Add Batch 86e84251-afd0-4a36-90f1-e1681024dcc7 data

.gitattributes

@@ -5219,3 +5219,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Code-Switched[[:space:]]Text[[:space:]]Synthesis[[:space:]]in[[:space:]]Unseen[[:space:]]Language[[:space:]]Pairs/b627b8b3-6d9e-49f2-a9f8-3a27be6b8015_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/CodePrompt_[[:space:]]Task-Agnostic[[:space:]]Prefix[[:space:]]Tuning[[:space:]]for[[:space:]]Program[[:space:]]and[[:space:]]Language[[:space:]]Generation/7f6aa9d3-5907-4e0f-8830-86a860ebc33b_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Coherent[[:space:]]or[[:space:]]Not_[[:space:]]Stressing[[:space:]]a[[:space:]]Neural[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Discourse[[:space:]]Coherence[[:space:]]in[[:space:]]Multiple[[:space:]]Languages/30e11dca-164c-4670-973d-dc3881674ea2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Common[[:space:]]Law[[:space:]]Annotations_[[:space:]]Investigating[[:space:]]the[[:space:]]Stability[[:space:]]of[[:space:]]Dialog[[:space:]]System[[:space:]]Output[[:space:]]Annotations/954eaa6a-8529-416a-8e69-2af9134bed4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Commonsense[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Completion[[:space:]]Via[[:space:]]Contrastive[[:space:]]Pretraining[[:space:]]and[[:space:]]Node[[:space:]]Clustering/8ad456f0-7f0d-40ee-8bd0-7b70b4a0c21c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Commonsense[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]for[[:space:]]Pre-trained[[:space:]]Language[[:space:]]Models/f96ae55a-934d-4cd1-b9c6-f4ed8849b12f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Communication[[:space:]]Efficient[[:space:]]Federated[[:space:]]Learning[[:space:]]for[[:space:]]Multilingual[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Adapter/6f207301-9f3f-466d-aed1-4bdd052903e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Complementary[[:space:]]Explanations[[:space:]]for[[:space:]]Effective[[:space:]]In-Context[[:space:]]Learning/a50d012a-1115-4961-8cce-819a0bcf40d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Compositional[[:space:]]Mathematical[[:space:]]Encoding[[:space:]]for[[:space:]]Math[[:space:]]Word[[:space:]]Problems/801bbb7b-d8b8-4c19-87bf-9a90de8c668a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Computer[[:space:]]says[[:space:]]“No”_[[:space:]]The[[:space:]]Case[[:space:]]Against[[:space:]]Empathetic[[:space:]]Conversational[[:space:]]AI/919d9a05-5ebb-4c0e-a1de-82ff5cd4abb0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/ConKI_[[:space:]]Contrastive[[:space:]]Knowledge[[:space:]]Injection[[:space:]]for[[:space:]]Multimodal[[:space:]]Sentiment[[:space:]]Analysis/b434c033-5cf4-43ce-8d1a-9159c5428ecf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Concept2Box_[[:space:]]Joint[[:space:]]Geometric[[:space:]]Embeddings[[:space:]]for[[:space:]]Learning[[:space:]]Two-View[[:space:]]Knowledge[[:space:]]Graphs/a7ac92f6-3015-4bc3-abcd-68379068a363_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Conformal[[:space:]]Nucleus[[:space:]]Sampling/39826fe0-3dde-4001-a9ca-9e6036bc530d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Connectivity[[:space:]]Patterns[[:space:]]are[[:space:]]Task[[:space:]]Embeddings/2edb6b34-fce4-4345-b22d-ee57d10cffcd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Constructing[[:space:]]Code-mixed[[:space:]]Universal[[:space:]]Dependency[[:space:]]Forest[[:space:]]for[[:space:]]Unbiased[[:space:]]Cross-lingual[[:space:]]Relation[[:space:]]Extraction/429c226a-56a4-4156-971d-3a66e0e9b66b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Constructing[[:space:]]Procedural[[:space:]]Graphs[[:space:]]with[[:space:]]Multiple[[:space:]]Dependency[[:space:]]Relations_[[:space:]]A[[:space:]]New[[:space:]]Dataset[[:space:]]and[[:space:]]Baseline/5b1bfda9-cf73-4295-b9e0-bfde530e97f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Constructing[[:space:]]Word-Context-Coupled[[:space:]]Space[[:space:]]Aligned[[:space:]]with[[:space:]]Associative[[:space:]]Knowledge[[:space:]]Relations[[:space:]]for[[:space:]]Interpretable[[:space:]]Language[[:space:]]Modeling/8b4df5c0-9402-4afb-be44-169f19da2689_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Context[[:space:]]or[[:space:]]Knowledge[[:space:]]is[[:space:]]Not[[:space:]]Always[[:space:]]Necessary_[[:space:]]A[[:space:]]Contrastive[[:space:]]Learning[[:space:]]Framework[[:space:]]for[[:space:]]Emotion[[:space:]]Recognition[[:space:]]in[[:space:]]Conversations/acb9a5c0-b7ea-46aa-b58f-31bf099745e8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Context-Aware[[:space:]]Document[[:space:]]Simplification/3a00946c-ec4c-4520-a60c-6f2337b5b9fc_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contextualized[[:space:]]Semantic[[:space:]]Distance[[:space:]]between[[:space:]]Highly[[:space:]]Overlapped[[:space:]]Texts/9fe0fdbd-465f-4d9b-bd88-b892cf6a144f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contextualized[[:space:]]Soft[[:space:]]Prompts[[:space:]]for[[:space:]]Extraction[[:space:]]of[[:space:]]Event[[:space:]]Arguments/a537edee-bb8c-41b5-b6f8-eccc0b23c307_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contrastive[[:space:]]Learning[[:space:]]of[[:space:]]Sociopragmatic[[:space:]]Meaning[[:space:]]in[[:space:]]Social[[:space:]]Media/e55c419a-045c-4945-830d-cfb5b7e031d0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contrastive[[:space:]]Learning[[:space:]]with[[:space:]]Generated[[:space:]]Representations[[:space:]]for[[:space:]]Inductive[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Embedding/343e40b9-1aff-40d5-b4ec-dbd1369b1299_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contrastive[[:space:]]Token-Wise[[:space:]]Meta-Learning[[:space:]]for[[:space:]]Unseen[[:space:]]Performer[[:space:]]Visual[[:space:]]Temporal-Aligned[[:space:]]Translation/958bf3a6-c967-4fc4-8fb2-fcb56acfa1a8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Contrastive[[:space:]]Training[[:space:]]Improves[[:space:]]Zero-Shot[[:space:]]Classification[[:space:]]of[[:space:]]Semi-structured[[:space:]]Documents/d276a959-8684-4308-919d-45092a02bda6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Controllable[[:space:]]Conversation[[:space:]]Generation[[:space:]]with[[:space:]]Conversation[[:space:]]Structures[[:space:]]via[[:space:]]Diffusion[[:space:]]Models/2b5fbf94-c9ae-4927-9078-e78e607c20c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Controlled[[:space:]]Text[[:space:]]Generation[[:space:]]with[[:space:]]Hidden[[:space:]]Representation[[:space:]]Transformations/4c66b10a-fd66-4289-847c-ba27e096727b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Controlling[[:space:]]Styles[[:space:]]in[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Activation[[:space:]]Prompt/c0680bdc-4c76-47f5-ae4c-4a0733d3fb0c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Correction[[:space:]]of[[:space:]]Errors[[:space:]]in[[:space:]]Preference[[:space:]]Ratings[[:space:]]from[[:space:]]Automated[[:space:]]Metrics[[:space:]]for[[:space:]]Text[[:space:]]Generation/f9517422-ff31-4ee6-b543-974f1a3bcbec_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cost-effective[[:space:]]Distillation[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models/265abe14-6ef5-4eb2-a29c-3219a00ee31b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Counterfactual[[:space:]]Probing[[:space:]]for[[:space:]]the[[:space:]]Influence[[:space:]]of[[:space:]]Affect[[:space:]]and[[:space:]]Specificity[[:space:]]on[[:space:]]Intergroup[[:space:]]Bias/b14b15cf-04a6-4ab7-a421-a395d36324f2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Counterfactuals[[:space:]]of[[:space:]]Counterfactuals_[[:space:]]a[[:space:]]back-translation-inspired[[:space:]]approach[[:space:]]to[[:space:]]analyse[[:space:]]counterfactual[[:space:]]editors/3a3957f5-d5b3-4162-9856-974e40ae3d89_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Coupling[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]with[[:space:]]Logic[[:space:]]Programming[[:space:]]for[[:space:]]Robust[[:space:]]and[[:space:]]General[[:space:]]Reasoning[[:space:]]from[[:space:]]Text/a7aeff65-0fe1-4520-b8dd-3902f7f2e3b0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Critic-Guided[[:space:]]Decoding[[:space:]]for[[:space:]]Controlled[[:space:]]Text[[:space:]]Generation/81402f2c-cabd-453f-8428-d7dd6f5f45ac_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross[[:space:]]Encoding[[:space:]]as[[:space:]]Augmentation_[[:space:]]Towards[[:space:]]Effective[[:space:]]Educational[[:space:]]Text[[:space:]]Classification/1cb6dcbd-a06c-4faa-a7d8-8437c9243be8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-Domain[[:space:]]Argument[[:space:]]Quality[[:space:]]Estimation/edd3df0c-d46b-4ae6-94b9-810e4dce4c9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-Lingual[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]for[[:space:]]Answer[[:space:]]Sentence[[:space:]]Selection[[:space:]]in[[:space:]]Low-Resource[[:space:]]Languages/1c072d91-6caa-4b05-8bcb-98f19a996249_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-Lingual[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Prompt[[:space:]]for[[:space:]]Low-Resource[[:space:]]Languages/25111c68-eed8-430a-a63d-eebb12c573d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-Lingual[[:space:]]Transfer[[:space:]]with[[:space:]]Target[[:space:]]Language-Ready[[:space:]]Task[[:space:]]Adapters/590ff230-f4d2-4c89-bbd2-57ac17933f66_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-lingual[[:space:]]AMR[[:space:]]Aligner_[[:space:]]Paying[[:space:]]Attention[[:space:]]to[[:space:]]Cross-Attention/c6ec5440-480a-49de-aa43-cbde925d1346_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Cross-task[[:space:]]Knowledge[[:space:]]Transfer[[:space:]]for[[:space:]]Extremely[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Text[[:space:]]Classification/c8f8c5aa-39ea-40f6-b9de-6c91f7a6ba3d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Curating[[:space:]]Datasets[[:space:]]for[[:space:]]Better[[:space:]]Performance[[:space:]]with[[:space:]]Example[[:space:]]Training[[:space:]]Dynamics/4d9b1605-a43c-46f6-a166-6873f3f856db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/D-CALM_[[:space:]]A[[:space:]]Dynamic[[:space:]]Clustering-based[[:space:]]Active[[:space:]]Learning[[:space:]]Approach[[:space:]]for[[:space:]]Mitigating[[:space:]]Bias/afde61a8-fe10-4965-8479-c70148009a80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DEnsity_[[:space:]]Open-domain[[:space:]]Dialogue[[:space:]]Evaluation[[:space:]]Metric[[:space:]]using[[:space:]]Density[[:space:]]Estimation/40d36ed3-20d6-45bb-8fb7-8aa55d911b82_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DKAF_[[:space:]]KB[[:space:]]Arbitration[[:space:]]for[[:space:]]Learning[[:space:]]Task-Oriented[[:space:]]Dialog[[:space:]]Systems[[:space:]]with[[:space:]]Dialog-KB[[:space:]]Inconsistencies/b221ac28-aa5f-4416-8f5a-0f37b594d14f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DLAMA_[[:space:]]A[[:space:]]Framework[[:space:]]for[[:space:]]Curating[[:space:]]Culturally[[:space:]]Diverse[[:space:]]Facts[[:space:]]for[[:space:]]Probing[[:space:]]the[[:space:]]Knowledge[[:space:]]of[[:space:]]Pretrained[[:space:]]Language[[:space:]]Models/eeb16766-7c13-4756-baad-fbd788f05546_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DMLM_[[:space:]]Descriptive[[:space:]]Masked[[:space:]]Language[[:space:]]Modeling/1cfec941-5a85-4563-9aa7-b6dbcf333d14_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DP-BART[[:space:]]for[[:space:]]Privatized[[:space:]]Text[[:space:]]Rewriting[[:space:]]under[[:space:]]Local[[:space:]]Differential[[:space:]]Privacy/cc468bce-8339-4836-98cf-48dfab039b0e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DSPM-NLG_[[:space:]]A[[:space:]]Dual[[:space:]]Supervised[[:space:]]Pre-trained[[:space:]]Model[[:space:]]for[[:space:]]Few-shot[[:space:]]Natural[[:space:]]Language[[:space:]]Generation[[:space:]]in[[:space:]]Task-oriented[[:space:]]Dialogue[[:space:]]System/aad8557f-c209-4d81-9cde-1703b3231533_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DSP_[[:space:]]Discriminative[[:space:]]Soft[[:space:]]Prompts[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Entity[[:space:]]and[[:space:]]Relation[[:space:]]Extraction/3d440a89-46fe-4618-a123-fd3355a4f7c6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DUB_[[:space:]]Discrete[[:space:]]Unit[[:space:]]Back-translation[[:space:]]for[[:space:]]Speech[[:space:]]Translation/ee55f951-6e95-4cd4-91e9-06a9789b84ab_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Low-Resource[[:space:]]Keyphrase[[:space:]]Generation/bfb0d361-2f82-4bb3-8cfc-1d9a0c30958b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Data[[:space:]]Sampling[[:space:]]and[[:space:]](In)stability[[:space:]]in[[:space:]]Machine[[:space:]]Translation[[:space:]]Evaluation/b7905956-85f1-4e80-b5f1-c64e6f94aa98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Data-Efficient[[:space:]]Finetuning[[:space:]]Using[[:space:]]Cross-Task[[:space:]]Nearest[[:space:]]Neighbors/53aa6bd8-b2b8-491c-8fb0-41aa0e44d63b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Data-Efficient[[:space:]]French[[:space:]]Language[[:space:]]Modeling[[:space:]]with[[:space:]]CamemBERTa/b1f59b13-e450-4b6d-8a5f-8e934e90fb9d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/DePlot_[[:space:]]One-shot[[:space:]]visual[[:space:]]language[[:space:]]reasoning[[:space:]]by[[:space:]]plot-to-table[[:space:]]translation/d26ecb38-9865-4528-9ec3-70dfa8c3a536_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Debiasing[[:space:]]should[[:space:]]be[[:space:]]Good[[:space:]]and[[:space:]]Bad_[[:space:]]Measuring[[:space:]]the[[:space:]]Consistency[[:space:]]of[[:space:]]Debiasing[[:space:]]Techniques[[:space:]]in[[:space:]]Language[[:space:]]Models/a9777a7b-6646-4124-87ff-6ba7ec88155f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Decker_[[:space:]]Double[[:space:]]Check[[:space:]]with[[:space:]]Heterogeneous[[:space:]]Knowledge[[:space:]]for[[:space:]]Commonsense[[:space:]]Fact[[:space:]]Verification/4b1d0827-a3c0-4745-9abc-2ca9cac4d5c5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Decouple[[:space:]]knowledge[[:space:]]from[[:space:]]paramters[[:space:]]for[[:space:]]plug-and-play[[:space:]]language[[:space:]]modeling/5bd03a64-ca5f-4ce9-b4e2-801af535b03b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Deep[[:space:]]Equilibrium[[:space:]]Non-Autoregressive[[:space:]]Sequence[[:space:]]Learning/0b65959f-c2d8-4bb5-85f3-25a7b15005af_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Deep[[:space:]]Span[[:space:]]Representations[[:space:]]for[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/723978ee-bbb8-495a-8d8c-94880d1fb3a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Deeply[[:space:]]Coupled[[:space:]]Cross-Modal[[:space:]]Prompt[[:space:]]Learning/5cbfd784-20c8-43aa-9536-e638a1b3531d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Defending[[:space:]]against[[:space:]]Insertion-based[[:space:]]Textual[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]via[[:space:]]Attribution/fdca1ae0-d1e8-417f-a0ab-cc1c4dc42fbb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Define,[[:space:]]Evaluate,[[:space:]]and[[:space:]]Improve[[:space:]]Task-Oriented[[:space:]]Cognitive[[:space:]]Capabilities[[:space:]]for[[:space:]]Instruction[[:space:]]Generation[[:space:]]Models/f3a90270-f098-41cf-886e-6143a5ef25c7_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Delving[[:space:]]into[[:space:]]the[[:space:]]Openness[[:space:]]of[[:space:]]CLIP/ed1bf016-42f7-4d79-bc29-546fb1136f00_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Denoising[[:space:]]Enhanced[[:space:]]Distantly[[:space:]]Supervised[[:space:]]Ultrafine[[:space:]]Entity[[:space:]]Typing/c5172952-75a5-4f7f-bd8c-cde461dd15d3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Detecting[[:space:]]Adversarial[[:space:]]Samples[[:space:]]through[[:space:]]Sharpness[[:space:]]of[[:space:]]Loss[[:space:]]Landscape/4b2e1f19-b274-4374-9187-0a2e5928490c_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Common Law Annotations_ Investigating the Stability of Dialog System Output Annotations/954eaa6a-8529-416a-8e69-2af9134bed4d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa07e96f8cdbafa8a82c6874603f517d3d58b32bce39be45d92f7ed74ab26ac0
+size 7660748

2023/Common Law Annotations_ Investigating the Stability of Dialog System Output Annotations/full.md

@@ -0,0 +1,922 @@
+# Common Law Annotations: Investigating the Stability of Dialog System Output Annotations
+
+Seunggun Lee $^{1}$  Alexandra DeLucia $^{2}$  Nikita Nangia $^{1}$  Praneeth S. Ganedi $^{1}$  Ryan Min Guan $^{3}$  Rubing Li $^{1}$  Britney A. Ngaw $^{1}$  Aditya Singhal $^{1}$  Shalaka Vaidya $^{1}$  Zijun Yuan $^{1}$  Lining Zhang $^{1}$  Joao Sedoc $^{1}$
+
+$^{1}$ New York University  $^{2}$ Johns Hopkins University  $^{3}$ University of Pennsylvania jsedoc@stern.nyu.edu
+
+# Abstract
+
+Metrics for Inter-Annotator Agreement (IAA), like Cohen's Kappa, are crucial for validating annotated datasets. Although high agreement is often used to show the reliability of annotation procedures, it is insufficient to ensure validity or reproducibility. While researchers are encouraged to increase annotator agreement, this can lead to specific and tailored annotation guidelines. We hypothesize that this may result in diverging annotations from different groups. To study this, we first propose the Lee et al. Protocol (LEAP), a standardized and codified annotation protocol. LEAP strictly enforces transparency in the annotation process, which ensures reproducibility of annotation guidelines. Using LEAP to annotate a dialog dataset, we empirically show that while research groups may create reliable guidelines by raising agreement, this can cause divergent annotations across different research groups, thus questioning the validity of the annotations. Therefore, we caution NLP researchers against using reliability as a proxy for reproducibility and validity.
+
+https://github.com/jsedoc/common-law-annotations
+
+# 1 Introduction
+
+The acquisition of reliable, valid, and reproducible human annotations is an essential component of Natural Language Processing (NLP) research. However, human annotations are inherently subjective (Basile et al., 2021) and each annotator has their own biases (Paun et al., 2022). To overcome this subjectivity, research groups aim to develop annotation guidelines that increase Inter-annotator Agreement (IAA) among annotators, also known as inter-rater reliability. Reliability—the level of agreement between the annotators—is a necessary, but not sufficient condition for reproducibility (Artstein, 2017). If the precise details
+
+of the annotation process—from creating the annotation guidelines to executing the annotations themselves—are not transparent, the annotations may not be reproducible. Furthermore, high reliability does not guarantee validity—the extent to which annotations accurately capture what is intended to be measured (Paun et al., 2022).
+
+To address these challenges, we first propose Lee et al. Protocol (LEAP) a codified annotation guideline creation process that standardizes the way research groups create, publicize, and implement annotation guidelines. LEAP ensures transparency in the annotation process through its step-by-step procedure, which is crucial to allow for better reproducibility and cross-paper analyses.
+
+Second, we use LEAP to investigate agreement by having pairs of researchers simulate the annotation procedure of different research groups on a common dataset, in order to observe the change in agreement within and between these groups. Within the simulation, we observe that each group creates their own unique guidelines, despite working on the same dataset and annotation categories. We leverage the metaphor of a common law, in which country/region-specific laws are based on precedent, much like researchers agreeing on common rules for edge cases to increase agreement. Similar to common laws differing between countries, the rules governing annotation guidelines can become increasingly research group-specific and divergent from other groups, as each group strives to raise their IAA. After developing annotation guidelines, we analyze if these observations persist when crowdsourcing the data with each guideline.
+
+While LEAP has broader applications, here we apply it to a conversational AI task, where common human annotation metrics include Appropriateness, Information content of output, and Humanlikeness (Howcroft et al., 2020). The popularity and recent advances in dialog agents, such as OpenAI's GPT-
+
+3 (Brown et al., 2020), ChatGPT, $^{1}$  and YouChat $^{2}$  motivated us to showcase our method in the dialog domain.
+
+In our investigation, we ask the following research questions:
+
+1. How are the agreement levels different for researchers within and across groups?  
+2. Do groups converge or diverge in their annotation guidelines?  
+3. Which groups are able to get the crowdsourced workers to agree most? Is it the same as the other groups?  
+4. Do crowdsourced workers converge or diverge within and between groups?
+
+Ultimately, we make the following contributions:
+
+- Empirically show that while groups may create reliable guidelines by artificially raising agreement, this can lead to divergent annotations across different research groups, thus questioning annotation validity.  
+- Propose LEAP as a standardized and transparent annotation protocol which ensures reproducibility of annotation guidelines, while also allowing for deeper analysis of validity influenced by divergent annotation guidelines.
+
+# 2 Related Work
+
+This paper contributes to the ongoing discourse on the 'science' of annotations (Hovy and Lavid, 2010). Similar to Hober et al. (2023), we call for improved reliability and transparency in annotations.
+
+# 2.1 Reporting Pitfalls & Errors
+
+The NLP / NLG community generally lacks error reporting (van Miltenburg et al., 2021). Agreement studies and works involving annotations are no exception to this problem. We assert that papers should report the caveats of their work, especially regarding agreement analysis, which we believe makes research more robust. We offer a standardized solution through LEAP, where our protocol ensures each published work exposes its entire annotation life-cycle.
+
+# 2.2 Annotation Protocols
+
+The benefits of crowdsourcing methods are widely recognized and used in fields beyond NLP, including healthcare studies (Hamilton et al., 1994) as
+
+well as Psychology (Cuccolo et al., 2021). In particular, the Psychology research community has established notable researcher crowdsourcing initiatives, such as CREP (Grahe et al., 2020), the Pipeline Project (Schweinsberg et al., 2016), and Psi Chi's Network for International Collaborative Exchange: Crowd component (NICE: Crowd) (Cuccolo et al., 2022), which outline standardized practices and methodologies to ensure quality data collection.
+
+Within the NLP field, there are several annotation protocols that outline steps within the annotation development cycle. The MATTER cycle (Model, Annotate, Train, Test, Evaluate, Revise) offers a high-level outline for collecting annotations to train and develop machine learning models (Pustejovsky and Stubbs, 2012). The MAMA (Model-Annotate-Model-Annotate) cycle-a subsection of the MATTER cycle-describes the iterative procedure of refining guidelines and collecting annotations to arrive at an optimal annotation model (Pustejovsky and Stubbs, 2012). The CASCADES model further extends the Model and Revise portion of MATTER, with the steps Conceptual Analysis, Abstract Syntax, Semantics, and Concrete Syntax (Bunt et al., 2010). For a deeper analysis of these protocols and their implementations, see Artstein (2017). With GENIE, Khashabi et al. (2022) address reproducibility concerns by providing a platform to run and study annotation results across a variety of text generation tasks.
+
+While such annotation protocols help standardize the annotation procedures, they do not entirely enforce the total transparency of the annotation procedures. To the best of our knowledge, LEAP is the first annotation protocol to strictly require complete transparency in the annotation guideline creation process through recorded discussions and transcripts to ensure full reproducibility and effective cross-paper analysis.
+
+# 2.3 Divergent Annotation Guidelines
+
+Though divergent annotation guidelines between research groups may seem natural due to each group's unique research purpose, this often occurs among research groups who have similar purposes (e.g., evaluating a new dialogue system). Researchers would benefit by using consistent standardized annotation approaches.
+
+For example, numerous papers created their own definitions for the category of Appropriateness (Reiter et al., 2000; van Deemter, 2000), Information
+
+![](images/dd5efbb7209e2dd12f3ead33c7defe43071f8b742545322935db42d557455f31.jpg)  
+Figure 1: A flowchart of LEAP used for our experiments.
+
+content of outputs (Carenini and Cheung, 2008; Filippova and Strube, 2008), and Humanlikeness (Agirrezabal et al., 2013; Cercas Curry et al., 2015) (See Appendix A.3 for more examples). Furthermore, though papers may use annotation categories that are different verbatim, the categories often overlap in meaning and purpose (Finch and Choi, 2020).
+
+# 2.4 Disagreement in Annotations
+
+Basile et al. (2021) emphasizes the importance of observing and embracing inherent disagreement in annotation tasks, arguing that focusing on a single 'ground truth' reference obscures the complexity and subjectivity of human-annotated data (Pavlick
+
+and Kwiatkowski, 2019; Uma et al., 2022).
