}} , \\mathbf {s} _ {\\bar {\\mathbf {x}}})\\right)}{\\exp \\left(D (\\mathbf {s} _ {\\mathbf {x}} , \\mathbf {s} _ {\\bar {\\mathbf {x}}})\\right) + \\sum_ {x ^ {\\prime}} \\exp (D (\\mathbf {s} _ {\\mathbf {x}}, \\hat {\\mathbf {s}} _ {\\mathbf {x} ^ {\\prime}}))}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            512,
+            834,
+            875,
+            878
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "$D(\\mathbf{u},\\mathbf{v})$  denotes the cosine distance between two vectors  $\\mathbf{u}$  and  $\\mathbf{v}$ .  $\\hat{\\mathbf{s}}_{\\mathbf{x}'}$  represents a negative example",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "9050",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "constructed for every other sentence  $\\mathbf{x}'$  in the batch. Given  $\\mathbf{x}' \\in (X \\cup \\bar{X}) \\setminus \\{\\mathbf{x}, \\bar{\\mathbf{x}}\\}$  and its sentence embedding  $\\mathbf{s}_{x'}$",
+        "bbox": [
+            112,
+            84,
+            489,
+            134
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {\\mathbf {s}} _ {\\mathbf {x} ^ {\\prime}} = \\left\\{ \\begin{array}{l l} \\mathbf {s} _ {\\mathbf {x}} + \\lambda_ {\\mathbf {x}} \\left(\\mathbf {s} _ {\\mathbf {x} ^ {\\prime}} - \\mathbf {s} _ {\\mathbf {x}}\\right) & \\text {i f ,} d _ {\\mathbf {x}} ^ {+} > d _ {\\mathbf {x}} ^ {-} \\\\ \\mathbf {s} _ {\\mathbf {x} ^ {\\prime}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            132,
+            158,
+            487,
+            200
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $d_{\\mathbf{x}}^{+} = \\| \\mathbf{s}_{\\mathbf{x}} - \\mathbf{s}_{\\overline{\\mathbf{x}}}\\|_{2}, d_{\\mathbf{x}}^{-} = \\| \\mathbf{s}_{\\mathbf{x}} - \\mathbf{s}_{\\mathbf{x}^{\\prime}}\\|_{2}$ , and  $\\lambda_{\\mathbf{x}} = \\frac{d_{\\mathbf{x}}^{+}}{d_{\\mathbf{x}}^{-}}$ . The above linear interpolation in Eq 2 with exponentially decaying  $\\lambda_{\\mathbf{x}}$  is implemented following Wei et al. (2021) to increase the hardness of negative examples by not selecting uninformative negative examples as training progresses.  $\\mathcal{L}_c(\\overline{\\mathbf{x}})$  follows similar formulation as  $\\mathcal{L}_c(\\mathbf{x})$  with  $\\mathbf{x}$  replaced by  $\\overline{\\mathbf{x}}$ , and vice versa.",
+        "bbox": [
+            112,
+            212,
+            489,
+            344
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To stabilize the NMT parameters before applying contrastive learning, we also use curriculum learning (Bengio et al., 2009). We first pre-train the model with NMT loss  $\\mathcal{L}_{\\mathrm{NMT}}$  only for the first  $N$  batches, followed by iteratively training alternate batches with  $\\mathcal{L}_{\\mathrm{CTL}}$  and  $\\mathcal{L}_{\\mathrm{NMT}}$  losses, respectively, until convergence. With this model and training regime, which alternates between training with translation and speech recognition datasets independently, we are able to build robust NMT models for cascaded ST without any requirement for speech translation data.",
+        "bbox": [
+            112,
+            348,
+            489,
+            539
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            554,
+            321,
+            571
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We experiment with English-German (En-De) and English-French (En-Fr) language directions.",
+        "bbox": [
+            112,
+            580,
+            485,
+            612
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Training Data: For parallel text translation data, we use WMT'16 En-De (Bojar et al., 2016) and WMT'14 En-Fr (Bojar et al., 2014) news translation task data for the corresponding language directions. To get the ASR data for contrastive learning, we use Open AI Whisper base ASR (Radford et al., 2022) to transcribe the speech in the training sets from Mozilla Common Voice 12.0 (Ardila et al., 2020), VoxPopuli (Wang et al., 2021), and the English set of Multilingual LibrSpeech (Pratap et al., 2020) for both language directions. As an additional experiment, we also try adding in-domain MuST-C training data. The full statistics of the datasets are in Appendix A.2 (Table 8).",
+        "bbox": [
+            112,
+            613,
+            489,
+            837
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Evaluation: We use MuST-C (Di Gangi et al., 2019) tst-COMMON as our test set and MuST-C dev as our validation set. For MT and ST tasks, the inputs to the model are the Whisper (base) ASR outputs and the corresponding human transcripts,",
+        "bbox": [
+            112,
+            838,
+            489,
+            920
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/227af530cf3b0c74951ef0c969c6389999af04acd8a16ba2a8375d0200a450ba.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>28.1</td><td>23.5</td><td>36.1</td><td>30.8</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>28.2</td><td>25.2</td><td>36.3</td><td>32.0</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>29.5</td><td>24.4</td><td>37.2</td><td>31.3</td></tr><tr><td>CLAD-ST</td><td>28.8</td><td>26.6</td><td>36.5</td><td>33.4</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            882,
+            158
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Performance of various models on MT and ST tasks on MuST-C tst-COMMON.",
+        "bbox": [
+            507,
+            167,
+            882,
+            195
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "respectively. We evaluate against the English and French reference translations for the respective translation directions. We use case-sensitive, detokenized BLEU computed using sacrebleu $^{1}$  (Post, 2018). We also compute statistical significance using paired bootstrap resampling (Koehn, 2004). Models: We use a base Transformer (Vaswani et al., 2017) as our baseline NMT model, which we extend to support our proposed approach, Contrastive Learning with Adversarial Data for Speech Translation (CLAD-ST). We also compare against two recent state-of-the-art approaches for NMT robustness which we re-implement: (1) CONF-ASR (Wang et al., 2022) generates adversarial examples by substituting random source tokens with candidate tokens from ASR confusion set based on their embedding distances to the original token. (2) CSA-NMT (Wei et al.,  $2022)^{2}$  uses a separate semantic encoder to model a semantic region around the source and target sentences. To improve NMT robustness, samples from this region are used to augment the NMT decoder representations.",
+        "bbox": [
+            505,
+            223,
+            884,
+            577
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We tokenize all datasets using a sentencepiece model (Kudo and Richardson, 2018) for each language pair on the text translation datasets with a shared source-target vocabulary size of 32,000. We use OpenAI Wispher (base) ASR as our main upstream model. We also evaluate the ST performance of our model using Whisper large ASR. Other training details and hyperparameters are in Appendix A.1.",
+        "bbox": [
+            507,
+            580,
+            882,
+            725
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            739,
+            722,
+            757
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Overall Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            768,
+            721,
+            783
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our proposed CLAD-ST achieves 26.6 and 33.4 BLEU on the En-De and En-Fr ST tasks, significantly  $(p < 0.001)$  outperforming the Transformer baseline by 3.1 BLEU points and 2.6 BLEU",
+        "bbox": [
+            507,
+            790,
+            884,
+            854
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ https://github.com/mjpost/sacrebleu. Signature: nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp",
+        "bbox": [
+            507,
+            866,
+            882,
+            892
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ Code and network hyperparameters are from https://github.com/pemywei/csanmt",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "9051",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/9e15bacd98b851205922f32a57bab18ced7f07b9e074648ab0bf4ca54a282a0d.jpg",
+        "image_caption": [
+            "Figure 2: BLEU scores compared to the number of token-level errors in the ASR outputs on En-De ST task on MuST-C tst-COMMON dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            119,
+            85,
+            495,
+            253
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "points (see Table 2). Our approach consistently performs better than other baselines (CONF-ASR and CSA-NMT), yielding gains of 1.4 BLEU points  $(p < 0.001)$  on both language directions. The improvement of our approach over CONF-ASR shows the importance of contextual modeling of the ASR errors instead of simulating them with ASR confusion lattices. CSA-NMT achieves the largest improvement in the MT task but fails to specialize to ASR errors, yielding lower improvements compared to our approach.",
+        "bbox": [
+            112,
+            344,
+            489,
+            521
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Robustness",
+        "text_level": 1,
+        "bbox": [
+            112,
+            550,
+            250,
+            564
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Figure 2, we report the performance of all the En-De models as the number of ASR errors increases in the input sequence. We group the ASR outputs based on the number of tokens different from the reference transcripts on the test set. The performance of the Transformer baseline model drastically drops as the number of ASR errors increases. CONF-ASR performs better than the Transformer baseline in all the ASR error conditions and closely with CLAD-ST on ASR outputs with one or two errors. However, CLAD-ST notably outperforms it as the number of errors increases. CSA-NMT performs slightly better than the baseline Transformer when the number of errors is few and performs slightly better than CONF-ASR when increasing the number of errors due to their semantic augmentation. Overall, our approach has shown to be robust against sequences having several ASR errors and performs consistently better than the other baselines, irrespective of the number of errors in the sequence.",
+        "bbox": [
+            112,
+            581,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Curriculum Learning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            727,
+            101
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We study the effect of the curriculum learning strategy (Section 2) and reports the results in Table 3. We find that curriculum learning strategy brings an improvement of 0.5 BLEU on En-De and 0.7 BLEU On En-Fr ST tasks, respectively.",
+        "bbox": [
+            507,
+            105,
+            884,
+            186
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/5897d39865bc589f32ffcba76150919b51178a78f391533a938942030fdf129a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>En-De</td><td>En-Fr</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>30.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>33.4</td></tr><tr><td>w/o Curriculum learning</td><td>26.1</td><td>32.7</td></tr></table>",
+        "bbox": [
+            542,
+            196,
+            850,
+            249
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.4 Out-of-distribution ASR",
+        "text_level": 1,
+        "bbox": [
+            507,
+            313,
+            747,
+            329
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Table 4, we report the performance of all the Ende systems using Whisper base as the ASR model during training and a higher-quality ASR (Whisper large) during testing. Although training on lower-quality ASR outputs with more errors, a similar BLEU score improvement is also noticed for all the systems when fed better ASR outputs. However, CLAD-ST achieves the largest improvements over the baseline, showing that it better generalizes over different ASR-quality outputs.",
+        "bbox": [
+            507,
+            335,
+            884,
+            495
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/b388d47b292ca2022285c8bb73a536f71ad68d82f05ea4add9568e7b06f484cc.jpg",
+        "table_caption": [
+            "Table 3: Performance on ST tasks with and without curriculum learning on MuST-C tst-COMMON."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">En-De</td></tr><tr><td>Wbase (WER = 14.3)</td><td>Wlarge (WER = 7.6)</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>24.4</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>25.2</td><td>26.3</td></tr><tr><td>CSA-NMT (Wei et al., 2022)</td><td>24.4</td><td>25.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>27.4</td></tr></table>",
+        "bbox": [
+            510,
+            507,
+            897,
+            595
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 4: BLEU scores on MuST-C En-De tst-COMMON ST task when using Whisper-large (Wlarge) vs. Whisper-base (Wbase) ASR. Word error rate (WER) of the ASR model is provided.",
+        "bbox": [
+            507,
+            605,
+            882,
+            662
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.5 Model Size",
+        "text_level": 1,
+        "bbox": [
+            507,
+            689,
+            643,
+            703
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To investigate if our results scale when we increase the size of the network, we use Transformer-large instead of Transformer-base for our baseline models as well as CLAD-ST. The results are reported in Table 5. We find that we get improvements of 1.5 and 1.3 BLEU score points for the ST task on En-De and En-Fr, respectively, compared to CONF-ASR, demonstrating its advantage even on larger NMT architectures.",
+        "bbox": [
+            507,
+            709,
+            884,
+            853
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.6 In-domain Training Data",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            754,
+            881
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We also experiment with adding in-domain training data from MuST-C for training the NMT models in",
+        "bbox": [
+            507,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "9052",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/7997a22651bac4774a3b1114a38c32cbe670e49a23d50b2de97567ac79248fe1.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-big</td><td>29.6</td><td>24.9</td><td>38.7</td><td>33.3</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>29.8</td><td>26.6</td><td>38.8</td><td>34.6</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>31.1</td><td>25.7</td><td>39.9</td><td>33.7</td></tr><tr><td>CLAD-ST</td><td>30.3</td><td>28.1</td><td>39.1</td><td>35.9</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            485,
+            156
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "CLAD-ST and the baselines (see Table 6). In line with our expectations, we find a notable improvement in performance of all the systems. Despite the improved baselines, we observe similar improvements for CLAD-ST of 1.4 BLEU on both En-De and En-Fr ST tasks compared to CONF-ASR, and 2.4 and 2.2 BLEU on En-De and En-Fr ST tasks, respectively, compared to CSA-NMT.",
+        "bbox": [
+            112,
+            237,
+            489,
+            367
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/126d26b75dbd75c5b0f44ea53b702ed259230c2f7d3d9607c4fb7c2c791d9c23.jpg",
+        "table_caption": [
+            "Table 5: Performance of using Transformer-large as the model architecture on MT and ST tasks on MuST-C tst-COMMON."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>31.0</td><td>26.1</td><td>37.9</td><td>32.9</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>30.9</td><td>27.9</td><td>38.0</td><td>33.9</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>32.1</td><td>26.9</td><td>38.8</td><td>33.1</td></tr><tr><td>CLAD-ST</td><td>31.6</td><td>29.4</td><td>38.2</td><td>35.3</td></tr></table>",
+        "bbox": [
+            115,
+            378,
+            485,
+            454
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Performance of various models on MT and ST tasks on MuST-C tst-COMMON when adding indomain MuST-C training data.",
+        "bbox": [
+            112,
+            464,
+            489,
+            508
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            537,
+            270,
+            552
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Several prior works tried to improve the robustness of MT models to noisy ASR inputs for cascaded speech translation. Sperber et al. (2017) randomly sampled synthetic noise in the form of insertions, deletions, and substitutions and injected them into the source side of MT training data. Li et al. (2018) encoded the syllables (Pinyin) for Chinese ST and injected synthetic character-level substitutions to simulate ASR errors. In contrast, Martucci et al. (2021) trained a lexical noise model using confusion sets extracted by aligning the actual ASR output to reference transcript to inject synthetic errors. Wang et al. (2022) picked substitution candidates from the ASR confusion set that are farthest away from the source token embeddings leading to more robust MT models. Cheng et al. (2019) also used the actual ASR output in training incorporating an additional discriminator network that tries to distinguish between human transcript and the ASR output following Cheng et al. (2018) to improve robustness. Padfield and Cherry (2021) simulated not only token-level errors but also segmentation errors",
+        "bbox": [
+            112,
+            565,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "and projected them on the target side as well. Bahar et al. (2021), on the other hand, proposes to jointly optimize the ASR and MT components in the cascaded system using speech-translation data. Similarly, Lam et al. (2021) also uses speech-translation data to reduce error propagation between the two components by mutually training the MT system on ASR n-best outputs and using cyclic feedback from MT outputs for self-training of the ASR model.",
+        "bbox": [
+            507,
+            84,
+            884,
+            230
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            240,
+            640,
+            255
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We improve the robustness of MT to ASR outputs using contrastive learning to bring the representations of clean and noisy examples closer in the semantic space. Our approach does not require any speech translation corpus. We significantly improve the translation accuracy on noisy ASR outputs without degrading translation accuracy on clean text. We also show that the approach is scalable to better-quality ASR models in the cascade other than the one used during training. The proposed approach is generic and is applicable beyond the context of speech translation alone, such as translating user-generated chat text or non-native text if paired noisy-clean data is available.",
+        "bbox": [
+            507,
+            265,
+            884,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            502,
+            615,
+            517
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The limitations of our paper are:",
+        "bbox": [
+            509,
+            527,
+            752,
+            543
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- As in any ST cascade architecture, the performance of our MT system depends on the quality of the ASR model outputs. Since we use examples generated by the ASR to train the MT model, the quality of ASR outputs directly affects the model performance. In our work we tested two ASR systems having different quality (base: 14.3 WER and large: 7.6 WER). This is particularly relevant when English is not the source language side.",
+            "- Evaluation is done using only the BLEU score. We did not use human evaluation or COMET.",
+            "- The evaluation is limited to two language pairs having English as the source. This selection of the source language is quite important because having a no-English language as a source will expose our MT model to a) ASR models of probably lower quality and b) different and more varied linguistic challenges that might affect the work of the adversarial method."
+        ],
+        "bbox": [
+            531,
+            552,
+            885,
+            892
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "9053",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Antonios Anastasopoulos, Loic Barrault, Luisa Bentivogli, Marcely Zanon Boito, Ondrej Bojar, Roldano Cattoni, Anna Currey, Georgiana Dinu, Kevin Duh, Maha Elbayad, Clara Emmanuel, Yannick Esteve, Marcello Federico, Christian Federmann, Souhir Gahbiche, Hongyu Gong, Roman Grundkiewicz, Barry Haddow, Benjamin Hsu, David Javorsky, Vera Kloudova, Surafel Lakew, Xutai Ma, Prashant Mathur, Paul McNamee, Kenton Murray, Maria Nadejde, Satoshi Nakamura, Matteo Negri, Jan Niehues, Xing Niu, John Ortega, Juan Pino, Elizabeth Salesky, Jiatong Shi, Matthias Sperber, Sebastian Stüker, Katsuhito Sudoh, Marco Turchi, Yogesh Virkar, Alexander Waibel, Changhan Wang, and Shinji Watanabe. 2022. Findings of the IWSLT 2022 evaluation campaign. In Proceedings of the 19th International Conference on Spoken Language Translation.",
+            "Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henrtty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference.",
+            "Parnia Bahar, Tobias Bieschke, Ralf Schlüter, and Hermann Ney. 2021. Tight integrated end-to-end training for cascaded speech translation. In Proceedings of the 2021 IEEE Spoken Language Technology Workshop.",
+            "Loic Barrault, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Muller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation.",
+            "Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In International Conference on Learning Representations.",
+            "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning.",
+            "Alexandre Bérard, Olivier Pietquin, Laurent Besacier, and Christophe Servan. 2016. Listen and translate: A proof of concept for end-to-end speech-to-text translation. In Proceedings of the NIPS Workshop on end-to-end learning for speech and audio processing.",
+            "Ondrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia, and Ales Tamchyna. 2014. Findings of the 2014 workshop on"
+        ],
+        "bbox": [
+            115,
+            105,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation.",
+            "Ondrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Aurélie Néveol, Mariana Neves, Martin Popel, Matt Post, Raphael Rubino, Carolina Scarton, Lucia Specia, Marco Turchi, Karin Verspoor, and Marcos Zampieri. 2016. Findings of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers, Berlin, Germany.",
+            "Qiao Cheng, Meiyuan Fan, Yaqian Han, Jin Huang, and Yitao Duan. 2019. Breaking the data barrier: Towards robust speech translation via adversarial stability training. In Proceedings of the 16th International Conference on Spoken Language Translation.",
+            "Yong Cheng, Lu Jiang, Wolfgang Macherey, and Jacob Eisenstein. 2020. AdvAug: Robust adversarial augmentation for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.",
+            "Yong Cheng, Zhaopeng Tu, Fandong Meng, Junjie Zhai, and Yang Liu. 2018. Towards robust neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics.",
+            "Marta R Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.",
+            "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.",
+            "Mattia A. Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019. MuST-C: a Multilingual Speech Translation Corpus. 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).",
+            "Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, Will Lewis, Mu Li, Shujie Liu, Tie-Yan Liu, Renqian Luo, Arul Menezes, Tao Qin, Frank Seide, Xu Tan, Fei Tian, Lijun Wu, Shuangzhi Wu, Yingce Xia, Dongdong Zhang, Zhirui Zhang, and Ming Zhou. 2018. Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "9054",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing.",
+            "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations.",
+            "Tsz Kin Lam, Shigehiko Schamoni, and Stefan Riezler. 2021. Cascaded models with cyclic feedback for direct speech translation. In Proceedings of the 2021 IEEE International Conference on Acoustics, Speech and Signal Processing.",
+            "Xiang Li, Haiyang Xue, Wei Chen, Yang Liu, Yang Feng, and Qun Liu. 2018. Improving the robustness of speech translation. arXiv preprint arXiv:1811.00728.",
+            "Giuseppe Martucci, Mauro Cettolo, Matteo Negri, and Marco Turchi. 2021. Lexical modeling of ASR errors for robust speech translation. In Proceedings of Interspeech.",
+            "Dirk Padfield and Colin Cherry. 2021. Inverted projection for robust speech translation. In Proceedings of the 18th International Conference on Spoken Language Translation. Association for Computational Linguistics.",
+            "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers.",
+            "Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. arXiv preprint arXiv:2012.03411.",
+            "Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356.",
+            "Motoki Sato, Jun Suzuki, and Shun Kiyono. 2019. *Effective adversarial regularization for neural machine translation*. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.",
+            "Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. 2019. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning.",
+            "Matthias Sperber, Jan Niehues, and Alex Waibel. 2017. Toward robust neural machine translation for noisy input sequences. In Proceedings of the International Workshop on Spoken Language Translation."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "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. In Advances in Neural Information Processing Systems, volume 30.",
+            "Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. 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).",
+            "Shuaibo Wang, Yufeng Chen, Songming Zhang, Deyi Xiong, and Jinan Xu. 2022. Adversarily improving NMT robustness to ASR errors with confusion sets. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers).",
+            "Xiangpeng Wei, Rongxiang Weng, Yue Hu, Luxi Xing, Heng Yu, and Weihua Luo. 2021. On learning universal representations across languages. In International Conference on Learning Representations.",
+            "Xiangpeng Wei, Heng Yu, Yue Hu, Rongxiang Weng, Weihua Luo, and Rong Jin. 2022. Learning to generalize to more: Continuous semantic augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.",
+            "Ron J. Weiss, Jan Chorowski, Navdeep Jaitly, Yonghui Wu, and Zhifeng Chen. 2017. Sequence-to-sequence models can directly translate foreign speech. In Proceedings of the Conference of the International Speech Communication Association."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            626
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            510,
+            640,
+            749,
+            657
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A.1 Training Details and Hyperparameters",
+        "text_level": 1,
+        "bbox": [
+            510,
+            667,
+            855,
+            683
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We implement our approach and (Wang et al., 2022) using  $\\text{Fairseq}^3$ . For Wang et al. (2021), we randomly selected  $14\\%$  tokens of sentences for a replacement to generate adversarial examples based on the ASR word error rate. All experiments are performed on A100 GPU, and gradient accumulation and batch size are set to 8 iterations and 8k tokens. For CLAD-ST, our NMT model is pretrained for the first 50K batches for NMT loss in the curriculum. For evaluation, the beam size and length penalty are set to 4 and 0.6 for both tasks. Hyperparameters for all the models are shown in Table 7.",
+        "bbox": [
+            509,
+            688,
+            882,
+            895
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://github.com/facebookresearch/fairseq",
+        "bbox": [
+            529,
+            903,
+            803,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "9055",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/707902e9cd0d63161ce418ae4dc1ef73659a5f25eae7fad06f0450c0139269c9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameter</td><td>Baseline</td><td>CONF-ASR</td><td>CSA-NMT</td><td>CLAD-ST</td></tr><tr><td>encoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>encoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>encoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>encoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>decoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>decoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>decoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>decoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>dropout</td><td></td><td>0.3</td><td></td><td></td></tr><tr><td>optimizer</td><td></td><td>adam</td><td></td><td></td></tr><tr><td>adam-β</td><td></td><td>(0.9, 0.98)</td><td></td><td></td></tr><tr><td>clip-norm</td><td></td><td>0.0</td><td></td><td></td></tr><tr><td>lr scheduler</td><td></td><td>invers-sqrt</td><td></td><td></td></tr><tr><td>learning rate</td><td></td><td>5e-4</td><td></td><td></td></tr><tr><td>warmup-updates</td><td></td><td>4000</td><td></td><td></td></tr><tr><td>label-smoothing</td><td></td><td>0.1</td><td></td><td></td></tr><tr><td>max tokens</td><td>12800</td><td>8192</td><td>6144</td><td>8192</td></tr><tr><td>num-initial-updates-nmt</td><td>N/A</td><td>50000</td><td>50000</td><td>50000</td></tr><tr><td>csanmt-semantic-samples</td><td>N/A</td><td>N/A</td><td>20</td><td>N/A</td></tr></table>",
+        "bbox": [
+            238,
+            80,
+            759,
+            337
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7da601e9d7facea7d6c71de58a835472075c467d2a1c76f949520d128fff4ea9.jpg",
+        "table_caption": [
+            "Table 7: Model hyperparameters"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Type</td><td>Name</td><td>Lang.</td><td>#Examples</td></tr><tr><td rowspan=\"4\">MT Data</td><td>WMT&#x27;16</td><td>En-De</td><td>4.5M</td></tr><tr><td>WMT&#x27;14</td><td>En-Fr</td><td>40 M</td></tr><tr><td>MuST-C</td><td>En-De</td><td>251K</td></tr><tr><td>train</td><td>En-Fr</td><td>275K</td></tr><tr><td rowspan=\"3\">ASR Data</td><td>VoxPopuli</td><td>En</td><td>0.18M</td></tr><tr><td>Commonvoice</td><td>En</td><td>0.99M</td></tr><tr><td>MLS</td><td>En</td><td>11M</td></tr><tr><td rowspan=\"2\">Dev</td><td rowspan=\"2\">MuST-C Dev</td><td>En-De</td><td>1415</td></tr><tr><td>En-Fr</td><td>1412</td></tr><tr><td rowspan=\"2\">Test</td><td>MuST-C</td><td>En-De</td><td>2580</td></tr><tr><td>tst-COMMON</td><td>En-Fr</td><td>2632</td></tr></table>",
+        "bbox": [
+            115,
+            382,
+            470,
+            532
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Statistics of datasets used in the experiments.",
+        "bbox": [
+            115,
+            541,
+            482,
+            556
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A.2 Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            581,
+            231,
+            595
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Dataset statistics can be found in Table 8. WMT'16 En-De training data<sup>4</sup> includes Europarl, Common Crawl corpus, and News Commentary corpora. WMT'14 En-Fr<sup>5</sup> training data includes Europarl, Common Crawl, UN corpus, News Commentary, and Fr-En Gigaword corpus.",
+        "bbox": [
+            112,
+            601,
+            489,
+            699
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://www.statmt.org/wmt16/ translation-task.html",
+        "bbox": [
+            112,
+            866,
+            364,
+            891
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{5}$ https://www.statmt.org/wmt14/translation-task.html",
+        "bbox": [
+            112,
+            891,
+            364,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "9056",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    }
+]

2023/CLAD-ST_ Contrastive Learning with Adversarial Data for Robust Speech Translation/3c4dc6ba-99d3-4dfa-ade5-337f84d977db_model.json

@@ -0,0 +1,1613 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.21,
+                0.089,
+                0.79,
+                0.13
+            ],
+            "angle": 0,
+            "content": "CLAD-ST: Contrastive Learning with Adversarial Data for Robust Speech Translation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.302,
+                0.143,
+                0.7,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Sathish Reddy Indurthi, Shamil Chollampatt, Ravi Agrawal, Marco Turchi"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.377,
+                0.178,
+                0.625,
+                0.191
+            ],
+            "angle": 0,
+            "content": "Zoom Video Communications"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.216,
+                0.193,
+                0.787,
+                0.209
+            ],
+            "angle": 0,
+            "content": "sathishreddy.indurthi@zoom.us, shamil.chollampatt@zoom.us,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.295,
+                0.211,
+                0.709,
+                0.226
+            ],
+            "angle": 0,
+            "content": "ravi.agrawal@zoom.us,marco.turchi@zoom.us"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.277,
+                0.461,
+                0.661
+            ],
+            "angle": 0,
+            "content": "The cascaded approach continues to be the most popular choice for speech translation (ST). This approach consists of an automatic speech recognition (ASR) model and a machine translation (MT) model that are used in a pipeline to translate speech in one language to text in another language. MT models are often trained on well-formed text and therefore lack robustness while translating noisy ASR outputs in the cascaded approach, degrading the overall translation quality significantly. We address this robustness problem in downstream MT models by forcing the MT encoder to bring the representations of a noisy input closer to its clean version in the semantic space. This is achieved by introducing a contrastive learning method that leverages adversarial examples in the form of ASR outputs paired with their corresponding human transcripts to optimize the network parameters. In addition, a curriculum learning strategy is then used to stabilize the training by alternating the standard MT log-likelihood loss and the contrastive losses. Our approach achieves significant gains of up to 3 BLEU scores in English-German and English-French speech translation without hurting the translation quality on clean text."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.669,
+                0.26,
+                0.685
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.694,
+                0.489,
+                0.838
+            ],
+            "angle": 0,
+            "content": "Neural machine translation (NMT) has made significant advancements over the past several years with claims of achieving 'human parity' (Hassan et al., 2018), and 'super-human performance' (Barrault et al., 2019). However, despite making tremendous progress in quality and coverage (Costa-jussa et al., 2022), NMT models have been identified to lack robustness in dealing with noisy inputs (Belinkov and Bisk, 2018)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Robustness is especially important in cascaded speech translation (ST) systems, where an NMT model works on the output of the upstream automatic speech recognition (ASR) system. In this scenario, significant MT performance degradation has"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.52,
+                0.253,
+                0.875,
+                0.315
+            ],
+            "angle": 0,
+            "content": "Human Transcript: I'm not sure that's wise, given the importance of the problem, but there's now the geoengineering discussion about: Should that be in the back pocket in case things happen faster, or this innovation goes a lot slower than we expect?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.519,
+                0.318,
+                0.875,
+                0.379
+            ],
+            "angle": 0,
+            "content": "ASR Output: I'm not sure that's why it's given the importance of the problem. But now that the geoengineering discussion about should that be in the back pocket in case things happen faster or this innovation goes a lot slower than we expect."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.391,
+                0.886,
+                0.448
+            ],
+            "angle": 0,
+            "content": "Table 1: Example human transcript and OpenAI Whisper ASR output from the MuST-C En-De devset. The underlined text indicates erroneous phrases and italicized indicates punctuation errors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.482,
+                0.885,
+                0.675
+            ],
+            "angle": 0,
+            "content": "been measured due to \\(i\\) error propagation from the ASR and ii) the mismatch between training-testing condition as the NMT model is typically trained on well-formed text making it weak in dealing with noisy inputs. For these reasons, there has been significant effort towards building end-to-end ST models (Bérard et al., 2016; Weiss et al., 2017). However, due to the lack of sufficient speech translation datasets in many languages and the flexibility to independently optimize ASR and MT systems, cascaded systems continue to be a dominant approach (Anastasopoulos et al., 2022)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.678,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Prior research has tried to tackle the robustness problem in NMT models independently by (1) synthetic noise injection (Belinkov and Bisk, 2018), including mimicking ASR errors for the ST task (Sperber et al., 2017; Li et al., 2018; Martucci et al., 2021; Wang et al., 2022) or (2) by adding noise to the continuous-space representations learned by the NMT model (Sato et al., 2019; Cheng et al., 2020; Wei et al., 2022). Some of the synthetic noise injection methods replace words with probable words from ASR confusion lattices (Martucci et al., 2021; Wang et al., 2022) but fail to consider the full context and to realistically replicate ASR errors, like the phrase-level errors and segmentation errors that ASR models often make (see Table 1). On the other"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.48,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9049"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.16,
+                0.946,
+                0.838,
+                0.974
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9049-9056 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.187,
+                0.085,
+                0.418,
+                0.182
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.195,
+                0.187,
+                0.391,
+                0.281
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.296,
+                0.49,
+                0.354
+            ],
+            "angle": 0,
+            "content": "Figure 1: Proposed model and the two alternating training objectives. \\(\\mathcal{L}_{\\mathrm{NMT}}\\) is the standard NMT negative log-likelihood loss and \\(\\mathcal{L}_{\\mathrm{CTL}}\\) is the contrastive loss for improving robustness to ASR outputs."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.38,
+                0.489,
+                0.442
+            ],
+            "angle": 0,
+            "content": "hand, the latter approaches predominantly generate arbitrary or random error distributions and overlook the real-world error signals from the actual upstream ASR model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.445,
+                0.49,
+                0.814
+            ],
+            "angle": 0,
+            "content": "To address the robustness problem, particularly in the context of cascaded ST, we propose to combine the best of the two approaches. This is obtained by training the NMT model with adversarial examples generated from the ASR outputs and encouraging the encoder representations of both the ASR outputs and their corresponding human transcripts to be closer to each other. This is done via contrastive learning (Saunshi et al., 2019), in addition to the standard log-likelihood NMT training, and through curriculum learning (Bengio et al., 2009) by pretraining on log-likelihood loss followed by iteratively alternating between the two objectives. Our method leads to significant improvements in the ST task on two language pairs (En-De, En-Fr) without hurting the performance on MT of well-formed texts. Our approach also follows the standard cascade approach that does not rely on any end-to-end ST data (speech aligned with their transcription and translation) unlike that of Bahar et al. (2021) and Lam et al. (2021), making it easier to apply to a wider range of language pairs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.829,
+                0.218,
+                0.843
+            ],
+            "angle": 0,
+            "content": "2 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Our NMT model is a Transformer model (Vaswani et al., 2017) with several encoder and decoder layers, with self-attention modules within each block and a cross-attention module to link the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.883,
+                0.233
+            ],
+            "angle": 0,
+            "content": "encoder and decoder components. Training the network parameters \\(\\Theta\\) is done by minimizing the mean negative log-likelihood over a batch of \\(n\\) source-target text translation pairs \\((X,Y) = (\\mathbf{x}^{(1)},\\mathbf{y}^{(1)}),\\ldots ,(\\mathbf{x}^{(n)},\\mathbf{y}^{(n)})\\). This objective encourages producing a target sentence \\(\\mathbf{y}\\) with a sequence of tokens \\(y_{1},y_{2},\\ldots ,y_{|\\mathbf{y}|}\\) given the source sentence \\(\\mathbf{x}\\) with sequence of tokens \\(x_{1},x_{2},\\ldots ,x_{|\\mathbf{x}|}\\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.527,
+                0.252,
+                0.866,
+                0.3
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {N M T}} = - \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {| \\mathbf {y} ^ {(i)} |} \\log P (y _ {j} ^ {(i)} | \\mathbf {x} ^ {(i)}, \\mathbf {y} _ {<   j} ^ {(i)}, \\Theta)\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.307,
+                0.884,
+                0.434
+            ],
+            "angle": 0,
+            "content": "To improve the robustness of the NMT model on noisy ASR outputs for cascaded speech translation, we use a contrastive learning method (Wei et al., 2021) that is aimed to bring the representations of a noisy sentence generated by the upstream ASR closer to its clean version (the human transcript for the same audio) in the semantic space modeled by the NMT encoder."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.436,
+                0.884,
+                0.581
+            ],
+            "angle": 0,
+            "content": "To get the encoder sentence representations efficiently for contrastive learning, a [CLS] token is prepended to the input sentences similar to the BERT model (Devlin et al., 2019). We found that using encoder output corresponding to the [CLS] token is slightly better than using the mean of encoder outputs as the sentence representation. \\( \\mathbf{s}_{\\mathbf{x}} \\) which also gets added to each of the decoder outputs before softmax (see Figure 1)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.581,
+                0.884,
+                0.723
+            ],
+            "angle": 0,
+            "content": "Contrastive learning uses speech transcription corpora (i.e., speech paired with human transcripts). The speech input is passed through the upstream ASR model to obtain the ASR outputs. Given the noisy ASR output \\(\\bar{\\mathbf{x}}\\) paired with its corresponding clean human transcript \\(\\mathbf{x}\\), we minimize the contrastive objective \\(\\mathcal{L}_{\\mathrm{CTL}}\\), which is an average of two symmetric sentence-level contrastive loss functions."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.511,
+                0.725,
+                0.882,
+                0.759
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} \\text {G i v e n a b a t c h o f} n \\text {e x a m p l e s} (X, \\bar {X}) = \\\\ \\left\\{\\left(\\mathbf {x} ^ {(1)}, \\bar {\\mathbf {x}} ^ {(1)}\\right), \\dots , \\left(\\mathbf {x} ^ {(n)}, \\bar {\\mathbf {x}} ^ {(n)}\\right) \\right\\} \\end{array}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.536,
+                0.767,
+                0.883,
+                0.807
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {\\mathrm {C T L}} = \\frac {1}{2 n} \\sum_ {i = 1} ^ {n} \\left(\\mathcal {L} _ {c} \\left(\\mathbf {x} ^ {(i)}\\right) + \\mathcal {L} _ {c} \\left(\\bar {\\mathbf {x}} ^ {(i)}\\right)\\right) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.816,
+                0.56,
+                0.828
+            ],
+            "angle": 0,
+            "content": "where"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.514,
+                0.835,
+                0.877,
+                0.879
+            ],
+            "angle": 0,
+            "content": "\\[\n\\mathcal {L} _ {c} (\\mathbf {x}) = \\frac {- \\exp \\left(D (\\mathbf {s} _ {\\mathbf {x}} , \\mathbf {s} _ {\\bar {\\mathbf {x}}})\\right)}{\\exp \\left(D (\\mathbf {s} _ {\\mathbf {x}} , \\mathbf {s} _ {\\bar {\\mathbf {x}}})\\right) + \\sum_ {x ^ {\\prime}} \\exp (D (\\mathbf {s} _ {\\mathbf {x}}, \\hat {\\mathbf {s}} _ {\\mathbf {x} ^ {\\prime}}))}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.888,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(D(\\mathbf{u},\\mathbf{v})\\) denotes the cosine distance between two vectors \\(\\mathbf{u}\\) and \\(\\mathbf{v}\\). \\(\\hat{\\mathbf{s}}_{\\mathbf{x}'}\\) represents a negative example"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9050"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.135
+            ],
+            "angle": 0,
+            "content": "constructed for every other sentence \\(\\mathbf{x}'\\) in the batch. Given \\(\\mathbf{x}' \\in (X \\cup \\bar{X}) \\setminus \\{\\mathbf{x}, \\bar{\\mathbf{x}}\\}\\) and its sentence embedding \\(\\mathbf{s}_{x'}\\)"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.134,
+                0.159,
+                0.489,
+                0.201
+            ],
+            "angle": 0,
+            "content": "\\[\n\\hat {\\mathbf {s}} _ {\\mathbf {x} ^ {\\prime}} = \\left\\{ \\begin{array}{l l} \\mathbf {s} _ {\\mathbf {x}} + \\lambda_ {\\mathbf {x}} \\left(\\mathbf {s} _ {\\mathbf {x} ^ {\\prime}} - \\mathbf {s} _ {\\mathbf {x}}\\right) & \\text {i f ,} d _ {\\mathbf {x}} ^ {+} > d _ {\\mathbf {x}} ^ {-} \\\\ \\mathbf {s} _ {\\mathbf {x} ^ {\\prime}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.214,
+                0.49,
+                0.346
+            ],
+            "angle": 0,
+            "content": "where \\(d_{\\mathbf{x}}^{+} = \\| \\mathbf{s}_{\\mathbf{x}} - \\mathbf{s}_{\\overline{\\mathbf{x}}}\\|_{2}, d_{\\mathbf{x}}^{-} = \\| \\mathbf{s}_{\\mathbf{x}} - \\mathbf{s}_{\\mathbf{x}^{\\prime}}\\|_{2}\\), and \\(\\lambda_{\\mathbf{x}} = \\frac{d_{\\mathbf{x}}^{+}}{d_{\\mathbf{x}}^{-}}\\). The above linear interpolation in Eq 2 with exponentially decaying \\(\\lambda_{\\mathbf{x}}\\) is implemented following Wei et al. (2021) to increase the hardness of negative examples by not selecting uninformative negative examples as training progresses. \\(\\mathcal{L}_c(\\overline{\\mathbf{x}})\\) follows similar formulation as \\(\\mathcal{L}_c(\\mathbf{x})\\) with \\(\\mathbf{x}\\) replaced by \\(\\overline{\\mathbf{x}}\\), and vice versa."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.349,
+                0.49,
+                0.541
+            ],
+            "angle": 0,
+            "content": "To stabilize the NMT parameters before applying contrastive learning, we also use curriculum learning (Bengio et al., 2009). We first pre-train the model with NMT loss \\(\\mathcal{L}_{\\mathrm{NMT}}\\) only for the first \\(N\\) batches, followed by iteratively training alternate batches with \\(\\mathcal{L}_{\\mathrm{CTL}}\\) and \\(\\mathcal{L}_{\\mathrm{NMT}}\\) losses, respectively, until convergence. With this model and training regime, which alternates between training with translation and speech recognition datasets independently, we are able to build robust NMT models for cascaded ST without any requirement for speech translation data."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.555,
+                0.322,
+                0.572
+            ],
+            "angle": 0,
+            "content": "3 Experimental Setup"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.581,
+                0.486,
+                0.613
+            ],
+            "angle": 0,
+            "content": "We experiment with English-German (En-De) and English-French (En-Fr) language directions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.614,
+                0.49,
+                0.838
+            ],
+            "angle": 0,
+            "content": "Training Data: For parallel text translation data, we use WMT'16 En-De (Bojar et al., 2016) and WMT'14 En-Fr (Bojar et al., 2014) news translation task data for the corresponding language directions. To get the ASR data for contrastive learning, we use Open AI Whisper base ASR (Radford et al., 2022) to transcribe the speech in the training sets from Mozilla Common Voice 12.0 (Ardila et al., 2020), VoxPopuli (Wang et al., 2021), and the English set of Multilingual LibrSpeech (Pratap et al., 2020) for both language directions. As an additional experiment, we also try adding in-domain MuST-C training data. The full statistics of the datasets are in Appendix A.2 (Table 8)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.839,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Evaluation: We use MuST-C (Di Gangi et al., 2019) tst-COMMON as our test set and MuST-C dev as our validation set. For MT and ST tasks, the inputs to the model are the Whisper (base) ASR outputs and the corresponding human transcripts,"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.082,
+                0.883,
+                0.159
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>28.1</td><td>23.5</td><td>36.1</td><td>30.8</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>28.2</td><td>25.2</td><td>36.3</td><td>32.0</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>29.5</td><td>24.4</td><td>37.2</td><td>31.3</td></tr><tr><td>CLAD-ST</td><td>28.8</td><td>26.6</td><td>36.5</td><td>33.4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.168,
+                0.883,
+                0.196
+            ],
+            "angle": 0,
+            "content": "Table 2: Performance of various models on MT and ST tasks on MuST-C tst-COMMON."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.224,
+                0.885,
+                0.579
+            ],
+            "angle": 0,
+            "content": "respectively. We evaluate against the English and French reference translations for the respective translation directions. We use case-sensitive, detokenized BLEU computed using sacrebleu\\(^{1}\\) (Post, 2018). We also compute statistical significance using paired bootstrap resampling (Koehn, 2004). Models: We use a base Transformer (Vaswani et al., 2017) as our baseline NMT model, which we extend to support our proposed approach, Contrastive Learning with Adversarial Data for Speech Translation (CLAD-ST). We also compare against two recent state-of-the-art approaches for NMT robustness which we re-implement: (1) CONF-ASR (Wang et al., 2022) generates adversarial examples by substituting random source tokens with candidate tokens from ASR confusion set based on their embedding distances to the original token. (2) CSA-NMT (Wei et al., \\(2022)^{2}\\) uses a separate semantic encoder to model a semantic region around the source and target sentences. To improve NMT robustness, samples from this region are used to augment the NMT decoder representations."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.581,
+                0.884,
+                0.726
+            ],
+            "angle": 0,
+            "content": "We tokenize all datasets using a sentencepiece model (Kudo and Richardson, 2018) for each language pair on the text translation datasets with a shared source-target vocabulary size of 32,000. We use OpenAI Wispher (base) ASR as our main upstream model. We also evaluate the ST performance of our model using Whisper large ASR. Other training details and hyperparameters are in Appendix A.1."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.74,
+                0.724,
+                0.758
+            ],
+            "angle": 0,
+            "content": "4 Results and Analysis"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.769,
+                0.722,
+                0.784
+            ],
+            "angle": 0,
+            "content": "4.1 Overall Performance"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.885,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Our proposed CLAD-ST achieves 26.6 and 33.4 BLEU on the En-De and En-Fr ST tasks, significantly \\((p < 0.001)\\) outperforming the Transformer baseline by 3.1 BLEU points and 2.6 BLEU"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.868,
+                0.883,
+                0.894
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)https://github.com/mjpost/sacrebleu. Signature: nrefs:1|case:mixed|eff:no|tok:13a|smooth:exp"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.894,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)Code and network hyperparameters are from https://github.com/pemywei/csanmt"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.868,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.519,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9051"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.121,
+                0.086,
+                0.496,
+                0.254
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.267,
+                0.49,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Figure 2: BLEU scores compared to the number of token-level errors in the ASR outputs on En-De ST task on MuST-C tst-COMMON dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.346,
+                0.49,
+                0.522
+            ],
+            "angle": 0,
+            "content": "points (see Table 2). Our approach consistently performs better than other baselines (CONF-ASR and CSA-NMT), yielding gains of 1.4 BLEU points \\((p < 0.001)\\) on both language directions. The improvement of our approach over CONF-ASR shows the importance of contextual modeling of the ASR errors instead of simulating them with ASR confusion lattices. CSA-NMT achieves the largest improvement in the MT task but fails to specialize to ASR errors, yielding lower improvements compared to our approach."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.551,
+                0.251,
+                0.565
+            ],
+            "angle": 0,
+            "content": "4.2 Robustness"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.582,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In Figure 2, we report the performance of all the En-De models as the number of ASR errors increases in the input sequence. We group the ASR outputs based on the number of tokens different from the reference transcripts on the test set. The performance of the Transformer baseline model drastically drops as the number of ASR errors increases. CONF-ASR performs better than the Transformer baseline in all the ASR error conditions and closely with CLAD-ST on ASR outputs with one or two errors. However, CLAD-ST notably outperforms it as the number of errors increases. CSA-NMT performs slightly better than the baseline Transformer when the number of errors is few and performs slightly better than CONF-ASR when increasing the number of errors due to their semantic augmentation. Overall, our approach has shown to be robust against sequences having several ASR errors and performs consistently better than the other baselines, irrespective of the number of errors in the sequence."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.085,
+                0.729,
+                0.102
+            ],
+            "angle": 0,
+            "content": "4.3 Curriculum Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.106,
+                0.885,
+                0.187
+            ],
+            "angle": 0,
+            "content": "We study the effect of the curriculum learning strategy (Section 2) and reports the results in Table 3. We find that curriculum learning strategy brings an improvement of 0.5 BLEU on En-De and 0.7 BLEU On En-Fr ST tasks, respectively."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.543,
+                0.197,
+                0.851,
+                0.25
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>En-De</td><td>En-Fr</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>30.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>33.4</td></tr><tr><td>w/o Curriculum learning</td><td>26.1</td><td>32.7</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.259,
+                0.882,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Table 3: Performance on ST tasks with and without curriculum learning on MuST-C tst-COMMON."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.315,
+                0.749,
+                0.33
+            ],
+            "angle": 0,
+            "content": "4.4 Out-of-distribution ASR"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.336,
+                0.885,
+                0.497
+            ],
+            "angle": 0,
+            "content": "In Table 4, we report the performance of all the Ende systems using Whisper base as the ASR model during training and a higher-quality ASR (Whisper large) during testing. Although training on lower-quality ASR outputs with more errors, a similar BLEU score improvement is also noticed for all the systems when fed better ASR outputs. However, CLAD-ST achieves the largest improvements over the baseline, showing that it better generalizes over different ASR-quality outputs."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.508,
+                0.898,
+                0.596
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Model</td><td colspan=\"2\">En-De</td></tr><tr><td>Wbase (WER = 14.3)</td><td>Wlarge (WER = 7.6)</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>24.4</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>25.2</td><td>26.3</td></tr><tr><td>CSA-NMT (Wei et al., 2022)</td><td>24.4</td><td>25.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>27.4</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.606,
+                0.884,
+                0.663
+            ],
+            "angle": 0,
+            "content": "Table 4: BLEU scores on MuST-C En-De tst-COMMON ST task when using Whisper-large (Wlarge) vs. Whisper-base (Wbase) ASR. Word error rate (WER) of the ASR model is provided."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.69,
+                0.644,
+                0.705
+            ],
+            "angle": 0,
+            "content": "4.5 Model Size"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.711,
+                0.885,
+                0.854
+            ],
+            "angle": 0,
+            "content": "To investigate if our results scale when we increase the size of the network, we use Transformer-large instead of Transformer-base for our baseline models as well as CLAD-ST. The results are reported in Table 5. We find that we get improvements of 1.5 and 1.3 BLEU score points for the ST task on En-De and En-Fr, respectively, compared to CONF-ASR, demonstrating its advantage even on larger NMT architectures."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.867,
+                0.755,
+                0.882
+            ],
+            "angle": 0,
+            "content": "4.6 In-domain Training Data"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We also experiment with adding in-domain training data from MuST-C for training the NMT models in"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9052"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.082,
+                0.486,
+                0.158
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-big</td><td>29.6</td><td>24.9</td><td>38.7</td><td>33.3</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>29.8</td><td>26.6</td><td>38.8</td><td>34.6</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>31.1</td><td>25.7</td><td>39.9</td><td>33.7</td></tr><tr><td>CLAD-ST</td><td>30.3</td><td>28.1</td><td>39.1</td><td>35.9</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.168,
+                0.488,
+                0.211
+            ],
+            "angle": 0,
+            "content": "Table 5: Performance of using Transformer-large as the model architecture on MT and ST tasks on MuST-C tst-COMMON."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.238,
+                0.49,
+                0.368
+            ],
+            "angle": 0,
+            "content": "CLAD-ST and the baselines (see Table 6). In line with our expectations, we find a notable improvement in performance of all the systems. Despite the improved baselines, we observe similar improvements for CLAD-ST of 1.4 BLEU on both En-De and En-Fr ST tasks compared to CONF-ASR, and 2.4 and 2.2 BLEU on En-De and En-Fr ST tasks, respectively, compared to CSA-NMT."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.379,
+                0.486,
+                0.455
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">NMT</td><td colspan=\"2\">En-De</td><td colspan=\"2\">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>31.0</td><td>26.1</td><td>37.9</td><td>32.9</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>30.9</td><td>27.9</td><td>38.0</td><td>33.9</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>32.1</td><td>26.9</td><td>38.8</td><td>33.1</td></tr><tr><td>CLAD-ST</td><td>31.6</td><td>29.4</td><td>38.2</td><td>35.3</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.465,
+                0.49,
+                0.509
+            ],
+            "angle": 0,
+            "content": "Table 6: Performance of various models on MT and ST tasks on MuST-C tst-COMMON when adding indomain MuST-C training data."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.538,
+                0.271,
+                0.554
+            ],
+            "angle": 0,
+            "content": "5 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.566,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Several prior works tried to improve the robustness of MT models to noisy ASR inputs for cascaded speech translation. Sperber et al. (2017) randomly sampled synthetic noise in the form of insertions, deletions, and substitutions and injected them into the source side of MT training data. Li et al. (2018) encoded the syllables (Pinyin) for Chinese ST and injected synthetic character-level substitutions to simulate ASR errors. In contrast, Martucci et al. (2021) trained a lexical noise model using confusion sets extracted by aligning the actual ASR output to reference transcript to inject synthetic errors. Wang et al. (2022) picked substitution candidates from the ASR confusion set that are farthest away from the source token embeddings leading to more robust MT models. Cheng et al. (2019) also used the actual ASR output in training incorporating an additional discriminator network that tries to distinguish between human transcript and the ASR output following Cheng et al. (2018) to improve robustness. Padfield and Cherry (2021) simulated not only token-level errors but also segmentation errors"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.231
+            ],
+            "angle": 0,
+            "content": "and projected them on the target side as well. Bahar et al. (2021), on the other hand, proposes to jointly optimize the ASR and MT components in the cascaded system using speech-translation data. Similarly, Lam et al. (2021) also uses speech-translation data to reduce error propagation between the two components by mutually training the MT system on ASR n-best outputs and using cyclic feedback from MT outputs for self-training of the ASR model."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.241,
+                0.642,
+                0.256
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.266,
+                0.885,
+                0.492
+            ],
+            "angle": 0,
+            "content": "We improve the robustness of MT to ASR outputs using contrastive learning to bring the representations of clean and noisy examples closer in the semantic space. Our approach does not require any speech translation corpus. We significantly improve the translation accuracy on noisy ASR outputs without degrading translation accuracy on clean text. We also show that the approach is scalable to better-quality ASR models in the cascade other than the one used during training. The proposed approach is generic and is applicable beyond the context of speech translation alone, such as translating user-generated chat text or non-native text if paired noisy-clean data is available."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.503,
+                0.616,
+                0.518
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.528,
+                0.753,
+                0.544
+            ],
+            "angle": 0,
+            "content": "The limitations of our paper are:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.553,
+                0.885,
+                0.715
+            ],
+            "angle": 0,
+            "content": "- As in any ST cascade architecture, the performance of our MT system depends on the quality of the ASR model outputs. Since we use examples generated by the ASR to train the MT model, the quality of ASR outputs directly affects the model performance. In our work we tested two ASR systems having different quality (base: 14.3 WER and large: 7.6 WER). This is particularly relevant when English is not the source language side."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.724,
+                0.887,
+                0.755
+            ],
+            "angle": 0,
+            "content": "- Evaluation is done using only the BLEU score. We did not use human evaluation or COMET."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.532,
+                0.766,
+                0.885,
+                0.894
+            ],
+            "angle": 0,
+            "content": "- The evaluation is limited to two language pairs having English as the source. This selection of the source language is quite important because having a no-English language as a source will expose our MT model to a) ASR models of probably lower quality and b) different and more varied linguistic challenges that might affect the work of the adversarial method."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.532,
+                0.553,
+                0.887,
+                0.894
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9053"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.107,
+                0.49,
+                0.341
+            ],
+            "angle": 0,
+            "content": "Antonios Anastasopoulos, Loic Barrault, Luisa Bentivogli, Marcely Zanon Boito, Ondrej Bojar, Roldano Cattoni, Anna Currey, Georgiana Dinu, Kevin Duh, Maha Elbayad, Clara Emmanuel, Yannick Esteve, Marcello Federico, Christian Federmann, Souhir Gahbiche, Hongyu Gong, Roman Grundkiewicz, Barry Haddow, Benjamin Hsu, David Javorsky, Vera Kloudova, Surafel Lakew, Xutai Ma, Prashant Mathur, Paul McNamee, Kenton Murray, Maria Nadejde, Satoshi Nakamura, Matteo Negri, Jan Niehues, Xing Niu, John Ortega, Juan Pino, Elizabeth Salesky, Jiatong Shi, Matthias Sperber, Sebastian Stüker, Katsuhito Sudoh, Marco Turchi, Yogesh Virkar, Alexander Waibel, Changhan Wang, and Shinji Watanabe. 2022. Findings of the IWSLT 2022 evaluation campaign. In Proceedings of the 19th International Conference on Spoken Language Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.351,
+                0.49,
+                0.442
+            ],
+            "angle": 0,
+            "content": "Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henrtty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.452,
+                0.49,
+                0.519
+            ],
+            "angle": 0,
+            "content": "Parnia Bahar, Tobias Bieschke, Ralf Schlüter, and Hermann Ney. 2021. Tight integrated end-to-end training for cascaded speech translation. In Proceedings of the 2021 IEEE Spoken Language Technology Workshop."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.527,
+                0.49,
+                0.631
+            ],
+            "angle": 0,
+            "content": "Loic Barrault, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Muller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.641,
+                0.49,
+                0.693
+            ],
+            "angle": 0,
+            "content": "Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.703,
+                0.49,
+                0.756
+            ],
+            "angle": 0,
+            "content": "Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.765,
+                0.49,
+                0.844
+            ],
+            "angle": 0,
+            "content": "Alexandre Bérard, Olivier Pietquin, Laurent Besacier, and Christophe Servan. 2016. Listen and translate: A proof of concept for end-to-end speech-to-text translation. In Proceedings of the NIPS Workshop on end-to-end learning for speech and audio processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.852,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Ondrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia, and Ales Tamchyna. 2014. Findings of the 2014 workshop on"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.116,
+                0.107,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.113
+            ],
+            "angle": 0,
+            "content": "statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.123,
+                0.885,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Ondrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Aurélie Néveol, Mariana Neves, Martin Popel, Matt Post, Raphael Rubino, Carolina Scarton, Lucia Specia, Marco Turchi, Karin Verspoor, and Marcos Zampieri. 2016. Findings of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers, Berlin, Germany."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.278,
+                0.885,
+                0.343
+            ],
+            "angle": 0,
+            "content": "Qiao Cheng, Meiyuan Fan, Yaqian Han, Jin Huang, and Yitao Duan. 2019. Breaking the data barrier: Towards robust speech translation via adversarial stability training. In Proceedings of the 16th International Conference on Spoken Language Translation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.354,
+                0.885,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Yong Cheng, Lu Jiang, Wolfgang Macherey, and Jacob Eisenstein. 2020. AdvAug: Robust adversarial augmentation for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.43,
+                0.885,
+                0.484
+            ],
+            "angle": 0,
+            "content": "Yong Cheng, Zhaopeng Tu, Fandong Meng, Junjie Zhai, and Yang Liu. 2018. Towards robust neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.493,
+                0.885,
+                0.572
+            ],
+            "angle": 0,
+            "content": "Marta R Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.583,
+                0.885,
+                0.675
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.685,
+                0.885,
+                0.778
+            ],
+            "angle": 0,
+            "content": "Mattia A. Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019. MuST-C: a Multilingual Speech Translation Corpus. 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)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.788,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, Will Lewis, Mu Li, Shujie Liu, Tie-Yan Liu, Renqian Luo, Arul Menezes, Tao Qin, Frank Seide, Xu Tan, Fei Tian, Lijun Wu, Shuangzhi Wu, Yingce Xia, Dongdong Zhang, Zhirui Zhang, and Ming Zhou. 2018. Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.885,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9054"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.139
+            ],
+            "angle": 0,
+            "content": "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.152,
+                0.487,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.244,
+                0.487,
+                0.31
+            ],
+            "angle": 0,
+            "content": "Tsz Kin Lam, Shigehiko Schamoni, and Stefan Riezler. 2021. Cascaded models with cyclic feedback for direct speech translation. In Proceedings of the 2021 IEEE International Conference on Acoustics, Speech and Signal Processing."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.324,
+                0.487,
+                0.375
+            ],
+            "angle": 0,
+            "content": "Xiang Li, Haiyang Xue, Wei Chen, Yang Liu, Yang Feng, and Qun Liu. 2018. Improving the robustness of speech translation. arXiv preprint arXiv:1811.00728."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.389,
+                0.487,
+                0.442
+            ],
+            "angle": 0,
+            "content": "Giuseppe Martucci, Mauro Cettolo, Matteo Negri, and Marco Turchi. 2021. Lexical modeling of ASR errors for robust speech translation. In Proceedings of Interspeech."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.456,
+                0.487,
+                0.522
+            ],
+            "angle": 0,
+            "content": "Dirk Padfield and Colin Cherry. 2021. Inverted projection for robust speech translation. In Proceedings of the 18th International Conference on Spoken Language Translation. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.535,
+                0.486,
+                0.575
+            ],
+            "angle": 0,
+            "content": "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.588,
+                0.487,
+                0.64
+            ],
+            "angle": 0,
+            "content": "Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. arXiv preprint arXiv:2012.03411."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.654,
+                0.487,
+                0.707
+            ],
+            "angle": 0,
+            "content": "Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.72,
+                0.487,
+                0.773
+            ],
+            "angle": 0,
+            "content": "Motoki Sato, Jun Suzuki, and Shun Kiyono. 2019. *Effective adversarial regularization for neural machine translation*. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.786,
+                0.487,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. 2019. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.865,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Matthias Sperber, Jan Niehues, and Alex Waibel. 2017. Toward robust neural machine translation for noisy input sequences. In Proceedings of the International Workshop on Spoken Language Translation."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                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. In Advances in Neural Information Processing Systems, volume 30."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.163,
+                0.882,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. 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)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.292,
+                0.882,
+                0.397
+            ],
+            "angle": 0,
+            "content": "Shuaibo Wang, Yufeng Chen, Songming Zhang, Deyi Xiong, and Jinan Xu. 2022. Adversarily improving NMT robustness to ASR errors with confusion sets. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.408,
+                0.882,
+                0.461
+            ],
+            "angle": 0,
+            "content": "Xiangpeng Wei, Rongxiang Weng, Yue Hu, Luxi Xing, Heng Yu, and Weihua Luo. 2021. On learning universal representations across languages. In International Conference on Learning Representations."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.472,
+                0.882,
+                0.551
+            ],
+            "angle": 0,
+            "content": "Xiangpeng Wei, Heng Yu, Yue Hu, Rongxiang Weng, Weihua Luo, and Rong Jin. 2022. Learning to generalize to more: Continuous semantic augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.562,
+                0.882,
+                0.627
+            ],
+            "angle": 0,
+            "content": "Ron J. Weiss, Jan Chorowski, Navdeep Jaitly, Yonghui Wu, and Zhifeng Chen. 2017. Sequence-to-sequence models can directly translate foreign speech. In Proceedings of the Conference of the International Speech Communication Association."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.627
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.642,
+                0.75,
+                0.658
+            ],
+            "angle": 0,
+            "content": "A Implementation Details"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.668,
+                0.856,
+                0.684
+            ],
+            "angle": 0,
+            "content": "A.1 Training Details and Hyperparameters"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.51,
+                0.689,
+                0.884,
+                0.896
+            ],
+            "angle": 0,
+            "content": "We implement our approach and (Wang et al., 2022) using \\( \\text{Fairseq}^3 \\). For Wang et al. (2021), we randomly selected \\( 14\\% \\) tokens of sentences for a replacement to generate adversarial examples based on the ASR word error rate. All experiments are performed on A100 GPU, and gradient accumulation and batch size are set to 8 iterations and 8k tokens. For CLAD-ST, our NMT model is pretrained for the first 50K batches for NMT loss in the curriculum. For evaluation, the beam size and length penalty are set to 4 and 0.6 for both tasks. Hyperparameters for all the models are shown in Table 7."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.904,
+                0.805,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>https://github.com/facebookresearch/fairseq"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9055"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.24,
+                0.081,
+                0.761,
+                0.338
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Hyperparameter</td><td>Baseline</td><td>CONF-ASR</td><td>CSA-NMT</td><td>CLAD-ST</td></tr><tr><td>encoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>encoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>encoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>encoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>decoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>decoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>decoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>decoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>dropout</td><td></td><td>0.3</td><td></td><td></td></tr><tr><td>optimizer</td><td></td><td>adam</td><td></td><td></td></tr><tr><td>adam-β</td><td></td><td>(0.9, 0.98)</td><td></td><td></td></tr><tr><td>clip-norm</td><td></td><td>0.0</td><td></td><td></td></tr><tr><td>lr scheduler</td><td></td><td>invers-sqrt</td><td></td><td></td></tr><tr><td>learning rate</td><td></td><td>5e-4</td><td></td><td></td></tr><tr><td>warmup-updates</td><td></td><td>4000</td><td></td><td></td></tr><tr><td>label-smoothing</td><td></td><td>0.1</td><td></td><td></td></tr><tr><td>max tokens</td><td>12800</td><td>8192</td><td>6144</td><td>8192</td></tr><tr><td>num-initial-updates-nmt</td><td>N/A</td><td>50000</td><td>50000</td><td>50000</td></tr><tr><td>csanmt-semantic-samples</td><td>N/A</td><td>N/A</td><td>20</td><td>N/A</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.386,
+                0.348,
+                0.61,
+                0.362
+            ],
+            "angle": 0,
+            "content": "Table 7: Model hyperparameters"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.384,
+                0.472,
+                0.533
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Type</td><td>Name</td><td>Lang.</td><td>#Examples</td></tr><tr><td rowspan=\"4\">MT Data</td><td>WMT&#x27;16</td><td>En-De</td><td>4.5M</td></tr><tr><td>WMT&#x27;14</td><td>En-Fr</td><td>40 M</td></tr><tr><td>MuST-C</td><td>En-De</td><td>251K</td></tr><tr><td>train</td><td>En-Fr</td><td>275K</td></tr><tr><td rowspan=\"3\">ASR Data</td><td>VoxPopuli</td><td>En</td><td>0.18M</td></tr><tr><td>Commonvoice</td><td>En</td><td>0.99M</td></tr><tr><td>MLS</td><td>En</td><td>11M</td></tr><tr><td rowspan=\"2\">Dev</td><td rowspan=\"2\">MuST-C Dev</td><td>En-De</td><td>1415</td></tr><tr><td>En-Fr</td><td>1412</td></tr><tr><td rowspan=\"2\">Test</td><td>MuST-C</td><td>En-De</td><td>2580</td></tr><tr><td>tst-COMMON</td><td>En-Fr</td><td>2632</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.117,
+                0.542,
+                0.483,
+                0.557
+            ],
+            "angle": 0,
+            "content": "Table 8: Statistics of datasets used in the experiments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.582,
+                0.233,
+                0.596
+            ],
+            "angle": 0,
+            "content": "A.2 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.602,
+                0.49,
+                0.7
+            ],
+            "angle": 0,
+            "content": "Dataset statistics can be found in Table 8. WMT'16 En-De training data<sup>4</sup> includes Europarl, Common Crawl corpus, and News Commentary corpora. WMT'14 En-Fr<sup>5</sup> training data includes Europarl, Common Crawl, UN corpus, News Commentary, and Fr-En Gigaword corpus."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.868,
+                0.365,
+                0.892
+            ],
+            "angle": 0,
+            "content": "4https://www.statmt.org/wmt16/ translation-task.html"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.892,
+                0.365,
+                0.918
+            ],
+            "angle": 0,
+            "content": "\\(^{5}\\)https://www.statmt.org/wmt14/translation-task.html"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.114,
+                0.868,
+                0.365,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "9056"
+        }
+    ]
+]

2023/CLAD-ST_ Contrastive Learning with Adversarial Data for Robust Speech Translation/3c4dc6ba-99d3-4dfa-ade5-337f84d977db_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d344ef66bfea86cfe1d3dbf82560c6fcc18e62df8e57ba6187dbf5fb69e7df51
+size 458868

2023/CLAD-ST_ Contrastive Learning with Adversarial Data for Robust Speech Translation/full.md

@@ -0,0 +1,222 @@
+# CLAD-ST: Contrastive Learning with Adversarial Data for Robust Speech Translation
+
+Sathish Reddy Indurthi, Shamil Chollampatt, Ravi Agrawal, Marco Turchi
+
+Zoom Video Communications
+
+sathishreddy.indurthi@zoom.us, shamil.chollampatt@zoom.us,
+
+ravi.agrawal@zoom.us,marco.turchi@zoom.us
+
+# Abstract
+
+The cascaded approach continues to be the most popular choice for speech translation (ST). This approach consists of an automatic speech recognition (ASR) model and a machine translation (MT) model that are used in a pipeline to translate speech in one language to text in another language. MT models are often trained on well-formed text and therefore lack robustness while translating noisy ASR outputs in the cascaded approach, degrading the overall translation quality significantly. We address this robustness problem in downstream MT models by forcing the MT encoder to bring the representations of a noisy input closer to its clean version in the semantic space. This is achieved by introducing a contrastive learning method that leverages adversarial examples in the form of ASR outputs paired with their corresponding human transcripts to optimize the network parameters. In addition, a curriculum learning strategy is then used to stabilize the training by alternating the standard MT log-likelihood loss and the contrastive losses. Our approach achieves significant gains of up to 3 BLEU scores in English-German and English-French speech translation without hurting the translation quality on clean text.
+
+# 1 Introduction
+
+Neural machine translation (NMT) has made significant advancements over the past several years with claims of achieving 'human parity' (Hassan et al., 2018), and 'super-human performance' (Barrault et al., 2019). However, despite making tremendous progress in quality and coverage (Costa-jussa et al., 2022), NMT models have been identified to lack robustness in dealing with noisy inputs (Belinkov and Bisk, 2018).
+
+Robustness is especially important in cascaded speech translation (ST) systems, where an NMT model works on the output of the upstream automatic speech recognition (ASR) system. In this scenario, significant MT performance degradation has
+
+Human Transcript: I'm not sure that's wise, given the importance of the problem, but there's now the geoengineering discussion about: Should that be in the back pocket in case things happen faster, or this innovation goes a lot slower than we expect?
+
+ASR Output: I'm not sure that's why it's given the importance of the problem. But now that the geoengineering discussion about should that be in the back pocket in case things happen faster or this innovation goes a lot slower than we expect.
+
+Table 1: Example human transcript and OpenAI Whisper ASR output from the MuST-C En-De devset. The underlined text indicates erroneous phrases and italicized indicates punctuation errors.
+
+been measured due to  $i$  error propagation from the ASR and ii) the mismatch between training-testing condition as the NMT model is typically trained on well-formed text making it weak in dealing with noisy inputs. For these reasons, there has been significant effort towards building end-to-end ST models (Bérard et al., 2016; Weiss et al., 2017). However, due to the lack of sufficient speech translation datasets in many languages and the flexibility to independently optimize ASR and MT systems, cascaded systems continue to be a dominant approach (Anastasopoulos et al., 2022).
+
+Prior research has tried to tackle the robustness problem in NMT models independently by (1) synthetic noise injection (Belinkov and Bisk, 2018), including mimicking ASR errors for the ST task (Sperber et al., 2017; Li et al., 2018; Martucci et al., 2021; Wang et al., 2022) or (2) by adding noise to the continuous-space representations learned by the NMT model (Sato et al., 2019; Cheng et al., 2020; Wei et al., 2022). Some of the synthetic noise injection methods replace words with probable words from ASR confusion lattices (Martucci et al., 2021; Wang et al., 2022) but fail to consider the full context and to realistically replicate ASR errors, like the phrase-level errors and segmentation errors that ASR models often make (see Table 1). On the other
+
+![](images/c1dcfeffafb9c5b1034ea7a4c031f9c6f73334a996f6901650f569b6520acb62.jpg)
+
+![](images/929bdd23670515232314d61d779cee3a1e7e69db6801a0f2fa31b9f60adbbbe1.jpg)  
+Figure 1: Proposed model and the two alternating training objectives.  $\mathcal{L}_{\mathrm{NMT}}$  is the standard NMT negative log-likelihood loss and  $\mathcal{L}_{\mathrm{CTL}}$  is the contrastive loss for improving robustness to ASR outputs.
+
+hand, the latter approaches predominantly generate arbitrary or random error distributions and overlook the real-world error signals from the actual upstream ASR model.
+
+To address the robustness problem, particularly in the context of cascaded ST, we propose to combine the best of the two approaches. This is obtained by training the NMT model with adversarial examples generated from the ASR outputs and encouraging the encoder representations of both the ASR outputs and their corresponding human transcripts to be closer to each other. This is done via contrastive learning (Saunshi et al., 2019), in addition to the standard log-likelihood NMT training, and through curriculum learning (Bengio et al., 2009) by pretraining on log-likelihood loss followed by iteratively alternating between the two objectives. Our method leads to significant improvements in the ST task on two language pairs (En-De, En-Fr) without hurting the performance on MT of well-formed texts. Our approach also follows the standard cascade approach that does not rely on any end-to-end ST data (speech aligned with their transcription and translation) unlike that of Bahar et al. (2021) and Lam et al. (2021), making it easier to apply to a wider range of language pairs.
+
+# 2 Method
+
+Our NMT model is a Transformer model (Vaswani et al., 2017) with several encoder and decoder layers, with self-attention modules within each block and a cross-attention module to link the
+
+encoder and decoder components. Training the network parameters  $\Theta$  is done by minimizing the mean negative log-likelihood over a batch of  $n$  source-target text translation pairs  $(X,Y) = (\mathbf{x}^{(1)},\mathbf{y}^{(1)}),\ldots ,(\mathbf{x}^{(n)},\mathbf{y}^{(n)})$ . This objective encourages producing a target sentence  $\mathbf{y}$  with a sequence of tokens  $y_{1},y_{2},\ldots ,y_{|\mathbf{y}|}$  given the source sentence  $\mathbf{x}$  with sequence of tokens  $x_{1},x_{2},\ldots ,x_{|\mathbf{x}|}$ .
+
+$$
+\mathcal {L} _ {\mathrm {N M T}} = - \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {| \mathbf {y} ^ {(i)} |} \log P (y _ {j} ^ {(i)} | \mathbf {x} ^ {(i)}, \mathbf {y} _ {<   j} ^ {(i)}, \Theta)
+$$
+
+To improve the robustness of the NMT model on noisy ASR outputs for cascaded speech translation, we use a contrastive learning method (Wei et al., 2021) that is aimed to bring the representations of a noisy sentence generated by the upstream ASR closer to its clean version (the human transcript for the same audio) in the semantic space modeled by the NMT encoder.
+
+To get the encoder sentence representations efficiently for contrastive learning, a [CLS] token is prepended to the input sentences similar to the BERT model (Devlin et al., 2019). We found that using encoder output corresponding to the [CLS] token is slightly better than using the mean of encoder outputs as the sentence representation.  $\mathbf{s}_{\mathbf{x}}$  which also gets added to each of the decoder outputs before softmax (see Figure 1).
+
+Contrastive learning uses speech transcription corpora (i.e., speech paired with human transcripts). The speech input is passed through the upstream ASR model to obtain the ASR outputs. Given the noisy ASR output  $\bar{\mathbf{x}}$  paired with its corresponding clean human transcript  $\mathbf{x}$ , we minimize the contrastive objective  $\mathcal{L}_{\mathrm{CTL}}$ , which is an average of two symmetric sentence-level contrastive loss functions.
+
+$$
+\begin{array}{l} \text {G i v e n a b a t c h o f} n \text {e x a m p l e s} (X, \bar {X}) = \\ \left\{\left(\mathbf {x} ^ {(1)}, \bar {\mathbf {x}} ^ {(1)}\right), \dots , \left(\mathbf {x} ^ {(n)}, \bar {\mathbf {x}} ^ {(n)}\right) \right\} \end{array}
+$$
+
+$$
+\mathcal {L} _ {\mathrm {C T L}} = \frac {1}{2 n} \sum_ {i = 1} ^ {n} \left(\mathcal {L} _ {c} \left(\mathbf {x} ^ {(i)}\right) + \mathcal {L} _ {c} \left(\bar {\mathbf {x}} ^ {(i)}\right)\right) \tag {1}
+$$
+
+where
+
+$$
+\mathcal {L} _ {c} (\mathbf {x}) = \frac {- \exp \left(D (\mathbf {s} _ {\mathbf {x}} , \mathbf {s} _ {\bar {\mathbf {x}}})\right)}{\exp \left(D (\mathbf {s} _ {\mathbf {x}} , \mathbf {s} _ {\bar {\mathbf {x}}})\right) + \sum_ {x ^ {\prime}} \exp (D (\mathbf {s} _ {\mathbf {x}}, \hat {\mathbf {s}} _ {\mathbf {x} ^ {\prime}}))}
+$$
+
+$D(\mathbf{u},\mathbf{v})$  denotes the cosine distance between two vectors  $\mathbf{u}$  and  $\mathbf{v}$ .  $\hat{\mathbf{s}}_{\mathbf{x}'}$  represents a negative example
+
+constructed for every other sentence  $\mathbf{x}'$  in the batch. Given  $\mathbf{x}' \in (X \cup \bar{X}) \setminus \{\mathbf{x}, \bar{\mathbf{x}}\}$  and its sentence embedding  $\mathbf{s}_{x'}$
+
+$$
+\hat {\mathbf {s}} _ {\mathbf {x} ^ {\prime}} = \left\{ \begin{array}{l l} \mathbf {s} _ {\mathbf {x}} + \lambda_ {\mathbf {x}} \left(\mathbf {s} _ {\mathbf {x} ^ {\prime}} - \mathbf {s} _ {\mathbf {x}}\right) & \text {i f ,} d _ {\mathbf {x}} ^ {+} > d _ {\mathbf {x}} ^ {-} \\ \mathbf {s} _ {\mathbf {x} ^ {\prime}} & \text {o t h e r w i s e} \end{array} \right. \tag {2}
+$$
+
+where  $d_{\mathbf{x}}^{+} = \| \mathbf{s}_{\mathbf{x}} - \mathbf{s}_{\overline{\mathbf{x}}}\|_{2}, d_{\mathbf{x}}^{-} = \| \mathbf{s}_{\mathbf{x}} - \mathbf{s}_{\mathbf{x}^{\prime}}\|_{2}$ , and  $\lambda_{\mathbf{x}} = \frac{d_{\mathbf{x}}^{+}}{d_{\mathbf{x}}^{-}}$ . The above linear interpolation in Eq 2 with exponentially decaying  $\lambda_{\mathbf{x}}$  is implemented following Wei et al. (2021) to increase the hardness of negative examples by not selecting uninformative negative examples as training progresses.  $\mathcal{L}_c(\overline{\mathbf{x}})$  follows similar formulation as  $\mathcal{L}_c(\mathbf{x})$  with  $\mathbf{x}$  replaced by  $\overline{\mathbf{x}}$ , and vice versa.
+
+To stabilize the NMT parameters before applying contrastive learning, we also use curriculum learning (Bengio et al., 2009). We first pre-train the model with NMT loss  $\mathcal{L}_{\mathrm{NMT}}$  only for the first  $N$  batches, followed by iteratively training alternate batches with  $\mathcal{L}_{\mathrm{CTL}}$  and  $\mathcal{L}_{\mathrm{NMT}}$  losses, respectively, until convergence. With this model and training regime, which alternates between training with translation and speech recognition datasets independently, we are able to build robust NMT models for cascaded ST without any requirement for speech translation data.
+
+# 3 Experimental Setup
+
+We experiment with English-German (En-De) and English-French (En-Fr) language directions.
+
+Training Data: For parallel text translation data, we use WMT'16 En-De (Bojar et al., 2016) and WMT'14 En-Fr (Bojar et al., 2014) news translation task data for the corresponding language directions. To get the ASR data for contrastive learning, we use Open AI Whisper base ASR (Radford et al., 2022) to transcribe the speech in the training sets from Mozilla Common Voice 12.0 (Ardila et al., 2020), VoxPopuli (Wang et al., 2021), and the English set of Multilingual LibrSpeech (Pratap et al., 2020) for both language directions. As an additional experiment, we also try adding in-domain MuST-C training data. The full statistics of the datasets are in Appendix A.2 (Table 8).
+
+Evaluation: We use MuST-C (Di Gangi et al., 2019) tst-COMMON as our test set and MuST-C dev as our validation set. For MT and ST tasks, the inputs to the model are the Whisper (base) ASR outputs and the corresponding human transcripts,
+
+<table><tr><td rowspan="2">NMT</td><td colspan="2">En-De</td><td colspan="2">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>28.1</td><td>23.5</td><td>36.1</td><td>30.8</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>28.2</td><td>25.2</td><td>36.3</td><td>32.0</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>29.5</td><td>24.4</td><td>37.2</td><td>31.3</td></tr><tr><td>CLAD-ST</td><td>28.8</td><td>26.6</td><td>36.5</td><td>33.4</td></tr></table>
+
+Table 2: Performance of various models on MT and ST tasks on MuST-C tst-COMMON.
+
+respectively. We evaluate against the English and French reference translations for the respective translation directions. We use case-sensitive, detokenized BLEU computed using sacrebleu $^{1}$  (Post, 2018). We also compute statistical significance using paired bootstrap resampling (Koehn, 2004). Models: We use a base Transformer (Vaswani et al., 2017) as our baseline NMT model, which we extend to support our proposed approach, Contrastive Learning with Adversarial Data for Speech Translation (CLAD-ST). We also compare against two recent state-of-the-art approaches for NMT robustness which we re-implement: (1) CONF-ASR (Wang et al., 2022) generates adversarial examples by substituting random source tokens with candidate tokens from ASR confusion set based on their embedding distances to the original token. (2) CSA-NMT (Wei et al.,  $2022)^{2}$  uses a separate semantic encoder to model a semantic region around the source and target sentences. To improve NMT robustness, samples from this region are used to augment the NMT decoder representations.
+
+We tokenize all datasets using a sentencepiece model (Kudo and Richardson, 2018) for each language pair on the text translation datasets with a shared source-target vocabulary size of 32,000. We use OpenAI Wispher (base) ASR as our main upstream model. We also evaluate the ST performance of our model using Whisper large ASR. Other training details and hyperparameters are in Appendix A.1.
+
+# 4 Results and Analysis
+
+# 4.1 Overall Performance
+
+Our proposed CLAD-ST achieves 26.6 and 33.4 BLEU on the En-De and En-Fr ST tasks, significantly  $(p < 0.001)$  outperforming the Transformer baseline by 3.1 BLEU points and 2.6 BLEU
+
+![](images/9e15bacd98b851205922f32a57bab18ced7f07b9e074648ab0bf4ca54a282a0d.jpg)  
+Figure 2: BLEU scores compared to the number of token-level errors in the ASR outputs on En-De ST task on MuST-C tst-COMMON dataset.
+
+points (see Table 2). Our approach consistently performs better than other baselines (CONF-ASR and CSA-NMT), yielding gains of 1.4 BLEU points  $(p < 0.001)$  on both language directions. The improvement of our approach over CONF-ASR shows the importance of contextual modeling of the ASR errors instead of simulating them with ASR confusion lattices. CSA-NMT achieves the largest improvement in the MT task but fails to specialize to ASR errors, yielding lower improvements compared to our approach.
+
+# 4.2 Robustness
+
+In Figure 2, we report the performance of all the En-De models as the number of ASR errors increases in the input sequence. We group the ASR outputs based on the number of tokens different from the reference transcripts on the test set. The performance of the Transformer baseline model drastically drops as the number of ASR errors increases. CONF-ASR performs better than the Transformer baseline in all the ASR error conditions and closely with CLAD-ST on ASR outputs with one or two errors. However, CLAD-ST notably outperforms it as the number of errors increases. CSA-NMT performs slightly better than the baseline Transformer when the number of errors is few and performs slightly better than CONF-ASR when increasing the number of errors due to their semantic augmentation. Overall, our approach has shown to be robust against sequences having several ASR errors and performs consistently better than the other baselines, irrespective of the number of errors in the sequence.
+
+# 4.3 Curriculum Learning
+
+We study the effect of the curriculum learning strategy (Section 2) and reports the results in Table 3. We find that curriculum learning strategy brings an improvement of 0.5 BLEU on En-De and 0.7 BLEU On En-Fr ST tasks, respectively.
+
+<table><tr><td>Model</td><td>En-De</td><td>En-Fr</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>30.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>33.4</td></tr><tr><td>w/o Curriculum learning</td><td>26.1</td><td>32.7</td></tr></table>
+
+# 4.4 Out-of-distribution ASR
+
+In Table 4, we report the performance of all the Ende systems using Whisper base as the ASR model during training and a higher-quality ASR (Whisper large) during testing. Although training on lower-quality ASR outputs with more errors, a similar BLEU score improvement is also noticed for all the systems when fed better ASR outputs. However, CLAD-ST achieves the largest improvements over the baseline, showing that it better generalizes over different ASR-quality outputs.
+
+Table 3: Performance on ST tasks with and without curriculum learning on MuST-C tst-COMMON.  
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">En-De</td></tr><tr><td>Wbase (WER = 14.3)</td><td>Wlarge (WER = 7.6)</td></tr><tr><td>Transformer-base</td><td>23.5</td><td>24.4</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>25.2</td><td>26.3</td></tr><tr><td>CSA-NMT (Wei et al., 2022)</td><td>24.4</td><td>25.8</td></tr><tr><td>CLAD-ST</td><td>26.6</td><td>27.4</td></tr></table>
+
+Table 4: BLEU scores on MuST-C En-De tst-COMMON ST task when using Whisper-large (Wlarge) vs. Whisper-base (Wbase) ASR. Word error rate (WER) of the ASR model is provided.
+
+# 4.5 Model Size
+
+To investigate if our results scale when we increase the size of the network, we use Transformer-large instead of Transformer-base for our baseline models as well as CLAD-ST. The results are reported in Table 5. We find that we get improvements of 1.5 and 1.3 BLEU score points for the ST task on En-De and En-Fr, respectively, compared to CONF-ASR, demonstrating its advantage even on larger NMT architectures.
+
+# 4.6 In-domain Training Data
+
+We also experiment with adding in-domain training data from MuST-C for training the NMT models in
+
+<table><tr><td rowspan="2">NMT</td><td colspan="2">En-De</td><td colspan="2">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-big</td><td>29.6</td><td>24.9</td><td>38.7</td><td>33.3</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>29.8</td><td>26.6</td><td>38.8</td><td>34.6</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>31.1</td><td>25.7</td><td>39.9</td><td>33.7</td></tr><tr><td>CLAD-ST</td><td>30.3</td><td>28.1</td><td>39.1</td><td>35.9</td></tr></table>
+
+CLAD-ST and the baselines (see Table 6). In line with our expectations, we find a notable improvement in performance of all the systems. Despite the improved baselines, we observe similar improvements for CLAD-ST of 1.4 BLEU on both En-De and En-Fr ST tasks compared to CONF-ASR, and 2.4 and 2.2 BLEU on En-De and En-Fr ST tasks, respectively, compared to CSA-NMT.
+
+Table 5: Performance of using Transformer-large as the model architecture on MT and ST tasks on MuST-C tst-COMMON.  
+
+<table><tr><td rowspan="2">NMT</td><td colspan="2">En-De</td><td colspan="2">En-Fr</td></tr><tr><td>MT</td><td>ST</td><td>MT</td><td>ST</td></tr><tr><td>Transformer-base</td><td>31.0</td><td>26.1</td><td>37.9</td><td>32.9</td></tr><tr><td>CONF-ASR (Wang et al., 2022)</td><td>30.9</td><td>27.9</td><td>38.0</td><td>33.9</td></tr><tr><td>CSA-NMT (Wei et al., 2021)</td><td>32.1</td><td>26.9</td><td>38.8</td><td>33.1</td></tr><tr><td>CLAD-ST</td><td>31.6</td><td>29.4</td><td>38.2</td><td>35.3</td></tr></table>
+
+Table 6: Performance of various models on MT and ST tasks on MuST-C tst-COMMON when adding indomain MuST-C training data.
+
+# 5 Related Work
+
+Several prior works tried to improve the robustness of MT models to noisy ASR inputs for cascaded speech translation. Sperber et al. (2017) randomly sampled synthetic noise in the form of insertions, deletions, and substitutions and injected them into the source side of MT training data. Li et al. (2018) encoded the syllables (Pinyin) for Chinese ST and injected synthetic character-level substitutions to simulate ASR errors. In contrast, Martucci et al. (2021) trained a lexical noise model using confusion sets extracted by aligning the actual ASR output to reference transcript to inject synthetic errors. Wang et al. (2022) picked substitution candidates from the ASR confusion set that are farthest away from the source token embeddings leading to more robust MT models. Cheng et al. (2019) also used the actual ASR output in training incorporating an additional discriminator network that tries to distinguish between human transcript and the ASR output following Cheng et al. (2018) to improve robustness. Padfield and Cherry (2021) simulated not only token-level errors but also segmentation errors
+
+and projected them on the target side as well. Bahar et al. (2021), on the other hand, proposes to jointly optimize the ASR and MT components in the cascaded system using speech-translation data. Similarly, Lam et al. (2021) also uses speech-translation data to reduce error propagation between the two components by mutually training the MT system on ASR n-best outputs and using cyclic feedback from MT outputs for self-training of the ASR model.
+
+# 6 Conclusion
+
+We improve the robustness of MT to ASR outputs using contrastive learning to bring the representations of clean and noisy examples closer in the semantic space. Our approach does not require any speech translation corpus. We significantly improve the translation accuracy on noisy ASR outputs without degrading translation accuracy on clean text. We also show that the approach is scalable to better-quality ASR models in the cascade other than the one used during training. The proposed approach is generic and is applicable beyond the context of speech translation alone, such as translating user-generated chat text or non-native text if paired noisy-clean data is available.
+
+# Limitations
+
+The limitations of our paper are:
+
+- As in any ST cascade architecture, the performance of our MT system depends on the quality of the ASR model outputs. Since we use examples generated by the ASR to train the MT model, the quality of ASR outputs directly affects the model performance. In our work we tested two ASR systems having different quality (base: 14.3 WER and large: 7.6 WER). This is particularly relevant when English is not the source language side.  
+- Evaluation is done using only the BLEU score. We did not use human evaluation or COMET.  
+- The evaluation is limited to two language pairs having English as the source. This selection of the source language is quite important because having a no-English language as a source will expose our MT model to a) ASR models of probably lower quality and b) different and more varied linguistic challenges that might affect the work of the adversarial method.
+
+# References
+
+Antonios Anastasopoulos, Loic Barrault, Luisa Bentivogli, Marcely Zanon Boito, Ondrej Bojar, Roldano Cattoni, Anna Currey, Georgiana Dinu, Kevin Duh, Maha Elbayad, Clara Emmanuel, Yannick Esteve, Marcello Federico, Christian Federmann, Souhir Gahbiche, Hongyu Gong, Roman Grundkiewicz, Barry Haddow, Benjamin Hsu, David Javorsky, Vera Kloudova, Surafel Lakew, Xutai Ma, Prashant Mathur, Paul McNamee, Kenton Murray, Maria Nadejde, Satoshi Nakamura, Matteo Negri, Jan Niehues, Xing Niu, John Ortega, Juan Pino, Elizabeth Salesky, Jiatong Shi, Matthias Sperber, Sebastian Stüker, Katsuhito Sudoh, Marco Turchi, Yogesh Virkar, Alexander Waibel, Changhan Wang, and Shinji Watanabe. 2022. Findings of the IWSLT 2022 evaluation campaign. In Proceedings of the 19th International Conference on Spoken Language Translation.  
+Rosana Ardila, Megan Branson, Kelly Davis, Michael Kohler, Josh Meyer, Michael Henrtty, Reuben Morais, Lindsay Saunders, Francis Tyers, and Gregor Weber. 2020. Common voice: A massively-multilingual speech corpus. In Proceedings of the Twelfth Language Resources and Evaluation Conference.  
+Parnia Bahar, Tobias Bieschke, Ralf Schlüter, and Hermann Ney. 2021. Tight integrated end-to-end training for cascaded speech translation. In Proceedings of the 2021 IEEE Spoken Language Technology Workshop.  
+Loic Barrault, Ondrej Bojar, Marta R. Costa-jussa, Christian Federmann, Mark Fishel, Yvette Graham, Barry Haddow, Matthias Huck, Philipp Koehn, Shervin Malmasi, Christof Monz, Mathias Muller, Santanu Pal, Matt Post, and Marcos Zampieri. 2019. Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation.  
+Yonatan Belinkov and Yonatan Bisk. 2018. Synthetic and natural noise both break neural machine translation. In International Conference on Learning Representations.  
+Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning.  
+Alexandre Bérard, Olivier Pietquin, Laurent Besacier, and Christophe Servan. 2016. Listen and translate: A proof of concept for end-to-end speech-to-text translation. In Proceedings of the NIPS Workshop on end-to-end learning for speech and audio processing.  
+Ondrej Bojar, Christian Buck, Christian Federmann, Barry Haddow, Philipp Koehn, Johannes Leveling, Christof Monz, Pavel Pecina, Matt Post, Herve Saint-Amand, Radu Soricut, Lucia Specia, and Ales Tamchyna. 2014. Findings of the 2014 workshop on
+
+statistical machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation.  
+Ondrej Bojar, Rajen Chatterjee, Christian Federmann, Yvette Graham, Barry Haddow, Matthias Huck, Antonio Jimeno Yepes, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Aurélie Néveol, Mariana Neves, Martin Popel, Matt Post, Raphael Rubino, Carolina Scarton, Lucia Specia, Marco Turchi, Karin Verspoor, and Marcos Zampieri. 2016. Findings of the 2016 conference on machine translation. In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Task Papers, Berlin, Germany.  
+Qiao Cheng, Meiyuan Fan, Yaqian Han, Jin Huang, and Yitao Duan. 2019. Breaking the data barrier: Towards robust speech translation via adversarial stability training. In Proceedings of the 16th International Conference on Spoken Language Translation.  
+Yong Cheng, Lu Jiang, Wolfgang Macherey, and Jacob Eisenstein. 2020. AdvAug: Robust adversarial augmentation for neural machine translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics.  
+Yong Cheng, Zhaopeng Tu, Fandong Meng, Junjie Zhai, and Yang Liu. 2018. Towards robust neural machine translation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics.  
+Marta R Costa-jussà, James Cross, Onur Celebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.  
+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.  
+Mattia A. Di Gangi, Roldano Cattoni, Luisa Bentivogli, Matteo Negri, and Marco Turchi. 2019. MuST-C: a Multilingual Speech Translation Corpus. 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).  
+Hany Hassan, Anthony Aue, Chang Chen, Vishal Chowdhary, Jonathan Clark, Christian Federmann, Xuedong Huang, Marcin Junczys-Dowmunt, Will Lewis, Mu Li, Shujie Liu, Tie-Yan Liu, Renqian Luo, Arul Menezes, Tao Qin, Frank Seide, Xu Tan, Fei Tian, Lijun Wu, Shuangzhi Wu, Yingce Xia, Dongdong Zhang, Zhirui Zhang, and Ming Zhou. 2018. Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567.
+
+Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing.  
+Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations.  
+Tsz Kin Lam, Shigehiko Schamoni, and Stefan Riezler. 2021. Cascaded models with cyclic feedback for direct speech translation. In Proceedings of the 2021 IEEE International Conference on Acoustics, Speech and Signal Processing.  
+Xiang Li, Haiyang Xue, Wei Chen, Yang Liu, Yang Feng, and Qun Liu. 2018. Improving the robustness of speech translation. arXiv preprint arXiv:1811.00728.  
+Giuseppe Martucci, Mauro Cettolo, Matteo Negri, and Marco Turchi. 2021. Lexical modeling of ASR errors for robust speech translation. In Proceedings of Interspeech.  
+Dirk Padfield and Colin Cherry. 2021. Inverted projection for robust speech translation. In Proceedings of the 18th International Conference on Spoken Language Translation. Association for Computational Linguistics.  
+Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on Machine Translation: Research Papers.  
+Vineel Pratap, Qiantong Xu, Anuroop Sriram, Gabriel Synnaeve, and Ronan Collobert. 2020. Mls: A large-scale multilingual dataset for speech research. arXiv preprint arXiv:2012.03411.  
+Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356.  
+Motoki Sato, Jun Suzuki, and Shun Kiyono. 2019. *Effective adversarial regularization for neural machine translation*. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics.  
+Nikunj Saunshi, Orestis Plevrakis, Sanjeev Arora, Mikhail Khodak, and Hrishikesh Khandeparkar. 2019. A theoretical analysis of contrastive unsupervised representation learning. In International Conference on Machine Learning.  
+Matthias Sperber, Jan Niehues, and Alex Waibel. 2017. Toward robust neural machine translation for noisy input sequences. In Proceedings of the International Workshop on Spoken Language Translation.
+
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.  
+Changhan Wang, Morgane Riviere, Ann Lee, Anne Wu, Chaitanya Talnikar, Daniel Haziza, Mary Williamson, Juan Pino, and Emmanuel Dupoux. 2021. VoxPopuli: A large-scale multilingual speech corpus for representation learning, semi-supervised learning and interpretation. 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).  
+Shuaibo Wang, Yufeng Chen, Songming Zhang, Deyi Xiong, and Jinan Xu. 2022. Adversarily improving NMT robustness to ASR errors with confusion sets. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 2: Short Papers).  
+Xiangpeng Wei, Rongxiang Weng, Yue Hu, Luxi Xing, Heng Yu, and Weihua Luo. 2021. On learning universal representations across languages. In International Conference on Learning Representations.  
+Xiangpeng Wei, Heng Yu, Yue Hu, Rongxiang Weng, Weihua Luo, and Rong Jin. 2022. Learning to generalize to more: Continuous semantic augmentation for neural machine translation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.  
+Ron J. Weiss, Jan Chorowski, Navdeep Jaitly, Yonghui Wu, and Zhifeng Chen. 2017. Sequence-to-sequence models can directly translate foreign speech. In Proceedings of the Conference of the International Speech Communication Association.
+
+# A Implementation Details
+
+# A.1 Training Details and Hyperparameters
+
+We implement our approach and (Wang et al., 2022) using  $\text{Fairseq}^3$ . For Wang et al. (2021), we randomly selected  $14\%$  tokens of sentences for a replacement to generate adversarial examples based on the ASR word error rate. All experiments are performed on A100 GPU, and gradient accumulation and batch size are set to 8 iterations and 8k tokens. For CLAD-ST, our NMT model is pretrained for the first 50K batches for NMT loss in the curriculum. For evaluation, the beam size and length penalty are set to 4 and 0.6 for both tasks. Hyperparameters for all the models are shown in Table 7.
+
+<table><tr><td>Hyperparameter</td><td>Baseline</td><td>CONF-ASR</td><td>CSA-NMT</td><td>CLAD-ST</td></tr><tr><td>encoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>encoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>encoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>encoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>decoder layers</td><td></td><td>6</td><td></td><td></td></tr><tr><td>decoder embed dim</td><td></td><td>256</td><td></td><td></td></tr><tr><td>decoder ffn embed dim</td><td></td><td>2048</td><td></td><td></td></tr><tr><td>decoder attention heads</td><td></td><td>8</td><td></td><td></td></tr><tr><td>dropout</td><td></td><td>0.3</td><td></td><td></td></tr><tr><td>optimizer</td><td></td><td>adam</td><td></td><td></td></tr><tr><td>adam-β</td><td></td><td>(0.9, 0.98)</td><td></td><td></td></tr><tr><td>clip-norm</td><td></td><td>0.0</td><td></td><td></td></tr><tr><td>lr scheduler</td><td></td><td>invers-sqrt</td><td></td><td></td></tr><tr><td>learning rate</td><td></td><td>5e-4</td><td></td><td></td></tr><tr><td>warmup-updates</td><td></td><td>4000</td><td></td><td></td></tr><tr><td>label-smoothing</td><td></td><td>0.1</td><td></td><td></td></tr><tr><td>max tokens</td><td>12800</td><td>8192</td><td>6144</td><td>8192</td></tr><tr><td>num-initial-updates-nmt</td><td>N/A</td><td>50000</td><td>50000</td><td>50000</td></tr><tr><td>csanmt-semantic-samples</td><td>N/A</td><td>N/A</td><td>20</td><td>N/A</td></tr></table>
+
+Table 7: Model hyperparameters  
+
+<table><tr><td>Type</td><td>Name</td><td>Lang.</td><td>#Examples</td></tr><tr><td rowspan="4">MT Data</td><td>WMT&#x27;16</td><td>En-De</td><td>4.5M</td></tr><tr><td>WMT&#x27;14</td><td>En-Fr</td><td>40 M</td></tr><tr><td>MuST-C</td><td>En-De</td><td>251K</td></tr><tr><td>train</td><td>En-Fr</td><td>275K</td></tr><tr><td rowspan="3">ASR Data</td><td>VoxPopuli</td><td>En</td><td>0.18M</td></tr><tr><td>Commonvoice</td><td>En</td><td>0.99M</td></tr><tr><td>MLS</td><td>En</td><td>11M</td></tr><tr><td rowspan="2">Dev</td><td rowspan="2">MuST-C Dev</td><td>En-De</td><td>1415</td></tr><tr><td>En-Fr</td><td>1412</td></tr><tr><td rowspan="2">Test</td><td>MuST-C</td><td>En-De</td><td>2580</td></tr><tr><td>tst-COMMON</td><td>En-Fr</td><td>2632</td></tr></table>
+
+Table 8: Statistics of datasets used in the experiments.
+
+# A.2 Datasets
+
+Dataset statistics can be found in Table 8. WMT'16 En-De training data<sup>4</sup> includes Europarl, Common Crawl corpus, and News Commentary corpora. WMT'14 En-Fr<sup>5</sup> training data includes Europarl, Common Crawl, UN corpus, News Commentary, and Fr-En Gigaword corpus.

2023/CLAD-ST_ Contrastive Learning with Adversarial Data for Robust Speech Translation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3492e9cc9b5e8f39705664f3c7139f8adbd5dafca6b8353851e850a23d3a1ffc
+size 275364

2023/CLAIR_ Evaluating Image Captions with Large Language Models/ed931ad6-a4d3-4ac0-93d7-f7030245fd98_content_list.json

@@ -0,0 +1,1154 @@
+[
+    {
+        "type": "text",
+        "text": "CLAIR: Evaluating Image Captions with Large Language Models",
+        "text_level": 1,
+        "bbox": [
+            159,
+            89,
+            836,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "David M. Chan, Suzanne Petryk, Joseph E. Gonzalez, Trevor Darrell, John Canny  \nUniversity of California, Berkeley",
+        "bbox": [
+            151,
+            141,
+            848,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{davidchan,spetryk,jegonzal,trevordarrell,canny}@berkeley.edu",
+        "bbox": [
+            191,
+            175,
+            808,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The evaluation of machine-generated image captions poses an interesting yet persistent challenge. Effective evaluation measures must consider numerous dimensions of similarity, including semantic relevance, visual structure, object interactions, caption diversity, and specificity. Existing highly-engineered measures attempt to capture specific aspects, but fall short in providing a holistic score that aligns closely with human judgments. Here, we propose CLAIR<sup>1</sup>, a novel method that leverages the zero-shot language modeling capabilities of large language models (LLMs) to evaluate candidate captions. In our evaluations, CLAIR demonstrates a stronger correlation with human judgments of caption quality compared to existing measures. Notably, on Flickr8K-Expert, CLAIR achieves relative correlation improvements over SPICE of  $39.6\\%$  and over image-augmented methods such as RefCLIP-S of  $18.3\\%$ . Moreover, CLAIR provides noisily interpretable results by allowing the language model to identify the underlying reasoning behind its assigned score. Code is available at https://davidmchan.github.io/clair/.",
+        "bbox": [
+            141,
+            285,
+            460,
+            640
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction & Background",
+        "text_level": 1,
+        "bbox": [
+            114,
+            659,
+            389,
+            676
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Automatically evaluating the quality of image captions is challenging. There are many dimensions to consider, such as grammatical quality, semantic relevance, correctness, and specificity, among others. To ensure fair evaluations, most image captioning works employ a suite of measures, each capturing different aspects. For instance, n-gram-based measures like BLEU (Papineni et al., 2002) or CIDEr (Vedantam et al., 2015) broadly measure content overlap, SPICE (Anderson et al., 2016) compares scene graph structures, and CLIPScore, TIFA, SeeTrue and VPEval (Hessel",
+        "bbox": [
+            112,
+            690,
+            489,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "You are trying to tell if a candidate set of captions is describing the same image as a reference set of captions.",
+        "bbox": [
+            552,
+            258,
+            840,
+            323
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Candidate set:",
+        "bbox": [
+            552,
+            324,
+            682,
+            337
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{candidate captions}",
+        "bbox": [
+            552,
+            338,
+            734,
+            354
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Reference set:",
+        "bbox": [
+            552,
+            356,
+            682,
+            369
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{reference captions}",
+        "bbox": [
+            552,
+            372,
+            734,
+            387
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "On a precise scale from 0 to 100, how likely is it that the candidate set is describing the same image as the reference set?",
+        "bbox": [
+            552,
+            388,
+            840,
+            451
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "(JSON format, with a key “score”, value between 0 and 100, and a key “reason” with a string value.)",
+        "bbox": [
+            552,
+            451,
+            840,
+            499
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Figure 1: CLAIR: a (surprisingly simple) large language model-based measure for image caption evaluation. We find that CLAIR not only correlates strongly with human judgments of caption quality but can also generate interpretable reasons for the generated scores.",
+        "bbox": [
+            507,
+            510,
+            882,
+            583
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "et al., 2021; Hu et al., 2023; Yarom et al., 2023; Cho et al., 2023) directly incorporate visual information. Unfortunately, while significant strides have been made in automated evaluation, human preference studies remain the most reliable (yet costly) source of caption evaluation.",
+        "bbox": [
+            507,
+            587,
+            884,
+            684
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Fortunately, recent advances in large language models (LLMs) have opened new avenues for automatic evaluation. Models trained with reinforcement learning from human feedback (RLHF, Christiano et al. (2017)) or similar methods are particularly useful for open-ended evaluation tasks, including image captioning, due to their explicit training to align with human preferences.",
+        "bbox": [
+            507,
+            689,
+            882,
+            818
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In our work, paralleling several recent works which find that LLMs can act as effective \"judges\" for selecting the better answer from two candidates (Bubeck et al., 2023; Dettmers et al., 2023; Chiang et al., 2023), we explore the ability of LLMs to evaluate caption quality in the multimodal setting.",
+        "bbox": [
+            507,
+            822,
+            885,
+            920
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{1}$ Meaning \"clear\" in French, in line with other colorful metric names (Papineni et al., 2002; Lin, 2004; Lita et al., 2005).",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We introduce CLAIR (Criterion using LLanguage models for Image caption Rating), a measure which scores a candidate caption based on the likelihood that it describes the same image as a set of references by directly asking an LLM to produce a numeric rating. We further query the LLM to provide a reason for its score, providing a level of interpretability to the scalar rating. As far as we are aware, this is the first paper to explore replacing measures of semantic text quality with directly obtained LLM judgments, however concurrently, Zheng et al. (2023) have shown that directly providing an answer rating can align highly with human preferences on a range of standard language-based tasks, such as conversational instruction following.",
+        "bbox": [
+            115,
+            84,
+            485,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Through several experiments on captioning datasets such as MS-COCO (Xu et al., 2016), Flickr8k (Mao et al., 2014), and PASCAL50S (Vedantam et al., 2015), we find that CLAIR correlates surprisingly well with human preferences, outperforming prior captioning measures. We additionally propose  $\\mathrm{CLAIR}_E$ , where we Ensemble the outputs of several LLMs by taking the average score, leading to further improvements.",
+        "bbox": [
+            115,
+            331,
+            485,
+            474
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Despite a simple pipeline using an LLM prompt with minimal output parsing, CLAIR's strong correlation with human preferences suggests that it captures multiple dimensions of caption similarity at once - a feature that prior measures struggle to achieve alone. More generally, CLAIR demonstrates how language-only models can evaluate vision-language tasks. We show LLMs can provide not only reliable scalar ratings but also corresponding reasoning for a given rating, offering a valuable combination of accuracy and interpretability.",
+        "bbox": [
+            115,
+            480,
+            485,
+            655
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 CLAIR: LLMs for Caption Evaluation",
+        "text_level": 1,
+        "bbox": [
+            115,
+            677,
+            475,
+            693
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In CLAIR, we adapt the zero-shot in-context learning approach described in Brown et al. (2020) to score candidate captions with large language models (LLMs). This involves converting the caption evaluation problem into a human-readable text completion task which is solved by the LLM. Using the prompt in Figure 1, CLAIR first generates completions from the LLM and then parses those completions into both candidate scores and an explainable reason for the score. We use a greedy sampling method  $(t = 0)$  to encourage reproducibility in the results, while acknowledging the inherent nondeterminism in LLMs (see section 4). CLAIR's",
+        "bbox": [
+            115,
+            709,
+            485,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "experimental implementation is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. See Appendix B for further implementation details.",
+        "bbox": [
+            512,
+            84,
+            882,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The choice of language model directly affects the quality of the CLAIR measure – more accurate models should produce evaluations that align better with human judgment. We explore three language models: GPT-3.5 (ChatGPT) (OpenAI, 2022), Claude (Instant) (Bai et al., 2022), and PaLM (Chowdhery et al., 2022). Unfortunately, we found for several open-weight language models including Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023) that CLAIR aligned poorly with human judgment.",
+        "bbox": [
+            512,
+            153,
+            882,
+            313
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As CLAIR is language model-agnostic, we can leverage the different distributions learned by each model and combine their decisions in an ensemble approach we term  $\\mathrm{CLAIR}_E$ . We calculate individual CLAIR scores for each model and compute an unweighted average to obtain the ensemble score.",
+        "bbox": [
+            512,
+            319,
+            882,
+            414
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Benchmark measures: We benchmark against several existing measure of caption similarity. BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), METEOR (Agarwal and Lavie, 2008) and CIDEr (Vedantam et al., 2015) all primarily measure n-gram overlap (however have different weighting schemes between n-grams, and across precision/recall). We also compare against SPICE (Anderson et al., 2016), which compares caption parse trees and focuses on matching perceived action and object relationships in addition to n-grams. While the aforementioned measures are commonly reported in image captioning works, we also compare against several modern measures, including CLIP-Score (Hessel et al., 2021) which uses the recent CLIP (Radford et al., 2021) model for reference-free evaluation.",
+        "bbox": [
+            512,
+            431,
+            882,
+            686
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Evaluation & Discussion",
+        "text_level": 1,
+        "bbox": [
+            512,
+            706,
+            751,
+            721
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the quality of the measure, we run several evaluations that compare scores generated by CLAIR to both human judgments of caption quality and other image captioning evaluation measures. We additionally provide several qualitative examples in Figure 2. A unique benefit of CLAIR is that it provides not only numeric scores but is also introspectable, as it can identify which details in the candidate caption set match the reference set.",
+        "bbox": [
+            512,
+            737,
+            882,
+            881
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Sample-level human correlation: We first ask the question, how well does CLAIR correlate with",
+        "bbox": [
+            512,
+            887,
+            880,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/7417d5ef07db21535534ee9344cada5add8fb593760d7ca11d0d3697447e1f9b.jpg",
+        "image_caption": [
+            "Candidate",
+            "A boy jumps into the blue pool water."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            152,
+            84,
+            336,
+            237
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLAIR Reason",
+        "text_level": 1,
+        "bbox": [
+            208,
+            263,
+            280,
+            271
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The candidate set and reference set both mention a boy jumping into water, but the candidate set does not provide as much detail about the boy's appearance or the location of the pool. Additionally, the reference set includes captions about diving off a pier and holding one's nose, which are not mentioned in the candidate set.",
+        "bbox": [
+            139,
+            272,
+            349,
+            341
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            215,
+            344,
+            273,
+            353
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "\"A blonde boy wearing a blue life vest and ca mo shorts jumps off a diveboard into a blue lake.\" \"A boy diving backward off a pier in a large lake.\" \"A boy in a blue life jacket jumps into the water.\" A boy in a blue life jacket jumps off a board into the lake. \"A boy is holding his nose and jumping off a diving board backwards into a lake.\"",
+        "bbox": [
+            149,
+            354,
+            337,
+            397
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/31972ec12e7b874e7b669083118e7acb8ce0691ec34a7c8f7e33b48d8002ea78.jpg",
+        "image_caption": [
+            "Candidate",
+            "A cyclist in a helmet and yellow jersey.",
+            "Figure 2: Several qualitative examples of CLAIR from the Flickr8K-Expert dataset. CLAIR not only correlates better with human judgments of caption quality but also provides detailed explanations for its score. CLAIR scores normalized by 100."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            411,
+            84,
+            596,
+            239
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLAIR Reason",
+        "text_level": 1,
+        "bbox": [
+            467,
+            262,
+            537,
+            269
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The candidate set only mentions a cyclist in a helmet and yellow jersey, while the reference set describes a man in a green shirt carrying a drink while riding a bicycle. Although there are some similarities, such as the fact that both sets mention a person riding a bike, the differences in the details make it uncertain whether they are describing the same image or not.",
+        "bbox": [
+            379,
+            271,
+            623,
+            340
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            473,
+            344,
+            532,
+            353
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "\"A biker enjoys a coffee.\" \"A man in a bright green shirt and sunglasses is riding a bicycle through the streets whilst drinking a latte.\" \"A man in a bright green shirt riding a bicycle in a paved courtyard, carrying a drink.\" \"A man rides on his bike with one hand and holds a drink with the other.\" \"A man riding a bike wearing a green shirt with a drink in his hand.\"",
+        "bbox": [
+            379,
+            353,
+            625,
+            397
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/50f9d861ee688dc38e0ae89fa571e3d73e37b589470a71d9ee02d47c059bc6fb.jpg",
+        "image_caption": [
+            "Candidate",
+            "Person snowboarding at a ski slope."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            670,
+            84,
+            855,
+            237
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLAIR Reason",
+        "text_level": 1,
+        "bbox": [
+            727,
+            263,
+            798,
+            269
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The candidate caption mentions only one person snowboarding, while all the reference captions mention multiple people skiing or climbing a mountain. Additionally, the candidate caption does not mention any details about the mountain or the surroundings, which are present in the reference captions.",
+        "bbox": [
+            663,
+            271,
+            857,
+            340
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            734,
+            344,
+            791,
+            353
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "\"a group of skiers going up a mountain.\" \"Four cross-country skiers climb uphill.\" \"Four people climbing a hill in the snow.\" \"Four skiers walking up a snow covered hill.\" \"four skiers climbing snow dropped mountain.\"",
+        "bbox": [
+            684,
+            353,
+            840,
+            385
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Sample-level correlation (Kendall's  $\\tau_b$ ) with human judgments. All p-values  $< 0.001$ .*: Model has access to additional visual context.",
+        "bbox": [
+            112,
+            435,
+            485,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/1360cd6662c52bceec0417c087c8626a58cc91efffc2844ac8001e62d8c36427.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Measure</td><td colspan=\"3\">Dataset</td></tr><tr><td>COMPOSITE</td><td>Flickr8K</td><td>MS-COCO</td></tr><tr><td>BLEU@1</td><td>0.313</td><td>0.323</td><td>0.265</td></tr><tr><td>BLEU@4</td><td>0.306</td><td>0.308</td><td>0.215</td></tr><tr><td>ROUGE-L</td><td>0.324</td><td>0.323</td><td>0.221</td></tr><tr><td>BERT-S</td><td>0.301</td><td>0.392</td><td>0.163</td></tr><tr><td>METEOR</td><td>0.389</td><td>0.418</td><td>0.239</td></tr><tr><td>CIDEr</td><td>0.377</td><td>0.439</td><td>0.262</td></tr><tr><td>SPICE</td><td>0.403</td><td>0.449</td><td>0.257</td></tr><tr><td>CLIP-S*</td><td>0.498</td><td>0.511</td><td>-</td></tr><tr><td>RefCLIP-S*</td><td>0.512</td><td>0.526</td><td>-</td></tr><tr><td>RefCLIP-X*</td><td>0.523</td><td>0.549</td><td>0.274</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.604</td><td>0.616</td><td>0.296</td></tr><tr><td>+Claude</td><td>0.542</td><td>0.563</td><td>0.320</td></tr><tr><td>+PaLM</td><td>0.580</td><td>0.546</td><td>0.355</td></tr><tr><td>CLAIRE</td><td>0.592</td><td>0.627</td><td>0.374</td></tr><tr><td>Inter-Human</td><td>-</td><td>0.736</td><td>-</td></tr></table>",
+        "bbox": [
+            115,
+            481,
+            485,
+            740
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "human judgments of caption quality at a sample level? We do so by exploring the performance on three datasets, COMPOSITE, Flickr8K-Expert, and MS-COCO (See Appendix B for details).",
+        "bbox": [
+            112,
+            769,
+            485,
+            833
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The results of our sample-level correlation experiments are shown in Table 1. We can see that CLAIR outperforms language-only measures (e.g., 0.604 to 0.403 for SPICE), and in most cases, outperforms vision-augmented measures.  $\\mathrm{CLAIR}_E$  achieves",
+        "bbox": [
+            112,
+            838,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "strong sample-level correlation on all datasets; for instance,  $\\mathrm{CLAIR}_E$  closes the gap to inter-human agreement by 0.097 over vision-based measures and 0.132 over language-based measures. The improvements of  $\\mathrm{CLAIR}_E$  over CLAIR suggest that each language model may have some bias (similar to each human), yet the ensemble of models correlates more strongly with human judgments. A reasonable concern might be that the models underlying existing approaches are significantly smaller than those in CLAIR, and trained on less data. To address this, we introduce and compare against RefCLIP-X, which replaces the CLIP model in RefCLIP with a CLIP ViT-bigG/14 model trained on LAION 2B (Ilharco et al., 2021). Even in this case, CLAIR demonstrates significantly improved performance. Note that in Table 1, we use  $\\tau_b$ , instead of the form  $\\tau_c$ , due to the ties generated by the CLAIR model.",
+        "bbox": [
+            507,
+            436,
+            882,
+            725
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "System-level human correlation: In addition to computing the sample-level correlation on the MS-COCO dataset, we use the annotations from the five models considered by Rohrbach et al. (2018) to compute the system-level correlation. For each of the methods, we compute the mean human score on the test samples, and mean metric score on the test samples, followed by the Kendall's rank correlation coefficient (Kendall's tau, strength of ordinal association) between these values (the set of five mean human scores, and the set of five metric scores). Our",
+        "bbox": [
+            507,
+            730,
+            882,
+            906
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/96ed4c7d5027c081e9cbe39cd93f61dda5fc9d2c9fb0f2b80539316f4bf86052.jpg",
+        "table_caption": [
+            "Table 2: System-level correlation between the average CLAIR score and human model evaluation for 5 models trained and evaluated on MS-COCO. All p-values  $< {0.05}$  ."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Measure</td><td>Kendall&#x27;s τb</td><td>Spearman&#x27;s ρ</td><td>Pearson r</td></tr><tr><td>BLEU@1</td><td>0.399</td><td>0.600</td><td>0.706</td></tr><tr><td>BLEU@4</td><td>0.799</td><td>0.899</td><td>0.910</td></tr><tr><td>ROUGE-L</td><td>0.600</td><td>0.700</td><td>0.792</td></tr><tr><td>METEOR</td><td>0.600</td><td>0.700</td><td>0.666</td></tr><tr><td>CIDEr</td><td>0.399</td><td>0.600</td><td>0.856</td></tr><tr><td>SPICE</td><td>0.399</td><td>0.600</td><td>0.690</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>0.799</td><td>0.899</td><td>0.869</td></tr><tr><td>+ Claude</td><td>1.000</td><td>1.000</td><td>0.868</td></tr><tr><td>+ PaLM</td><td>1.000</td><td>1.000</td><td>0.954</td></tr><tr><td>CLAIR_E</td><td>1.000</td><td>1.000</td><td>0.903</td></tr></table>",
+        "bbox": [
+            115,
+            131,
+            485,
+            297
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/9821143f4d49ae7c0dbff7a94e392dad5a6fe7f914430a3557a78659e3739232.jpg",
+        "table_caption": [
+            "Table 3: Accuracy of measures when matching human decisions for PASCAL-50S (5 reference captions). *: Model has access to additional visual context."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Measure</td><td>HC</td><td>HI</td><td>HM</td><td>MM</td><td>All</td></tr><tr><td>BLEU@1</td><td>51.20</td><td>95.70</td><td>91.20</td><td>58.20</td><td>74.08</td></tr><tr><td>BLEU@4</td><td>53.00</td><td>92.40</td><td>86.70</td><td>59.40</td><td>72.88</td></tr><tr><td>ROUGE-L</td><td>51.50</td><td>94.50</td><td>92.50</td><td>57.70</td><td>74.05</td></tr><tr><td>METEOR</td><td>56.70</td><td>97.60</td><td>94.20</td><td>63.40</td><td>77.98</td></tr><tr><td>CIDEr</td><td>53.00</td><td>98.00</td><td>91.50</td><td>64.50</td><td>76.75</td></tr><tr><td>SPICE</td><td>52.60</td><td>93.90</td><td>83.60</td><td>48.10</td><td>69.55</td></tr><tr><td>TIGEr*</td><td>56.00</td><td>99.80</td><td>92.80</td><td>74.20</td><td>80.70</td></tr><tr><td>CLIP-S*</td><td>56.50</td><td>99.30</td><td>96.40</td><td>70.40</td><td>80.70</td></tr><tr><td>RefCLIP-S*</td><td>64.50</td><td>99.60</td><td>95.40</td><td>72.80</td><td>83.10</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>52.40</td><td>99.50</td><td>89.80</td><td>73.00</td><td>78.67</td></tr><tr><td>+ Claude</td><td>57.90</td><td>98.50</td><td>91.30</td><td>62.90</td><td>77.65</td></tr><tr><td>+ PaLM</td><td>54.70</td><td>98.30</td><td>87.30</td><td>64.00</td><td>76.08</td></tr><tr><td>CLAIRE</td><td>57.70</td><td>99.80</td><td>94.60</td><td>75.60</td><td>81.93</td></tr></table>",
+        "bbox": [
+            115,
+            365,
+            485,
+            574
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "results, given in Table 2, demonstrate that CLAIR ranks the five methods in a novel way that is more accordant with human rankings of the methods. These results further suggest that CLAIR has the potential to redefine which methods are preferable to humans compared to existing n-gram approaches.",
+        "bbox": [
+            112,
+            592,
+            487,
+            690
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Decision Making: In addition to evaluating the correlation with human judgments, we also evaluate the capability of the measure to perform discriminative analysis. The PASCAL-50S dataset (Vedantam et al., 2015) contains a set of 4000 human-annotated caption pairs. For each pair of captions, humans label which caption in the pair is closest to the reference set for the image. The caption pairs fall into four groups: \"HC:\" two human-written captions matching the image, \"HI:\" one human caption, and one machine-generated caption, with only one matching the image, \"HM:\" a matching human caption and a matching machine-generated caption and \"MM:\" two matching machine-generated captions.",
+        "bbox": [
+            112,
+            694,
+            489,
+            920
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/059e7bab77f0e2d87b53dfac17bb880b311fae12350d0d2a0f9b77d0e7209ed0.jpg",
+        "table_caption": [
+            "Table 4: Pearson correlation with human judgments when evaluating sets of captions on MS-COCO ( $\\bar{N} = 794$ )."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Measure</td><td colspan=\"2\">Coveragep-value</td><td colspan=\"2\">Correctnessp-value</td></tr><tr><td>BLEU@4</td><td>0.004</td><td>0.816</td><td>0.003</td><td>0.888</td></tr><tr><td>ROUGE-L</td><td>0.011</td><td>0.563</td><td>0.038</td><td>0.184</td></tr><tr><td>METEOR</td><td>0.016</td><td>0.398</td><td>0.006</td><td>0.765</td></tr><tr><td>CIDEr</td><td>0.004</td><td>0.844</td><td>0.026</td><td>0.173</td></tr><tr><td>TRM-METEOR</td><td colspan=\"2\">0.128&lt;0.001</td><td colspan=\"2\">0.108&lt;0.001</td></tr><tr><td>TRM-BLEU</td><td colspan=\"2\">0.127&lt;0.001</td><td colspan=\"2\">0.151&lt;0.001</td></tr><tr><td>MMD-BERT</td><td colspan=\"2\">0.129&lt;0.001</td><td colspan=\"2\">0.124&lt;0.001</td></tr><tr><td>FID-BERT</td><td>0.081</td><td>0.011</td><td colspan=\"2\">0.098&lt;0.001</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.195</td><td>0.011</td><td>0.187</td><td>0.014</td></tr><tr><td>+Claude</td><td>0.110</td><td>0.099</td><td>0.124</td><td>0.145</td></tr><tr><td>+PaLM</td><td>0.129</td><td>0.081</td><td>0.085</td><td>0.172</td></tr><tr><td>CLAIR_E</td><td>0.183</td><td>0.027</td><td>0.156</td><td>0.018</td></tr><tr><td>Inter-Human</td><td colspan=\"2\">0.225&lt;0.001</td><td colspan=\"2\">0.274&lt;0.001</td></tr></table>",
+        "bbox": [
+            510,
+            118,
+            880,
+            335
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "See Appendix B for more dataset information.",
+        "bbox": [
+            507,
+            363,
+            847,
+            378
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The performance on PASCAL-50S is given in Table 3. We can see that  $\\mathrm{CLAIR}_E$  outperforms all existing text-only measures (e.g., by  $5.18\\%$  overall score over CIDEr), and in many cases, even outperforms measures that have access to the image at test time. Note that it is relatively weaker than image-augmented models in the HC setting; however, since both captions are correct, the model often cannot judge which is better purely the text. Models such as RefCLIP-S that have access to the image are naturally better discriminators in this case. We suspect that CLAIR's discriminative performance could be further improved by giving the LLM a choice between the two captions; however, we leave this optimization to future work.",
+        "bbox": [
+            505,
+            382,
+            882,
+            625
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Groups of Captions: While CLAIR is capable of comparing a single candidate caption to a set of reference captions, it is also capable of comparing sets of candidate captions to sets of reference captions. This task is necessary when evaluating the ability of a model to generate captions that are diverse and that fully describe the conditional text distribution. We evaluate on the COCO-Sets dataset (Chan et al., 2022), 794 caption sets rated by AMT workers on two scales: how closely a candidate set matches the reference set in terms of both correctness and content coverage (See Appendix B for details). The results of this experiment are given in Table 4. We can see that CLAIR outperforms well when measuring the quality of a group of captions, and approaches the inter-human correlation on the (very) challenging task. CLAIR also outperforms TRM-Meteor and TRM-BLEU (Chan et al.,",
+        "bbox": [
+            507,
+            627,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2022), suggesting that LLMs can judge both the content and diversity of the caption sets.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            133,
+            250,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While CLAIR correlates well with human judgments of caption quality, it has several limitations:",
+        "bbox": [
+            112,
+            164,
+            489,
+            197
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Non-Determinism and Parsing Errors: Because CLAIR depends on the output of a language model, the measure can be non-deterministic and noisy. For instance, it may fail to elicit a judgment (e.g., \"As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references\"), or rarely, generate malformed JSON output. To address these issues, we perform multiple queries to the LLM, sometimes at higher temperatures if necessary. As a consequence, the measure may differ between runs, although we found the variance to be relatively insignificant (< 0.01 in many of the experiments). Additionally, since the language models used are not open-source, the models are subject to arbitrary change, replacement, or removal, which limits the efficacy of the measure as a long-term comparable measurement. We hope that increasing open access to language models with efforts such as Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023), will help to alleviate these challenges in the future.",
+        "bbox": [
+            112,
+            200,
+            489,
+            523
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Increased Cost: CLAIR relies on language models which contain many billions of parameters. These language models have not only monetary cost but also human and environmental costs (Bender et al., 2021) which can reduce its utility as a target during training, such as for self-critical sequence training (Rennie et al., 2017). While API-based LLMs may be considered costly, even open-source LLMs have a cost (which can often be hard to quantify). CLAIR on the MS-COCO dataset uses an average of 226.148 tokens per sample (on OpenAI's API), representing a cost of  $0.0067 per sample (GPT-4), or$ 0.00033 per sample (GPT 3.5). For PALM, this drops to $0.000113 per sample. We hope that over time, advances in LLM inference (such as quantization and distillation), coupled with improvements in architecture will continue to yield lower-cost alternatives with strong performance on the caption evaluation task.",
+        "bbox": [
+            112,
+            527,
+            489,
+            834
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Hallucination: While CLAIR does suffer from potential hallucination, we strongly believe that this weakness does not diminish the fact that CLAIR still correlates strongly with human judgment. In CLAIR, hallucinations in the score manifest as",
+        "bbox": [
+            112,
+            839,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "\"incorrect\" judgements of similarity, while hallucinations in the explanations manifest as poorly grounded explanations of the score/quality. Hallucinations in the score should be considered false negatives (blind spots instead of hallucinations). In the case of hallucinations in the explanations, such hallucinations may lead to misinterpretation, but arguably less misinterpretation than a black box method, and may even indicate misunderstandings in the model. Hallucination is a well-known challenge of current LLMs and is the subject of a great amount of research on instruction-tuning, RLHF, RLAIF, and other methods. As hallucination and instruction-following performance of the base models improves, CLAIR inherit similar improvements.",
+        "bbox": [
+            505,
+            84,
+            884,
+            326
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Explainability: While CLAIR generates explanations for each rating, CLAIR has no strict scoring rubric. Much like human judgments, there is no direct way of attributing changes in score to changes in caption quality. For similar reasons, it is difficult to evaluate the quality of the generated explanations. Qualitatively, the explanations are generally reasonable and consider multiple axes of judgment.",
+        "bbox": [
+            507,
+            330,
+            884,
+            460
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            475,
+            640,
+            491
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This work introduces CLAIR, an LLM-based evaluation measure for image captioning. CLAIR's superior performance compared to highly-engineered measures indicates a remarkable fact: LLMs are well aligned with human judgments of caption quality, even more so than some measures designed specifically for semantic similarity. CLAIR is only a glimpse into how LLMs can be used for evaluation tasks, and image captioning is only the beginning. We hope that our work will inspire further exploration of similar measures in other vision and language domains, such as visual storytelling (Huang et al., 2016), where human evaluation of generated text remains a challenging task.",
+        "bbox": [
+            505,
+            506,
+            884,
+            731
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Somak Aditya, Yezhou Yang, Chitta Baral, Cornelia Fermuller, and Yiannis Aloimonos. 2015. From images to sentences through scene description graphs using commonsense reasoning and knowledge. ArXiv preprint, abs/1511.03292.",
+            "Abhaya Agarwal and Alon Lavie. 2008. Meteor, M-BLEU and M-TER: Evaluation metrics for high-correlation with human rankings of machine translation output. In Proceedings of the Third Workshop on Statistical Machine Translation, pages 115-118. Association for Computational Linguistics.",
+            "Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2016. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer.",
+            "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. ArXiv preprint, abs/2204.05862.",
+            "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, pages 610-623.",
+            "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.",
+            "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, abs/2303.12712.",
+            "David M Chan, Yiming Ni, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. 2022. Distribution aware metrics for conditional natural language generation. ArXiv preprint, abs/2209.07518.",
+            "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and"
+        ],
+        "bbox": [
+            115,
+            111,
+            485,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\\%$  * chatgpt quality.",
+            "Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. Visual programming for text-to-image generation and evaluation. arXiv preprint arXiv:2305.15328.",
+            "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. ArXiv preprint, abs/2204.02311.",
+            "Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4299-4307.",
+            "Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. *ArXiv preprint*, abs/2305.14314.",
+            "Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. Koala: A dialogue model for academic research. Blog post.",
+            "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics.",
+            "Micah Hodosh, Peter Young, and Julia Hockenmaier. 2013. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899.",
+            "Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. 2023. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897.",
+            "Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, and Margaret Mitchell. 2016. Visual storytelling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1233-1239, San Diego, California. Association for Computational Linguistics.",
+            "Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Schmidt. 2021. Openclip. If you use this software, please cite it as below.",
+            "Ming Jiang, Qiuyuan Huang, Lei Zhang, Xin Wang, Pengchuan Zhang, Zhe Gan, Jana Diesner, and Jianfeng Gao. 2019. TIGEr: Text-to-image grounding for image caption evaluation. 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 2141-2152. Association for Computational Linguistics.",
+            "Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR.",
+            "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics.",
+            "Lucian Vlad Lita, Monica Rogati, and Alon Lavie. 2005. Blanc: Learning evaluation metrics for mt. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 740-747.",
+            "Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L Yuille. 2014. Explain images with multimodal recurrent neural networks. arXiv preprint arXiv:1410.1090.",
+            "OpenAI. 2022. Introducing chatgpt.",
+            "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. 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, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.",
+            "Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. 2017. Self-critical sequence training for image captioning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1179-1195. IEEE Computer Society."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045. Association for Computational Linguistics.",
+            "Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575. IEEE Computer Society.",
+            "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 23318-23340. PMLR.",
+            "Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society.",
+            "Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. 2023. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400.",
+            "Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78.",
+            "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. ArXiv preprint, abs/2306.05685."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            695
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            512,
+            714,
+            596,
+            730
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            512,
+            744,
+            714,
+            760
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We thank Suhong Moon and Kate Saenko for their helpful comments on the work. Authors, as part of their affiliation with UC Berkeley, were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program, as well as gifts from Anyscale, Astronomer, Google, IBM, Intel, Lacework, Microsoft, Mohamed Bin Zayed University of Artificial Intelligence, Samsung SDS, Uber, and VMware.",
+        "bbox": [
+            510,
+            774,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Additional Experimental Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            430,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we provide several additional details for the experiments in section 3 run with the CLAIR measure.",
+        "bbox": [
+            112,
+            118,
+            487,
+            165
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.1 Input Prompt Formatting",
+        "text_level": 1,
+        "bbox": [
+            114,
+            190,
+            366,
+            206
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The CLAIR prompt is given in its entirety in Figure 1. During run-time, candidate and reference captions are prefixed with a “-” and inserted into the prompt, one per line. The resulting query is passed to the large language model. In addition, for models which were not RLHF-tuned to perform conversation (such as PaLM), we found that it was helpful to append an additional prefix {\"score\": to the beginning of the output, to encourage the correct output formatting. CLAIR is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. The model checkpoint metadata is generally unknown (as the APIs are somewhat fluid and evolving).",
+        "bbox": [
+            112,
+            219,
+            489,
+            445
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.2 LLM Output Post-Processing",
+        "text_level": 1,
+        "bbox": [
+            114,
+            468,
+            394,
+            483
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Because CLAIR relies on an LLM to produce output, there is no guarantee that the output will be in the format we expect (i.e. valid, parsable JSON). To extract both the score and the reason, we first extract the first set of paired braces from the output of the LLM and attempt to parse the result as JSON. In most cases (99.997% for GPT-3, 99.991% for Claude, and 99.94% for PaLM during the course of our experiments), this is successful, and the score and reason are returned. In the case that the JSON output is malformed, we attempt to extract any sequence of digits from the LLM to use as a score, and set the reason to \"Unknown.\" When this fails, as can be the case when the models produce an output such as \"As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references\", we retry the prompt at a higher temperature ( $t = 1.0$ ) several times. Failing this (which occurred only three times in the entire evaluation of this paper, across several hundred thousand calls), we set the score to 0 for the caption.",
+        "bbox": [
+            112,
+            495,
+            489,
+            834
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "B.3 Datasets",
+        "text_level": 1,
+        "bbox": [
+            114,
+            858,
+            231,
+            872
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this section, we provide additional detail regarding the datasets used in the evaluations in section 3.",
+        "bbox": [
+            112,
+            885,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "COMPOSITE: The COMPOSITE dataset (Aditya et al., 2015) contains machine-generated test captions for 3995 images spread across the MS-COCO (Xu et al., 2016), Flickr8K (Mao et al., 2014) and Flickr30k (Young et al., 2014) datasets. Each image has three test captions, one written by a human, and two that are model generated. The candidate captions are graded by annotators on Amazon Mechanical Turk (AMT) on a scale of 1 (not relevant) to 5 (very relevant). Inter-human correlations are not available for this dataset.",
+        "bbox": [
+            507,
+            83,
+            884,
+            261
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Flickr8K-Expert: The Flickr8K-Expert dataset (Hodosh et al., 2013) contains 5822 captions associated with 1000 images. The dataset is annotated with expert human judgments of quality, where images are rated from 1 (caption is unrelated to the image) to 4 (caption describes the image without errors). Unlike the composite and MS-COCO datasets, the captions here are selected using an image retrieval system, instead of generated using a learned image captioning model. Following Jiang et al. (2019), we exclude any candidate captions that overlap the reference set.",
+        "bbox": [
+            507,
+            275,
+            884,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "MS-COCO: Following experiments by Rohrbach et al. (2018), we compute the sample-level correlation between our method and human ratings on a 500-image subset of the MS-COCO Karpathy test set. Each image in the subset contains candidate captions generated by 5 models, and each caption is labeled with the average three human ratings generated by AMT workers which range from 1 (very bad) to 5 (very good). Inter-human correlations are not available for this dataset.",
+        "bbox": [
+            507,
+            485,
+            882,
+            645
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "PASCAL-50S: PASCAL-50S contains 1000 images drawn from the PASCAL sentence dataset. Each image is associated with at least 50 (and as many as 120) reference captions. In addition to the reference captions, PASCAL-50S contains a set of 4000 human annotated image/caption pairs containing an image, and two candidate captions. The caption pairs fall into four groups:",
+        "bbox": [
+            507,
+            661,
+            882,
+            791
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. HC: In the HC group, both captions in the pair are human written, and describe the content of the target image correctly.",
+            "2. HI: In the HI group, both captions in the pair are human written, but one caption correctly describes the content of the image, and the"
+        ],
+        "bbox": [
+            522,
+            808,
+            880,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "other caption describes the content of a different image.",
+        "bbox": [
+            149,
+            85,
+            485,
+            116
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "3. HM: In the HM group, one caption is written by a human, and one caption is written by a machine, but both correctly describe the content of the image.",
+            "4. MM: In the MM group, both captions are written by a machine, and both correctly describe the image content."
+        ],
+        "bbox": [
+            129,
+            127,
+            485,
+            249
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In PASCAL-50S, the task is to decide which caption in the pair humans prefer more (a subjective task, hopefully indicating caption quality). Following previous work (Jiang et al., 2019; Hessel et al., 2021), we limit the number of reference sentences to five during evaluation.",
+        "bbox": [
+            112,
+            266,
+            487,
+            361
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "COCO-Sets: The COCO-Sets dataset (Chan et al., 2022) is a set of samples that are designed to evaluate the correlation of distribution-aware image captioning measures with human judgments of distributional distance. In this dataset, humans were presented with two candidate caption sets (two image captioning models, OFA (Wang et al., 2022) and BLIP (Li et al., 2022) using different temperatures), and asked which candidate caption set correlated better with a reference caption set on two measures: how much they overlapped factually (correctness), and how much information they provided about the references (coverage). It consists of 794 AMT worker-generated judgments of caption quality for images in the MS-COCO dataset.",
+        "bbox": [
+            112,
+            376,
+            487,
+            617
+        ],
+        "page_idx": 8
+    }
+]