+
+In fact, in SummEval (Fabbri et al., 2021) they found that crowd worker annotations had reasonable IAA but were uncorrelated to expert annotators who also had high IAA. This suggests a flaw in the current annotation paradigm. Instead, we propose in our work that a pair of researchers should first converge with high IAA on a subset of the dataset. Then the pair should create the instructions and design for the crowd annotation task and validate the agreement.
+
+In our work, we extend this study of disagreement by empirically illustrating how artificially eradicating irreconcilable disagreement can harm accuracy (and thus potentially harm validity).
+
+# 3 Experiment Design
+
+# 3.1 LEAP
+
+Figure 1 illustrates the codified steps of LEAP for our experiment. In the following paragraphs, we explain the core components of LEAP. While the procedure outlined below is tailored for dialog annotations, the overall method can be adapted to other tasks (see Appendix subsection A.1).
+
+Parameters To customize LEAP for a specific annotation task, several parameters need to be selected. These parameters include:
+
+- Minimum and maximum number of rounds  
+- Agreement criteria and threshold  
+Common law discussion time limit  
+- Number of researchers involved (minimum of 2)  
+Number of items per small iteration  
+- Number of items for the larger annotation
+
+Our advice to practitioners is that while each of these can be modified during an annotation process, the best practice is to set them a priori based on a smaller pilot, prior studies, or budget limitations.
+
+Annotations Annotations are done independently, on the same subset of data. During the annotation, annotators are not allowed to communicate with each other. After each iteration of annotations, the agreement score is calculated for each annotation category. The agreement scores are shared with the annotators.
+
+Annotation Discussions Each pair of annotators in a research group use discussions to walk through and compare their annotations. During discussions, annotators are asked to resolve edge cases that are causing disagreement, ultimately working towards a shared understanding of each category's annotation guidelines.
+
+All discussions are conducted using a recorded video-conference platform, such as Zoom, $^{3}$  to ensure full transparency of the annotation process. Discussions are limited to a pre-specified amount of time. As researchers compare individual annotation examples, screen-share is enabled to make the process transparent, while transcript tools are enabled to allow for efficient analysis post-experiment. $^{4}$
+
+The quintessential idea for the records is to ensure that the decisions made during the meeting are documented as they may provide insights into
+
+construct validity and also help in understanding survey design. Since recording may not be available for all situations (e.g., automatic transcription does not support all languages), an alternative is to maintain detailed notes during the discussions.
+
+Rounds & Iterations Prior to developing the final annotation guidelines, LEAP requires researchers to annotate multiple subsets of data.
+
+Each round consists of a subset of a given dataset. Each annotation session is termed as an iteration. After a pair of researchers complete an iteration of annotations, the agreement score for each annotation category is calculated. The average of the agreement scores across the annotation categories is used to compare against a pre-designated threshold level of agreement.
+
+If the category average agreement score meets the threshold, the researchers move on to the next round of annotations. This next round uses a new subset of the dataset. However, if the category average agreement score does not meet the threshold, the researchers are unable to move to the next round of annotations. Rather, the researchers discuss the most recent iteration of annotations to fine-tune their shared understanding of the annotation categories. Then the researchers conduct the next iteration of annotations. In the new iteration, researchers annotate the same subset of data, however, the presentation order is shuffled. Iterations allow researchers to test their level of convergence. This step is repeated until the researchers are able to meet their desired threshold, upon which they move on to the next round of annotations.
+
+Post Convergence Round Once the researchers complete their rounds of annotations, they annotate a larger set of new items. This larger round of items is used to compare the crowd worker ratings with the researchers and evaluate consistency over a large set of annotations.
+
+Creating the Annotation Guidelines The final component of the protocol is creating the annotation guidelines. Similar to the discussions, this process is made transparent through recorded screen share and live transcripts.
+
+There are several benefits to such an iterative annotation procedure. First, researchers are able to find and fix pitfalls and mistakes in the anno
+
+tation process by experiencing them directly. Furthermore, through the iterative process, researchers are able to systemically fine-tune their annotations to construct a shared understanding of the annotation categories. Finally, the iterative process allows the researchers to retroactively analyze the discussions conducted after each annotation session in a structured manner.
+
+# 3.2 Experimental Design
+
+Data For this task, we generated model responses using prompts from the English as a Second Language (ESL) (Sedoc et al., 2019) and Daily Dialog (Li et al., 2017) evaluation sets (1,323 prompts). For each prompt, we generated model responses using eight state-of-the-art conversational models, including DialoGPT (Zhang et al., 2020), GPT-3 (Brown et al., 2020), Plato2 (Bao et al., 2021), and BlenderBot 2 (Weston and Shuster, 2021; Komeili et al., 2022; Xu et al., 2022). In total, we created 11,907 prompt-response pairs. The prompts and model responses have been detokenized to avoid revealing the model origins to the annotator. We used the dialog prompts and the language generation systems within their intended usage. For more information on the model parameters, see Appendix A.4.
+
+Instructions The experiment followed the LEAP architecture. The goal of each group was to create annotation guidelines that would help other annotators annotate conversational text data as similarly as possible. The annotations consisted of static evaluations, as they are one of the most used forms of human evaluations in NLP (Finch and Choi, 2020). Following Howcroft et al. (2020), we provided the following base definitions for three annotation categories:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Humanlikeness: The degree to which the output could have been produced by a human.
+
+We intentionally kept the category definitions simple to give each group freedom in devising their own annotation guidelines. See Appendix Figure 6 for an example of the prompt and response annotated by the researchers.
+
+See Appendix Figure 8 for the tabular step-by-step instructions-created using LEAP-shared with all researcher annotators. For specific instructions
+
+on creating annotation guidelines, shared with all researcher annotators, see Appendix Figure 7.
+
+LEAP parameters In order to maintain intergroup consistency, each group was instructed to use a five-point ordinal scale. For our agreement criteria, we chose linear Cohen's  $\kappa$  as it is commonly used. We ran a small pilot and estimated that the Cohen's  $\kappa$ $95\%$  confidence interval was  $\pm 0.1$  with 50 annotations and  $\pm 0.05$  with 400 annotations. Cohen's  $\kappa$  of 0.6 to 0.8 is commonly regarded as a threshold for sufficient inter-annotator agreement in NLP research (Landis and Koch, 1977) thus we chose a category average Cohen's  $\kappa$  of 0.7 as the threshold. To time-bound the process, we chose a minimum of 2 rounds and a maximum of 5 rounds. In our pilot, we also found that 30 minutes was sufficient to discuss annotation differences.
+
+Groups We simulated the process of six individual research groups (Group 1-6) defining guidelines for human annotation of conversational data. Each group consisted of two researchers. The group pairings had diverse members in terms of gender identity and annotation experience.[7]
+
+# 3.3 Crowdsourced Annotation Parameters
+
+Once all the annotation guidelines have been created, we used Amazon Mechanical Turk (MTurk) to collect crowdsourced data. For our experiment, groups did not iterate over the annotation guidelines with crowd workers.
+
+Instructions Each crowd worker was given the following instructions:
+
+The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt.
+
+Then the annotators were given the annotation guideline based on the group task chosen (see Figures 9 to 14 in the Appendix).
+
+All MTurk tasks were deployed using the same portion of the dataset as the round of 400 prompts and responses that were annotated by the researchers. This choice was made because the round
+
+of 400 annotations was the latest set of annotations done by the researchers, meaning the researchers' annotations were most calibrated with the annotation guidelines.9 Groups 1 and 2 used the full LEAP protocol with iterations.
+
+# 3.4 Testing Iteration-Free LEAP
+
+While the iterations in LEAP give researchers the opportunity to converge on their common law annotation guidelines, we acknowledge that this may require additional time and resources. Furthermore, it reduces the independent relationship between annotations. Thus, we tested an iteration-free version of LEAP (see Figure 5 for a flowchart).
+
+The iteration-free version of LEAP excludes the iteration component. If a group is unable to reach the pre-designated agreement threshold, they move on to the next round of annotations. This allows researchers to annotate more data while converging; however, they cannot discuss a subset of data multiple times. Iteration-free LEAP favors coverage over convergence. A new round of annotations consists of a new subset of data. Groups 3, 4, 5, and 6 used the iteration-free LEAP.
+
+# 4 Results & Discussion
+
+# 4.1 Agreement Analysis - LEAP
+
+Within Group By using the iterative annotation procedure of LEAP, Group 1 and Group 2 were able to achieve a high level of agreement on the second iteration of the second round of annotations. Figure 2 illustrates the change in agreement for Groups 1 and 2.
+
+We also observed a drop in agreement for both groups when moving from round 1 to round 2. This is expected, as the change in annotated data introduces new edge cases, causing divergence between annotators. However, as both groups were able to calibrate their annotations via the iterations in round 1, round 2 required substantially fewer amounts of iterations to achieve the threshold of 0.7.
+
+Between Groups Taking advantage of the standardized annotation protocol codified through LEAP, we analyzed the changes in the agreement between annotators of different groups. Figure 3 illustrates the changes in agreement for annota
+
+![](images/09400688475d4430faede9b1473f9efc3a4c328654a950760a96fbb440feb2ec.jpg)
+
+![](images/a7e72ad354e15bf1ef1c179350dc7366f92ffaf0bb963205eff863494802c898.jpg)  
+Figure 2: Agreement scores for Groups 1 (above) and 2 (below) with using LEAP.
+
+tors within the same group and between different groups.
+
+In round 1 and round 2, for all three categories, within-group agreement—that is the level of agreement between annotators of the same group—was relatively higher than between-group agreement, or the level of agreement between annotators of different groups. Such observation suggests that raising agreement levels through fine-tuned annotation guidelines can cause divergence across different research groups.
+
+Interestingly, we observed a relatively higher level of between-group agreement for Appropriateness, despite the fact that researchers in Group 1 and Group 2 never communicated with one another. This suggests that certain annotation categories, such as Appropriatness, have a stronger shared construct than others.
+
+# 4.2 Agreement Analysis - Iteration-Free LEAP
+
+Groups 3, 4, 5, and 6 tested the iteration-free LEAP. None of the groups were able to reach the designated threshold of an average Cohen's  $\kappa > 0.7$ . In addition, we found supporting evidence of divergence across annotators of different groups.
+
+![](images/efdb8f00dd423cb48cfbe13ad62d2875a32d2df3b406c1c26273e1dc40ad0ebd.jpg)  
+Figure 3: Contingency table of annotations for Group 1 (researchers 1 and 2) and Group 2 (researchers 3 and 4) - From top to bottom: Appropriateness, Information content of outputs, and Humanlikeness. The graphs for Round 1 and Round 2 show the figures for the final iteration of each round. Round 400 indicates the final round of annotations in LEAP with 400 items. Red borders indicate within-group agreement. Darker blue indicates higher agreement (Cohen's  $\kappa$ ).
+
+We present the detailed results in Appendix A.9. Remarkably, the iterations were important to solidify common law rules, since moving to new samples (i.e., new rounds) caused more confusion and the rules did not ground well.
+
+# 4.3 Annotation Guidelines
+
+We analyzed each group's annotation guideline and its creation process by examining the Zoom recordings of discussions. For the final version of the guidelines for all groups see Figures 9 to 14 in the Appendix.
+
+Appropriateness The group discussion transcripts and written guidelines showed that the different groups took a similar approach when annotating Appropriateness. Primarily, all groups based their Appropriateness score on whether the model response "made sense" in relation to the prompt itself. Also, all groups considered the contextual relevance of the response in relation to the prompt. This reinforces our observation that annotators overall
+
+had a strong shared construct of Appropriateness, which resulted in high levels of agreement for the category.
+
+Information content of output Unlike Appropriateness, agreement levels between groups for Information content of output were relatively low. While Group 1 gauged the category based on the specificity of the information provided by the response, Group 2 based the category score on the length of the response (ie. the number of sentences), as well as the correctness of the response (ie. if the information provided is factually correct). Such divergences in annotation guidelines explain the low level of agreement between annotators of different groups.
+
+We conducted a similar analysis on Groups 3, 4, 5, and 6. As discussed in Appendix A.9, we observed two distinct silos of convergence in agreement. The annotation discussion transcripts revealed that Group 3 and Group 6 quantified the
+
+amount of new information stated in the response to score Information of content, while Group 4 and Group 5 did not. For example, if a response did not reveal any new information, but was relevant to the prompt, Group 4 and Group 5 would give at least a 3 for Information content of output. However, as Group 3 and Group 6 focused on the quantity of new information when annotating Information content of output, they would give it a low score.
+
+Furthermore, Groups 3 and 6 solely looked at the response field to judge Information content of output, meaning a short, generic response would receive a low score for this category. In comparison, Groups 4 and 5 created guidelines that looked at both the prompt and response to judge the level of information given, meaning a short, generic response could still receive a higher score depending on the broader context.
+
+The divergence in annotation guidelines not only explains the low average agreement between groups for Information content of output but also uncovers why different clusters of agreement occur between certain groups.
+
+Humanlikeness While both Groups 1 and 2 based Humanlikeness on whether a real human would have said the response, both groups had diverging approaches for the annotation category. Group 1 emphasized that the annotator should not consider the appropriateness of the response when judging Humanlikeness. On the other hand, Group 2 simply evaluated whether a real human could have said the response, while also taking into consideration grammatical errors.
+
+For Groups 3, 4, 5, and 6 two separate clusters of agreement occurred between the groups—one between Group 3 and Group 6, another between Group 4 and Group 5. The clusters of agreement can be attributed to the differing annotation procedures that emerged between these silos. Group 3 and Group 6 annotated by ignoring the prompt and judging solely the Humanlikeness of the response. On the other hand, Group 4 and Group 5 took into consideration the response's context. For example, following Group 3 and Group 6's guidelines, even if the response was a complete replica of an utterance in the prompt, the response could receive a high score for Humanlikeness. In contrast, if the response repeated content from the prompt, Group 4 and Group 5 gave the response a low Humanlikeness score.
+
+The two different interpretations of a category re
+
+inforce the notion that a "ground truth" annotation value is difficult to reach, especially for categories that have less of a shared construct - like Information content of output and Humanlikeness.
+
+# 4.4 Crowdsourced Data
+
+In order to examine how diverging annotation guidelines impact agreement levels for crowdsourced annotations, we employed batches of Human Intelligence Tasks (HITs) on Amazon Mechanical Turk (MTurk). We recruited and filtered MTurk workers who were able to achieve a category average  $\kappa > 0.7$  agreement with the researchers on a pilot HIT. These workers were then given a larger MTurk task of annotating the same set of 400 prompt-response questions from the guideline creations, with 55 prompt-response questions per HIT (for details see subsection A.7).
+
+Agreement Between Researchers & Crowdsourced Workers The average agreement between the crowdsourced workers and the researcher for each Group is illustrated in Figure 15 in the Appendix. For all Groups except Group 1, Appropriateness was the category with the highest agreement between the researchers and the HIT workers. Overall, HIT workers who used Group 4's guideline had the highest average agreement scores with the Group's researchers. Furthermore, the variable levels of agreement for LEAP indicate that annotations are relatively noisy even with a well-defined protocol.
+
+Group 1 & Group 2 We calculated the agreement between MTurk annotators of the same group's annotation guidelines, as well as the agreement between annotators of Groups 1 and 2 (see Table 1.
+
+<table><tr><td>Groups</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>Group 1</td><td>0.37</td><td>0.09</td><td>0.19</td></tr><tr><td>Group 2</td><td>0.58</td><td>0.20</td><td>0.30</td></tr><tr><td>Between Groups 1 &amp; 2</td><td>0.37</td><td>0.13</td><td>0.09</td></tr></table>
+
+Table 1: IAA within and between crowd workers using Group 1's and Group 2's guidelines
+
+Of the three categories, again, Appropriateness had the strongest shared construct with the highest level of agreement. Group 1 and Group 2 had higher agreement within groups for Humanlikeness compared to the IAA from between Groups 1 and 2. As with the researcher annotators, crowd workers
+
+who followed different annotation guidelines were unable to achieve high agreement.
+
+Groups 3, 4, 5, & 6 Similarly, we analyzed the differences in agreement levels for crowd workers using guidelines created by Groups 3, 4, 5, and 6 (see Table 2 and Table 3).
+
+<table><tr><td>Groups</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>Group 3</td><td>0.38</td><td>0.16</td><td>0.25</td></tr><tr><td>Group 4</td><td>0.46</td><td>0.54</td><td>0.56</td></tr><tr><td>Group 5</td><td>0.38</td><td>0.30</td><td>0.23</td></tr><tr><td>Group 6</td><td>0.47</td><td>0.22</td><td>0.43</td></tr><tr><td>Average</td><td>0.42</td><td>0.31</td><td>0.37</td></tr></table>
+
+Table 2: IAA within group for crowdsourced workers using guidelines created by Groups 3, 4, 5, and 6.  
+
+<table><tr><td>Groups</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>Groups 3 &amp; 4</td><td>0.38</td><td>0.20</td><td>0.20</td></tr><tr><td>Groups 3 &amp; 5</td><td>0.32</td><td>0.23</td><td>0.18</td></tr><tr><td>Groups 3 &amp; 6</td><td>0.37</td><td>0.27</td><td>0.23</td></tr><tr><td>Groups 4 &amp; 5</td><td>0.32</td><td>0.17</td><td>0.22</td></tr><tr><td>Groups 4 &amp; 6</td><td>0.3</td><td>0.15</td><td>0.24</td></tr><tr><td>Groups 5 &amp; 6</td><td>0.57</td><td>0.27</td><td>0.27</td></tr><tr><td>Average</td><td>0.38</td><td>0.22</td><td>0.22</td></tr></table>
+
+Table 3: IAA between groups for crowdsourced workers using guidelines created by Groups 3, 4, 5, and 6.
+
+Similar to Groups 1 and 2, crowd workers for Groups 3, 4, 5, and 6 had relatively higher agreement within group compared to between different groups.
+
+# 5 Conclusion
+
+In this paper, we caution NLP researchers against using reliability as a proxy for reproducibility and validity. While LEAP does not strictly enforce validity, it creates transparency in the "common law" annotation rules. This transparency can enable others to assess the validity of the choices. We propose and encourage researchers to use LEAP as a solution to ensure reproducibility by rendering the annotation protocol completely transparent while allowing for deeper cross-paper analysis on validity through the standardized annotation procedure.
+
+Using LEAP, we simulated a parallel series of independent annotation procedures, illustrating how even if a research group achieves agreement, their agreement with annotators from differ
+
+ent groups can be low for certain categories due to diverging annotation guidelines.
+
+Overall, research groups should use agreement metrics with care. While a high agreement score is often a community-recognized threshold required for research groups to publish their annotated datasets, research groups should be aware of the pitfalls in raising agreement metrics. Furthermore, research groups should follow a standardized annotation guideline creation process, such as LEAP, and make the entire procedure transparent. With such standardization and transparency, we will be able to better understand the issues associated with simply using agreement metrics as the main threshold to cross for publications.
+
+# 6 Limitations
+
+LEAP requires access to a telecommunication platform, such as Zoom, which can record, screen-share, and save live transcripts of the discussions. The dialogue data used in the annotations, as well as the annotation categories and their respective guidelines, were all in English. Furthermore, the researcher participants of the study were all co-authors of the paper and did not include professional annotators. We tested LEAP using only conversational dialogue. We only used three annotation categories. Though there are other protocols that could have helped in the analysis, we only experimented with LEAP and an ablation of LEAP. Some model responses may have contained bias.
+
+# Acknowledgements
+
+We thank the reviewers for their comments and suggestions. We thank Mark Liberman for suggesting the common law metaphor and for his helpful feedback. We also thank Nicholas Lourie for his help with the project. We would like to thank Claire Daniele for her help with figures and editorial support.
+
+# References
+
+Manex Agirrezabal, Bertol Arrieta, Aitzol Astigarraga, and Mans Hulden. 2013. POS-tag based poetry generation with WordNet. In Proceedings of the 14th European Workshop on Natural Language Generation, pages 162-166, Sofia, Bulgaria. Association for Computational Linguistics.
+
+Ron Artstein. 2017. Inter-annotator Agreement. In Nancy Ide and James Pustejovsky, editors, Hand
+
+book of Linguistic Annotation, pages 297-313. Springer Netherlands, Dordrecht.  
+Siqi Bao, Huang He, Fan Wang, Hua Wu, Haifeng Wang, Wenquan Wu, Zhen Guo, Zhibin Liu, and Xinchao Xu. 2021. PLATO-2: Towards Building an Open-Domain Chatbot via Curriculum Learning. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2513-2525, Online. Association for Computational Linguistics.  
+Valerio Basile, Michael Fell, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, Massimo Poesio, and Alexandra Uma. 2021. We Need to Consider Disagreement in Evaluation. In Proceedings of the 1st Workshop on Benchmarking: Past, Present and Future, pages 15-21, Online. Association for Computational Linguistics.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877-1901. Curran Associates, Inc.  
+Harry Bunt, Alex Chengyu Fang, Nancy M. Ide, and Jonathan J. Webster. 2010. A methodology for designing semantic annotation languages exploiting syntactic-semantic iso-morphisms. In Proceedings of the Second International Conference on Global Interoperability for Language Resources (ICGL 2010), Hong Kong, pages 29-46.  
+Joan Byamugisha, C. Maria Keet, and Brian DeRenzi. 2017. Evaluation of a Runyankore grammar engine for healthcare messages. In Proceedings of the 10th International Conference on Natural Language Generation, pages 105-113, Santiago de Compostela, Spain. Association for Computational Linguistics.  
+Giuseppe Carenini and Jackie C. K. Cheung. 2008. Extractive vs. NLG-based abstractive summarization of evaluative text: The effect of corpus controversiality. In Proceedings of the Fifth International Natural Language Generation Conference, pages 33-41, Salt Fork, Ohio, USA. Association for Computational Linguistics.  
+Amanda Cercas Curry, Dimitra Gkatzia, and Verena Rieser. 2015. Generating and Evaluating Landmark-Based Navigation Instructions in Virtual Environments. In Proceedings of the 15th European Workshop on Natural Language Generation (ENLG), pages 90-94, Brighton, UK. Association for Computational Linguistics.
+
+Hyungtak Choi, Siddarth K.M., Haehun Yang, Heesik Jeon, Inchul Hwang, and Jihie Kim. 2018. Self-Learning Architecture for Natural Language Generation. In Proceedings of the 11th International Conference on Natural Language Generation, pages 165-170, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+Philipp Cimiano, Janna Lüker, David Nagel, and Christina Unger. 2013. Exploiting ontology lexic a for generating natural language texts from RDF data. In Proceedings of the 14th European Workshop on Natural Language Generation, pages 10-19, Sofia, Bulgaria. Association for Computational Linguistics.  
+Jacob Cohen. 1968. Weighted kappa: Nominal scale agreement provision for scaled disagreement or partial credit. Psychological Bulletin, 70(4):213-220.  
+Kelly Cuccolo, Jon E Grahe, Martha S Zlokovich, John Edlund, rick miller, Susana Gallor, Jordan R Wagge, Kaitlyn M Werner, Albert L Ly, Fanli Jia, and et al. 2022. NICE: CROWD.  
+Kelly Cuccolo, Megan S. Irgens, Martha S. Zlokovich, Jon Grahe, and John E. Edlund. 2021. What Crowdsourcing Can Offer to Cross-Cultural Psychological Science. *Cross-Cultural Research*, 55(1):3-28. Publisher: SAGE Publications Inc.  
+Seniz Demir, Sandra Carberry, and Kathleen F. McCoy. 2008. Generating textual summaries of bar charts. In Proceedings of the Fifth International Natural Language Generation Conference on - INLG '08, page 7, Salt Fork, Ohio. Association for Computational Linguistics.  
+Jan Milan Deriu and Mark Cieliebak. 2018. Syntactic Manipulation for Generating more Diverse and Interesting Texts. In Proceedings of the 11th International Conference on Natural Language Generation, pages 22-34, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+Alexander R. Fabbri, Wojciech Krysciński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev. 2021. SummEval: Re-evaluating summarization evaluation. Transactions of the Association for Computational Linguistics, 9:391-409.  
+Abdurrisyad Fikri, Hiroya Takamura, and Manabu Okumura. 2018. Stylistically User-Specific Generation. In Proceedings of the 11th International Conference on Natural Language Generation, pages 89-98, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+Katja Filippova and Michael Strube. 2008. Dependency tree based sentence compression. In Proceedings of the Fifth International Natural Language Generation Conference on - INLG '08, page 25, Salt Fork, Ohio. Association for Computational Linguistics.
+
+Sarah E. Finch and Jinho D. Choi. 2020. Towards Unified Dialogue System Evaluation: A Comprehensive Analysis of Current Evaluation Protocols. In Proceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 236-245, 1st virtual meeting. Association for Computational Linguistics.  
+Dimitra Gkatzia, Helen Hastie, Srinivasan Janarthanam, and Oliver Lemon. 2013. Generating student feedback from time-series data using reinforcement learning. In Proceedings of the 14th European Workshop on Natural Language Generation, pages 115-124, Sofia, Bulgaria. Association for Computational Linguistics.  
+Jon E. Grahe, Kelly Cuccolo, Dana C. Leighton, and Leslie D. Cromblet Alvarez. 2020. Open Science Promotes Diverse, Just, and Sustainable Research and Educational Outcomes. Psychology Learning & Teaching, 19(1):5-20. _eprint: https://doi.org/10.1177/1475725719869164.  
+Byron B. Hamilton, Judith A. Laughlin, Roger C. Fiedler, and Carl V. Granger. 1994. Interrater reliability of the 7-level functional independence measure (FIM). Scandinavian journal of rehabilitation medicine, 26(3):115-119.  