2023/CLAIR_ Evaluating Image Captions with Large Language Models/ed931ad6-a4d3-4ac0-93d7-f7030245fd98_model.json

@@ -0,0 +1,1626 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.16,
+                0.09,
+                0.838,
+                0.112
+            ],
+            "angle": 0,
+            "content": "CLAIR: Evaluating Image Captions with Large Language Models"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.142,
+                0.85,
+                0.176
+            ],
+            "angle": 0,
+            "content": "David M. Chan, Suzanne Petryk, Joseph E. Gonzalez, Trevor Darrell, John Canny  \nUniversity of California, Berkeley"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.192,
+                0.177,
+                0.81,
+                0.193
+            ],
+            "angle": 0,
+            "content": "{davidchan,spetryk,jegonzal,trevordarrell,canny}@berkeley.edu"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.262,
+                0.253,
+                0.341,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.286,
+                0.461,
+                0.642
+            ],
+            "angle": 0,
+            "content": "The evaluation of machine-generated image captions poses an interesting yet persistent challenge. Effective evaluation measures must consider numerous dimensions of similarity, including semantic relevance, visual structure, object interactions, caption diversity, and specificity. Existing highly-engineered measures attempt to capture specific aspects, but fall short in providing a holistic score that aligns closely with human judgments. Here, we propose CLAIR<sup>1</sup>, a novel method that leverages the zero-shot language modeling capabilities of large language models (LLMs) to evaluate candidate captions. In our evaluations, CLAIR demonstrates a stronger correlation with human judgments of caption quality compared to existing measures. Notably, on Flickr8K-Expert, CLAIR achieves relative correlation improvements over SPICE of \\(39.6\\%\\) and over image-augmented methods such as RefCLIP-S of \\(18.3\\%\\). Moreover, CLAIR provides noisily interpretable results by allowing the language model to identify the underlying reasoning behind its assigned score. Code is available at https://davidmchan.github.io/clair/."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.66,
+                0.391,
+                0.677
+            ],
+            "angle": 0,
+            "content": "1 Introduction & Background"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.691,
+                0.49,
+                0.884
+            ],
+            "angle": 0,
+            "content": "Automatically evaluating the quality of image captions is challenging. There are many dimensions to consider, such as grammatical quality, semantic relevance, correctness, and specificity, among others. To ensure fair evaluations, most image captioning works employ a suite of measures, each capturing different aspects. For instance, n-gram-based measures like BLEU (Papineni et al., 2002) or CIDEr (Vedantam et al., 2015) broadly measure content overlap, SPICE (Anderson et al., 2016) compares scene graph structures, and CLIPScore, TIFA, SeeTrue and VPEval (Hessel"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.553,
+                0.259,
+                0.841,
+                0.324
+            ],
+            "angle": 0,
+            "content": "You are trying to tell if a candidate set of captions is describing the same image as a reference set of captions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.325,
+                0.684,
+                0.338
+            ],
+            "angle": 0,
+            "content": "Candidate set:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.34,
+                0.736,
+                0.355
+            ],
+            "angle": 0,
+            "content": "{candidate captions}"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.357,
+                0.684,
+                0.37
+            ],
+            "angle": 0,
+            "content": "Reference set:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.373,
+                0.736,
+                0.388
+            ],
+            "angle": 0,
+            "content": "{reference captions}"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.389,
+                0.841,
+                0.452
+            ],
+            "angle": 0,
+            "content": "On a precise scale from 0 to 100, how likely is it that the candidate set is describing the same image as the reference set?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.554,
+                0.453,
+                0.841,
+                0.5
+            ],
+            "angle": 0,
+            "content": "(JSON format, with a key “score”, value between 0 and 100, and a key “reason” with a string value.)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.511,
+                0.884,
+                0.584
+            ],
+            "angle": 0,
+            "content": "Figure 1: CLAIR: a (surprisingly simple) large language model-based measure for image caption evaluation. We find that CLAIR not only correlates strongly with human judgments of caption quality but can also generate interpretable reasons for the generated scores."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.588,
+                0.885,
+                0.685
+            ],
+            "angle": 0,
+            "content": "et al., 2021; Hu et al., 2023; Yarom et al., 2023; Cho et al., 2023) directly incorporate visual information. Unfortunately, while significant strides have been made in automated evaluation, human preference studies remain the most reliable (yet costly) source of caption evaluation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.69,
+                0.884,
+                0.819
+            ],
+            "angle": 0,
+            "content": "Fortunately, recent advances in large language models (LLMs) have opened new avenues for automatic evaluation. Models trained with reinforcement learning from human feedback (RLHF, Christiano et al. (2017)) or similar methods are particularly useful for open-ended evaluation tasks, including image captioning, due to their explicit training to align with human preferences."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.887,
+                0.921
+            ],
+            "angle": 0,
+            "content": "In our work, paralleling several recent works which find that LLMs can act as effective \"judges\" for selecting the better answer from two candidates (Bubeck et al., 2023; Dettmers et al., 2023; Chiang et al., 2023), we explore the ability of LLMs to evaluate caption quality in the multimodal setting."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.892,
+                0.488,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{1}\\)Meaning \"clear\" in French, in line with other colorful metric names (Papineni et al., 2002; Lin, 2004; Lita et al., 2005)."
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.085,
+                0.486,
+                0.325
+            ],
+            "angle": 0,
+            "content": "We introduce CLAIR (Criterion using LLanguage models for Image caption Rating), a measure which scores a candidate caption based on the likelihood that it describes the same image as a set of references by directly asking an LLM to produce a numeric rating. We further query the LLM to provide a reason for its score, providing a level of interpretability to the scalar rating. As far as we are aware, this is the first paper to explore replacing measures of semantic text quality with directly obtained LLM judgments, however concurrently, Zheng et al. (2023) have shown that directly providing an answer rating can align highly with human preferences on a range of standard language-based tasks, such as conversational instruction following."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.332,
+                0.486,
+                0.475
+            ],
+            "angle": 0,
+            "content": "Through several experiments on captioning datasets such as MS-COCO (Xu et al., 2016), Flickr8k (Mao et al., 2014), and PASCAL50S (Vedantam et al., 2015), we find that CLAIR correlates surprisingly well with human preferences, outperforming prior captioning measures. We additionally propose \\(\\mathrm{CLAIR}_E\\), where we Ensemble the outputs of several LLMs by taking the average score, leading to further improvements."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.481,
+                0.486,
+                0.656
+            ],
+            "angle": 0,
+            "content": "Despite a simple pipeline using an LLM prompt with minimal output parsing, CLAIR's strong correlation with human preferences suggests that it captures multiple dimensions of caption similarity at once - a feature that prior measures struggle to achieve alone. More generally, CLAIR demonstrates how language-only models can evaluate vision-language tasks. We show LLMs can provide not only reliable scalar ratings but also corresponding reasoning for a given rating, offering a valuable combination of accuracy and interpretability."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.117,
+                0.678,
+                0.477,
+                0.694
+            ],
+            "angle": 0,
+            "content": "2 CLAIR: LLMs for Caption Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.71,
+                0.486,
+                0.918
+            ],
+            "angle": 0,
+            "content": "In CLAIR, we adapt the zero-shot in-context learning approach described in Brown et al. (2020) to score candidate captions with large language models (LLMs). This involves converting the caption evaluation problem into a human-readable text completion task which is solved by the LLM. Using the prompt in Figure 1, CLAIR first generates completions from the LLM and then parses those completions into both candidate scores and an explainable reason for the score. We use a greedy sampling method \\((t = 0)\\) to encourage reproducibility in the results, while acknowledging the inherent nondeterminism in LLMs (see section 4). CLAIR's"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.085,
+                0.883,
+                0.149
+            ],
+            "angle": 0,
+            "content": "experimental implementation is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. See Appendix B for further implementation details."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.154,
+                0.883,
+                0.315
+            ],
+            "angle": 0,
+            "content": "The choice of language model directly affects the quality of the CLAIR measure – more accurate models should produce evaluations that align better with human judgment. We explore three language models: GPT-3.5 (ChatGPT) (OpenAI, 2022), Claude (Instant) (Bai et al., 2022), and PaLM (Chowdhery et al., 2022). Unfortunately, we found for several open-weight language models including Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023) that CLAIR aligned poorly with human judgment."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.32,
+                0.883,
+                0.415
+            ],
+            "angle": 0,
+            "content": "As CLAIR is language model-agnostic, we can leverage the different distributions learned by each model and combine their decisions in an ensemble approach we term \\(\\mathrm{CLAIR}_E\\). We calculate individual CLAIR scores for each model and compute an unweighted average to obtain the ensemble score."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.432,
+                0.883,
+                0.687
+            ],
+            "angle": 0,
+            "content": "Benchmark measures: We benchmark against several existing measure of caption similarity. BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), METEOR (Agarwal and Lavie, 2008) and CIDEr (Vedantam et al., 2015) all primarily measure n-gram overlap (however have different weighting schemes between n-grams, and across precision/recall). We also compare against SPICE (Anderson et al., 2016), which compares caption parse trees and focuses on matching perceived action and object relationships in addition to n-grams. While the aforementioned measures are commonly reported in image captioning works, we also compare against several modern measures, including CLIP-Score (Hessel et al., 2021) which uses the recent CLIP (Radford et al., 2021) model for reference-free evaluation."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.514,
+                0.707,
+                0.752,
+                0.722
+            ],
+            "angle": 0,
+            "content": "3 Evaluation & Discussion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.738,
+                0.883,
+                0.882
+            ],
+            "angle": 0,
+            "content": "To evaluate the quality of the measure, we run several evaluations that compare scores generated by CLAIR to both human judgments of caption quality and other image captioning evaluation measures. We additionally provide several qualitative examples in Figure 2. A unique benefit of CLAIR is that it provides not only numeric scores but is also introspectable, as it can identify which details in the candidate caption set match the reference set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.514,
+                0.888,
+                0.881,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Sample-level human correlation: We first ask the question, how well does CLAIR correlate with"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.153,
+                0.085,
+                0.337,
+                0.239
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.218,
+                0.243,
+                0.272,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Candidate"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.156,
+                0.251,
+                0.333,
+                0.26
+            ],
+            "angle": 0,
+            "content": "A boy jumps into the blue pool water."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.209,
+                0.264,
+                0.281,
+                0.272
+            ],
+            "angle": 0,
+            "content": "CLAIR Reason"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.14,
+                0.273,
+                0.351,
+                0.342
+            ],
+            "angle": 0,
+            "content": "The candidate set and reference set both mention a boy jumping into water, but the candidate set does not provide as much detail about the boy's appearance or the location of the pool. Additionally, the reference set includes captions about diving off a pier and holding one's nose, which are not mentioned in the candidate set."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.216,
+                0.346,
+                0.274,
+                0.354
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.355,
+                0.339,
+                0.398
+            ],
+            "angle": 0,
+            "content": "\"A blonde boy wearing a blue life vest and ca mo shorts jumps off a diveboard into a blue lake.\" \"A boy diving backward off a pier in a large lake.\" \"A boy in a blue life jacket jumps into the water.\" A boy in a blue life jacket jumps off a board into the lake. \"A boy is holding his nose and jumping off a diving board backwards into a lake.\""
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.412,
+                0.085,
+                0.597,
+                0.24
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.477,
+                0.243,
+                0.53,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Candidate"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.412,
+                0.251,
+                0.595,
+                0.26
+            ],
+            "angle": 0,
+            "content": "A cyclist in a helmet and yellow jersey."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.468,
+                0.263,
+                0.539,
+                0.271
+            ],
+            "angle": 0,
+            "content": "CLAIR Reason"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.38,
+                0.272,
+                0.625,
+                0.341
+            ],
+            "angle": 0,
+            "content": "The candidate set only mentions a cyclist in a helmet and yellow jersey, while the reference set describes a man in a green shirt carrying a drink while riding a bicycle. Although there are some similarities, such as the fact that both sets mention a person riding a bike, the differences in the details make it uncertain whether they are describing the same image or not."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.474,
+                0.346,
+                0.534,
+                0.354
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.38,
+                0.354,
+                0.626,
+                0.398
+            ],
+            "angle": 0,
+            "content": "\"A biker enjoys a coffee.\" \"A man in a bright green shirt and sunglasses is riding a bicycle through the streets whilst drinking a latte.\" \"A man in a bright green shirt riding a bicycle in a paved courtyard, carrying a drink.\" \"A man rides on his bike with one hand and holds a drink with the other.\" \"A man riding a bike wearing a green shirt with a drink in his hand.\""
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.672,
+                0.085,
+                0.857,
+                0.239
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.738,
+                0.243,
+                0.79,
+                0.251
+            ],
+            "angle": 0,
+            "content": "Candidate"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.677,
+                0.251,
+                0.848,
+                0.26
+            ],
+            "angle": 0,
+            "content": "Person snowboarding at a ski slope."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.728,
+                0.264,
+                0.799,
+                0.271
+            ],
+            "angle": 0,
+            "content": "CLAIR Reason"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.665,
+                0.272,
+                0.858,
+                0.341
+            ],
+            "angle": 0,
+            "content": "The candidate caption mentions only one person snowboarding, while all the reference captions mention multiple people skiing or climbing a mountain. Additionally, the candidate caption does not mention any details about the mountain or the surroundings, which are present in the reference captions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.735,
+                0.346,
+                0.792,
+                0.354
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.685,
+                0.354,
+                0.842,
+                0.386
+            ],
+            "angle": 0,
+            "content": "\"a group of skiers going up a mountain.\" \"Four cross-country skiers climb uphill.\" \"Four people climbing a hill in the snow.\" \"Four skiers walking up a snow covered hill.\" \"four skiers climbing snow dropped mountain.\""
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.405,
+                0.882,
+                0.43
+            ],
+            "angle": 0,
+            "content": "Figure 2: Several qualitative examples of CLAIR from the Flickr8K-Expert dataset. CLAIR not only correlates better with human judgments of caption quality but also provides detailed explanations for its score. CLAIR scores normalized by 100."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.436,
+                0.487,
+                0.472
+            ],
+            "angle": 0,
+            "content": "Table 1: Sample-level correlation (Kendall's \\( \\tau_b \\)) with human judgments. All p-values \\( < 0.001 \\).*: Model has access to additional visual context."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.116,
+                0.482,
+                0.487,
+                0.741
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Measure</td><td colspan=\"3\">Dataset</td></tr><tr><td>COMPOSITE</td><td>Flickr8K</td><td>MS-COCO</td></tr><tr><td>BLEU@1</td><td>0.313</td><td>0.323</td><td>0.265</td></tr><tr><td>BLEU@4</td><td>0.306</td><td>0.308</td><td>0.215</td></tr><tr><td>ROUGE-L</td><td>0.324</td><td>0.323</td><td>0.221</td></tr><tr><td>BERT-S</td><td>0.301</td><td>0.392</td><td>0.163</td></tr><tr><td>METEOR</td><td>0.389</td><td>0.418</td><td>0.239</td></tr><tr><td>CIDEr</td><td>0.377</td><td>0.439</td><td>0.262</td></tr><tr><td>SPICE</td><td>0.403</td><td>0.449</td><td>0.257</td></tr><tr><td>CLIP-S*</td><td>0.498</td><td>0.511</td><td>-</td></tr><tr><td>RefCLIP-S*</td><td>0.512</td><td>0.526</td><td>-</td></tr><tr><td>RefCLIP-X*</td><td>0.523</td><td>0.549</td><td>0.274</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.604</td><td>0.616</td><td>0.296</td></tr><tr><td>+Claude</td><td>0.542</td><td>0.563</td><td>0.320</td></tr><tr><td>+PaLM</td><td>0.580</td><td>0.546</td><td>0.355</td></tr><tr><td>CLAIRE</td><td>0.592</td><td>0.627</td><td>0.374</td></tr><tr><td>Inter-Human</td><td>-</td><td>0.736</td><td>-</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.77,
+                0.487,
+                0.834
+            ],
+            "angle": 0,
+            "content": "human judgments of caption quality at a sample level? We do so by exploring the performance on three datasets, COMPOSITE, Flickr8K-Expert, and MS-COCO (See Appendix B for details)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.839,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The results of our sample-level correlation experiments are shown in Table 1. We can see that CLAIR outperforms language-only measures (e.g., 0.604 to 0.403 for SPICE), and in most cases, outperforms vision-augmented measures. \\(\\mathrm{CLAIR}_E\\) achieves"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.437,
+                0.884,
+                0.726
+            ],
+            "angle": 0,
+            "content": "strong sample-level correlation on all datasets; for instance, \\(\\mathrm{CLAIR}_E\\) closes the gap to inter-human agreement by 0.097 over vision-based measures and 0.132 over language-based measures. The improvements of \\(\\mathrm{CLAIR}_E\\) over CLAIR suggest that each language model may have some bias (similar to each human), yet the ensemble of models correlates more strongly with human judgments. A reasonable concern might be that the models underlying existing approaches are significantly smaller than those in CLAIR, and trained on less data. To address this, we introduce and compare against RefCLIP-X, which replaces the CLIP model in RefCLIP with a CLIP ViT-bigG/14 model trained on LAION 2B (Ilharco et al., 2021). Even in this case, CLAIR demonstrates significantly improved performance. Note that in Table 1, we use \\(\\tau_b\\), instead of the form \\(\\tau_c\\), due to the ties generated by the CLAIR model."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.731,
+                0.884,
+                0.907
+            ],
+            "angle": 0,
+            "content": "System-level human correlation: In addition to computing the sample-level correlation on the MS-COCO dataset, we use the annotations from the five models considered by Rohrbach et al. (2018) to compute the system-level correlation. For each of the methods, we compute the mean human score on the test samples, and mean metric score on the test samples, followed by the Kendall's rank correlation coefficient (Kendall's tau, strength of ordinal association) between these values (the set of five mean human scores, and the set of five metric scores). Our"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.115,
+                0.084,
+                0.486,
+                0.121
+            ],
+            "angle": 0,
+            "content": "Table 2: System-level correlation between the average CLAIR score and human model evaluation for 5 models trained and evaluated on MS-COCO. All p-values \\( < {0.05} \\) ."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.132,
+                0.486,
+                0.298
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Measure</td><td>Kendall&#x27;s τb</td><td>Spearman&#x27;s ρ</td><td>Pearson r</td></tr><tr><td>BLEU@1</td><td>0.399</td><td>0.600</td><td>0.706</td></tr><tr><td>BLEU@4</td><td>0.799</td><td>0.899</td><td>0.910</td></tr><tr><td>ROUGE-L</td><td>0.600</td><td>0.700</td><td>0.792</td></tr><tr><td>METEOR</td><td>0.600</td><td>0.700</td><td>0.666</td></tr><tr><td>CIDEr</td><td>0.399</td><td>0.600</td><td>0.856</td></tr><tr><td>SPICE</td><td>0.399</td><td>0.600</td><td>0.690</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>0.799</td><td>0.899</td><td>0.869</td></tr><tr><td>+ Claude</td><td>1.000</td><td>1.000</td><td>0.868</td></tr><tr><td>+ PaLM</td><td>1.000</td><td>1.000</td><td>0.954</td></tr><tr><td>CLAIR_E</td><td>1.000</td><td>1.000</td><td>0.903</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.319,
+                0.486,
+                0.356
+            ],
+            "angle": 0,
+            "content": "Table 3: Accuracy of measures when matching human decisions for PASCAL-50S (5 reference captions). *: Model has access to additional visual context."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.366,
+                0.486,
+                0.575
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Measure</td><td>HC</td><td>HI</td><td>HM</td><td>MM</td><td>All</td></tr><tr><td>BLEU@1</td><td>51.20</td><td>95.70</td><td>91.20</td><td>58.20</td><td>74.08</td></tr><tr><td>BLEU@4</td><td>53.00</td><td>92.40</td><td>86.70</td><td>59.40</td><td>72.88</td></tr><tr><td>ROUGE-L</td><td>51.50</td><td>94.50</td><td>92.50</td><td>57.70</td><td>74.05</td></tr><tr><td>METEOR</td><td>56.70</td><td>97.60</td><td>94.20</td><td>63.40</td><td>77.98</td></tr><tr><td>CIDEr</td><td>53.00</td><td>98.00</td><td>91.50</td><td>64.50</td><td>76.75</td></tr><tr><td>SPICE</td><td>52.60</td><td>93.90</td><td>83.60</td><td>48.10</td><td>69.55</td></tr><tr><td>TIGEr*</td><td>56.00</td><td>99.80</td><td>92.80</td><td>74.20</td><td>80.70</td></tr><tr><td>CLIP-S*</td><td>56.50</td><td>99.30</td><td>96.40</td><td>70.40</td><td>80.70</td></tr><tr><td>RefCLIP-S*</td><td>64.50</td><td>99.60</td><td>95.40</td><td>72.80</td><td>83.10</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>52.40</td><td>99.50</td><td>89.80</td><td>73.00</td><td>78.67</td></tr><tr><td>+ Claude</td><td>57.90</td><td>98.50</td><td>91.30</td><td>62.90</td><td>77.65</td></tr><tr><td>+ PaLM</td><td>54.70</td><td>98.30</td><td>87.30</td><td>64.00</td><td>76.08</td></tr><tr><td>CLAIRE</td><td>57.70</td><td>99.80</td><td>94.60</td><td>75.60</td><td>81.93</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.593,
+                0.489,
+                0.691
+            ],
+            "angle": 0,
+            "content": "results, given in Table 2, demonstrate that CLAIR ranks the five methods in a novel way that is more accordant with human rankings of the methods. These results further suggest that CLAIR has the potential to redefine which methods are preferable to humans compared to existing n-gram approaches."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.695,
+                0.49,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Decision Making: In addition to evaluating the correlation with human judgments, we also evaluate the capability of the measure to perform discriminative analysis. The PASCAL-50S dataset (Vedantam et al., 2015) contains a set of 4000 human-annotated caption pairs. For each pair of captions, humans label which caption in the pair is closest to the reference set for the image. The caption pairs fall into four groups: \"HC:\" two human-written captions matching the image, \"HI:\" one human caption, and one machine-generated caption, with only one matching the image, \"HM:\" a matching human caption and a matching machine-generated caption and \"MM:\" two matching machine-generated captions."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.084,
+                0.882,
+                0.11
+            ],
+            "angle": 0,
+            "content": "Table 4: Pearson correlation with human judgments when evaluating sets of captions on MS-COCO (\\( \\bar{N} = 794 \\))."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.512,
+                0.12,
+                0.882,
+                0.336
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Measure</td><td colspan=\"2\">Coveragep-value</td><td colspan=\"2\">Correctnessp-value</td></tr><tr><td>BLEU@4</td><td>0.004</td><td>0.816</td><td>0.003</td><td>0.888</td></tr><tr><td>ROUGE-L</td><td>0.011</td><td>0.563</td><td>0.038</td><td>0.184</td></tr><tr><td>METEOR</td><td>0.016</td><td>0.398</td><td>0.006</td><td>0.765</td></tr><tr><td>CIDEr</td><td>0.004</td><td>0.844</td><td>0.026</td><td>0.173</td></tr><tr><td>TRM-METEOR</td><td colspan=\"2\">0.128&lt;0.001</td><td colspan=\"2\">0.108&lt;0.001</td></tr><tr><td>TRM-BLEU</td><td colspan=\"2\">0.127&lt;0.001</td><td colspan=\"2\">0.151&lt;0.001</td></tr><tr><td>MMD-BERT</td><td colspan=\"2\">0.129&lt;0.001</td><td colspan=\"2\">0.124&lt;0.001</td></tr><tr><td>FID-BERT</td><td>0.081</td><td>0.011</td><td colspan=\"2\">0.098&lt;0.001</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.195</td><td>0.011</td><td>0.187</td><td>0.014</td></tr><tr><td>+Claude</td><td>0.110</td><td>0.099</td><td>0.124</td><td>0.145</td></tr><tr><td>+PaLM</td><td>0.129</td><td>0.081</td><td>0.085</td><td>0.172</td></tr><tr><td>CLAIR_E</td><td>0.183</td><td>0.027</td><td>0.156</td><td>0.018</td></tr><tr><td>Inter-Human</td><td colspan=\"2\">0.225&lt;0.001</td><td colspan=\"2\">0.274&lt;0.001</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.364,
+                0.848,
+                0.379
+            ],
+            "angle": 0,
+            "content": "See Appendix B for more dataset information."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.384,
+                0.884,
+                0.626
+            ],
+            "angle": 0,
+            "content": "The performance on PASCAL-50S is given in Table 3. We can see that \\(\\mathrm{CLAIR}_E\\) outperforms all existing text-only measures (e.g., by \\(5.18\\%\\) overall score over CIDEr), and in many cases, even outperforms measures that have access to the image at test time. Note that it is relatively weaker than image-augmented models in the HC setting; however, since both captions are correct, the model often cannot judge which is better purely the text. Models such as RefCLIP-S that have access to the image are naturally better discriminators in this case. We suspect that CLAIR's discriminative performance could be further improved by giving the LLM a choice between the two captions; however, we leave this optimization to future work."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.629,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Groups of Captions: While CLAIR is capable of comparing a single candidate caption to a set of reference captions, it is also capable of comparing sets of candidate captions to sets of reference captions. This task is necessary when evaluating the ability of a model to generate captions that are diverse and that fully describe the conditional text distribution. We evaluate on the COCO-Sets dataset (Chan et al., 2022), 794 caption sets rated by AMT workers on two scales: how closely a candidate set matches the reference set in terms of both correctness and content coverage (See Appendix B for details). The results of this experiment are given in Table 4. We can see that CLAIR outperforms well when measuring the quality of a group of captions, and approaches the inter-human correlation on the (very) challenging task. CLAIR also outperforms TRM-Meteor and TRM-BLEU (Chan et al.,"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.085,
+                0.488,
+                0.117
+            ],
+            "angle": 0,
+            "content": "2022), suggesting that LLMs can judge both the content and diversity of the caption sets."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.134,
+                0.251,
+                0.15
+            ],
+            "angle": 0,
+            "content": "4 Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.165,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "While CLAIR correlates well with human judgments of caption quality, it has several limitations:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.202,
+                0.49,
+                0.524
+            ],
+            "angle": 0,
+            "content": "Non-Determinism and Parsing Errors: Because CLAIR depends on the output of a language model, the measure can be non-deterministic and noisy. For instance, it may fail to elicit a judgment (e.g., \"As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references\"), or rarely, generate malformed JSON output. To address these issues, we perform multiple queries to the LLM, sometimes at higher temperatures if necessary. As a consequence, the measure may differ between runs, although we found the variance to be relatively insignificant (< 0.01 in many of the experiments). Additionally, since the language models used are not open-source, the models are subject to arbitrary change, replacement, or removal, which limits the efficacy of the measure as a long-term comparable measurement. We hope that increasing open access to language models with efforts such as Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023), will help to alleviate these challenges in the future."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.529,
+                0.49,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Increased Cost: CLAIR relies on language models which contain many billions of parameters. These language models have not only monetary cost but also human and environmental costs (Bender et al., 2021) which can reduce its utility as a target during training, such as for self-critical sequence training (Rennie et al., 2017). While API-based LLMs may be considered costly, even open-source LLMs have a cost (which can often be hard to quantify). CLAIR on the MS-COCO dataset uses an average of 226.148 tokens per sample (on OpenAI's API), representing a cost of \\(0.0067 per sample (GPT-4), or \\)0.00033 per sample (GPT 3.5). For PALM, this drops to $0.000113 per sample. We hope that over time, advances in LLM inference (such as quantization and distillation), coupled with improvements in architecture will continue to yield lower-cost alternatives with strong performance on the caption evaluation task."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Hallucination: While CLAIR does suffer from potential hallucination, we strongly believe that this weakness does not diminish the fact that CLAIR still correlates strongly with human judgment. In CLAIR, hallucinations in the score manifest as"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.085,
+                0.885,
+                0.327
+            ],
+            "angle": 0,
+            "content": "\"incorrect\" judgements of similarity, while hallucinations in the explanations manifest as poorly grounded explanations of the score/quality. Hallucinations in the score should be considered false negatives (blind spots instead of hallucinations). In the case of hallucinations in the explanations, such hallucinations may lead to misinterpretation, but arguably less misinterpretation than a black box method, and may even indicate misunderstandings in the model. Hallucination is a well-known challenge of current LLMs and is the subject of a great amount of research on instruction-tuning, RLHF, RLAIF, and other methods. As hallucination and instruction-following performance of the base models improves, CLAIR inherit similar improvements."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.331,
+                0.885,
+                0.461
+            ],
+            "angle": 0,
+            "content": "Explainability: While CLAIR generates explanations for each rating, CLAIR has no strict scoring rubric. Much like human judgments, there is no direct way of attributing changes in score to changes in caption quality. For similar reasons, it is difficult to evaluate the quality of the generated explanations. Qualitatively, the explanations are generally reasonable and consider multiple axes of judgment."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.476,
+                0.642,
+                0.492
+            ],
+            "angle": 0,
+            "content": "5 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.507,
+                0.885,
+                0.732
+            ],
+            "angle": 0,
+            "content": "This work introduces CLAIR, an LLM-based evaluation measure for image captioning. CLAIR's superior performance compared to highly-engineered measures indicates a remarkable fact: LLMs are well aligned with human judgments of caption quality, even more so than some measures designed specifically for semantic similarity. CLAIR is only a glimpse into how LLMs can be used for evaluation tasks, and image captioning is only the beginning. We hope that our work will inspire further exploration of similar measures in other vision and language domains, such as visual storytelling (Huang et al., 2016), where human evaluation of generated text remains a challenging task."
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.112,
+                0.487,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Somak Aditya, Yezhou Yang, Chitta Baral, Cornelia Fermuller, and Yiannis Aloimonos. 2015. From images to sentences through scene description graphs using commonsense reasoning and knowledge. ArXiv preprint, abs/1511.03292."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.188,
+                0.487,
+                0.268
+            ],
+            "angle": 0,
+            "content": "Abhaya Agarwal and Alon Lavie. 2008. Meteor, M-BLEU and M-TER: Evaluation metrics for high-correlation with human rankings of machine translation output. In Proceedings of the Third Workshop on Statistical Machine Translation, pages 115-118. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.277,
+                0.486,
+                0.332
+            ],
+            "angle": 0,
+            "content": "Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2016. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.341,
+                0.487,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. ArXiv preprint, abs/2204.05862."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.43,
+                0.487,
+                0.508
+            ],
+            "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, pages 610-623."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.519,
+                0.487,
+                0.716
+            ],
+            "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,
+                0.726,
+                0.487,
+                0.793
+            ],
+            "angle": 0,
+            "content": "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, abs/2303.12712."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.803,
+                0.487,
+                0.868
+            ],
+            "angle": 0,
+            "content": "David M Chan, Yiming Ni, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. 2022. Distribution aware metrics for conditional natural language generation. ArXiv preprint, abs/2209.07518."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.878,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.112,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.53,
+                0.086,
+                0.882,
+                0.114
+            ],
+            "angle": 0,
+            "content": "Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.122,
+                0.882,
+                0.162
+            ],
+            "angle": 0,
+            "content": "Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. Visual programming for text-to-image generation and evaluation. arXiv preprint arXiv:2305.15328."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.172,
+                0.882,
+                0.249
+            ],
+            "angle": 0,
+            "content": "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. ArXiv preprint, abs/2204.02311."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.26,
+                0.882,
+                0.352
+            ],
+            "angle": 0,
+            "content": "Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4299-4307."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.362,
+                0.882,
+                0.402
+            ],
+            "angle": 0,
+            "content": "Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. *ArXiv preprint*, abs/2305.14314."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.411,
+                0.882,
+                0.464
+            ],
+            "angle": 0,
+            "content": "Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. Koala: A dialogue model for academic research. Blog post."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.473,
+                0.882,
+                0.565
+            ],
+            "angle": 0,
+            "content": "Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.575,
+                0.882,
+                0.628
+            ],
+            "angle": 0,
+            "content": "Micah Hodosh, Peter Young, and Julia Hockenmaier. 2013. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.637,
+                0.882,
+                0.703
+            ],
+            "angle": 0,
+            "content": "Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. 2023. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.713,
+                0.882,
+                0.856
+            ],
+            "angle": 0,
+            "content": "Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, and Margaret Mitchell. 2016. Visual storytelling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1233-1239, San Diego, California. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.866,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.135,
+                0.086,
+                0.489,
+                0.113
+            ],
+            "angle": 0,
+            "content": "Schmidt. 2021. Openclip. If you use this software, please cite it as below."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.125,
+                0.487,
+                0.243
+            ],
+            "angle": 0,
+            "content": "Ming Jiang, Qiuyuan Huang, Lei Zhang, Xin Wang, Pengchuan Zhang, Zhe Gan, Jana Diesner, and Jianfeng Gao. 2019. TIGEr: Text-to-image grounding for image caption evaluation. 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 2141-2152. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.255,
+                0.487,
+                0.358
+            ],
+            "angle": 0,
+            "content": "Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.372,
+                0.487,
+                0.424
+            ],
+            "angle": 0,
+            "content": "Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.437,
+                0.487,
+                0.503
+            ],
+            "angle": 0,
+            "content": "Lucian Vlad Lita, Monica Rogati, and Alon Lavie. 2005. Blanc: Learning evaluation metrics for mt. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 740-747."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.515,
+                0.487,
+                0.567
+            ],
+            "angle": 0,
+            "content": "Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L Yuille. 2014. Explain images with multimodal recurrent neural networks. arXiv preprint arXiv:1410.1090."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.58,
+                0.357,
+                0.594
+            ],
+            "angle": 0,
+            "content": "OpenAI. 2022. Introducing chatgpt."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.606,
+                0.487,
+                0.685
+            ],
+            "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. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.696,
+                0.487,
+                0.827
+            ],
+            "angle": 0,
+            "content": "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.84,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. 2017. Self-critical sequence training for image captioning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1179-1195. IEEE Computer Society."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.489,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.513,
+                0.086,
+                0.882,
+                0.165
+            ],
+            "angle": 0,
+            "content": "Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.175,
+                0.882,
+                0.254
+            ],
+            "angle": 0,
+            "content": "Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575. IEEE Computer Society."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.264,
+                0.882,
+                0.381
+            ],
+            "angle": 0,
+            "content": "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 23318-23340. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.391,
+                0.882,
+                0.47
+            ],
+            "angle": 0,
+            "content": "Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.48,
+                0.882,
+                0.545
+            ],
+            "angle": 0,
+            "content": "Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. 2023. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.556,
+                0.882,
+                0.621
+            ],
+            "angle": 0,
+            "content": "Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.631,
+                0.882,
+                0.696
+            ],
+            "angle": 0,
+            "content": "Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. ArXiv preprint, abs/2306.05685."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.882,
+                0.696
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.715,
+                0.598,
+                0.731
+            ],
+            "angle": 0,
+            "content": "Appendix"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.513,
+                0.745,
+                0.715,
+                0.761
+            ],
+            "angle": 0,
+            "content": "A Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.775,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We thank Suhong Moon and Kate Saenko for their helpful comments on the work. Authors, as part of their affiliation with UC Berkeley, were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program, as well as gifts from Anyscale, Astronomer, Google, IBM, Intel, Lacework, Microsoft, Mohamed Bin Zayed University of Artificial Intelligence, Samsung SDS, Uber, and VMware."
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.084,
+                0.431,
+                0.101
+            ],
+            "angle": 0,
+            "content": "B Additional Experimental Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.119,
+                0.488,
+                0.166
+            ],
+            "angle": 0,
+            "content": "In this section, we provide several additional details for the experiments in section 3 run with the CLAIR measure."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.191,
+                0.367,
+                0.208
+            ],
+            "angle": 0,
+            "content": "B.1 Input Prompt Formatting"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.22,
+                0.49,
+                0.446
+            ],
+            "angle": 0,
+            "content": "The CLAIR prompt is given in its entirety in Figure 1. During run-time, candidate and reference captions are prefixed with a “-” and inserted into the prompt, one per line. The resulting query is passed to the large language model. In addition, for models which were not RLHF-tuned to perform conversation (such as PaLM), we found that it was helpful to append an additional prefix {\"score\": to the beginning of the output, to encourage the correct output formatting. CLAIR is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. The model checkpoint metadata is generally unknown (as the APIs are somewhat fluid and evolving)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.469,
+                0.395,
+                0.485
+            ],
+            "angle": 0,
+            "content": "B.2 LLM Output Post-Processing"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.497,
+                0.49,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Because CLAIR relies on an LLM to produce output, there is no guarantee that the output will be in the format we expect (i.e. valid, parsable JSON). To extract both the score and the reason, we first extract the first set of paired braces from the output of the LLM and attempt to parse the result as JSON. In most cases (99.997% for GPT-3, 99.991% for Claude, and 99.94% for PaLM during the course of our experiments), this is successful, and the score and reason are returned. In the case that the JSON output is malformed, we attempt to extract any sequence of digits from the LLM to use as a score, and set the reason to \"Unknown.\" When this fails, as can be the case when the models produce an output such as \"As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references\", we retry the prompt at a higher temperature (\\(t = 1.0\\)) several times. Failing this (which occurred only three times in the entire evaluation of this paper, across several hundred thousand calls), we set the score to 0 for the caption."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.859,
+                0.232,
+                0.873
+            ],
+            "angle": 0,
+            "content": "B.3 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.887,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this section, we provide additional detail regarding the datasets used in the evaluations in section 3."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.084,
+                0.885,
+                0.262
+            ],
+            "angle": 0,
+            "content": "COMPOSITE: The COMPOSITE dataset (Aditya et al., 2015) contains machine-generated test captions for 3995 images spread across the MS-COCO (Xu et al., 2016), Flickr8K (Mao et al., 2014) and Flickr30k (Young et al., 2014) datasets. Each image has three test captions, one written by a human, and two that are model generated. The candidate captions are graded by annotators on Amazon Mechanical Turk (AMT) on a scale of 1 (not relevant) to 5 (very relevant). Inter-human correlations are not available for this dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.277,
+                0.885,
+                0.472
+            ],
+            "angle": 0,
+            "content": "Flickr8K-Expert: The Flickr8K-Expert dataset (Hodosh et al., 2013) contains 5822 captions associated with 1000 images. The dataset is annotated with expert human judgments of quality, where images are rated from 1 (caption is unrelated to the image) to 4 (caption describes the image without errors). Unlike the composite and MS-COCO datasets, the captions here are selected using an image retrieval system, instead of generated using a learned image captioning model. Following Jiang et al. (2019), we exclude any candidate captions that overlap the reference set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.486,
+                0.884,
+                0.646
+            ],
+            "angle": 0,
+            "content": "MS-COCO: Following experiments by Rohrbach et al. (2018), we compute the sample-level correlation between our method and human ratings on a 500-image subset of the MS-COCO Karpathy test set. Each image in the subset contains candidate captions generated by 5 models, and each caption is labeled with the average three human ratings generated by AMT workers which range from 1 (very bad) to 5 (very good). Inter-human correlations are not available for this dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.662,
+                0.884,
+                0.792
+            ],
+            "angle": 0,
+            "content": "PASCAL-50S: PASCAL-50S contains 1000 images drawn from the PASCAL sentence dataset. Each image is associated with at least 50 (and as many as 120) reference captions. In addition to the reference captions, PASCAL-50S contains a set of 4000 human annotated image/caption pairs containing an image, and two candidate captions. The caption pairs fall into four groups:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.525,
+                0.809,
+                0.881,
+                0.856
+            ],
+            "angle": 0,
+            "content": "1. HC: In the HC group, both captions in the pair are human written, and describe the content of the target image correctly."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.524,
+                0.871,
+                0.881,
+                0.918
+            ],
+            "angle": 0,
+            "content": "2. HI: In the HI group, both captions in the pair are human written, but one caption correctly describes the content of the image, and the"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.524,
+                0.809,
+                0.881,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.086,
+                0.486,
+                0.117
+            ],
+            "angle": 0,
+            "content": "other caption describes the content of a different image."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.128,
+                0.486,
+                0.191
+            ],
+            "angle": 0,
+            "content": "3. HM: In the HM group, one caption is written by a human, and one caption is written by a machine, but both correctly describe the content of the image."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.203,
+                0.486,
+                0.25
+            ],
+            "angle": 0,
+            "content": "4. MM: In the MM group, both captions are written by a machine, and both correctly describe the image content."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.13,
+                0.128,
+                0.486,
+                0.25
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.267,
+                0.489,
+                0.362
+            ],
+            "angle": 0,
+            "content": "In PASCAL-50S, the task is to decide which caption in the pair humans prefer more (a subjective task, hopefully indicating caption quality). Following previous work (Jiang et al., 2019; Hessel et al., 2021), we limit the number of reference sentences to five during evaluation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.377,
+                0.489,
+                0.618
+            ],
+            "angle": 0,
+            "content": "COCO-Sets: The COCO-Sets dataset (Chan et al., 2022) is a set of samples that are designed to evaluate the correlation of distribution-aware image captioning measures with human judgments of distributional distance. In this dataset, humans were presented with two candidate caption sets (two image captioning models, OFA (Wang et al., 2022) and BLIP (Li et al., 2022) using different temperatures), and asked which candidate caption set correlated better with a reference caption set on two measures: how much they overlapped factually (correctness), and how much information they provided about the references (coverage). It consists of 794 AMT worker-generated judgments of caption quality for images in the MS-COCO dataset."
+        }
+    ]
+]

2023/CLAIR_ Evaluating Image Captions with Large Language Models/ed931ad6-a4d3-4ac0-93d7-f7030245fd98_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1472ffa190737aa2822f17242c293d83817f30441de2e7c0abbcf48f37cdfa2
+size 2523998

2023/CLAIR_ Evaluating Image Captions with Large Language Models/full.md

@@ -0,0 +1,236 @@
+# CLAIR: Evaluating Image Captions with Large Language Models
+
+David M. Chan, Suzanne Petryk, Joseph E. Gonzalez, Trevor Darrell, John Canny  
+University of California, Berkeley
+
+{davidchan,spetryk,jegonzal,trevordarrell,canny}@berkeley.edu
+
+# Abstract
+
+The evaluation of machine-generated image captions poses an interesting yet persistent challenge. Effective evaluation measures must consider numerous dimensions of similarity, including semantic relevance, visual structure, object interactions, caption diversity, and specificity. Existing highly-engineered measures attempt to capture specific aspects, but fall short in providing a holistic score that aligns closely with human judgments. Here, we propose CLAIR<sup>1</sup>, a novel method that leverages the zero-shot language modeling capabilities of large language models (LLMs) to evaluate candidate captions. In our evaluations, CLAIR demonstrates a stronger correlation with human judgments of caption quality compared to existing measures. Notably, on Flickr8K-Expert, CLAIR achieves relative correlation improvements over SPICE of  $39.6\%$  and over image-augmented methods such as RefCLIP-S of  $18.3\%$ . Moreover, CLAIR provides noisily interpretable results by allowing the language model to identify the underlying reasoning behind its assigned score. Code is available at https://davidmchan.github.io/clair/.
+
+# 1 Introduction & Background
+
+Automatically evaluating the quality of image captions is challenging. There are many dimensions to consider, such as grammatical quality, semantic relevance, correctness, and specificity, among others. To ensure fair evaluations, most image captioning works employ a suite of measures, each capturing different aspects. For instance, n-gram-based measures like BLEU (Papineni et al., 2002) or CIDEr (Vedantam et al., 2015) broadly measure content overlap, SPICE (Anderson et al., 2016) compares scene graph structures, and CLIPScore, TIFA, SeeTrue and VPEval (Hessel
+
+You are trying to tell if a candidate set of captions is describing the same image as a reference set of captions.
+
+Candidate set:
+
+{candidate captions}
+
+Reference set:
+
+{reference captions}
+
+On a precise scale from 0 to 100, how likely is it that the candidate set is describing the same image as the reference set?
+
+(JSON format, with a key “score”, value between 0 and 100, and a key “reason” with a string value.)
+
+Figure 1: CLAIR: a (surprisingly simple) large language model-based measure for image caption evaluation. We find that CLAIR not only correlates strongly with human judgments of caption quality but can also generate interpretable reasons for the generated scores.
+
+et al., 2021; Hu et al., 2023; Yarom et al., 2023; Cho et al., 2023) directly incorporate visual information. Unfortunately, while significant strides have been made in automated evaluation, human preference studies remain the most reliable (yet costly) source of caption evaluation.
+
+Fortunately, recent advances in large language models (LLMs) have opened new avenues for automatic evaluation. Models trained with reinforcement learning from human feedback (RLHF, Christiano et al. (2017)) or similar methods are particularly useful for open-ended evaluation tasks, including image captioning, due to their explicit training to align with human preferences.
+
+In our work, paralleling several recent works which find that LLMs can act as effective "judges" for selecting the better answer from two candidates (Bubeck et al., 2023; Dettmers et al., 2023; Chiang et al., 2023), we explore the ability of LLMs to evaluate caption quality in the multimodal setting.
+
+We introduce CLAIR (Criterion using LLanguage models for Image caption Rating), a measure which scores a candidate caption based on the likelihood that it describes the same image as a set of references by directly asking an LLM to produce a numeric rating. We further query the LLM to provide a reason for its score, providing a level of interpretability to the scalar rating. As far as we are aware, this is the first paper to explore replacing measures of semantic text quality with directly obtained LLM judgments, however concurrently, Zheng et al. (2023) have shown that directly providing an answer rating can align highly with human preferences on a range of standard language-based tasks, such as conversational instruction following.
+
+Through several experiments on captioning datasets such as MS-COCO (Xu et al., 2016), Flickr8k (Mao et al., 2014), and PASCAL50S (Vedantam et al., 2015), we find that CLAIR correlates surprisingly well with human preferences, outperforming prior captioning measures. We additionally propose  $\mathrm{CLAIR}_E$ , where we Ensemble the outputs of several LLMs by taking the average score, leading to further improvements.
+
+Despite a simple pipeline using an LLM prompt with minimal output parsing, CLAIR's strong correlation with human preferences suggests that it captures multiple dimensions of caption similarity at once - a feature that prior measures struggle to achieve alone. More generally, CLAIR demonstrates how language-only models can evaluate vision-language tasks. We show LLMs can provide not only reliable scalar ratings but also corresponding reasoning for a given rating, offering a valuable combination of accuracy and interpretability.
+
+# 2 CLAIR: LLMs for Caption Evaluation
+
+In CLAIR, we adapt the zero-shot in-context learning approach described in Brown et al. (2020) to score candidate captions with large language models (LLMs). This involves converting the caption evaluation problem into a human-readable text completion task which is solved by the LLM. Using the prompt in Figure 1, CLAIR first generates completions from the LLM and then parses those completions into both candidate scores and an explainable reason for the score. We use a greedy sampling method  $(t = 0)$  to encourage reproducibility in the results, while acknowledging the inherent nondeterminism in LLMs (see section 4). CLAIR's
+
+experimental implementation is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. See Appendix B for further implementation details.
+
+The choice of language model directly affects the quality of the CLAIR measure – more accurate models should produce evaluations that align better with human judgment. We explore three language models: GPT-3.5 (ChatGPT) (OpenAI, 2022), Claude (Instant) (Bai et al., 2022), and PaLM (Chowdhery et al., 2022). Unfortunately, we found for several open-weight language models including Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023) that CLAIR aligned poorly with human judgment.
+
+As CLAIR is language model-agnostic, we can leverage the different distributions learned by each model and combine their decisions in an ensemble approach we term  $\mathrm{CLAIR}_E$ . We calculate individual CLAIR scores for each model and compute an unweighted average to obtain the ensemble score.
+
+Benchmark measures: We benchmark against several existing measure of caption similarity. BLEU (Papineni et al., 2002), ROUGE (Lin, 2004), METEOR (Agarwal and Lavie, 2008) and CIDEr (Vedantam et al., 2015) all primarily measure n-gram overlap (however have different weighting schemes between n-grams, and across precision/recall). We also compare against SPICE (Anderson et al., 2016), which compares caption parse trees and focuses on matching perceived action and object relationships in addition to n-grams. While the aforementioned measures are commonly reported in image captioning works, we also compare against several modern measures, including CLIP-Score (Hessel et al., 2021) which uses the recent CLIP (Radford et al., 2021) model for reference-free evaluation.
+
+# 3 Evaluation & Discussion
+
+To evaluate the quality of the measure, we run several evaluations that compare scores generated by CLAIR to both human judgments of caption quality and other image captioning evaluation measures. We additionally provide several qualitative examples in Figure 2. A unique benefit of CLAIR is that it provides not only numeric scores but is also introspectable, as it can identify which details in the candidate caption set match the reference set.
+
+Sample-level human correlation: We first ask the question, how well does CLAIR correlate with
+
+![](images/7417d5ef07db21535534ee9344cada5add8fb593760d7ca11d0d3697447e1f9b.jpg)  
+Candidate  
+A boy jumps into the blue pool water.
+
+# CLAIR Reason
+
+The candidate set and reference set both mention a boy jumping into water, but the candidate set does not provide as much detail about the boy's appearance or the location of the pool. Additionally, the reference set includes captions about diving off a pier and holding one's nose, which are not mentioned in the candidate set.
+
+# References
+
+"A blonde boy wearing a blue life vest and ca mo shorts jumps off a diveboard into a blue lake." "A boy diving backward off a pier in a large lake." "A boy in a blue life jacket jumps into the water." A boy in a blue life jacket jumps off a board into the lake. "A boy is holding his nose and jumping off a diving board backwards into a lake."
+
+![](images/31972ec12e7b874e7b669083118e7acb8ce0691ec34a7c8f7e33b48d8002ea78.jpg)  
+Candidate  
+A cyclist in a helmet and yellow jersey.  
+Figure 2: Several qualitative examples of CLAIR from the Flickr8K-Expert dataset. CLAIR not only correlates better with human judgments of caption quality but also provides detailed explanations for its score. CLAIR scores normalized by 100.
+
+# CLAIR Reason
+
+The candidate set only mentions a cyclist in a helmet and yellow jersey, while the reference set describes a man in a green shirt carrying a drink while riding a bicycle. Although there are some similarities, such as the fact that both sets mention a person riding a bike, the differences in the details make it uncertain whether they are describing the same image or not.
+
+# References
+
+"A biker enjoys a coffee." "A man in a bright green shirt and sunglasses is riding a bicycle through the streets whilst drinking a latte." "A man in a bright green shirt riding a bicycle in a paved courtyard, carrying a drink." "A man rides on his bike with one hand and holds a drink with the other." "A man riding a bike wearing a green shirt with a drink in his hand."
+
+![](images/50f9d861ee688dc38e0ae89fa571e3d73e37b589470a71d9ee02d47c059bc6fb.jpg)  
+Candidate  
+Person snowboarding at a ski slope.
+
+# CLAIR Reason
+
+The candidate caption mentions only one person snowboarding, while all the reference captions mention multiple people skiing or climbing a mountain. Additionally, the candidate caption does not mention any details about the mountain or the surroundings, which are present in the reference captions.
+
+# References
+
+"a group of skiers going up a mountain." "Four cross-country skiers climb uphill." "Four people climbing a hill in the snow." "Four skiers walking up a snow covered hill." "four skiers climbing snow dropped mountain."
+
+Table 1: Sample-level correlation (Kendall's  $\tau_b$ ) with human judgments. All p-values  $< 0.001$ .*: Model has access to additional visual context.
+
+<table><tr><td rowspan="2">Measure</td><td colspan="3">Dataset</td></tr><tr><td>COMPOSITE</td><td>Flickr8K</td><td>MS-COCO</td></tr><tr><td>BLEU@1</td><td>0.313</td><td>0.323</td><td>0.265</td></tr><tr><td>BLEU@4</td><td>0.306</td><td>0.308</td><td>0.215</td></tr><tr><td>ROUGE-L</td><td>0.324</td><td>0.323</td><td>0.221</td></tr><tr><td>BERT-S</td><td>0.301</td><td>0.392</td><td>0.163</td></tr><tr><td>METEOR</td><td>0.389</td><td>0.418</td><td>0.239</td></tr><tr><td>CIDEr</td><td>0.377</td><td>0.439</td><td>0.262</td></tr><tr><td>SPICE</td><td>0.403</td><td>0.449</td><td>0.257</td></tr><tr><td>CLIP-S*</td><td>0.498</td><td>0.511</td><td>-</td></tr><tr><td>RefCLIP-S*</td><td>0.512</td><td>0.526</td><td>-</td></tr><tr><td>RefCLIP-X*</td><td>0.523</td><td>0.549</td><td>0.274</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.604</td><td>0.616</td><td>0.296</td></tr><tr><td>+Claude</td><td>0.542</td><td>0.563</td><td>0.320</td></tr><tr><td>+PaLM</td><td>0.580</td><td>0.546</td><td>0.355</td></tr><tr><td>CLAIRE</td><td>0.592</td><td>0.627</td><td>0.374</td></tr><tr><td>Inter-Human</td><td>-</td><td>0.736</td><td>-</td></tr></table>
+
+human judgments of caption quality at a sample level? We do so by exploring the performance on three datasets, COMPOSITE, Flickr8K-Expert, and MS-COCO (See Appendix B for details).
+
+The results of our sample-level correlation experiments are shown in Table 1. We can see that CLAIR outperforms language-only measures (e.g., 0.604 to 0.403 for SPICE), and in most cases, outperforms vision-augmented measures.  $\mathrm{CLAIR}_E$  achieves
+
+strong sample-level correlation on all datasets; for instance,  $\mathrm{CLAIR}_E$  closes the gap to inter-human agreement by 0.097 over vision-based measures and 0.132 over language-based measures. The improvements of  $\mathrm{CLAIR}_E$  over CLAIR suggest that each language model may have some bias (similar to each human), yet the ensemble of models correlates more strongly with human judgments. A reasonable concern might be that the models underlying existing approaches are significantly smaller than those in CLAIR, and trained on less data. To address this, we introduce and compare against RefCLIP-X, which replaces the CLIP model in RefCLIP with a CLIP ViT-bigG/14 model trained on LAION 2B (Ilharco et al., 2021). Even in this case, CLAIR demonstrates significantly improved performance. Note that in Table 1, we use  $\tau_b$ , instead of the form  $\tau_c$ , due to the ties generated by the CLAIR model.
+
+System-level human correlation: In addition to computing the sample-level correlation on the MS-COCO dataset, we use the annotations from the five models considered by Rohrbach et al. (2018) to compute the system-level correlation. For each of the methods, we compute the mean human score on the test samples, and mean metric score on the test samples, followed by the Kendall's rank correlation coefficient (Kendall's tau, strength of ordinal association) between these values (the set of five mean human scores, and the set of five metric scores). Our
+
+Table 2: System-level correlation between the average CLAIR score and human model evaluation for 5 models trained and evaluated on MS-COCO. All p-values  $< {0.05}$  .  
+
+<table><tr><td>Measure</td><td>Kendall&#x27;s τb</td><td>Spearman&#x27;s ρ</td><td>Pearson r</td></tr><tr><td>BLEU@1</td><td>0.399</td><td>0.600</td><td>0.706</td></tr><tr><td>BLEU@4</td><td>0.799</td><td>0.899</td><td>0.910</td></tr><tr><td>ROUGE-L</td><td>0.600</td><td>0.700</td><td>0.792</td></tr><tr><td>METEOR</td><td>0.600</td><td>0.700</td><td>0.666</td></tr><tr><td>CIDEr</td><td>0.399</td><td>0.600</td><td>0.856</td></tr><tr><td>SPICE</td><td>0.399</td><td>0.600</td><td>0.690</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>0.799</td><td>0.899</td><td>0.869</td></tr><tr><td>+ Claude</td><td>1.000</td><td>1.000</td><td>0.868</td></tr><tr><td>+ PaLM</td><td>1.000</td><td>1.000</td><td>0.954</td></tr><tr><td>CLAIR_E</td><td>1.000</td><td>1.000</td><td>0.903</td></tr></table>
+
+Table 3: Accuracy of measures when matching human decisions for PASCAL-50S (5 reference captions). *: Model has access to additional visual context.  
+
+<table><tr><td>Measure</td><td>HC</td><td>HI</td><td>HM</td><td>MM</td><td>All</td></tr><tr><td>BLEU@1</td><td>51.20</td><td>95.70</td><td>91.20</td><td>58.20</td><td>74.08</td></tr><tr><td>BLEU@4</td><td>53.00</td><td>92.40</td><td>86.70</td><td>59.40</td><td>72.88</td></tr><tr><td>ROUGE-L</td><td>51.50</td><td>94.50</td><td>92.50</td><td>57.70</td><td>74.05</td></tr><tr><td>METEOR</td><td>56.70</td><td>97.60</td><td>94.20</td><td>63.40</td><td>77.98</td></tr><tr><td>CIDEr</td><td>53.00</td><td>98.00</td><td>91.50</td><td>64.50</td><td>76.75</td></tr><tr><td>SPICE</td><td>52.60</td><td>93.90</td><td>83.60</td><td>48.10</td><td>69.55</td></tr><tr><td>TIGEr*</td><td>56.00</td><td>99.80</td><td>92.80</td><td>74.20</td><td>80.70</td></tr><tr><td>CLIP-S*</td><td>56.50</td><td>99.30</td><td>96.40</td><td>70.40</td><td>80.70</td></tr><tr><td>RefCLIP-S*</td><td>64.50</td><td>99.60</td><td>95.40</td><td>72.80</td><td>83.10</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>+ GPT3.5</td><td>52.40</td><td>99.50</td><td>89.80</td><td>73.00</td><td>78.67</td></tr><tr><td>+ Claude</td><td>57.90</td><td>98.50</td><td>91.30</td><td>62.90</td><td>77.65</td></tr><tr><td>+ PaLM</td><td>54.70</td><td>98.30</td><td>87.30</td><td>64.00</td><td>76.08</td></tr><tr><td>CLAIRE</td><td>57.70</td><td>99.80</td><td>94.60</td><td>75.60</td><td>81.93</td></tr></table>
+
+results, given in Table 2, demonstrate that CLAIR ranks the five methods in a novel way that is more accordant with human rankings of the methods. These results further suggest that CLAIR has the potential to redefine which methods are preferable to humans compared to existing n-gram approaches.
+
+Decision Making: In addition to evaluating the correlation with human judgments, we also evaluate the capability of the measure to perform discriminative analysis. The PASCAL-50S dataset (Vedantam et al., 2015) contains a set of 4000 human-annotated caption pairs. For each pair of captions, humans label which caption in the pair is closest to the reference set for the image. The caption pairs fall into four groups: "HC:" two human-written captions matching the image, "HI:" one human caption, and one machine-generated caption, with only one matching the image, "HM:" a matching human caption and a matching machine-generated caption and "MM:" two matching machine-generated captions.
+
+Table 4: Pearson correlation with human judgments when evaluating sets of captions on MS-COCO ( $\bar{N} = 794$ ).  
+
+<table><tr><td>Measure</td><td colspan="2">Coveragep-value</td><td colspan="2">Correctnessp-value</td></tr><tr><td>BLEU@4</td><td>0.004</td><td>0.816</td><td>0.003</td><td>0.888</td></tr><tr><td>ROUGE-L</td><td>0.011</td><td>0.563</td><td>0.038</td><td>0.184</td></tr><tr><td>METEOR</td><td>0.016</td><td>0.398</td><td>0.006</td><td>0.765</td></tr><tr><td>CIDEr</td><td>0.004</td><td>0.844</td><td>0.026</td><td>0.173</td></tr><tr><td>TRM-METEOR</td><td colspan="2">0.128&lt;0.001</td><td colspan="2">0.108&lt;0.001</td></tr><tr><td>TRM-BLEU</td><td colspan="2">0.127&lt;0.001</td><td colspan="2">0.151&lt;0.001</td></tr><tr><td>MMD-BERT</td><td colspan="2">0.129&lt;0.001</td><td colspan="2">0.124&lt;0.001</td></tr><tr><td>FID-BERT</td><td>0.081</td><td>0.011</td><td colspan="2">0.098&lt;0.001</td></tr><tr><td>CLAIR</td><td></td><td></td><td></td><td></td></tr><tr><td>+GPT3.5</td><td>0.195</td><td>0.011</td><td>0.187</td><td>0.014</td></tr><tr><td>+Claude</td><td>0.110</td><td>0.099</td><td>0.124</td><td>0.145</td></tr><tr><td>+PaLM</td><td>0.129</td><td>0.081</td><td>0.085</td><td>0.172</td></tr><tr><td>CLAIR_E</td><td>0.183</td><td>0.027</td><td>0.156</td><td>0.018</td></tr><tr><td>Inter-Human</td><td colspan="2">0.225&lt;0.001</td><td colspan="2">0.274&lt;0.001</td></tr></table>
+
+See Appendix B for more dataset information.
+
+The performance on PASCAL-50S is given in Table 3. We can see that  $\mathrm{CLAIR}_E$  outperforms all existing text-only measures (e.g., by  $5.18\%$  overall score over CIDEr), and in many cases, even outperforms measures that have access to the image at test time. Note that it is relatively weaker than image-augmented models in the HC setting; however, since both captions are correct, the model often cannot judge which is better purely the text. Models such as RefCLIP-S that have access to the image are naturally better discriminators in this case. We suspect that CLAIR's discriminative performance could be further improved by giving the LLM a choice between the two captions; however, we leave this optimization to future work.
+
+Groups of Captions: While CLAIR is capable of comparing a single candidate caption to a set of reference captions, it is also capable of comparing sets of candidate captions to sets of reference captions. This task is necessary when evaluating the ability of a model to generate captions that are diverse and that fully describe the conditional text distribution. We evaluate on the COCO-Sets dataset (Chan et al., 2022), 794 caption sets rated by AMT workers on two scales: how closely a candidate set matches the reference set in terms of both correctness and content coverage (See Appendix B for details). The results of this experiment are given in Table 4. We can see that CLAIR outperforms well when measuring the quality of a group of captions, and approaches the inter-human correlation on the (very) challenging task. CLAIR also outperforms TRM-Meteor and TRM-BLEU (Chan et al.,
+
+2022), suggesting that LLMs can judge both the content and diversity of the caption sets.
+
+# 4 Limitations
+
+While CLAIR correlates well with human judgments of caption quality, it has several limitations:
+
+Non-Determinism and Parsing Errors: Because CLAIR depends on the output of a language model, the measure can be non-deterministic and noisy. For instance, it may fail to elicit a judgment (e.g., "As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references"), or rarely, generate malformed JSON output. To address these issues, we perform multiple queries to the LLM, sometimes at higher temperatures if necessary. As a consequence, the measure may differ between runs, although we found the variance to be relatively insignificant (< 0.01 in many of the experiments). Additionally, since the language models used are not open-source, the models are subject to arbitrary change, replacement, or removal, which limits the efficacy of the measure as a long-term comparable measurement. We hope that increasing open access to language models with efforts such as Koala (Geng et al., 2023) and Vicuna (Chiang et al., 2023), will help to alleviate these challenges in the future.
+
+Increased Cost: CLAIR relies on language models which contain many billions of parameters. These language models have not only monetary cost but also human and environmental costs (Bender et al., 2021) which can reduce its utility as a target during training, such as for self-critical sequence training (Rennie et al., 2017). While API-based LLMs may be considered costly, even open-source LLMs have a cost (which can often be hard to quantify). CLAIR on the MS-COCO dataset uses an average of 226.148 tokens per sample (on OpenAI's API), representing a cost of  $0.0067 per sample (GPT-4), or$ 0.00033 per sample (GPT 3.5). For PALM, this drops to $0.000113 per sample. We hope that over time, advances in LLM inference (such as quantization and distillation), coupled with improvements in architecture will continue to yield lower-cost alternatives with strong performance on the caption evaluation task.
+
+Hallucination: While CLAIR does suffer from potential hallucination, we strongly believe that this weakness does not diminish the fact that CLAIR still correlates strongly with human judgment. In CLAIR, hallucinations in the score manifest as
+
+"incorrect" judgements of similarity, while hallucinations in the explanations manifest as poorly grounded explanations of the score/quality. Hallucinations in the score should be considered false negatives (blind spots instead of hallucinations). In the case of hallucinations in the explanations, such hallucinations may lead to misinterpretation, but arguably less misinterpretation than a black box method, and may even indicate misunderstandings in the model. Hallucination is a well-known challenge of current LLMs and is the subject of a great amount of research on instruction-tuning, RLHF, RLAIF, and other methods. As hallucination and instruction-following performance of the base models improves, CLAIR inherit similar improvements.
+
+Explainability: While CLAIR generates explanations for each rating, CLAIR has no strict scoring rubric. Much like human judgments, there is no direct way of attributing changes in score to changes in caption quality. For similar reasons, it is difficult to evaluate the quality of the generated explanations. Qualitatively, the explanations are generally reasonable and consider multiple axes of judgment.
+
+# 5 Conclusion
+
+This work introduces CLAIR, an LLM-based evaluation measure for image captioning. CLAIR's superior performance compared to highly-engineered measures indicates a remarkable fact: LLMs are well aligned with human judgments of caption quality, even more so than some measures designed specifically for semantic similarity. CLAIR is only a glimpse into how LLMs can be used for evaluation tasks, and image captioning is only the beginning. We hope that our work will inspire further exploration of similar measures in other vision and language domains, such as visual storytelling (Huang et al., 2016), where human evaluation of generated text remains a challenging task.
+
+# References
+
+Somak Aditya, Yezhou Yang, Chitta Baral, Cornelia Fermuller, and Yiannis Aloimonos. 2015. From images to sentences through scene description graphs using commonsense reasoning and knowledge. ArXiv preprint, abs/1511.03292.  
+Abhaya Agarwal and Alon Lavie. 2008. Meteor, M-BLEU and M-TER: Evaluation metrics for high-correlation with human rankings of machine translation output. In Proceedings of the Third Workshop on Statistical Machine Translation, pages 115-118. Association for Computational Linguistics.  
+Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. 2016. Spice: Semantic propositional image caption evaluation. In European conference on computer vision, pages 382-398. Springer.  
+Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. ArXiv preprint, abs/2204.05862.  
+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, pages 610-623.  
+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.  
+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, abs/2303.12712.  
+David M Chan, Yiming Ni, Austin Myers, Sudheendra Vijayanarasimhan, David A Ross, and John Canny. 2022. Distribution aware metrics for conditional natural language generation. ArXiv preprint, abs/2209.07518.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and
+
+Eric P. Xing. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. Visual programming for text-to-image generation and evaluation. arXiv preprint arXiv:2305.15328.  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. ArXiv preprint, abs/2204.02311.  
+Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2017. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 4299-4307.  
+Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. Qlora: Efficient finetuning of quantized llms. *ArXiv preprint*, abs/2305.14314.  
+Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. Koala: A dialogue model for academic research. Blog post.  
+Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528. Association for Computational Linguistics.  
+Micah Hodosh, Peter Young, and Julia Hockenmaier. 2013. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research, 47:853-899.  
+Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. 2023. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. arXiv preprint arXiv:2303.11897.  
+Ting-Hao Kenneth Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, C. Lawrence Zitnick, Devi Parikh, Lucy Vanderwende, Michel Galley, and Margaret Mitchell. 2016. Visual storytelling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1233-1239, San Diego, California. Association for Computational Linguistics.  
+Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig
+
+Schmidt. 2021. Openclip. If you use this software, please cite it as below.  
+Ming Jiang, Qiuyuan Huang, Lei Zhang, Xin Wang, Pengchuan Zhang, Zhe Gan, Jana Diesner, and Jianfeng Gao. 2019. TIGEr: Text-to-image grounding for image caption evaluation. 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 2141-2152. Association for Computational Linguistics.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. 2022. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 12888-12900. PMLR.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81. Association for Computational Linguistics.  
+Lucian Vlad Lita, Monica Rogati, and Alon Lavie. 2005. Blanc: Learning evaluation metrics for mt. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 740-747.  
+Junhua Mao, Wei Xu, Yi Yang, Jiang Wang, and Alan L Yuille. 2014. Explain images with multimodal recurrent neural networks. arXiv preprint arXiv:1410.1090.  
+OpenAI. 2022. Introducing chatgpt.  
+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. 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, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748-8763. PMLR.  
+Steven J. Rennie, Etienne Marcheret, Youssef Mroueh, Jerret Ross, and Vaibhava Goel. 2017. Self-critical sequence training for image captioning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 1179-1195. IEEE Computer Society.
+
+Anna Rohrbach, Lisa Anne Hendricks, Kaylee Burns, Trevor Darrell, and Kate Saenko. 2018. Object hallucination in image captioning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4035-4045. Association for Computational Linguistics.  
+Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. 2015. Cider: Consensus-based image description evaluation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2015, Boston, MA, USA, June 7-12, 2015, pages 4566-4575. IEEE Computer Society.  
+Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. OFA: unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 23318-23340. PMLR.  
+Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. MSR-VTT: A large video description dataset for bridging video and language. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 5288-5296. IEEE Computer Society.  
+Michal Yarom, Yonatan Bitton, Soravit Changpinyo, Roee Aharoni, Jonathan Herzig, Oran Lang, Eran Ofek, and Idan Szpektor. 2023. What you see is what you read? improving text-image alignment evaluation. arXiv preprint arXiv:2305.10400.  
+Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. 2014. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78.  
+Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. ArXiv preprint, abs/2306.05685.
+
+# Appendix
+
+# A Acknowledgements
+
+We thank Suhong Moon and Kate Saenko for their helpful comments on the work. Authors, as part of their affiliation with UC Berkeley, were supported in part by the NSF, DoD, and/or the Berkeley Artificial Intelligence Research (BAIR) industrial alliance program, as well as gifts from Anyscale, Astronomer, Google, IBM, Intel, Lacework, Microsoft, Mohamed Bin Zayed University of Artificial Intelligence, Samsung SDS, Uber, and VMware.
+
+# B Additional Experimental Details
+
+In this section, we provide several additional details for the experiments in section 3 run with the CLAIR measure.
+
+# B.1 Input Prompt Formatting
+
+The CLAIR prompt is given in its entirety in Figure 1. During run-time, candidate and reference captions are prefixed with a “-” and inserted into the prompt, one per line. The resulting query is passed to the large language model. In addition, for models which were not RLHF-tuned to perform conversation (such as PaLM), we found that it was helpful to append an additional prefix {"score": to the beginning of the output, to encourage the correct output formatting. CLAIR is surprisingly simple: it uses no in-context examples (is entirely zero-shot), and default inference parameters for the APIs. The model checkpoint metadata is generally unknown (as the APIs are somewhat fluid and evolving).
+
+# B.2 LLM Output Post-Processing
+
+Because CLAIR relies on an LLM to produce output, there is no guarantee that the output will be in the format we expect (i.e. valid, parsable JSON). To extract both the score and the reason, we first extract the first set of paired braces from the output of the LLM and attempt to parse the result as JSON. In most cases (99.997% for GPT-3, 99.991% for Claude, and 99.94% for PaLM during the course of our experiments), this is successful, and the score and reason are returned. In the case that the JSON output is malformed, we attempt to extract any sequence of digits from the LLM to use as a score, and set the reason to "Unknown." When this fails, as can be the case when the models produce an output such as "As an AI language model, I cannot see, and thus, cannot determine if the image captions match the references", we retry the prompt at a higher temperature ( $t = 1.0$ ) several times. Failing this (which occurred only three times in the entire evaluation of this paper, across several hundred thousand calls), we set the score to 0 for the caption.
+
+# B.3 Datasets
+
+In this section, we provide additional detail regarding the datasets used in the evaluations in section 3.
+
+COMPOSITE: The COMPOSITE dataset (Aditya et al., 2015) contains machine-generated test captions for 3995 images spread across the MS-COCO (Xu et al., 2016), Flickr8K (Mao et al., 2014) and Flickr30k (Young et al., 2014) datasets. Each image has three test captions, one written by a human, and two that are model generated. The candidate captions are graded by annotators on Amazon Mechanical Turk (AMT) on a scale of 1 (not relevant) to 5 (very relevant). Inter-human correlations are not available for this dataset.
+
+Flickr8K-Expert: The Flickr8K-Expert dataset (Hodosh et al., 2013) contains 5822 captions associated with 1000 images. The dataset is annotated with expert human judgments of quality, where images are rated from 1 (caption is unrelated to the image) to 4 (caption describes the image without errors). Unlike the composite and MS-COCO datasets, the captions here are selected using an image retrieval system, instead of generated using a learned image captioning model. Following Jiang et al. (2019), we exclude any candidate captions that overlap the reference set.
+
+MS-COCO: Following experiments by Rohrbach et al. (2018), we compute the sample-level correlation between our method and human ratings on a 500-image subset of the MS-COCO Karpathy test set. Each image in the subset contains candidate captions generated by 5 models, and each caption is labeled with the average three human ratings generated by AMT workers which range from 1 (very bad) to 5 (very good). Inter-human correlations are not available for this dataset.
+
+PASCAL-50S: PASCAL-50S contains 1000 images drawn from the PASCAL sentence dataset. Each image is associated with at least 50 (and as many as 120) reference captions. In addition to the reference captions, PASCAL-50S contains a set of 4000 human annotated image/caption pairs containing an image, and two candidate captions. The caption pairs fall into four groups:
+
+1. HC: In the HC group, both captions in the pair are human written, and describe the content of the target image correctly.  
+2. HI: In the HI group, both captions in the pair are human written, but one caption correctly describes the content of the image, and the
+
+other caption describes the content of a different image.
+
+3. HM: In the HM group, one caption is written by a human, and one caption is written by a machine, but both correctly describe the content of the image.  
+4. MM: In the MM group, both captions are written by a machine, and both correctly describe the image content.
+
+In PASCAL-50S, the task is to decide which caption in the pair humans prefer more (a subjective task, hopefully indicating caption quality). Following previous work (Jiang et al., 2019; Hessel et al., 2021), we limit the number of reference sentences to five during evaluation.
+
+COCO-Sets: The COCO-Sets dataset (Chan et al., 2022) is a set of samples that are designed to evaluate the correlation of distribution-aware image captioning measures with human judgments of distributional distance. In this dataset, humans were presented with two candidate caption sets (two image captioning models, OFA (Wang et al., 2022) and BLIP (Li et al., 2022) using different temperatures), and asked which candidate caption set correlated better with a reference caption set on two measures: how much they overlapped factually (correctness), and how much information they provided about the references (coverage). It consists of 794 AMT worker-generated judgments of caption quality for images in the MS-COCO dataset.

2023/CLAIR_ Evaluating Image Captions with Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01c5a45620da40442658942bf6e1765d516683a0433989048f9b8dee1e3382cc
+size 258557

2023/CLEME_ Debiasing Multi-reference Evaluation for Grammatical Error Correction/ef2f484a-30da-421b-8e1e-02febc3b3494_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3766271fa5e1e4555b4dffeb52ad346169e9d888566bc6a30c32a62919e03b43
+size 817628

2023/CLEME_ Debiasing Multi-reference Evaluation for Grammatical Error Correction/full.md