+Vrindavan Harrison and Marilyn Walker. 2018. Neural Generation of Diverse Questions using Answer Focus, Contextual and Linguistic Features. In Proceedings of the 11th International Conference on Natural Language Generation, pages 296-306, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+Nicole Hober, Tulay Dixon, and Tove Larsson. 2023. Toward increased reliability and transparency in projects with manual linguistic coding. Corpora, 18.  
+Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The Curious Case of Neural Text Degeneration. In International Conference on Learning Representations.  
+Eduard Hovy and Julia Lavid. 2010. Towards a 'Science' of Corpus Annotation: A New Methodological Challenge for Corpus Linguistics. International journal of translation, 22(1):13-36.  
+David M. Howcroft, Anya Belz, Miruna Clinciu, Dimitra Gkatzia, Sadid A. Hasan, Saad Mahamood, Simon Mille, Emiel van Miltenburg, Sashank Santhanam, and Verena Rieser. 2020. Twenty years of confusion in human evaluation: NLG needs evaluation sheets and standardised definitions. In Proceedings of the 13th International Conference on Natural Language Generation, pages 169-182, Dublin, Ireland. Association for Computational Linguistics.  
+Stephanie Inglis. 2015. Summarising Unreliable Data. In Proceedings of the 15th European Workshop on
+
+Natural Language Generation (ENLG), pages 95-99, Brighton, UK. Association for Computational Linguistics.  
+Stephanie Inglis, Ehud Reiter, and Somayajulu Sripada. 2017. Textually Summarising Incomplete Data. In Proceedings of the 10th International Conference on Natural Language Generation, pages 228-232, Santiago de Compostela, Spain. Association for Computational Linguistics.  
+Daniel Khashabi, Gabriel Stanovsky, Jonathan Bragg, Nicholas Lourie, Jungo Kasai, Yejin Choi, Noah A. Smith, and Daniel Weld. 2022. GENIE: Toward reproducible and standardized human evaluation for text generation. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11444-11458, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-Augmented Dialogue Generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8460-8478, Dublin, Ireland. Association for Computational Linguistics.  
+Kittipitch Kuptavanich, Ehud Reiter, Kees Van Deemter, and Advaith Siddharthan. 2018. Generating Summaries of Sets of Consumer Products: Learning from Experiments. In Proceedings of the 11th International Conference on Natural Language Generation, pages 403-407, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+J. Richard Landis and Gary G. Koch. 1977. The Measurement of Observer Agreement for Categorical Data. Biometrics, 33(1):159-174. Publisher: [Wiley, International Biometric Society].  
+Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao, and Shuzi Niu. 2017. DailyDialog: A manually labelled multi-turn dialogue dataset. In Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 986-995, Taipei, Taiwan. Asian Federation of Natural Language Processing.  
+Rensis Likert. 1932. A technique for the measurement of attitudes. Archives of Psychology, 22 140:55-55.  
+Saad Mahamood and Ehud Reiter. 2011. Generating affective natural language for parents of neonatal infants. In Proceedings of the 13th European Workshop on Natural Language Generation, pages 12-21, Nancy, France. Association for Computational Linguistics.  
+Saad Mahamood and Ehud Reiter. 2012. Working with clinicians to improve a patient-information NLG system. In INLG 2012 Proceedings of the Seventh International Natural Language Generation Conference, pages 100-104, Utica, IL. Association for Computational Linguistics.
+
+Alexander Miller, Will Feng, Dhruv Batra, Antoine Bordes, Adam Fisch, Jiasen Lu, Devi Parikh, and Jason Weston. 2017. ParlAI: A Dialog Research Software Platform. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 79-84, Copenhagen, Denmark. Association for Computational Linguistics.  
+Priscilla Moraes, Kathy McCoy, and Sandra Carberry. 2014. Adapting Graph Summaries to the Users' Reading Levels. In Proceedings of the 8th International Natural Language Generation Conference (INLG), pages 64-73, Philadelphia, Pennsylvania, U.S.A. Association for Computational Linguistics.  
+Yusuke Mori, Hiroaki Yamane, Yusuke Mukuta, and Tatsuya Harada. 2019. Toward a Better Story End: Collecting Human Evaluation with Reasons. In Proceedings of the 12th International Conference on Natural Language Generation, pages 383-390, Tokyo, Japan. Association for Computational Linguistics.  
+Gabriel Murray, Giuseppe Carenini, and Raymond Ng. 2010. Generating and validating abstracts of meeting conversations: a user study. In Proceedings of the 6th International Natural Language Generation Conference. Association for Computational Linguistics.  
+Alice Oh and Howard Shrobe. 2008. Generating baseball summaries from multiple perspectives by reordering content. In Proceedings of the Fifth International Natural Language Generation Conference on - INLG '08, page 173, Salt Fork, Ohio. Association for Computational Linguistics.  
+Silviu Paun, Ron Artstein, and Massimo Poesio. 2022. Learning from Multi-Annotated Corpora. In *Statistical Methods for Annotation Analysis*, pages 147-165. Springer International Publishing, Cham.  
+Ellie Pavlick and Tom Kwiatkowski. 2019. Inherent Disagreements in Human Textual Inferences. Transactions of the Association for Computational Linguistics, 7:677-694. Place: Cambridge, MA Publisher: MIT Press.  
+James Pustejovsky and Amber Stubbs. 2012. *Natural Language Annotation for Machine Learning*. O'Reilly Media, Inc.  
+Raheel Qader, Khoder Jneid, François Portet, and Cyril Labbe. 2018. Generation of Company descriptions using concept-to-text and text-to-text deep models: dataset collection and systems evaluation. In Proceedings of the 11th International Conference on Natural Language Generation, pages 254-263, Tilburg University, The Netherlands. Association for Computational Linguistics.  
+Ehud Reiter, Albert Gatt, François Portet, and Marian van der Meulen. 2008. The importance of narrative and other lessons from an evaluation of an NLG system that summarizes clinical data. In Proceedings
+
+of the Fifth International Natural Language Generation Conference on - INLG '08, page 147, Salt Fork, Ohio. Association for Computational Linguistics.  
+Ehud Reiter, Roma Robertson, and Liesl Osman. 2000. Knowledge acquisition for natural language generation. In Proceedings of the first international conference on Natural language generation - INLG '00, volume 14, page 217, Mitzpe Ramon, Israel. Association for Computational Linguistics.  
+Sashank Santhanam and Samira Shaikh. 2019. Towards Best Experiment Design for Evaluating Dialogue System Output. In Proceedings of the 12th International Conference on Natural Language Generation, pages 88-94, Tokyo, Japan. Association for Computational Linguistics.  
+Björn Schlünder and Ralf Klabunde. 2013. Greetings generation in video role playing games. In Proceedings of the 14th European Workshop on Natural Language Generation, pages 167-171, Sofia, Bulgaria. Association for Computational Linguistics.
+
+Martin Schweinsberg, Nikhil Madan, Michelangelo Vianello, S. Amy Sommer, Jennifer Jordan, Warren Tierney, Eli Awtrey, Luke Lei Zhu, Daniel Diermeier, Justin E. Heinze, Malavika Srinivasan, David Tannenbaum, Eliza Bivolaru, Jason Dana, Clintin P. Davis-Stober, Christilene du Plessis, Quentin F. Gronau, Andrew C. Hafenbrack, Eko Yi Liao, Alexander Ly, Maarten Marsman, Toshio Murase, Israr Qureshi, Michael Schaerer, Nico Thornley, Christina M. Tworek, Eric-Jan Wagenmakers, Lynn Wong, Tabitha Anderson, Christopher W. Bauman, Wendy L. Bedwell, Victoria Brescoll, Andrew Canavan, Jesse J. Chandler, Erik Cheries, Sapna Cheryan, Felix Cheung, Andrei Cimpian, Mark A. Clark, Diana Cordon, Fiery Cushman, Peter H. Ditto, Thomas Donahue, Sarah E. Frick, Monica Gamez-Djokic, Rebecca Hofstein Grady, Jesse Graham, Jun Gu, Adam Hahn, Brittany E. Hanson, Nicole J. Hartwich, Kristie Hein, Yoel Inbar, Lily Jiang, Tehlyr Kellogg, Deanna M. Kennedy, Nicole Legate, Timo P. Luoma, Heidi Maibuecher, Peter Meindl, Jennifer Miles, Alexandra Mislin, Daniel C. Molden, Matt Motyl, George Newman, Hoai Huong Ngo, Harvey Packham, Philip S. Ramsay, Jennifer L. Ray, Aaron M. Sackett, Anne-Laure Sellier, Tatiana Sokolova, Walter Sowden, Daniel Storage, Xiaomin Sun, Jay J. Van Bavel, Anthony N. Washburn, Cong Wei, Erik Wetter, Carlos T. Wilson, Sophie-Charlotte Darroux, and Eric Luis Uhlmann. 2016. The pipeline project: Pre-publication independent replications of a single laboratory's research pipeline. Journal of Experimental Social Psychology, 66:55-67.
+
+João Sedoc, Daphne Ippolito, Arun Kirubarajan, Jai Thirani, Lyle Ungar, and Chris Callison-Burch. 2019. ChatEval: A tool for chatbot evaluation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), pages 60-65, Minneapolis, Minnesota. Association for Computational Linguistics.
+
+Advaith Siddharthan, Matthew Green, Kees van Deemter, Chris Mellish, and René van der Wal. 2012. Blogging birds: Generating narratives about reintroduced species to promote public engagement. In INLG 2012 Proceedings of the Seventh International Natural Language Generation Conference, pages 120-124, Utica, IL. Association for Computational Linguistics.  
+Alexandra N. Uma, Tommaso Fornaciari, Dirk Hovy, Silviu Paun, Barbara Plank, and Massimo Poesio. 2022. Learning from Disagreement: A Survey. J. Artif. Int. Res., 72:1385-1470. Place: El Segundo, CA, USA Publisher: AI Access Foundation.  
+Kees van Deemter. 2000. Generating vague descriptions. In Proceedings of the first international conference on Natural language generation - INLG '00, volume 14, page 179, Mitzpe Ramon, Israel. Association for Computational Linguistics.  
+Emiel van Miltenburg, Miruna Clinciu, Ondrej Dušek, Dimitra Gkatzia, Stephanie Inglis, Leo Leppänen, Saad Mahamood, Emma Manning, Stephanie Schoch, Craig Thomson, and Luou Wen. 2021. Underreporting of errors in NLG output, and what to do about it. In Proceedings of the 14th International Conference on Natural Language Generation, pages 140-153, Aberdeen, Scotland, UK. Association for Computational Linguistics.  
+Sebastian Varges. 2006. Overgeneration and ranking for spoken dialogue systems. In Proceedings of the Fourth International Natural Language Generation Conference on - INLG '06, page 20, Sydney, Australia. Association for Computational Linguistics.  
+Jason Weston and Kurt Shuster. 2021. Blender Bot 2.0: An open source chatbot that builds long-term memory and searches the internet.  
+Jing Xu, Arthur Szlam, and Jason Weston. 2022. Beyond Goldfish Memory: Long-Term Open-Domain Conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5180-5197, Dublin, Ireland. Association for Computational Linguistics.  
+Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT: Large-Scale Generative Pre-training for Conversational Response Generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270-278, Online. Association for Computational Linguistics.  
+Tianyu Zhao, Divesh Lala, and Tatsuya Kawahara. 2020. Designing Precise and Robust Dialogue Response Evaluators. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 26-33, Online. Association for Computational Linguistics.
+
+![](images/f8226cdf1f1c238984552fc3127dfc8a0d076fe7fb158da638048322096c6e61.jpg)  
+Figure 4: A flowchart of LEAP - a standardized and transparent annotation protocol. It has the following parameters: the minimum and maximum number of rounds ( $N_{min}$  and  $N_{max}$ ), agreement criteria and threshold, common law discussion time limit ( $T$ ), number of researchers involved (minimum of 2), the number of items per small iteration (A), the number of items for the larger annotation (B).
+
+# A Appendix
+
+# A.1 Generic LEAP
+
+As stated in subsection 3.1 has task-specific properties that need to be chosen by researchers. We used LEAP with a particular set of parameters for our experiment. Figure 4 is a generic flowchart for LEAP. We acknowledge that LEAP is extremely rigid. For our experiments, this was important as we desired to minimize variance between groups. In practice, researchers may desire more flexibility (e.g., discussion time cap or the number of examples per iteration). If this is the case then we encourage the documentation of deviations from the protocol.
+
+![](images/929a7461db00359d69fbf499107e04fd177023c447ab16788ddcedd6917a2f1b.jpg)  
+Figure 5: A flowchart of iteration-free LEAP.
+
+# A.2 Iteration-Free LEAP
+
+Figure 5 shows iteration-free LEAP. Iteration-free LEAP might be more attractive for researchers who are concerned about the independent nature of annotations of each round. We suggest that iterations within a round are important since researchers may not have explicitly agreed on the common law rule.
+
+# A.3 NLP Work Using Appropriateness, Information content of output, and Humanlikeness
+
+Various papers created their own definitions for the category of Appropriateness (Varges, 2006; Reiter et al., 2008; Oh and Shrobe, 2008; Murray et al., 2010; Mahamood and Reiter, 2011; Schlünder and Klabunde, 2013; Gatzia et al., 2013; Cimiano et al., 2013; Inglis et al., 2017; Harrison and Walker, 2018; Mori et al., 2019; Santhanam and Shaikh, 2019), Information content of outputs (Demir et al., 2008; Siddharthan et al., 2012; Mahamood and Reiter, 2012; Moraes et al., 2014; Inglis, 2015; Kuptavanich et al., 2018; Qader et al., 2018; Choi et al., 2018), and for Humanlikeness (Byamugisha et al., 2017; Deriu and Cieliebak, 2018; Fikri et al., 2018).
+
+# A.4 Dialog Model Parameters
+
+For DialogGPT, which was trained on 147M dialogue instances created from Reddit threads (Zhang et al., 2020), we used the pre-trained model with the medium (345M) model checkpoint, using the top-K sorting algorithm. For GPT-3 (Brown et al., 2020), we used a temperature of 0.9 and a top-p decoding strategy (Holtzman et al., 2019) with  $p = 0.92$ . We used the following format for the prompt for GPT-3:
+
+The following is a conversation between A and B.
+
+A: Oh, I am so tired.
+
+B: I know what you mean.
+
+A: I don't know if I can continue working like this.
+
+B:
+
+For Plato2, we used two model sizes, 24L (with 310M parameters), and 32L (with 1.6B parameters) (Bao et al., 2021). For BlenderBot, two model sizes were used: 2.7B and 9B (Miller et al., 2017). For BlenderBot 2, two model sizes were used as well: 400M and 3B (Weston and Shuster, 2021; Komeili et al., 2022; Xu et al., 2022). Finally, we used the original human responses that are a part of the ESL (Sedoc et al., 2019) and Daily Dialog (Li et al., 2017) evaluation sets.
+
+<table><tr><td>prompt</td><td>response</td></tr><tr><td>A: Oh, I am so tired.</td><td></td></tr><tr><td>B: I know what you mean.</td><td>Why don&#x27;t</td></tr><tr><td>A: I don&#x27;t know if I can</td><td>you take a</td></tr><tr><td>continue working like this.</td><td>break?</td></tr></table>
+
+Figure 6: An example prompt and response annotated by the researchers and crowdsourced workers.
+
+# A.5 Instructions
+
+# Common Law Annotations
+
+# Creating Annotation Guidelines
+
+The goal is to create guidelines that help people annotate conversational text data as similarly as possible. In order to increase agreement with your annotation partner, you will meet with them to discuss a common annotation methodology.
+
+The annotation task is to label chatbot responses to prompts, using three annotation criteria:
+
+- Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+Information content of outputs: The amount of information conveyed by an output.  
+Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Annotating Model Responses
+
+For each round of annotations, you will be provided with a Google Sheets document containing 50 prompts and responses. During these annotations, you may not communicate with your partner annotator.
+
+The prompt column contains a single utterance or multiple-utterance conversation. The response column contains the chatbot's response to the last utterance in the prompt.
+
+Utterances are separated by A: and B:, which indicate two speakers. There are at most two speakers per prompt, though there may be prompts with only one speaker.
+
+For example,
+
+<table><tr><td>prompt</td><td>response</td><td>Appropriateness</td><td>Information content of outputs</td><td>Humanlikeness</td></tr><tr><td>A: Oh, I am so tired. 
+B: I know what you mean. 
+A: I don&#x27;t know if I can 
+continue working like this.</td><td>Why don&#x27;t you take a break?</td><td>enter annotation here</td><td>enter annotation here</td><td>enter annotation here</td></tr></table>
+
+Remember that the annotation values should be a number between 1 and 5. You will annotate 50 prompt-response pairs each round. Please time yourself at the start and end of each annotation session.
+
+Figure 7: The annotation and discussion instructions shared to all groups.
+
+# Annotation & Discussion Plan [Annotator]
+
+Objective: Annotators repeat annotation and discussion in order to increase their inter-annotator agreement.
+
+<table><tr><td>Step</td><td>Title</td><td>Time Needed (Approx.)</td><td>Instructions</td><td>Notes</td></tr><tr><td>1</td><td>Discuss initial annotation methodology</td><td>30 min.</td><td>Schedule a common time using the Doodle poll
+Join this public Zoom call on your scheduled time and discuss annotation methodologies</td><td></td></tr><tr><td>2</td><td>1st Annotation Session</td><td>-</td><td>Annotate 50 model responses</td><td></td></tr><tr><td>3</td><td>Discuss annotation methodology</td><td>30 min.</td><td>Schedule a common time using the Doodle poll
+Join this public Zoom call on your scheduled time and discuss annotation methodologies</td><td></td></tr><tr><td>4</td><td>2nd Annotation Session</td><td>-</td><td>Annotate 50 model responses</td><td></td></tr><tr><td>5</td><td>Discuss annotation methodology</td><td>30 min.</td><td>Schedule a common time using the Doodle poll
+Join your designated Zoom call on your scheduled time and discuss annotation methodologies</td><td></td></tr><tr><td colspan="5">If Inter-Annotator Agreement is below 0.7: proceed to STEP 6 and STEP 7
+If Inter-Annotator Agreement is above 0.7: proceed to STEP 8</td></tr><tr><td>6</td><td>Annotation Session</td><td>-</td><td>Annotate 50 model responses</td><td></td></tr><tr><td>7</td><td>Discuss annotation methodology</td><td>30 min.</td><td>Schedule a common time using the Doodle poll
+Join your designated Zoom call on your scheduled time and discuss annotation methodologies</td><td></td></tr><tr><td colspan="5">If Inter-Annotator Agreement is below 0.7: repeat STEP 6 and STEP 7
+If Inter-Annotator Agreement is above 0.7: proceed to STEP 8
+→ Maximum of 5 annotation- discussion repetitions</td></tr><tr><td>8</td><td>Annotate 400 responses</td><td>-</td><td>Annotate 400 model responses</td><td></td></tr><tr><td>9</td><td>Create Individual Annotation Guideline</td><td>-</td><td>Each annotator creates their own annotation guideline</td><td></td></tr><tr><td>10</td><td>Merge Annotation Guideline</td><td>-</td><td>The annotator pair merges their annotation guideline</td><td></td></tr><tr><td>11</td><td>150 AMT Annotations</td><td>-</td><td>Annotate 150 items through the Amazon Mechanical Turk (AMT) platform - 50 using your own guideline, 50 using a different group&#x27;s guideline, and 50 using another group&#x27;s guideline</td><td></td></tr></table>
+
+Figure 8: The step-by-step LEAP instructions shared among researcher annotators.
+
+# A.6 Annotation Guidelines
+
+# Instructions
+
+# NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs
+
+The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt. The three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+# Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best. Specific Definitions
+
+# Tips:
+
+- Do not consider the humanlkeness of the response when judging its appropriateness. If the response fits the context, but sounds weird, mark it as highly appropriate.  
+- Do not consider the appropriateness of the response when judging its humanlikeness. Humanlikeness should depend only on the response itself, ignoring the context.
+
+<table><tr><td>#</td><td>Appropriateness</td><td>Information content of output</td><td>Humanlikeness</td></tr><tr><td></td><td>Does the response fit the context well?</td><td>Is the response specific or generic?</td><td>Could a human have said this response?</td></tr><tr><td>1</td><td>The response is completely irrelevant. It does not match or reference the context at all.</td><td>The response could follow almost any statement or context, for example “How so?” or “Yeah”.</td><td>The response is garbled text or nonsensical. It has no reasonable interpretation. For example, “He green a”.</td></tr><tr><td>2</td><td>The response matches or references the context, but is highly unexpected. For example, it might contradict a previous statement, repeat word-for-word something already said, or respond to an earlier part of the conversation.</td><td>The response could follow a large category of statements, but not any statement, for example, “Why not?” or “I&#x27;m down for that.”</td><td>The response has some well-formed parts that you can interpret, but it&#x27;s completely unbelievable that the response came from a human. For example, “It&#x27;s just north of here UUUU.”</td></tr><tr><td>3</td><td>The response has at least one interpretation that makes sense in the conversation; however, the interpretation is a stretch or strange given the context.</td><td>The response talks about something generic in a vague way, for example: “Yeah, I love that.” or “It really is great.”</td><td>The response is well-formed, but doesn&#x27;t seem realistic, sounds like a poorly written fictional person, or over-explains. For example, “I am sleepy because it&#x27;s night time, so now I will nap in my bed.”</td></tr><tr><td>4</td><td>The first interpretation of the response makes sense in the conversation; however, it&#x27;s a bit strange, awkward, or unexpected.</td><td>The response talks about something specific to the conversation in a vague way, for example: “Park street is great.”</td><td>The response is mostly believable as coming from a person, but seems a bit strange. For example, the response is unreasonably polite.</td></tr><tr><td>5</td><td>The response perfectly fits the conversation.</td><td>The response is highly specific to the conversation. It discusses something from the conversation in a specific way, for example “I really miss them, now that they&#x27;re gone.”</td><td>The response is perfectly believable as coming from a human. You would not be surprised at all if a person had actually said it.</td></tr></table>
+
+# Examples:
+
+<table><tr><td>Prompt</td><td>Response</td><td>Appropriateness</td><td>Information content of output</td><td>Humanlikeness</td></tr><tr><td>A: This floor is falling apart. 
+B: How can you tell? 
+A: Are you kidding me? Step over here.</td><td>That&#x27;s what I was thinking.</td><td>1</td><td>2</td><td>5</td></tr><tr><td>A: The prices on fruit have really gone up this year! 
+B: Yes, they have. It&#x27;s ridiculous! 
+A: The fruit is not of good quality, either.</td><td>I&#x27;ve also noticed the drop in quality.</td><td>5</td><td>5</td><td>5</td></tr><tr><td>A: I&#x27;m making the food for the party tomorrow. 
+B: I heard you are a fantastic cook! 
+A: Thank you! I do my best.</td><td>I&#x27;m sure you do!</td><td>4</td><td>2</td><td>4</td></tr></table>
+
+Figure 9: Annotation Guideline for Group 1.
+
+# Instructions
+
+# NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs
+
+The a   t the three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Specific Definitions
+
+# Appropriateness
+
+Given the context in the prompt, we will consider the following aspects when assigning the score for appropriateness:
+
+1. Answer the question  
+2. Talk about the same thing in the prompt
+
+If the response satisfies all the above requirements, we will assign a score of 5. If the response somehow answers the question but does not satisfy one of the other two requirements, we will assign a score of 4. If the response only answers the question partially, we will assign a score of 3. If the response does not answer the question but satisfies one of the other two requirements, we will assign a score of 2. If the response does not satisfy all the above requirements, we will assign a score of 1.
+
+# Information content of outputs
+
+For this part, we will take the information conveyed in the response and the length of the response into consideration. To be detailed, we will consider the following aspects when assigning the score for information content of outputs:
+
+1. The information covered for the question  
+2. The number of sentence  $(>= 4$  long;  $>= 3$  median;  $>= 1$  short)  
+3. Information is valid (even if it is not related to the prompt)
+
+If the response satisfies all the above requirements (long sentences), we will assign a score of 5. If the response contains enough and valid information but does not have a reasonable length (median), we will assign a score of 4. If the response contains some information, but the information may not be valid or the response does not have a reasonable length (median), we will assign a score of 3. If the response contains limited information and the length of the response is short, we will assign a score of 2. If the response does not satisfy all the above requirements, we will assign a score of 1.
+
+# Humanlikeness
+
+We will evaluate the degree to which the response looks like a human sentence. We will consider the following aspects when assigning the score for human-likeness:
+
+1. First impression of reading as a human sentence  
+2. Check grammar and syntax error
+
+If the response satisfies all the above requirements, we will assign a score of 5. If the response contains minor grammar or syntax errors but overall looks like a human sentence, we will assign a score of 4. If the response contains a few grammar or syntax errors but still looks like a human sentence, we will assign a score of 3. If the response somehow does not look like a human sentence slightly but there are few grammar or syntax errors in the response, we will assign a score of 2. If the response does not satisfy all the above requirements, we will assign a score of 1.
+
+Figure 10: Annotation Guideline for Group 2.
+
+# Instructions
+
+# NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs
+
+The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt. The three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Specific Definitions
+
+Laurel: Ben let out the cats this morning but one of them didn't come back into the house. Dara: oh no, was it Tom?
+
+# Appropriateness:
+
+1. Completely irrelevant and non-topical response
+Laurel: Did you get dinner?  
+2. Response is on topic, but not appropriate.  
+Laurel: Cats are cute  
+3. Half of response is appropriate, half is not
+Laurel: Thankfully, it wasn't Tom, but I want some hotdogs  
+4. Response is mostly appropriate, albeit slightly awkward.  