@@ -0,0 +1,386 @@
+# CLEME: Debiasing Multi-reference Evaluation for Grammatical Error Correction
+
+Jingheng Ye $^{1*}$ , Yinghui Li $^{1*}$ , Qingyu Zhou $^{2}$ , Yangning Li $^{1,3}$ , Shirong Ma $^{1}$ , Hai-Tao Zheng $^{1,3\dagger}$ , Ying Shen $^{4}$
+
+$^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University
+
+$^{2}$ OPPO Research Institute,  $^{3}$ Peng Cheng Laboratory
+
+$^{4}$ School of Intelligent Systems Engineering, Sun-Yat Sen University
+
+{yejh22, liyinghu20}@mails.tsinghua.edu.cn
+
+# Abstract
+
+Evaluating the performance of Grammatical Error Correction (GEC) systems is a challenging task due to its subjectivity. Designing an evaluation metric that is as objective as possible is crucial to the development of GEC task. However, mainstream evaluation metrics, i.e., reference-based metrics, introduce bias into the multi-reference evaluation by extracting edits without considering the presence of multiple references. To overcome this issue, we propose ChunkLEvel Multi-reference Evaluation (CLEME), designed to evaluate GEC systems in the multi-reference evaluation setting. CLEME builds chunk sequences with consistent boundaries for the source, the hypothesis and references, thus eliminating the bias caused by inconsistent edit boundaries. Furthermore, we observe the consistent boundary could also act as the boundary of grammatical errors, based on which the  $\mathrm{F}_{0.5}$  score is then computed following the correction independence assumption. We conduct experiments on six English reference sets based on the CoNLL-2014 shared task. Extensive experiments and detailed analyses demonstrate the correctness of our discovery and the effectiveness of CLEME. Further analysis reveals that CLEME is robust to evaluate GEC systems across reference sets with varying numbers of references and annotation styles<sup>1</sup>.
+
+# 1 Introduction
+
+Grammatical Error Correction (GEC) is a task that involves making local substitutions to correct grammatical errors in a given ungrammatical text (Bryant et al., 2022; Ma et al., 2022; Ye et al., 2022; Ma et al., 2023). The practical value of GEC in daily life has led to increasing attention being paid to this task (Li et al., 2021, 2022a,b; Kaneko
+
+![](images/5689e0eff71ef9215b3c5fa42debb751b0500f725f869b75943bc593b50e68c4.jpg)  
+Figure 1: A comparison of edits automatically extracted by ERRANT and CLEME. An orange block is an edit.
+
+et al., 2022; Li et al., 2023; Ye et al., 2023a,b; Zhang et al., 2023). However, it is intractable to evaluate GEC systems due to the highly subjective nature of the task and the low inter-annotator agreement (IAA) (Choshen and Abend, 2018). Therefore, most datasets improve compatibility by incorporating multiple references to guarantee a more realistic evaluation of the model performance.
+
+There are two broad categories of GEC metrics: reference-based and reference-less. Reference-based metrics evaluate GEC systems by comparing their hypotheses and human-annotated references in terms of edits (Dahlmeier and Ng, 2012; Bryant et al., 2017) or n-grams (Napoles et al., 2015). Reference-less metrics are proposed to evaluate GEC systems without references. However, Deutsch et al. (2022) demonstrate that reference-less metrics are inherently biased and limited in their ability to evaluate generated text. Therefore, we focus on reference-based metrics, which can evaluate in an interpretable manner, thus providing useful insights for model analysis.
+
+Figure 1 illustrates how existing reference-based metrics, such as ERRANT, extract the edit and then compute the  $\mathrm{F}_{0.5}$  score by comparing hypotheses and references. However, these metrics often fail to consider multiple references, which can result in bias during multi-reference evaluation. We argue that this bias arises because the current approach rewards equally good corrections unfairly. For in-
+
+stance, the ungrammatical phrase the technologies were is equally well-corrected by both Ref. 1 and Ref. 2. However, if a hypothesis aligns with Ref. 1's corrections (i.e.,  $[the\to \epsilon ]$  and  $[were\to have]$ $\mathrm{TP} = 2)$  , it will be rewarded less than the corrections of Ref. 2 (i.e.,  $[the\to \epsilon ]$  , [technologies  $\rightarrow$  technology] and  $[were\to has]$ $\mathrm{TP} = 3)$
+
+In this paper, we propose Chunk-LEvel Multi-reference Evaluation (CLEME), which enables unbiased  $\mathrm{F}_{0.5}$  scores for GEC multi-reference evaluation. Inspired by (Gotou et al., 2020), CLEME transforms the source, the hypothesis and all the references into chunk sequences with consistent boundaries, thereby eliminating the bias in GEC multi-reference evaluation.
+
+Existing metrics assume that corrections of grammatical errors are dependent. That is, whenever there is more than one reference for a source, the metrics try each reference in turn, and then the highest score is taken as the final score. However, we observe that grammatical errors corrections in terms of chunks can be considered approximately independent. For example, the ungrammatical phrases the technologies were and for shown in Figure 1 can be corrected independently, i.e., the correction of the technologies were has no bearing on the correction of for. Based on this observation, we compute  $\mathrm{F}_{0.5}$  scores following the assumption that corrections of grammatical errors are independent. Specifically, we iterate through the chunks of a hypothesis and consider a chunk correct if it matches any of the corresponding chunks in the references. In this case, the hypothesis in Figure 1 would be rewarded 2TP, rather than 1TP and 1FP, which is the traditional case. To demonstrate the effectiveness and robustness of CLEME, we conduct experiments on six English reference sets with varying numbers of references and annotation styles, either calculating the  $\mathrm{F}_{0.5}$  score at the corpus- or sentence-level.
+
+In summary, our contributions are three folds:
+
+(1) We propose CLEME, a reference-based metric that evaluates GEC systems at the chunk-level, aiming to provide unbiased  $\mathrm{F}_{0.5}$  scores for GEC multi-reference evaluation.  
+(2) We observe that the corrections of grammatical errors in terms of chunks are approximately independent. Therefore, we propose to compute  $\mathrm{F}_{0.5}$  scores based on the correction independence assumption.
+
+(3) Extensive experiments and human evaluation are conducted to confirm the effectiveness and robustness of our approach.
+
+# 2 Preliminary Study
+
+# 2.1 Consistent Boundaries
+
+We determine consistent chunk-level boundaries by chunk partition process to debias the multi-reference evaluation, as depicted in Figure 2. We first extract the edit sets of the hypothesis and references, and then merge the overlapping edits into a chunk. It's worth noting that the source, hypothesis and references are all segmented into chunk sequences with the same number of chunks, regardless of the number of their tokens. This process is straightforward since we can locate and examine all possible corrections of an erroneous chunk. For example, the chunk by the can be corrected in two ways, i.e., with in Ref. 1 and through in Ref. 2. The resulting chunks fall into three categories: 1) unchanged chunks, which contain the same text segments as the source sentence, 2) corrected chunks, which consist of non-empty text segments different from the source sentence, and 3) dummy chunks are empty chunks.
+
+# 2.2 Boundaries of Grammatical Errors
+
+Figure 2 illustrates the merging of overlapping edits into either corrected or dummy chunks, which are then separated by unchanged chunks. This raises the question, are chunk boundaries the boundaries of grammatical errors?
+
+Dataset. To answer the question, we conduct experiments on BN-10GEC (Bryant and Ng, 2015). The dataset comprises 1,312 source sentences that are identical to the CoNLL-2014 test data (Ng et al., 2014). Each source sentence is associated with 10 references annotated by 10 native English speakers, including two official annotators of CoNLL-2014, the first author of the paper, and seven freelancers recruited via an online recruitment website.
+
+Experiment Setup. For each source sentence, we sample 9 references and run the chunk partition process described in Section 2.1. The resulting chunk sequences are determined collectively by all 9 references. The edits of the remaining reference  $\{e_1,\dots ,e_M\}$  are then used to calculate the following three statistics: 1) The In-Corrected-Chunk (ICC) ratio indicates the proportion of edits included by corrected/dummy chunks of the other
+
+references. An edit is included by a chunk if the interval of the edit falls within that of the chunk. 2) The In-Unchanged-Chunk (IUC) ratio gives the proportion of edits included by unchanged chunks of the other references. 3) The Cross-Chunk (CC) ratio computes the proportion of edits that extend beyond the original boundaries. These statistics are calculated as follows:
+
+$$
+\mathrm {I C C} = \frac {1}{M} \sum_ {i = 1} ^ {M} f _ {1} \left(e _ {i}\right), \tag {1}
+$$
+
+$$
+I U C = \frac {1}{M} \sum_ {i = 1} ^ {M} f _ {2} \left(e _ {i}\right), \tag {2}
+$$
+
+$$
+\mathrm {C C} = 1 - \mathrm {I C C} - \mathrm {I U C}, \tag {3}
+$$
+
+where  $M$  is the number of edits from the remaining reference. If the edit  $e_i$  is included in a corrected/dummy chunk, the function  $f_1(e_i)$  returns 1, otherwise 0. Likewise, if the edit  $e_i$  is included in an unchanged chunk, the function  $f_2(e_i)$  returns 1, otherwise 0. We sample 9 different references for chunk partition in each run and repeatedly calculate the statistics using the remaining reference.
+
+Results. As shown in Table 1, the number of corrected and dummy chunks are less than that of edits since overlapping edits are merged into a chunk. A total of  $90.66\%$  edits are included by the corrected/dummy chunks, which suggests the grammatical errors to be corrected have been considered by the other references. However, only  $7.74\%$  edits are included by corrected chunks, indicating that these edits may be over-corrected since the other references believe no grammatical errors needed correction. Interestingly,  $1.61\%$  edits cross the chunk boundaries, suggesting that the chunk boundaries are stable enough to serve as the boundaries of grammatical errors to some extent. Additionally, human evaluation in Section 4.2 could be used as another argument to support this conclusion. Therefore, we have the following assumption.
+
+![](images/f3ce436ff7a48511edf6ddf51662388e04893da137763fe434cc93eeb4c6f883.jpg)
+
+Correction independence assumption: grammatical error corrections are independent.
+
+That is, the correction of a grammatical error does not impact the correction of other grammatical errors. With this assumption,  $\mathrm{F}_{0.5}$  scores can be calculated using an alternate method, which will be introduced in Section 3.
+
+<table><tr><td>Item</td><td>Number (perc.)</td><td>Length</td></tr><tr><td>Sentences</td><td>1,312</td><td>23.0</td></tr><tr><td>References</td><td>13,120</td><td>22.9</td></tr><tr><td>Edits</td><td>36,677</td><td>1.0</td></tr><tr><td>UnchangedChunks</td><td>93,469 (77.63%)</td><td>2.5</td></tr><tr><td>Corrected/DummyChunks</td><td>26,948 (22.37%)</td><td>2.4</td></tr><tr><td>ICC</td><td>33,251 (90.66%)</td><td>-</td></tr><tr><td>IUC</td><td>2,837 (7.74%)</td><td>-</td></tr><tr><td>CC</td><td>589 (1.61%)</td><td>-</td></tr></table>
+
+Table 1: Statistics of the BN-10GEC dataset.
+
+# 3 Method
+
+# 3.1 Chunk Evaluation
+
+As shown in Figure 2, each chunk consists of edit operation(s), start index, end index, and correct tokens. Conventional reference-based metrics such as MaxMatch  $(\mathbf{M}^2)$  and ERRANT compute  $\mathrm{F_{0.5}}$  scores based on the correction dependence assumption. They evaluate the performance for each reference separately and select the one that yields the best result for the source sentence. CLEME-dependent also computes  $\mathrm{F_{0.5}}$  scores in this way by treating corrected/dummy chunks as edits. On the other hand, CLEME-independent is proposed to compute  $\mathrm{F_{0.5}}$  scores based on the correction independence assumption. A corrected/dummy chunk from the hypothesis is considered correct if it matches one of the corresponding chunks from the references. It is worth noting that CLEME is able to fully inherit pre-classified errors from ERRANT, where each corrected/dummy chunk may consist of multiple error with different types.
+
+# 3.2 Length Weighting
+
+The average length of chunks is much longer than that of edits shown in Table 1, resulting in the unfairness of chunk evaluation if a longer chunk is rewarded equally with a shorter one. Therefore, we add length weighting to the chunk evaluation. The intuition of length weighting is to compensate for long chunk matching. The weights of True Positives (TPs), False Positives (FPs), and False Negatives (FNs) are computed as follows:2
+
+$$
+w ^ {\mathrm {T P}} = \operatorname {c l i p} \left(\frac {\alpha_ {1}}{1 + (\alpha_ {1} - 1) \exp (\ell - x)}, c _ {\min }, c _ {\max }\right), \tag {4}
+$$
+
+$$
+w ^ {\mathrm {F P}} = \operatorname {c l i p} \left(\frac {\alpha_ {2}}{1 + \left(\alpha_ {2} - 1\right) \exp (x - \ell)}, c _ {\min }, c _ {\max }\right), \tag {5}
+$$
+
+![](images/474cbfaf536fc30bfa38e4aa97436adb8bb9f16090cd8b901d55aed4b8f2e19b.jpg)
+
+![](images/cd8e14596165fc74d8feb075e08f14583ebe75ae8703181490f22120a2a8ede4.jpg)  
+Figure 2: Overview of our approach CLEME. CLEME first 1) extracts edits of the hypothesis and the references, 2) merges the overlapping edits into chunks, and then 3) computes the  $\mathrm{F}_{0.5}$  scores based on two different assumptions.
+
+![](images/9c7d1f803588e0d6d7c56e27685f87c0a16ce0ac268bacd9f4710501c3affbcd.jpg)
+
+$$
+w ^ {\mathrm {F N}} = \operatorname {c l i p} \left(\frac {\alpha_ {3}}{1 + \left(\alpha_ {3} - 1\right) \exp (\ell - x)}, c _ {\min }, c _ {\max }\right), \tag {6}
+$$
+
+where  $\alpha_{1},\alpha_{2}$  and  $\alpha_{3}$  are scale factors for TPs, FPs and FNs respectively,  $x$  is the length of the chunk,  $\ell$  is the average length of chunks, and the function  $\mathrm{clip}(v,a,b)$  clips the value  $v$  between  $a$  and  $b$ . The curves of length weighting are depicted in Figure 3. Formally, given a system corrected/dummy chunk set  $C^H$  and a gold corrected/dummy chunk set  $C^R$ , we apply length weighting on each chunk to compute precision, recall and  $\mathrm{F}_{0.5}$  as follows:
+
+$$
+P = \frac {\sum_ {c \in C ^ {H} \cap C ^ {R}} w _ {c} ^ {\mathrm {T P}}}{\sum_ {c \in C ^ {H} \cap C ^ {R}} w _ {c} ^ {\mathrm {T P}} + \sum_ {c \in C ^ {H} \backslash C ^ {R}} w _ {c} ^ {\mathrm {F P}}}, \tag {7}
+$$
+
+$$
+R = \frac {\sum_ {c \in C ^ {H} \cap C ^ {R}} w _ {c} ^ {\mathrm {T P}}}{\sum_ {c \in C ^ {H} \cap C ^ {R}} w _ {c} ^ {\mathrm {T P}} + \sum_ {c \in C ^ {R} \backslash C ^ {H}} w _ {c} ^ {\mathrm {F N}}}, \tag {8}
+$$
+
+$$
+F _ {\beta} = \left(1 + \beta^ {2}\right) \cdot \frac {P \cdot R}{\left(\beta^ {2} \cdot P\right) + R}, \tag {9}
+$$
+
+where  $\beta = 0.5$  is usually used, which weighs precision twice as much as recall.
+
+![](images/ee833dc676f1c0f65a850b45bbd4541065c83bead421035ecda25219e26e96d5.jpg)  
+Figure 3: Curves of length weighting with different  $\alpha$  for  $\ell = 2$ . All the curves pass through the point  $(\ell, 1.0)$ . A curve with a larger scale factor has a greater slope.
+
+# 3.3 Corpus-level v.s. Sentence-level
+
+We compute  $\mathrm{F}_{0.5}$  scores of GEC systems at both corpus-level and sentence-level following (Gong et al., 2022). Corpus-level metrics compute an  $\mathrm{F}_{0.5}$  score over the entire dataset. Sentence-level metrics compute an  $\mathrm{F}_{0.5}$  score over each sentence of the dataset and evaluate GEC systems by us
+
+ing the average  $\mathrm{F}_{0.5}$  score. CLEME-dependent and CLEME-independent are corpus-level metrics, and their sentence-level variants are respectively SentCLEME-dependent and SentCLEME-independent. Both levels of the GEC metric are developed to provide more user-friendly options. Sentence-level metrics should be used if consistent evaluation weight for each sample is desired. This ensures that the evaluation result of each sample has the same influence on the final score. On the other hand, if harder samples containing more edits should have larger weight, then corpus-level metrics should be used instead.
+
+# 4 Experiments
+
+# 4.1 Correlations with Human Judgments
+
+Dataset. To verify the effectiveness of CLEME, we measure correlations between reference-based metrics and human judgments on multiple English reference sets, including CoNLL-2014 (Grundkiewicz et al., 2015), BN-10GEC (Bryant and Ng, 2015) and SN-8GEC (Sakaguchi et al., 2016). All the reference sets are based on CoNLL-2014 (Ng et al., 2014), consisting of 1,312 source sentences. SN-8GEC collected 8 references sets of annotations from both experts and non-experts, including 4 sets of minimal edits and 4 sets of fluency edits (2 by experts and 2 by non-experts). Reference sets statistics are reported in Appendix A.
+
+The human judgments for the outputs of 13 GEC systems (including the unchanged source text) are presented by (Grundkiewicz et al., 2015), where eight native speakers were asked to rank the output of all the systems from best to worst. Two system ranking lists are generated using Expected Wins (EW) (Macháček and Bojar, 2013) and TrueSkill (TS) (Sakaguchi et al., 2014) respectively.
+
+Experiment Settings. Following (Gong et al., 2022; Chollampatt and Ng, 2018), we compute the Pearson  $\gamma$  and Spearman correlation coefficient  $\rho$  between reference-based metrics and human judgments based on corpus-level ranking. We tune the hyperparameters on CoNLL-2014 and keep the hyperparameters on the other reference sets, in order to demonstrate the adaptability of our approach. The detailed hyperparameters of CLEME are reported in Appendix B.
+
+Evaluation Metrics. We compare our approach with the following reference-based metrics, includ
+
+ing corpus- and sentence-level variants 3:
+
+- GLEU and SentGLEU (Napoles et al., 2015) are n-gram based metrics, which reward hypothesis n-grams that overlap with the reference but not the source and penalize hypothesis n-grams that overlap with the source but not the reference.  
+-  $\mathbf{M}^2$  and SentM $^2$  (Dahlmeier and Ng, 2012) dynamically extract the hypothesis edits with the maximum overlap of gold annotations.  
+- ERRANAT and SentERRANT (Bryant et al., 2017) extract edits by utilizing a linguistically-enhanced alignment algorithm.  
+- PT-M $^2$  and SentPT-M $^2$  (Gong et al., 2022) are recently proposed reference and PLM-based GEC metric, which score edits using the knowledge of pre-trained language model.
+
+Additionally, CLEME can evaluate GEC systems by accuracy scores, which is usually not implemented by conventional reference-based metrics. Please refer to Appendix C for the introduction and analyses of evaluating GEC systems by accuracy.
+
+Results. Table 2 reports the correlations between reference-based metrics and human judgments. For the corpus-level metrics, GLEU achieves the highest correlations on BN-10GEC and NE-fluency reference sets. However, GLEU suffers from negative correlations on NE-Minimal, which is caused by low-quality annotations  $^{4}$  of NE-Minimal, indicating that GLEU may not be a robust metric, consistent with the findings of (Sakaguchi et al., 2016). ERRANT performs slightly better than  $\mathbf{M}^2$  on most reference sets, while PT- $\mathbf{M}^2$  is a strong corpus-level metric, which achieves the highest or comparable correlations on all reference sets at the cost of more than  $10\times$  running time than other reference-based metrics. Our proposed CLEME-dependent and CLEME-independent make better use of consistent chunk boundaries, thus performing slightly better than  $\mathbf{M}^2$  and ERRANT on most reference sets. Notably, CLEME-independent achieves comparable performance to CLEME-dependent, showing the
+
+<table><tr><td rowspan="2" colspan="2">Metric</td><td colspan="2">CoNLL-2014</td><td colspan="2">BN-10GEC</td><td colspan="2">E-Minimal</td><td colspan="2">E-Fluency</td><td colspan="2">NE-Minimal</td><td colspan="2">NE-Fluency</td></tr><tr><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td></tr><tr><td rowspan="2">M2</td><td>γ</td><td>0.623</td><td>0.672</td><td>0.547</td><td>0.610</td><td>0.597</td><td>0.650</td><td>0.590</td><td>0.659</td><td>0.575</td><td>0.634</td><td>0.582</td><td>0.649</td></tr><tr><td>ρ</td><td>0.687</td><td>0.720</td><td>0.648</td><td>0.692</td><td>0.654</td><td>0.703</td><td>0.654</td><td>0.709</td><td>0.577</td><td>0.648</td><td>0.648</td><td>0.703</td></tr><tr><td rowspan="2">GLEU</td><td>γ</td><td>0.701</td><td>0.750</td><td>0.678</td><td>0.761</td><td>0.533</td><td>0.513</td><td>0.693</td><td>0.771</td><td>-0.044</td><td>-0.113</td><td>0.674</td><td>0.767</td></tr><tr><td>ρ</td><td>0.467</td><td>0.555</td><td>0.754</td><td>0.806</td><td>0.577</td><td>0.511</td><td>0.710</td><td>0.757</td><td>-0.005</td><td>-0.055</td><td>0.725</td><td>0.819</td></tr><tr><td rowspan="2">ERRANT</td><td>γ</td><td>0.642</td><td>0.688</td><td>0.586</td><td>0.644</td><td>0.578</td><td>0.631</td><td>0.594</td><td>0.663</td><td>0.585</td><td>0.637</td><td>0.597</td><td>0.659</td></tr><tr><td>ρ</td><td>0.659</td><td>0.698</td><td>0.637</td><td>0.698</td><td>0.742</td><td>0.786</td><td>0.720</td><td>0.775</td><td>0.747</td><td>0.797</td><td>0.753</td><td>0.797</td></tr><tr><td rowspan="2">PT-M2</td><td>γ</td><td>0.693</td><td>0.737</td><td>0.650</td><td>0.706</td><td>0.626</td><td>0.667</td><td>0.621</td><td>0.681</td><td>0.630</td><td>0.675</td><td>0.620</td><td>0.682</td></tr><tr><td>ρ</td><td>0.758</td><td>0.769</td><td>0.690</td><td>0.824</td><td>0.709</td><td>0.736</td><td>0.758</td><td>0.802</td><td>0.736</td><td>0.758</td><td>0.758</td><td>0.802</td></tr><tr><td rowspan="2">CLEME-dependent (Ours)</td><td>γ</td><td>0.648</td><td>0.691</td><td>0.602</td><td>0.656</td><td>0.594</td><td>0.644</td><td>0.589</td><td>0.654</td><td>0.595</td><td>0.643</td><td>0.612</td><td>0.673</td></tr><tr><td>ρ</td><td>0.709</td><td>0.742</td><td>0.692</td><td>0.747</td><td>0.797</td><td>0.813</td><td>0.714</td><td>0.775</td><td>0.786</td><td>0.835</td><td>0.720</td><td>0.791</td></tr><tr><td rowspan="2">CLEME-independent (Ours)</td><td>γ</td><td>0.649</td><td>0.691</td><td>0.609</td><td>0.659</td><td>0.593</td><td>0.643</td><td>0.587</td><td>0.653</td><td>0.601</td><td>0.647</td><td>0.611</td><td>0.672</td></tr><tr><td>ρ</td><td>0.709</td><td>0.731</td><td>0.692</td><td>0.747</td><td>0.791</td><td>0.802</td><td>0.731</td><td>0.791</td><td>0.797</td><td>0.841</td><td>0.714</td><td>0.786</td></tr><tr><td rowspan="2">SentM2</td><td>γ</td><td>0.871</td><td>0.864</td><td>0.567</td><td>0.646</td><td>0.805♣</td><td>0.836♣</td><td>0.655</td><td>0.732</td><td>0.729♣</td><td>0.785♣</td><td>0.621</td><td>0.699</td></tr><tr><td>ρ</td><td>0.731</td><td>0.758</td><td>0.593</td><td>0.648</td><td>0.806♣</td><td>0.845♣</td><td>0.731</td><td>0.764</td><td>0.797♣</td><td>0.846♣</td><td>0.632</td><td>0.687</td></tr><tr><td rowspan="2">SentGLEU</td><td>γ</td><td>0.784</td><td>0.828</td><td>0.756</td><td>0.826</td><td>0.742♣</td><td>0.773♣</td><td>0.785</td><td>0.846</td><td>0.723♣</td><td>0.762♣</td><td>0.778</td><td>0.848</td></tr><tr><td>ρ</td><td>0.720</td><td>0.775</td><td>0.769</td><td>0.824</td><td>0.764♣</td><td>0.797♣</td><td>0.791</td><td>0.846</td><td>0.764♣</td><td>0.830♣</td><td>0.768</td><td>0.846</td></tr><tr><td rowspan="2">SentERRANT</td><td>γ</td><td>0.870</td><td>0.846</td><td>0.885</td><td>0.896</td><td>0.768♣</td><td>0.803♣</td><td>0.806</td><td>0.732</td><td>0.710♣</td><td>0.765♣</td><td>0.793</td><td>0.847</td></tr><tr><td>ρ</td><td>0.742</td><td>0.747</td><td>0.786</td><td>0.830</td><td>0.775♣</td><td>0.819♣</td><td>0.813</td><td>0.764</td><td>0.780♣</td><td>0.841♣</td><td>0.830</td><td>0.857</td></tr><tr><td rowspan="2">SentPT-M2</td><td>γ</td><td>0.949</td><td>0.938</td><td>0.602♣</td><td>0.682♣</td><td>0.831♣</td><td>0.855♣</td><td>0.689</td><td>0.763</td><td>0.770♣</td><td>0.822♣</td><td>0.648</td><td>0.725</td></tr><tr><td>ρ</td><td>0.907</td><td>0.874</td><td>0.626♣</td><td>0.670♣</td><td>0.808♣</td><td>0.819♣</td><td>0.797</td><td>0.841</td><td>0.813♣</td><td>0.857♣</td><td>0.742</td><td>0.786</td></tr><tr><td rowspan="2">SentCLEME-dependent (Ours)</td><td>γ</td><td>0.876</td><td>0.844</td><td>0.915</td><td>0.913</td><td>0.806♣</td><td>0.838♣</td><td>0.849</td><td>0.886</td><td>0.742♣</td><td>0.795♣</td><td>0.876</td><td>0.921</td></tr><tr><td>ρ</td><td>0.824</td><td>0.808</td><td>0.835</td><td>0.874</td><td>0.775♣</td><td>0.819♣</td><td>0.824</td><td>0.863</td><td>0.797♣</td><td>0.846♣</td><td>0.791</td><td>0.846</td></tr><tr><td rowspan="2">SentCLEME-independent (Ours)</td><td>γ</td><td>0.868</td><td>0.857</td><td>0.855♣</td><td>0.876♣</td><td>0.821♣</td><td>0.856♣</td><td>0.841</td><td>0.877</td><td>0.782♣</td><td>0.831♣</td><td>0.852</td><td>0.896</td></tr><tr><td>ρ</td><td>0.725</td><td>0.758</td><td>0.659♣</td><td>0.714♣</td><td>0.775♣</td><td>0.819♣</td><td>0.808</td><td>0.846</td><td>0.819♣</td><td>0.874♣</td><td>0.762</td><td>0.825</td></tr></table>
+
+Table 2: Overview of correlations between mainstream GEC metrics and human judgments. We highlight the highest score in bold and the second-highest score with underlines. SN-8GEC consists of four reference sets, i.e., E-Minimal, E-Fluency, NE-Minimal and NE-Fluency.  $\clubsuit$  We remove unchanged reference sentences for higher correlations due to low-quality annotations. Otherwise, negative correlations are possible.
+
+effectiveness of computing  $\mathrm{F}_{0.5}$  scores based on the correction independence assumption.
+
+The majority of the sentence-level metrics outperform their corpus-level counterparts because they weigh samples equally, which is in line with the bias of human annotation. Despite the strong performance of PT-M², SentPT-M² achieves lower correlations on BN-10GEC, E-Fluency and NE-Fluency compared to other sentence-level metrics. It suggests that scoring edits using pre-trained language models may not generalize well to unseen reference sets for sentence-level metrics. Our approach aligns better with human judgments than existing reference-based metrics for most reference sets. Specifically, SentCLEME-dependent performs best on BN-10GEC and NE-Fluency, and performs on a par with the best metric on E-Fluency, indicating it is more suitable for fluent reference sets. This phenomenon aligns with our intuition since fluent editing is more likely to follow the correction dependence assumption. In contrast, SentCLEME-independent achieves higher correlations on E-Minimal and NE-Minimal, as we would expect from minimal editing that is more likely to follow the correction independence assumption. These results suggest that reference sets may have a preference towards one of the correction assump
+
+tions. Additionally, our approach achieves higher correlations on (N)E-Fluency rather than (N)E-Minimal, while SentM² and SentERRANT perform worse on E-Fluency than E-Minimal. This is because CLEME evaluates GEC systems using longer chunks rather than crappy edits, which could better reflect whether a grammatical error is fluently corrected. Overall, our approach achieves higher or comparable correlations on sentence-level than existing reference-based methods.
+
+# 4.2 Human Evaluation
+
+Experiments have shown the effectiveness of evaluating GEC systems based on the correction independence assumption. In this section, we aim to demonstrate whether the correction independence assumption makes sense for humans. We define the correction independence of a pair of chunks as the irrelevance of the correction of one chunk to the correction of the other. A simple case is presented in Appendix E. To evaluate this assumption, we conduct human evaluation experiments on 1,000 sentences randomly sampled from BN-10GEC (Bryant and Ng, 2015). Three annotators were asked to judge whether a pair of chunks is correction-independent.
+
+Table 3 reports the ratio of correction independen
+
+![](images/465589ceb2d504285c3a2f0516ac63050ef8a0bc3926c488f7fdaf2730303f4f.jpg)  
+Figure 4: Effect of references number on BN-10GEC.
+
+<table><tr><td>Annotator</td><td>Ratio of Correction Independence</td></tr><tr><td>A1</td><td>90.85%</td></tr><tr><td>A2</td><td>93.55%</td></tr><tr><td>A3</td><td>91.14%</td></tr><tr><td>Annotator</td><td>Cohen&#x27;s-κ</td></tr><tr><td>A1 v.s. A2</td><td>38.66%</td></tr><tr><td>A1 v.s. A3</td><td>43.10%</td></tr><tr><td>A2 v.s. A3</td><td>39.34%</td></tr></table>
+
+dence and Cohen's- $\kappa$  (Cohen, 1960) inter-annotator agreement (IAA) across the three annotators. Results show that more than  $90\%$  pairs of chunks are correction-independent for all the annotators, indicating that it is reasonable to evaluate GEC systems based on the correction independence assumption. Moreover, considering the subjectivity of GEC task, the IAA statistics show that it is relatively easy to judge whether a pair of chunks is correction-independent, compared with the previous study (Bryant and Ng, 2015) $^{5}$ .
+
+# 5 Analysis
+
+# 5.1 False Negative
+
+We observe that the number of false negatives (FNs) identified by CLEME is significantly lower than that of ERRANT. This difference can be attributed to the distinct definitions used by each system. While ERRANT considers FNs as edits in the reference that do not match those made in the hypothesis, CLEME identifies FNs as corrected/dummy chunks in the reference that do not match the chunks in the hypothesis. We argue the definition of ERRANT is problematic, as it tends to
+
+![](images/cc37ba916f3f583794ee51eea16ace214b4a397d562a23274aae1ce859abae40.jpg)  
+Figure 5: Effect of scale factors on BN-10GEC.
+
+Table 3: A comparison of correction independence annotations across three annotators.  
+
+<table><tr><td colspan="2">Text</td><td>FP</td><td>FN</td></tr><tr><td rowspan="2">Source 
+Hyp.</td><td>It has improved compared for the last century.</td><td></td><td></td></tr><tr><td>It has improved compared between the last century.</td><td></td><td></td></tr><tr><td colspan="2">ERRANT</td><td></td><td></td></tr><tr><td>Ref. 1</td><td>It has improved compared to the century.</td><td>1</td><td>1</td></tr><tr><td>Ref. 2</td><td>It has improved compared with the century.</td><td>1</td><td>1</td></tr><tr><td colspan="2">CLEME</td><td></td><td></td></tr><tr><td>Ref. 1</td><td>It has improved compared to the century.</td><td>1</td><td>0</td></tr><tr><td>Ref. 2</td><td>It has improved compared with the century.</td><td>1</td><td>0</td></tr></table>
+
+Table 4: Cases of ERRANT and CLEME. ERRANT gives FP=1 and FN=1 since the hypothesis does not match one of the edits of references. CLEME gives only FP=1 since the hypothesis tries to correct the error.
+
+overestimate FN counts in grammatical error correction (GEC) systems, which is evident from the examples presented in Table 4. On the other hand, CLEME's definition also includes true negatives (TNs), making it possible to calculate accuracy.
+
+# 5.2 Ablation Study
+
+We present ablation analyses of our approaches on BN-10GEC - we have similar findings on other reference sets. We report Pearson correlations  $\gamma$  using Expected Wins ranking. The trend is similar for Spearson correlations and TrueSkill ranking.
+
+Number of References. Since CLEME is designed for multi-reference evaluation, it degrades to conventional reference-based metrics such as  $\mathbf{M}^2$  and ERRANT when only one reference is available. Here we demonstrate how correlations change against an increasing number of available references. The results reported in Figure 4 indicate that the correlations of corpus-level metrics do not change significantly with the increasing number of available references. However, except for SentGLEU, correlations of sentence-level metric are consistently higher than corpus-level metrics, and steadily increase with more references. Therefore, we recommend evaluating GEC systems using sentence-level metrics rather than corpus-level
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td><td>Chunk 5</td><td>Chunk 6</td></tr><tr><td>Source</td><td>On the other hand, if there are</td><td>ways can</td><td>help us to control</td><td>or</td><td>cure the disease, we can</td><td>going.</td></tr><tr><td>Hyp.</td><td>On the other hand, if there are</td><td>ways that can</td><td>help us to control</td><td>and</td><td>cure the disease, we can</td><td>go.</td></tr><tr><td>Ref. 1</td><td>On the other hand, if there are</td><td>ways that can</td><td>help us to control</td><td>or</td><td>cure the disease, we can</td><td>go.</td></tr><tr><td>Ref. 2</td><td>On the other hand, if there are</td><td>things that can</td><td>help us to control</td><td>and</td><td>cure the disease, we can</td><td>go.</td></tr></table>
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td><td>Chunk 5</td></tr><tr><td>Source</td><td>On</td><td></td><td>one hand, we do not want this potential danger</td><td>causing frightening affects in</td><td>our lives .</td></tr><tr><td>Hyp.</td><td>On</td><td></td><td>one hand, we do not want this potential danger</td><td>causing frightening affects in</td><td>our lives .</td></tr><tr><td>Ref. 1</td><td>On</td><td></td><td>one hand, we do not want this potential danger</td><td>having frightening effects in</td><td>our lives .</td></tr><tr><td>Ref. 2</td><td>On</td><td>the</td><td>one hand, we do not want this potential danger</td><td>to have frightening effects on</td><td>our lives .</td></tr></table>
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td></tr><tr><td>Source</td><td>Especially for the</td><td>young people</td><td>without marrige, if he/she is</td><td>known to have some genetic risk.</td></tr><tr><td>Hyp.</td><td>Especially for the</td><td>young people</td><td>without marriage, if the latter is</td><td>known to have some genetic risk.</td></tr><tr><td>Ref. 1</td><td>Especially for unmarried</td><td>young people</td><td>marriage who are</td><td>known to have some genetic risk.</td></tr><tr><td>Ref. 2</td><td>This is especially the case for</td><td>young people</td><td>who are unmarried, if he/she is</td><td>known to have some genetic risk.</td></tr></table>
+
+metrics for the multi-reference evaluation setting.
+
+Parameter Sensitivity Analysis. The scale factors introduced in Section 3.2 dictate how much the weights of chunks change with their length. We report the corrections for various scale factors, as shown in Figure 5. The results demonstrate that CLEME is resilient to hyperparameter selection.
+
+# 5.3 Case Study
+
+Table 5 presents additional examples of CLEME. In the top group, chunk 2 and chunk 4 of the hypothesis respectively match those of Ref. 1 and Ref. 2. In this case, CLEME-dependent gives  $\mathrm{TP} = 1$  and  $\mathrm{FP} = 1$ , while CLEME-independent gives  $\mathrm{TP} = 2$ . In the second group, the hypothesis exactly corrects the ungrammatical word *frightening* in chunk 4. However, it cannot be rewarded since the entire chunk is not corrected. In the bottom group, two given references have made extensive modifications, with an unchanged chunk *young people*. Evaluating hypotheses in terms of chunks is generally more challenging than fragmented edits, but it provides a more comprehensive diagnosis.
+
+Even though there are larger grammatical errors spanning a significant portion of a sentence, CLEME would not necessarily collapse, i.e., producing one single correction chunk spanning the entire sentence. If collapse happens, the quality of the reference set should be checked first. This is because that collapse happens only if the input sentences of chunk partition are completely different, resulting in a trivial chunk partition result, which is an extreme case that has not been observed in our
+
+Table 5: Cases of chunk partition. These tables are automatically generated by CLEME. More cases from multiple datasets and language are provided in Appendix F.  
+
+<table><tr><td>Reference-based Metrics</td><td>Granularity</td><td>Score</td><td>Deterministic</td></tr><tr><td>M2 (Dahlmeier and Ng, 2012)</td><td>Phrase-level Edit</td><td>Fβ</td><td>●</td></tr><tr><td>GLEU (Napoles et al., 2015)</td><td>N-gram</td><td>Weighted Precision</td><td>○</td></tr><tr><td>ERRANT (Bryant et al., 2017)</td><td>Phrase-level Edit</td><td>Fβ</td><td>●</td></tr><tr><td>CLEME (Ours)</td><td>Chunk-level Edit</td><td>Fβ</td><td>●</td></tr></table>
+
+Table 6: A comparison of mainstream reference-based GEC metrics. GLEU is indeterministic since it involves sampling operation.
+
+experiments.
+
+# 6 Related Work
+
+# 6.1 Reference-based Metrics
+
+Reference-based metrics score GEC systems under the guidance of manually written references.  $\mathbf{M}^2$  scorer (Dahlmeier and Ng, 2012) determines an optimal edit sequence between a source sentence and a system hypothesis that achieves the highest overlap with the gold-standard annotation. The performance of each system is then represented using the  $\mathrm{F}_{0.5}$  score. However, optimality in terms of overlap does not guarantee optimality in GEC evaluation. Bryant et al. (2017) showed that  $\mathbf{M}^2$  scorer exploits its dynamic edit boundary prediction to artificially maximize true positives and minimize false positives, thus producing slightly inflated scores. Therefore, (Bryant et al., 2017) proposed ERRANT, which improves edit extraction using a linguistically-enhanced alignment algorithm and merging rules, improving the alignment of tokens with similar linguistic properties. Despite its effectiveness, ERRANT is language-dependent and bias still exists in multi-reference evaluation. Inspired by BLEU (Papineni et al., 2002) in NMT, Naples
+
+et al. (2015) proposed GLEU, an n-gram based metric for GEC evaluation. To remedy the shortcoming that  $\mathrm{F}_{0.5}$  is unable to differentiate a do-nothing system and a bad system unless  $\mathrm{TP} > 0$ , I-measure (Felice and Briscoe, 2015) generates an exact (global optimal) alignment using a three-way alignment algorithm and computes weighted accuracy to score GEC systems in terms of relative textual improvement. The comparison of reference-based GEC metrics is shown in Table 6.
+
+# 6.2 Reference-less Metrics
+
+To avoid the prerequisite of references for GEC evaluation, recent works focus on scoring GEC systems without human-annotated references. Inspired by quality estimation in neural machine translation, Naples et al. (2016) propose three Grammaticality-Based Metrics (GBMs), which are calculated by a benchmark GEC system or a pretrained ridge regression model. Asano et al. (2017) extend GBMs by introducing three assessment criteria for Grammaticality, Fluency and Meaning preservation (GFM). SOME (Yoshimura et al., 2020) further improves GFM by optimizing each Grammaticality, Fluency and Meaning preservation metric to more closely correlate with human judgements. Scribendi Score (Islam and Magnani, 2021) overcomes the limitations of SOME, requiring neither a benchmark GEC system nor fine-tuning. IMPARA (Maeda et al., 2022) comprises a quality estimator (QE) and similarity estimator (SE) based on BERT (Devlin et al., 2019), which evaluate the quality of GEC output and semantic similarity of two sentences, respectively.
+
+Although recent reference-less metrics may be highly consistent with human judgments, they always suffer from the lack of interpretability and robustness (Bryant et al., 2022), which are crucial factors for GEC evaluation. Additionally, evaluating GEC systems by leveraging pre-trained or fine-tuned language models could pose potential risks. Efficiency of reference-less metrics that rely on language models is also critical if they are used for GEC benchmarks.
+
+# 6.3 Meta Evaluation Methods
+
+It is intractable to determine the best GEC metric. A reasonable GEC metric should take into account multiple factors, including correlation with human judgments, interpretability and efficiency. Inspired by WMT human evaluation campaigns (Callison-Burch et al., 2008), 13 system outputs (incl-
+
+ing the unchanged source) from the CoNLL-2014 shared task (Ng et al., 2014) were ranked based on human rankings collected by two ranking methods: Expected Wins (EW) and TrueSkill (TS) (Grundkiewicz et al., 2015). Sakaguchi et al. (2016) collected  $8(2\times 2\times 2)$  annotations (minimal and fluency, expert and non-expert, with two corrections each), revealing that GEC metrics work differently across reference sets. Naples et al. (2019) explored how GEC metrics operate in new domains (formal and informal writing by native English speakers). They constructed a multi-reference GEC test set called GMEG-Data and a new ensemble metric GMEG-Metric.
+
+# 7 Conclusion
+
+This paper proposes CLEME, a reference-based GEC metric that aims to provide unbiased  $\mathrm{F}_{0.5}$  scores for multi-reference evaluation. We explore evaluating GEC systems based on either the correction dependence assumption or the correction independence assumption. Several possible approaches can be suggested to further improve CLEME. For example, developing (1) a GEC metric that adaptively combines dependent and independent assumptions, and (2) a weighting strategy by utilizing the knowledge of pre-trained model. In the future, we would like to develop CLEME for all languages and demonstrate the effectiveness of CLEME on other languages. It is also worthwhile to explore accuracy-based metrics.
+
+# Limitations
+
+Although CLEME can be extended to other languages, we have not tested its effectiveness in any language other than English. Furthermore, all the reference sets used in our experiments are based on the CoNLL-2014 shared task, a second-language dataset. To demonstrate the robustness of our approaches, further experiments on evaluation datasets with multiple text domains are required. We believe that introducing the correction independence assumption perspective into GEC datasets of other languages and domains could lead to more in-depth analysis and exploration.
+
+While recent PLM-based metrics have shown superior correlations compared to reference-based metrics, including ours on some reference sets, our approach enables the evaluation of GEC systems in an interpretable manner, which is a significant advantage over reference-less metrics. We leave the
+
+exploration of incorporating the PLM's knowledge into CLEME for future work.
+
+# Ethics Statement
+
+In this paper, we verify the effectiveness of our proposed approach using CoNLL-2014, BN-10GEC, and SN-8GEC reference sets, all of which are from publicly available datasets or resources on legitimate websites without sensitive data involved. All the baselines used in our experiments are also publicly available GEC metrics and we have cited the corresponding authors. We confirm that all datasets and baselines used in our experiments are consistent with their intended use.
+
+Additionally, we conduct human evaluation experiments to show the rationality of correction independence assumption. To do so, three postgraduate students specializing in foreign linguistics and applied linguistics were employed as part-time annotators. Each annotator could complete the entire annotation process within approximately 6 working hours. All annotators were paid for their work, with an average salary of approximately $5 per hour.
+
+# Acknowledgements
+
+This research is supported by National Natural Science Foundation of China (Grant No.62276154), Research Center for Computer Network (Shenzhen) Ministry of Education, the Natural Science Foundation of Guangdong Province (Grant No. 2023A1515012914), Basic Research Fund of Shenzhen City (Grant No. JCYJ20210324120012033 and JSGG20210802154402007), the Major Key Project of PCL for Experiments and Applications (PCL2021A06), and Overseas Cooperation Research Fund of Tsinghua Shenzhen International Graduate School (HW2021008).
+
+# References
+
+Hiroki Asano, Tomoya Mizumoto, and Kentaro Inui. 2017. Reference-based metrics can be replaced with reference-less metrics in evaluating grammatical error correction systems. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 343-348, Taipei, Taiwan. Asian Federation of Natural Language Processing.  
+Christopher Bryant, Mariano Felice, and Ted Briscoe. 2017. Automatic annotation and evaluation of error types for grammatical error correction. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),
+
+pages 793-805, Vancouver, Canada. Association for Computational Linguistics.  
+Christopher Bryant and Hwee Tou Ng. 2015. How far are we from fully automatic high quality grammatical error correction? In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 697-707, Beijing, China. Association for Computational Linguistics.  
+Christopher Bryant, Zheng Yuan, Muhammad Reza Qorib, Hannan Cao, Hwee Tou Ng, and Ted Briscoe. 2022. Grammatical error correction: A survey of the state of the art. arXiv preprint arXiv:2211.05166.  
+Chris Callison-Burch, Cameron Fordyce, Philipp Koehn, Christof Monz, and Josh Schroeder. 2008. Further meta-evaluation of machine translation. In Proceedings of the Third Workshop on Statistical Machine Translation, pages 70-106, Columbus, Ohio. Association for Computational Linguistics.  
+Martin Chodorow, Markus Dickinson, Ross Israel, and Joel Tetreault. 2012. Problems in evaluating grammatical error detection systems. In Proceedings of COLING 2012, pages 611-628, Mumbai, India. The COLING 2012 Organizing Committee.  
+Shamil Chollampatt and Hwee Tou Ng. 2018. A reassessment of reference-based grammatical error correction metrics. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2730-2741, Santa Fe, New Mexico, USA. Association for Computational Linguistics.  
+Leshem Choshen and Omri Abend. 2018. Automatic metric validation for grammatical error correction. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1372-1382, Melbourne, Australia. Association for Computational Linguistics.  
+Jacob Cohen. 1960. A coefficient of agreement for nominal scales. Educational and psychological measurement, 20(1):37-46.  
+Daniel Dahlmeier and Hwee Tou Ng. 2012. Better evaluation for grammatical error correction. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 568-572, Montreal, Canada. Association for Computational Linguistics.  
+Daniel Deutsch, Rotem Dror, and Dan Roth. 2022. On the limitations of reference-free evaluations of generated text. arXiv preprint arXiv:2210.12563.  
+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.  
+Mariano Felice and Ted Briscoe. 2015. Towards a standard evaluation method for grammatical error detection and correction. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 578-587, Denver, Colorado. Association for Computational Linguistics.  
+Peiyuan Gong, Xuebo Liu, Heyan Huang, and Min Zhang. 2022. Revisiting grammatical error correction evaluation and beyond. arXiv preprint arXiv:2211.01635.  
+Takumi Gotou, Ryo Nagata, Masato Mita, and Kazuaki Hanawa. 2020. Taking the correction difficulty into account in grammatical error correction evaluation. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2085-2095, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Roman Grundkiewicz, Marcin Junczys-Dowmunt, and Edward Gillian. 2015. Human evaluation of grammatical error correction systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 461-470.  
+Md Asadul Islam and Enrico Magnani. 2021. Is this the end of the gold standard? a straightforward referenceless grammatical error correction metric. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3009-3015, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Masahiro Kaneko, Sho Takase, Ayana Niwa, and Naoaki Okazaki. 2022. Interpretability for language learners using example-based grammatical error correction. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7176-7187, Dublin, Ireland. Association for Computational Linguistics.  
+Yinghui Li, Haojing Huang, Shirong Ma, Yong Jiang, Yangning Li, Feng Zhou, Hai-Tao Zheng, and Qingyu Zhou. 2023. On the (in)effectiveness of large language models for chinese text correction. CoRR, abs/2307.09007.  
+Yinghui Li, Shirong Ma, Qingyu Zhou, Zhongli Li, Li Yangning, Shulin Huang, Ruiyang Liu, Chao Li, Yunbo Cao, and Haitao Zheng. 2022a. Learning from the dictionary: Heterogeneous knowledge guided fine-tuning for Chinese spell checking. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, pages 238-249, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
+
+Yinghui Li, Qingyu Zhou, Yangning Li, Zhongli Li, Ruiyang Liu, Rongyi Sun, Zizhen Wang, Chao Li, Yunbo Cao, and Hai-Tao Zheng. 2022b. The past mistake is the future wisdom: Error-driven contrastive probability optimization for Chinese spell checking. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3202-3213, Dublin, Ireland. Association for Computational Linguistics.  
+Zuchao Li, Kevin Parnow, Masao Utiyama, Eiichiro Sumita, and Hai Zhao. 2021. MiSS: An assistant for multi-style simultaneous translation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 1-10, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Shirong Ma, Yinghui Li, Haojing Huang, Shulin Huang, Yangning Li, Hai-Tao Zheng, and Ying Shen. 2023. Progressive multi-task learning framework for Chinese text error correction. CoRR, abs/2306.17447.  
+Shirong Ma, Yinghui Li, Rongyi Sun, Qingyu Zhou, Shulin Huang, Ding Zhang, Li Yangning, Ruiyang Liu, Zhongli Li, Yunbo Cao, Haitao Zheng, and Ying Shen. 2022. Linguistic rules-based corpus generation for native Chinese grammatical error correction. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 576-589, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Matouš Macháček and Ondřej Bojar. 2013. Results of the WMT13 metrics shared task. In Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 45-51, Sofia, Bulgaria. Association for Computational Linguistics.  
+Koki Maeda, Masahiro Kaneko, and Naoaki Okazaki. 2022. IMPARA: Impact-based metric for GEC using parallel data. In Proceedings of the 29th International Conference on Computational Linguistics, pages 3578-3588, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.  
+Courtney Napolos, Maria Nadejde, and Joel Tetreault. 2019. Enabling robust grammatical error correction in new domains: Data sets, metrics, and analyses. Transactions of the Association for Computational Linguistics, 7:551-566.  
+Courtney Naples, Keisuke Sakaguchi, Matt Post, and Joel Tetreault. 2015. Ground truth for grammatical error correction metrics. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 588-593, Beijing, China. Association for Computational Linguistics.  
+Courtney Naples, Keisuke Sakaguchi, and Joel Tetreault. 2016. There's no comparison: Referenceless evaluation metrics in grammatical error correction. In Proceedings of the 2016 Conference on
+
+Empirical Methods in Natural Language Processing, pages 2109-2115, Austin, Texas. Association for Computational Linguistics.  
+Courtney Naples, Keisuke Sakaguchi, and Joel Tetreault. 2017. JFLEG: A fluency corpus and benchmark for grammatical error correction. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 229-234, Valencia, Spain. Association for Computational Linguistics.  
+Hwee Tou Ng, Siew Mei Wu, Ted Briscoe, Christian Hadiwinoto, Raymond Hendy Susanto, and Christopher Bryant. 2014. The CoNLL-2014 shared task on grammatical error correction. In Proceedings of the Eighteenth Conference on Computational Natural Language Learning: Shared Task, pages 1-14, Baltimore, Maryland. 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.  
+Keisuke Sakaguchi, Courtney Naples, Matt Post, and Joel Tetreault. 2016. Reassessing the goals of grammatical error correction: Fluency instead of grammaticality. Transactions of the Association for Computational Linguistics, 4:169-182.  
+Keisuke Sakaguchi, Matt Post, and Benjamin Van Durme. 2014. Efficient elicitation of annotations for human evaluation of machine translation. In Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 1-11, Baltimore, Maryland, USA. Association for Computational Linguistics.  
+Jingheng Ye, Yinghui Li, Yangning Li, and Hai-Tao Zheng. 2023a. Mixedit: Revisiting data augmentation and beyond for grammatical error correction.  
+Jingheng Ye, Yinghui Li, Shirong Ma, Rui Xie, Wei Wu, and Hai-Tao Zheng. 2022. Focus is what you need for chinese grammatical error correction. CoRR, abs/2210.12692.  
+Jingheng Ye, Yinghui Li, and Haitao Zheng. 2023b. System report for CCL23-eval task 7: THU KELab (sz) - exploring data augmentation and denoising for Chinese grammatical error correction. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics (Volume 3: Evaluations), pages 262-270, Harbin, China. Chinese Information Processing Society of China.  
+Ryoma Yoshimura, Masahiro Kaneko, Tomoyuki Kajiwara, and Mamoru Komachi. 2020. SOME: Reference-less sub-metrics optimized for manual evaluations of grammatical error correction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6516-6522, Barcelona, Spain (Online). International Committee on Computational Linguistics.
+
+Ding Zhang, Yinghui Li, Qingyu Zhou, Shirong Ma, Yangning Li, Yunbo Cao, and Hai-Tao Zheng. 2023. Contextual similarity is more valuable than character similarity: An empirical study for chinese spell checking. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5.  
+Yue Zhang, Zhenghua Li, Zuyi Bao, Jiacheng Li, Bo Zhang, Chen Li, Fei Huang, and Min Zhang. 2022. MuCGEC: a multi-reference multi-source evaluation dataset for Chinese grammatical error correction. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3118-3130, Seattle, United States. Association for Computational Linguistics.
+
+# A Statistics of Reference Sets
+
+Table 7 presents the statistics of all reference sets involved in our experiments, including In-Corrected-Chunk (ICC) ratio, Unchanged-Chunk (IUC) ratio and Cross-Chunk (CC) ratio. It is worth noting that all reference sets exhibit a low CC ratio with varying ICC and IUC ratios, indicating the rationality and feasibility of evaluating GEC systems following the correction independence assumption.
+
+# B Hyperparameters
+
+The hyperparameters of our proposed CLEME consist of scale factors  $\alpha$  and thresholds. We tune the hyperparameters on CoNLL-2014 and keep them on the other reference sets to demonstrate the adaptability of our method. The hyperparameters of CLEME are listed in Table 8.
+
+# C Evaluate by Accuracy
+
+Conventional reference-based metrics such as Max-Match  $(\mathbf{M}^2)$  and ERRANT are unable to calculate accuracy because they do not define True Negatives (TNs)  $^{6}$ . In order to implement the computation of accuracy, CLEME defines TNs as hypothesis unchanged chunks that match the chunks of references. Similar to  $\mathrm{F}_{0.5}$ , accuracy can be computed based on correction dependence or independence assumptions in both corpus- and sentence-level, resulting in four new variants: 1) CLEME-dependent-acc, 2) CLEME-independent-acc, 3) SentCLEME-dependent-acc, and 4) SentCLEME-independent-acc.
+
+<table><tr><td>Item</td><td>CoNLL-2014</td><td>BN-10GEC</td><td>E-Minimal</td><td>E-Fluency</td><td>NE-Minimal</td><td>NE-Fluency</td></tr><tr><td># Sents (Length)</td><td>1,312 (23.0)</td><td>1,312 (23.0)</td><td>1,312 (23.0)</td><td>1,312 (23.0)</td><td>1,312 (23.0)</td><td>1,312 (23.0)</td></tr><tr><td># Refs (Length)</td><td>2,624 (22.8)</td><td>13,120 (22.9)</td><td>2,624 (23.2)</td><td>2,624 (22.8)</td><td>2,624 (23.0)</td><td>2,624 (22.2)</td></tr><tr><td># Edits (Length)</td><td>5,937 (1.0)</td><td>36,677 (1.0)</td><td>4,500 (1.0)</td><td>8,373 (1.1)</td><td>4,964 (0.9)</td><td>11,033 (1.2)</td></tr><tr><td># UnchangedChunks (Length)</td><td>11,174 (4.8)</td><td>93,496 (2.5)</td><td>8,887 (6.3)</td><td>12,823 (3.8)</td><td>10,748 (5.1)</td><td>14,086 (2.9)</td></tr><tr><td># Corrected/DummyChunks (Length)</td><td>4,994 (1.3)</td><td>26,948 (2.4)</td><td>3,963 (1.2)</td><td>6,305 (1.7)</td><td>4,221 (1.2)</td><td>6,892 (2.6)</td></tr><tr><td>ICC (Number)</td><td>51.05% (3,031)</td><td>90.66% (33,251)</td><td>62.47% (2,811)</td><td>52.69% (4,412)</td><td>43.84% (2,176)</td><td>43.32% (4,780)</td></tr><tr><td>IUC (Number)</td><td>45.61% (2,708)</td><td>7.74% (2,837)</td><td>36.6% (1,647)</td><td>42.51% (3,559)</td><td>54.3% (2,698)</td><td>48.92% (5,397)</td></tr><tr><td>CC (Number)</td><td>3.34% (198)</td><td>1.61% (589)</td><td>0.93% (42)</td><td>4.8% (402)</td><td>1.81% (90)</td><td>7.76% (856)</td></tr></table>
+
+Table 7: Statistics of CoNLL-2014 (Ng et al., 2014), BN-10GEC (Bryant and Ng, 2015) and SN-8GEC (Sakaguchi et al., 2016) reference sets. We use ERRANT (Bryant et al., 2017) for edit extraction.  
+
+<table><tr><td>Configuration</td><td>CLEME-dependent(-acc)</td><td>CLEME-independent(-acc)</td><td>SentCLEME-dependent(-acc)</td><td>SentCLEME-independent(-acc)</td></tr><tr><td>Scale factor of TPs α1</td><td>2</td><td>2</td><td>10</td><td>10</td></tr><tr><td>Scale factor of FPs α2</td><td>2</td><td>2</td><td>10</td><td>10</td></tr><tr><td>Scale factor of FNs α3</td><td>2</td><td>2</td><td>-</td><td>-</td></tr><tr><td>Scale factor of TNs α4</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>Threshold of TPs</td><td>(0.75, 1.25)</td><td>(0.75, 1.25)</td><td>(1.00, 10.00)</td><td>(2.50, 10.00)</td></tr><tr><td>Threshold of FPs</td><td>(0.75, 1.25)</td><td>(0.75, 1.25)</td><td>(0.25, 10.00)</td><td>(0.25, 1.00)</td></tr><tr><td>Threshold of FNs</td><td>(0.75, 1.25)</td><td>(0.75, 1.25)</td><td>(1.00, 1.00)</td><td>(1.00, 1.00)</td></tr><tr><td>Threshold of TNs</td><td>(1.00, 1.00)</td><td>(1.00, 1.00)</td><td>(1.00, 1.00)</td><td>(1.00, 1.00)</td></tr></table>
+
+Table 8: Hyperparameter values used in our experiments.
+
+The results of human correlations are reported in Table 9. Accuracy-based metrics perform very differently at the corpus- and sentence-level, which is similar to the findings (Napoles et al., 2016, 2019). Surprisingly, two accuracy-based corpus-level metrics, i.e., CLEME-dependent-acc and CLEME-independent-acc, result in negative correlations on all reference sets. However, their sentence-level variants, i.e., SentCLEME-dependent-acc and SentCLEME-independent-acc, perform well and achieve the highest correlations on some reference sets. Regarding the disparity between the performance of accuracy-based metrics and  $\mathrm{F_{0.5}}$ -based metrics at the sentence level, one notable difference is their stability or robustness on reference sets with varying numbers of references and annotation styles.  $\mathrm{F_{0.5}}$ -based metrics are more robust to different reference sets, where SentCLEME(in)dependent achieves comparable correlations with the best metric on all reference sets. However, the performance of accuracy-based metrics lags far behind other metrics on some reference sets (BN10GEC, E-Minimal and NE-Fluency). A deeper investigation into this phenomenon is needed to understand the instability of accuracy-based metrics. We leave the exploration and further analysis of accuracy-based metric for future work.
+
+# D Detailed Analysis
+
+Table 10 reports the detailed evaluation results of 13 systems on CoNLL-2014. The reason behind the lower TP and FP counts of CLEME as compared
+
+to ERRANT is attributed to the chunk partition process, where overlapping edits are merged into chunks. It is worth noting that the FN counts of CLEME are significantly lower than those of ERRANT because of their distinct definition. While ERRANT considers FNs as the edits of references that are not identical to hypotheses, CLEME defines them as the corrected/dummy chunks of references that do not exactly match the chunks of hypotheses. We believe that the definition of ERRANT could be problematic, as it has a tendency to overestimate the FN counts of GEC systems. This may result in an underestimated Recall rate in turn.
+
+-dependent v.s. -independent. Comparing the Precision and Recall of (Sent)CLEME-independent to those of (Sent)CLEME-dependent, it is observed that the former has a slightly higher value. This is because (Sent)CLEME-independent has the potential to overestimate the performance of GEC systems, whereas (Sent)CLEME-dependent could result in underestimating the same. It is noteworthy that both metrics provide an upper bound and lower bound for GEC performance, respectively.
+
+Corpus-level v.s. Sentence-level. The precision, recall, and  $\mathrm{F}_{0.5}$  scores of sentence-level metrics are considerably higher than those of corpus-level variants. There might be several factors contributing to this difference, but one possible explanation is that precision and recall values get affected by a limited number of challenging samples that contain numerous corrected/dummy chunks.
+
+<table><tr><td rowspan="2" colspan="2">Metric</td><td colspan="2">CoNLL-2014</td><td colspan="2">BN-10GEC</td><td colspan="2">E-Minimal</td><td colspan="2">E-Fluency</td><td colspan="2">NE-Minimal</td><td colspan="2">NE-Fluency</td></tr><tr><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td><td>EW</td><td>TS</td></tr><tr><td rowspan="2">M2</td><td>γ</td><td>0.623</td><td>0.672</td><td>0.547</td><td>0.610</td><td>0.597</td><td>0.650</td><td>0.590</td><td>0.659</td><td>0.575</td><td>0.634</td><td>0.582</td><td>0.649</td></tr><tr><td>ρ</td><td>0.687</td><td>0.720</td><td>0.648</td><td>0.692</td><td>0.654</td><td>0.703</td><td>0.654</td><td>0.709</td><td>0.577</td><td>0.648</td><td>0.648</td><td>0.703</td></tr><tr><td rowspan="2">GLEU</td><td>γ</td><td>0.701</td><td>0.750</td><td>0.678</td><td>0.761</td><td>0.533</td><td>0.513</td><td>0.693</td><td>0.771</td><td>-0.044</td><td>-0.113</td><td>0.674</td><td>0.767</td></tr><tr><td>ρ</td><td>0.467</td><td>0.555</td><td>0.754</td><td>0.806</td><td>0.577</td><td>0.511</td><td>0.710</td><td>0.757</td><td>-0.005</td><td>-0.055</td><td>0.725</td><td>0.819</td></tr><tr><td rowspan="2">ERRANT</td><td>γ</td><td>0.642</td><td>0.688</td><td>0.586</td><td>0.644</td><td>0.578</td><td>0.631</td><td>0.594</td><td>0.663</td><td>0.585</td><td>0.637</td><td>0.597</td><td>0.659</td></tr><tr><td>ρ</td><td>0.659</td><td>0.698</td><td>0.637</td><td>0.698</td><td>0.742</td><td>0.786</td><td>0.720</td><td>0.775</td><td>0.747</td><td>0.797</td><td>0.753</td><td>0.797</td></tr><tr><td rowspan="2">CLEME-dependent (Ours)</td><td>γ</td><td>0.648</td><td>0.691</td><td>0.602</td><td>0.656</td><td>0.594</td><td>0.644</td><td>0.589</td><td>0.654</td><td>0.595</td><td>0.643</td><td>0.612</td><td>0.673</td></tr><tr><td>ρ</td><td>0.709</td><td>0.742</td><td>0.692</td><td>0.747</td><td>0.797</td><td>0.813</td><td>0.714</td><td>0.775</td><td>0.786</td><td>0.835</td><td>0.720</td><td>0.791</td></tr><tr><td rowspan="2">CLEME-independent (Ours)</td><td>γ</td><td>0.649</td><td>0.691</td><td>0.609</td><td>0.659</td><td>0.593</td><td>0.643</td><td>0.587</td><td>0.653</td><td>0.601</td><td>0.647</td><td>0.611</td><td>0.672</td></tr><tr><td>ρ</td><td>0.709</td><td>0.731</td><td>0.692</td><td>0.747</td><td>0.791</td><td>0.802</td><td>0.731</td><td>0.791</td><td>0.797</td><td>0.841</td><td>0.714</td><td>0.786</td></tr><tr><td rowspan="2">CLEME-dependent-acc (Ours)</td><td>γ</td><td>-0.261</td><td>-0.342</td><td>-0.288</td><td>-0.371</td><td>-0.222</td><td>-0.313</td><td>-0.216</td><td>-0.302</td><td>-0.370</td><td>-0.453</td><td>-0.430</td><td>-0.513</td></tr><tr><td>ρ</td><td>-0.407</td><td>-0.478</td><td>-0.445</td><td>-0.516</td><td>-0.335</td><td>-0.423</td><td>-0.347</td><td>-0.437</td><td>-0.429</td><td>-0.516</td><td>-0.473</td><td>-0.555</td></tr><tr><td rowspan="2">CLEME-independent-acc (Ours)</td><td>γ</td><td>-0.175</td><td>-0.262</td><td>-0.206</td><td>-0.284</td><td>-0.195</td><td>-0.283</td><td>-0.105</td><td>-0.189</td><td>-0.335</td><td>-0.420</td><td>-0.328</td><td>-0.412</td></tr><tr><td>ρ</td><td>-0.176</td><td>-0.264</td><td>-0.341</td><td>-0.418</td><td>-0.291</td><td>-0.379</td><td>-0.132</td><td>-0.231</td><td>-0.429</td><td>-0.516</td><td>-0.451</td><td>-0.522</td></tr><tr><td rowspan="2">SentM2</td><td>γ</td><td>0.871</td><td>0.864</td><td>0.567</td><td>0.646</td><td>0.805*</td><td>0.836*</td><td>0.655</td><td>0.732</td><td>0.729*</td><td>0.785*</td><td>0.621</td><td>0.699</td></tr><tr><td>ρ</td><td>0.731</td><td>0.758</td><td>0.593</td><td>0.648</td><td>0.806*</td><td>0.845*</td><td>0.731</td><td>0.764</td><td>0.797*</td><td>0.846*</td><td>0.632</td><td>0.687</td></tr><tr><td rowspan="2">SentGLEU</td><td>γ</td><td>0.784</td><td>0.828</td><td>0.756</td><td>0.826</td><td>0.742*</td><td>0.773*</td><td>0.785</td><td>0.846</td><td>0.723*</td><td>0.762*</td><td>0.778</td><td>0.848</td></tr><tr><td>ρ</td><td>0.720</td><td>0.775</td><td>0.769</td><td>0.824</td><td>0.764*</td><td>0.797*</td><td>0.791</td><td>0.846</td><td>0.764*</td><td>0.830*</td><td>0.768</td><td>0.846</td></tr><tr><td rowspan="2">SentERRANT</td><td>γ</td><td>0.870</td><td>0.846</td><td>0.885</td><td>0.896</td><td>0.768*</td><td>0.803*</td><td>0.806</td><td>0.732</td><td>0.710*</td><td>0.765*</td><td>0.793</td><td>0.847</td></tr><tr><td>ρ</td><td>0.742</td><td>0.747</td><td>0.786</td><td>0.830</td><td>0.775*</td><td>0.819*</td><td>0.813</td><td>0.764</td><td>0.780*</td><td>0.841*</td><td>0.830</td><td>0.857</td></tr><tr><td rowspan="2">SentCLEME-dependent (Ours)</td><td>γ</td><td>0.876</td><td>0.844</td><td>0.915</td><td>0.913</td><td>0.806*</td><td>0.838*</td><td>0.849</td><td>0.886</td><td>0.742*</td><td>0.795*</td><td>0.876</td><td>0.921</td></tr><tr><td>ρ</td><td>0.824</td><td>0.808</td><td>0.835</td><td>0.874</td><td>0.775*</td><td>0.819*</td><td>0.824</td><td>0.863</td><td>0.797*</td><td>0.846*</td><td>0.791</td><td>0.846</td></tr><tr><td rowspan="2">SentCLEME-independent (Ours)</td><td>γ</td><td>0.868</td><td>0.857</td><td>0.855*</td><td>0.876*</td><td>0.821*</td><td>0.856*</td><td>0.841</td><td>0.877</td><td>0.782*</td><td>0.831*</td><td>0.852</td><td>0.896</td></tr><tr><td>ρ</td><td>0.725</td><td>0.758</td><td>0.659*</td><td>0.714*</td><td>0.775*</td><td>0.819*</td><td>0.808</td><td>0.846</td><td>0.819*</td><td>0.874*</td><td>0.762</td><td>0.825</td></tr><tr><td rowspan="2">SentCLEME-dependent-acc (Ours)</td><td>γ</td><td>0.828</td><td>0.857</td><td>0.650</td><td>0.719</td><td>0.808*</td><td>0.838*</td><td>0.679</td><td>0.740</td><td>0.757*</td><td>0.811*</td><td>0.557</td><td>0.641</td></tr><tr><td>ρ</td><td>0.813</td><td>0.841</td><td>0.682</td><td>0.740</td><td>0.830*</td><td>0.852*</td><td>0.731</td><td>0.786</td><td>0.853*</td><td>0.894*</td><td>0.655</td><td>0.702</td></tr><tr><td rowspan="2">SentCLEME-independent-acc (Ours)</td><td>γ</td><td>0.900*</td><td>0.920*</td><td>0.604*</td><td>0.555*</td><td>0.693*</td><td>0.637*</td><td>0.840*</td><td>0.891*</td><td>0.791*</td><td>0.763*</td><td>0.756*</td><td>0.822*</td></tr><tr><td>ρ</td><td>0.830*</td><td>0.849*</td><td>0.363*</td><td>0.303*</td><td>0.588*</td><td>0.544*</td><td>0.857*</td><td>0.890*</td><td>0.654*</td><td>0.626*</td><td>0.747*</td><td>0.819*</td></tr></table>
+
+Table 9: Overview of correlations between reference-based metrics and human judgments. We highlight the highest score in bold and the second-highest score with underlines.  $\clubsuit$  We remove unchanged reference sentences for higher correlations due to low-quality annotations. Otherwise, negative correlations are possible.  
+
+<table><tr><td>Metric</td><td></td><td>AMU</td><td>CAMB</td><td>CUUI</td><td>IITB</td><td>INPUT</td><td>IPN</td><td>NTHU</td><td>PKU</td><td>POST</td><td>RAC</td><td>SJTU</td><td>UFC</td><td>UMC</td></tr><tr><td rowspan="6">ERRANT</td><td>TP</td><td>483</td><td>725</td><td>607</td><td>28</td><td>0</td><td>52</td><td>409</td><td>294</td><td>508</td><td>319</td><td>104</td><td>36</td><td>311</td></tr><tr><td>FP</td><td>795</td><td>1329</td><td>985</td><td>65</td><td>0</td><td>488</td><td>991</td><td>697</td><td>1152</td><td>794</td><td>261</td><td>14</td><td>774</td></tr><tr><td>FN</td><td>1934</td><td>1886</td><td>1946</td><td>2064</td><td>2070</td><td>2078</td><td>1976</td><td>2007</td><td>1985</td><td>2044</td><td>2036</td><td>2069</td><td>2020</td></tr><tr><td>P</td><td>37.79</td><td>35.30</td><td>38.13</td><td>30.11</td><td>100.0</td><td>9.63</td><td>29.21</td><td>29.67</td><td>30.60</td><td>28.66</td><td>28.49</td><td>72.00</td><td>28.66</td></tr><tr><td>R</td><td>19.98</td><td>27.77</td><td>23.78</td><td>1.34</td><td>0.00</td><td>2.44</td><td>17.15</td><td>12.78</td><td>20.38</td><td>13.50</td><td>4.86</td><td>1.71</td><td>13.34</td></tr><tr><td>F0.5</td><td>32.08</td><td>33.48</td><td>34.02</td><td>5.68</td><td>0.00</td><td>6.06</td><td>25.61</td><td>23.46</td><td>27.81</td><td>23.40</td><td>14.44</td><td>7.81</td><td>23.31</td></tr><tr><td rowspan="10">CLEME-dependent</td><td>TP</td><td>314</td><td>482</td><td>379</td><td>17</td><td>0</td><td>33</td><td>266</td><td>195</td><td>333</td><td>203</td><td>69</td><td>24</td><td>213</td></tr><tr><td>w/o LW</td><td>382</td><td>588</td><td>471</td><td>22</td><td>0</td><td>39</td><td>330</td><td>246</td><td>412</td><td>254</td><td>85</td><td>32</td><td>216</td></tr><tr><td>FP</td><td>872</td><td>1392</td><td>1034</td><td>72</td><td>0</td><td>529</td><td>975</td><td>776</td><td>1246</td><td>782</td><td>292</td><td>19</td><td>844</td></tr><tr><td>w/o LW</td><td>815</td><td>1303</td><td>964</td><td>67</td><td>0</td><td>488</td><td>905</td><td>709</td><td>1144</td><td>782</td><td>272</td><td>18</td><td>788</td></tr><tr><td>FN</td><td>1182</td><td>987</td><td>1169</td><td>1564</td><td>1592</td><td>1445</td><td>1191</td><td>1259</td><td>1158</td><td>1278</td><td>1471</td><td>1583</td><td>1266</td></tr><tr><td>w/o LW</td><td>1345</td><td>1132</td><td>1333</td><td>1751</td><td>1782</td><td>1634</td><td>1366</td><td>1426</td><td>1333</td><td>1453</td><td>1657</td><td>1772</td><td>1439</td></tr><tr><td>TN</td><td>6312</td><td>6347</td><td>6245</td><td>6313</td><td>6308</td><td>6412</td><td>6295</td><td>6314</td><td>6449</td><td>6310</td><td>6324</td><td>6280</td><td>6377</td></tr><tr><td>P</td><td>26.45</td><td>25.74</td><td>26.81</td><td>19.29</td><td>100.0</td><td>5.85</td><td>21.42</td><td>20.06</td><td>21.07</td><td>20.60</td><td>19.02</td><td>56.40</td><td>20.14</td></tr><tr><td>R</td><td>20.97</td><td>32.84</td><td>24.48</td><td>1.09</td><td>0.00</td><td>2.22</td><td>18.23</td><td>13.39</td><td>22.31</td><td>13.71</td><td>4.45</td><td>1.52</td><td>14.40</td></tr><tr><td>F0.5</td><td>25.14</td><td>26.90</td><td>26.31</td><td>4.45</td><td>0.00</td><td>4.41</td><td>20.69</td><td>18.24</td><td>21.31</td><td>18.72</td><td>11.50</td><td>6.85</td><td>18.65</td></tr><tr><td rowspan="3">SentCLEME-dependent</td><td>P</td><td>63.05</td><td>41.07</td><td>57.60</td><td>95.27</td><td>100.0</td><td>67.45</td><td>55.17</td><td>63.62</td><td>53.26</td><td>65.05</td><td>82.90</td><td>98.70</td><td>61.78</td></tr><tr><td>R</td><td>48.87</td><td>59.94</td><td>51.21</td><td>32.37</td><td>31.33</td><td>36.07</td><td>49.71</td><td>44.84</td><td>50.61</td><td>44.57</td><td>36.06</td><td>32.15</td><td>44.43</td></tr><tr><td>F0.5</td><td>36.24</td><td>32.94</td><td>37.39</td><td>31.51</td><td>31.33</td><td>23.25</td><td>32.24</td><td>32.56</td><td>33.34</td><td>32.37</td><td>31.93</td><td>32.30</td><td>31.46</td></tr><tr><td rowspan="10">CLEME-independent</td><td>TP</td><td>318</td><td>488</td><td>392</td><td>17</td><td>0</td><td>33</td><td>272</td><td>196</td><td>339</td><td>204</td><td>69</td><td>24</td><td>214</td></tr><tr><td>w/o LW</td><td>388</td><td>596</td><td>487</td><td>22</td><td>0</td><td>39</td><td>338</td><td>248</td><td>420</td><td>255</td><td>85</td><td>32</td><td>262</td></tr><tr><td>FP</td><td>864</td><td>1382</td><td>1016</td><td>72</td><td>0</td><td>529</td><td>965</td><td>773</td><td>1236</td><td>781</td><td>292</td><td>19</td><td>843</td></tr><tr><td>w/o LW</td><td>809</td><td>1295</td><td>948</td><td>67</td><td>0</td><td>488</td><td>897</td><td>707</td><td>1136</td><td>781</td><td>272</td><td>18</td><td>787</td></tr><tr><td>FN</td><td>928</td><td>701</td><td>884</td><td>1362</td><td>1393</td><td>1246</td><td>937</td><td>1022</td><td>883</td><td>1025</td><td>1266</td><td>1372</td><td>1026</td></tr><tr><td>w/o LW</td><td>1029</td><td>778</td><td>984</td><td>1497</td><td>1530</td><td>1382</td><td>1045</td><td>1129</td><td>990</td><td>1135</td><td>1398</td><td>1506</td><td>1136</td></tr><tr><td>TN</td><td>6629</td><td>6701</td><td>6597</td><td>6567</td><td>6560</td><td>6664</td><td>6617</td><td>6611</td><td>6793</td><td>6628</td><td>6583</td><td>6546</td><td>6680</td></tr><tr><td>P</td><td>26.90</td><td>26.11</td><td>27.85</td><td>19.29</td><td>100.0</td><td>5.85</td><td>22.00</td><td>20.23</td><td>21.50</td><td>20.69</td><td>19.02</td><td>56.40</td><td>20.22</td></tr><tr><td>R</td><td>25.53</td><td>41.06</td><td>30.71</td><td>1.25</td><td>0.00</td><td>2.57</td><td>22.52</td><td>16.10</td><td>27.72</td><td>16.59</td><td>5.14</td><td>1.75</td><td>17.23</td></tr><tr><td>F0.5</td><td>26.61</td><td>28.16</td><td>28.38</td><td>4.97</td><td>0.00</td><td>4.66</td><td>22.10</td><td>19.24</td><td>22.51</td><td>19.71</td><td>12.35</td><td>7.77</td><td>19.54</td></tr><tr><td rowspan="3">SentCLEME-independent</td><td>P</td><td>65.36</td><td>45.39</td><td>60.92</td><td>95.27</td><td>100.0</td><td>67.51</td><td>57.44</td><td>65.03</td><td>56.26</td><td>66.65</td><td>83.17</td><td>98.70</td><td>63.36</td></tr><tr><td>R</td><td>57.20</td><td>70.15</td><td>60.76</td><td>36.87</td><td>35.29</td><td>40.56</td><td>58.08</td><td>52.20</td><td>59.83</td><td>52.14</td><td>41.31</td><td>36.59</td><td>51.94</td></tr><tr><td>F0.5</td><td>42.00</td><td>39.63</td><td>43.76</td><td>35.49</td><td>35.29</td><td>25.90</td><td>38.13</td><td>37.43</td><td>39.38</td><td>37.04</td><td>36.25</td><td>36.48</td><td>36.59</td></tr></table>
+
+Table 10: Detailed evaluation results across 13 GEC systems on CoNLL-2014.
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td><td>Chunk 5</td><td>Chunk 6</td><td>Chunk 7</td></tr><tr><td>Source</td><td>If</td><td></td><td>not</td><td>their family then</td><td>who else</td><td>that are</td><td>willing to do that?</td></tr><tr><td>Ref. 1</td><td>If</td><td></td><td>not</td><td>their family then</td><td>who else</td><td>will be</td><td>willing to do that?</td></tr><tr><td>Ref. 2</td><td>If</td><td></td><td>not</td><td>their family then</td><td>who else</td><td>would be</td><td>willing to do that?</td></tr><tr><td>Ref. 3</td><td>If</td><td></td><td>not</td><td>from your family then</td><td>who else</td><td>is</td><td>willing to do that?</td></tr><tr><td>Ref. 4</td><td>If</td><td></td><td>not</td><td>their family , then</td><td>who else</td><td>will be</td><td>willing to do that?</td></tr><tr><td>Ref. 5</td><td>If</td><td></td><td>not</td><td>their family then</td><td>who else</td><td>will be</td><td>willing to do that?</td></tr><tr><td>Ref. 6</td><td>If</td><td></td><td>not</td><td>their family</td><td>who else</td><td>would be</td><td>willing to do that?</td></tr><tr><td>Ref. 7</td><td>If</td><td></td><td>not</td><td>their family then</td><td>who else</td><td>will be</td><td>willing to do that?</td></tr><tr><td>Ref. 8</td><td>If</td><td></td><td>not</td><td>their family ,</td><td>who else</td><td>is</td><td>willing to do that?</td></tr><tr><td>Ref. 9</td><td>If</td><td>family do</td><td>not</td><td>help then</td><td>who else</td><td>would be</td><td>willing to do that?</td></tr><tr><td>Ref. 10</td><td>If</td><td></td><td>not</td><td>their family , then</td><td>who else</td><td>is</td><td>willing to do that?</td></tr></table>
+
+Table 11: A case of correction independence. We apply chunk partition to the source and all the references.  
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td></tr><tr><td>Source</td><td>For not</td><td>use</td><td>car</td><td>.</td></tr><tr><td>Ref. 1</td><td>Not for</td><td>use</td><td>with a car</td><td>.</td></tr><tr><td>Ref. 2</td><td>Do not</td><td>use</td><td>in the car</td><td>.</td></tr><tr><td>Ref. 3</td><td>Car not for</td><td>use</td><td></td><td>.</td></tr><tr><td>Ref. 4</td><td>Can not</td><td>use</td><td>the car</td><td>.</td></tr></table>
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td><td>Chunk 5</td><td>Chunk 6</td></tr><tr><td>Source</td><td>One person if do n’t</td><td>have good health</td><td>that means</td><td>so many things</td><td>they could lost</td><td>.</td></tr><tr><td>Ref. 1</td><td>If a person does n’t</td><td>have good health</td><td>,</td><td>so many things</td><td>could be lost</td><td>.</td></tr><tr><td>Ref. 2</td><td>If one person does not</td><td>have good health</td><td>, that means they could lose</td><td>so many things</td><td></td><td>.</td></tr><tr><td>Ref. 3</td><td>If one person does n’t</td><td>have good health</td><td>, that means they could lose</td><td>so many things</td><td></td><td>.</td></tr><tr><td>Ref. 4</td><td>If one person does n’t</td><td>have good health</td><td>, that means</td><td>so many things</td><td>they could lost</td><td>.</td></tr></table>
+
+<table><tr><td></td><td>Chunk 1</td><td>Chunk 2</td><td>Chunk 3</td><td>Chunk 4</td><td>Chunk 5</td><td>Chunk 6</td></tr><tr><td>Source</td><td>今天</td><td>听天气预报说</td><td>今</td><td>天</td><td>还有天气</td><td>冷。</td></tr><tr><td>Ref. 1</td><td></td><td>听天气预报说</td><td>今</td><td>天</td><td>天气</td><td>冷。</td></tr><tr><td>Ref. 2</td><td>今天</td><td>听天气预报说</td><td></td><td>天</td><td>气还会变</td><td>冷。</td></tr><tr><td>Ref. 3</td><td></td><td>听天气预报说</td><td>今</td><td>天</td><td>天气还会变</td><td>冷。</td></tr></table>
+
+<table><tr><td></td><td colspan="2">Chunk 1</td><td colspan="2">Chunk 2</td><td colspan="2">Chunk 3</td><td colspan="2">Chunk 4</td><td colspan="2">Chunk 5</td><td colspan="2">Chunk 6</td><td colspan="2">Chunk 7</td><td colspan="2">Chunk 8</td></tr><tr><td>Source</td><td>所以</td><td>我从小到现在在这些快餐店</td><td>店</td><td>吃饭的机会很少</td><td>。</td><td>对我来说每</td><td>次顿</td><td>饭都很重要。</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Ref. 1</td><td>所以</td><td>我从小到现在在这些快餐店</td><td>店</td><td>吃饭的机会很少</td><td>。</td><td>对我来说每</td><td>次吃</td><td>饭都很重要。</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Ref. 2</td><td>所以</td><td>我从小到现在在这些快餐店</td><td>店</td><td>吃饭的机会很少</td><td>。</td><td>对我来说每</td><td>次吃</td><td>饭都很重要。</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Ref. 3</td><td></td><td>我从小到现在在这些快餐店</td><td>店</td><td>吃饭的机会很少</td><td>，所以</td><td>对我来说每</td><td>次吃</td><td>饭都很重要。</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>Ref. 4</td><td></td><td>我从小到现在在这些快餐店</td><td>店</td><td>吃饭的机会很少</td><td>，所以</td><td>对我来说每</td><td>顿</td><td>饭都很重要。</td><td></td><td></td><td></td><td></td><td></td><td></td><td></td><td></td></tr></table>
+
+Table 12: More cases of chunk partition. These tables are automatically generated by CLEME. The first two cases are from JELEG-dev, and the next two cases are from MuCGEC-dev.
+
+# E Correction Independence
+
+We introduce the term correction independence to describe a pair of chunks where the correction of each chunk is not related to the correction of the other, as illustrated in Table 11. Specifically, chunk 2 and chunk 4 are considered correction-dependent because the correction of chunk 2 family do from Ref.9 must be matched with the correction of chunk 4 help then from Ref.9. However, chunk 6 is correction-independent with chunk 2 (or 4) since the correction of chunk 6 has no impact on the correction of chunk 2 (or 4).
+
+# F More cases
+
+We list more cases in Table 12, which involve JFLEG (Napoles et al., 2017) for English, and MuCGEC (Zhang et al., 2022) for Chinese.

2023/CLEME_ Debiasing Multi-reference Evaluation for Grammatical Error Correction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6691095527c8293090324fe0677e73967eb2682f377bbe42c28f1ed76c3405a6
+size 1447125

2023/CLEVR-Implicit_ A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension/7a840248-0cf8-41af-aee2-5d7039744b4d_content_list.json