+Laurel: Jerry  
+5. Response is
+Laurel: No, it was Jerry ... He's always crawling under the house.
+
+# Information content of output
+
+1. Repetition of something said in the prompt Laurel: The cats were let out by Ben  
+2. A generic answer or a question
+Laurel: Maybe
+Laurel: Do you like cats?  
+3. One type of information conveyed (information about self or the world)  
+Laurel: No, it was Jerry.  
+Laurel: Cats often run away.  
+4. Both types of information conveyed (information about self and the world)  
+Laurel: No Jerry; (I am really worried.  
+5. 3 or more distinct pieces of information conveyed.  
+Laurel: No it is Jerry. I'm really worried.  $90\%$  of cats that aren't found within 5 hours are roadkill.
+
+Humanlikeness The extent to which the response BY ITSELF (ignore context) could have been written by a person
+
+1. Self-contradictory or ungrammatical Laurel: Yes it's Jerry. But Jerry is Tom.  
+2. Incorrect obvious facts about the world
+Laurel: Jerry is mostly a cat name as in Tom and Jerry  
+3. Too long or unnatural sounding (lack of conversationalist properties)  
+Laurel: Jerry got lost then he might have gone under the house just like I drew in my drawing. Maybe this was all a dream  
+4. Close but not right almost like a non-native English speaker.  
+Laurel: Jerry ran away ... Sad feelings.  
+5. Perfectly fluent. You could imagine yourself saying this.  
+Laurel: Unfortunately both Jerry and Tom that ran away. I'm not sure what to do.
+
+Figure 11: Annotation Guideline for Group 3.
+
+# Instructions
+
+# NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs
+
+The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt. The three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Specific Definitions
+
+Information content of output, Appropriateness:
+
+Appropriateness: 1
+
+Information content of output: 1
+
+If the response doesn't make sense, doesn't relate to the previous conversation, doesn't have some new information
+
+Appropriateness: 1
+
+Information content of output: 2
+
+If the response doesn't make sense, doesn't relate to the previous conversation, has some new information
+
+Appropriateness: 2
+
+Information content of output: 1
+
+If the response doesn't make sense, but still relates to the previous conversation, doesn't have new information
+
+Appropriateness: 4
+
+Information content of output: 3
+
+If the response does make sense, but still relate to the previous conversation, doesn't have new information
+
+Appropriateness: 5
+
+Information content of output: 5
+
+If the response does make sense, but still relate to the previous conversation, has some new information
+
+Appropriateness: 4
+
+Information content of output: 3
+
+If the response does make sense, not quite appropriate, has new information
+
+Appropriateness: 4
+
+Information content of output: 4
+
+If the response does make sense, not as appropriate, has new information
+
+Humanlikeness
+
+Humanlikeness: 1
+
+If the response doesn't make sense, doesn't relate to the previous conversation, repeat previous information,
+
+Humanlikeness: 1
+
+If the response doesn't make sense, doesn't relate to the previous conversation, doesn't repeat previous information,
+
+Humanlikeness: 2
+
+If the response doesn't make sense, but still relates to the previous conversation, repeats previous information,
+
+Humanlikeness: 2
+
+If the response doesn't make sense, but still relates to the previous conversation, doesn't repeat previous information,
+
+Humanlikeness: 1
+
+If the response does make sense, doesn't relate to the previous conversation, repeats previous information,
+
+Humanlikeness: 1
+
+If the response does make sense, doesn't relate to the previous conversation, doesn't repeat previous information,
+
+Humanlikeness: 3
+
+If the response does make sense, but still relates to the previous conversation, repeats previous information,
+
+Humanlikeness: 4
+
+If the response does make sense, relates to the previous conversation, and paraphrases previous information
+
+Humanlikeness: 5
+
+If the response does make sense, still relates to the previous conversation, doesn't repeat previous information,
+
+Figure 12: Annotation Guideline for Group 4.
+
+# Instructions
+
+NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt. The three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Specific Definitions
+
+Appropriateness:
+
+Lower score:
+
+Confusing response, off-topic  
+- Offensive, aggressive  
+- Condtradiction
+
+Higher score:
+
+- Empathetic, compassionate responses  
+- Apt responses, matching emotional toll of the situation
+
+Information Content:
+
+Lower score:
+
+- Contradiction  
+- Off-topic  
+- Repetition  
+- Standalone response doesn't make sense
+
+Higher score:
+
+- Reasoning, could be indicated by joint statements/multiple clauses
+
+Humanlikeness
+
+Lower score:
+
+- Extensive repetition  
+- Contradiction  
+- Off-topic  
+- Generic responses ("I am sorry to hear that", "How can i help you?".
+
+Higher score:
+
+- Appropriate emojis  
+- First person pronouns ("I", "We")  
+- Referring to familial relationships  
+Colloquial language ("wanna", etc.)  
+- Contractions ("I'm", "aren't", etc.)  
+- Discuss of emotions ("I feel" statements for example)  
+- Expression of surprise ("oh!", etc.)
+
+Figure 13: Annotation Guideline for Group 5.
+
+# Instructions
+
+# NOTE: THERE ARE A MAXIMUM OF 5 HITs YOU CAN COMPLETE. COMPLETING ALL 5 HITs WILL GIVE YOU A BONUS! WE ENCOURAGE YOU TO DO ALL 5 HITs
+
+The annotation task is to label responses to a given prompt. The prompt consists of two people (A and B) talking to each other. The response is the next utterance after the final utterance in the prompt. The three base annotation criteria are:
+
+1. Appropriateness: The degree to which the output is appropriate in the given context/situation.  
+2. Information content of outputs: The amount of information conveyed by an output.  
+3. Human-likeness: The degree to which an output could have been produced by a human.
+
+Each criteria is annotated on a 5-point scale where 1 is worst and 5 is best.
+
+# Specific Definitions
+
+# Appropriateness:
+
+1. "this had nothing to do with the conversation whatsoever"  
+2. "huh, wait, that's very weird"
+
+$\mathrm{O}$  Response has at least a mild relevance to the topic discussed in prompt but otherwise is a very unusual sentence
+
+3. "that's absurd but lets move on"  
+An unusual sentence given topic-prompt or sentence construction but people would just ignore it if someone said that - because its not THAT weird
+
+4. "Hm. I guess that makes sense"  
+- Follows logically from the prompt; typical response given the context in the prompt. Sort of what was expected.
+
+5. "That helped the conversation"
+
+This adds new and relevant information to the conversation. This is going above and beyond in the right direction for this conversation.
+
+# Information Content:
+
+1. Hard to infer anything about the conversation. Very generic.  
+Example:"okay".
+
+2. I know one thing that this conversations is about
+
+Example:"Yeah I've been there"  
+- You know they are talking about a place which counts as 1 thing
+
+3. I know 2 things. Longer sentence.
+
+Example: "Yeah I've been there. I thought it was quite nice actually"
+
+4. Multiple sentences with with 3 or more things.  
+Example:"Yeah I've been there. I thought it was quite nice. My mom liked it too."
+
+5. A long informative sentence ( $\sim > 10$  words) with LOTs to add in terms of specific.
+
+Example: "The thing I like about the Taj Mahal is that it is all one block or marble [...]
+
+# Human-likeness:
+
+1. Impossible for a human to say in any context
+
+Gibberish: "trhaiotjhoath  $\wedge \wedge$  blah_cat"
+
+2. Correct words but very wrong
+
+○ grammar or word order "Go blue stuff very, I said?"
+
+3. some beginner ESL person could say this, i guess
+
+- ESL: "Very nice that thing is"
+
+4. Someone would only say this in a weirdly specific situation
+
+"it would have been nice if they had dunked it"  
+Technical language: "compile down your source code into binary bits via the JVM"
+
+5. literally me or my circle would legitimately say this "Yeah i get you"
+
+Figure 14: Annotation Guideline for Group 6.
+
+Figure 15: Average agreement between Researchers and Amazon Mechanical Turk Workers, using each Group's guidelines.  
+Figure 16: An example attention check question asked to crowdsourced workers.  
+![](images/6b43de6b71ab6589e5cc171a9fc726ab29d1b43cd91010d7723d5055836ed3d7.jpg)  
+To verify that you've read the instructions, please read the following prompt and write an example Response that satisfies the following: Appropriateness: 5, Information content of Output: 2, Humanlikeness: 5
+
+A: Oh, I am so tired.  
+B: I know what you mean.  
+A: I don't know if I can continue working like this.
+
+# A.7 Crowdsourced Annotations
+
+Metadata Using guidelines created by Groups 1 and 2, which were created using LEAP, we deployed an initial screening round of annotations to distinguish the workers who were able to have high agreement with the researchers of the respective groups. Each screening round consisted of 1 HIT task and 10 unique workers completed the HITs. Workers who were able to achieve a category average  $\kappa > 0.7$  agreement with the researchers were noted as quality workers. The qualified workers were then given a larger MTurk task of 400 prompt-response questions, where each HIT asked 55 prompt-response questions.
+
+Three workers qualified for Group 1 and four workers qualified for Group 2. A total of 24 HITs were created for the three workers using Group 1's guidelines and a total of 32 HITs were published for the four workers using Group 2's guidelines. The workers for Group 1 completed a total of 23 HITs and the workers for Group 2 completed a total of 19 HITs.
+
+The workers were notified the annotations will be used for research purposes.
+
+Compensation We conducted an initial pilot run of a HIT and learned the workers took an average of 25 minutes to complete a HIT of 55 items. We paid each worker $6.25 per HIT.
+
+Qualifications We required a minimum of 500 approved tasks on MTurk. Second, the workers were chosen from a group of workers whose quality was verified for other text-generation evaluation tasks (e.g., summarization evaluation).
+
+Quality Checks In order to ensure the quality of the crowdsourced data, we implemented several different quality and attention checks. For each HIT, we asked two quality-check questions to confirm that the
+
+worker read and understood the annotation guidelines (Figure 16). We asked an attention-check question to ensure the worker was not randomly participating in the HIT without reading the prompt and responses. Finally, we excluded all workers who did not pass the attention checks or had a category average  $\kappa < 0.1$ .
+
+# A.8 Average Annotation Ratings per Conversational Model
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>3.58</td><td>3.12</td><td>4.78</td></tr><tr><td>BlenderBot 2 - 400m</td><td>3.10</td><td>4.10</td><td>4.32</td></tr><tr><td>BlenderBot - 3b</td><td>2.38</td><td>4.75</td><td>3.30</td></tr><tr><td>BlenderBot - 9b</td><td>3.62</td><td>4.05</td><td>4.25</td></tr><tr><td>DiaLoGPT</td><td>3.19</td><td>4.13</td><td>3.86</td></tr><tr><td>GPT-3</td><td>4.42</td><td>3.90</td><td>4.63</td></tr><tr><td>Ground truth</td><td>4.08</td><td>4.48</td><td>4.50</td></tr><tr><td>Plato 2</td><td>3.16</td><td>4.07</td><td>3.57</td></tr><tr><td>Plato 2 - 24L</td><td>4.30</td><td>4.70</td><td>3.80</td></tr><tr><td>Plato 2 - 32L</td><td>4.40</td><td>5.00</td><td>4.60</td></tr></table>
+
+Table 4: Average annotation ratings per conversational model for Group 1.  
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>3.63</td><td>2.53</td><td>4.30</td></tr><tr><td>BlenderBot 2 - 400m</td><td>3.24</td><td>3.52</td><td>4.58</td></tr><tr><td>BlenderBot - 3b</td><td>2.84</td><td>3.87</td><td>4.21</td></tr><tr><td>BlenderBot - 9b</td><td>3.54</td><td>3.46</td><td>4.25</td></tr><tr><td>DiaLoGPT</td><td>3.54</td><td>3.31</td><td>4.41</td></tr><tr><td>GPT-3</td><td>3.95</td><td>3.33</td><td>4.53</td></tr><tr><td>Ground truth</td><td>3.89</td><td>3.23</td><td>4.87</td></tr><tr><td>Plato 2</td><td>3.47</td><td>3.61</td><td>4.27</td></tr><tr><td>Plato 2 - 24L</td><td>3.85</td><td>4.30</td><td>4.50</td></tr><tr><td>Plato 2 - 32L</td><td>4.40</td><td>4.50</td><td>4.80</td></tr></table>
+
+Table 5: Average annotation ratings per conversational model for Group 2.  
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>2.74</td><td>2.88</td><td>4.61</td></tr><tr><td>BlenderBot 2 - 400m</td><td>2.45</td><td>2.94</td><td>4.64</td></tr><tr><td>BlenderBot - 3b</td><td>3.22</td><td>3.70</td><td>4.71</td></tr><tr><td>BlenderBot - 9b</td><td>3.21</td><td>3.49</td><td>4.97</td></tr><tr><td>DiaLoGPT</td><td>3.35</td><td>2.77</td><td>4.60</td></tr><tr><td>GPT-3</td><td>4.75</td><td>2.77</td><td>4.86</td></tr><tr><td>Ground truth</td><td>4.31</td><td>3.01</td><td>4.95</td></tr><tr><td>Plato 2</td><td>3.64</td><td>3.30</td><td>4.28</td></tr><tr><td>Plato 2 - 24L</td><td>3.02</td><td>3.42</td><td>3.68</td></tr><tr><td>Plato 2 - 32L</td><td>3.61</td><td>3.40</td><td>4.25</td></tr></table>
+
+Table 6: Average annotation ratings per conversational model for Group 3.
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>2.72</td><td>2.19</td><td>2.42</td></tr><tr><td>BlenderBot 2 - 400m</td><td>2.24</td><td>1.95</td><td>2.17</td></tr><tr><td>BlenderBot - 3b</td><td>3.64</td><td>3.85</td><td>3.60</td></tr><tr><td>BlenderBot - 9b</td><td>3.16</td><td>3.18</td><td>3.18</td></tr><tr><td>DiaLoGPT</td><td>3.24</td><td>2.94</td><td>3.13</td></tr><tr><td>GPT-3</td><td>4.54</td><td>4.31</td><td>4.53</td></tr><tr><td>Ground truth</td><td>4.16</td><td>3.94</td><td>4.14</td></tr><tr><td>Plato 2</td><td>3.86</td><td>3.78</td><td>3.83</td></tr><tr><td>Plato 2 - 24L</td><td>2.65</td><td>2.63</td><td>2.65</td></tr><tr><td>Plato 2 - 32L</td><td>3.92</td><td>3.99</td><td>3.65</td></tr></table>
+
+Table 7: Average annotation ratings per conversational model for Group 4.  
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>2.94</td><td>3.02</td><td>3.16</td></tr><tr><td>BlenderBot 2 - 400m</td><td>2.95</td><td>3.52</td><td>2.92</td></tr><tr><td>BlenderBot - 3b</td><td>3.99</td><td>4.26</td><td>3.94</td></tr><tr><td>BlenderBot - 9b</td><td>3.57</td><td>4.10</td><td>3.79</td></tr><tr><td>DiaLoGPT</td><td>3.58</td><td>3.52</td><td>3.67</td></tr><tr><td>GPT-3</td><td>4.49</td><td>4.20</td><td>4.60</td></tr><tr><td>Ground truth</td><td>4.48</td><td>4.35</td><td>4.57</td></tr><tr><td>Plato 2</td><td>4.08</td><td>4.18</td><td>4.19</td></tr><tr><td>Plato 2 - 24L</td><td>3.17</td><td>3.90</td><td>3.57</td></tr><tr><td>Plato 2 - 32L</td><td>4.12</td><td>4.50</td><td>3.57</td></tr></table>
+
+Table 8: Average annotation ratings per conversational model for Group 5.  
+
+<table><tr><td>Model</td><td>App.</td><td>Info.</td><td>Human.</td></tr><tr><td>BlenderBot 2 - 3b</td><td>2.85</td><td>2.59</td><td>4.79</td></tr><tr><td>BlenderBot 2 - 400m</td><td>2.82</td><td>2.96</td><td>4.68</td></tr><tr><td>BlenderBot - 3b</td><td>3.70</td><td>3.38</td><td>4.74</td></tr><tr><td>BlenderBot - 9b</td><td>3.42</td><td>3.48</td><td>4.81</td></tr><tr><td>DiaLoGPT</td><td>3.40</td><td>2.37</td><td>4.66</td></tr><tr><td>GPT-3</td><td>4.56</td><td>2.56</td><td>4.95</td></tr><tr><td>Ground truth</td><td>4.31</td><td>2.77</td><td>4.92</td></tr><tr><td>Plato 2</td><td>3.77</td><td>3.46</td><td>4.27</td></tr><tr><td>Plato 2 - 24L</td><td>3.12</td><td>4.18</td><td>3.92</td></tr><tr><td>Plato 2 - 32L</td><td>3.63</td><td>3.90</td><td>4.30</td></tr></table>
+
+Table 9: Average annotation ratings per conversational model for Group 6.
+
+![](images/e14fcfa2b3c1a6365e717e01c9ba0e70dd6f092def44f2ab5d6e9996931c6fb0.jpg)  
+Figure 17: Contingency table of annotations for Groups 3, 4, 5, and 6 - From top to bottom: Appropriateness, Information content of outputs, and Humanlikeness. Round 400 indicates the final round of annotations in LEAP with 400 items. Red borders indicate within-group agreement.
+
+# A.9 IAA analysis - Iteration-free LEAP
+
+Within Group The red borders in Figure 17 show the change in within-group agreement for Groups 3, 4, 5, and 6. We observed that agreement scores for Appropriateness were relatively higher than other categories for most rounds across all groups. This coincides with our earlier findings that certain categories, such as Appropriateness, may have stronger shared constructs than others.
+
+Between Groups While each group's annotation guideline helped the researchers achieve high agreement within-group, Figure 17 shows that agreement between annotators of different groups remained low throughout the five rounds. Surprisingly, we can observe that agreement between annotators across different groups remained high throughout all five rounds for Appropriateness, suggesting that certain annotation categories have a strong shared understanding across annotators of the different groups.
+
+Another interesting observation can be seen in Figure 18, which shows the level of agreement for Information content of output during Round 4. The green border shows a distinct silo of agreement between annotators of Groups 4 and 5. We can see that Researcher 10 (Group 5) has low agreement scores of 0.09 and 0.01 with Researchers 5 and 6 (Group 3) and 0.07 and 0.02 with Researchers 11 and 12 (Group 6).
+
+However, Researcher 10 has a relatively high agreement of 0.5 and 0.43 with Researchers 5 and 6 (Group 5). With Researcher 7, who also belongs to Group 4, Researcher 10 has an agreement score of 0.39. While the distinction is not as clear, annotators of Group 3 (Researchers 5 and 6) show higher agreement with annotators of Group 6 (Researchers 11 and 12) compared to annotators of Group 4 and Group 5.
+
+Similar distinct silos of agreement can be observed in Figure 17 for Humanlikeness, one between Groups 4 and 5 and another between Groups 3 and 6.
+
+![](images/699322f68b5cf2008fdd8b1fbe1dbfb57da9851645fab6da8a8aa5acc51fe41e.jpg)  
+Figure 18: Example of distinct silo of agreement between Group 4 and Group 5 for Information content of output, Round 4. The Green border show agreement between annotators of Group 4 and Group 5.
+
+# A.10 Cohen's Kappa
+
+Counting the raw number of matching annotations is one of the simplest ways to measure agreement. However, the raw agreement fails to account for the possibility of random chance agreement, which becomes problematic when the random chance is very high (Artstein, 2017). To overcome this limitation, Cohen's Kappa  $(\kappa)$  measures observed agreement above the expected agreement (Cohen, 1968), more formally stated,
+
+$$
+\kappa = \frac {p _ {o} - p _ {e}}{1 - p _ {e}}
+$$
+
+where  $p_{o}$  is the relative observed agreement among annotators, and  $p_{e}$  is the expected probability of random chance agreement. Cohen's Kappa measures agreement between two annotators, treating any disagreement linearly. If a pair of annotators matches on all annotations (thus  $p_{o} = 1$ ), then  $\kappa = 1$ . On the other hand, if the pair has no agreement other than what is expected by chance (thus  $p_{o} = p_{e}$ ), then  $\kappa = 0$ .  $\kappa < 0$  is also possible when the pair annotates worse than expected chance agreement ( $p_{0} < p_{e}$ ).
+
+Some annotation studies require different weights to be applied to different levels of agreement between annotators. For example, on a 5-point Likert scale (Likert, 1932), annotation scores 4 and 5 should be regarded as being in higher agreement than annotation scores 1 and 5. To account for this, the weighted Cohen's Kappa (Cohen, 1968) is often used to measure IAA in annotation tasks, in order to weigh disagreement differently, thus,
+
+$$
+\kappa = 1 - \frac {\sum_ {i = 1} ^ {k} \sum_ {j = 1} ^ {k} w _ {i j} x _ {i j}}{\sum_ {i = 1} ^ {k} \sum_ {j = 1} ^ {k} w _ {i j} m _ {i j}},
+$$
+
+where  $w_{ij}$  is the weight matrix,  $x_{ij}$  is the observed matrix, and  $m_{ij}$  is the expected matrix.
+
+Cohen's Kappa of 0.6 to 0.8 is commonly regarded as a threshold for sufficient inter-annotator agreement in NLP research (Landis and Koch, 1977). In order to strengthen the reliability of annotation guidelines, various methods have been used to raise the kappa above the threshold, such as taking out outlier anomalous annotations from the dataset (Zhao et al., 2020). However, this is no guarantee that the validity of the dataset is improved by the discarded outliers.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? Section 6  
+A2. Did you discuss any potential risks of your work? Section 6 - particularly some models generated responses containing biases.  