@@ -0,0 +1,1787 @@
+[
+    {
+        "type": "text",
+        "text": "CLEVR-Implicit: A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension",
+        "text_level": 1,
+        "bbox": [
+            163,
+            87,
+            835,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Jingwei Zhang $^{1,\\ast}$ , Xin Wu $^{1,\\ast}$ , Yi Cai $^{1,2,\\dagger}$",
+        "bbox": [
+            332,
+            141,
+            673,
+            159
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>School of Software Engineering, South China University of Technology",
+        "bbox": [
+            206,
+            159,
+            794,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Key Laboratory of Big Data and Intelligent Robot (South China University of Technology) Ministry of Education",
+        "bbox": [
+            194,
+            175,
+            806,
+            208
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ycai@scut.edu.cn",
+        "bbox": [
+            418,
+            211,
+            583,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recently, pre-trained vision-language (VL) models have achieved remarkable success in various cross-modal tasks, including referring expression comprehension (REC). These models are pre-trained on the large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the corresponding images and then fine-tuned on downstream tasks. However, the performance of VL models is hindered when dealing with implicit text, which describes objects through comparisons between two or more objects rather than explicitly mentioning them. This is because the models struggle to align the implicit text with the objects in the images. To address the challenge, we introduce CLEVR-Implicit, a dataset consisting of synthetic images and corresponding two types of implicit text for the REC task. Additionally, to enhance the performance of VL models on implicit text, we propose a method called Transform Implicit Text into Explicit Text (TIE), which enables VL models to process with the implicit text. TIE consists of two modules: (1) the prompt design module builds prompts for implicit text by adding masked tokens, and (2) the cloze procedure module fine-tunes the prompts by utilizing masked language modeling (MLM) to predict the explicit words with the implicit prompts. Experimental results on our dataset demonstrate a significant improvement of  $37.94\\%$  in the performance of VL models on implicit text after employing our TIE method.",
+        "bbox": [
+            141,
+            279,
+            460,
+            747
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            762,
+            258,
+            777
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In recent years, pre-trained vision-language models have demonstrated remarkable achievements in various vision-language downstream tasks (Lu et al., 2019; Su et al., 2019; Kim et al., 2021; Yao et al., 2021; Kamath et al., 2021; Liu et al., 2022), including referring expression comprehension (REC).",
+        "bbox": [
+            112,
+            788,
+            490,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/77222350d0df425f570fe3c29d53208ff917f0073d72d9c8c398573d991a08d6.jpg",
+        "image_caption": [
+            "The left one of the only two objects with the same color and material"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            250,
+            816,
+            363
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/fe8d46a1c8337935d4cd0d7fbf538785d0ed7c9d16d5b87a45bec261f3c50a47.jpg",
+        "image_caption": [
+            "The right one of the only two objects with different shape and material",
+            "Figure 1: Examples from CLEVR-Implicit. There are two types of implicit text: Same and Different."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            576,
+            395,
+            815,
+            507
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "REC involves identifying a specific object in an image based on corresponding textual descriptions. A common approach is to pre-train a vision-language model on large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the images. Fine-tuning is then performed to optimize model performance (Lu et al., 2019; Su et al., 2019; Kamath et al., 2021).",
+        "bbox": [
+            507,
+            609,
+            884,
+            738
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite the state-of-the-art results demonstrated by these methods, they still exhibit limitations. Pre-trained models are good at processing explicit text, which describes the appearance attributes of the objects in the image, whether they are reference or target objects. For example, consider the phrase \"The object left on the red cube.\", models rely on the explicit word \"red cube\" to locate the referred object in the image. However, they encounter challenges when it comes to implicit reasoning, which is processing the implicit text in REC. Implicit",
+        "bbox": [
+            507,
+            741,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Equal contribution.",
+        "bbox": [
+            141,
+            892,
+            272,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "† Corresponding author.",
+        "bbox": [
+            142,
+            904,
+            287,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12820",
+        "bbox": [
+            475,
+            927,
+            526,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12820-12830 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            152,
+            945,
+            843,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "reasoning occurs when the textual expression indirectly describes an object through a comparison between two or more objects, one type of implicit text is formatted like \"The left one of two same color objects\". The phrase \"two same color objects\" does not describe the appearance attributes of any objects, only the relationships between them, and there is no direct reference for the models to infer the target. Consequently, models struggle to align \"two same color objects\" with any objects in the image. To evaluate the performance gap between explicit and implicit text, we conducted experiments using existing methods on both types of text separately. Preliminary results indicate that the performance on implicit text is approximately  $42\\%$  lower compared to explicit text.",
+        "bbox": [
+            110,
+            84,
+            492,
+            343
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Besides, the implicit text also plays a significant role in the domain of human-computer interaction. However, existing referring expression datasets, such as RefCOCO (Yu et al., 2016), RefCOCO+ (Yu et al., 2016), RefCOCOg (Mao et al., 2016), and CLEVR-Ref+ (Liu et al., 2019), primarily consist of explicit text. Unfortunately, there is currently no dataset available that is specifically designed to provide research data on implicit text.",
+        "bbox": [
+            112,
+            346,
+            489,
+            491
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To address the lack of research data and methods for implicit text, we propose a synthetic dataset called CLEVR-Implicit, specifically for implicit reasoning in REC. Unlike existing datasets, the implicit text we constructed is a subset of all types of implicit text, since they describe synthetic images, and do not need to consider external knowledge. Specifically, We focus on the more difficult types (only containing relationships between objects without any description of their appearance attributes). We evaluate the current pre-trained REC models on our CLEVR-Implicit dataset with these types of implicit text.",
+        "bbox": [
+            110,
+            495,
+            489,
+            705
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In real-world scenarios, most humans will first extract explicit information from visual context and text information when processing these types of implicit text, and then reason the target object from the extracted explicit information. Inspired by the thinking mode of humans, we simulate the process of humans reasoning about the implicit text and introduce a method called TIE (Transform the Implicit text into Explicit text), which enables pre-trained models to process the implicit text by extracting the explicit information. By employing TIE, REC models facilitate alignment between explicit words and corresponding images on CLEVR-",
+        "bbox": [
+            112,
+            709,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "table",
+        "img_path": "images/cc14284366a9c657f6ce5eef35833ee25b3499a2480333e8a76ec03c727b5134.jpg",
+        "table_caption": [
+            "Table 1: The attributes and example values of the objects in CLEVR. We build the CLEVR-Implicit based on these objects."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Attribute</td><td>Example values</td></tr><tr><td>color</td><td>red,blue,cyan,purple,gray,green</td></tr><tr><td>material</td><td>metal,rubber</td></tr><tr><td>shape</td><td>cube,cylinder,sphere</td></tr><tr><td>size</td><td>large,small</td></tr><tr><td>spatial</td><td>left,right,behind,front</td></tr></table>",
+        "bbox": [
+            519,
+            135,
+            870,
+            237
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Implicit. Experimental results demonstrate a substantial difference in the performance of REC models on our dataset before and after applying TIE.",
+        "bbox": [
+            507,
+            259,
+            884,
+            307
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We summarize our contributions as follows:",
+        "bbox": [
+            527,
+            307,
+            855,
+            322
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We analyze the limitations of existing referring expression comprehension models on the implicit text and construct the CLEVR-Implicit dataset.",
+            "- We propose the TIE method to enable pretrained REC models to process the implicit text by transforming the implicit text to explicit text.",
+            "- We evaluate the existing models on the CLEVR-Implicit and analyze the differences between the models before and after applying TIE."
+        ],
+        "bbox": [
+            527,
+            332,
+            885,
+            544
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            509,
+            567,
+            672,
+            582
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Referring Expression Comprehension",
+        "text_level": 1,
+        "bbox": [
+            507,
+            592,
+            852,
+            608
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Referring expression comprehension (REC) aims to localize the region in an image that corresponds to a given sentence description. In recent years, there has been a significant increase in research attention towards this task, driven by the increasing applications of REC in the domain of human-computer interaction. The related task of referring expression generation (REG) has been studied for many years (Krahmer and Van Deemter, 2012; Mitchell et al., 2013; Reiter and Dale, 1992) in the natural language processing (NLP) domain.",
+        "bbox": [
+            505,
+            613,
+            884,
+            789
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Several real-world datasets, such as RefCOCO, RefCOCO+, and RefCOCOg, have been proposed based on the MSCOCO dataset (Vinyals et al., 2016) for this task. CLEVR-Ref+ (Liu et al., 2019) is a synthetic dataset that contains geometric objects. These datasets have been widely adopted as benchmarks for evaluating the performance of referring expression models. MMI (Mao et al., 2016)",
+        "bbox": [
+            507,
+            790,
+            885,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12821",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "directly predicted the bounding box by ranking the region proposals. SLR (Yu et al., 2017) utilized a joint embedding model to predict the referred object by learning the visual and language representations in an embedding space. MattNet (Yu et al., 2018) adapted the attention mechanism to parse expressions into three modules and predict the visual regions by a weighted combination of module scores. VL-BERT (Su et al., 2019) adopted a Transformer model and concatenates the visual and language embeddings as inputs. MDETR (Kamath et al., 2021) constructed an end-to-end system based on the DETR (Carion et al., 2020) detector and a language model to learn the alignment between objects and phrases. VLTVG (Yang et al., 2022) improves comprehension ability by utilizing a visual-linguistic verification module to generate discriminative feature representations.",
+        "bbox": [
+            112,
+            84,
+            492,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Prompt Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            387,
+            299,
+            403
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Prompt learning, initially introduced in the field of NLP, is used to reformulate downstream tasks to the form of the pre-training task masked language modeling (MLM) (Devlin et al., 2018). This approach enables models to effectively utilize the knowledge during the pre-training stage (Liu et al., 2023). Concretely, prompt engineering reformulates the downstream task into a \"fill-in-the-blank\" cloze task with the same objective function as in the MLM task. The effectiveness of prompt learning heavily relies on the design of appropriate prompt templates. Even slight variations in the prompt template can yield significant differences in performance. Therefore, the selection of suitable prompt templates becomes crucial for achieving desirable results.",
+        "bbox": [
+            112,
+            409,
+            489,
+            665
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Prompt learning has also been explored in the multi-modal domain through various studies. CLIP (Radford et al., 2021) pre-trained on large-scale image-text pairs and designs manual prompts like \"The photo of [MASK]\" to facilitate image classification by predicting the masked token. DPT (Liu et al., 2022) reformulated the visual question answering (VQA) task as the MLM by converting the input (question) to a declaration with the masked token.",
+        "bbox": [
+            112,
+            668,
+            489,
+            828
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Methodology",
+        "text_level": 1,
+        "bbox": [
+            112,
+            843,
+            263,
+            860
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This section begins by explaining the construction process of the CLEVR-Implicit dataset. Subsequently, we introduce our proposed TIE method,",
+        "bbox": [
+            112,
+            870,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/5244bd8759603dbf5cd6cdc91be4616fcc29835655807db34725a19b3618e72c.jpg",
+        "table_caption": [
+            "Table 2: The statistic data of CLEVR-Implicit."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Attribute num</td><td>Same</td><td>Different</td></tr><tr><td>1</td><td>3000</td><td>-</td></tr><tr><td>2</td><td>3000</td><td>1553</td></tr><tr><td>3</td><td>3000</td><td>3000</td></tr><tr><td>4</td><td>2905</td><td>3000</td></tr></table>",
+        "bbox": [
+            547,
+            105,
+            840,
+            193
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "which involves converting the implicit text within the CLEVR-Implicit dataset into explicit text. This explicit text is then utilized in the referring expression comprehension task through the fine-tuning process of pre-trained vision-language models.",
+        "bbox": [
+            507,
+            217,
+            884,
+            299
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            507,
+            312,
+            722,
+            326
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.1 CLEVR-Implicit Overview",
+        "text_level": 1,
+        "bbox": [
+            507,
+            334,
+            781,
+            349
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "CLEVR-Implicit Dataset uses a subset of the CLEVR (Johnson et al., 2017) scenes (10K images for all,  $60\\%$  for the train set and  $40\\%$  for the validation and test set). Table 1 provides an overview of the attributes and values associated with the objects in CLEVR. In order to enhance the difficulty of the dataset, we filter and choose the images that contain 3 or more objects. Given that CLEVR-Ref+ provides scene graphs about the attributes (such as color, material, shape, and size) of objects, as well as the spatial relationships between them within the image, we generate corresponding referring expressions for each image. A total of 19K referring expressions are generated, averaging 1.9 referring expressions for each image. Finally, the bounding box of the referred object is assigned as the label for each referring expression in the CLEVR-Implicit Dataset.",
+        "bbox": [
+            507,
+            354,
+            884,
+            644
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1.2 Implicit Text Generation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            656,
+            764,
+            671
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To generate implicit text that describes objects through a comparison between two or more objects, we utilize the scene graphs provided by CLEVR-Ref+. A program is employed to extract the attributes of all the objects present in the images. This allows for the comparison and analysis of their attributes, enabling the generation of implicit text that captures the relationships between objects without explicitly mentioning any specific object in the image.",
+        "bbox": [
+            507,
+            676,
+            884,
+            837
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For an image, we compare the attributes of each two objects and find each set of objects that satisfies the requirement for generating implicit text. In CLEVR-Implicit, we set up two types of implicit text: \"same\" and \"different\". The \"same\"",
+        "bbox": [
+            507,
+            839,
+            885,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12822",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/6d0e1fdff11b465a75365b4c58d62a849d447d1cf7802f27c0303a46d47f69d3.jpg",
+        "image_caption": [
+            "Figure 2: The framework of our TIE method and subsequent REC process. The implicit text in CLEVR-Implicit is converted to an implicit prompt which contains a masked token before the attribute. Then vision-language model predicts the explicit words by Multi-step and Path-select prediction strategies in the MLM task. After TIE, the output explicit text and the image are fed to referring expression model to predict the result."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            80,
+            884,
+            296
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "type represents that the text refers to the object by describing the same of certain attributes between objects (e.g. \"The left one of the only two objects with the same color and material\"), while the \"different\" type is the opposite of \"same\" (e.g., \"The left object that is different shape and size from the other one). Table 2 provides statistics regarding the distribution of \"same\" and \"different\" type samples in the CLEVR-Implicit dataset.",
+        "bbox": [
+            112,
+            388,
+            487,
+            533
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "It is worth mentioning that the \"same\" type of referring expressions also exist in the CLEVR-Ref+ dataset. The key difference in \"same\" type text between CLEVR-Implicit and CLEVR-Ref+ is whether the explicit words are hidden. The \"same\" type of text in CLEVR-Ref+ is like \"The objects that have the same size as the red sphere\", which contains the explicit word \"red sphere\". The explicit context provides REC models with a reference object and helps them locate the target. Oppositely, the \"same\" type of implicit text in CLEVR-Implicit is formatted like \"The two objects with the same size\", which hides explicit words. Without explicit context and hard to find the reference object, REC models are forced to compare every two objects to identify what the \"size\" is and whether they are the same. That illustrates that implicit text in CLEVR-Implicit is more challenging.",
+        "bbox": [
+            112,
+            541,
+            489,
+            832
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We pay more attention to the reference to a single object in the context of REC. During the generation process, we take precautions to avoid text ambiguity, ensuring that the implicit text describes only one object in the image. To achieve this, we",
+        "bbox": [
+            112,
+            839,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "prioritize maximizing the number of attributes involved in attribute comparisons. For example, if the target is in a group of objects that share the same color only, we must ensure no other group of objects that can satisfy the same color and material. In cases where such a group of objects exists, we will change the target and modify the implicit text to describe the corresponding number of attributes until no additional attributes can be included. More details of the implicit text generation are shown in Appendix A.",
+        "bbox": [
+            507,
+            388,
+            884,
+            565
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The implicit text in CLEVR-Implicit follows a standardized format based on several templates and rules. Especially, it is common for the same sentence to describe different target objects across different images. By employing homogeneous text, we purposefully direct the model's focus toward the image itself, thereby effectively testing the model's reasoning ability.",
+        "bbox": [
+            507,
+            567,
+            882,
+            696
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 TIE Method",
+        "text_level": 1,
+        "bbox": [
+            507,
+            712,
+            653,
+            726
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2.1 TIE Overview",
+        "text_level": 1,
+        "bbox": [
+            507,
+            736,
+            680,
+            750
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To address the challenge of implicit text comprehension, we propose a method called Transform Implicit text to Explicit text (TIE). The objective of TIE is to incorporate explicit words into the original implicit text, enabling the model to reason more effectively. Our approach consists of two key modules: (1) The prompt design module builds prompts for implicit text by adding masked tokens. By strategically designing the prompts, we prepare the groundwork for the next module. (2)",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12823",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The cloze procedure module uses a pre-trained vision-language model to predict the explicit words with the implicit prompts.",
+        "bbox": [
+            112,
+            84,
+            487,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By leveraging these two modules, TIE enables the integration of explicit words into the implicit text and facilitates indirect reasoning on the implicit text effectively.",
+        "bbox": [
+            112,
+            133,
+            489,
+            198
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.2 Prompt Design Module",
+        "text_level": 1,
+        "bbox": [
+            112,
+            208,
+            357,
+            223
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The prompt design module converts the implicit text into a format suitable for the Masked Language Modeling (MLM) task. One approach to adding explicit words is by obtaining the specific attribute values involved in the comparison of attributes in the implicit text (e.g., \"What is the color in 'The left one of the only two objects with the same color'?\"). Consequently, the prompt design module generates the prompt like \"The left one of the only two objects with the same [MASK] color\" for an implicit text sample. It's important to note that the conversion process differs between the \"same\" and \"different\" types of implicit text. In the \"same\" type, only one identical attribute value needs to be predicted, while in the \"different\" type, two different attribute values must be predicted. Therefore, in the \"different\" type, two masked tokens are added before each attribute in the implicit text. This process of converting the implicit text can be seen as a translation task, where the source text is the original implicit text, and the target text is the input format required for the MLM task. To accomplish this conversion process, we create an implicit prompt dataset using the implicit text from CLEVR-Implicit as the source text, and the target text is a text with the masked token preceding each attribute in the implicit text. Subsequently, we train an encoder-decoder model (T5, Raffel et al., 2020) on the implicit prompt dataset to add the appropriate number of masked tokens corresponding to the attributes for each sample in CLEVR-Implicit.",
+        "bbox": [
+            112,
+            227,
+            489,
+            728
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2.3 Cloze Procedure Module",
+        "text_level": 1,
+        "bbox": [
+            112,
+            738,
+            369,
+            752
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The training objective of the cloze procedure module aligns with the MLM task, a commonly employed pre-training task in various pre-trained models, so another cloze dataset is devised for the MLM training stage. In this dataset, the input consists of implicit text containing masked tokens and an image description, while the target comprises explicit text containing the ground-truth attribute values. In this module, we fine-tune a pre-trained vision-language model ViLT to predict a specific word",
+        "bbox": [
+            112,
+            758,
+            490,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "within the masked tokens. And we will introduce three prediction strategies: Base, Multi-step, and Path-select.",
+        "bbox": [
+            507,
+            84,
+            882,
+            131
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Base In the prediction stage, the prediction of multiple attribute values in a sample is different from the prediction of a single attribute value. For implicit text containing multiple attributes, a straightforward prediction strategy is to predict all attribute values. However, this strategy tends to decrease the correlation among the predicted values, resulting in potential discrepancies where multiple predicted values may not refer to the same group of objects. Consequently, incorrect prediction results may arise.",
+        "bbox": [
+            507,
+            136,
+            884,
+            313
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Multi-step To address the aforementioned issue, we propose an enhanced strategy named multi-step prediction. Specifically, when predicting a sample with multiple attributes, the model outputs only one attribute value at a time. Subsequently, it incorporates the previous prediction as input to predict the next attribute value. Compared with the method of predicting all attribute values at one time, this strategy enhances the correlation among multiple attribute values. For instance, if the original text is \"The left one of the only two objects with the same [MASK] color and [MASK] material,\" the corresponding first-step output is adjusted to \"The left one of the only two objects with the same red color and [MASK] material.\" And the second-step output is \"The left one of the only two objects with the same red color and metal material.\"",
+        "bbox": [
+            507,
+            316,
+            884,
+            589
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Path-select We find that the Multi-step strategy has a potential challenge where the accuracy of predictions relies on the first attribute. If the first attribute prediction is incorrect, it might trigger a chain reaction leading to subsequent attributes being errors. To mitigate this issue, we search for some objects that are not the targets as negative samples. They are characterized by partially satisfying the constraints of implicit text. Part of the attributes that do not satisfy are marked [NONE]. For instance, for a target sample \"The same red color and rubber material\" in cloze dataset, the corresponding negative sample might be \"The same blue color and [NONE] material\".",
+        "bbox": [
+            507,
+            594,
+            882,
+            818
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "During the prediction stage, when the model obtains the optimal Softmax result for an attribute and predicts [NONE] for the next attribute, it recognizes this as an incorrect predicted path. Subsequently, the model reverts to the prediction of the preceding attribute, considering the suboptimal",
+        "bbox": [
+            507,
+            822,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12824",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/c2e0e595e6cec1de375b30c049c1b98e552b22a0765b620c5c5931573e9ffd20.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">ViLT (Kim et al., 2021) MLM</td><td colspan=\"2\">Same</td><td colspan=\"2\">Different</td><td colspan=\"2\">Total</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>Base</td><td>80.54</td><td>81.97</td><td>25.73</td><td>23.35</td><td>61.99</td><td>60.48</td></tr><tr><td>Multi-step</td><td>92.55</td><td>91.42</td><td>87.05</td><td>85.47</td><td>91.45</td><td>91.50</td></tr><tr><td>Path-select</td><td>91.76</td><td>91.35</td><td>85.58</td><td>87.50</td><td>91.48</td><td>90.37</td></tr></table>",
+        "bbox": [
+            216,
+            80,
+            781,
+            167
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/7120df46020b4bf494e74639ab9d3453049898b72b5a8072d2f73d51cc3a161d.jpg",
+        "table_caption": [
+            "Table 3: The TIE experiment results on CLEVR-Implicit. We use three different prediction strategies to compare the performance of ViLT."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Before TIE</td><td colspan=\"2\">After TIE</td><td colspan=\"2\">Ground-truth</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>MDETR (Kamath et al., 2021)</td><td>53.57</td><td>53.89</td><td>91.51</td><td>92.19</td><td>96.01</td><td>95.66</td></tr><tr><td>VLTVG (Yang et al., 2022)</td><td>58.83</td><td>58.46</td><td>84.48</td><td>86.15</td><td>95.97</td><td>96.85</td></tr></table>",
+        "bbox": [
+            188,
+            216,
+            808,
+            287
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Referring Expression Comprehension results performed by MDETR and VLTVG. Before TIE means that the models are trained on the implicit text before the input transformation of the TIE method, and After TIE means that the models are trained on the explicit text after TIE.",
+        "bbox": [
+            112,
+            296,
+            882,
+            338
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Softmax result, which guides the model toward the correct prediction.",
+        "bbox": [
+            112,
+            365,
+            485,
+            395
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            407,
+            260,
+            424
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.1 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            432,
+            329,
+            448
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For the REC task on the CLEVR-Implicit dataset, we employ two models: MDETR (Kamath et al., 2021) and VLTVG (Yang et al., 2022). MDETR utilizes weights pre-trained on various datasets such as COCO (RefCOCO, RefCOCO+, RefCOCOg), Visual Genome (Krishna et al., 2017), and Flickr30k (Plummer et al., 2015). and VLTVG uses weights pre-trained on RefCOCOg. And we fine-tune them on the CLEVR-Implicit dataset. During the training stage, we use the settings of learning rate 1e-4, learning rate of backbone 1e-4, batch size 16 and weight decay 1e-5 to train for 60 epochs. In the input transformation experiment, we utilize the T5-small model for the prompt design module. As for the cloze procedure module, we select ViLT (Kim et al., 2021). To ensure a fair comparison, all different prediction methods are configured with the AdamW optimizer, a learning rate of 1e-4, a batch size of 32, and an image size of  $384 \\times 384$  to train for 70 epochs.",
+        "bbox": [
+            112,
+            453,
+            489,
+            776
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Results of TIE",
+        "text_level": 1,
+        "bbox": [
+            112,
+            785,
+            273,
+            800
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 3 presents the results of the TIE experiment. The evaluation method is to assess whether the predicted explicit words match the corresponding ground-truth values. For samples containing multiple attributes, it is necessary to ensure the accurate prediction of all explicit words. otherwise, it is considered incorrect. In the case of \"different\" type",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "samples, where attribute values differ between objects, the order of the two predicted values for a single attribute can be reversed.",
+        "bbox": [
+            507,
+            365,
+            882,
+            412
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We present the results separately for training on the \"same\" and \"different\" types of samples, and the \"Total\" column represents the combined performance on both types of implicit text. In the Base strategy, we observe that \"different\" samples are more challenging to understand compared to \"same\" samples, resulting in a decrease in performance of up to  $54.81\\%$ . For \"different\" samples, the model needs to output two values to predict a single attribute, resulting in twice as many explicit words to be predicted compared to \"same\" samples. In contrast, the Multi-step strategy exhibits significant improvements for both types of text. It shows a  $12.01\\%$  increase in performance for \"same\" samples and a remarkable  $61.32\\%$  increase for \"different\" samples. A reasonable explanation is that the prediction of each explicit word takes into account all previously predicted results. Multi-step strategy mitigates cases where the initial predictions are correct, but subsequent predictions suffer from errors, leading to more accurate overall predictions.",
+        "bbox": [
+            507,
+            417,
+            884,
+            755
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From the comparing results of the Path-select and Multi-step strategies, we observe some differences when incorporating negative samples into the training set. The Path-select results on the validation set exhibit a slight decrease of  $0.79\\%$  and  $1.47\\%$  compared to Multi-step. We speculate that this is because the majority of incorrect positions in wrongly predicted samples tend to occur in the second attribute or later, and the prediction of the first attribute in these samples is generally accurate.",
+        "bbox": [
+            507,
+            758,
+            884,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12825",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2db1b42fe41556b45219dc3e907eabb7e8a3688e217c282c58a436445ad816fd.jpg",
+        "image_caption": [
+            "(a) same attributes"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            157,
+            105,
+            463,
+            268
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/224046e4ef00d24df1e3c70f69b53067799e18f96600cad47dd810288363c7ca.jpg",
+        "image_caption": [
+            "(b) different attributes",
+            "Figure 3: 3(a) is the performance of three prediction strategies on \"same\" type samples in test set, and 3(b) is the performance of three prediction strategies on \"different\" type samples in test set."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            105,
+            821,
+            269
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "However, on the test set of \"different\" samples, Path-select surpasses Multi-step by  $2.03\\%$  in terms of performance. Particularly, the performance of Multi-step and Path-select in the \"Total\" column is comparable, indicating that the addition of negative samples can enhance the model's robustness. And negative samples enable the model to choose the correct path when the first attribute prediction is incorrect.",
+        "bbox": [
+            112,
+            359,
+            487,
+            502
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.3 Results of Referring Expression",
+        "text_level": 1,
+        "bbox": [
+            112,
+            514,
+            410,
+            530
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4 presents the results of the referring expression experiment. We evaluate the performance of MDETR and VLTVG, both before and after applying the TIE based on the Multi-step prediction strategy. Additionally, we provide the performance of the models trained on ground-truth explicit samples as a reference. The evaluation metric is Intersection over Union (IoU), with a threshold value of 0.5. We observe that prior to TIE, the performance of both models trained on implicit text is relatively poor, indicating limited reasoning ability when dealing with implicit text. However, after applying TIE, there is a significant improvement in performance for both models (53.57% vs. 91.51% for MDETR and 58.83% vs. 84.48% for VLTVG). However, compared with Ground-truth, there is a certain gap in performance, because it is limited by the impact of TIE performance.",
+        "bbox": [
+            112,
+            533,
+            489,
+            825
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "4.4 Results of Contrast Experiment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            834,
+            410,
+            850
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Difficulty of CLEVR-Implicit In order to contextualize the difficulty of CLEVR-Implicit, we evaluate the performance of REC models on CLEVR-Implicit when the models are trained on CLEVR-",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/422ccd457d5b22c3c3b1b578a731052890aa92174c722c4b7eb70a5f0bea87cb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td colspan=\"2\">Train on CLEVR-Ref+</td></tr><tr><td></td><td>val</td><td>test</td></tr><tr><td>MDETR</td><td>47.02</td><td>45.85</td></tr><tr><td>VLTVG</td><td>47.99</td><td>47.11</td></tr></table>",
+        "bbox": [
+            557,
+            356,
+            835,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ced57f14b829c2acb969572b6442fda52a91f29b513a4eac97c02cc1ed7e70f.jpg",
+        "table_caption": [
+            "Table 5: The performances of REC models which are trained on CLEVR-Ref+ and tested on CLEVR-Implicit."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>VLTVG on CLEVR-Ref+</td><td>val</td><td>test</td></tr><tr><td>Before TIE</td><td>66.43</td><td>65.22</td></tr><tr><td>After TIE</td><td>68.05</td><td>66.46</td></tr></table>",
+        "bbox": [
+            527,
+            494,
+            865,
+            546
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6: The performances of VLTVG respectively on original CLEVR-Ref+ (Before TIE) and CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit.",
+        "bbox": [
+            507,
+            556,
+            882,
+            601
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ref+. Table 5 shows that the results of MDETR and VLTVG models are all lower than \"Before TIE\" Column results in Table4. That indicates that the experience that the models learn from CLEVR-Ref+ cannot help them process the implicit text, because the samples in CLEVR-Ref+ contain explicit information to help the model locate the target, while the samples in CLEVR-Implicit hide the information. Therefore, we believe that it's challenging to process the implicit text in our CLEVR-Implicit dataset.",
+        "bbox": [
+            505,
+            627,
+            882,
+            803
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Affect of TIE on explicit text Since the TIE method converts implicit text into explicit text, in addition to implicit text, we also need to analyze whether the explicit text after applying TIE will affect the performance of REC models. To verify this problem, we evaluate the performance of VLTVG respectively on original CLEVR-Ref+ and",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12826",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/4542ea89e2f3d33f5e3496e8ed38951ea0d330ac3f1ed3ca770175e98686f4b3.jpg",
+        "image_caption": [
+            "(a) An example of the comparison from Base, Multi-step and Path-select prediction strategies"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            80,
+            295,
+            167
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "O: The left one of the only two objects with the same shape, color and size in the image",
+        "bbox": [
+            305,
+            86,
+            549,
+            105
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "P: The left one of the only two objects with the same [MASK] shape, [MASK] color and [MASK] size in the image",
+        "bbox": [
+            305,
+            114,
+            549,
+            142
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "GT: cylinder, yellow, large",
+        "bbox": [
+            305,
+            149,
+            438,
+            161
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Base: sphere, brown, large",
+        "bbox": [
+            594,
+            91,
+            734,
+            102
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-step: cylinder, yellow, large",
+        "bbox": [
+            594,
+            118,
+            769,
+            129
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Path-select: cylinder, yellow, large",
+        "bbox": [
+            594,
+            145,
+            771,
+            156
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/3fcf88c1bbcfbfbf23cca254074f611ca8b029de58cead01d3969871479acb9e.jpg",
+        "image_caption": [
+            "(b) REC before TIE",
+            "Before: The left one of the only two objects with the same shape, color and size in the image",
+            "Figure 4: Examples from TIE and referring expression comprehension experiments. (a) an example of the comparison from Base, Multi-step and Path-select prediction strategies. O, P and GT denote original implicit text, generated prompt text and ground-truth explicit words. (b) Visualization of the predicted bounding box from VLTVG before TIE and (c) after TIE. green and red rectangles denote ground-truth and predicted bounding box."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            200,
+            295,
+            286
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/19825c8cb9209c6bbc9cd197168e16b2754dfc0c3c097586893372296c1a22ec.jpg",
+        "image_caption": [
+            "(c) REC after TIE",
+            "After: The left one of the only two objects with the same cylinder shape, yellow color and large size in the image"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            497,
+            200,
+            675,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit. Table 6 shows that TIE does not affect performance (even better on VLTVG) on other types of referring expressions besides the implicit text, and TIE is robust in a broader class of referring expressions.",
+        "bbox": [
+            112,
+            400,
+            485,
+            495
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Contrast between explicit and implicit As shown in Table 7, we evaluate the performance of MDETR and VLTVG models separately on predicted explicit text and implicit text when the models are trained on implicit text and ground-truth explicit text respectively.",
+        "bbox": [
+            112,
+            497,
+            487,
+            593
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "When both validation sets provide predicted explicit words (Column 1 vs Column 3), the performance of the model trained on the implicit training set is significantly decreased (MDETR:  $31.48\\%$  vs  $91.98\\%$ , VLTVG:  $53.85\\%$  vs  $85.90\\%$ ) compared with that on the explicit training set, which is evidence that explicit words play a crucial role in the models' comprehension and reasoning processes. The same situation occurs when exchanging implicit and explicit states on train and validation sets (Column 2 vs Column 3).",
+        "bbox": [
+            112,
+            595,
+            489,
+            771
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.5 Performance on different number of attributes",
+        "text_level": 1,
+        "bbox": [
+            112,
+            784,
+            445,
+            815
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 illustrates the performance of ViLT on \"same\" and \"different\" samples with different numbers of attributes. In Figure 3(a), we observe that the model achieves the highest performance when predicting one-attribute samples, while the lowest performance is observed when predicting three-",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/09255481e0d05e8109e9e91b3f9bcc6138cb427ed8628fde1b01695b4d909b7e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Train-i Val-pe</td><td>Train-ge Val-i</td><td>Train-ge Val-pe</td></tr><tr><td>MDETR</td><td>31.48</td><td>46.25</td><td>91.98</td></tr><tr><td>VLTVG</td><td>53.85</td><td>48.34</td><td>85.90</td></tr></table>",
+        "bbox": [
+            532,
+            397,
+            858,
+            467
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 7: The contrast experiment for REC. The i suffix denotes the set with implicit samples. The pe suffix denotes that the explicit words in samples are predicted by TIE. The ge denotes the explicit words in samples is ground-truth.",
+        "bbox": [
+            507,
+            476,
+            882,
+            546
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "attribute samples. Interestingly, the performance on four-attribute samples is second only to that of one-attribute samples. A reasonable conjecture is that a set of objects is identical when all four properties are the same in CLEVR images, so the model only needs to find two objects that are identical. Figure 3(b) shows that predicting four-attribute \"different\" samples has the most significant challenge on the Base strategy. However, on the Multi-step strategy, the performance on four-attribute \"different\" samples surpasses that of three-attribute samples. This indicates that the Multi-step prediction strategy effectively mitigates the impact of an increased number of predicted words, resulting in improved performance.",
+        "bbox": [
+            507,
+            575,
+            884,
+            816
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4.6 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            832,
+            643,
+            847
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 4 indicates an example of the comparison of different prediction strategies in the TIE experiment, and we visualize the ground-truth bounding box and the bounding box predicted from VLTVG",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "12827",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 8: Examples about the application of implicit text in daily life.",
+        "bbox": [
+            112,
+            82,
+            485,
+            110
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Real-world Implicit Text Examples",
+        "text_level": 1,
+        "bbox": [
+            134,
+            123,
+            413,
+            137
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Two persons are doing the same movement.",
+        "bbox": [
+            134,
+            140,
+            460,
+            154
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "His height is different from each other.",
+        "bbox": [
+            136,
+            156,
+            423,
+            170
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "His pants and shoes are the same color.",
+        "bbox": [
+            136,
+            172,
+            425,
+            187
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The same brand of cosmetics on the table.",
+        "bbox": [
+            136,
+            187,
+            448,
+            202
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Take the drugs used for the same symptom.",
+        "bbox": [
+            136,
+            204,
+            455,
+            219
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "before and after TIE in the REC experiment. In Figure 4(a), with the attributes of the example \"shape\" and \"color\", the Base strategy mistakenly predicts \"sphere\" and \"brown\". We observe that from the image there is a group of objects that is both \"sphere\" and another group of objects that is both \"brown\", but there are no such two objects that satisfy \"sphere\" and \"brown\" simultaneously. This indicates that the Base strategy performs poorly on multi-attribute samples due to the weak correlation between attributes, leading to the error that different attributes refer to different groups of objects. In contrast, the Multi-step and Path-select strategies enhance the correlation between attributes and make predictions step-by-step, leading to more accurate predictions.",
+        "bbox": [
+            112,
+            247,
+            487,
+            504
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In Figure 4(b), by taking the implicit text as input, VLTVG incorrectly predicts the referred object as a brown cube. While VLTVG predicts the referred object accurately when taking the explicit text after TIE as input. The comparison result indicates that current models are being confused when performing implicit reasoning based on the logical relation conveyed by the implicit text. Only when explicit words are explicitly included, the models can align the explicit words with the corresponding objects in the image for explicit reasoning.",
+        "bbox": [
+            112,
+            506,
+            489,
+            683
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5 Implicit Text in Real-world Scenarios",
+        "text_level": 1,
+        "bbox": [
+            112,
+            697,
+            470,
+            714
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The implicit descriptions in real-world scenarios are more complex and difficult because they not only describe the targets implicitly through the geometrical attributes of color, material, etc., but also contain more attributes as well as external factors. Table 8 shows the implicit text examples in real-world daily life. It is important that even though our generated implicit text relies on the CLEVR images and the provided annotations, our proposed TIE method can still extend the scope beyond the CLEVR domain to real-world scenarios. For example, based on Figure 5 which is chosen from Visual",
+        "bbox": [
+            112,
+            726,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/fdb3d565039b1a5d940f5783edb3568056d3ebf3eae3adb2b64ad43b803ead53.jpg",
+        "image_caption": [
+            "Original Implicit Text: The left one of two persons with the same posture and gender",
+            "Figure 5: A real-world image from Visual Genome"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            536,
+            80,
+            867,
+            258
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "After Prompt Design Module: The left one of two persons with the same [MASK] posture and [MASK] gender",
+        "bbox": [
+            515,
+            296,
+            884,
+            319
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "After Cloze Procedure Module: The left one of two persons with the same walking posture and woman gender",
+        "bbox": [
+            515,
+            328,
+            884,
+            351
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Genome, consider the implicit text like \"The left one of two persons with the same posture and gender\". Through the TIE method, the prompt can be built before the \"posture\" and \"gender\" by prompt design module, like \"The left one of two persons with the same [MASK] posture and [MASK] gender\". The prompt guides the cloze procedure module to predict the explicit words \"walking\" and \"woman\" like \"The left one of two persons with the same walking posture and woman gender\". By integrating TIE approach, current models can also process the implicit text on the real-world scenarios.",
+        "bbox": [
+            505,
+            409,
+            882,
+            615
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            633,
+            640,
+            648
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we analyze the significance of implicit reasoning, along with the lack of investigation in this domain. To address this, we introduce the CLEVR-Implicit dataset for implicit reasoning in the REC task. This dataset utilizes synthesized images to generate two types of implicit text: \"same\" and \"different.\" Evaluating the performance of existing pre-trained models on CLEVR-Implicit, we observe model confusion when attempting to align implicit text with images. To overcome this challenge, we propose the Transform Implicit text to Explicit text (TIE) method, which converts implicit text into explicit text. Our experiments demonstrate that the reasoning capability of the models improves significantly through the alignment of explicit words and images.",
+        "bbox": [
+            505,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "12828",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "The limitation of our work is that CLEVR-Implicit is built on synthetic images, so it's straightforward to generate corresponding implicit text based on the annotation of the synthetic images. Besides, since the synthetic images contain only geometric objects, our implicit text can only be generated within the scope of the object's attributes. Therefore, in future work, we will annotate more types of implicit text based on real-world scenarios. This will allow us to explore the mechanisms of model understanding of implicit text.",
+        "bbox": [
+            112,
+            108,
+            490,
+            287
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            296,
+            285,
+            313
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the National Natural Science Foundation of China (62076100), Fundamental Research Funds for the Central Universities, SCUT (x2rjD2230080), the Science and Technology Planning Project of Guangdong Province (2020B0101100002), CAAI-Huawei MindSpore Open Fund, CCF-Zhipu AI Large Model Fund.",
+        "bbox": [
+            112,
+            319,
+            489,
+            434
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            458,
+            213,
+            474
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.",
+            "Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910.",
+            "Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. 2021. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790.",
+            "Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR.",
+            "Emiel Krahmer and Kees Van Deemter. 2012. Computational generation of referring expressions: A survey. Computational Linguistics, 38(1):173-218."
+        ],
+        "bbox": [
+            115,
+            481,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73.",
+            "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35.",
+            "Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. 2019. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194.",
+            "Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. 2022. Declaration-based prompt tuning for visual question answering. arXiv preprint arXiv:2205.02456.",
+            "Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32.",
+            "Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. 2016. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20.",
+            "Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. 2013. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184.",
+            "Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649.",
+            "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.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67."
+        ],
+        "bbox": [
+            510,
+            85,
+            885,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "12829",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "ref_text",
+        "text": "Ehud Reiter and Robert Dale. 1992. A fast algorithm for the generation of referring expressions. In Proceedings of the 14th Conference on Computational Linguistics - Volume 1, page 232-238. Association for Computational Linguistics.",
+        "bbox": [
+            115,
+            85,
+            485,
+            151
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. 2019. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530.",
+        "bbox": [
+            115,
+            161,
+            485,
+            214
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2016. Show and tell: Lessons learned from the 2015 mscoco image captioning challenge. IEEE transactions on pattern analysis and machine intelligence, 39(4):652-663.",
+        "bbox": [
+            115,
+            224,
+            485,
+            290
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. 2022. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499-9508.",
+        "bbox": [
+            115,
+            300,
+            485,
+            378
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2021. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797.",
+        "bbox": [
+            115,
+            388,
+            485,
+            442
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. 2018. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315.",
+        "bbox": [
+            115,
+            451,
+            485,
+            531
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. 2016. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer.",
+        "bbox": [
+            115,
+            541,
+            485,
+            595
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "ref_text",
+        "text": "Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. 2017. A joint speaker-listener-reinforcer model for referring expressions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7282-7290.",
+        "bbox": [
+            115,
+            605,
+            485,
+            670
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A More Details of Implicit Text Generation",
+        "text_level": 1,
+        "bbox": [
+            115,
+            684,
+            401,
+            715
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "To clarify on our implicit referring expression generation process, we'd like to elaborate on the method used to create the CLEVR-Implicit dataset from the existing CLEVR images and their associated annotations. These annotations contain object attributes, as well as spatial relationships between objects. For each image, we will traverse all the objects in the image by comparing whether the attribute values between two objects are the same/different each time. For example, if an image contains 4 objects, this leads to  $\\mathrm{C}(4,2) = 6$  object pair comparisons (e.g., (1,2), (1,3), (1,4), (2,3),",
+        "bbox": [
+            115,
+            726,
+            485,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "(2,4), (3,4), the numbers are object ids). We will generate the corresponding implicit text if a certain pair satisfies the \"same\" or \"different\" condition. Suppose the (2,3) combination objects' shape and color are both cube and blue, and guarantee that no other pairs satisfy the condition \"same shape and color\" to avoid the ambiguity in the REC task, then we will generate the \"same\" type of implicit text as \"The two objects with the same shape and color\". Due to the only one target of the REC task, we must choose one of the (2,3) objects to be the target. We employ a random selection for the target based on the spatial relationship between 2 and 3. If object 2 is left of object 3 and object 2 is selected, the finally generated implicit text is \"The left one of the two objects with the same shape and color\". The \"different\" types of text are generated in the same way. This consistent pattern of implicit text generation in the CLEVR-Implicit dataset ensures both the concealment of explicit information and the reference of target objects in the images.",
+        "bbox": [
+            510,
+            84,
+            880,
+            420
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "12830",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    }
+]

2023/CLEVR-Implicit_ A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension/7a840248-0cf8-41af-aee2-5d7039744b4d_model.json

@@ -0,0 +1,2114 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.164,
+                0.089,
+                0.836,
+                0.131
+            ],
+            "angle": 0,
+            "content": "CLEVR-Implicit: A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.334,
+                0.142,
+                0.674,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Jingwei Zhang\\(^{1,\\ast}\\), Xin Wu\\(^{1,\\ast}\\), Yi Cai\\(^{1,2,\\dagger}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.207,
+                0.16,
+                0.796,
+                0.177
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>School of Software Engineering, South China University of Technology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.195,
+                0.177,
+                0.808,
+                0.209
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)Key Laboratory of Big Data and Intelligent Robot (South China University of Technology) Ministry of Education"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.419,
+                0.212,
+                0.584,
+                0.226
+            ],
+            "angle": 0,
+            "content": "ycai@scut.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,
+                0.749
+            ],
+            "angle": 0,
+            "content": "Recently, pre-trained vision-language (VL) models have achieved remarkable success in various cross-modal tasks, including referring expression comprehension (REC). These models are pre-trained on the large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the corresponding images and then fine-tuned on downstream tasks. However, the performance of VL models is hindered when dealing with implicit text, which describes objects through comparisons between two or more objects rather than explicitly mentioning them. This is because the models struggle to align the implicit text with the objects in the images. To address the challenge, we introduce CLEVR-Implicit, a dataset consisting of synthetic images and corresponding two types of implicit text for the REC task. Additionally, to enhance the performance of VL models on implicit text, we propose a method called Transform Implicit Text into Explicit Text (TIE), which enables VL models to process with the implicit text. TIE consists of two modules: (1) the prompt design module builds prompts for implicit text by adding masked tokens, and (2) the cloze procedure module fine-tunes the prompts by utilizing masked language modeling (MLM) to predict the explicit words with the implicit prompts. Experimental results on our dataset demonstrate a significant improvement of \\(37.94\\%\\) in the performance of VL models on implicit text after employing our TIE method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.763,
+                0.26,
+                0.778
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.789,
+                0.491,
+                0.886
+            ],
+            "angle": 0,
+            "content": "In recent years, pre-trained vision-language models have demonstrated remarkable achievements in various vision-language downstream tasks (Lu et al., 2019; Su et al., 2019; Kim et al., 2021; Yao et al., 2021; Kamath et al., 2021; Liu et al., 2022), including referring expression comprehension (REC)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.251,
+                0.817,
+                0.364
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.571,
+                0.369,
+                0.816,
+                0.39
+            ],
+            "angle": 0,
+            "content": "The left one of the only two objects with the same color and material"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.578,
+                0.397,
+                0.816,
+                0.508
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.567,
+                0.514,
+                0.822,
+                0.536
+            ],
+            "angle": 0,
+            "content": "The right one of the only two objects with different shape and material"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.551,
+                0.883,
+                0.58
+            ],
+            "angle": 0,
+            "content": "Figure 1: Examples from CLEVR-Implicit. There are two types of implicit text: Same and Different."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.611,
+                0.885,
+                0.739
+            ],
+            "angle": 0,
+            "content": "REC involves identifying a specific object in an image based on corresponding textual descriptions. A common approach is to pre-train a vision-language model on large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the images. Fine-tuning is then performed to optimize model performance (Lu et al., 2019; Su et al., 2019; Kamath et al., 2021)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Despite the state-of-the-art results demonstrated by these methods, they still exhibit limitations. Pre-trained models are good at processing explicit text, which describes the appearance attributes of the objects in the image, whether they are reference or target objects. For example, consider the phrase \"The object left on the red cube.\", models rely on the explicit word \"red cube\" to locate the referred object in the image. However, they encounter challenges when it comes to implicit reasoning, which is processing the implicit text in REC. Implicit"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.142,
+                0.893,
+                0.273,
+                0.905
+            ],
+            "angle": 0,
+            "content": "*Equal contribution."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.143,
+                0.905,
+                0.289,
+                0.919
+            ],
+            "angle": 0,
+            "content": "† Corresponding author."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.142,
+                0.893,
+                0.289,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.527,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12820"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.153,
+                0.946,
+                0.845,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12820-12830 December 6-10, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.085,
+                0.493,
+                0.344
+            ],
+            "angle": 0,
+            "content": "reasoning occurs when the textual expression indirectly describes an object through a comparison between two or more objects, one type of implicit text is formatted like \"The left one of two same color objects\". The phrase \"two same color objects\" does not describe the appearance attributes of any objects, only the relationships between them, and there is no direct reference for the models to infer the target. Consequently, models struggle to align \"two same color objects\" with any objects in the image. To evaluate the performance gap between explicit and implicit text, we conducted experiments using existing methods on both types of text separately. Preliminary results indicate that the performance on implicit text is approximately \\(42\\%\\) lower compared to explicit text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.347,
+                0.49,
+                0.492
+            ],
+            "angle": 0,
+            "content": "Besides, the implicit text also plays a significant role in the domain of human-computer interaction. However, existing referring expression datasets, such as RefCOCO (Yu et al., 2016), RefCOCO+ (Yu et al., 2016), RefCOCOg (Mao et al., 2016), and CLEVR-Ref+ (Liu et al., 2019), primarily consist of explicit text. Unfortunately, there is currently no dataset available that is specifically designed to provide research data on implicit text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.497,
+                0.49,
+                0.706
+            ],
+            "angle": 0,
+            "content": "To address the lack of research data and methods for implicit text, we propose a synthetic dataset called CLEVR-Implicit, specifically for implicit reasoning in REC. Unlike existing datasets, the implicit text we constructed is a subset of all types of implicit text, since they describe synthetic images, and do not need to consider external knowledge. Specifically, We focus on the more difficult types (only containing relationships between objects without any description of their appearance attributes). We evaluate the current pre-trained REC models on our CLEVR-Implicit dataset with these types of implicit text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.71,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "In real-world scenarios, most humans will first extract explicit information from visual context and text information when processing these types of implicit text, and then reason the target object from the extracted explicit information. Inspired by the thinking mode of humans, we simulate the process of humans reasoning about the implicit text and introduce a method called TIE (Transform the Implicit text into Explicit text), which enables pre-trained models to process the implicit text by extracting the explicit information. By employing TIE, REC models facilitate alignment between explicit words and corresponding images on CLEVR-"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.083,
+                0.885,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Table 1: The attributes and example values of the objects in CLEVR. We build the CLEVR-Implicit based on these objects."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.52,
+                0.136,
+                0.871,
+                0.238
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Attribute</td><td>Example values</td></tr><tr><td>color</td><td>red,blue,cyan,purple,gray,green</td></tr><tr><td>material</td><td>metal,rubber</td></tr><tr><td>shape</td><td>cube,cylinder,sphere</td></tr><tr><td>size</td><td>large,small</td></tr><tr><td>spatial</td><td>left,right,behind,front</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.26,
+                0.885,
+                0.308
+            ],
+            "angle": 0,
+            "content": "Implicit. Experimental results demonstrate a substantial difference in the performance of REC models on our dataset before and after applying TIE."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.528,
+                0.309,
+                0.856,
+                0.323
+            ],
+            "angle": 0,
+            "content": "We summarize our contributions as follows:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.529,
+                0.334,
+                0.887,
+                0.398
+            ],
+            "angle": 0,
+            "content": "- We analyze the limitations of existing referring expression comprehension models on the implicit text and construct the CLEVR-Implicit dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.529,
+                0.409,
+                0.886,
+                0.473
+            ],
+            "angle": 0,
+            "content": "- We propose the TIE method to enable pretrained REC models to process the implicit text by transforming the implicit text to explicit text."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.53,
+                0.483,
+                0.884,
+                0.545
+            ],
+            "angle": 0,
+            "content": "- We evaluate the existing models on the CLEVR-Implicit and analyze the differences between the models before and after applying TIE."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.529,
+                0.334,
+                0.887,
+                0.545
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.568,
+                0.673,
+                0.583
+            ],
+            "angle": 0,
+            "content": "2 Related Works"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.593,
+                0.853,
+                0.609
+            ],
+            "angle": 0,
+            "content": "2.1 Referring Expression Comprehension"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.614,
+                0.885,
+                0.79
+            ],
+            "angle": 0,
+            "content": "Referring expression comprehension (REC) aims to localize the region in an image that corresponds to a given sentence description. In recent years, there has been a significant increase in research attention towards this task, driven by the increasing applications of REC in the domain of human-computer interaction. The related task of referring expression generation (REG) has been studied for many years (Krahmer and Van Deemter, 2012; Mitchell et al., 2013; Reiter and Dale, 1992) in the natural language processing (NLP) domain."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.791,
+                0.886,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Several real-world datasets, such as RefCOCO, RefCOCO+, and RefCOCOg, have been proposed based on the MSCOCO dataset (Vinyals et al., 2016) for this task. CLEVR-Ref+ (Liu et al., 2019) is a synthetic dataset that contains geometric objects. These datasets have been widely adopted as benchmarks for evaluating the performance of referring expression models. MMI (Mao et al., 2016)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12821"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.376
+            ],
+            "angle": 0,
+            "content": "directly predicted the bounding box by ranking the region proposals. SLR (Yu et al., 2017) utilized a joint embedding model to predict the referred object by learning the visual and language representations in an embedding space. MattNet (Yu et al., 2018) adapted the attention mechanism to parse expressions into three modules and predict the visual regions by a weighted combination of module scores. VL-BERT (Su et al., 2019) adopted a Transformer model and concatenates the visual and language embeddings as inputs. MDETR (Kamath et al., 2021) constructed an end-to-end system based on the DETR (Carion et al., 2020) detector and a language model to learn the alignment between objects and phrases. VLTVG (Yang et al., 2022) improves comprehension ability by utilizing a visual-linguistic verification module to generate discriminative feature representations."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.388,
+                0.3,
+                0.404
+            ],
+            "angle": 0,
+            "content": "2.2 Prompt Learning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.41,
+                0.49,
+                0.666
+            ],
+            "angle": 0,
+            "content": "Prompt learning, initially introduced in the field of NLP, is used to reformulate downstream tasks to the form of the pre-training task masked language modeling (MLM) (Devlin et al., 2018). This approach enables models to effectively utilize the knowledge during the pre-training stage (Liu et al., 2023). Concretely, prompt engineering reformulates the downstream task into a \"fill-in-the-blank\" cloze task with the same objective function as in the MLM task. The effectiveness of prompt learning heavily relies on the design of appropriate prompt templates. Even slight variations in the prompt template can yield significant differences in performance. Therefore, the selection of suitable prompt templates becomes crucial for achieving desirable results."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.669,
+                0.49,
+                0.829
+            ],
+            "angle": 0,
+            "content": "Prompt learning has also been explored in the multi-modal domain through various studies. CLIP (Radford et al., 2021) pre-trained on large-scale image-text pairs and designs manual prompts like \"The photo of [MASK]\" to facilitate image classification by predicting the masked token. DPT (Liu et al., 2022) reformulated the visual question answering (VQA) task as the MLM by converting the input (question) to a declaration with the masked token."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.844,
+                0.265,
+                0.861
+            ],
+            "angle": 0,
+            "content": "3 Methodology"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.871,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "This section begins by explaining the construction process of the CLEVR-Implicit dataset. Subsequently, we introduce our proposed TIE method,"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.537,
+                0.083,
+                0.855,
+                0.098
+            ],
+            "angle": 0,
+            "content": "Table 2: The statistic data of CLEVR-Implicit."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.548,
+                0.107,
+                0.842,
+                0.194
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Attribute num</td><td>Same</td><td>Different</td></tr><tr><td>1</td><td>3000</td><td>-</td></tr><tr><td>2</td><td>3000</td><td>1553</td></tr><tr><td>3</td><td>3000</td><td>3000</td></tr><tr><td>4</td><td>2905</td><td>3000</td></tr></table>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.218,
+                0.885,
+                0.3
+            ],
+            "angle": 0,
+            "content": "which involves converting the implicit text within the CLEVR-Implicit dataset into explicit text. This explicit text is then utilized in the referring expression comprehension task through the fine-tuning process of pre-trained vision-language models."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.313,
+                0.724,
+                0.327
+            ],
+            "angle": 0,
+            "content": "3.1 Dataset Construction"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.335,
+                0.783,
+                0.35
+            ],
+            "angle": 0,
+            "content": "3.1.1 CLEVR-Implicit Overview"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.355,
+                0.885,
+                0.645
+            ],
+            "angle": 0,
+            "content": "CLEVR-Implicit Dataset uses a subset of the CLEVR (Johnson et al., 2017) scenes (10K images for all, \\(60\\%\\) for the train set and \\(40\\%\\) for the validation and test set). Table 1 provides an overview of the attributes and values associated with the objects in CLEVR. In order to enhance the difficulty of the dataset, we filter and choose the images that contain 3 or more objects. Given that CLEVR-Ref+ provides scene graphs about the attributes (such as color, material, shape, and size) of objects, as well as the spatial relationships between them within the image, we generate corresponding referring expressions for each image. A total of 19K referring expressions are generated, averaging 1.9 referring expressions for each image. Finally, the bounding box of the referred object is assigned as the label for each referring expression in the CLEVR-Implicit Dataset."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.657,
+                0.766,
+                0.672
+            ],
+            "angle": 0,
+            "content": "3.1.2 Implicit Text Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.677,
+                0.885,
+                0.838
+            ],
+            "angle": 0,
+            "content": "To generate implicit text that describes objects through a comparison between two or more objects, we utilize the scene graphs provided by CLEVR-Ref+. A program is employed to extract the attributes of all the objects present in the images. This allows for the comparison and analysis of their attributes, enabling the generation of implicit text that captures the relationships between objects without explicitly mentioning any specific object in the image."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.84,
+                0.887,
+                0.919
+            ],
+            "angle": 0,
+            "content": "For an image, we compare the attributes of each two objects and find each set of objects that satisfies the requirement for generating implicit text. In CLEVR-Implicit, we set up two types of implicit text: \"same\" and \"different\". The \"same\""
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12822"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.115,
+                0.081,
+                0.885,
+                0.297
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.112,
+                0.306,
+                0.884,
+                0.365
+            ],
+            "angle": 0,
+            "content": "Figure 2: The framework of our TIE method and subsequent REC process. The implicit text in CLEVR-Implicit is converted to an implicit prompt which contains a masked token before the attribute. Then vision-language model predicts the explicit words by Multi-step and Path-select prediction strategies in the MLM task. After TIE, the output explicit text and the image are fed to referring expression model to predict the result."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.39,
+                0.489,
+                0.534
+            ],
+            "angle": 0,
+            "content": "type represents that the text refers to the object by describing the same of certain attributes between objects (e.g. \"The left one of the only two objects with the same color and material\"), while the \"different\" type is the opposite of \"same\" (e.g., \"The left object that is different shape and size from the other one). Table 2 provides statistics regarding the distribution of \"same\" and \"different\" type samples in the CLEVR-Implicit dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.542,
+                0.49,
+                0.833
+            ],
+            "angle": 0,
+            "content": "It is worth mentioning that the \"same\" type of referring expressions also exist in the CLEVR-Ref+ dataset. The key difference in \"same\" type text between CLEVR-Implicit and CLEVR-Ref+ is whether the explicit words are hidden. The \"same\" type of text in CLEVR-Ref+ is like \"The objects that have the same size as the red sphere\", which contains the explicit word \"red sphere\". The explicit context provides REC models with a reference object and helps them locate the target. Oppositely, the \"same\" type of implicit text in CLEVR-Implicit is formatted like \"The two objects with the same size\", which hides explicit words. Without explicit context and hard to find the reference object, REC models are forced to compare every two objects to identify what the \"size\" is and whether they are the same. That illustrates that implicit text in CLEVR-Implicit is more challenging."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.84,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We pay more attention to the reference to a single object in the context of REC. During the generation process, we take precautions to avoid text ambiguity, ensuring that the implicit text describes only one object in the image. To achieve this, we"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.39,
+                0.885,
+                0.566
+            ],
+            "angle": 0,
+            "content": "prioritize maximizing the number of attributes involved in attribute comparisons. For example, if the target is in a group of objects that share the same color only, we must ensure no other group of objects that can satisfy the same color and material. In cases where such a group of objects exists, we will change the target and modify the implicit text to describe the corresponding number of attributes until no additional attributes can be included. More details of the implicit text generation are shown in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.568,
+                0.884,
+                0.697
+            ],
+            "angle": 0,
+            "content": "The implicit text in CLEVR-Implicit follows a standardized format based on several templates and rules. Especially, it is common for the same sentence to describe different target objects across different images. By employing homogeneous text, we purposefully direct the model's focus toward the image itself, thereby effectively testing the model's reasoning ability."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.713,
+                0.655,
+                0.727
+            ],
+            "angle": 0,
+            "content": "3.2 TIE Method"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.737,
+                0.682,
+                0.751
+            ],
+            "angle": 0,
+            "content": "3.2.1 TIE Overview"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "To address the challenge of implicit text comprehension, we propose a method called Transform Implicit text to Explicit text (TIE). The objective of TIE is to incorporate explicit words into the original implicit text, enabling the model to reason more effectively. Our approach consists of two key modules: (1) The prompt design module builds prompts for implicit text by adding masked tokens. By strategically designing the prompts, we prepare the groundwork for the next module. (2)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12823"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.132
+            ],
+            "angle": 0,
+            "content": "The cloze procedure module uses a pre-trained vision-language model to predict the explicit words with the implicit prompts."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.134,
+                0.49,
+                0.199
+            ],
+            "angle": 0,
+            "content": "By leveraging these two modules, TIE enables the integration of explicit words into the implicit text and facilitates indirect reasoning on the implicit text effectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.209,
+                0.358,
+                0.224
+            ],
+            "angle": 0,
+            "content": "3.2.2 Prompt Design Module"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.228,
+                0.49,
+                0.729
+            ],
+            "angle": 0,
+            "content": "The prompt design module converts the implicit text into a format suitable for the Masked Language Modeling (MLM) task. One approach to adding explicit words is by obtaining the specific attribute values involved in the comparison of attributes in the implicit text (e.g., \"What is the color in 'The left one of the only two objects with the same color'?\"). Consequently, the prompt design module generates the prompt like \"The left one of the only two objects with the same [MASK] color\" for an implicit text sample. It's important to note that the conversion process differs between the \"same\" and \"different\" types of implicit text. In the \"same\" type, only one identical attribute value needs to be predicted, while in the \"different\" type, two different attribute values must be predicted. Therefore, in the \"different\" type, two masked tokens are added before each attribute in the implicit text. This process of converting the implicit text can be seen as a translation task, where the source text is the original implicit text, and the target text is the input format required for the MLM task. To accomplish this conversion process, we create an implicit prompt dataset using the implicit text from CLEVR-Implicit as the source text, and the target text is a text with the masked token preceding each attribute in the implicit text. Subsequently, we train an encoder-decoder model (T5, Raffel et al., 2020) on the implicit prompt dataset to add the appropriate number of masked tokens corresponding to the attributes for each sample in CLEVR-Implicit."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.739,
+                0.37,
+                0.753
+            ],
+            "angle": 0,
+            "content": "3.2.3 Cloze Procedure Module"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.491,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The training objective of the cloze procedure module aligns with the MLM task, a commonly employed pre-training task in various pre-trained models, so another cloze dataset is devised for the MLM training stage. In this dataset, the input consists of implicit text containing masked tokens and an image description, while the target comprises explicit text containing the ground-truth attribute values. In this module, we fine-tune a pre-trained vision-language model ViLT to predict a specific word"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.085,
+                0.883,
+                0.132
+            ],
+            "angle": 0,
+            "content": "within the masked tokens. And we will introduce three prediction strategies: Base, Multi-step, and Path-select."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.137,
+                0.885,
+                0.314
+            ],
+            "angle": 0,
+            "content": "Base In the prediction stage, the prediction of multiple attribute values in a sample is different from the prediction of a single attribute value. For implicit text containing multiple attributes, a straightforward prediction strategy is to predict all attribute values. However, this strategy tends to decrease the correlation among the predicted values, resulting in potential discrepancies where multiple predicted values may not refer to the same group of objects. Consequently, incorrect prediction results may arise."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.317,
+                0.885,
+                0.59
+            ],
+            "angle": 0,
+            "content": "Multi-step To address the aforementioned issue, we propose an enhanced strategy named multi-step prediction. Specifically, when predicting a sample with multiple attributes, the model outputs only one attribute value at a time. Subsequently, it incorporates the previous prediction as input to predict the next attribute value. Compared with the method of predicting all attribute values at one time, this strategy enhances the correlation among multiple attribute values. For instance, if the original text is \"The left one of the only two objects with the same [MASK] color and [MASK] material,\" the corresponding first-step output is adjusted to \"The left one of the only two objects with the same red color and [MASK] material.\" And the second-step output is \"The left one of the only two objects with the same red color and metal material.\""
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.595,
+                0.884,
+                0.819
+            ],
+            "angle": 0,
+            "content": "Path-select We find that the Multi-step strategy has a potential challenge where the accuracy of predictions relies on the first attribute. If the first attribute prediction is incorrect, it might trigger a chain reaction leading to subsequent attributes being errors. To mitigate this issue, we search for some objects that are not the targets as negative samples. They are characterized by partially satisfying the constraints of implicit text. Part of the attributes that do not satisfy are marked [NONE]. For instance, for a target sample \"The same red color and rubber material\" in cloze dataset, the corresponding negative sample might be \"The same blue color and [NONE] material\"."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.823,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "During the prediction stage, when the model obtains the optimal Softmax result for an attribute and predicts [NONE] for the next attribute, it recognizes this as an incorrect predicted path. Subsequently, the model reverts to the prediction of the preceding attribute, considering the suboptimal"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12824"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.217,
+                0.081,
+                0.782,
+                0.168
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">ViLT (Kim et al., 2021) MLM</td><td colspan=\"2\">Same</td><td colspan=\"2\">Different</td><td colspan=\"2\">Total</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>Base</td><td>80.54</td><td>81.97</td><td>25.73</td><td>23.35</td><td>61.99</td><td>60.48</td></tr><tr><td>Multi-step</td><td>92.55</td><td>91.42</td><td>87.05</td><td>85.47</td><td>91.45</td><td>91.50</td></tr><tr><td>Path-select</td><td>91.76</td><td>91.35</td><td>85.58</td><td>87.50</td><td>91.48</td><td>90.37</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.177,
+                0.884,
+                0.207
+            ],
+            "angle": 0,
+            "content": "Table 3: The TIE experiment results on CLEVR-Implicit. We use three different prediction strategies to compare the performance of ViLT."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.189,
+                0.217,
+                0.809,
+                0.288
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">Method</td><td colspan=\"2\">Before TIE</td><td colspan=\"2\">After TIE</td><td colspan=\"2\">Ground-truth</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>MDETR (Kamath et al., 2021)</td><td>53.57</td><td>53.89</td><td>91.51</td><td>92.19</td><td>96.01</td><td>95.66</td></tr><tr><td>VLTVG (Yang et al., 2022)</td><td>58.83</td><td>58.46</td><td>84.48</td><td>86.15</td><td>95.97</td><td>96.85</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.297,
+                0.884,
+                0.34
+            ],
+            "angle": 0,
+            "content": "Table 4: Referring Expression Comprehension results performed by MDETR and VLTVG. Before TIE means that the models are trained on the implicit text before the input transformation of the TIE method, and After TIE means that the models are trained on the explicit text after TIE."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.366,
+                0.486,
+                0.397
+            ],
+            "angle": 0,
+            "content": "Softmax result, which guides the model toward the correct prediction."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.409,
+                0.262,
+                0.425
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.434,
+                0.331,
+                0.449
+            ],
+            "angle": 0,
+            "content": "4.1 Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.454,
+                0.49,
+                0.777
+            ],
+            "angle": 0,
+            "content": "For the REC task on the CLEVR-Implicit dataset, we employ two models: MDETR (Kamath et al., 2021) and VLTVG (Yang et al., 2022). MDETR utilizes weights pre-trained on various datasets such as COCO (RefCOCO, RefCOCO+, RefCOCOg), Visual Genome (Krishna et al., 2017), and Flickr30k (Plummer et al., 2015). and VLTVG uses weights pre-trained on RefCOCOg. And we fine-tune them on the CLEVR-Implicit dataset. During the training stage, we use the settings of learning rate 1e-4, learning rate of backbone 1e-4, batch size 16 and weight decay 1e-5 to train for 60 epochs. In the input transformation experiment, we utilize the T5-small model for the prompt design module. As for the cloze procedure module, we select ViLT (Kim et al., 2021). To ensure a fair comparison, all different prediction methods are configured with the AdamW optimizer, a learning rate of 1e-4, a batch size of 32, and an image size of \\(384 \\times 384\\) to train for 70 epochs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.787,
+                0.275,
+                0.801
+            ],
+            "angle": 0,
+            "content": "4.2 Results of TIE"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.807,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Table 3 presents the results of the TIE experiment. The evaluation method is to assess whether the predicted explicit words match the corresponding ground-truth values. For samples containing multiple attributes, it is necessary to ensure the accurate prediction of all explicit words. otherwise, it is considered incorrect. In the case of \"different\" type"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.366,
+                0.884,
+                0.413
+            ],
+            "angle": 0,
+            "content": "samples, where attribute values differ between objects, the order of the two predicted values for a single attribute can be reversed."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.418,
+                0.885,
+                0.756
+            ],
+            "angle": 0,
+            "content": "We present the results separately for training on the \"same\" and \"different\" types of samples, and the \"Total\" column represents the combined performance on both types of implicit text. In the Base strategy, we observe that \"different\" samples are more challenging to understand compared to \"same\" samples, resulting in a decrease in performance of up to \\(54.81\\%\\). For \"different\" samples, the model needs to output two values to predict a single attribute, resulting in twice as many explicit words to be predicted compared to \"same\" samples. In contrast, the Multi-step strategy exhibits significant improvements for both types of text. It shows a \\(12.01\\%\\) increase in performance for \"same\" samples and a remarkable \\(61.32\\%\\) increase for \"different\" samples. A reasonable explanation is that the prediction of each explicit word takes into account all previously predicted results. Multi-step strategy mitigates cases where the initial predictions are correct, but subsequent predictions suffer from errors, leading to more accurate overall predictions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "From the comparing results of the Path-select and Multi-step strategies, we observe some differences when incorporating negative samples into the training set. The Path-select results on the validation set exhibit a slight decrease of \\(0.79\\%\\) and \\(1.47\\%\\) compared to Multi-step. We speculate that this is because the majority of incorrect positions in wrongly predicted samples tend to occur in the second attribute or later, and the prediction of the first attribute in these samples is generally accurate."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12825"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.159,
+                0.106,
+                0.465,
+                0.269
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.263,
+                0.278,
+                0.38,
+                0.291
+            ],
+            "angle": 0,
+            "content": "(a) same attributes"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.514,
+                0.106,
+                0.822,
+                0.27
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.609,
+                0.278,
+                0.746,
+                0.291
+            ],
+            "angle": 0,
+            "content": "(b) different attributes"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.306,
+                0.882,
+                0.336
+            ],
+            "angle": 0,
+            "content": "Figure 3: 3(a) is the performance of three prediction strategies on \"same\" type samples in test set, and 3(b) is the performance of three prediction strategies on \"different\" type samples in test set."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.36,
+                0.489,
+                0.504
+            ],
+            "angle": 0,
+            "content": "However, on the test set of \"different\" samples, Path-select surpasses Multi-step by \\(2.03\\%\\) in terms of performance. Particularly, the performance of Multi-step and Path-select in the \"Total\" column is comparable, indicating that the addition of negative samples can enhance the model's robustness. And negative samples enable the model to choose the correct path when the first attribute prediction is incorrect."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.515,
+                0.411,
+                0.531
+            ],
+            "angle": 0,
+            "content": "4.3 Results of Referring Expression"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.535,
+                0.49,
+                0.826
+            ],
+            "angle": 0,
+            "content": "Table 4 presents the results of the referring expression experiment. We evaluate the performance of MDETR and VLTVG, both before and after applying the TIE based on the Multi-step prediction strategy. Additionally, we provide the performance of the models trained on ground-truth explicit samples as a reference. The evaluation metric is Intersection over Union (IoU), with a threshold value of 0.5. We observe that prior to TIE, the performance of both models trained on implicit text is relatively poor, indicating limited reasoning ability when dealing with implicit text. However, after applying TIE, there is a significant improvement in performance for both models (53.57% vs. 91.51% for MDETR and 58.83% vs. 84.48% for VLTVG). However, compared with Ground-truth, there is a certain gap in performance, because it is limited by the impact of TIE performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.835,
+                0.411,
+                0.851
+            ],
+            "angle": 0,
+            "content": "4.4 Results of Contrast Experiment"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Difficulty of CLEVR-Implicit In order to contextualize the difficulty of CLEVR-Implicit, we evaluate the performance of REC models on CLEVR-Implicit when the models are trained on CLEVR-"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.558,
+                0.357,
+                0.836,
+                0.428
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td colspan=\"2\">Train on CLEVR-Ref+</td></tr><tr><td></td><td>val</td><td>test</td></tr><tr><td>MDETR</td><td>47.02</td><td>45.85</td></tr><tr><td>VLTVG</td><td>47.99</td><td>47.11</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.436,
+                0.884,
+                0.48
+            ],
+            "angle": 0,
+            "content": "Table 5: The performances of REC models which are trained on CLEVR-Ref+ and tested on CLEVR-Implicit."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.528,
+                0.495,
+                0.866,
+                0.548
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>VLTVG on CLEVR-Ref+</td><td>val</td><td>test</td></tr><tr><td>Before TIE</td><td>66.43</td><td>65.22</td></tr><tr><td>After TIE</td><td>68.05</td><td>66.46</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.557,
+                0.884,
+                0.602
+            ],
+            "angle": 0,
+            "content": "Table 6: The performances of VLTVG respectively on original CLEVR-Ref+ (Before TIE) and CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.629,
+                0.884,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Ref+. Table 5 shows that the results of MDETR and VLTVG models are all lower than \"Before TIE\" Column results in Table4. That indicates that the experience that the models learn from CLEVR-Ref+ cannot help them process the implicit text, because the samples in CLEVR-Ref+ contain explicit information to help the model locate the target, while the samples in CLEVR-Implicit hide the information. Therefore, we believe that it's challenging to process the implicit text in our CLEVR-Implicit dataset."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Affect of TIE on explicit text Since the TIE method converts implicit text into explicit text, in addition to implicit text, we also need to analyze whether the explicit text after applying TIE will affect the performance of REC models. To verify this problem, we evaluate the performance of VLTVG respectively on original CLEVR-Ref+ and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12826"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.082,
+                0.296,
+                0.168
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.306,
+                0.087,
+                0.551,
+                0.107
+            ],
+            "angle": 0,
+            "content": "O: The left one of the only two objects with the same shape, color and size in the image"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.307,
+                0.115,
+                0.55,
+                0.143
+            ],
+            "angle": 0,
+            "content": "P: The left one of the only two objects with the same [MASK] shape, [MASK] color and [MASK] size in the image"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.307,
+                0.151,
+                0.439,
+                0.162
+            ],
+            "angle": 0,
+            "content": "GT: cylinder, yellow, large"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.595,
+                0.092,
+                0.735,
+                0.103
+            ],
+            "angle": 0,
+            "content": "Base: sphere, brown, large"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.595,
+                0.119,
+                0.77,
+                0.13
+            ],
+            "angle": 0,
+            "content": "Multi-step: cylinder, yellow, large"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.595,
+                0.146,
+                0.773,
+                0.157
+            ],
+            "angle": 0,
+            "content": "Path-select: cylinder, yellow, large"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.149,
+                0.173,
+                0.733,
+                0.185
+            ],
+            "angle": 0,
+            "content": "(a) An example of the comparison from Base, Multi-step and Path-select prediction strategies"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.117,
+                0.201,
+                0.296,
+                0.287
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.181,
+                0.292,
+                0.294,
+                0.303
+            ],
+            "angle": 0,
+            "content": "(b) REC before TIE"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.299,
+                0.224,
+                0.486,
+                0.262
+            ],
+            "angle": 0,
+            "content": "Before: The left one of the only two objects with the same shape, color and size in the image"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.498,
+                0.201,
+                0.677,
+                0.286
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.562,
+                0.292,
+                0.663,
+                0.303
+            ],
+            "angle": 0,
+            "content": "(c) REC after TIE"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.682,
+                0.224,
+                0.872,
+                0.263
+            ],
+            "angle": 0,
+            "content": "After: The left one of the only two objects with the same cylinder shape, yellow color and large size in the image"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.318,
+                0.882,
+                0.377
+            ],
+            "angle": 0,
+            "content": "Figure 4: Examples from TIE and referring expression comprehension experiments. (a) an example of the comparison from Base, Multi-step and Path-select prediction strategies. O, P and GT denote original implicit text, generated prompt text and ground-truth explicit words. (b) Visualization of the predicted bounding box from VLTVG before TIE and (c) after TIE. green and red rectangles denote ground-truth and predicted bounding box."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.401,
+                0.487,
+                0.497
+            ],
+            "angle": 0,
+            "content": "CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit. Table 6 shows that TIE does not affect performance (even better on VLTVG) on other types of referring expressions besides the implicit text, and TIE is robust in a broader class of referring expressions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.498,
+                0.489,
+                0.594
+            ],
+            "angle": 0,
+            "content": "Contrast between explicit and implicit As shown in Table 7, we evaluate the performance of MDETR and VLTVG models separately on predicted explicit text and implicit text when the models are trained on implicit text and ground-truth explicit text respectively."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.596,
+                0.49,
+                0.772
+            ],
+            "angle": 0,
+            "content": "When both validation sets provide predicted explicit words (Column 1 vs Column 3), the performance of the model trained on the implicit training set is significantly decreased (MDETR: \\(31.48\\%\\) vs \\(91.98\\%\\), VLTVG: \\(53.85\\%\\) vs \\(85.90\\%\\)) compared with that on the explicit training set, which is evidence that explicit words play a crucial role in the models' comprehension and reasoning processes. The same situation occurs when exchanging implicit and explicit states on train and validation sets (Column 2 vs Column 3)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.785,
+                0.447,
+                0.816
+            ],
+            "angle": 0,
+            "content": "4.5 Performance on different number of attributes"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.823,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Figure 3 illustrates the performance of ViLT on \"same\" and \"different\" samples with different numbers of attributes. In Figure 3(a), we observe that the model achieves the highest performance when predicting one-attribute samples, while the lowest performance is observed when predicting three-"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.534,
+                0.398,
+                0.86,
+                0.468
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Method</td><td>Train-i Val-pe</td><td>Train-ge Val-i</td><td>Train-ge Val-pe</td></tr><tr><td>MDETR</td><td>31.48</td><td>46.25</td><td>91.98</td></tr><tr><td>VLTVG</td><td>53.85</td><td>48.34</td><td>85.90</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.477,
+                0.883,
+                0.548
+            ],
+            "angle": 0,
+            "content": "Table 7: The contrast experiment for REC. The i suffix denotes the set with implicit samples. The pe suffix denotes that the explicit words in samples are predicted by TIE. The ge denotes the explicit words in samples is ground-truth."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.576,
+                0.885,
+                0.817
+            ],
+            "angle": 0,
+            "content": "attribute samples. Interestingly, the performance on four-attribute samples is second only to that of one-attribute samples. A reasonable conjecture is that a set of objects is identical when all four properties are the same in CLEVR images, so the model only needs to find two objects that are identical. Figure 3(b) shows that predicting four-attribute \"different\" samples has the most significant challenge on the Base strategy. However, on the Multi-step strategy, the performance on four-attribute \"different\" samples surpasses that of three-attribute samples. This indicates that the Multi-step prediction strategy effectively mitigates the impact of an increased number of predicted words, resulting in improved performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.833,
+                0.644,
+                0.848
+            ],
+            "angle": 0,
+            "content": "4.6 Case Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.855,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Figure 4 indicates an example of the comparison of different prediction strategies in the TIE experiment, and we visualize the ground-truth bounding box and the bounding box predicted from VLTVG"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12827"
+        }
+    ],
+    [
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.083,
+                0.486,
+                0.111
+            ],
+            "angle": 0,
+            "content": "Table 8: Examples about the application of implicit text in daily life."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.135,
+                0.124,
+                0.414,
+                0.139
+            ],
+            "angle": 0,
+            "content": "Real-world Implicit Text Examples"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.136,
+                0.141,
+                0.461,
+                0.155
+            ],
+            "angle": 0,
+            "content": "Two persons are doing the same movement."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.157,
+                0.425,
+                0.171
+            ],
+            "angle": 0,
+            "content": "His height is different from each other."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.173,
+                0.426,
+                0.188
+            ],
+            "angle": 0,
+            "content": "His pants and shoes are the same color."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.189,
+                0.449,
+                0.203
+            ],
+            "angle": 0,
+            "content": "The same brand of cosmetics on the table."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.137,
+                0.205,
+                0.457,
+                0.22
+            ],
+            "angle": 0,
+            "content": "Take the drugs used for the same symptom."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.248,
+                0.489,
+                0.505
+            ],
+            "angle": 0,
+            "content": "before and after TIE in the REC experiment. In Figure 4(a), with the attributes of the example \"shape\" and \"color\", the Base strategy mistakenly predicts \"sphere\" and \"brown\". We observe that from the image there is a group of objects that is both \"sphere\" and another group of objects that is both \"brown\", but there are no such two objects that satisfy \"sphere\" and \"brown\" simultaneously. This indicates that the Base strategy performs poorly on multi-attribute samples due to the weak correlation between attributes, leading to the error that different attributes refer to different groups of objects. In contrast, the Multi-step and Path-select strategies enhance the correlation between attributes and make predictions step-by-step, leading to more accurate predictions."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.507,
+                0.49,
+                0.684
+            ],
+            "angle": 0,
+            "content": "In Figure 4(b), by taking the implicit text as input, VLTVG incorrectly predicts the referred object as a brown cube. While VLTVG predicts the referred object accurately when taking the explicit text after TIE as input. The comparison result indicates that current models are being confused when performing implicit reasoning based on the logical relation conveyed by the implicit text. Only when explicit words are explicitly included, the models can align the explicit words with the corresponding objects in the image for explicit reasoning."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.699,
+                0.471,
+                0.715
+            ],
+            "angle": 0,
+            "content": "5 Implicit Text in Real-world Scenarios"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.727,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The implicit descriptions in real-world scenarios are more complex and difficult because they not only describe the targets implicitly through the geometrical attributes of color, material, etc., but also contain more attributes as well as external factors. Table 8 shows the implicit text examples in real-world daily life. It is important that even though our generated implicit text relies on the CLEVR images and the provided annotations, our proposed TIE method can still extend the scope beyond the CLEVR domain to real-world scenarios. For example, based on Figure 5 which is chosen from Visual"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.537,
+                0.082,
+                0.868,
+                0.259
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.516,
+                0.264,
+                0.862,
+                0.289
+            ],
+            "angle": 0,
+            "content": "Original Implicit Text: The left one of two persons with the same posture and gender"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.297,
+                0.885,
+                0.321
+            ],
+            "angle": 0,
+            "content": "After Prompt Design Module: The left one of two persons with the same [MASK] posture and [MASK] gender"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.516,
+                0.329,
+                0.885,
+                0.353
+            ],
+            "angle": 0,
+            "content": "After Cloze Procedure Module: The left one of two persons with the same walking posture and woman gender"
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.522,
+                0.368,
+                0.869,
+                0.382
+            ],
+            "angle": 0,
+            "content": "Figure 5: A real-world image from Visual Genome"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.41,
+                0.884,
+                0.617
+            ],
+            "angle": 0,
+            "content": "Genome, consider the implicit text like \"The left one of two persons with the same posture and gender\". Through the TIE method, the prompt can be built before the \"posture\" and \"gender\" by prompt design module, like \"The left one of two persons with the same [MASK] posture and [MASK] gender\". The prompt guides the cloze procedure module to predict the explicit words \"walking\" and \"woman\" like \"The left one of two persons with the same walking posture and woman gender\". By integrating TIE approach, current models can also process the implicit text on the real-world scenarios."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.634,
+                0.642,
+                0.649
+            ],
+            "angle": 0,
+            "content": "6 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.662,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In this paper, we analyze the significance of implicit reasoning, along with the lack of investigation in this domain. To address this, we introduce the CLEVR-Implicit dataset for implicit reasoning in the REC task. This dataset utilizes synthesized images to generate two types of implicit text: \"same\" and \"different.\" Evaluating the performance of existing pre-trained models on CLEVR-Implicit, we observe model confusion when attempting to align implicit text with images. To overcome this challenge, we propose the Transform Implicit text to Explicit text (TIE) method, which converts implicit text into explicit text. Our experiments demonstrate that the reasoning capability of the models improves significantly through the alignment of explicit words and images."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12828"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.221,
+                0.1
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.109,
+                0.491,
+                0.288
+            ],
+            "angle": 0,
+            "content": "The limitation of our work is that CLEVR-Implicit is built on synthetic images, so it's straightforward to generate corresponding implicit text based on the annotation of the synthetic images. Besides, since the synthetic images contain only geometric objects, our implicit text can only be generated within the scope of the object's attributes. Therefore, in future work, we will annotate more types of implicit text based on real-world scenarios. This will allow us to explore the mechanisms of model understanding of implicit text."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.297,
+                0.287,
+                0.314
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.321,
+                0.49,
+                0.435
+            ],
+            "angle": 0,
+            "content": "This work was supported by the National Natural Science Foundation of China (62076100), Fundamental Research Funds for the Central Universities, SCUT (x2rjD2230080), the Science and Technology Planning Project of Guangdong Province (2020B0101100002), CAAI-Huawei MindSpore Open Fund, CCF-Zhipu AI Large Model Fund."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.46,
+                0.214,
+                0.475
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.482,
+                0.489,
+                0.549
+            ],
+            "angle": 0,
+            "content": "Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.557,
+                0.488,
+                0.61
+            ],
+            "angle": 0,
+            "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.618,
+                0.487,
+                0.709
+            ],
+            "angle": 0,
+            "content": "Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.718,
+                0.489,
+                0.797
+            ],
+            "angle": 0,
+            "content": "Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. 2021. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.805,
+                0.49,
+                0.87
+            ],
+            "angle": 0,
+            "content": "Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.879,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Emiel Krahmer and Kees Van Deemter. 2012. Computational generation of referring expressions: A survey. Computational Linguistics, 38(1):173-218."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.482,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.886,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.189,
+                0.885,
+                0.256
+            ],
+            "angle": 0,
+            "content": "Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.265,
+                0.885,
+                0.332
+            ],
+            "angle": 0,
+            "content": "Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. 2019. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.342,
+                0.883,
+                0.394
+            ],
+            "angle": 0,
+            "content": "Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. 2022. Declaration-based prompt tuning for visual question answering. arXiv preprint arXiv:2205.02456."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.405,
+                0.885,
+                0.459
+            ],
+            "angle": 0,
+            "content": "Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.469,
+                0.885,
+                0.548
+            ],
+            "angle": 0,
+            "content": "Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. 2016. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.558,
+                0.885,
+                0.638
+            ],
+            "angle": 0,
+            "content": "Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. 2013. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.648,
+                0.885,
+                0.739
+            ],
+            "angle": 0,
+            "content": "Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.75,
+                0.885,
+                0.841
+            ],
+            "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": [
+                0.511,
+                0.852,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12829"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.116,
+                0.086,
+                0.487,
+                0.152
+            ],
+            "angle": 0,
+            "content": "Ehud Reiter and Robert Dale. 1992. A fast algorithm for the generation of referring expressions. In Proceedings of the 14th Conference on Computational Linguistics - Volume 1, page 232-238. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.162,
+                0.487,
+                0.215
+            ],
+            "angle": 0,
+            "content": "Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. 2019. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.225,
+                0.487,
+                0.291
+            ],
+            "angle": 0,
+            "content": "Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2016. Show and tell: Lessons learned from the 2015 mscoco image captioning challenge. IEEE transactions on pattern analysis and machine intelligence, 39(4):652-663."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.301,
+                0.487,
+                0.379
+            ],
+            "angle": 0,
+            "content": "Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. 2022. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499-9508."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.39,
+                0.487,
+                0.443
+            ],
+            "angle": 0,
+            "content": "Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2021. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.453,
+                0.487,
+                0.532
+            ],
+            "angle": 0,
+            "content": "Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. 2018. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.542,
+                0.487,
+                0.596
+            ],
+            "angle": 0,
+            "content": "Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. 2016. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.606,
+                0.487,
+                0.671
+            ],
+            "angle": 0,
+            "content": "Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. 2017. A joint speaker-listener-reinforcer model for referring expressions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7282-7290."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.685,
+                0.402,
+                0.716
+            ],
+            "angle": 0,
+            "content": "A More Details of Implicit Text Generation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.727,
+                0.487,
+                0.919
+            ],
+            "angle": 0,
+            "content": "To clarify on our implicit referring expression generation process, we'd like to elaborate on the method used to create the CLEVR-Implicit dataset from the existing CLEVR images and their associated annotations. These annotations contain object attributes, as well as spatial relationships between objects. For each image, we will traverse all the objects in the image by comparing whether the attribute values between two objects are the same/different each time. For example, if an image contains 4 objects, this leads to \\( \\mathrm{C}(4,2) = 6 \\) object pair comparisons (e.g., (1,2), (1,3), (1,4), (2,3),"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.085,
+                0.882,
+                0.422
+            ],
+            "angle": 0,
+            "content": "(2,4), (3,4), the numbers are object ids). We will generate the corresponding implicit text if a certain pair satisfies the \"same\" or \"different\" condition. Suppose the (2,3) combination objects' shape and color are both cube and blue, and guarantee that no other pairs satisfy the condition \"same shape and color\" to avoid the ambiguity in the REC task, then we will generate the \"same\" type of implicit text as \"The two objects with the same shape and color\". Due to the only one target of the REC task, we must choose one of the (2,3) objects to be the target. We employ a random selection for the target based on the spatial relationship between 2 and 3. If object 2 is left of object 3 and object 2 is selected, the finally generated implicit text is \"The left one of the two objects with the same shape and color\". The \"different\" types of text are generated in the same way. This consistent pattern of implicit text generation in the CLEVR-Implicit dataset ensures both the concealment of explicit information and the reference of target objects in the images."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12830"
+        }
+    ]
+]