+A3. Do the abstract and introduction summarize the paper's main claims? Abstract and Section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Yes, Section 3
+
+B1. Did you cite the creators of artifacts you used? Yes, Section 3.2  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? Not relevant.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? Section 1  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? Section 3.2  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? Section 3.2  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. Section 3.2
+
+C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? No response.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? No response.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? No response.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? No response.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Section 3.3
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Section 3.3, Appendix A.5  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Appendix A.5  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Appendix A.5  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? All of the expert annotators are co-authors of the paper. The annotations from AMT workers would be considered IRB exempt.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Section 3.3

2023/Common Law Annotations_ Investigating the Stability of Dialog System Output Annotations/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e72a117e70f9af1b1916936e104164c612ff0dd42d7a9b22300d206450d4295
+size 1315375

2023/Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering/8ad456f0-7f0d-40ee-8bd0-7b70b4a0c21c_content_list.json

@@ -0,0 +1,2450 @@
+[
+    {
+        "type": "text",
+        "text": "Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering",
+        "text_level": 1,
+        "bbox": [
+            243,
+            90,
+            754,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Siwei Wu, Xiangqing Shen, and Rui Xia*",
+        "bbox": [
+            326,
+            149,
+            680,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "School of Computer Science and Engineering, Nanjing University of Science and Technology, China {wusiwei, xiangqing.shen, rxia}@njust.edu.cn",
+        "bbox": [
+            278,
+            167,
+            722,
+            216
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The nodes in the commonsense knowledge graph (CSKG) are normally represented by free-form short text (e.g., word or phrase). Different nodes may represent the same concept. This leads to the problems of edge sparsity and node redundancy, which challenges CSKG representation and completion. On the one hand, edge sparsity limits the performance of graph representation learning; On the other hand, node redundancy makes different nodes corresponding to the same concept have inconsistent relations with other nodes. To address the two problems, we propose a new CSKG completion framework based on Contrastive Pretraining and Node Clustering (CPNC). Contrastive Pretraining constructs positive and negative head-tail node pairs on CSKG and utilizes contrastive learning to obtain better semantic node representation. Node Clustering aggregates nodes with the same concept into a latent concept, assisting the task of CSKG completion. We evaluate our CPNC approach on two CSKG completion benchmarks (CN-100K and ATOMIC), where CPNC outperforms the state-of-the-art methods. Extensive experiments demonstrate that both Contrastive Pretraining and Node Clustering can significantly improve the performance of CSKG completion. The source code of CPNC is publicly available on https://github.com/NUSTM/CPNC.",
+        "bbox": [
+            141,
+            277,
+            460,
+            703
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            713,
+            258,
+            727
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Commonsense knowledge graphs (CSKG) have been widely used to build commonsense-grounded AI applications, such as question answering (Lv et al., 2020), visual question answering (Zhu et al., 2020), sentiment analysis (Li et al., 2022), dialogue system (Tu et al., 2022), etc. Commonsense knowledge graphs such as ConceptNet (Speer et al., 2017) and ATOMIC (Sap et al., 2019) provide a structured way of representing a commonsense concept, which consists of a head node, a tail node, and",
+        "bbox": [
+            112,
+            738,
+            489,
+            898
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/477f90750452fd158a58c16e870cff0691e813899b344b3a6b73d148925898f3.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            250,
+            875,
+            326
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1fa42bc1f39770a20c77bd3edd2c2862265687be3fdb675ccadbdc2a71a03c16.jpg",
+        "image_caption": [
+            "Figure 1: Illustration of a subgraph and the latent concept in ConceptNet. The nodes of the same color belong to the same latent concept."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            326,
+            875,
+            476
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "the relation edge. Nodes in commonsense knowledge graphs are typically represented by free-form short text (word or phrase), resulting in many different nodes representing the same concept. Figure 1 shows a subgraph of ConceptNet, where nodes in the same color indicate the same concept. E.g., \"have breakfast\", \"take breakfast\" and \"eat breakfast\" all express the concept of \"eat breakfast\". This problem also results in a large number of missing edges between nodes, as illustrated by the dashed lines in Figure 1. E.g., \"eat breakfast\" and \"prepare food\" have a \"HasPrerequisite\" relation, but such relation is missing between \"eat breakfast\" and \"cook food\".",
+        "bbox": [
+            505,
+            562,
+            884,
+            785
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "On the one hand, node redundancy in CSKG makes different nodes of the same concept have inconsistent relations with other nodes. In CSKG representation and completion, it would be beneficial to make use of the latent concept information behind different nodes to help learn more semantic-general representations. However, this intuition was ignored by most of the existing work in",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Corresponding author",
+        "bbox": [
+            141,
+            904,
+            285,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "13977",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 13977-13989 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "CSKG completion. On the other hand, as analyzed by Malaviya et al. (2020), edge sparsity in CSKG limits information propagation in graph neural networks. Upon graph neural networks, researchers further incorporated pre-trained language model such as BERT to enhance the semantic representation of nodes (Malaviya et al., 2020; Ju et al., 2022; Wang et al., 2021). However, fine-tuning BERT is still imperfect in representing the commonsense knowledge graph, which is made up of linked nodes represented in a free-form short text.",
+        "bbox": [
+            112,
+            84,
+            490,
+            261
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To tackle the two issues, we propose a new CSKG completion framework based on the Encoder-Decoder architecture, which contains two core modules Contrastive Pretraining and Node Clustering (CPNC). Contrastive Pretraining is to alleviate the difficulty of node representation learning induced by edge sparsity. Through contrastive learning on positive and negative head-tail node pairs, the embedding distance between related nodes becomes closer, and that between unrelated nodes becomes farther so as to learn better node representations. Node Clustering aims to address the edge inconsistency issue caused by node redundancy. We cluster nodes with close semantic representations and take the mean vector as the latent concept representation for nodes in this cluster. To assist CSKG completion, the latent concept representation is fused with the node representation.",
+        "bbox": [
+            115,
+            261,
+            489,
+            549
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our CPNC framework on two CSKG completion benchmarks, i.e., CN-100K and ATOMIC. The results show that our model outperforms the state-of-the-art models for this task significantly. Ablation studies demonstrate that both Contrastive Pretraining and Node Clustering modules can significantly improve the performance of CSKG completion. Further experiments verify that our model can consistently improve as the sparsity of knowledge graphs increases: the higher the sparsity, the more improvement our model achieves.",
+        "bbox": [
+            112,
+            551,
+            489,
+            728
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            740,
+            270,
+            755
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We briefly review traditional knowledge graph completion, and then pay more attention to commonsense knowledge graph completion.",
+        "bbox": [
+            112,
+            765,
+            490,
+            813
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional Knowledge Graph Completion Many knowledge graph completion methods have been proposed, which can be classified into three types: embedding-based completion methods, path-finding-based completion methods, and logical rule-based completion methods. The embedding-",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "based method learned the relation and node embedding by shortening the distance between the head-relation pair representation and the tail node representation, which had good scalability. Moreover, convolution has been proven to be an effective operation for acquiring the head-relation representation (Dettmers et al., 2018; Shang et al., 2019) in the embedding-based method. The path-based methods used the random walk inference based algorithm to find the related path. They achieved the prediction of missing tuples by comparing the related path with the relation to be predicted (Lao and Cohen, 2010; Khot et al., 2011). The rule-based methods utilized the induction rules to simplify the path-finding process in the knowledge graphs (Ren et al., 2020; Yang and Song, 2020). We refer the reader to (Ji et al., 2022) for more details about knowledge graph completion.",
+        "bbox": [
+            507,
+            84,
+            884,
+            375
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Knowledge Graph Completion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            386,
+            882,
+            401
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The existing methods assume specific relations, dense edges, and sufficient training samples in the knowledge graph. However, the sparsity of edges and the abstract nature of relations pose challenges for directly applying these methods to the CSKG.",
+        "bbox": [
+            507,
+            401,
+            882,
+            482
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Early methods in this field employed a strategy where, for a given head and tail, all relations were used to generate a large number of tuples. BiLSTMs and linear transformations were then utilized to score these tuples, allowing for the prediction of missing tuples in the CSKG (Saito et al., 2018; Li et al., 2016; Jastrzebski et al., 2018). Building upon this foundation, Shen et al. (2022) took a step further. They normalized the tail nodes and used RoBERTa to score these tuples. This approach resulted in the curation of Dense-ATOMIC, a CSKG with increased coverage and a richer set of multi-hop paths.",
+        "bbox": [
+            507,
+            483,
+            882,
+            692
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "these methods did not take into account structural information and may suffer from computational inefficiency during inference. Another line of studies attempted to use the translated-based approach to reduce computational consumption (Malaviya et al., 2020; Ju et al., 2022; Wang et al., 2021). Furthermore, they incorporated GCN for extracting graph representations and fine-tuned BERT in the CSKG to obtain semantic representations. This integration capitalizes on the complementary nature of these two representations. In addition, these methods have applied convolutional layers to enhance the performance of CSKG completion. This further improves the effectiveness of the com",
+        "bbox": [
+            507,
+            694,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "13978",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/3366c71a2d3c7a89e62ef37b89eada80383cdc3ece5460df6cd96e6ffd2b2603.jpg",
+        "image_caption": [
+            "Figure 2: Overall architecture of CPNC. The Encoder is to get node embedding. Harnessing the power of GCN and Contractive Pretraining, the Encoder transforms CSKG into rich semantic and structural representations. In detail, the lower left section utilizes the mechanism of Contrastive Pretraining. Then, we describe the node clustering process in the lower right, driven by semantic representations from Contractive Pretraining. Finally, the fusion of semantic, structure, and latent concepts in the upper right yields node embeddings. The Decoder part is to rank the nodes in candidates set by Node Embedding."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            82,
+            877,
+            294
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "pletion task. However, it was observed that simply fine-tuning BERT (Devlin et al., 2019) with pretraining tasks is hard to effectively capture the sentence-level semantic connection between two nodes. In response to this limitation, Su et al. (2022) proposed MICO, which learns distinct node representations for head nodes under varying relations, proving evidence that high-quality node representation can significantly aid CSKG completion. Moreover, different nodes in the CSKG may express the same concept. The node redundancy will bring difficulties to the commonsense inference on the CSKG (Jung et al., 2022), while the previous CSKG completion methods ignored.",
+        "bbox": [
+            110,
+            416,
+            487,
+            640
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Compared to these methods, we address the issue of node redundancy by employing a clustering algorithm. Our work stands out by learning the semantic representations of nodes through sentence-level semantic connections and integrating latent concept information into node representations.",
+        "bbox": [
+            112,
+            640,
+            489,
+            737
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Task Definition",
+        "text_level": 1,
+        "bbox": [
+            112,
+            749,
+            282,
+            764
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given a CSKG  $G = (N, V, R)$  where  $N$  is the set of nodes,  $V$  is the set of edges and  $R$  is the set of relations. We regard each tuple  $v_{i} = (h, rel, t)$  in the CSKG as a sample, which is composed of head node  $h$ , tail node  $t$  and relation  $rel$ , where  $h, t \\in N$ ,  $v_{i} \\in V$ , and  $rel \\in R$ . Given a query  $(h, rel)$  formed by a head node  $h$  and a relation  $rel$ , the target of the CSKG completion is to maximize the score of the tail node  $t$ . Following the",
+        "bbox": [
+            112,
+            772,
+            487,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "previous work (Malaviya et al., 2020), for an edge  $(h,rel,t)$  existing in the CSKG, we also add an inverse edge  $(t,rel^{-1},h)$  to the graph, where  $rel^{-1}$  is the inverse relation of rel.",
+        "bbox": [
+            507,
+            416,
+            882,
+            479
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Approach",
+        "text_level": 1,
+        "bbox": [
+            507,
+            492,
+            631,
+            508
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose a new framework, CPNC, for CSKG completion using an Encoder-Decoder architecture (see Figure 2). The Encoder incorporates semantic, graph structure, and latent concept representation obtained from Contrastive Pretraining, Graph Convolutional Network (GCN), and Node Clustering, respectively, to acquire node representations. Given query  $(h,rel)$ , the Decoder ranks the nodes in the candidate set  $N$  and finds tail nodes for the query by the rank.",
+        "bbox": [
+            505,
+            517,
+            882,
+            678
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Contrastive Pretraining",
+        "text_level": 1,
+        "bbox": [
+            507,
+            689,
+            742,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The current mainstream approach for completing CSKG utilizes translation-based methods and relies on GCN to represent the graph structure of nodes. Although GCNs are effective in modeling graph structure, they have limitations in capturing graph structure information in CSKG due to sparse edges. To address this issue, these methods incorporate semantic information by fine-tuning BERT on CSKG using Masked Language Model (MLM) and Next Sentence Prediction (NSP) tasks.",
+        "bbox": [
+            505,
+            709,
+            882,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, those methods are imperfect in modeling CSKGs composed of nodes presented as short text. MLM, which is a token-level task, focuses on",
+        "bbox": [
+            507,
+            870,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "13979",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "modeling the connection between masked words and other words, making it unsuitable for achieving sentence-level node representation. Similarly, NSP, while capable of modeling the semantic connection between two nodes at the sentence level, suffers from a mismatch between the input of the pretraining phase and the CSKG completion phase. During pretraining, NSP requires a pair of head and tail nodes for prediction, but during CSKG completion, only a single node is used for sentence representation. Consequently, NSP is not well-suited for the task of learning node representations, and similar observations have been made in the field of sentence representation learning (Li et al., 2020).",
+        "bbox": [
+            112,
+            84,
+            489,
+            309
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To obtain better node representations for the CSKG completion, we introduce Contrastive Pretraining (CP), a new method that sufficiently leverages semantic information at the sentence level. CP focuses on CSKG completion and differs from MICO (Su et al., 2022) by not incorporating relation categories. Instead, it fine-tunes BERT's node representation using contrastive learning and capture sentence-level connections between nodes.",
+        "bbox": [
+            112,
+            310,
+            489,
+            455
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.1 Building Contrastive Learning Samples",
+        "text_level": 1,
+        "bbox": [
+            112,
+            464,
+            484,
+            481
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Assuming node pairs linked in the CSKG are semantically related, we consider node pairs without a link in the CSKG as semantically unrelated.",
+        "bbox": [
+            112,
+            485,
+            489,
+            533
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each arbitrary edge  $(h,rel,t)$  in CSKG, we randomly sample a node  $\\bar{t}$  having no linking with the head node  $h$  as its hard negative tail and construct a contrastive learning training sample  $c = (h,t,\\bar{t})$ . As shown in Figure 3, \"take breakfast\" and \"cook food\" are the head and tail nodes in a edge of CSKG. We randomly select a node \"find information\" having no linking with \"take breakfast\" to construct a training sample (\"take breakfast\", \"cook food\", \"find information\").",
+        "bbox": [
+            110,
+            533,
+            489,
+            695
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The previous methods primarily focused on the head nodes while constructing negative samples, neglecting the importance of capturing unrelated semantics among the tail node set. To tackle this problem, we additionally constructed a contrastive learning sample  $c_{\\text{reverse}} = (t, h, \\overline{h})$  for the tail node  $t$  in the edge  $(h, \\text{rel}, t)$ .",
+        "bbox": [
+            112,
+            696,
+            489,
+            808
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1.2 Multiple Negatives Ranking Loss",
+        "text_level": 1,
+        "bbox": [
+            112,
+            819,
+            433,
+            834
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We employ BERT for node embedding by inputting all nodes in a batch separately and applying Mean Pooling on the final layer's output. Then, we use  $x(n)$  to denote the representation of a specific node  $n$  obtained through BERT.",
+        "bbox": [
+            112,
+            838,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/64cdbc3434e2f846addd358851f9d42c4924505760880259037f1f93c4b2929e.jpg",
+        "image_caption": [
+            "Figure 3: Main process of building Constructive Pretraining samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            534,
+            82,
+            863,
+            401
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In order to ensure that semantically related nodes are positioned closely in the embedding space, we utilize contrastive learning and employ the Multiple Negatives Ranking Loss for contrastive learning.",
+        "bbox": [
+            507,
+            472,
+            882,
+            536
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For a given sample  $c_{i} = (h_{i}, t_{i}, \\bar{t}_{i})$  in a batch, we employ the Multiple Negatives Ranking Loss. We consider  $h_{i}$  paired with  $t_{i}$  as the positive sample while treating  $h_{i}$  paired with other tail nodes  $t_{j}$  in the batch as negative samples. Additionally, we consider  $h_{i}$  paired with all hard negative tails in the batch as negative samples as well. The objective is to minimize the semantic distance between the nodes in positive samples within the batch. The formulation of the multiple negatives ranking loss is as follows:",
+        "bbox": [
+            507,
+            537,
+            882,
+            712
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = - \\sum_ {i = 1} ^ {M} \\log \\frac {e ^ {D \\left(h _ {i} , t _ {i}\\right))}}{\\sum_ {j = 1} ^ {M} e ^ {D \\left(h _ {i} , t _ {j}\\right)} + e ^ {D \\left(h _ {i} , \\bar {t} _ {j}\\right)}}, \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            526,
+            736,
+            882,
+            780
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $D(h,t) = (x(h)\\cdot x(t)^T) / (\\| x(h)\\| \\cdot \\| x(t)\\|)$  represents the cosine similarity of two nodes,  $M$  is the number of samples in the a batch.",
+        "bbox": [
+            507,
+            790,
+            882,
+            839
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Encoder",
+        "text_level": 1,
+        "bbox": [
+            507,
+            850,
+            621,
+            864
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The Encoder aims to generate node representations for CSKG completion, as illustrated in the left part of Figure 2.",
+        "bbox": [
+            507,
+            871,
+            880,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "13980",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The node representation comprises two components,",
+        "bbox": [
+            112,
+            84,
+            489,
+            115
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "1) the semantic representation matrix of nodes  $E_{sem}$  obtained through Contrastive Pretraining:",
+        "bbox": [
+            114,
+            116,
+            485,
+            148
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {s e m} = \\text {C o n t r a s t i v e - P r e t r a i n i n g} (N), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            161,
+            158,
+            487,
+            174
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2) the graph structure representation matrix of nodes acquired using a Graph Convolutional Network (GCN),",
+        "bbox": [
+            112,
+            185,
+            489,
+            231
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nE _ {\\text {g r a p h}} = \\operatorname {G C N} (G). \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            223,
+            242,
+            487,
+            260
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Specifically,  $e_{sem}$  and  $e_{graph}$  represent the semantic and graph representations of a particular node, respectively, which can be derived from  $E_{sem}$  and  $E_{graph}$ .",
+        "bbox": [
+            112,
+            269,
+            487,
+            335
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.3 Node Clustering",
+        "text_level": 1,
+        "bbox": [
+            112,
+            343,
+            290,
+            359
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "On the basis of the semantic node representation  $e_{sem}$ , we use the K-means algorithm to cluster all nodes  $N$  in the graph as follows:",
+        "bbox": [
+            112,
+            363,
+            487,
+            412
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\left\\{\\mathcal {C} _ {1}, \\mathcal {C} _ {2}, \\dots , \\mathcal {C} _ {K} \\right\\} = \\mathrm {K} - \\text {m e a n s} (N), \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            422,
+            487,
+            439
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathcal{C}_i$  is the  $i$ -th cluster, and  $K$  represents the number of clusters. Nodes in  $\\mathcal{C}_i$  are considered to have the same latent concept. As shown in the Node Clustering part of Figure 2, \"eat breakfast\", \"have breakfast\" and \"take breakfast\" are in the same cluster and have the same latent concept.",
+        "bbox": [
+            112,
+            448,
+            487,
+            543
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We define the representation of the latent concept as the mean of semantic representation of all nodes in a cluster, denoted as  $\\mathrm{mean}(\\mathcal{C}_k)$ , where  $\\mathcal{C}_k$  contains a set of nodes in the same cluster.",
+        "bbox": [
+            112,
+            545,
+            487,
+            608
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To acquire the final node representation, we combine the node semantic representation,  $e_{sem}$ , the graph representation,  $e_{graph}$ , and the latent concept representation,  $\\mathrm{mean}(\\mathcal{C}_k)$ . This fusion is performed as follows:",
+        "bbox": [
+            112,
+            609,
+            487,
+            688
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\ne = \\left[ e _ {s e m}; e _ {g r a p h}; \\operatorname {m e a n} \\left(\\mathcal {C} _ {k}\\right) \\right] \\cdot W _ {\\text {e m b e d d i n g}}, \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            699,
+            487,
+            717
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{embedding}$  is a weight matrix. The resulting node embeddings are considered as the node embedding space.",
+        "bbox": [
+            112,
+            726,
+            487,
+            772
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Intuitively, Node Clustering enhances CSKG completion by leveraging information from nodes within the same latent concept.",
+        "bbox": [
+            112,
+            774,
+            487,
+            822
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We also use a progressive masking process, following Malaviya et al. (2020), to integrate latent concept information. Initially, the latent concept representation is fully masked, and throughout the first 100 training epochs, it is gradually unmasked, improving the overall performance.",
+        "bbox": [
+            112,
+            822,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.4 Decoder",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            621,
+            98
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The goal of Decoder is to rank the nodes in the candidate set, which is considered as  $N$  in our work.",
+        "bbox": [
+            507,
+            107,
+            880,
+            153
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to obtain the query representation, we first use a convolutional layer to fuse the relation and head node representation:",
+        "bbox": [
+            507,
+            156,
+            880,
+            203
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nq = \\operatorname {C o n v} \\left(e _ {h}, e _ {r e l}\\right), \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            620,
+            219,
+            880,
+            237
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $e_h$  is the representation of the head nodes in node embedding space obtained from Eq. 5,  $e_{rel}$  is the representation of relation rel and Conv is a convolution operation.",
+        "bbox": [
+            507,
+            252,
+            880,
+            315
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Then, we predict a 0-1 distribution vector, based on the representation of the query  $q$ , indicating the likelihood of each node in the candidate set  $N$  becoming the tail node:",
+        "bbox": [
+            507,
+            317,
+            880,
+            380
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\np _ {t} = \\delta \\left(q W _ {\\text {c o n v}} E _ {N}\\right), \\tag {7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            616,
+            397,
+            880,
+            414
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{conv}$  is a weight matrix,  $E_N$  is the representation matrix of candidate nodes in node embedding space obtained from Eq. 5, and  $\\delta$  is a sigmoid function.",
+        "bbox": [
+            507,
+            429,
+            882,
+            492
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            508,
+            653,
+            524
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            536,
+            712,
+            552
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            507,
+            558,
+            636,
+            571
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate our CPNC framework on two CSKG completion benchmarks, i.e., CN-100K (Speer et al., 2017) and ATOMIC (Sap et al., 2019).",
+        "bbox": [
+            507,
+            579,
+            880,
+            626
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "CN-100K is a dataset that encompasses general commonsense knowledge. This version contains 36 relation types and the Open Mind Common Sense (OMCS) entries from ConceptNet (Speer and Havasi, 2013). The average length of the nodes in CN-100K is 2.85 words. Following Malaviya et al. (2020)'s split, the training set contains 10,000 tuples, and the validation set and test set both contain 1200 tuples.",
+        "bbox": [
+            507,
+            627,
+            882,
+            772
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ATOMIC is an atlas of everyday commonsense reasoning and primarily focuses on event-level commonsense knowledge in the form of if-then relations. It comprises 9 relation types, with an average of 4.40 words per node. We split ATOMIC following Malaviya et al. (2020)'s work, where the training set consists of 610,536 tuples, while the validation and test sets contain 87,700 tuples and 87,701 tuples, respectively.",
+        "bbox": [
+            507,
+            774,
+            880,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "13981",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/968994aebd9802d11e02be88fcca8e62e48a5cdb327eff338549c6d572a3ec91.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Methods</td><td colspan=\"4\">CN-100K</td><td colspan=\"4\">ATOMIC</td></tr><tr><td>Type</td><td>Model</td><td>MRR</td><td>HITs@1</td><td>@3</td><td>@10</td><td>MRR</td><td>HITs@1</td><td>@3</td><td>@10</td></tr><tr><td rowspan=\"4\">KG-adapted</td><td>DistMult</td><td>8.97</td><td>4.51</td><td>9.76</td><td>17.44</td><td>12.39</td><td>9.24</td><td>15.18</td><td>18.30</td></tr><tr><td>ComplEx</td><td>11.40</td><td>7.42</td><td>12.45</td><td>19.01</td><td>14.24</td><td>13.27</td><td>14.13</td><td>15.96</td></tr><tr><td>ConvE</td><td>20.88</td><td>13.97</td><td>22.91</td><td>34.02</td><td>10.07</td><td>8.24</td><td>10.29</td><td>13.37</td></tr><tr><td>ConvTransE</td><td>18.68</td><td>7.87</td><td>23.87</td><td>38.95</td><td>12.94</td><td>12.92</td><td>12.95</td><td>12.98</td></tr><tr><td rowspan=\"2\">Generation-based</td><td>COMeT-Normalized</td><td>6.07</td><td>0.08</td><td>2.92</td><td>21.17</td><td>3.36</td><td>0.00</td><td>2.15</td><td>15.75</td></tr><tr><td>COMeT-Total</td><td>6.21</td><td>0.00</td><td>0.00</td><td>24.00</td><td>4.91</td><td>0.00</td><td>2.40</td><td>21.60</td></tr><tr><td rowspan=\"3\">CSKG-dedicated</td><td>RGAT</td><td>43.97</td><td>30.75</td><td>51.54</td><td>69.34</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SGBC</td><td>49.12*</td><td>37.71*</td><td>56.67*</td><td>71.29*</td><td>10.25*</td><td>8.72*</td><td>10.54*</td><td>13.26*</td></tr><tr><td>Inductive</td><td>56.92*</td><td>45.54*</td><td>63.38*</td><td>78.63*</td><td>13.19*</td><td>10.26*</td><td>13.61*</td><td>18.83*</td></tr><tr><td rowspan=\"2\">Ours</td><td>CPNC-S</td><td>54.52</td><td>45.33</td><td>61.46</td><td>75.92</td><td>13.14</td><td>10.11</td><td>13.75</td><td>18.80</td></tr><tr><td>CPNC-I</td><td>59.00</td><td>48.29</td><td>65.04</td><td>79.13</td><td>14.38</td><td>10.53</td><td>15.22</td><td>21.79</td></tr></table>",
+        "bbox": [
+            119,
+            82,
+            885,
+            237
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: CSKG completion results on CN-100K and ATOMIC. The result of HITS@1 was not reported in Wang et al. (2021). To fairly compare our method with the previous method in the same setting, we rerun the code of InductivE and SGBC on ATOMIC and CN-100K and mark the results of rerun experiments with *. Those results are a little lower than that in the original paper. For KG-adapted and Generation methods, we reuse the results reported in Malaviya et al. (2020).",
+        "bbox": [
+            112,
+            252,
+            884,
+            324
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1.2 Evaluation Metric",
+        "text_level": 1,
+        "bbox": [
+            112,
+            349,
+            317,
+            362
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the performance of our method using MRR and HITS, following previous CSKG completion methods (Wang et al., 2021; Malaviya et al., 2020; Ju et al., 2022). The results are reported by averaging over both forward tuples  $(h, rel, t)$  and inverse tuples  $(t, rel^{-1}, h)$ . Moreover, because the nodes in CSKG are represented in free-form text, it is possible for nodes other than the golden tail nodes to be considered reasonable tail nodes. To address this, we conduct a human evaluation to assess the predictions made by our models.",
+        "bbox": [
+            112,
+            370,
+            487,
+            546
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1.3 Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            558,
+            359,
+            574
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To perform Contrastive Pretraining, we use the contrastive learning framework provided in https://github.com/UKPLab/sentence-transformers. Our approach employs BERT-large as the base model, which contains 340M parameters. During training, we utilized a batch size of 128 and conducted training for 3 epochs on the CN-100K and ATOMIC datasets. We choose the Adam optimizer for optimization, setting the learning rate to 1e-4 for BERT-large and 5e-5 for the MLP. For the remaining hyperparameters, we used the default values provided by the framework.",
+        "bbox": [
+            112,
+            580,
+            489,
+            772
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "For the CPNC-I and CPNC-S model, we use experimental settings proposed by Malaviya et al. (2020) and Wang et al. (2021), respectively. Both models are trained for a minimum of 200 epochs using BERT-large, which consists of 340M parameters, to encode semantic representations. During training, we evaluate the MRR on the development set every 10 epochs for CN-100K and ATOMIC. Training continues until no further improvement in",
+        "bbox": [
+            112,
+            774,
+            489,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "MRR is observed. We select the model checkpoint that achieves the highest MRR on the development set for testing.",
+        "bbox": [
+            507,
+            349,
+            882,
+            397
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Compared Systems",
+        "text_level": 1,
+        "bbox": [
+            507,
+            418,
+            707,
+            434
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare our approach with nine baseline systems across three categories.",
+        "bbox": [
+            507,
+            444,
+            882,
+            476
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.1 KG-adapted Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            495,
+            741,
+            511
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We adapt classic knowledge graph completion methods for CSKG completion. DistMult (Yang et al., 2015) employed a bi-linear product to calculate score of a tuple; ComplE (Trouillon et al., 2016) utilized imaginary number representation to effectively handle a large number of relations in the knowledge graph; ConvE (Dettmers et al., 2018) fused the representation of the source node and the relation through a 2D convolution layer to obtain the representation of query (h, rel); ConvTransE extended ConvE by incorporating the translational properties of TransE.",
+        "bbox": [
+            507,
+            521,
+            882,
+            714
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.2 Generation-based Methods",
+        "text_level": 1,
+        "bbox": [
+            507,
+            733,
+            781,
+            747
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "COMeT (Bosselut et al., 2019) is a Transformer-based knowledge generation model. Following Malaviya et al. (2020), we adapt COMeT for the CSKG completion and only evaluate in the forward direction. Additionally, we also use their modifications to the ranking method used in COMeT. Specifically, COMET-Normalized and COMET-Total use the normalized negative log-likelihood scores and total log-likelihood scores, respectively, to rank nodes in the candidate set.",
+        "bbox": [
+            507,
+            757,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "13982",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2.3 CSKG-dedicated Methods",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            381,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "SIM+GCN+BERT+ConvTransE (SGBC) densified the CSKG by connecting the synthetic edges between similar semantic nodes, improving the graph structure representation (Malaviya et al., 2020). Moreover, they used fine-tuned BERT to encode the semantic information of the nodes.",
+        "bbox": [
+            112,
+            103,
+            487,
+            198
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Inductive (Wang et al., 2021) is proposed to enhance the unseen entity representation with neighboring structural information by densifying Graph. Relational graph attention networks (RGAT) are proposed weighted the importance of neighbor nodes of each node to obtain a better node representation.",
+        "bbox": [
+            112,
+            200,
+            489,
+            311
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2.4 Our CPNC Methods",
+        "text_level": 1,
+        "bbox": [
+            112,
+            321,
+            337,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "By incorporating CPNC with two CSKG-dedicated methods (SGBC and InductivE), we obtain two models, i.e., CPNC-S and CPNC-I.",
+        "bbox": [
+            112,
+            340,
+            487,
+            388
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Main Results",
+        "text_level": 1,
+        "bbox": [
+            112,
+            399,
+            265,
+            413
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In Table 1, we report experimental results of KG-adapted methods, Generation methods, CSKG-dedicated methods and our methods on CN-100K and ATOMIC.",
+        "bbox": [
+            112,
+            420,
+            489,
+            482
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On CN-100K, both KG-adapted methods and Generation methods exhibit unsatisfactory performance, with MRR values below  $21\\%$  and HITS@1 values below  $14\\%$ . These results demonstrate that directly applying KG-adapted and Generation methods to CSKG completion is ineffective. In contrast, CSKG-dedicated methods achieve significantly better results on CN-100K, with MRR values above  $43\\%$  and HITS@1 values above  $30\\%$ . These metrics are twice as high as those obtained by KG-adapted and Generation methods, highlighting the advantage of CSKG-dedicated methods. Comparing our proposed CPNC-S and CPNC-I models with mainstream CSKG-dedicated methods (SGBC and Inductive), we observe performance improvements of  $5.40\\%$  and  $2.08\\%$  on MRR, respectively. Notably, our CPNC-I model sets a new state-of-the-art result on CN-100K.",
+        "bbox": [
+            112,
+            483,
+            489,
+            772
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On ATOMIC, we draw a similar conclusion. The KG-adapted methods exhibit poor performance, while CSKG-dedicated methods achieve better results. In addition, compared with the SGBC and Inductive models, our CPNC-S and CPNC-I models outperform the SGBC and Inductive models by  $2.89\\%$  and  $1.19\\%$  on MRR, respectively. And the CPNC-I model achieves the state-of-the-art result on ATOMIC.",
+        "bbox": [
+            112,
+            774,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We conduct Paired t-Test and the result proves that the improvement of CPNC is significant.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            509,
+            127,
+            673,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/b590e1ca7184abac8b5af4ef9343528a3add76f24ede50ee6fe5d3aa5935cd40.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td></td><td>MRR</td><td>HITS@10</td></tr><tr><td rowspan=\"8\">CN-100K</td><td>CPNC-S</td><td>54.52</td><td>75.92</td></tr><tr><td>-w/o CP</td><td>51.90</td><td>76.08</td></tr><tr><td>-w/o NC</td><td>49.74</td><td>71.38</td></tr><tr><td>SGBC</td><td>49.12</td><td>71.29</td></tr><tr><td>CPNC-I</td><td>59.00</td><td>79.13</td></tr><tr><td>-w/o CP</td><td>57.16</td><td>74.90</td></tr><tr><td>-w/o NC</td><td>58.30</td><td>78.75</td></tr><tr><td>InductivE</td><td>56.92</td><td>78.63</td></tr><tr><td rowspan=\"8\">ATOMIC</td><td>CPNC-S</td><td>13.14</td><td>18.80</td></tr><tr><td>-w/o CP</td><td>12.77</td><td>17.15</td></tr><tr><td>-w/o NC</td><td>12.46</td><td>17.72</td></tr><tr><td>SGBC</td><td>10.25</td><td>13.26</td></tr><tr><td>CPNC-I</td><td>14.38</td><td>21.79</td></tr><tr><td>-w/o CP</td><td>13.22</td><td>19.16</td></tr><tr><td>-w/o NC</td><td>14.21</td><td>21.38</td></tr><tr><td>InductivE</td><td>13.19</td><td>18.83</td></tr></table>",
+        "bbox": [
+            546,
+            158,
+            848,
+            366
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: MRR and HITS@10 of removing CP and NC from CPNC on CN-100K and ATOMIC.",
+        "bbox": [
+            507,
+            374,
+            880,
+            403
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the effectiveness of Contrastive Pretraining (CP) and Node Clustering (NC) on CSKG completion, we perform the ablation study where we removed CP and NC. The results are presented in Table 2.",
+        "bbox": [
+            507,
+            420,
+            880,
+            500
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "When NC was removed, we can observe that the performance of both CPNC-S and CPNC-I decreases on CN-100K and ATOMIC. This indicates that NC's latent concept representation effectively assists CSKG completion. It is worth noting that CP can still bring improvement compared with the previous method, showing that the semantic representation provided by CP is crucial for CSKG completion.",
+        "bbox": [
+            507,
+            501,
+            882,
+            645
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "CPNC, which combines NC and CP, consistently yields the highest result. By replacing CP with finetuned BERT (Malaviya et al., 2020), the performance of both CPNC-S and CPNC-I drops. However, they still achieve comparable results with the previous CSKG-dedicated methods. This demonstrates that the CP requires a high-quality semantic representation of nodes to acquire good latent concept representation.",
+        "bbox": [
+            507,
+            646,
+            884,
+            790
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.5 Human Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            801,
+            705,
+            816
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "For a given query  $(h,rel)$ , additional nodes besides the golden tail node could also become reasonable tail nodes. For example, given a query “(do house-work, Causes)”, besides golden tail node “clean the house”, “house get clean” and “clean the room” are also reasonable tail nodes. However, automated",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "13983",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "metrics fail to cover these nodes. To address this issue, we randomly select 200 queries from the test set of the CN-100K and ATOMIC. For each query, we use our methods and current mainstream methods to rank the nodes in the candidate set, We manually calculate the number of reasonable tail nodes in the top 10 candidates. The results are shown in Table 3.",
+        "bbox": [
+            112,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/f4b90a52c70e7729a2d48f119fe14ce868cc429af96343b89396da0d2b604c1e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>ATOMIC</td><td>CN-100K</td></tr><tr><td>CPNC-I</td><td>70.22</td><td>83.20</td></tr><tr><td>Inductive</td><td>64.45</td><td>76.40</td></tr><tr><td>CPNC-S</td><td>60.23</td><td>69.40</td></tr><tr><td>SGBC</td><td>62.11</td><td>65.40</td></tr></table>",
+        "bbox": [
+            181,
+            222,
+            421,
+            286
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Average percentage of reasonable tuples by human evaluation results.",
+        "bbox": [
+            112,
+            296,
+            485,
+            324
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Observing human evaluation results, the CPNC model has an excellent performance in CSKG completion. Besides, our method is significantly higher than the current mainstream methods under the human evaluation results, which further validates the effectiveness of our approach.",
+        "bbox": [
+            112,
+            344,
+            487,
+            442
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.6 Model Adaptability",
+        "text_level": 1,
+        "bbox": [
+            112,
+            456,
+            314,
+            470
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The CPNC can be effectively applied to various methods. In our experiment, we incorporate CPNC into SGBC and InductivE, resulting in significant improvements.",
+        "bbox": [
+            112,
+            479,
+            485,
+            542
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "The SGBC method initializes the input of GCN randomly and combines graph structural representation acquired by GCN and semantic representation obtained by BERT as the node representation; InductiveE initializes the input of GCN with BERT and uses the output of GCN as the node representation. Despite the differences in their structures, both methods benefit from the application of CP and NC.",
+        "bbox": [
+            112,
+            543,
+            487,
+            687
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "On CN-100K, CPNC-S and CPNC-I bring improvements of  $5.40\\%$  and  $2.08\\%$  in MRR, respectively; On ATOMIC, CPNC-S and CPNC-I yield MRR improvements of  $2.89\\%$  and  $1.19\\%$ , respectively. This result demonstrates the adaptability of our method.",
+        "bbox": [
+            112,
+            689,
+            487,
+            784
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.7 Discussion on the Effect of Relieving Sparsity",
+        "text_level": 1,
+        "bbox": [
+            112,
+            800,
+            445,
+            832
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "To evaluate the effectiveness of CPNC in mitigating the impact of edge sparsity in the CSKG, we conducted a comparative study between CPNC-S and SGBC on CN-100K with varying sparsity levels. The removal of edges in the CSKG results in",
+        "bbox": [
+            112,
+            839,
+            487,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "increased sparsity, with a higher number of deleted edges corresponding to a higher sparsity degree.",
+        "bbox": [
+            507,
+            84,
+            880,
+            116
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/bd60b82d789d20f700afa3335a1117e68142408a574f2e866c234edb6a749002.jpg",
+        "image_caption": [
+            "Figure 4: CSKG completion results of SGBC and CPNC-S on CN-100K with various degrees of sparsity."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            534,
+            131,
+            857,
+            300
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 4, when the sparsity degree increases, CSKG completion performance deteriorates. CPNC-S consistently outperforms SGBC in terms of MRR on CN-100K across different sparsity levels, indicating CPNC's ability to mitigate edge sparsity in the CSKG. CPNC-S achieves a  $12.48\\%$  and  $3.95\\%$  increase in MRR compared to SGBC at sparsity extents of  $25\\%$  and  $50\\%$  respectively. These results demonstrate that CPNC is more effective in alleviating edge sparsity in the CSKG, with greater improvements observed for sparser graphs.",
+        "bbox": [
+            507,
+            374,
+            882,
+            567
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.8 Discussion on the Number of Latent Concepts",
+        "text_level": 1,
+        "bbox": [
+            507,
+            577,
+            836,
+            609
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/38e969457fec383c734630948f379dee9ade0538ffd83ca7cec056cbce55514e.jpg",
+        "image_caption": [
+            "Figure 5: CSKG completion performance with different cluster numbers on CN-100K and ATOMIC."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            534,
+            630,
+            852,
+            778
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Different numbers of clusters can lead to varying levels of granularity in latent concepts, which in turn can impact the performance of CSKG completion. In order to investigate this, we conducted experiments with the different numbers of clusters",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "13984",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "on CN-100K and ATOMIC datasets and present the results in Figure 5.",
+        "bbox": [
+            112,
+            84,
+            485,
+            115
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "On CN-100K, we observed an overall increasing trend in completion results as the number of clusters increased. However, the rate of improvement tends to diminish after reaching 900 clusters. The maximum MRR achieved when using the number of clusters is 1200.",
+        "bbox": [
+            112,
+            116,
+            485,
+            211
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "On ATOMIC, the MRR initially increases with an increasing number of clusters but starts to decline once the number of clusters reaches 900.",
+        "bbox": [
+            112,
+            212,
+            485,
+            260
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Based on our main results, we selected 1200 clusters for CN-100K and 900 clusters for ATOMIC, allowing us to effectively capture latent concepts through cluster nodes.",
+        "bbox": [
+            112,
+            262,
+            485,
+            326
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.9 Discussing on Contrastive Learning Loss",
+        "text_level": 1,
+        "bbox": [
+            112,
+            338,
+            480,
+            354
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In order to obtain a better semantic representation of nodes, we experiment with five kinds of contrastive learning methods and observe their performance in CSKG completion. The result is shown in Table 4.",
+        "bbox": [
+            112,
+            359,
+            485,
+            437
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "table",
+        "img_path": "images/fe43d802489fecc5c3a42d00b5dbd74bcb411859fedae62286618222228a4791.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>MRR</td></tr><tr><td>Contrastive Loss</td><td>40.03</td></tr><tr><td>MTriplet Loss</td><td>44.36</td></tr><tr><td>MICO</td><td>45.93</td></tr><tr><td>Batch Semi Hard Triplet Loss</td><td>27.92</td></tr><tr><td>Multiple Negatives Ranking Loss</td><td>49.74</td></tr></table>",
+        "bbox": [
+            137,
+            447,
+            463,
+            533
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 4: CSKG completion performance of CPNC-S using different Contrastive Learning Loss on CN-100K.",
+        "bbox": [
+            112,
+            542,
+            485,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In order to compare those contrastive learning methods, we use the node embedding obtained by those methods to complete CSKG without node clustering. Each contrastive learning method yields distinct results, and our main result selects Multiple Negatives Ranking Loss due to its highest MRR.",
+        "bbox": [
+            112,
+            590,
+            485,
+            686
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            700,
+            245,
+            714
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose a new CSKG completion framework CPNC to address issues arising from node redundancy and edge sparsity. CPNC obtains better semantic node information through Contrastive Pretraining, which alleviates the problems caused by edge sparsity. CPNC also utilizes the latent concept representation acquired through Node Clustering to alleviate the problem caused by node redundancy. On CN-100K and ATOMIC, experimental results and extensive analysis demonstrate the effectiveness of Contrastive Pretraining and Node Clustering.",
+        "bbox": [
+            112,
+            726,
+            487,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            613,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Due to the limitation of time and resources, in this work, we select a relatively small number of clusters during the clustering process, which results in coarse-grained clustering. Fine-grained clustering can provide a better latent concept but will also lead to increased computational resources and time consumption. We will attempt to trade-off between the cost and the granularity of clustering in future work to further explore the impact of the latent concept on CSKG completion. Besides, our Node Clustering module is not integrated in an end-to-end manner in our work; We will consider using the topic neural network to construct an end-to-end model.",
+        "bbox": [
+            507,
+            111,
+            882,
+            335
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            351,
+            660,
+            366
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We would like to thank Malaviya et al. (2020) and Wang et al. (2021) for their code on commonsense knowledge graph completion. Their models are licensed under MIT, which allows copying, modifying, merging, publishing, and distributing of the material.",
+        "bbox": [
+            507,
+            378,
+            882,
+            473
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In the stage of human evaluation, we employed three graduate students experienced in natural language processing for human evaluation. We paid the graduate students about \\(8 per hour, well above the local average wage, and engaged in constructive discussions if they had concerns about the process.",
+        "bbox": [
+            507,
+            476,
+            882,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            588,
+            680,
+            604
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was supported by the Natural Science Foundation of China (No. 62076133), and the Natural Science Foundation of Jiangsu Province for Distinguished Young Scholars (No. BK20200018).",
+        "bbox": [
+            507,
+            615,
+            882,
+            678
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            709,
+            608,
+            722
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. COMET: commonsense transformers for automatic knowledge graph construction. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 4762-4779. Association for Computational Linguistics.",
+            "Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence,"
+        ],
+        "bbox": [
+            509,
+            733,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "13985",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "(AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 1811-1818. AAAI Press.",
+            "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, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.",
+            "Stanislaw Jastrzebski, Dzmitry Bahdanau, Seyedarian Hosseini, Michael Noukhovitch, Yoshua Bengio, and Jackie Chi Kit Cheung. 2018. Commonsense mining as knowledge base completion? A study on the impact of novelty. CoRR, abs/1804.09259.",
+            "Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Martinen, and Philip S. Yu. 2022. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Trans. Neural Networks Learn. Syst., 33(2):494-514.",
+            "Jinhao Ju, Deqing Yang, and Jingping Liu. 2022. Commonsense knowledge base completion with relational graph attention network and pre-trained language model. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 4104-4108. ACM.",
+            "Yong-Ho Jung, Jun-Hyung Park, Joon-Young Choi, Mingyu Lee, Junho Kim, Kang-Min Kim, and SangKeun Lee. 2022. Learning from missing relations: Contrastive learning with commonsense knowledge graphs for commonsense inference. In *Findings of the Association for Computational Linguistics: ACL* 2022, Dublin, Ireland, May 22-27, 2022, pages 1514-1523. Association for Computational Linguistics.",
+            "Tushar Khot, Sriraam Natarajan, Kristian Kersting, and Jude W. Shavlik. 2011. Learning markov logic networks via functional gradient boosting. In 11th IEEE International Conference on Data Mining, ICDM 2011, Vancouver, BC, Canada, December 11-14, 2011, pages 320-329. IEEE Computer Society.",
+            "Ni Lao and William W. Cohen. 2010. Relational retrieval using a combination of path-constrained random walks. Mach. Learn., 81(1):53-67.",
+            "Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 9119-9130. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jiangnan Li, Fandong Meng, Zheng Lin, Rui Liu, Peng Fu, Yanan Cao, Weiping Wang, and Jie Zhou. 2022. Neutral utterances are also causes: Enhancing conversational causal emotion entailment with social commonsense knowledge. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 4209-4215. ijcai.org.",
+            "Xiang Li, Aynaz Taheri, Lifu Tu, and Kevin Gimpel. 2016. Commonsense knowledge base completion. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.",
+            "Shangwen Lv, Daya Guo, Jingjing Xu, Duyu Tang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, and Songlin Hu. 2020. Graph-based reasoning over heterogeneous external knowledge for commonsense question answering. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 8449-8456. AAAI Press.",
+            "Chaitanya Malaviya, Chandra Bhagavatula, Antoine Bosselut, and Yejin Choi. 2020. Commonsense knowledge base completion with structural and semantic context. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 2925-2933. AAAI Press.",
+            "Hongyu Ren, Weihua Hu, and Jure Leskovec. 2020. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.",
+            "Itsumi Saito, Kyosuke Nishida, Hisako Asano, and Junji Tomita. 2018. Commonsense knowledge base completion and generation. In Proceedings of the 22nd Conference on Computational Natural Language Learning, CoNLL 2018, Brussels, Belgium, October 31 - November 1, 2018, pages 141-150. Association for Computational Linguistics.",
+            "Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019. ATOMIC: an atlas of machine commonsense for if-then reasoning. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA,"
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "13986",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "January 27 - February 1, 2019, pages 3027-3035. AAAI Press.",
+            "Chao Shang, Yun Tang, Jing Huang, Jinbo Bi, Xiaodong He, and Bowen Zhou. 2019. End-to-end structure-aware convolutional networks for knowledge base completion. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 3060-3067. AAAI Press.",
+            "Xiangqing Shen, Siwei Wu, and Rui Xia. 2022. Dense-atomic: Construction of densely-connected and multihop commonsense knowledge graph upon ATOMIC. CoRR, abs/2210.07621.",
+            "Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4444-4451. AAAI Press.",
+            "Robyn Speer and Catherine Havasi. 2013. Conceptnet 5: A large semantic network for relational knowledge. In Iryna Gurevych and Jungi Kim, editors, The People's Web Meets NLP, Collaboratively Constructed Language Resources, Theory and Applications of Natural Language Processing, pages 161-176. Springer.",
+            "Ying Su, Zihao Wang, Tianqing Fang, Hongming Zhang, Yangqiu Song, and Tong Zhang. 2022. MICO: A multi-alternative contrastive learning framework for commonsense knowledge representation. CoRR, abs/2210.07570.",
+            "Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 2071-2080. JMLR.org.",
+            "Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. MISC: A mixed strategy-aware model integrating COMET for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 308-319. Association for Computational Linguistics.",
+            "Bin Wang, Guangtao Wang, Jing Huang, Jiaxuan You, Jure Leskovec, and C.-C. Jay Kuo. 2021. Inductive learning on commonsense knowledge graph completion. In International Joint Conference on Neural Networks, IJCNN 2021, Shenzhen, China, July 18-22, 2021, pages 1-8. IEEE."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.",
+            "Yuan Yang and Le Song. 2020. Learn to explain efficiently via neural logic inductive learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.",
+            "Zihao Zhu, Jing Yu, Yujing Wang, Yajing Sun, Yue Hu, and Qi Wu. 2020. Mucko: Multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 1097-1103. ijcai.org."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            329
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "13987",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            143
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Limitations section",
+        "bbox": [
+            149,
+            145,
+            295,
+            156
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            127,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            187,
+            349,
+            200
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Abstract section and section 1 Introduction",
+        "bbox": [
+            149,
+            230,
+            470,
+            243
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            255,
+            668,
+            272
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            299,
+            487,
+            316
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            132,
+            321,
+            410,
+            336
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            129,
+            346,
+            529,
+            363
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            149,
+            363,
+            428,
+            379
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            129,
+            388,
+            778,
+            406
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement Section",
+        "bbox": [
+            149,
+            407,
+            336,
+            420
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            129,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            149,
+            497,
+            428,
+            512
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            129,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We use publically available datasets, and the authors of the dataset have made the corresponding declaration.",
+        "bbox": [
+            149,
+            573,
+            880,
+            602
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            129,
+            615,
+            880,
+            646
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We will provide those data after the review process.",
+        "bbox": [
+            149,
+            648,
+            529,
+            663
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            129,
+            673,
+            880,
+            753
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "5.1.1 section Datasets",
+        "bbox": [
+            149,
+            755,
+            317,
+            769
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            780,
+            492,
+            797
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "5.1.3 section Implementation Details",
+        "bbox": [
+            132,
+            803,
+            406,
+            816
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            828,
+            880,
+            860
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            149,
+            860,
+            428,
+            876
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            884,
+            877,
+            908
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "13988",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            129,
+            83,
+            878,
+            115
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            149,
+            117,
+            428,
+            131
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            129,
+            142,
+            882,
+            190
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "5.3 Main Results",
+        "bbox": [
+            149,
+            192,
+            278,
+            206
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            129,
+            217,
+            882,
+            265
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Section 5.1.3 Implementation Details",
+        "bbox": [
+            149,
+            267,
+            428,
+            282
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            114,
+            292,
+            877,
+            310
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "3",
+        "bbox": [
+            134,
+            315,
+            146,
+            326
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            129,
+            338,
+            882,
+            372
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We perform simple human annotation and evaluation, there is no need of providing the full text.",
+        "bbox": [
+            149,
+            373,
+            852,
+            388
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            129,
+            398,
+            882,
+            447
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement section",
+        "bbox": [
+            149,
+            449,
+            334,
+            462
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            129,
+            474,
+            882,
+            521
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            523,
+            231,
+            538
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            129,
+            549,
+            875,
+            565
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            565,
+            349,
+            581
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            129,
+            590,
+            880,
+            623
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement section",
+        "bbox": [
+            149,
+            625,
+            334,
+            639
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "13989",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    }
+]

2023/Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering/8ad456f0-7f0d-40ee-8bd0-7b70b4a0c21c_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83cdff279fb940cf78ab8a28913f0a122ff60c4addb098333c051a0c6bf79c0c
+size 645248

2023/Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering/full.md

@@ -0,0 +1,433 @@
+# Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering
+
+Siwei Wu, Xiangqing Shen, and Rui Xia*
+
+School of Computer Science and Engineering, Nanjing University of Science and Technology, China {wusiwei, xiangqing.shen, rxia}@njust.edu.cn
+
+# Abstract
+
+The nodes in the commonsense knowledge graph (CSKG) are normally represented by free-form short text (e.g., word or phrase). Different nodes may represent the same concept. This leads to the problems of edge sparsity and node redundancy, which challenges CSKG representation and completion. On the one hand, edge sparsity limits the performance of graph representation learning; On the other hand, node redundancy makes different nodes corresponding to the same concept have inconsistent relations with other nodes. To address the two problems, we propose a new CSKG completion framework based on Contrastive Pretraining and Node Clustering (CPNC). Contrastive Pretraining constructs positive and negative head-tail node pairs on CSKG and utilizes contrastive learning to obtain better semantic node representation. Node Clustering aggregates nodes with the same concept into a latent concept, assisting the task of CSKG completion. We evaluate our CPNC approach on two CSKG completion benchmarks (CN-100K and ATOMIC), where CPNC outperforms the state-of-the-art methods. Extensive experiments demonstrate that both Contrastive Pretraining and Node Clustering can significantly improve the performance of CSKG completion. The source code of CPNC is publicly available on https://github.com/NUSTM/CPNC.
+
+# 1 Introduction
+
+Commonsense knowledge graphs (CSKG) have been widely used to build commonsense-grounded AI applications, such as question answering (Lv et al., 2020), visual question answering (Zhu et al., 2020), sentiment analysis (Li et al., 2022), dialogue system (Tu et al., 2022), etc. Commonsense knowledge graphs such as ConceptNet (Speer et al., 2017) and ATOMIC (Sap et al., 2019) provide a structured way of representing a commonsense concept, which consists of a head node, a tail node, and
+
+![](images/477f90750452fd158a58c16e870cff0691e813899b344b3a6b73d148925898f3.jpg)
+
+![](images/1fa42bc1f39770a20c77bd3edd2c2862265687be3fdb675ccadbdc2a71a03c16.jpg)  
+Figure 1: Illustration of a subgraph and the latent concept in ConceptNet. The nodes of the same color belong to the same latent concept.
+
+the relation edge. Nodes in commonsense knowledge graphs are typically represented by free-form short text (word or phrase), resulting in many different nodes representing the same concept. Figure 1 shows a subgraph of ConceptNet, where nodes in the same color indicate the same concept. E.g., "have breakfast", "take breakfast" and "eat breakfast" all express the concept of "eat breakfast". This problem also results in a large number of missing edges between nodes, as illustrated by the dashed lines in Figure 1. E.g., "eat breakfast" and "prepare food" have a "HasPrerequisite" relation, but such relation is missing between "eat breakfast" and "cook food".