2023/CLEVR-Implicit_ A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension/7a840248-0cf8-41af-aee2-5d7039744b4d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3981c05e3665d4f94c8d7aa2ea7b3aba97441b13f8d1a4bf858afb4c48e3f5b
+size 2771656

2023/CLEVR-Implicit_ A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension/full.md

@@ -0,0 +1,312 @@
+# CLEVR-Implicit: A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension
+
+Jingwei Zhang $^{1,\ast}$ , Xin Wu $^{1,\ast}$ , Yi Cai $^{1,2,\dagger}$
+
+<sup>1</sup>School of Software Engineering, South China University of Technology
+
+$^{2}$ Key Laboratory of Big Data and Intelligent Robot (South China University of Technology) Ministry of Education
+
+ycai@scut.edu.cn
+
+# Abstract
+
+Recently, pre-trained vision-language (VL) models have achieved remarkable success in various cross-modal tasks, including referring expression comprehension (REC). These models are pre-trained on the large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the corresponding images and then fine-tuned on downstream tasks. However, the performance of VL models is hindered when dealing with implicit text, which describes objects through comparisons between two or more objects rather than explicitly mentioning them. This is because the models struggle to align the implicit text with the objects in the images. To address the challenge, we introduce CLEVR-Implicit, a dataset consisting of synthetic images and corresponding two types of implicit text for the REC task. Additionally, to enhance the performance of VL models on implicit text, we propose a method called Transform Implicit Text into Explicit Text (TIE), which enables VL models to process with the implicit text. TIE consists of two modules: (1) the prompt design module builds prompts for implicit text by adding masked tokens, and (2) the cloze procedure module fine-tunes the prompts by utilizing masked language modeling (MLM) to predict the explicit words with the implicit prompts. Experimental results on our dataset demonstrate a significant improvement of  $37.94\%$  in the performance of VL models on implicit text after employing our TIE method.
+
+# 1 Introduction
+
+In recent years, pre-trained vision-language models have demonstrated remarkable achievements in various vision-language downstream tasks (Lu et al., 2019; Su et al., 2019; Kim et al., 2021; Yao et al., 2021; Kamath et al., 2021; Liu et al., 2022), including referring expression comprehension (REC).
+
+![](images/77222350d0df425f570fe3c29d53208ff917f0073d72d9c8c398573d991a08d6.jpg)  
+The left one of the only two objects with the same color and material
+
+![](images/fe8d46a1c8337935d4cd0d7fbf538785d0ed7c9d16d5b87a45bec261f3c50a47.jpg)  
+The right one of the only two objects with different shape and material  
+Figure 1: Examples from CLEVR-Implicit. There are two types of implicit text: Same and Different.
+
+REC involves identifying a specific object in an image based on corresponding textual descriptions. A common approach is to pre-train a vision-language model on large-scale image-text pairs to learn the alignment between words in textual descriptions and objects in the images. Fine-tuning is then performed to optimize model performance (Lu et al., 2019; Su et al., 2019; Kamath et al., 2021).
+
+Despite the state-of-the-art results demonstrated by these methods, they still exhibit limitations. Pre-trained models are good at processing explicit text, which describes the appearance attributes of the objects in the image, whether they are reference or target objects. For example, consider the phrase "The object left on the red cube.", models rely on the explicit word "red cube" to locate the referred object in the image. However, they encounter challenges when it comes to implicit reasoning, which is processing the implicit text in REC. Implicit
+
+reasoning occurs when the textual expression indirectly describes an object through a comparison between two or more objects, one type of implicit text is formatted like "The left one of two same color objects". The phrase "two same color objects" does not describe the appearance attributes of any objects, only the relationships between them, and there is no direct reference for the models to infer the target. Consequently, models struggle to align "two same color objects" with any objects in the image. To evaluate the performance gap between explicit and implicit text, we conducted experiments using existing methods on both types of text separately. Preliminary results indicate that the performance on implicit text is approximately  $42\%$  lower compared to explicit text.
+
+Besides, the implicit text also plays a significant role in the domain of human-computer interaction. However, existing referring expression datasets, such as RefCOCO (Yu et al., 2016), RefCOCO+ (Yu et al., 2016), RefCOCOg (Mao et al., 2016), and CLEVR-Ref+ (Liu et al., 2019), primarily consist of explicit text. Unfortunately, there is currently no dataset available that is specifically designed to provide research data on implicit text.
+
+To address the lack of research data and methods for implicit text, we propose a synthetic dataset called CLEVR-Implicit, specifically for implicit reasoning in REC. Unlike existing datasets, the implicit text we constructed is a subset of all types of implicit text, since they describe synthetic images, and do not need to consider external knowledge. Specifically, We focus on the more difficult types (only containing relationships between objects without any description of their appearance attributes). We evaluate the current pre-trained REC models on our CLEVR-Implicit dataset with these types of implicit text.
+
+In real-world scenarios, most humans will first extract explicit information from visual context and text information when processing these types of implicit text, and then reason the target object from the extracted explicit information. Inspired by the thinking mode of humans, we simulate the process of humans reasoning about the implicit text and introduce a method called TIE (Transform the Implicit text into Explicit text), which enables pre-trained models to process the implicit text by extracting the explicit information. By employing TIE, REC models facilitate alignment between explicit words and corresponding images on CLEVR-
+
+Table 1: The attributes and example values of the objects in CLEVR. We build the CLEVR-Implicit based on these objects.  
+
+<table><tr><td>Attribute</td><td>Example values</td></tr><tr><td>color</td><td>red,blue,cyan,purple,gray,green</td></tr><tr><td>material</td><td>metal,rubber</td></tr><tr><td>shape</td><td>cube,cylinder,sphere</td></tr><tr><td>size</td><td>large,small</td></tr><tr><td>spatial</td><td>left,right,behind,front</td></tr></table>
+
+Implicit. Experimental results demonstrate a substantial difference in the performance of REC models on our dataset before and after applying TIE.
+
+We summarize our contributions as follows:
+
+- We analyze the limitations of existing referring expression comprehension models on the implicit text and construct the CLEVR-Implicit dataset.  
+- We propose the TIE method to enable pretrained REC models to process the implicit text by transforming the implicit text to explicit text.  
+- We evaluate the existing models on the CLEVR-Implicit and analyze the differences between the models before and after applying TIE.
+
+# 2 Related Works
+
+# 2.1 Referring Expression Comprehension
+
+Referring expression comprehension (REC) aims to localize the region in an image that corresponds to a given sentence description. In recent years, there has been a significant increase in research attention towards this task, driven by the increasing applications of REC in the domain of human-computer interaction. The related task of referring expression generation (REG) has been studied for many years (Krahmer and Van Deemter, 2012; Mitchell et al., 2013; Reiter and Dale, 1992) in the natural language processing (NLP) domain.
+
+Several real-world datasets, such as RefCOCO, RefCOCO+, and RefCOCOg, have been proposed based on the MSCOCO dataset (Vinyals et al., 2016) for this task. CLEVR-Ref+ (Liu et al., 2019) is a synthetic dataset that contains geometric objects. These datasets have been widely adopted as benchmarks for evaluating the performance of referring expression models. MMI (Mao et al., 2016)
+
+directly predicted the bounding box by ranking the region proposals. SLR (Yu et al., 2017) utilized a joint embedding model to predict the referred object by learning the visual and language representations in an embedding space. MattNet (Yu et al., 2018) adapted the attention mechanism to parse expressions into three modules and predict the visual regions by a weighted combination of module scores. VL-BERT (Su et al., 2019) adopted a Transformer model and concatenates the visual and language embeddings as inputs. MDETR (Kamath et al., 2021) constructed an end-to-end system based on the DETR (Carion et al., 2020) detector and a language model to learn the alignment between objects and phrases. VLTVG (Yang et al., 2022) improves comprehension ability by utilizing a visual-linguistic verification module to generate discriminative feature representations.
+
+# 2.2 Prompt Learning
+
+Prompt learning, initially introduced in the field of NLP, is used to reformulate downstream tasks to the form of the pre-training task masked language modeling (MLM) (Devlin et al., 2018). This approach enables models to effectively utilize the knowledge during the pre-training stage (Liu et al., 2023). Concretely, prompt engineering reformulates the downstream task into a "fill-in-the-blank" cloze task with the same objective function as in the MLM task. The effectiveness of prompt learning heavily relies on the design of appropriate prompt templates. Even slight variations in the prompt template can yield significant differences in performance. Therefore, the selection of suitable prompt templates becomes crucial for achieving desirable results.
+
+Prompt learning has also been explored in the multi-modal domain through various studies. CLIP (Radford et al., 2021) pre-trained on large-scale image-text pairs and designs manual prompts like "The photo of [MASK]" to facilitate image classification by predicting the masked token. DPT (Liu et al., 2022) reformulated the visual question answering (VQA) task as the MLM by converting the input (question) to a declaration with the masked token.
+
+# 3 Methodology
+
+This section begins by explaining the construction process of the CLEVR-Implicit dataset. Subsequently, we introduce our proposed TIE method,
+
+Table 2: The statistic data of CLEVR-Implicit.  
+
+<table><tr><td>Attribute num</td><td>Same</td><td>Different</td></tr><tr><td>1</td><td>3000</td><td>-</td></tr><tr><td>2</td><td>3000</td><td>1553</td></tr><tr><td>3</td><td>3000</td><td>3000</td></tr><tr><td>4</td><td>2905</td><td>3000</td></tr></table>
+
+which involves converting the implicit text within the CLEVR-Implicit dataset into explicit text. This explicit text is then utilized in the referring expression comprehension task through the fine-tuning process of pre-trained vision-language models.
+
+# 3.1 Dataset Construction
+
+# 3.1.1 CLEVR-Implicit Overview
+
+CLEVR-Implicit Dataset uses a subset of the CLEVR (Johnson et al., 2017) scenes (10K images for all,  $60\%$  for the train set and  $40\%$  for the validation and test set). Table 1 provides an overview of the attributes and values associated with the objects in CLEVR. In order to enhance the difficulty of the dataset, we filter and choose the images that contain 3 or more objects. Given that CLEVR-Ref+ provides scene graphs about the attributes (such as color, material, shape, and size) of objects, as well as the spatial relationships between them within the image, we generate corresponding referring expressions for each image. A total of 19K referring expressions are generated, averaging 1.9 referring expressions for each image. Finally, the bounding box of the referred object is assigned as the label for each referring expression in the CLEVR-Implicit Dataset.
+
+# 3.1.2 Implicit Text Generation
+
+To generate implicit text that describes objects through a comparison between two or more objects, we utilize the scene graphs provided by CLEVR-Ref+. A program is employed to extract the attributes of all the objects present in the images. This allows for the comparison and analysis of their attributes, enabling the generation of implicit text that captures the relationships between objects without explicitly mentioning any specific object in the image.
+
+For an image, we compare the attributes of each two objects and find each set of objects that satisfies the requirement for generating implicit text. In CLEVR-Implicit, we set up two types of implicit text: "same" and "different". The "same"
+
+![](images/6d0e1fdff11b465a75365b4c58d62a849d447d1cf7802f27c0303a46d47f69d3.jpg)  
+Figure 2: The framework of our TIE method and subsequent REC process. The implicit text in CLEVR-Implicit is converted to an implicit prompt which contains a masked token before the attribute. Then vision-language model predicts the explicit words by Multi-step and Path-select prediction strategies in the MLM task. After TIE, the output explicit text and the image are fed to referring expression model to predict the result.
+
+type represents that the text refers to the object by describing the same of certain attributes between objects (e.g. "The left one of the only two objects with the same color and material"), while the "different" type is the opposite of "same" (e.g., "The left object that is different shape and size from the other one). Table 2 provides statistics regarding the distribution of "same" and "different" type samples in the CLEVR-Implicit dataset.
+
+It is worth mentioning that the "same" type of referring expressions also exist in the CLEVR-Ref+ dataset. The key difference in "same" type text between CLEVR-Implicit and CLEVR-Ref+ is whether the explicit words are hidden. The "same" type of text in CLEVR-Ref+ is like "The objects that have the same size as the red sphere", which contains the explicit word "red sphere". The explicit context provides REC models with a reference object and helps them locate the target. Oppositely, the "same" type of implicit text in CLEVR-Implicit is formatted like "The two objects with the same size", which hides explicit words. Without explicit context and hard to find the reference object, REC models are forced to compare every two objects to identify what the "size" is and whether they are the same. That illustrates that implicit text in CLEVR-Implicit is more challenging.
+
+We pay more attention to the reference to a single object in the context of REC. During the generation process, we take precautions to avoid text ambiguity, ensuring that the implicit text describes only one object in the image. To achieve this, we
+
+prioritize maximizing the number of attributes involved in attribute comparisons. For example, if the target is in a group of objects that share the same color only, we must ensure no other group of objects that can satisfy the same color and material. In cases where such a group of objects exists, we will change the target and modify the implicit text to describe the corresponding number of attributes until no additional attributes can be included. More details of the implicit text generation are shown in Appendix A.
+
+The implicit text in CLEVR-Implicit follows a standardized format based on several templates and rules. Especially, it is common for the same sentence to describe different target objects across different images. By employing homogeneous text, we purposefully direct the model's focus toward the image itself, thereby effectively testing the model's reasoning ability.
+
+# 3.2 TIE Method
+
+# 3.2.1 TIE Overview
+
+To address the challenge of implicit text comprehension, we propose a method called Transform Implicit text to Explicit text (TIE). The objective of TIE is to incorporate explicit words into the original implicit text, enabling the model to reason more effectively. Our approach consists of two key modules: (1) The prompt design module builds prompts for implicit text by adding masked tokens. By strategically designing the prompts, we prepare the groundwork for the next module. (2)
+
+The cloze procedure module uses a pre-trained vision-language model to predict the explicit words with the implicit prompts.
+
+By leveraging these two modules, TIE enables the integration of explicit words into the implicit text and facilitates indirect reasoning on the implicit text effectively.
+
+# 3.2.2 Prompt Design Module
+
+The prompt design module converts the implicit text into a format suitable for the Masked Language Modeling (MLM) task. One approach to adding explicit words is by obtaining the specific attribute values involved in the comparison of attributes in the implicit text (e.g., "What is the color in 'The left one of the only two objects with the same color'?"). Consequently, the prompt design module generates the prompt like "The left one of the only two objects with the same [MASK] color" for an implicit text sample. It's important to note that the conversion process differs between the "same" and "different" types of implicit text. In the "same" type, only one identical attribute value needs to be predicted, while in the "different" type, two different attribute values must be predicted. Therefore, in the "different" type, two masked tokens are added before each attribute in the implicit text. This process of converting the implicit text can be seen as a translation task, where the source text is the original implicit text, and the target text is the input format required for the MLM task. To accomplish this conversion process, we create an implicit prompt dataset using the implicit text from CLEVR-Implicit as the source text, and the target text is a text with the masked token preceding each attribute in the implicit text. Subsequently, we train an encoder-decoder model (T5, Raffel et al., 2020) on the implicit prompt dataset to add the appropriate number of masked tokens corresponding to the attributes for each sample in CLEVR-Implicit.
+
+# 3.2.3 Cloze Procedure Module
+
+The training objective of the cloze procedure module aligns with the MLM task, a commonly employed pre-training task in various pre-trained models, so another cloze dataset is devised for the MLM training stage. In this dataset, the input consists of implicit text containing masked tokens and an image description, while the target comprises explicit text containing the ground-truth attribute values. In this module, we fine-tune a pre-trained vision-language model ViLT to predict a specific word
+
+within the masked tokens. And we will introduce three prediction strategies: Base, Multi-step, and Path-select.
+
+Base In the prediction stage, the prediction of multiple attribute values in a sample is different from the prediction of a single attribute value. For implicit text containing multiple attributes, a straightforward prediction strategy is to predict all attribute values. However, this strategy tends to decrease the correlation among the predicted values, resulting in potential discrepancies where multiple predicted values may not refer to the same group of objects. Consequently, incorrect prediction results may arise.
+
+Multi-step To address the aforementioned issue, we propose an enhanced strategy named multi-step prediction. Specifically, when predicting a sample with multiple attributes, the model outputs only one attribute value at a time. Subsequently, it incorporates the previous prediction as input to predict the next attribute value. Compared with the method of predicting all attribute values at one time, this strategy enhances the correlation among multiple attribute values. For instance, if the original text is "The left one of the only two objects with the same [MASK] color and [MASK] material," the corresponding first-step output is adjusted to "The left one of the only two objects with the same red color and [MASK] material." And the second-step output is "The left one of the only two objects with the same red color and metal material."
+
+Path-select We find that the Multi-step strategy has a potential challenge where the accuracy of predictions relies on the first attribute. If the first attribute prediction is incorrect, it might trigger a chain reaction leading to subsequent attributes being errors. To mitigate this issue, we search for some objects that are not the targets as negative samples. They are characterized by partially satisfying the constraints of implicit text. Part of the attributes that do not satisfy are marked [NONE]. For instance, for a target sample "The same red color and rubber material" in cloze dataset, the corresponding negative sample might be "The same blue color and [NONE] material".
+
+During the prediction stage, when the model obtains the optimal Softmax result for an attribute and predicts [NONE] for the next attribute, it recognizes this as an incorrect predicted path. Subsequently, the model reverts to the prediction of the preceding attribute, considering the suboptimal
+
+<table><tr><td rowspan="2">ViLT (Kim et al., 2021) MLM</td><td colspan="2">Same</td><td colspan="2">Different</td><td colspan="2">Total</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>Base</td><td>80.54</td><td>81.97</td><td>25.73</td><td>23.35</td><td>61.99</td><td>60.48</td></tr><tr><td>Multi-step</td><td>92.55</td><td>91.42</td><td>87.05</td><td>85.47</td><td>91.45</td><td>91.50</td></tr><tr><td>Path-select</td><td>91.76</td><td>91.35</td><td>85.58</td><td>87.50</td><td>91.48</td><td>90.37</td></tr></table>
+
+Table 3: The TIE experiment results on CLEVR-Implicit. We use three different prediction strategies to compare the performance of ViLT.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Before TIE</td><td colspan="2">After TIE</td><td colspan="2">Ground-truth</td></tr><tr><td>val</td><td>test</td><td>val</td><td>test</td><td>val</td><td>test</td></tr><tr><td>MDETR (Kamath et al., 2021)</td><td>53.57</td><td>53.89</td><td>91.51</td><td>92.19</td><td>96.01</td><td>95.66</td></tr><tr><td>VLTVG (Yang et al., 2022)</td><td>58.83</td><td>58.46</td><td>84.48</td><td>86.15</td><td>95.97</td><td>96.85</td></tr></table>
+
+Table 4: Referring Expression Comprehension results performed by MDETR and VLTVG. Before TIE means that the models are trained on the implicit text before the input transformation of the TIE method, and After TIE means that the models are trained on the explicit text after TIE.
+
+Softmax result, which guides the model toward the correct prediction.
+
+# 4 Experiments
+
+# 4.1 Implementation Details
+
+For the REC task on the CLEVR-Implicit dataset, we employ two models: MDETR (Kamath et al., 2021) and VLTVG (Yang et al., 2022). MDETR utilizes weights pre-trained on various datasets such as COCO (RefCOCO, RefCOCO+, RefCOCOg), Visual Genome (Krishna et al., 2017), and Flickr30k (Plummer et al., 2015). and VLTVG uses weights pre-trained on RefCOCOg. And we fine-tune them on the CLEVR-Implicit dataset. During the training stage, we use the settings of learning rate 1e-4, learning rate of backbone 1e-4, batch size 16 and weight decay 1e-5 to train for 60 epochs. In the input transformation experiment, we utilize the T5-small model for the prompt design module. As for the cloze procedure module, we select ViLT (Kim et al., 2021). To ensure a fair comparison, all different prediction methods are configured with the AdamW optimizer, a learning rate of 1e-4, a batch size of 32, and an image size of  $384 \times 384$  to train for 70 epochs.
+
+# 4.2 Results of TIE
+
+Table 3 presents the results of the TIE experiment. The evaluation method is to assess whether the predicted explicit words match the corresponding ground-truth values. For samples containing multiple attributes, it is necessary to ensure the accurate prediction of all explicit words. otherwise, it is considered incorrect. In the case of "different" type
+
+samples, where attribute values differ between objects, the order of the two predicted values for a single attribute can be reversed.
+
+We present the results separately for training on the "same" and "different" types of samples, and the "Total" column represents the combined performance on both types of implicit text. In the Base strategy, we observe that "different" samples are more challenging to understand compared to "same" samples, resulting in a decrease in performance of up to  $54.81\%$ . For "different" samples, the model needs to output two values to predict a single attribute, resulting in twice as many explicit words to be predicted compared to "same" samples. In contrast, the Multi-step strategy exhibits significant improvements for both types of text. It shows a  $12.01\%$  increase in performance for "same" samples and a remarkable  $61.32\%$  increase for "different" samples. A reasonable explanation is that the prediction of each explicit word takes into account all previously predicted results. Multi-step strategy mitigates cases where the initial predictions are correct, but subsequent predictions suffer from errors, leading to more accurate overall predictions.
+
+From the comparing results of the Path-select and Multi-step strategies, we observe some differences when incorporating negative samples into the training set. The Path-select results on the validation set exhibit a slight decrease of  $0.79\%$  and  $1.47\%$  compared to Multi-step. We speculate that this is because the majority of incorrect positions in wrongly predicted samples tend to occur in the second attribute or later, and the prediction of the first attribute in these samples is generally accurate.
+
+![](images/2db1b42fe41556b45219dc3e907eabb7e8a3688e217c282c58a436445ad816fd.jpg)  
+(a) same attributes
+
+![](images/224046e4ef00d24df1e3c70f69b53067799e18f96600cad47dd810288363c7ca.jpg)  
+(b) different attributes  
+Figure 3: 3(a) is the performance of three prediction strategies on "same" type samples in test set, and 3(b) is the performance of three prediction strategies on "different" type samples in test set.
+
+However, on the test set of "different" samples, Path-select surpasses Multi-step by  $2.03\%$  in terms of performance. Particularly, the performance of Multi-step and Path-select in the "Total" column is comparable, indicating that the addition of negative samples can enhance the model's robustness. And negative samples enable the model to choose the correct path when the first attribute prediction is incorrect.
+
+# 4.3 Results of Referring Expression
+
+Table 4 presents the results of the referring expression experiment. We evaluate the performance of MDETR and VLTVG, both before and after applying the TIE based on the Multi-step prediction strategy. Additionally, we provide the performance of the models trained on ground-truth explicit samples as a reference. The evaluation metric is Intersection over Union (IoU), with a threshold value of 0.5. We observe that prior to TIE, the performance of both models trained on implicit text is relatively poor, indicating limited reasoning ability when dealing with implicit text. However, after applying TIE, there is a significant improvement in performance for both models (53.57% vs. 91.51% for MDETR and 58.83% vs. 84.48% for VLTVG). However, compared with Ground-truth, there is a certain gap in performance, because it is limited by the impact of TIE performance.
+
+# 4.4 Results of Contrast Experiment
+
+Difficulty of CLEVR-Implicit In order to contextualize the difficulty of CLEVR-Implicit, we evaluate the performance of REC models on CLEVR-Implicit when the models are trained on CLEVR-
+
+<table><tr><td>Method</td><td colspan="2">Train on CLEVR-Ref+</td></tr><tr><td></td><td>val</td><td>test</td></tr><tr><td>MDETR</td><td>47.02</td><td>45.85</td></tr><tr><td>VLTVG</td><td>47.99</td><td>47.11</td></tr></table>
+
+Table 5: The performances of REC models which are trained on CLEVR-Ref+ and tested on CLEVR-Implicit.  
+
+<table><tr><td>VLTVG on CLEVR-Ref+</td><td>val</td><td>test</td></tr><tr><td>Before TIE</td><td>66.43</td><td>65.22</td></tr><tr><td>After TIE</td><td>68.05</td><td>66.46</td></tr></table>
+
+Table 6: The performances of VLTVG respectively on original CLEVR-Ref+ (Before TIE) and CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit.
+
+Ref+. Table 5 shows that the results of MDETR and VLTVG models are all lower than "Before TIE" Column results in Table4. That indicates that the experience that the models learn from CLEVR-Ref+ cannot help them process the implicit text, because the samples in CLEVR-Ref+ contain explicit information to help the model locate the target, while the samples in CLEVR-Implicit hide the information. Therefore, we believe that it's challenging to process the implicit text in our CLEVR-Implicit dataset.
+
+Affect of TIE on explicit text Since the TIE method converts implicit text into explicit text, in addition to implicit text, we also need to analyze whether the explicit text after applying TIE will affect the performance of REC models. To verify this problem, we evaluate the performance of VLTVG respectively on original CLEVR-Ref+ and
+
+![](images/4542ea89e2f3d33f5e3496e8ed38951ea0d330ac3f1ed3ca770175e98686f4b3.jpg)  
+(a) An example of the comparison from Base, Multi-step and Path-select prediction strategies
+
+O: The left one of the only two objects with the same shape, color and size in the image
+
+P: The left one of the only two objects with the same [MASK] shape, [MASK] color and [MASK] size in the image
+
+GT: cylinder, yellow, large
+
+Base: sphere, brown, large
+
+Multi-step: cylinder, yellow, large
+
+Path-select: cylinder, yellow, large
+
+![](images/3fcf88c1bbcfbfbf23cca254074f611ca8b029de58cead01d3969871479acb9e.jpg)  
+(b) REC before TIE  
+Before: The left one of the only two objects with the same shape, color and size in the image  
+Figure 4: Examples from TIE and referring expression comprehension experiments. (a) an example of the comparison from Base, Multi-step and Path-select prediction strategies. O, P and GT denote original implicit text, generated prompt text and ground-truth explicit words. (b) Visualization of the predicted bounding box from VLTVG before TIE and (c) after TIE. green and red rectangles denote ground-truth and predicted bounding box.
+
+![](images/19825c8cb9209c6bbc9cd197168e16b2754dfc0c3c097586893372296c1a22ec.jpg)  
+(c) REC after TIE  
+After: The left one of the only two objects with the same cylinder shape, yellow color and large size in the image
+
+CLEVR-Ref+ after inference by TIE trained on CLEVR-Implicit. Table 6 shows that TIE does not affect performance (even better on VLTVG) on other types of referring expressions besides the implicit text, and TIE is robust in a broader class of referring expressions.
+
+Contrast between explicit and implicit As shown in Table 7, we evaluate the performance of MDETR and VLTVG models separately on predicted explicit text and implicit text when the models are trained on implicit text and ground-truth explicit text respectively.
+
+When both validation sets provide predicted explicit words (Column 1 vs Column 3), the performance of the model trained on the implicit training set is significantly decreased (MDETR:  $31.48\%$  vs  $91.98\%$ , VLTVG:  $53.85\%$  vs  $85.90\%$ ) compared with that on the explicit training set, which is evidence that explicit words play a crucial role in the models' comprehension and reasoning processes. The same situation occurs when exchanging implicit and explicit states on train and validation sets (Column 2 vs Column 3).
+
+# 4.5 Performance on different number of attributes
+
+Figure 3 illustrates the performance of ViLT on "same" and "different" samples with different numbers of attributes. In Figure 3(a), we observe that the model achieves the highest performance when predicting one-attribute samples, while the lowest performance is observed when predicting three-
+
+<table><tr><td>Method</td><td>Train-i Val-pe</td><td>Train-ge Val-i</td><td>Train-ge Val-pe</td></tr><tr><td>MDETR</td><td>31.48</td><td>46.25</td><td>91.98</td></tr><tr><td>VLTVG</td><td>53.85</td><td>48.34</td><td>85.90</td></tr></table>
+
+Table 7: The contrast experiment for REC. The i suffix denotes the set with implicit samples. The pe suffix denotes that the explicit words in samples are predicted by TIE. The ge denotes the explicit words in samples is ground-truth.
+
+attribute samples. Interestingly, the performance on four-attribute samples is second only to that of one-attribute samples. A reasonable conjecture is that a set of objects is identical when all four properties are the same in CLEVR images, so the model only needs to find two objects that are identical. Figure 3(b) shows that predicting four-attribute "different" samples has the most significant challenge on the Base strategy. However, on the Multi-step strategy, the performance on four-attribute "different" samples surpasses that of three-attribute samples. This indicates that the Multi-step prediction strategy effectively mitigates the impact of an increased number of predicted words, resulting in improved performance.
+
+# 4.6 Case Study
+
+Figure 4 indicates an example of the comparison of different prediction strategies in the TIE experiment, and we visualize the ground-truth bounding box and the bounding box predicted from VLTVG
+
+Table 8: Examples about the application of implicit text in daily life.
+
+# Real-world Implicit Text Examples
+
+Two persons are doing the same movement.
+
+His height is different from each other.
+
+His pants and shoes are the same color.
+
+The same brand of cosmetics on the table.
+
+Take the drugs used for the same symptom.
+
+before and after TIE in the REC experiment. In Figure 4(a), with the attributes of the example "shape" and "color", the Base strategy mistakenly predicts "sphere" and "brown". We observe that from the image there is a group of objects that is both "sphere" and another group of objects that is both "brown", but there are no such two objects that satisfy "sphere" and "brown" simultaneously. This indicates that the Base strategy performs poorly on multi-attribute samples due to the weak correlation between attributes, leading to the error that different attributes refer to different groups of objects. In contrast, the Multi-step and Path-select strategies enhance the correlation between attributes and make predictions step-by-step, leading to more accurate predictions.
+
+In Figure 4(b), by taking the implicit text as input, VLTVG incorrectly predicts the referred object as a brown cube. While VLTVG predicts the referred object accurately when taking the explicit text after TIE as input. The comparison result indicates that current models are being confused when performing implicit reasoning based on the logical relation conveyed by the implicit text. Only when explicit words are explicitly included, the models can align the explicit words with the corresponding objects in the image for explicit reasoning.
+
+# 5 Implicit Text in Real-world Scenarios
+
+The implicit descriptions in real-world scenarios are more complex and difficult because they not only describe the targets implicitly through the geometrical attributes of color, material, etc., but also contain more attributes as well as external factors. Table 8 shows the implicit text examples in real-world daily life. It is important that even though our generated implicit text relies on the CLEVR images and the provided annotations, our proposed TIE method can still extend the scope beyond the CLEVR domain to real-world scenarios. For example, based on Figure 5 which is chosen from Visual
+
+![](images/fdb3d565039b1a5d940f5783edb3568056d3ebf3eae3adb2b64ad43b803ead53.jpg)  
+Original Implicit Text: The left one of two persons with the same posture and gender  
+Figure 5: A real-world image from Visual Genome
+
+After Prompt Design Module: The left one of two persons with the same [MASK] posture and [MASK] gender
+
+After Cloze Procedure Module: The left one of two persons with the same walking posture and woman gender
+
+Genome, consider the implicit text like "The left one of two persons with the same posture and gender". Through the TIE method, the prompt can be built before the "posture" and "gender" by prompt design module, like "The left one of two persons with the same [MASK] posture and [MASK] gender". The prompt guides the cloze procedure module to predict the explicit words "walking" and "woman" like "The left one of two persons with the same walking posture and woman gender". By integrating TIE approach, current models can also process the implicit text on the real-world scenarios.
+
+# 6 Conclusion
+
+In this paper, we analyze the significance of implicit reasoning, along with the lack of investigation in this domain. To address this, we introduce the CLEVR-Implicit dataset for implicit reasoning in the REC task. This dataset utilizes synthesized images to generate two types of implicit text: "same" and "different." Evaluating the performance of existing pre-trained models on CLEVR-Implicit, we observe model confusion when attempting to align implicit text with images. To overcome this challenge, we propose the Transform Implicit text to Explicit text (TIE) method, which converts implicit text into explicit text. Our experiments demonstrate that the reasoning capability of the models improves significantly through the alignment of explicit words and images.
+
+# Limitations
+
+The limitation of our work is that CLEVR-Implicit is built on synthetic images, so it's straightforward to generate corresponding implicit text based on the annotation of the synthetic images. Besides, since the synthetic images contain only geometric objects, our implicit text can only be generated within the scope of the object's attributes. Therefore, in future work, we will annotate more types of implicit text based on real-world scenarios. This will allow us to explore the mechanisms of model understanding of implicit text.
+
+# Acknowledgements
+
+This work was supported by the National Natural Science Foundation of China (62076100), Fundamental Research Funds for the Central Universities, SCUT (x2rjD2230080), the Science and Technology Planning Project of Guangdong Province (2020B0101100002), CAAI-Huawei MindSpore Open Fund, CCF-Zhipu AI Large Model Fund.
+
+# References
+
+Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.  
+Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. 2017. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910.  
+Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. 2021. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1780-1790.  
+Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR.  
+Emiel Krahmer and Kees Van Deemter. 2012. Computational generation of referring expressions: A survey. Computational Linguistics, 38(1):173-218.
+
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123(1):32-73.  
+Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pretrain, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35.  
+Runtao Liu, Chenxi Liu, Yutong Bai, and Alan L Yuille. 2019. Clevr-ref+: Diagnosing visual reasoning with referring expressions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4185-4194.  
+Yuhang Liu, Wei Wei, Daowan Peng, and Feida Zhu. 2022. Declaration-based prompt tuning for visual question answering. arXiv preprint arXiv:2205.02456.  
+Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32.  
+Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. 2016. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20.  
+Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. 2013. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184.  
+Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE international conference on computer vision, pages 2641-2649.  
+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.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.
+
+Ehud Reiter and Robert Dale. 1992. A fast algorithm for the generation of referring expressions. In Proceedings of the 14th Conference on Computational Linguistics - Volume 1, page 232-238. Association for Computational Linguistics.
+
+Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. 2019. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530.
+
+Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. 2016. Show and tell: Lessons learned from the 2015 mscoco image captioning challenge. IEEE transactions on pattern analysis and machine intelligence, 39(4):652-663.
+
+Li Yang, Yan Xu, Chunfeng Yuan, Wei Liu, Bing Li, and Weiming Hu. 2022. Improving visual grounding with visual-linguistic verification and iterative reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9499-9508.
+
+Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2021. Cpt: Colorful prompt tuning for pre-trained vision-language models. arXiv preprint arXiv:2109.11797.
+
+Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. 2018. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1307-1315.
+
+Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. 2016. Modeling context in referring expressions. In European Conference on Computer Vision, pages 69-85. Springer.
+
+Licheng Yu, Hao Tan, Mohit Bansal, and Tamara L Berg. 2017. A joint speaker-listener-reinforcer model for referring expressions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7282-7290.
+
+# A More Details of Implicit Text Generation
+
+To clarify on our implicit referring expression generation process, we'd like to elaborate on the method used to create the CLEVR-Implicit dataset from the existing CLEVR images and their associated annotations. These annotations contain object attributes, as well as spatial relationships between objects. For each image, we will traverse all the objects in the image by comparing whether the attribute values between two objects are the same/different each time. For example, if an image contains 4 objects, this leads to  $\mathrm{C}(4,2) = 6$  object pair comparisons (e.g., (1,2), (1,3), (1,4), (2,3),
+
+(2,4), (3,4), the numbers are object ids). We will generate the corresponding implicit text if a certain pair satisfies the "same" or "different" condition. Suppose the (2,3) combination objects' shape and color are both cube and blue, and guarantee that no other pairs satisfy the condition "same shape and color" to avoid the ambiguity in the REC task, then we will generate the "same" type of implicit text as "The two objects with the same shape and color". Due to the only one target of the REC task, we must choose one of the (2,3) objects to be the target. We employ a random selection for the target based on the spatial relationship between 2 and 3. If object 2 is left of object 3 and object 2 is selected, the finally generated implicit text is "The left one of the two objects with the same shape and color". The "different" types of text are generated in the same way. This consistent pattern of implicit text generation in the CLEVR-Implicit dataset ensures both the concealment of explicit information and the reference of target objects in the images.

2023/CLEVR-Implicit_ A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00893b209ff345213da03613cde08036d08eefc6e510034524693cad3fb4e227
+size 285470

2023/COFFEE_ Counterfactual Fairness for Personalized Text Generation in Explainable Recommendation/495d6ac8-cce4-4b58-8d12-6baca0a4fa1d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:835866edb126cb173c647bfb9fc4b3e493bccb2b10bd05100cb3fc95d2467b07
+size 5369386

2023/COFFEE_ Counterfactual Fairness for Personalized Text Generation in Explainable Recommendation/full.md