+
+On the one hand, node redundancy in CSKG makes different nodes of the same concept have inconsistent relations with other nodes. In CSKG representation and completion, it would be beneficial to make use of the latent concept information behind different nodes to help learn more semantic-general representations. However, this intuition was ignored by most of the existing work in
+
+CSKG completion. On the other hand, as analyzed by Malaviya et al. (2020), edge sparsity in CSKG limits information propagation in graph neural networks. Upon graph neural networks, researchers further incorporated pre-trained language model such as BERT to enhance the semantic representation of nodes (Malaviya et al., 2020; Ju et al., 2022; Wang et al., 2021). However, fine-tuning BERT is still imperfect in representing the commonsense knowledge graph, which is made up of linked nodes represented in a free-form short text.
+
+To tackle the two issues, we propose a new CSKG completion framework based on the Encoder-Decoder architecture, which contains two core modules Contrastive Pretraining and Node Clustering (CPNC). Contrastive Pretraining is to alleviate the difficulty of node representation learning induced by edge sparsity. Through contrastive learning on positive and negative head-tail node pairs, the embedding distance between related nodes becomes closer, and that between unrelated nodes becomes farther so as to learn better node representations. Node Clustering aims to address the edge inconsistency issue caused by node redundancy. We cluster nodes with close semantic representations and take the mean vector as the latent concept representation for nodes in this cluster. To assist CSKG completion, the latent concept representation is fused with the node representation.
+
+We evaluate our CPNC framework on two CSKG completion benchmarks, i.e., CN-100K and ATOMIC. The results show that our model outperforms the state-of-the-art models for this task significantly. Ablation studies demonstrate that both Contrastive Pretraining and Node Clustering modules can significantly improve the performance of CSKG completion. Further experiments verify that our model can consistently improve as the sparsity of knowledge graphs increases: the higher the sparsity, the more improvement our model achieves.
+
+# 2 Related Work
+
+We briefly review traditional knowledge graph completion, and then pay more attention to commonsense knowledge graph completion.
+
+Traditional Knowledge Graph Completion Many knowledge graph completion methods have been proposed, which can be classified into three types: embedding-based completion methods, path-finding-based completion methods, and logical rule-based completion methods. The embedding-
+
+based method learned the relation and node embedding by shortening the distance between the head-relation pair representation and the tail node representation, which had good scalability. Moreover, convolution has been proven to be an effective operation for acquiring the head-relation representation (Dettmers et al., 2018; Shang et al., 2019) in the embedding-based method. The path-based methods used the random walk inference based algorithm to find the related path. They achieved the prediction of missing tuples by comparing the related path with the relation to be predicted (Lao and Cohen, 2010; Khot et al., 2011). The rule-based methods utilized the induction rules to simplify the path-finding process in the knowledge graphs (Ren et al., 2020; Yang and Song, 2020). We refer the reader to (Ji et al., 2022) for more details about knowledge graph completion.
+
+# Commonsense Knowledge Graph Completion
+
+The existing methods assume specific relations, dense edges, and sufficient training samples in the knowledge graph. However, the sparsity of edges and the abstract nature of relations pose challenges for directly applying these methods to the CSKG.
+
+Early methods in this field employed a strategy where, for a given head and tail, all relations were used to generate a large number of tuples. BiLSTMs and linear transformations were then utilized to score these tuples, allowing for the prediction of missing tuples in the CSKG (Saito et al., 2018; Li et al., 2016; Jastrzebski et al., 2018). Building upon this foundation, Shen et al. (2022) took a step further. They normalized the tail nodes and used RoBERTa to score these tuples. This approach resulted in the curation of Dense-ATOMIC, a CSKG with increased coverage and a richer set of multi-hop paths.
+
+these methods did not take into account structural information and may suffer from computational inefficiency during inference. Another line of studies attempted to use the translated-based approach to reduce computational consumption (Malaviya et al., 2020; Ju et al., 2022; Wang et al., 2021). Furthermore, they incorporated GCN for extracting graph representations and fine-tuned BERT in the CSKG to obtain semantic representations. This integration capitalizes on the complementary nature of these two representations. In addition, these methods have applied convolutional layers to enhance the performance of CSKG completion. This further improves the effectiveness of the com
+
+![](images/3366c71a2d3c7a89e62ef37b89eada80383cdc3ece5460df6cd96e6ffd2b2603.jpg)  
+Figure 2: Overall architecture of CPNC. The Encoder is to get node embedding. Harnessing the power of GCN and Contractive Pretraining, the Encoder transforms CSKG into rich semantic and structural representations. In detail, the lower left section utilizes the mechanism of Contrastive Pretraining. Then, we describe the node clustering process in the lower right, driven by semantic representations from Contractive Pretraining. Finally, the fusion of semantic, structure, and latent concepts in the upper right yields node embeddings. The Decoder part is to rank the nodes in candidates set by Node Embedding.
+
+pletion task. However, it was observed that simply fine-tuning BERT (Devlin et al., 2019) with pretraining tasks is hard to effectively capture the sentence-level semantic connection between two nodes. In response to this limitation, Su et al. (2022) proposed MICO, which learns distinct node representations for head nodes under varying relations, proving evidence that high-quality node representation can significantly aid CSKG completion. Moreover, different nodes in the CSKG may express the same concept. The node redundancy will bring difficulties to the commonsense inference on the CSKG (Jung et al., 2022), while the previous CSKG completion methods ignored.
+
+Compared to these methods, we address the issue of node redundancy by employing a clustering algorithm. Our work stands out by learning the semantic representations of nodes through sentence-level semantic connections and integrating latent concept information into node representations.
+
+# 3 Task Definition
+
+Given a CSKG  $G = (N, V, R)$  where  $N$  is the set of nodes,  $V$  is the set of edges and  $R$  is the set of relations. We regard each tuple  $v_{i} = (h, rel, t)$  in the CSKG as a sample, which is composed of head node  $h$ , tail node  $t$  and relation  $rel$ , where  $h, t \in N$ ,  $v_{i} \in V$ , and  $rel \in R$ . Given a query  $(h, rel)$  formed by a head node  $h$  and a relation  $rel$ , the target of the CSKG completion is to maximize the score of the tail node  $t$ . Following the
+
+previous work (Malaviya et al., 2020), for an edge  $(h,rel,t)$  existing in the CSKG, we also add an inverse edge  $(t,rel^{-1},h)$  to the graph, where  $rel^{-1}$  is the inverse relation of rel.
+
+# 4 Approach
+
+In this paper, we propose a new framework, CPNC, for CSKG completion using an Encoder-Decoder architecture (see Figure 2). The Encoder incorporates semantic, graph structure, and latent concept representation obtained from Contrastive Pretraining, Graph Convolutional Network (GCN), and Node Clustering, respectively, to acquire node representations. Given query  $(h,rel)$ , the Decoder ranks the nodes in the candidate set  $N$  and finds tail nodes for the query by the rank.
+
+# 4.1 Contrastive Pretraining
+
+The current mainstream approach for completing CSKG utilizes translation-based methods and relies on GCN to represent the graph structure of nodes. Although GCNs are effective in modeling graph structure, they have limitations in capturing graph structure information in CSKG due to sparse edges. To address this issue, these methods incorporate semantic information by fine-tuning BERT on CSKG using Masked Language Model (MLM) and Next Sentence Prediction (NSP) tasks.
+
+However, those methods are imperfect in modeling CSKGs composed of nodes presented as short text. MLM, which is a token-level task, focuses on
+
+modeling the connection between masked words and other words, making it unsuitable for achieving sentence-level node representation. Similarly, NSP, while capable of modeling the semantic connection between two nodes at the sentence level, suffers from a mismatch between the input of the pretraining phase and the CSKG completion phase. During pretraining, NSP requires a pair of head and tail nodes for prediction, but during CSKG completion, only a single node is used for sentence representation. Consequently, NSP is not well-suited for the task of learning node representations, and similar observations have been made in the field of sentence representation learning (Li et al., 2020).
+
+To obtain better node representations for the CSKG completion, we introduce Contrastive Pretraining (CP), a new method that sufficiently leverages semantic information at the sentence level. CP focuses on CSKG completion and differs from MICO (Su et al., 2022) by not incorporating relation categories. Instead, it fine-tunes BERT's node representation using contrastive learning and capture sentence-level connections between nodes.
+
+# 4.1.1 Building Contrastive Learning Samples
+
+Assuming node pairs linked in the CSKG are semantically related, we consider node pairs without a link in the CSKG as semantically unrelated.
+
+For each arbitrary edge  $(h,rel,t)$  in CSKG, we randomly sample a node  $\bar{t}$  having no linking with the head node  $h$  as its hard negative tail and construct a contrastive learning training sample  $c = (h,t,\bar{t})$ . As shown in Figure 3, "take breakfast" and "cook food" are the head and tail nodes in a edge of CSKG. We randomly select a node "find information" having no linking with "take breakfast" to construct a training sample ("take breakfast", "cook food", "find information").
+
+The previous methods primarily focused on the head nodes while constructing negative samples, neglecting the importance of capturing unrelated semantics among the tail node set. To tackle this problem, we additionally constructed a contrastive learning sample  $c_{\text{reverse}} = (t, h, \overline{h})$  for the tail node  $t$  in the edge  $(h, \text{rel}, t)$ .
+
+# 4.1.2 Multiple Negatives Ranking Loss
+
+We employ BERT for node embedding by inputting all nodes in a batch separately and applying Mean Pooling on the final layer's output. Then, we use  $x(n)$  to denote the representation of a specific node  $n$  obtained through BERT.
+
+![](images/64cdbc3434e2f846addd358851f9d42c4924505760880259037f1f93c4b2929e.jpg)  
+Figure 3: Main process of building Constructive Pretraining samples.
+
+In order to ensure that semantically related nodes are positioned closely in the embedding space, we utilize contrastive learning and employ the Multiple Negatives Ranking Loss for contrastive learning.
+
+For a given sample  $c_{i} = (h_{i}, t_{i}, \bar{t}_{i})$  in a batch, we employ the Multiple Negatives Ranking Loss. We consider  $h_{i}$  paired with  $t_{i}$  as the positive sample while treating  $h_{i}$  paired with other tail nodes  $t_{j}$  in the batch as negative samples. Additionally, we consider  $h_{i}$  paired with all hard negative tails in the batch as negative samples as well. The objective is to minimize the semantic distance between the nodes in positive samples within the batch. The formulation of the multiple negatives ranking loss is as follows:
+
+$$
+L = - \sum_ {i = 1} ^ {M} \log \frac {e ^ {D \left(h _ {i} , t _ {i}\right))}}{\sum_ {j = 1} ^ {M} e ^ {D \left(h _ {i} , t _ {j}\right)} + e ^ {D \left(h _ {i} , \bar {t} _ {j}\right)}}, \tag {1}
+$$
+
+where  $D(h,t) = (x(h)\cdot x(t)^T) / (\| x(h)\| \cdot \| x(t)\|)$  represents the cosine similarity of two nodes,  $M$  is the number of samples in the a batch.
+
+# 4.2 Encoder
+
+The Encoder aims to generate node representations for CSKG completion, as illustrated in the left part of Figure 2.
+
+The node representation comprises two components,
+
+1) the semantic representation matrix of nodes  $E_{sem}$  obtained through Contrastive Pretraining:
+
+$$
+E _ {s e m} = \text {C o n t r a s t i v e - P r e t r a i n i n g} (N), \tag {2}
+$$
+
+2) the graph structure representation matrix of nodes acquired using a Graph Convolutional Network (GCN),
+
+$$
+E _ {\text {g r a p h}} = \operatorname {G C N} (G). \tag {3}
+$$
+
+Specifically,  $e_{sem}$  and  $e_{graph}$  represent the semantic and graph representations of a particular node, respectively, which can be derived from  $E_{sem}$  and  $E_{graph}$ .
+
+# 4.3 Node Clustering
+
+On the basis of the semantic node representation  $e_{sem}$ , we use the K-means algorithm to cluster all nodes  $N$  in the graph as follows:
+
+$$
+\left\{\mathcal {C} _ {1}, \mathcal {C} _ {2}, \dots , \mathcal {C} _ {K} \right\} = \mathrm {K} - \text {m e a n s} (N), \tag {4}
+$$
+
+where  $\mathcal{C}_i$  is the  $i$ -th cluster, and  $K$  represents the number of clusters. Nodes in  $\mathcal{C}_i$  are considered to have the same latent concept. As shown in the Node Clustering part of Figure 2, "eat breakfast", "have breakfast" and "take breakfast" are in the same cluster and have the same latent concept.
+
+We define the representation of the latent concept as the mean of semantic representation of all nodes in a cluster, denoted as  $\mathrm{mean}(\mathcal{C}_k)$ , where  $\mathcal{C}_k$  contains a set of nodes in the same cluster.
+
+To acquire the final node representation, we combine the node semantic representation,  $e_{sem}$ , the graph representation,  $e_{graph}$ , and the latent concept representation,  $\mathrm{mean}(\mathcal{C}_k)$ . This fusion is performed as follows:
+
+$$
+e = \left[ e _ {s e m}; e _ {g r a p h}; \operatorname {m e a n} \left(\mathcal {C} _ {k}\right) \right] \cdot W _ {\text {e m b e d d i n g}}, \tag {5}
+$$
+
+where  $W_{embedding}$  is a weight matrix. The resulting node embeddings are considered as the node embedding space.
+
+Intuitively, Node Clustering enhances CSKG completion by leveraging information from nodes within the same latent concept.
+
+We also use a progressive masking process, following Malaviya et al. (2020), to integrate latent concept information. Initially, the latent concept representation is fully masked, and throughout the first 100 training epochs, it is gradually unmasked, improving the overall performance.
+
+# 4.4 Decoder
+
+The goal of Decoder is to rank the nodes in the candidate set, which is considered as  $N$  in our work.
+
+In order to obtain the query representation, we first use a convolutional layer to fuse the relation and head node representation:
+
+$$
+q = \operatorname {C o n v} \left(e _ {h}, e _ {r e l}\right), \tag {6}
+$$
+
+where  $e_h$  is the representation of the head nodes in node embedding space obtained from Eq. 5,  $e_{rel}$  is the representation of relation rel and Conv is a convolution operation.
+
+Then, we predict a 0-1 distribution vector, based on the representation of the query  $q$ , indicating the likelihood of each node in the candidate set  $N$  becoming the tail node:
+
+$$
+p _ {t} = \delta \left(q W _ {\text {c o n v}} E _ {N}\right), \tag {7}
+$$
+
+where  $W_{conv}$  is a weight matrix,  $E_N$  is the representation matrix of candidate nodes in node embedding space obtained from Eq. 5, and  $\delta$  is a sigmoid function.
+
+# 5 Experiments
+
+# 5.1 Experimental Setup
+
+# 5.1.1 Datasets
+
+We evaluate our CPNC framework on two CSKG completion benchmarks, i.e., CN-100K (Speer et al., 2017) and ATOMIC (Sap et al., 2019).
+
+CN-100K is a dataset that encompasses general commonsense knowledge. This version contains 36 relation types and the Open Mind Common Sense (OMCS) entries from ConceptNet (Speer and Havasi, 2013). The average length of the nodes in CN-100K is 2.85 words. Following Malaviya et al. (2020)'s split, the training set contains 10,000 tuples, and the validation set and test set both contain 1200 tuples.
+
+ATOMIC is an atlas of everyday commonsense reasoning and primarily focuses on event-level commonsense knowledge in the form of if-then relations. It comprises 9 relation types, with an average of 4.40 words per node. We split ATOMIC following Malaviya et al. (2020)'s work, where the training set consists of 610,536 tuples, while the validation and test sets contain 87,700 tuples and 87,701 tuples, respectively.
+
+<table><tr><td colspan="2">Methods</td><td colspan="4">CN-100K</td><td colspan="4">ATOMIC</td></tr><tr><td>Type</td><td>Model</td><td>MRR</td><td>HITs@1</td><td>@3</td><td>@10</td><td>MRR</td><td>HITs@1</td><td>@3</td><td>@10</td></tr><tr><td rowspan="4">KG-adapted</td><td>DistMult</td><td>8.97</td><td>4.51</td><td>9.76</td><td>17.44</td><td>12.39</td><td>9.24</td><td>15.18</td><td>18.30</td></tr><tr><td>ComplEx</td><td>11.40</td><td>7.42</td><td>12.45</td><td>19.01</td><td>14.24</td><td>13.27</td><td>14.13</td><td>15.96</td></tr><tr><td>ConvE</td><td>20.88</td><td>13.97</td><td>22.91</td><td>34.02</td><td>10.07</td><td>8.24</td><td>10.29</td><td>13.37</td></tr><tr><td>ConvTransE</td><td>18.68</td><td>7.87</td><td>23.87</td><td>38.95</td><td>12.94</td><td>12.92</td><td>12.95</td><td>12.98</td></tr><tr><td rowspan="2">Generation-based</td><td>COMeT-Normalized</td><td>6.07</td><td>0.08</td><td>2.92</td><td>21.17</td><td>3.36</td><td>0.00</td><td>2.15</td><td>15.75</td></tr><tr><td>COMeT-Total</td><td>6.21</td><td>0.00</td><td>0.00</td><td>24.00</td><td>4.91</td><td>0.00</td><td>2.40</td><td>21.60</td></tr><tr><td rowspan="3">CSKG-dedicated</td><td>RGAT</td><td>43.97</td><td>30.75</td><td>51.54</td><td>69.34</td><td>-</td><td>-</td><td>-</td><td>-</td></tr><tr><td>SGBC</td><td>49.12*</td><td>37.71*</td><td>56.67*</td><td>71.29*</td><td>10.25*</td><td>8.72*</td><td>10.54*</td><td>13.26*</td></tr><tr><td>Inductive</td><td>56.92*</td><td>45.54*</td><td>63.38*</td><td>78.63*</td><td>13.19*</td><td>10.26*</td><td>13.61*</td><td>18.83*</td></tr><tr><td rowspan="2">Ours</td><td>CPNC-S</td><td>54.52</td><td>45.33</td><td>61.46</td><td>75.92</td><td>13.14</td><td>10.11</td><td>13.75</td><td>18.80</td></tr><tr><td>CPNC-I</td><td>59.00</td><td>48.29</td><td>65.04</td><td>79.13</td><td>14.38</td><td>10.53</td><td>15.22</td><td>21.79</td></tr></table>
+
+Table 1: CSKG completion results on CN-100K and ATOMIC. The result of HITS@1 was not reported in Wang et al. (2021). To fairly compare our method with the previous method in the same setting, we rerun the code of InductivE and SGBC on ATOMIC and CN-100K and mark the results of rerun experiments with *. Those results are a little lower than that in the original paper. For KG-adapted and Generation methods, we reuse the results reported in Malaviya et al. (2020).
+
+# 5.1.2 Evaluation Metric
+
+We evaluate the performance of our method using MRR and HITS, following previous CSKG completion methods (Wang et al., 2021; Malaviya et al., 2020; Ju et al., 2022). The results are reported by averaging over both forward tuples  $(h, rel, t)$  and inverse tuples  $(t, rel^{-1}, h)$ . Moreover, because the nodes in CSKG are represented in free-form text, it is possible for nodes other than the golden tail nodes to be considered reasonable tail nodes. To address this, we conduct a human evaluation to assess the predictions made by our models.
+
+# 5.1.3 Implementation Details
+
+To perform Contrastive Pretraining, we use the contrastive learning framework provided in https://github.com/UKPLab/sentence-transformers. Our approach employs BERT-large as the base model, which contains 340M parameters. During training, we utilized a batch size of 128 and conducted training for 3 epochs on the CN-100K and ATOMIC datasets. We choose the Adam optimizer for optimization, setting the learning rate to 1e-4 for BERT-large and 5e-5 for the MLP. For the remaining hyperparameters, we used the default values provided by the framework.
+
+For the CPNC-I and CPNC-S model, we use experimental settings proposed by Malaviya et al. (2020) and Wang et al. (2021), respectively. Both models are trained for a minimum of 200 epochs using BERT-large, which consists of 340M parameters, to encode semantic representations. During training, we evaluate the MRR on the development set every 10 epochs for CN-100K and ATOMIC. Training continues until no further improvement in
+
+MRR is observed. We select the model checkpoint that achieves the highest MRR on the development set for testing.
+
+# 5.2 Compared Systems
+
+We compare our approach with nine baseline systems across three categories.
+
+# 5.2.1 KG-adapted Methods
+
+We adapt classic knowledge graph completion methods for CSKG completion. DistMult (Yang et al., 2015) employed a bi-linear product to calculate score of a tuple; ComplE (Trouillon et al., 2016) utilized imaginary number representation to effectively handle a large number of relations in the knowledge graph; ConvE (Dettmers et al., 2018) fused the representation of the source node and the relation through a 2D convolution layer to obtain the representation of query (h, rel); ConvTransE extended ConvE by incorporating the translational properties of TransE.
+
+# 5.2.2 Generation-based Methods
+
+COMeT (Bosselut et al., 2019) is a Transformer-based knowledge generation model. Following Malaviya et al. (2020), we adapt COMeT for the CSKG completion and only evaluate in the forward direction. Additionally, we also use their modifications to the ranking method used in COMeT. Specifically, COMET-Normalized and COMET-Total use the normalized negative log-likelihood scores and total log-likelihood scores, respectively, to rank nodes in the candidate set.
+
+# 5.2.3 CSKG-dedicated Methods
+
+SIM+GCN+BERT+ConvTransE (SGBC) densified the CSKG by connecting the synthetic edges between similar semantic nodes, improving the graph structure representation (Malaviya et al., 2020). Moreover, they used fine-tuned BERT to encode the semantic information of the nodes.
+
+Inductive (Wang et al., 2021) is proposed to enhance the unseen entity representation with neighboring structural information by densifying Graph. Relational graph attention networks (RGAT) are proposed weighted the importance of neighbor nodes of each node to obtain a better node representation.
+
+# 5.2.4 Our CPNC Methods
+
+By incorporating CPNC with two CSKG-dedicated methods (SGBC and InductivE), we obtain two models, i.e., CPNC-S and CPNC-I.
+
+# 5.3 Main Results
+
+In Table 1, we report experimental results of KG-adapted methods, Generation methods, CSKG-dedicated methods and our methods on CN-100K and ATOMIC.
+
+On CN-100K, both KG-adapted methods and Generation methods exhibit unsatisfactory performance, with MRR values below  $21\%$  and HITS@1 values below  $14\%$ . These results demonstrate that directly applying KG-adapted and Generation methods to CSKG completion is ineffective. In contrast, CSKG-dedicated methods achieve significantly better results on CN-100K, with MRR values above  $43\%$  and HITS@1 values above  $30\%$ . These metrics are twice as high as those obtained by KG-adapted and Generation methods, highlighting the advantage of CSKG-dedicated methods. Comparing our proposed CPNC-S and CPNC-I models with mainstream CSKG-dedicated methods (SGBC and Inductive), we observe performance improvements of  $5.40\%$  and  $2.08\%$  on MRR, respectively. Notably, our CPNC-I model sets a new state-of-the-art result on CN-100K.
+
+On ATOMIC, we draw a similar conclusion. The KG-adapted methods exhibit poor performance, while CSKG-dedicated methods achieve better results. In addition, compared with the SGBC and Inductive models, our CPNC-S and CPNC-I models outperform the SGBC and Inductive models by  $2.89\%$  and  $1.19\%$  on MRR, respectively. And the CPNC-I model achieves the state-of-the-art result on ATOMIC.
+
+We conduct Paired t-Test and the result proves that the improvement of CPNC is significant.
+
+# 5.4 Ablation Study
+
+<table><tr><td></td><td></td><td>MRR</td><td>HITS@10</td></tr><tr><td rowspan="8">CN-100K</td><td>CPNC-S</td><td>54.52</td><td>75.92</td></tr><tr><td>-w/o CP</td><td>51.90</td><td>76.08</td></tr><tr><td>-w/o NC</td><td>49.74</td><td>71.38</td></tr><tr><td>SGBC</td><td>49.12</td><td>71.29</td></tr><tr><td>CPNC-I</td><td>59.00</td><td>79.13</td></tr><tr><td>-w/o CP</td><td>57.16</td><td>74.90</td></tr><tr><td>-w/o NC</td><td>58.30</td><td>78.75</td></tr><tr><td>InductivE</td><td>56.92</td><td>78.63</td></tr><tr><td rowspan="8">ATOMIC</td><td>CPNC-S</td><td>13.14</td><td>18.80</td></tr><tr><td>-w/o CP</td><td>12.77</td><td>17.15</td></tr><tr><td>-w/o NC</td><td>12.46</td><td>17.72</td></tr><tr><td>SGBC</td><td>10.25</td><td>13.26</td></tr><tr><td>CPNC-I</td><td>14.38</td><td>21.79</td></tr><tr><td>-w/o CP</td><td>13.22</td><td>19.16</td></tr><tr><td>-w/o NC</td><td>14.21</td><td>21.38</td></tr><tr><td>InductivE</td><td>13.19</td><td>18.83</td></tr></table>
+
+Table 2: MRR and HITS@10 of removing CP and NC from CPNC on CN-100K and ATOMIC.
+
+To demonstrate the effectiveness of Contrastive Pretraining (CP) and Node Clustering (NC) on CSKG completion, we perform the ablation study where we removed CP and NC. The results are presented in Table 2.
+
+When NC was removed, we can observe that the performance of both CPNC-S and CPNC-I decreases on CN-100K and ATOMIC. This indicates that NC's latent concept representation effectively assists CSKG completion. It is worth noting that CP can still bring improvement compared with the previous method, showing that the semantic representation provided by CP is crucial for CSKG completion.