@@ -0,0 +1,435 @@
+# COFFEE: Counterfactual Fairness for Personalized Text Generation in Explainable Recommendation
+
+Nan Wang $^{1,3}$ , Qifan Wang $^{2}$ , Yi-Chia Wang $^{2}$ , Maziar Sanjabi $^{2}$ , Jingzhou Liu $^{2}$ , Hamed Firooz $^{2}$ , Hongning Wang $^{3}$ , Shaoliang Nie $^{2}$
+
+<sup>1</sup>Netflix Inc., Los Gatos, California, USA
+
+$^{2}$ Meta AI, Menlo Park, CA, USA
+
+<sup>3</sup>University of Virginia, VA, USA
+
+# Abstract
+
+As language models become increasingly integrated into our digital lives, Personalized Text Generation (PTG) has emerged as a pivotal component with a wide range of applications. However, the bias inherent in user written text, often used for PTG model training, can inadvertently associate different levels of linguistic quality with users' protected attributes. The model can inherit the bias and perpetuate inequality in generating text w.r.t. users' protected attributes, leading to unfair treatment when serving users. In this work, we investigate fairness of PTG in the context of personalized explanation generation for recommendations. We first discuss the biases in generated explanations and their fairness implications. To promote fairness, we introduce a general framework to achieve measure-specific counterfactual fairness in explanation generation. Extensive experiments and human evaluations demonstrate the effectiveness of our method.
+
+# 1 Introduction
+
+Personalized text generation (PTG) has extensive applications, such as explainable recommendation (Zhang and Chen, 2020; Chen et al., 2021), post generation (Yuan and Huang, 2019; He et al., 2021), and conversational systems (Zhang et al., 2018, 2019; Lee et al., 2021). The auto-generated text, functioning at the frontier of human-machine interaction, influences users' decisions and transforms their way of thinking and behaving. However, due to its immense power and wide reach, PTG can inadvertently give rise to fairness issues and lead to unintended consequences (Alim et al., 2016; Borda and Bowman, 2019; Blodgett et al., 2020).
+
+In this work, we investigate the fairness issues in PTG, focusing on one of the mostly studied settings: generating natural language explanations for
+
+![](images/52d637406bb2d62133bfe8ea5b85db8bbb9feaaf8ae3dc1a6d1661682db7809e.jpg)  
+Figure 1: Left: average FeatCov of ground-truth explanations in the train set. Right: average FeatCov of explanations generated on the test set by PETER.
+
+![](images/405b77173e59142a97f6badbb193789d50b824ea194cd11fa9cb5d4ea8ef6f48.jpg)
+
+recommendations (Wang et al., 2018; Chen et al., 2021; Yang et al., 2021; Li et al., 2021a; Yang et al., 2022). Personalized explanation generation aims to provide a user with descriptive paragraphs on recommended items that align with his/her preferences, enabling more informative and accurate decision-making. The generators are typically language models trained on user written reviews from e-commerce platforms (Zhang et al., 2018; Ni et al., 2019; Yang et al., 2021), where sentences related to item descriptions are retained to construct the ground-truth explanations. However, due to historical, social, or behavioral reasons, inherent bias may exist within the review text, associating specific linguistic characteristics with the users' protected attributes such as gender or race (Newman et al., 2008; Alim et al., 2016; Volz et al., 2020). While certain linguistic features that capture the diversity of language use (Newman et al., 2008; Groenwold et al., 2020) are suitable for personalization, others pertaining to the linguistic quality of explanations, such as informativeness or detailedness, (Louis, 2013), should be excluded. Failure to do so can result in unfair treatment when serving users.
+
+As an example, we investigate the explanation generation on Amazon Movies<sup>1</sup> with the personalized transformer model PETER (Li et al., 2021a). We adopt feature coverage (FeatCov, the number of unique features mentioned about a movie) as an automatically measurable metric of explanation
+
+![](images/be9016a9baecdd7f19dbf236f5f3797fb93a1044f9fcc70a6112d1ffbed295fe.jpg)  
+Figure 2: FeatCov vs. human evaluated quality measures. Each measure is rated in a five-category scale.  $p$ -value  $< 0.05$  in Kruskal-Wallis H test for all three criteria (see Appendix D for detailed test results).
+
+quality. As shown in Figure 1 (left), reviews from male users generally have higher FeatCov on the target movies than those from female users (Fan-Osuala, 2023) $^{2}$ . A model trained on such data can inherit the bias and generate explanations discriminately when serving users—higher FeatCov for males than females, as indicated in Figure 1 (right). To further substantiate the bias issue, we conducted a user study and found strong positive correlation, shown in Figure 2, between FeatCov and human evaluated quality criteria including informativeness, detailedness and helpfulness. The observations remain consistent when results from male and female human evaluators are analyzed separately, affirming that both genders consider FeatCov a significant indicator of quality. This study demonstrates the bias in FeatCov as a concerning issue for PTG. Details of the user study are in Section 7.
+
+In particular, the bias observed in training data originates from various factors such as different demographic behavior patterns that are beyond our control. But the system should not inherit the bias and act discriminately by generating explanations of different quality for different users—the lack of informative reviews from users of a demographic group should not prevent them from receiving informative explanations. Without proper intervention, the system can exhibit such bias, thus adversarily affecting the users' experience, reliance and trust in the system (Tintarev and Masthoff, 2015). From a broader perspective, the bias can reinforce itself by influencing how users write online, and jeopardize fairness in the long run (Schwartz et al., 2016; Alim et al., 2016; Cordia and Bowman, 2019).
+
+To mitigate the issue, we take a causal perspective to dissect the bias and enforce counterfactual fairness (CF) (Kusner et al., 2017) in personalized explanation generation: the quality of generated
+
+explanations should not differentiate a user in the real world vs. in the counterfactual world where only the user's protected attribute (e.g., gender) is changed. This problem is essential and unique compared to fairness problems in the literature (Section 2), and imposes specific technical challenges (Section 4). To achieve the goal, we develop a general framework, COFFEE, for COunterFactual FairnEss in Explanation generation. COFFEE treats a user's protected attribute value as a separate token input to the model, and disentangles its representation from the user's representation (Ma et al., 2019; Locatello et al., 2019b; Zheng et al., 2021). By controlling the input attribute values for counterfactual inference (CI), we impose a measure-specific CF constraint (Russell et al., 2017) on generated explanations. Then a novel fair policy learning scheme is developed to optimize CF with carefully designed rewards, which generalizes to any single or combination of quality measures. We use user-side fairness by default in discussions, but COFFEE is general to ensure fairness on either user or item side in the two-sided market (Wang and Joachims, 2021), and be adapted to different models. Extensive experiments and rigorous user studies demonstrate COFFEE's superiority in achieving fairness and maintain high generation performance compared to baselines.
+
+# 2 Background
+
+Uniqueness and Importance: Fairness in machine learning (ML) is originally studied in automatic decision-making systems that directly impose "significant" or "legal" effects on individuals (Voigt and Bussche, 2017). Such fairness considerations often revolve around resource allocation by ML models, exemplified in contexts like loan assessments (Lee and Floridi, 2021) or job applications (Singh and Joachims, 2018), where model predictions can unfavorably affect a protected group (Du et al., 2021; Mehrabi et al., 2021).
+
+In contrast to resource allocation fairness, NLP researchers primarily examine the representational fairness (Blodgett et al., 2020; Liang et al., 2021; Sheng et al., 2021) regarding how language models shape social biases and stereotypes through natural language understanding (NLU) or generation (NLG). In the realm of NLG, fairness particularly concerns how generated text may contain biased information about a specific demographic group. For instance, Huang et al. (2020) analyze the sentence completion by a GPT-2 model, and find dif
+
+ferent sentiment distributions of completed sentences when the occupation word is counterfactually changed in the prompts. Bordia and Bowman (2019) revealed a more frequent co-occurrence of the 'doctor' with male pronouns and 'nurse' with female pronouns in generated text. However, these biases, directly encapsulated within the text, can be more easily analyzed. To the best of our knowledge, our work is pioneering in exploring how personalization, bridging NLP and recommendation, associates the bias in NLG with protected attributes of users.
+
+Fairness Notions: Various fairness notions exist in the literature, with group-wise fairness notions being the firstly studied ones (Zafar et al., 2015; Hardt et al., 2016; Zafar et al., 2017). Yet, group-wise fairness has different quantitative definitions that are generally incompatible (Kleinberg et al., 2017; Berk et al., 2021). Some definitions can even exacerbate discrimination (Kusner et al., 2017). Individual fairness (Zemel et al., 2013; Joseph et al., 2016) requires similar users to receive similar predictions. But it relies on carefully chosen domain-specific similarity metrics (Dwork et al., 2012). In contrast, counterfactual fairness (CF) (Kusner et al., 2017), considering fairness from a causal perspective, has gained prominence recently as a more robust fairness notion (Russell et al., 2017; Wu et al., 2019; Makhlouf et al., 2020), which can also enhance group-wise fairness in certain scenarios (Zhang and Baireinboim, 2018; Khademi et al., 2019).
+
+Though CF has been studied in some non-personalized NLP tasks (Huang et al., 2020; Garg et al., 2019), most existing works study the dependency of model outputs on attribute-specific words within the input text (Blodgett et al., 2020; Liang et al., 2021; Sheng et al., 2021). In such cases, CI can be easily performed on the input text itself, such as changing male pronouns to female pronouns (Huang et al., 2020; Garg et al., 2019). However, CF in PTG necessitates CI on the protected attributes of users being served-an area yet to be thoroughly explored.
+
+# 3 Problem Formulation
+
+In the following discussions, we consider a single protected attribute on the user side for simplicity, but our proposed framework is versatile to accommodate multiple attributes on either the user or the item side. The value of a user's protected attribute is denoted by a variable  $A \in \mathcal{A}$ ,
+
+where  $\mathcal{A}$  is the set of possible attribute values, e.g.,  $\mathcal{A} = \{\text{male, female, other}\}$  for gender. Each dataset entry is a tuple of  $(u,i,a,e)$ , corresponding to user ID, item ID, observed attribute value, and ground-truth explanation. The explanation generator  $G_{\theta}$  is a language model parameterized by  $\theta$ . Given a user  $u$ , an item  $i$ , and observed attribute value  $a$ , an explanation can be sampled from the generator as  $Y \sim G_{\theta}(u,i|A = a)$ . The linguistic quality of any explanation  $y$  is measured by a function  $Q(y)$ . Notably, we treat  $Q$  as a black box oracle—a quality measure that can only be queried. This is essential in practice and offers the flexibility to arbitrarily tailor  $Q$  based on the fairness requirements of the application. An explanation can be gauged in various ways by customizing  $Q$ , such as using an explicit function, human evaluation, or a tool provided by authorities. We assume, without loss of generality, that higher  $Q$  values represent superior quality. CF on any measure  $Q$  of explanations is achieved when, given a user  $u$  and an item  $i$ ,
+
+$$
+P \left(Q \left(Y _ {A \leftarrow a}\right) \mid u, i, a\right) = P \left(Q \left(Y _ {A \leftarrow a ^ {\prime}}\right) \mid u, i, a\right), \tag {1}
+$$
+
+where  $Y_{A\leftarrow a'} \sim G_{\theta}(u, i | A = a')$  is the explanation generated when we counterfactually assign the value of the user's protected attribute by  $A \leftarrow a', a' \neq a$ . The right side of Eq. (1) evaluates the quality distribution of explanations generated had the user's protected attribute value been  $a'$ , given that the observed attribute value is  $a$  (Kusner et al., 2017; Li et al., 2021b).
+
+Denote the loss of the generator for a given user  $u$  and item  $i$  by  $\mathcal{L}_{gen}(G_{\theta}(u,i|A = a),e)$ , which is typically the negative log-likelihood (NLL) loss or a combination of several losses (Li et al., 2017; Yang et al., 2021; Li et al., 2021a). We consider training the generator for fair explanation generation as a constrained optimization problem:
+
+$$
+\begin{array}{l} \min  \mathcal {L} _ {g e n} (G _ {\theta} (u, i | A = a), e) \\ s. t. \mathbb {E} _ {Y _ {A \leftarrow a}} [ Q (Y _ {A \leftarrow a}) | u, i, a ] = \tag {2} \\ \mathbb {E} _ {Y _ {A \leftarrow a ^ {\prime}}} [ Q (Y _ {A \leftarrow a ^ {\prime}}) | u, i, a ] \\ \end{array}
+$$
+
+For ease of presentation, we consider a single user-item pair, and the total loss on a dataset is simply summed over all user-item pairs with the constraint applied to every pair. In this work, we apply the first-order moment of the quality of generated explanations to construct the constraint, and leave the extension to other moment-matching constraints for future work. We further simplify the expression of the constraint as  $\mathbb{E}[Q(Y_{A\leftarrow a})] = \mathbb{E}[Q(Y_{A\leftarrow a^{\prime}})]$
+
+![](images/b6c701d377fc40fefad681a579501fe8bb9ec79cac946ae71b5a1f18497803ce.jpg)  
+Figure 3: Given a user with attribute value  $a$  and a recommended item, COFFEE performs CI by switching the attribute value to  $a'$  to get the counterfactual user, which is achieved by disentangled attribute embeddings. Explanations  $Y$  for both the real user and the counterfactual user are sampled from the generator, evaluated by  $Q$ . COFFEE then updates the generator's parameters  $\theta$  by policy learning from the fairness constraint.
+
+# 4 COFFEE
+
+User and item IDs are conventionally input to the generator for explanation generation, represented as learned embedding vectors (Li et al., 2017; Wang et al., 2018; Li et al., 2021a). However, the user's protected attribute, entangled with their preference, is implicitly encoded in the representations (Locatello et al., 2019a), hindering its direct manipulation for CI. One way explored for CF in personalization involves removing all information about the protected attribute in representations via discriminators (Zemel et al., 2013; Li et al., 2021b; Wang et al., 2022). Although this improves fairness, it detrimentally affects personalization and eliminates desired characteristics linked to the protected attribute. In contrast, we aim to enforce the independence of the protected attribute from any specified quality measure  $Q$  on generated explanations, while preserving explanation content-dependence to sustain high personalization performance.
+
+To enable CI, COFFEE considers a user's protected attribute value as a separate token input to the model, along with user and item IDs, and learns disentangled attribute representations from the user representation to encapsulate the effect of each attribute value in explanation generation (Locatello et al., 2019a,b; Ma et al., 2019; Zheng et al., 2021). This mirrors methods in controllable text generation (Oraby et al., 2018; Shu et al., 2020; Dathathri et al., 2020), where disentangled attribute tokens are utilized as inputs to modulate the topic, sentiment, or style of generated text. CI is then achieved by altering the attribute token input to the model. Subsequently, we enforce a fairness constraint based on the explanations generated pre and post CI, and establish a policy learning method for optimizing this constraint. An illustration of the COFFEE framework is shown in Figure 3.
+
+# 4.1 Disentangled Attribute Representation
+
+For a given tuple  $(u,i,a)$ , we denote the representation for the attribute value  $a$  as  $\mathbf{r}_a$ , the user's preference representation (independent from the protected attribute) as  $\mathbf{r}_u$  and item representation as  $\mathbf{r}_i$ . The complete user representation is  $\mathbf{r}_u^a = [\mathbf{r}_a,\mathbf{r}_u]$ . Correspondingly, when performing  $A\gets a'$  on user  $u$ , we change the input attribute token from  $a$  to  $a'$ , and the new user representation becomes  $\mathbf{r}_u^{A\leftarrow a'} = [\mathbf{r}_a',\mathbf{r}_u]$ . Note that each attribute value has its own representation, and is shared across all users having that same attribute value. For instance, all male users' attribute representation is the same vector  $\mathbf{r}_{male}$ . We can do the same for item-side attributes as  $\mathbf{r}_i^a = [\mathbf{r}_a,\mathbf{r}_i]$ .<sup>3</sup>
+
+Simply separating the user's protected attribute and preference representations does not guarantee that  $\mathbf{r}_u$  will not contain any information about the protected attribute, inhibiting the accuracy of CI. To further enforce the disentanglement, we introduce a discriminator  $D(\mathbf{r}_u)$ , and add an adversarial loss on  $\mathbf{r}_u$  in Eq. (2) as
+
+$$
+\begin{array}{l} \min  \mathcal {L} _ {\text {g e n}} \left(G _ {\theta} (u, i | A = a), e\right) + \lambda_ {D} \log \left(D \left(r _ {u}, a\right)\right) \\ s. t. \quad \mathbb {E} [ Q \left(Y _ {A \leftarrow a}\right) ] = \mathbb {E} [ Q \left(Y _ {A \leftarrow a ^ {\prime}}\right) ], \forall a ^ {\prime} \in \mathcal {A}, a ^ {\prime} \neq a, \end{array} \tag {3}
+$$
+
+where  $D(r_u,a)$  is the probability of predicting the correct attribute value  $a$ . In this way, we adversarially remove the protected attribute information from  $\mathbf{r}_u$ , and enforce  $\mathbf{r}_a$  to capture all the attribute information. During mini-batch optimization, we alternate between the parameter updates of the model and the discriminator as follows: (1)  $X$  batches of updates minimizing the loss Eq. (3) with  $D$  fixed, and (2)  $Z$  batches of updates maximizing the loss Eq. (3) with the generator  $G_{\theta}$  fixed.
+
+# 4.2 Policy Learning for Fairness
+
+Once the user, item, and attribute representations are learned and fixed, we can optimize the generator w.r.t. the CF constraint. Due to the nonconvexity of the constrained fairness optimization problem, closed-form solutions are unattainable. We add the constraint with Lagrangian relaxation as a regularization in the loss (Russell et al., 2017)
+
+$$
+\left. \min  \mathcal {L} _ {a l l} + \lambda \left| \mathbb {E} \left[ Q \left(Y _ {A \leftarrow a}\right) \right] - \mathbb {E} \left[ Q \left(Y _ {A \leftarrow a ^ {\prime}}\right) \right] \right|, \right. \tag {4}
+$$
+
+with  $\mathcal{L}_{all} = \mathcal{L}_{gen}(G_{\theta}(u,i|A = a),e) + \lambda_D\log (D(r_u,a))$  in Eq. (3) denotes all other losses except for the CF constraint, and  $\lambda$  is a hyperparameter for fairness-utility trade-off.
+
+However, standard gradient methods cannot be directly applied to optimize the constraint: explanations are discretely sampled from the generator, and  $Q$  is an oracle. Instead, we consider policy learning for fairness optimization, where the explanation distribution imposed by the generator is considered as the policy for explanation generation, and sampled explanations are actions. Concretely, for estimating the expectations in the regularization, we sample  $N$  explanations and calculate the average quality of explanations sampled in both the real-world and the counterfactual world. Denote the regularization term as  $\mathcal{L}_{fair}$ , the expectations in the regularization can be estimated as:
+
+$$
+\begin{array}{l} \mathcal {L} _ {f a i r} = \left| \mathbb {E} \left[ Q \left(Y _ {A \leftarrow a}\right) \right] - \mathbb {E} \left[ Q \left(Y _ {A \leftarrow a ^ {\prime}}\right) \right] \right| \\ \approx \operatorname {s i g n} (\Delta) \left(\frac {1}{N} \sum_ {k = 1} ^ {N} Q \left(y _ {A \leftarrow a} ^ {k}\right) - \frac {1}{N} \sum_ {k = 1} ^ {N} Q \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right)\right), \tag {5} \\ \end{array}
+$$
+
+$\Delta = \frac{1}{N}\sum_{k = 1}^{N}Q\big(y_{A\leftarrow a}^{k}\big) - \frac{1}{N}\sum_{k = 1}^{N}Q\big(y_{A\leftarrow a^{\prime}}^{k}\big).$  For an explanation  $y_{A\leftarrow a}^{k}$  sampled in the realworld, its contribution to the unfairness in regularization term is thus  $\frac{1}{N}\mathrm{sign}(\Delta)Q\big(y_{A\leftarrow a}^{k}\big)$  . To improve fairness by minimizing the regularization, the reward for  $y_{A\leftarrow a}^{k}$  is considered as  $r\big(y_{A\leftarrow a}^{k}\big) =$ $-\frac{1}{N}\mathrm{sign}(\Delta)Q\big(y_{A\leftarrow a}^{k}\big)$  . Similarly, the reward for a sampled explanation  $y_{A\leftarrow a^{\prime}}^{k}$  in the counterfactual world is  $r\big(y_{A\leftarrow a^{\prime}}^{k}\big) = \frac{1}{N}\mathrm{sign}(\Delta)Q\big(y_{A\leftarrow a^{\prime}}^{k}\big)$  . In this way, we convert the CF optimization into maximizing the rewards of generated explanations.
+
+Although the rewards are designed to minimize the difference in the qualities of explanations from the real vs. counterfactual world, we lack direct control over how the difference should be minimized during optimization. The optimization may arbitrarily improve or decrease the quality of explanations to achieve fairness, e.g., always generating low quality explanations, which greatly hurt
+
+their utility in practice. To address this, we further design a weighting mechanism to calibrate the rewards such that the optimization focuses more on improving (or decreasing) the quality measure for an attribute value for achieving fairness. This empowers the designer to better control the optimization and the utility-fairness trade-off. Specifically, we introduce a quality promotion weight  $\eta \in [0,1]$  to re-weigh the rewards of explanations in the world with lower expected quality, and use  $1 - \eta$  to reweigh the rewards of explanations in the other world. The resulting calibrated rewards are:
+
+$$
+r _ {w} \left(y _ {A \leftarrow a} ^ {k}\right) = - \frac {1}{N} \operatorname {s i g n} (\Delta) Q \left(y _ {A \leftarrow a} ^ {k}\right) \cdot w (\Delta)
+$$
+
+$$
+r _ {w} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) = \frac {1}{N} \operatorname {s i g n} (\Delta) Q \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) \cdot (1 - w (\Delta))
+$$
+
+where  $w = \frac{\mathrm{sign}(\Delta) + 1}{2} (1 - \eta) + \left(1 - \frac{\mathrm{sign}(\Delta) + 1}{2}\right)\eta$ . We can leverage this weight to guide the fairness optimization: if  $\eta > 0.5$ , the algorithm will focus more on improving the quality of low-quality explanations for fairness. Finally, for stability and faster convergence, we apply the advantage trick (Mnih et al., 2016), and each reward is its difference from the average reward in its corresponding world:
+
+$$
+r _ {a d v} \left(y _ {A \leftarrow a} ^ {k}\right) = r _ {w} \left(y _ {A \leftarrow a} ^ {k}\right) - \bar {r} _ {w} \left(y _ {A \leftarrow a} ^ {k}\right) \tag {6}
+$$
+
+$$
+r _ {a d v} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) = r _ {w} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) - \bar {r} _ {w} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right)
+$$
+
+The policy gradient (Sutton et al., 1999; Yang et al., 2021) on  $\theta$  can then be estimated as:
+
+$$
+\begin{array}{l} \nabla_ {\theta} \mathcal {L} _ {f a i r} \approx \sum_ {k = 1} ^ {N} \nabla_ {\theta} \log G _ {\theta} \left(y _ {A \leftarrow a} ^ {k}\right) r _ {a d v} \left(y _ {A \leftarrow a} ^ {k}\right) \tag {7} \\ + \sum_ {k = 1} ^ {N} \nabla_ {\theta} \log G _ {\theta} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) r _ {a d v} \left(y _ {A \leftarrow a ^ {\prime}} ^ {k}\right) \\ \end{array}
+$$
+
+Intuitively, during optimization, the probability of sampled explanations that lead to the unfairness will be demoted, while the probability of those that contribute to fairness will be promoted.
+
+To train a COFFEE model, we first pre-train the model without the fairness constraint. Then we fix the latent representations of users, items and attributes, and add the fairness constraint in the objective function for fine-tuning the model.
+
+# 5 Applying COFFEE to Existing Models
+
+Existing models for personalized explanation generation mostly adopt either Transformer (Li et al., 2021a) or RNN (Li et al., 2017; Chen et al., 2021; Yang et al., 2021). In this section, we demonstrate the application of COFFEE to PETER (Li
+
+Table 1: Statistics of the datasets and the protected attributes used in experiments. The candidate values of the attributes are indicated in parentheses.  
+
+<table><tr><td></td><td>#Users</td><td>#Items</td><td>#Reviews</td><td>Attribute Values</td></tr><tr><td>Video Games</td><td>9,511</td><td>5,173</td><td>59,374</td><td>male, female</td></tr><tr><td>Movies &amp; TV</td><td>28,001</td><td>12,888</td><td>265,646</td><td>male, female</td></tr><tr><td>Yelp</td><td>35,714</td><td>24,900</td><td>1,499,291</td><td>$, $$, $$$$</td></tr></table>
+
+et al., 2021a), which is an explanation generation model based on personalized Transformers. In Appendix A, we also illustrate the application of COFFEE to NRT (Li et al., 2017), which is based on RNN. However, it is worth noting that COFFEE is a general framework that can be applied to most of existing explanation generation models, as long as the model is based on user/item representation learning and can be fine-tuned.
+
+In PETER (Li et al., 2021a), user and item IDs, represented as special tokens, are appended to the start of each explanation sequence before being inputted to transformer layers. A context prediction loss is introduced for enhanced personalization in addition to the NLL loss, which predicts the words in the explanations regardless of their order. We denote the overall loss as  $\mathcal{L}_{\text {peter }}$  , which is the sum of the NLL loss and context prediction loss (Li et al., 2021a). To apply COFFEE to PETER, we introduce an attribute token at the start of each sequence for disentangled attribute embedding. The discriminator is applied at the input embedding layer to enhance disentanglement. User, item, and attribute tokens can attend to each other, while the explanation words can only attend to past tokens. When applying COFFEE to PETER, the loss is constructed by replacing  $\mathcal{L}_{\text {gen }}$  in Eq. (3) by  $\mathcal{L}_{\text {peter }}$ .
+
+# 6 Experiments
+
+We conduct experiments on three public review datasets: Amazon Video Games, Amazon Movies & TV, and Yelp Restaurant, and compare COFFEE with multiple baseline models. Due to space limit, we present the results based on PETER, and put example explanations and results based on NRT in Appendix E.
+
+# 6.1 Experiment Setup
+
+Datasets. In Amazon Games and Amazon Movies review datasets<sup>4</sup>, we use users' binary gender (male, female) as the protected attribute. We study the fairness on the item side on the Yelp dataset<sup>5</sup>,
+
+where a restaurant's price range ( $,$ \ $,$ \ $\$ ) is used as the protected attribute, as the system may not discriminate a restaurant simply because the restaurant owner sets lower prices for equally good food and service than its counterparts. The statistics of the datasets are summarized in Table 1. The details of data processing are described in Appendix B.
+
+On both Amazon Games and Movies, the raw model tends to generate higher FeatCov explanations for male users than female users, aligned with the existing findings that male users usually write more detailed reviews (Fan-Osuala, 2023). On Yelp, the model generates more thorough feature descriptions for pricier restaurants than their less expensive counterparts, as shown in Table 4 in Appendix B. This could be a reflection of users' heightened attention to experiences in more expensive restaurants due to higher expenditure. These observations also align with the biases present in the respective training datasets.
+
+Baselines. As we are the first to study the quality fairness in PTG, there is no existing work that directly addresses this problem. We adapt popular methods in the fairness literature for the purpose. Below we briefly introduce them and describe their detailed implementations in Appendix B.
+
+- RAW: Original PETER model without modification, which is also the base for other baselines.  
+- ADV (in-processing): Adversarially removing the sensitive information in user or item representations by adding a discriminator (Li et al., 2021b).  
+- NORM (pre-processing): Normalizing the training data to remove the bias on group-level.  
+- BT (pre-processing): Back-translation has been shown to help normalize text and reduce bias (Rabinovich et al., 2017; Christiansen et al., 2021). We pre-process the training data by translating the explanations to Chinese and then back to English.  
+- ATTR: To evaluate the effectiveness of the disentanglement and optimizing the fairness constraint in COFFEE, we use the model trained without the constraint (equivalent to  $\lambda = 0$  in COFFEE) but with disentangled attribute representations.  
+- NATTR (post-processing): We disable the protected attribute token in ATTR for generating explanations during inference.
+
+We train each model on the training set, tune hyper-parameters on the validation set, and report the results on the testing set. Each reported result point is averaged over 3 runs. We put the detailed model specifications and parameter tuning in Ap
+
+pendix C.
+
+Metrics. Our evaluation consists of two parts: fairness and utility. Fairness requirements are subjective, contingent on both the application and the designer's objective (Hu and Chen, 2020; Khademi et al., 2019). COFFEE is designed to accommodate different fairness specifications by appropriately defining  $Q$ . To select meaningful quality evaluation metrics in experiments, we evaluated explanations with different measures such as FeatCov, length, grammar, redundancy, structure (Zhu and Bhat, 2020). We found that only FeatCov consistently shows significant correlations to human perceptions of explanation quality such as informativeness, detailedness, and helpfulness. The detailed correlation results are shown in Table 6 in Appendix D. Therefore, we specify the quality measure function in COFFEE as  $Q_{\text{feat}}$ , which directly corresponds to an explanation's FeatCov. Furthermore, experiment findings suggest a substantial influence of the number of tokens in explanations on FeatCov. Hence, we additionally explore a combination of FeatCov and number of tokens to facilitate the fairness optimization on FeatCov, specifying  $Q$  as  $Q_{\text{feat}} + Q_{\text{numtoken}}$ . We refer to this variant, with  $Q$  augmented by the number of tokens, as COFFEE-NT. This modification also helps validate COFFEE's ability to adapt to different quality measures in optimizing fairness.
+
+- Fairness metrics. We adopt both individual and group-wise fairness metrics evaluated on FeatCov, with and without counterfactual perspectives. Detailed metric formulas are in Appendix B. The first metric, Ind-CF, assesses individual CF as defined by (Kusner et al., 2017), which is equivalent to the regularization term in Eq. (4) averaged over all test set user-item pairs. We also evaluate the counterfactual effects on group-level (Coston et al., 2020). For a demographic group with attribute value  $a$ , we counterfactually alter all group members' attribute value to  $a' \neq a$ , and compute the change of the average FeatCov of the group. The average change is denoted as Grp-CF. Additionally, we evaluate COFFEE's generalization to improve group-wise fairness evaluated by demographic disparity (DDP) (Zafar et al., 2015). Metrics based on CF can only be evaluated on baselines that feature disentangled attribute representations, such as ATTR and COFFEE. All fairness metrics are lower the better.
+
+- Utility metrics. Besides commonly used BLEU-{1,4} (Papineni et al., 2002) and ROUGE-{1,2,L}
+
+![](images/147e7de7934470d1ca4de261a208476a043fef5de4c52047c201bb38fe3669f5.jpg)  
+Figure 4: Trade-off between fairness and utility on Amazon Movies & TV (left) and Yelp (right) datasets. A method should give low DDP and high BERTScore (top-left corner of each plot) for better trade-offs.
+
+![](images/c631027c979195eaac9ffcf618edc097c1577d243d8267abffa9f85351ad2519.jpg)
+
+(Lin, 2004) scores, we also employ the increasingly popular BERTScore (Zhang et al., 2020), known for better semantic alignment between target and generated text and stronger correlation with human judgments. We compute BERTScore using RoBERTa-large with re-scaled scores, and report F-1 scores for ROUGE and BERTScore. All utility metrics are higher the better. Lastly, we incorporate FeatCov into the utility metrics to evaluate the impact of various fairness attainment methods.
+
+# 6.2 Comparison to Baselines
+
+An ideal method should enhance fairness without hurting explanation utility. Due to space limits, we display complete results on Amazon Games in Table 2, and sample results for Amazon Movies and Yelp in Figure 4; full results are in Appendix E. As indicated in Table 2, COFFEE excels in fairness improvement across all metrics, significantly outperforming baselines. Moreover, comparing COFFEE and ATTR in terms of Ind-CF and Grp-CF validates the effectiveness of our fair policy learning scheme in optimizing CF constraints. Additionally, introducing  $Q_{numtoken}$  and  $Q_{feat}$  in COFFEE-NT bolsters fairness optimization in FeatCov, suggesting that enhanced quality measures can aid fairness optimization, echoing findings in (Bose and Hamilton, 2019). Crucially, COFFEE secures strong fairness outcomes while preserving high utility in explanation generation.
+
+We plot results on Amazon Movies and Yelp in Figure 4, where we use DDP on FeatCov for fairness and BERTScore for utility. Notably, COFFEE achieves the best trade-off by drastically reducing DDP with minimal impact on BERTScore. ADV and NORM moderately improve fairness, albeit less significantly than COFFEE or at the cost of a greater decrease on utility. Lastly, BT also effectively enhances fairness, showing its interesting
+
+Table 2: Comparison between COFFEE and baselines. BL stand for BLEU and RG denotes ROUGE. BLEU, ROUGE and BERTScore are in percentage values and others are in absolute values. The best results are boldfaced, and the second best are underlined. * indicates  $p < {0.05}$  for significance test over the second best baseline.  
+
+<table><tr><td rowspan="2">PETER</td><td colspan="3">Fairness onFeatCov</td><td colspan="7">Utility</td></tr><tr><td>Ind-CF↓</td><td>Grp-CF↓</td><td>DDP↓</td><td>BL1↑</td><td>BL4↑</td><td>RG1↑</td><td>RG2↑</td><td>RGL↑</td><td>BERT↑</td><td>FeatCov</td></tr><tr><td colspan="11">Amazon Games (User&#x27;s Gender as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>1.18</td><td>8.82</td><td>1.49</td><td>19.95*</td><td>5.37*</td><td>15.54*</td><td>14.14</td><td>6.53</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>0.62</td><td>8.78</td><td>1.50</td><td>16.13</td><td>3.96</td><td>13.06</td><td>12.84</td><td>6.12</td></tr><tr><td>NORM-F</td><td>-</td><td>-</td><td>1.13</td><td>8.83</td><td>1.51</td><td>18.20</td><td>4.66</td><td>14.42</td><td>13.91</td><td>4.28</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>0.65</td><td>10.43*</td><td>1.55</td><td>13.42</td><td>2.08</td><td>10.44</td><td>9.55</td><td>5.42</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.63</td><td>8.76</td><td>1.52</td><td>16.55</td><td>3.86</td><td>13.02</td><td>12.72</td><td>5.26</td></tr><tr><td>ATTR</td><td>2.68</td><td>0.75</td><td>1.40</td><td>8.91</td><td>1.57</td><td>17.73</td><td>4.45</td><td>14.13</td><td>14.08</td><td>5.78</td></tr><tr><td>COFFEE</td><td>1.26</td><td>0.32</td><td>0.68</td><td>9.77</td><td>1.58</td><td>16.73</td><td>3.98</td><td>13.43</td><td>14.12</td><td>6.34</td></tr><tr><td>COFFEE-NT</td><td>1.15</td><td>0.01*</td><td>0.08*</td><td>10.23</td><td>1.24</td><td>17.38</td><td>3.70</td><td>13.62</td><td>14.02</td><td>6.27</td></tr></table>
+
+![](images/dcf4de5bc3a108c2b08d4fc259d59f09e820e3890449d427fcc1649ed36e1baa.jpg)  
+Figure 5: Relative Ind-CF and BERTScore ratio w.r.t. to  $\lambda = 0$  in COFFEE. The Ind-CF and BERTScore of ATTR-PETER are 12.42 and 14.08, respectively.
+
+normalization effect. But BT causes a dramatic sacrifice on utility, as the translation may eliminate much information about personal preference.
+
+# 6.3 Effect of Tuning  $\lambda$  in COFFEE
+
+We further study the fairness-utility trade-off in COFFEE by tuning the weight  $\lambda$  on fairness constraint in Eq. (4). We use Ind-CF on FeatCov for fairness evaluation which corresponds to the constraint that COFFEE directly optimizes, and use BERTScore for utility evaluation. We plot the results on Amazon Movies in Fig. 5, where the y-axis is the metric values ratio to the base results when  $\lambda = 0$  (equivalent to ATTR). As  $\lambda$  increases from 0 to 0.5, COFFEE significantly improves fairness, with a drop of about  $77\%$  in Ind-CF, while barely hurts the BERTScore, with a drop of about  $3\%$ . This study shows COFFEE's outstanding efficiency and effectiveness in fairness optimization.
+
+# 7 Human Evaluation
+
+We conduct human evaluations to justify the use of FeatCov for fairness and confirm COFFEE's efficacy. Human evaluators, recruited on Mechanical Turk, are asked to assess explanations generated from Amazon Movies' test data, as movies are
+
+Table 3: Fairness and quality based on human evaluations of explanations. Smaller Ind-CF means better fairness. * indicates  $p$ -value  $< 0.05$  for paired t-test.  
+
+<table><tr><td rowspan="2"></td><td colspan="2">Ind-CF↓</td><td colspan="2">Average↑</td></tr><tr><td>RAW</td><td>COFFEE</td><td>RAW</td><td>COFFEE</td></tr><tr><td>Infor</td><td>0.8476</td><td>0.5673*</td><td>2.6381</td><td>2.6712</td></tr><tr><td>Detail</td><td>0.8190</td><td>0.6442*</td><td>2.5238</td><td>2.3788</td></tr><tr><td>Helpful</td><td>0.8857</td><td>0.5786*</td><td>2.5904</td><td>2.5577</td></tr></table>
+
+more universally understood. In each questionnaire, we present the title of a recommended movie for a user, randomly sampled from the dataset, along with two explanations: one generated for the real user and the other for a counterfactual user with a modified gender. Evaluators then assess each explanation from three criteria: informativeness, detailedness, and helpfulness. Each criterion is rated on a scale of 5: 1: "not at all", 2: "somewhat", 3: "moderate", 4: "very", 5: "extremely". We also collect evaluators' gender (only for aggregated analysis) to examine if humans with different genders evaluate the explanations differently. Rigorous privacy protection and quality control mechanisms are implemented. We compare the explanations generated from the raw model and COFFEE, with 150 valid questionnaires collected for each.
+
+We measure fairness in the three human evaluated criteria of explanations. The results are shown in Table 3, where we use Ind-CF as the fairness metric. COFFEE can effectively achieve better fairness on human evaluated quality measures by improving fairness on FeatCov. Moreover, as indicated by the average quality measure values, the quality of explanations generated by COFFEE is comparable to that of the RAW model. These results demonstrate COFFEE's ability and practical use in achieving fairness in explanation generation.
+
+# 8 Conclusion
+
+We investigate the fairness problem in PTG by focusing on the bias in FeatCov of explanation generation. We propose COFFEE for achieving counterfactual fairness in explanation generation. Comprehensive experiments and human evaluations show the efficacy of COFFEE in achieving fairness without hurting the utility of explanations.
+
+This work opens the new direction of quality fairness in PTG. We anticipate that this work will encourage researchers to investigate novel fairness notions in this problem, based on different quality measures, and conduct in-depth analyses on their social impacts. A promising direction is to involve humans in the loop for direct fairness optimization on human evaluated measures. We also plan to generalize the COFFEE to other PTG settings, such as conversational systems.
+
+# Limitations
+
+In this study, we formulated a counterfactual fairness constraint on explanation qualities, which is grounded in sampled explanations from both real and counterfactual worlds. We employed straightforward first moment matching to minimize the disparity between these two sample sets. Future work could explore the use of higher moment matching to more effectively align the quality distributions from the two worlds. For our experiments, we utilized FeatCov of explanations as a quality measure, given its simplicity and significance in the context of explainable recommendations. However, exploring other domain-specific measures (e.g., sentiment, emotion) to define  $Q$  and assess COFFEE's versatility in different PTG settings would be intriguing. We have also achieved encouraging results in terms of sentiment fairness using COFFEE. Nevertheless, a more detailed discussion on this falls outside the scope of this paper's focus on quality fairness, and thus we earmark it for future investigation.
+
+# Acknowledgements
+
+This work was partially supported by NSF IIS-2007492 and NSF IIS-1838615.
+
+# References
+
+H. Samy Alim, John R. Rickford, and Arnetha F. Ball. 2016. Racial linguistics: How Language Shapes Our Ideas About Race. Oxford University Press.
+
+Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth. 2021. Fairness in criminal justice risk assessments: The state of the art. *Sociological Methods & Research*.  
+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. Association for Computational Linguistics.  
+Shikha Bordia and Samuel R. Bowman. 2019. Identifying and reducing gender bias in word-level language models. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Avishek Joey Bose and William Hamilton. 2019. Compositional fairness constraints for graph embeddings.  
+Hanxiong Chen, Xu Chen, Shaoyun Shi, and Yongfeng Zhang. 2021. Generate natural language explanations for recommendation.  
+Jonathan Gabel Christiansen, Mathias Gammelgaard, and Anders Søgaard. 2021. The effect of round-trip translation on fairness in sentiment analysis. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.  
+Amanda Coston, Alan Mishler, Edward H. Kennedy, and Alexandra Chouldechova. 2020. Counterfactual risk assessments, evaluation, and fairness. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* '20. Association for Computing Machinery.  
+Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representations.  
+Mengnan Du, Fan Yang, Na Zou, and Xia Hu. 2021. Fairness in deep learning: A computational perspective. IEEE Intelligent Systems.  
+Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference, ITCS '12. Association for Computing Machinery.  
+Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics.
+
+Onochie Fan-Osuala. 2023. Women's online opinions are still not as influential as those of their male peers in buying decisions. *Humanities and Social Sciences Communications*, 10:40.  
+Sahaj Garg, Vincent Perot, Nicole Limtiaco, Ankur Taly, Ed H. Chi, and Alex Beutel. 2019. Counterfactual fairness in text classification through robustness. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, AIES '19, New York, NY, USA. Association for Computing Machinery.  
+Sophie Groenwold, Lily Ou, Aesha Parekh, Samhita Honnavalli, Sharon Levy, Diba Mirza, and William Yang Wang. 2020. Investigating african-american vernacular english in transformer-based text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics.  
+Moritz Hardt, Eric Price, Eric Price, and Nati Srebro. 2016. Equality of opportunity in supervised learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc.  
+Bing He, Mustaque Ahamad, and Srijan Kumar. 2021. Petgen: Personalized text generation attack on deep sequence embedding-based classification models. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, KDD '21, New York, NY, USA. Association for Computing Machinery.  
+Lily Hu and Yiling Chen. 2020. Fair classification and social welfare. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, FAT* '20. Association for Computing Machinery.  
+Po-Sen Huang, Huan Zhang, Ray Jiang, Robert Stanforth, Johannes Welbl, Jack Rae, Vishal Maini, Dani Yogatama, and Pushmeet Kohli. 2020. Reducing sentiment bias in language models via counterfactual evaluation. In *Findings of the Association for Computational Linguistics: EMNLP* 2020. Association for Computational Linguistics.  
+Matthew Joseph, Michael Kearns, Jamie H. Morgenstern, Seth Neel, and Aaron Roth. 2016. Rawlsian fairness for machine learning. *ArXiv*, abs/1610.09559.  
+Aria Khademi, Sanghack Lee, David Foley, and Vasant Honavar. 2019. Fairness in algorithmic decision making: An excursion through the lens of causality. In *The World Wide Web Conference*, WWW '19. Association for Computing Machinery.  
+Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2017. Inherent trade-offs in the fair determination of risk scores. In 8th Innovations in Theoretical Computer Science Conference (ITCS 2017), Leibniz International Proceedings in Informatics. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik.
+
+Matt J Kusner, Joshua Loftus, Chris Russell, and Ricardo Silva. 2017. Counterfactual fairness. In Advances in Neural Information Processing Systems. Curran Associates, Inc.  
+Jing Yang Lee, Kong Aik Lee, and Woon Seng Gan. 2021. Dlvgen: A dual latent variable approach to personalized dialogue generation.  
+Michelle Seng Ah Lee and Luciano Floridi. 2021. Algorithmic fairness in mortgage lending: From absolute conditions to relational trade-offs. *Minds Mach*.  
+Lei Li, Yongfeng Zhang, and Li Chen. 2021a. Personalized transformer for explainable recommendation. 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). Association for Computational Linguistics.  
+Piji Li, Zihao Wang, Zhaochun Ren, Lidong Bing, and Wai Lam. 2017. Neural rating regression with abstractive tips generation for recommendation. In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '17. Association for Computing Machinery.  
+Yunqi Li, Hanxiong Chen, Shuyuan Xu, Yingqiang Ge, and Yongfeng Zhang. 2021b. Towards personalized fairness based on causal notion. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, New York, NY, USA. Association for Computing Machinery.  
+Paul Pu Liang, Chiyu Wu, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2021. Towards understanding and mitigating social biases in language models. In ICML.  
+Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text Summarization Branches Out. Association for Computational Linguistics.  
+Francesco Locatello, Gabriele Abbati, Thomas Rainforth, Stefan Bauer, Bernhard Scholkopf, and Olivier Bachem. 2019a. On the fairness of disentangled representations. In Advances in Neural Information Processing Systems. Curran Associates, Inc.  
+Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Ratsch, Sylvain Gelly, Bernhard Scholkopf, and Olivier Bachem. 2019b. Challenging common assumptions in the unsupervised learning of disentangled representations. In ICML.  
+Annie Louis. 2013. Predicting text quality: Metrics for content, organization and reader interest. *Dissertations available from ProQuest*.  
+Jianxin Ma, Chang Zhou, Peng Cui, Hongxia Yang, and Wenwu Zhu. 2019. Learning Disentangled Representations for Recommendation. Curran Associates Inc., Red Hook, NY, USA.
+
+Karima Makhlouf, Sami Zhioua, and Catuscia Palamidessi. 2020. Survey on causal-based machine learning fairness notions.  
+Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. 2021. A survey on bias and fairness in machine learning. ACM Comput. Surv.  
+Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. 2016. Asynchronous methods for deep reinforcement learning.  
+Matthew Newman, Carla Groom, Lori Handelman, and James Pennebaker. 2008. Gender differences in language use: An analysis of 14,000 text samples. *Discourse Processes - DISCOURSE PROCESS*.  
+Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. 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), Hong Kong, China. Association for Computational Linguistics.  
+Shereen Oraby, Lena Reed, Shubhangi Tandon, Sharath T.S., Stephanie Lukin, and Marilyn Walker. 2018. Controlling personality-based stylistic variation with neural natural language generators. In Proceedings of the 19th Annual SIGdial Meeting on Discourse and Dialogue. 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 on Association for Computational Linguistics, ACL '02, USA. Association for Computational Linguistics.  
+Ella Rabinovich, Raj Nath Patel, Shachar Mirkin, Lucia Specia, and Shuly Wintner. 2017. Personalized machine translation: Preserving original author traits. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers. Association for Computational Linguistics.  
+Chris Russell, Matt J. Kusner, Joshua R. Loftus, and Ricardo Silva. 2017. When worlds collide: Integrating different counterfactual assumptions in fairness. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17. Curran Associates Inc.  
+H. Schwartz, Maarten Sap, Margaret Kern, Johannes Eichstaedt, Adam Kapelner, Megha Agrawal, Eduardo Blanco, Lukasz Dziurzynski, Gregory Park, David Stillwell, Michal Kosinski, Martin Seligman, and Lyle Ungar. 2016. Predicting individual wellbeing through the language of social media.
+
+Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. 2021. Societal biases in language generation: Progress and challenges. In Proceedings of the Conference of the 59th Annual Meeting of the Association for Computational Linguistics (ACL).  
+Lei Shu, Alexandros Papangelis, Yi-Chia Wang, Gokhan Tur, Hu Xu, Zhaleh Feizollahi, Bing Liu, and Piero Molino. 2020. Controllable text generation with focused variation. In Findings of the Association for Computational Linguistics: EMNLP 2020. Association for Computational Linguistics.  
+Ashudeep Singh and Thorsten Joachims. 2018. Fairness of exposure in rankings. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '18. Association for Computing Machinery.  
+Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. 1999. Policy gradient methods for reinforcement learning with function approximation. In Advances in Neural Information Processing Systems. MIT Press.  
+Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning.  
+Nava Tintarev and Judith Masthoff. 2015. *Explaining Recommendations: Design and Evaluation*. Springer US.  
+Paul Voigt and Axel von dem Bussche. 2017. The EU General Data Protection Regulation (GDPR): A Practical Guide. Springer Publishing Company, Incorporated.  
+Sarah Volz, Marc-Andre Reinhard, and Patrick Müller. 2020. Why don't you believe me? detecting deception in messages written by nonnative and native speakers. Applied Cognitive Psychology.  
+Lequn Wang and Thorsten Joachims. 2021. User fairness, item fairness, and diversity for rankings in two-sided markets. In Proceedings of the 2021 ACM SIGIR International Conference on Theory of Information Retrieval, ICTIR '21. Association for Computing Machinery.  
+Nan Wang, Lu Lin, Jundong Li, and Hongning Wang. 2022. Unbiased graph embedding with biased graph observations. In Proceedings of the ACM Web Conference 2022, WWW '22, New York, NY, USA. Association for Computing Machinery.  
+Nan Wang, Hongning Wang, Yiling Jia, and Yue Yin. 2018. Explainable recommendation via multi-task learning in opinionated text data. In The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '18. Association for Computing Machinery.
+
+Yongkai Wu, Lu Zhang, Xintao Wu, and Hanghang Tong. 2019. PC-Fairness: A Unified Framework for Measuring Causality-Based Fairness. Curran Associates Inc.  
+Aobo Yang, Nan Wang, Renqin Cai, Hongbo Deng, and Hongning Wang. 2022. Comparative explanations of recommendations. In Proceedings of the ACM Web Conference 2022, WWW '22. Association for Computing Machinery.  
+Aobo Yang, Nan Wang, Hongbo Deng, and Hongning Wang. 2021. Explanation as a defense of recommendation. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining, WSDM '21, New York, NY, USA. Association for Computing Machinery.  
+Chenhan Yuan and Yi-Chin Huang. 2019. Personalized sentence generation using generative adversarial networks with author-specific word usage. CoRR, abs/1904.09442.  
+Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi. 2017. Fairness beyond disparate treatment and disparate impact: Learning classification without disparate mistreatment. In Proceedings of the 26th International Conference on World Wide Web, WWW '17. International World Wide Web Conferences Steering Committee.  
+Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P. Gummadi. 2015. Learning fair classifiers. arXiv: Machine Learning.  
+Rich Zemel, Yu Wu, Kevin Swersky, Toni Pitassi, and Cynthia Dwork. 2013. Learning fair representations. In Proceedings of the 30th International Conference on Machine Learning, Proceedings of Machine Learning Research. PMLR.  
+Junzhe Zhang and Elias Bareinboim. 2018. Equality of opportunity in classification: A causal approach. In Advances in Neural Information Processing Systems. Curran Associates, Inc.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations.  
+Wei-Nan Zhang, Qingfu Zhu, Yifa Wang, Yanyan Zhao, and Ting Liu. 2019. Neural personalized response generation as domain adaptation. World Wide Web, 22(4).  
+Yongfeng Zhang and Xu Chen. 2020. Explanable recommendation: A survey and new perspectives. Found. Trends Inf. Retr., 14:1-101.  
+Yongfeng Zhang, Xu Chen, Qingyao Ai, Liu Yang, and W. Bruce Croft. 2018. Towards conversational search and recommendation: System ask, user respond. In Proceedings of the 27th ACM International Conference on Information and Knowledge Management,
+
+CIKM '18, New York, NY, USA. Association for Computing Machinery.  
+Yu Zheng, Chen Gao, Xiang Li, Xiangnan He, Yong Li, and Depeng Jin. 2021. Disentangling user interest and conformity for recommendation with causal embeddings. In Proceedings of the Web Conference 2021, WWW '21. Association for Computing Machinery.  
+Wanzheng Zhu and Suma Bhat. 2020. GRUEN for evaluating linguistic quality of generated text. In Findings of the Association for Computational Linguistics: EMNLP 2020, Online. Association for Computational Linguistics.
+
+# A Applying COFFEE to NRT
+
+There are mainly two modules in NRT (Li et al., 2017)—a MLP for rating prediction, and an RNN for explanation generation. Both modules share the same user and item embeddings mapped from input user and item IDs for personalization. In particular, the user and item embeddings are used in the initial hidden state of the RNN for explanation generation (see Figure 2 in (Li et al., 2017)). To apply COFFEE to NRT, we simply need to use the concatenation of disentangled user preference embedding and the attribute embedding instead of the holistic user embedding, exactly as introduced in Section 4.1. Similar to PETER, there are also three loss terms in the original NRT, corresponding to the explanation, context, and rating prediction (see Section 3 in (Li et al., 2017)). We denote the total loss of NRT by  $\mathcal{L}_{nrt}$ , which is used to replace  $\mathcal{L}_{gen}$  in Eq. (3) for applying COFFEE.
+
+# B Experiment Setup
+
+# B.1 Dataset Processing
+
+The first two datasets are the Video Games and the Movies & TV categories from Amazon reviews. These two datasets only provide the names of users. We use a gender classification tool to predict the gender of users from their names. To guarantee the accuracy of the attribute values, we only reserve the users whose names can be confidently classified as male or female and remove those classified as unknown. In the Yelp dataset, we use a restaurant's price range as the protected attribute. There are originally four price ranges ( $,$ 1, $1, $2, $3, $4, $5) provided. However, we found that there are much less four-dollar restaurants than in the other price ranges, and thus we merge the four-dollar restaurants into the group of three-dollar ones for experiments. Using restaurant price as a sensitive attribute is based on the counterfactual argument: if a restaurant chose to raise/lower the price of its dishes without other changes, the quality of generated explanations shouldn't change. The rationale is that users may care more about experiences in expensive restaurants since they are paying more, and thus writing more detailed reviews. As shown in Table 4, the explanations for expensive restaurants have higher FeatCov and thus are more informative and detailed, which helps recommendations on them be accepted more often. However, this
+
+Table 4: Average FeatCov of explanations generated for restaurants of different price tags.  
+
+<table><tr><td>FeatCov</td><td>$</td><td>$$</td><td>$$$$</td></tr><tr><td>Ground-Truth</td><td>3.15</td><td>3.72</td><td>4.38</td></tr><tr><td>RAW PETER</td><td>3.26</td><td>3.68</td><td>4.03</td></tr></table>
+
+will also leave the lower-priced restaurants at a disadvantage.
+
+We split each dataset into training  $(80\%)$ , validation  $(10\%)$ , and testing  $(10\%)$  sets, and ensure that there is at least one record in each subset for every user and item.
+
+# B.2 Details of Baselines
+
+- ADV (in-processing): Adversarially removing the sensitive information in user or item representations by adding a discriminator (Li et al., 2021b). We add discriminators on the user's (or item's) embedding in PETER and NRT to remove the information about protected attributes.  
+- NORM (pre-processing): We normalize the training data to remove the bias on group-level. With two demographic groups, we remove the explanations with the higher (lower) quality in the group with more reviews, until the difference of average quality between the two groups is below  $10\%$  of the original difference. With more than two groups, we recursively apply the procedure to the two groups with the maximum difference until the maximum difference is below the threshold. NORM removes less than  $2\%$  of training data on Amazon datasets, and less than  $18\%$  on Yelp.  
+- We pre-process the training data by translating the explanations to Chinese and then back to English. We use the multilingual model mBART (Tang et al., 2020) from EasyNMT<sup>7</sup> to perform the translation.  
+- ATTR: In order to evaluate the effectiveness of optimizing the fairness constraint in COFFEE, we use the model trained without the constraint  $(\lambda = 0)$  in Eq. (4)) as a baseline. We call the model ATTR, which means the attribute is disentangled from the user's or item's representations as introduced in Section 5 but without adding the fairness constraint.  
+- NATTR (post-processing): We disable the protected attribute token in ATTR for generating explanations during inference. Specifically, for NATR-NRT, we replace the learned attribute embeddings
+
+with random embeddings with values uniformly sampled in  $[-1, 1]$  for explanation generation. For NATTR-PETER, we disable the other tokens' attention on the attribute token during inference.
+
+# B.3 Evaluation Metrics
+
+We evaluate both individual-level and group-level fairness, with and without counterfactual perspectives. The first metric denoted as Ind-CF evaluates the originally defined CF on individuals (Kusner et al., 2017), which is the same as the regularization term in Eq. (4) averaged over all user-item pairs on the test set  $\mathcal{D}$ :
+
+Ind-CF (8)
+
+$$
+= \frac {1}{| \mathcal {D} |} \sum_ {u, i \in \mathcal {D}} \left| \mathbb {E} \left[ Q \left(y _ {A \leftarrow a}\right) \mid u, i \right] - \mathbb {E} \left[ Q \left(y _ {A \leftarrow a ^ {\prime}}\right) \mid u, i \right] \right|.
+$$
+
+We also evaluate the counterfactual effect on group-level (Coston et al., 2020). For each attribute value  $a$  and the corresponding demographic group, we counterfactually change the attribute value of all users in this group by  $a' \neq a$ , and calculate the change on the average quality of this group. We call this metric Group-CF:
+
+Group-CF (9)
+
+$$
+= \frac {1}{| \mathcal {A} | (| \mathcal {A} | - 1)} \sum_ {a \in \mathcal {A}} \sum_ {a ^ {\prime} \neq a} \frac {1}{| \mathcal {D} _ {a} |}
+$$
+
+$$
+\left| \sum_ {u, i \in \mathcal {D} _ {a}} \mathbb {E} [ Q (y _ {A \leftarrow a}) | u, i ] - \sum_ {u, i \in \mathcal {D} _ {a}} \mathbb {E} [ Q (y _ {A \leftarrow a ^ {\prime}}) | u, i ] \right|
+$$
+
+where  $\mathcal{D}_a$  denotes the demographic group with attribute value  $a$ .
+
+Finally, besides counterfactual metrics, we also evaluate COFFEE's generalization to improve group-wise fairness by the popular notion of demographic disparity (DDP) (Zafar et al., 2015):
+
+DDP (10)
+
+$$
+= \frac {2}{| \mathcal {A} | (| \mathcal {A} | - 1)} \sum_ {a, a ^ {\prime} \in \mathcal {A}}
+$$
+
+$$
+\left| \frac {1}{| \mathcal {D} _ {a} |} \sum_ {u, i \in \mathcal {D} _ {a}} \mathbb {E} [ Q (y) | u, i ] - \frac {1}{| \mathcal {D} _ {a ^ {\prime}} |} \sum_ {u, i \in \mathcal {D} _ {a ^ {\prime}}} \mathbb {E} [ Q (y) | u, i ] \right|
+$$
+
+All fairness metrics are lower the better, and the quality expectation on each user-item pair is calculated over  $N = 3$  sampled explanations.
+
+<table><tr><td rowspan="2">COFFEE</td><td colspan="3">PETER</td><td colspan="3">NRT</td></tr><tr><td>Games</td><td>Movies</td><td>Yelp</td><td>Games</td><td>Movies</td><td>Yelp</td></tr><tr><td>λ</td><td>0.2</td><td>0.2</td><td>0.2</td><td>0.3</td><td>0.1</td><td>0.1</td></tr><tr><td>η</td><td>0.6</td><td>0.6</td><td>0.5</td><td>0.5</td><td>0.5</td><td>0.5</td></tr></table>
+
+Table 5: Weight  $\lambda$  for the fairness constraint and the promotion weight  $\eta$  in reward calibration of COFFEE when applied to PETER and NRT on the three datasets.
+
+# C Experiment Settings
+
+Here we elaborate the experimental protocols, model specifications, and hyper-parameters. For all models, we set the size of vocabulary to 20,000 by keeping the most frequent words. By default, we use top-5 sampling (Fan et al., 2018) as the decoding strategy, and the maximum decoding sequence length is 128.
+
+For the raw PETER model (Li et al., 2021a), we mostly follow the original paper's hyperparameters. The token embedding dimension is 512 and the dimension of feed-forward network is 2,048. The number of transformer layers and attention heads are both two. The dropout rate during training is 0.2. For NRT (Li et al., 2017), we follow the original paper and set the embedding dimension to 300 for all users, items and words. The dimension of hidden layers is 400. The dropout rate during training is 0.1. COFFEE consists of pre-training the ATTR models with disentangled attribute representations and then fine-tune them with the counterfactual fairness constraints. For ATTR-PETER, we just add an additional attribute token to PETER, with the same embedding dimension for the attribute. For ATTR-NRT, we disentangle the attribute representation by concatenating to the user or item embedding an attribute embedding of dimension 100. The other model specifications for ATTR-PETER and ATTR-NRT are the same as raw PETER and NRT. When applying the discriminator for removing the information about the protected attribute from user or item embeddings, we use a 2-layer MLP with hidden size of 512 as the attribute discriminator. For both ATTR-PETER and ATTR-NRT, the weight  $\lambda_{D}$  of the adversarial loss is set to 0.5 on Amazon Games and Yelp datasets, and 0 on Amazon Movies. We iterate between one epoch of model training and one epoch of discriminator training until convergence. After the pre-training of ATTR models, we fix the user, item and attribute embeddings, remove the loss on rating prediction but keep the loss on explanation generation, and tune the model with the fairness constraint. The
+
+Table 6: Kruskal-Wallis H test between FeatCov and quality and utility evaluated by human.  
+
+<table><tr><td>K-W H test</td><td></td><td>Infor</td><td>Detail</td><td>Help</td></tr><tr><td rowspan="2">All</td><td>H-stats</td><td>39.44</td><td>45.67</td><td>23.83</td></tr><tr><td>p-value</td><td>5.6e-8</td><td>2.9e-9</td><td>8.6e-5</td></tr><tr><td rowspan="2">Male</td><td>H-stats</td><td>24.28</td><td>28.07</td><td>14.37</td></tr><tr><td>p-value</td><td>7.0e-5</td><td>1.2e-5</td><td>6.2e-3</td></tr><tr><td rowspan="2">Female</td><td>H-stats</td><td>16.58</td><td>15.74</td><td>12.24</td></tr><tr><td>p-value</td><td>2.3e-4</td><td>3.4e-4</td><td>5.6e-3</td></tr></table>
+
+weight  $\lambda$  for the fairness constraint and the promotion weight  $\eta$  in reward calibration are tuned for different models and datasets, and are listed in Table 5.
+
+For model training, we use Adam as optimizer and the initial learning rate for training raw models and ATTR models is 1e-4. The initial learning rate is 1e-5 for COFFEE during fine tuning by the fairness constraint. By default, the batch size is 16. But we use batch size of 8 during the tuning phase of COFFEE when applying to PETER models, mainly due to memory issues. For training raw models and ATTR models, we evaluate the total loss on the validation set after each epoch, and use the epoch checkpoint when the total loss is higher in the next 5 consecutive epochs. When COFFEE fine-tunes the pre-trained ATTR models, we always use the checkpoint after a single epoch, which already yields promising results from COFFEE.
+
+# D Results from Kruskal-Wallis H Tests
+
+We perform Kruskal-Wallis H test between FeatCov and three human evaluated measures including Informativeness, Detailiness, and Helpfulness. The results are shown in Table 6, where a H-stats larger than one means a strong correlation, and a p-value smaller than 0.05 means the correlation is significant. FeatCov strongly and significantly correlate with humans' (both male and female) judgement of the quality and helpfulness of explanations. Thus, the bias can lead to unfair treatments to users of different genders, as the quality of explanations, evaluated by FeatCov in this case, may affect the utility of explanations to users.
+
+# E Additional Results
+
+We show the complete results on the Amazon Movies and Yelp datasets based on the PETER
+
+model in Table 7. Importantly, COFFEE achieves strong fairness results while maintaining high utility on explanation generation. Although COFFEE may slightly drop the utility compared to the baselines, it can sometimes even outperform the baselines, as shown in the results on Amazon Movies and Yelp datasets. This is because the disentanglement mechanism enables better representation learning (Ma et al., 2019; Zheng et al., 2021) and increases the flexibility and accuracy of interactions between the user and item for better explanation generation.
+
+We also present some examples of generated explanations on Amazon Games in Table 8. In particular, we select a male user Chadwick who wrote detailed and informative reviews with high FeatCov and a female user Noemi who wrote generic reviews with low FeatCov for the same item "Wii Nunchuk Controller". The raw PETER model follows the ground-truth reviews and generate detailed explanation for Chadwick but short and generic explanation for Noemi. In contrast, after applying COFFEE to improve fairness, the model tend to generate and more descriptive and informative explanation for Noemi. These examples demonstrate the effectiveness of COFFEE in generating fair and high-quality explanations for users with different protected attribute values.
+
+# E.1 Entanglement of Protected Attribute with Preference
+
+In Section 4.1, we employ a disentanglement approach where each user's representation is separated into two components: the attribute representation and the preference representation. This setup enables us to perform counterfactual inference by switching the attribute representation from one attribute value to another. Following this method, we analyzed how much changing a user's learnt attribute representation will influence the FeatCov of generated explanations. Take the Amazon Games dataset for example, without imposing any fairness constraints, we observed that the average FeatCov change is 2.68 when altering the attribute representation. Furthermore, over  $92\%$  of users exhibit a FeatCov change greater than 1. This indicates that, in most cases, the attribute and preference representations are intertwined, and the disentanglement approach effectively separates these two aspects.
+
+Table 7: Comparison between COFFEE and baselines based on the PETER model. BL stand for BLEU and RG denotes ROUGE. BLEU, ROUGE and BERTScore are in percentage values and others are in absolute values. The best results are boldfaced, and the second best are underlined. * indicates  $p < {0.05}$  for significance test over the second best baseline.  
+
+<table><tr><td rowspan="2">PETER</td><td colspan="3">Fairness on FeatCov</td><td colspan="7">Utility</td></tr><tr><td>Ind-CF↓</td><td>Grp-CF↓</td><td>DDP↓</td><td>BL1↑</td><td>BL4↑</td><td>RG1↑</td><td>RG2↑</td><td>RGL↑</td><td>BERT↑</td><td>FeatCov</td></tr><tr><td colspan="11">Amazon Movies &amp; TV (User&#x27;s Gender as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>0.57</td><td>11.25</td><td>2.20</td><td>18.04</td><td>4.53</td><td>15.48</td><td>19.93</td><td>5.61</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>0.26</td><td>11.08</td><td>2.16</td><td>17.20</td><td>4.32</td><td>14.37</td><td>18.94</td><td>5.14</td></tr><tr><td>NORM</td><td>-</td><td>-</td><td>0.27</td><td>11.23</td><td>2.17</td><td>17.46</td><td>4.42</td><td>15.00</td><td>19.67</td><td>4.34</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>0.16</td><td>11.25</td><td>2.18</td><td>12.91</td><td>1.66</td><td>11.02</td><td>15.01</td><td>4.13</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.27</td><td>11.24</td><td>2.19</td><td>17.35</td><td>4.40</td><td>14.92</td><td>19.34</td><td>4.71</td></tr><tr><td>ATTR</td><td>1.47</td><td>0.10</td><td>0.37</td><td>11.19</td><td>2.16</td><td>17.43</td><td>4.63</td><td>15.98</td><td>19.72</td><td>5.31</td></tr><tr><td>COFFEE</td><td>1.13</td><td>0.01</td><td>0.07</td><td>11.28</td><td>2.12</td><td>17.77</td><td>4.60</td><td>15.35</td><td>20.13*</td><td>5.45</td></tr><tr><td>COFFEE-NT</td><td>1.16</td><td>0.01</td><td>0.00*</td><td>10.05</td><td>1.39</td><td>19.34*</td><td>4.56</td><td>15.86</td><td>19.52</td><td>5.08</td></tr><tr><td colspan="11">Yelp (Restaurant&#x27;s Price as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>0.51</td><td>10.24</td><td>1.57</td><td>18.71</td><td>3.11</td><td>14.64</td><td>19.82</td><td>3.78</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>0.28</td><td>9.87</td><td>1.60</td><td>17.25</td><td>2.64</td><td>13.99</td><td>18.34</td><td>3.18</td></tr><tr><td>NORM</td><td>-</td><td>-</td><td>0.09</td><td>9.96</td><td>1.54</td><td>17.56</td><td>2.81</td><td>14.10</td><td>19.67</td><td>3.06</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>0.27</td><td>10.69</td><td>1.61</td><td>16.42</td><td>2.36</td><td>13.29</td><td>18.89</td><td>3.20</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.32</td><td>10.08</td><td>1.55</td><td>18.53</td><td>3.06</td><td>14.57</td><td>19.75</td><td>3.37</td></tr><tr><td>ATTR</td><td>1.50</td><td>0.44</td><td>0.64</td><td>10.00</td><td>1.53</td><td>18.83*</td><td>3.16</td><td>14.75</td><td>19.81</td><td>3.52</td></tr><tr><td>COFFEE</td><td>0.84</td><td>0.08</td><td>0.13</td><td>11.17</td><td>1.77</td><td>17.97</td><td>3.00</td><td>14.39</td><td>20.58*</td><td>3.68</td></tr><tr><td>COFFEE-NT</td><td>0.54*</td><td>0.03</td><td>0.03*</td><td>11.61*</td><td>1.72</td><td>17.41</td><td>2.93</td><td>14.31</td><td>20.21</td><td>3.61</td></tr></table>
+
+Table 8: Generated explanations by different models on Amazon Games. We select one male user Chadwick and one female user Noemi, and present the explanations for them on the same item "Wii Nunchuk Controller".  
+
+<table><tr><td>PETER</td><td>user: Chadwick, item: Wii Nunchuk Controller</td><td>user: Noemi, item: Wii Nunchuk Controller</td></tr><tr><td>Ground-truth</td><td>this review is for the white wii nunchuk controller. it is a necessary component in most wii games, and attaches to the bottom port of every wii remote. it is well-designed, sturdy, and comfortable to hold. the price is relatively low for such a controller.</td><td>great controller, arrived as expected.</td></tr><tr><td>RAW</td><td>i bought it to play wii u and this is a great addition to the wii remote. the gamepad is a nice addition to the wii u gamepad. the wii u is a must have for any wii u console owner.</td><td>the controller worked very well.</td></tr><tr><td>COFFEE</td><td>i bought this controller to use the nunchuck. this is a great controller for wii - u owners especially the wii u. the controller is very responsive and the triggers are awesome.</td><td>i bought this controller for my wii u. it was exactly what it would be expected. i love the feel of the box and it works great.</td></tr></table>
+
+# E.2 Results based on NRT
+
+We present the complete results based on the NRT model in Table 9. Similar to the results on PETER, COFFEE when applied to NRT can significantly outperform the baselines in terms of fairness improvement. Again, when optimizing both  $Q_{numtoken}$  and  $Q_{feat}$  together, COFFEE achieves the best fairness improvements, verifying the effect of correlations between different quality measures. In general, PETER can generate better explanations than NRT, e.g., on BERTScore, with or without fairness optimizations, showing the capability of transformers over RNNs. However, COFFEE still maintains high generation utility based on NRT when compared to baselines, showing its advantage in generalizing the results to different models.
+
+The observations and conclusions are similar to the PETER based results. These results show COFFEE's ability to generalize the fairness improvement to different types of models, and indicate its flexibility and practicality in real-world uses.
+
+Table 9: Comparison between COFFEE and baselines based on the NRT model. BL stand for BLEU and RG denotes ROUGE. BLEU, ROUGE and BERTScore are in percentage values and others are in absolute values. The best results are boldfaced, and the second best are underlined. * indicates  $p < {0.05}$  for significance test over the second best baseline.  
+
+<table><tr><td rowspan="2">NRT</td><td colspan="3">Fairness on Qfeat</td><td colspan="7">Utility</td></tr><tr><td>Ind-CF↓</td><td>Grp-CF↓</td><td>DDP↓</td><td>BL1↑</td><td>BL4↑</td><td>RG1↑</td><td>RG2↑</td><td>RGL↑</td><td>BERT↑</td><td>FeatCov</td></tr><tr><td colspan="11">Amazon Games (User&#x27;s Gender as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>1.42</td><td>8.33</td><td>1.30</td><td>19.21</td><td>3.97</td><td>14.35*</td><td>13.79</td><td>6.48</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>1.25</td><td>8.13</td><td>1.29</td><td>17.67</td><td>3.37</td><td>13.27</td><td>12.85</td><td>6.08</td></tr><tr><td>NORM</td><td>-</td><td>-</td><td>1.79</td><td>8.17</td><td>1.26</td><td>19.89*</td><td>4.27*</td><td>13.65</td><td>13.49</td><td>4.80</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>1.47</td><td>8.34</td><td>1.17</td><td>16.44</td><td>2.71</td><td>12.23</td><td>10.73</td><td>5.29</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.12</td><td>8.39</td><td>1.39</td><td>13.07</td><td>2.71</td><td>10.83</td><td>11.11</td><td>5.83</td></tr><tr><td>ATTR</td><td>6.70</td><td>1.46</td><td>1.65</td><td>8.98</td><td>1.29</td><td>18.94</td><td>3.89</td><td>14.29</td><td>13.84</td><td>6.23</td></tr><tr><td>COFFEE</td><td>1.81</td><td>0.52</td><td>0.51</td><td>8.91</td><td>1.30</td><td>17.94</td><td>3.50</td><td>13.63</td><td>13.83</td><td>6.30</td></tr><tr><td>COFFEE-NT</td><td>0.77*</td><td>0.02*</td><td>0.04</td><td>8.67</td><td>1.26</td><td>17.11</td><td>3.11</td><td>12.46</td><td>13.69</td><td>6.27</td></tr><tr><td colspan="11">Amazon Movies &amp; TV (User&#x27;s Gender as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>0.46</td><td>10.55</td><td>1.86</td><td>18.37</td><td>3.98</td><td>15.16</td><td>19.20</td><td>5.46</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>0.24</td><td>10.48</td><td>1.83</td><td>18.44</td><td>3.95</td><td>15.10</td><td>19.09</td><td>5.18</td></tr><tr><td>NORM</td><td>-</td><td>-</td><td>0.46</td><td>10.53</td><td>1.91</td><td>18.09</td><td>3.89</td><td>14.96</td><td>19.03</td><td>4.62</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>0.43</td><td>10.63</td><td>1.62</td><td>14.84</td><td>1.87</td><td>11.95</td><td>14.36</td><td>5.06</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.33</td><td>9.84</td><td>2.07</td><td>15.90</td><td>3.37</td><td>13.77</td><td>16.57</td><td>4.75</td></tr><tr><td>ATTR</td><td>1.74</td><td>0.14</td><td>0.48</td><td>10.53</td><td>1.84</td><td>18.68*</td><td>4.10</td><td>15.34*</td><td>19.21</td><td>5.07</td></tr><tr><td>COFFEE</td><td>1.23</td><td>0.04</td><td>0.24</td><td>11.05</td><td>2.08</td><td>17.90</td><td>4.02</td><td>15.11</td><td>19.27</td><td>5.48</td></tr><tr><td>COFFEE-NT</td><td>1.10*</td><td>0.04</td><td>0.19*</td><td>10.52</td><td>2.02</td><td>17.04</td><td>3.92</td><td>14.77</td><td>19.35*</td><td>5.27</td></tr><tr><td colspan="11">Yelp (Restaurant&#x27;s Price as Protected Attribute)</td></tr><tr><td>RAW</td><td>-</td><td>-</td><td>0.56</td><td>10.09</td><td>1.49</td><td>19.14</td><td>3.17</td><td>14.83</td><td>19.50</td><td>3.76</td></tr><tr><td>ADV</td><td>-</td><td>-</td><td>0.21</td><td>10.35</td><td>1.97</td><td>16.78</td><td>2.49</td><td>13.14</td><td>17.88</td><td>3.24</td></tr><tr><td>NORM</td><td>-</td><td>-</td><td>0.15</td><td>9.88</td><td>1.42</td><td>19.20</td><td>3.14</td><td>14.85</td><td>19.48</td><td>3.04</td></tr><tr><td>BT</td><td>-</td><td>-</td><td>0.30</td><td>10.56*</td><td>1.52</td><td>16.44</td><td>2.31</td><td>13.24</td><td>18.95</td><td>3.10</td></tr><tr><td>NATTR</td><td>-</td><td>-</td><td>0.19</td><td>10.38</td><td>2.23*</td><td>15.89</td><td>2.26</td><td>12.58</td><td>17.26</td><td>3.04</td></tr><tr><td>ATTR</td><td>1.58</td><td>0.42</td><td>0.50</td><td>10.08</td><td>1.85</td><td>18.92</td><td>3.09</td><td>14.72</td><td>19.74</td><td>3.22</td></tr><tr><td>COFFEE</td><td>0.93</td><td>0.17</td><td>0.10</td><td>10.25</td><td>1.75</td><td>18.34</td><td>2.92</td><td>14.33</td><td>19.49</td><td>3.70</td></tr><tr><td>COFFEE-NT</td><td>0.69*</td><td>0.09*</td><td>0.02</td><td>10.45</td><td>2.07</td><td>18.65</td><td>2.86</td><td>14.25</td><td>19.67</td><td>3.52</td></tr></table>

2023/COFFEE_ Counterfactual Fairness for Personalized Text Generation in Explainable Recommendation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f1a906e61d62ad6962506d087f75d51eef2cb899c79f1fadb7373c21f05b1af
+size 791916