+
+CPNC, which combines NC and CP, consistently yields the highest result. By replacing CP with finetuned BERT (Malaviya et al., 2020), the performance of both CPNC-S and CPNC-I drops. However, they still achieve comparable results with the previous CSKG-dedicated methods. This demonstrates that the CP requires a high-quality semantic representation of nodes to acquire good latent concept representation.
+
+# 5.5 Human Evaluation
+
+For a given query  $(h,rel)$ , additional nodes besides the golden tail node could also become reasonable tail nodes. For example, given a query “(do house-work, Causes)”, besides golden tail node “clean the house”, “house get clean” and “clean the room” are also reasonable tail nodes. However, automated
+
+metrics fail to cover these nodes. To address this issue, we randomly select 200 queries from the test set of the CN-100K and ATOMIC. For each query, we use our methods and current mainstream methods to rank the nodes in the candidate set, We manually calculate the number of reasonable tail nodes in the top 10 candidates. The results are shown in Table 3.
+
+<table><tr><td></td><td>ATOMIC</td><td>CN-100K</td></tr><tr><td>CPNC-I</td><td>70.22</td><td>83.20</td></tr><tr><td>Inductive</td><td>64.45</td><td>76.40</td></tr><tr><td>CPNC-S</td><td>60.23</td><td>69.40</td></tr><tr><td>SGBC</td><td>62.11</td><td>65.40</td></tr></table>
+
+Table 3: Average percentage of reasonable tuples by human evaluation results.
+
+Observing human evaluation results, the CPNC model has an excellent performance in CSKG completion. Besides, our method is significantly higher than the current mainstream methods under the human evaluation results, which further validates the effectiveness of our approach.
+
+# 5.6 Model Adaptability
+
+The CPNC can be effectively applied to various methods. In our experiment, we incorporate CPNC into SGBC and InductivE, resulting in significant improvements.
+
+The SGBC method initializes the input of GCN randomly and combines graph structural representation acquired by GCN and semantic representation obtained by BERT as the node representation; InductiveE initializes the input of GCN with BERT and uses the output of GCN as the node representation. Despite the differences in their structures, both methods benefit from the application of CP and NC.
+
+On CN-100K, CPNC-S and CPNC-I bring improvements of  $5.40\%$  and  $2.08\%$  in MRR, respectively; On ATOMIC, CPNC-S and CPNC-I yield MRR improvements of  $2.89\%$  and  $1.19\%$ , respectively. This result demonstrates the adaptability of our method.
+
+# 5.7 Discussion on the Effect of Relieving Sparsity
+
+To evaluate the effectiveness of CPNC in mitigating the impact of edge sparsity in the CSKG, we conducted a comparative study between CPNC-S and SGBC on CN-100K with varying sparsity levels. The removal of edges in the CSKG results in
+
+increased sparsity, with a higher number of deleted edges corresponding to a higher sparsity degree.
+
+![](images/bd60b82d789d20f700afa3335a1117e68142408a574f2e866c234edb6a749002.jpg)  
+Figure 4: CSKG completion results of SGBC and CPNC-S on CN-100K with various degrees of sparsity.
+
+As shown in Figure 4, when the sparsity degree increases, CSKG completion performance deteriorates. CPNC-S consistently outperforms SGBC in terms of MRR on CN-100K across different sparsity levels, indicating CPNC's ability to mitigate edge sparsity in the CSKG. CPNC-S achieves a  $12.48\%$  and  $3.95\%$  increase in MRR compared to SGBC at sparsity extents of  $25\%$  and  $50\%$  respectively. These results demonstrate that CPNC is more effective in alleviating edge sparsity in the CSKG, with greater improvements observed for sparser graphs.
+
+# 5.8 Discussion on the Number of Latent Concepts
+
+![](images/38e969457fec383c734630948f379dee9ade0538ffd83ca7cec056cbce55514e.jpg)  
+Figure 5: CSKG completion performance with different cluster numbers on CN-100K and ATOMIC.
+
+Different numbers of clusters can lead to varying levels of granularity in latent concepts, which in turn can impact the performance of CSKG completion. In order to investigate this, we conducted experiments with the different numbers of clusters
+
+on CN-100K and ATOMIC datasets and present the results in Figure 5.
+
+On CN-100K, we observed an overall increasing trend in completion results as the number of clusters increased. However, the rate of improvement tends to diminish after reaching 900 clusters. The maximum MRR achieved when using the number of clusters is 1200.
+
+On ATOMIC, the MRR initially increases with an increasing number of clusters but starts to decline once the number of clusters reaches 900.
+
+Based on our main results, we selected 1200 clusters for CN-100K and 900 clusters for ATOMIC, allowing us to effectively capture latent concepts through cluster nodes.
+
+# 5.9 Discussing on Contrastive Learning Loss
+
+In order to obtain a better semantic representation of nodes, we experiment with five kinds of contrastive learning methods and observe their performance in CSKG completion. The result is shown in Table 4.
+
+<table><tr><td>Methods</td><td>MRR</td></tr><tr><td>Contrastive Loss</td><td>40.03</td></tr><tr><td>MTriplet Loss</td><td>44.36</td></tr><tr><td>MICO</td><td>45.93</td></tr><tr><td>Batch Semi Hard Triplet Loss</td><td>27.92</td></tr><tr><td>Multiple Negatives Ranking Loss</td><td>49.74</td></tr></table>
+
+Table 4: CSKG completion performance of CPNC-S using different Contrastive Learning Loss on CN-100K.
+
+In order to compare those contrastive learning methods, we use the node embedding obtained by those methods to complete CSKG without node clustering. Each contrastive learning method yields distinct results, and our main result selects Multiple Negatives Ranking Loss due to its highest MRR.
+
+# 6 Conclusion
+
+In this work, we propose a new CSKG completion framework CPNC to address issues arising from node redundancy and edge sparsity. CPNC obtains better semantic node information through Contrastive Pretraining, which alleviates the problems caused by edge sparsity. CPNC also utilizes the latent concept representation acquired through Node Clustering to alleviate the problem caused by node redundancy. On CN-100K and ATOMIC, experimental results and extensive analysis demonstrate the effectiveness of Contrastive Pretraining and Node Clustering.
+
+# Limitations
+
+Due to the limitation of time and resources, in this work, we select a relatively small number of clusters during the clustering process, which results in coarse-grained clustering. Fine-grained clustering can provide a better latent concept but will also lead to increased computational resources and time consumption. We will attempt to trade-off between the cost and the granularity of clustering in future work to further explore the impact of the latent concept on CSKG completion. Besides, our Node Clustering module is not integrated in an end-to-end manner in our work; We will consider using the topic neural network to construct an end-to-end model.
+
+# Ethics Statement
+
+We would like to thank Malaviya et al. (2020) and Wang et al. (2021) for their code on commonsense knowledge graph completion. Their models are licensed under MIT, which allows copying, modifying, merging, publishing, and distributing of the material.
+
+In the stage of human evaluation, we employed three graduate students experienced in natural language processing for human evaluation. We paid the graduate students about \(8 per hour, well above the local average wage, and engaged in constructive discussions if they had concerns about the process.
+
+# Acknowledgements
+
+This work was supported by the Natural Science Foundation of China (No. 62076133), and the Natural Science Foundation of Jiangsu Province for Distinguished Young Scholars (No. BK20200018).
+
+# References
+
+Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. COMET: commonsense transformers for automatic knowledge graph construction. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 4762-4779. Association for Computational Linguistics.  
+Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence,
+
+(AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 1811-1818. AAAI Press.  
+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, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4171-4186. Association for Computational Linguistics.  
+Stanislaw Jastrzebski, Dzmitry Bahdanau, Seyedarian Hosseini, Michael Noukhovitch, Yoshua Bengio, and Jackie Chi Kit Cheung. 2018. Commonsense mining as knowledge base completion? A study on the impact of novelty. CoRR, abs/1804.09259.  
+Shaoxiong Ji, Shirui Pan, Erik Cambria, Pekka Martinen, and Philip S. Yu. 2022. A survey on knowledge graphs: Representation, acquisition, and applications. IEEE Trans. Neural Networks Learn. Syst., 33(2):494-514.  
+Jinhao Ju, Deqing Yang, and Jingping Liu. 2022. Commonsense knowledge base completion with relational graph attention network and pre-trained language model. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022, pages 4104-4108. ACM.  
+Yong-Ho Jung, Jun-Hyung Park, Joon-Young Choi, Mingyu Lee, Junho Kim, Kang-Min Kim, and SangKeun Lee. 2022. Learning from missing relations: Contrastive learning with commonsense knowledge graphs for commonsense inference. In *Findings of the Association for Computational Linguistics: ACL* 2022, Dublin, Ireland, May 22-27, 2022, pages 1514-1523. Association for Computational Linguistics.  
+Tushar Khot, Sriraam Natarajan, Kristian Kersting, and Jude W. Shavlik. 2011. Learning markov logic networks via functional gradient boosting. In 11th IEEE International Conference on Data Mining, ICDM 2011, Vancouver, BC, Canada, December 11-14, 2011, pages 320-329. IEEE Computer Society.  
+Ni Lao and William W. Cohen. 2010. Relational retrieval using a combination of path-constrained random walks. Mach. Learn., 81(1):53-67.  
+Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the sentence embeddings from pre-trained language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 9119-9130. Association for Computational Linguistics.
+
+Jiangnan Li, Fandong Meng, Zheng Lin, Rui Liu, Peng Fu, Yanan Cao, Weiping Wang, and Jie Zhou. 2022. Neutral utterances are also causes: Enhancing conversational causal emotion entailment with social commonsense knowledge. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vienna, Austria, 23-29 July 2022, pages 4209-4215. ijcai.org.  
+Xiang Li, Aynaz Taheri, Lifu Tu, and Kevin Gimpel. 2016. Commonsense knowledge base completion. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics.  
+Shangwen Lv, Daya Guo, Jingjing Xu, Duyu Tang, Nan Duan, Ming Gong, Linjun Shou, Daxin Jiang, Guihong Cao, and Songlin Hu. 2020. Graph-based reasoning over heterogeneous external knowledge for commonsense question answering. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 8449-8456. AAAI Press.  
+Chaitanya Malaviya, Chandra Bhagavatula, Antoine Bosselut, and Yejin Choi. 2020. Commonsense knowledge base completion with structural and semantic context. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 2925-2933. AAAI Press.  
+Hongyu Ren, Weihua Hu, and Jure Leskovec. 2020. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.  
+Itsumi Saito, Kyosuke Nishida, Hisako Asano, and Junji Tomita. 2018. Commonsense knowledge base completion and generation. In Proceedings of the 22nd Conference on Computational Natural Language Learning, CoNLL 2018, Brussels, Belgium, October 31 - November 1, 2018, pages 141-150. Association for Computational Linguistics.  
+Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019. ATOMIC: an atlas of machine commonsense for if-then reasoning. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA,
+
+January 27 - February 1, 2019, pages 3027-3035. AAAI Press.  
+Chao Shang, Yun Tang, Jing Huang, Jinbo Bi, Xiaodong He, and Bowen Zhou. 2019. End-to-end structure-aware convolutional networks for knowledge base completion. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 3060-3067. AAAI Press.  
+Xiangqing Shen, Siwei Wu, and Rui Xia. 2022. Dense-atomic: Construction of densely-connected and multihop commonsense knowledge graph upon ATOMIC. CoRR, abs/2210.07621.  
+Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, February 4-9, 2017, San Francisco, California, USA, pages 4444-4451. AAAI Press.  
+Robyn Speer and Catherine Havasi. 2013. Conceptnet 5: A large semantic network for relational knowledge. In Iryna Gurevych and Jungi Kim, editors, The People's Web Meets NLP, Collaboratively Constructed Language Resources, Theory and Applications of Natural Language Processing, pages 161-176. Springer.  
+Ying Su, Zihao Wang, Tianqing Fang, Hongming Zhang, Yangqiu Song, and Tong Zhang. 2022. MICO: A multi-alternative contrastive learning framework for commonsense knowledge representation. CoRR, abs/2210.07570.  
+Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, volume 48 of JMLR Workshop and Conference Proceedings, pages 2071-2080. JMLR.org.  
+Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. MISC: A mixed strategy-aware model integrating COMET for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 308-319. Association for Computational Linguistics.  
+Bin Wang, Guangtao Wang, Jing Huang, Jiaxuan You, Jure Leskovec, and C.-C. Jay Kuo. 2021. Inductive learning on commonsense knowledge graph completion. In International Joint Conference on Neural Networks, IJCNN 2021, Shenzhen, China, July 18-22, 2021, pages 1-8. IEEE.
+
+Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings.  
+Yuan Yang and Le Song. 2020. Learn to explain efficiently via neural logic inductive learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.  
+Zihao Zhu, Jing Yu, Yujing Wang, Yajing Sun, Yue Hu, and Qi Wu. 2020. Mucko: Multi-layer cross-modal knowledge reasoning for fact-based visual question answering. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI 2020, pages 1097-1103. ijcai.org.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Limitations section
+
+A2. Did you discuss any potential risks of your work?
+
+Not applicable. Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Abstract section and section 1 Introduction
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Section 5.1.3 Implementation Details
+
+B1. Did you cite the creators of artifacts you used?
+
+Section 5.1.3 Implementation Details
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Ethics Statement Section
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Section 5.1.3 Implementation Details
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+We use publically available datasets, and the authors of the dataset have made the corresponding declaration.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+We will provide those data after the review process.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+5.1.1 section Datasets
+
+C Did you run computational experiments?
+
+5.1.3 section Implementation Details
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Section 5.1.3 Implementation Details
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Section 5.1.3 Implementation Details
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+5.3 Main Results
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Section 5.1.3 Implementation Details
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+3
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+We perform simple human annotation and evaluation, there is no need of providing the full text.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Ethics Statement section
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Ethics Statement section

2023/Commonsense Knowledge Graph Completion Via Contrastive Pretraining and Node Clustering/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:02a4ee15955e91dd34ae6c358017cafa7a23bae442fda61cc58ff63c78ffb33c
+size 435678

2023/Commonsense Knowledge Transfer for Pre-trained Language Models/f96ae55a-934d-4cd1-b9c6-f4ed8849b12f_content_list.json

@@ -0,0 +1,1784 @@
+[
+    {
+        "type": "text",
+        "text": "Commonsense Knowledge Transfer for Pre-trained Language Models",
+        "text_level": 1,
+        "bbox": [
+            139,
+            89,
+            860,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Wangchunshu Zhou*¹ Ronan Le Bras² Yejin Choi²³",
+        "bbox": [
+            231,
+            129,
+            771,
+            148
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ ETH Zurich",
+        "bbox": [
+            447,
+            148,
+            557,
+            162
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2Allen Institute for AI",
+        "bbox": [
+            400,
+            164,
+            583,
+            179
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>3</sup>Paul G. Allen School of Computer Science & Engineering, University of Washington wangchunshu.zhou@inf.ethz.ch",
+        "bbox": [
+            151,
+            180,
+            850,
+            215
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite serving as the foundation models for a wide range of NLP benchmarks, pre-trained language models have shown limited capabilities of acquiring implicit commonsense knowledge from self-supervision alone, compared to learning linguistic and factual knowledge that appear more explicitly in the surface patterns in text.",
+        "bbox": [
+            141,
+            279,
+            460,
+            394
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In this work, we introduce commonsense knowledge transfer, a framework to transfer the commonsense knowledge stored in a neural commonsense knowledge model to a general-purpose pre-trained language model. It first exploits general texts to form queries for extracting commonsense knowledge from the neural commonsense knowledge model and then refines the language model with two self-supervised objectives: commonsense mask infilling and commonsense relation prediction, which align human language with the underlying commonsense knowledge.",
+        "bbox": [
+            139,
+            398,
+            460,
+            583
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Empirical results show that our approach consistently improves the model's performance on downstream tasks that require commonsense reasoning. Moreover, we find that the improvement is more significant in the few-shot setting. This suggests that our approach helps language models better transfer to downstream tasks without extensive supervision by injecting commonsense knowledge into their parameters.",
+        "bbox": [
+            141,
+            588,
+            460,
+            731
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            743,
+            260,
+            758
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent advances in pre-trained language models have transformed the landscape of natural language processing. Self-supervised pre-training objectives including masked language modeling (Devlin et al., 2019) and masked span infilling (Lewis et al., 2020) enable pre-trained models to acquire linguistic (Hewitt and Manning, 2019; Manning et al., 2020) and",
+        "bbox": [
+            112,
+            769,
+            490,
+            881
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/1832410ace14b469811e8e90ee7af2805ccfeacefbfaf920aef3a7e20737e420.jpg",
+        "image_caption": [
+            "Figure 1: Illustration of the commonsense knowledge transfer framework. We first extract commonsense knowledge related to sentences in general text corpus from a neural commonsense knowledge model. We then use natural texts and the extracted commonsense knowledge to form self-supervised training data to refine a pre-trained model with commonsense knowledge."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            250,
+            878,
+            411
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "factual knowledge (Petroni et al., 2019) by modeling the distribution of naturally occurring texts. However, most of these objectives are limited to exploiting the surface form of human language, and the lack of grounded supervision calls into question how well these representations can ever capture meaning (Bender and Koller, 2020), not to mention the underlying commonsense knowledge which is often reasoned implicitly and does not appear in the surface form of human language (Merrill et al., 2021; Zhou et al., 2020a; Hwang et al., 2021). On the other hand, commonsense reasoning is important for building generalizable models because it enables the model to reason about a great number of events, causes, and effects, while observing only a small fraction of them. The ineffectiveness of self-supervised language model pre-training on acquiring commonsense knowledge makes them require a relatively large number of labeled examples to succeed in a downstream task and prune to overfit task-specific correlations (Tu et al., 2020).",
+        "bbox": [
+            507,
+            546,
+            884,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Therefore, equipping pre-trained language models with commonsense reasoning ability has at",
+        "bbox": [
+            507,
+            887,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Work done while interning at the Allen Institute for AI",
+        "bbox": [
+            147,
+            903,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "5946",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 5946-5960",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            959,
+            700,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "tracted much attention. To this end, two distinct lines of research focus on improving commonsense reasoning ability of pre-trained language models. The first one focuses on incorporating external commonsense knowledge graph for commonsense reasoning (Lin et al., 2019; Liu et al., 2021; Cui and Chen, 2021) while the other attempts to inject commonsense knowledge into the parameters of pretrained models (Li et al., 2019; Zhou et al., 2021; Klein and Nabi, 2021). In this work we focus on the second type of method because it alleviates the need for external knowledge bases for training and inference on downstream tasks, thus simpler, more efficient, and not limited by the coverage issue of external knowledge bases.",
+        "bbox": [
+            115,
+            84,
+            485,
+            324
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Prior work injects commonsense knowledge into pre-trained models either on symbolic commonsense knowledge graphs with manually defined rules (Li et al., 2019) or masked language modeling (Hosseini et al., 2021) or on general text corpus with concept-centric self-supervised objectives (Zhou et al., 2021). The former method is limited by the coverage of knowledge graphs and human-written rules. It also fails to make use of large-scale diverse natural text corpus. Therefore, the training is limited to short and synthetic commonsense tuples, which affects its generalization ability on diverse downstream tasks. The latter method, however, only captures surface-level order relations between concepts and fails to learn commonsense relations between concepts such as cause, effect, intent, requirement, etc., which are crucial for commonsense reasoning but often implicitly reasoned, thus do not appear in the surface form of natural language.",
+        "bbox": [
+            115,
+            330,
+            485,
+            651
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we propose commonsense knowledge transfer, an alternative framework to refine a general purpose pre-trained model's commonsense reasoning ability. In contrast to previous work, it aims to transfer the commonsense knowledge stored in a neural commonsense knowledge model (e.g., COMET (Bosselut et al., 2019)) to a general purpose pre-trained model on large scale general text corpus. In this way, our approach combines the best of both worlds from prior art: the dense and informative commonsense knowledge from commonsense knowledge graphs and the accessibility of large-scale diverse general corpus.",
+        "bbox": [
+            115,
+            657,
+            485,
+            865
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Commonsense knowledge transfer is conceptually related to knowledge distillation (KD) (Hinton et al., 2015) since they both aim to trans",
+        "bbox": [
+            115,
+            871,
+            485,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "fer knowledge from a knowledge-rich model to another model that lacks it. However, different from conventional KD, in commonsense knowledge transfer, the source model (i.e., neural commonsense model) and the target model (i.e., pretrained model) are heterogeneous. Moreover, instead of simply mimicking the teacher model, commonsense knowledge transfer requires the target model to learn specialized knowledge from the source model while retaining its own capability. This poses unique challenges since the knowledge transfer can not be accomplished by simply matching the logits or feature distribution between the student and the teacher. To this end, we propose to first extract commonsense knowledge in textual form from the source model and then exploit the extracted knowledge to form self-supervised training data for the target model. As illustrated in Figure 1, commonsense knowledge transfer first exploits general texts to form queries for retrieving commonsense knowledge from the neural commonsense knowledge model. Then it refines a pretrained model with two self-supervised objectives that align the surface form of human language with its underlying commonsense inference: commonsense text infilling and commonsense relation prediction. The former objective concatenates natural text with its commonsense inference to form an input example, masks certain spans in it, and trains the model to reconstruct the original input. The latter method instead trains the model to distinguish valid commonsense inference from carefully constructed spurious commonsense inference given the original text and commonsense relation. Refining a pre-trained model by multi-tasking on both generation (former) and understanding (latter) tasks enables the model to better adapt to different kinds of downstream tasks.",
+        "bbox": [
+            510,
+            84,
+            882,
+            694
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We refine T5 (Raffel et al., 2020) with commonsense knowledge transfer and fine-tune the resulting model downstream tasks requiring commonsense reasoning ability in both the fully supervised setting and few-shot settings where only a percentage of labeled examples are available. Experimental results show substantial improvements in downstream tasks requiring commonsense reasoning, especially in the few-shot setting, demonstrating the effectiveness of our approach.",
+        "bbox": [
+            512,
+            758,
+            882,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "5947",
+        "bbox": [
+            482,
+            928,
+            517,
+            939
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Methodology",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            263,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Our proposed commonsense knowledge transfer framework consists of a neural commonsense knowledge model (e.g., COMET) and a pre-trained model (e.g., T5). The goal of commonsense knowledge transfer is to transfer the commonsense knowledge from the neural commonsense knowledge model (i.e., source model) to the pre-trained model (i.e., target model) so that it can generalize better to downstream tasks requiring commonsense reasoning ability.",
+        "bbox": [
+            112,
+            111,
+            489,
+            271
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Compared to conventional knowledge transfer methods such as knowledge distillation, commonsense knowledge transfer faces a unique challenge: the source model and the target model are heterogeneous because they are trained on different data with different objectives. As such, we can not simply feed a batch of data to both of the models and train the target model to match the source model's logits or feature distribution. To alleviate this problem, we propose a two-stage knowledge transfer scheme as illustrated in Figure 1. To be specific, we first use natural texts to form queries for retrieving commonsense knowledge (in text form) from the neural commonsense knowledge model. We then construct training data with two novel commonsense-related self-supervised objectives based on the retrieved commonsense knowledge and the corresponding natural text. Finally, we train the target model on the constructed training data to inject commonsense knowledge retrieved from the source model. We describe our method to extract commonsense knowledge from a neural commonsense knowledge model and the proposed commonsense-related self-supervised objectives in detail in this section.",
+        "bbox": [
+            115,
+            273,
+            489,
+            674
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1 Commonsense Knowledge Extraction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            688,
+            453,
+            703
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We first describe the source model, i.e., neural commonsense knowledge model, in the commonsense knowledge transfer framework. It is a transformer (Vaswani et al., 2017) language model trained on commonsense knowledge graphs like ATOMIC (Sap et al., 2019a) and ConceptNet (Speer et al., 2017) with the objective of predicting the object (i.e., commonsense inference) with the subject (i.e., natural text) and relation as input. For example, given a commonsense tuple (s=\"take a nap\", r=Causes, o=\"have energy\"), the neural commonsense knowledge model is trained to generate  $o$  given  $s$  and  $r$  as inputs. After training, it can gen",
+        "bbox": [
+            112,
+            709,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "erate accurate, representative knowledge for new, unseen entities and events.",
+        "bbox": [
+            507,
+            84,
+            882,
+            115
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To extract commonsense knowledge stored in a neural commonsense knowledge model, we use a natural sentence as the subject  $s$  (e.g., he wants to cook a meal) and concatenate it with a randomly selected commonsense relation  $r$  (e.g., xNeed) from a pre-defined set to form a prompt (e.g., he wants to cook a meal xNeed). We then feed the prompt to the neural commonsense knowledge model and use it to generate a commonsense inference (e.g., to buy ingredients). In this way, the commonsense knowledge generation process resembles the way in which the neural commonsense knowledge model is trained. As such, we can get commonsense inferences of relatively high qualities.",
+        "bbox": [
+            507,
+            118,
+            884,
+            343
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Using a neural commonsense knowledge model as a knowledge source has two advantages. On one hand, compared to the previous method (Li et al., 2019) using a symbolic commonsense knowledge graph, a neural commonsense knowledge model can generalize to unseen subjects, thus enabling us to refine the target pre-trained model on large-scale natural text corpus together with its commonsense inferences. As such, the resulting model can better adapt to downstream tasks which are formulated in diverse natural texts. On the other hand, compared to another method (Zhou et al., 2021) that only uses plain text and is thus limited to the surface form of naturally occurring text, the use of a neural commonsense knowledge model provides much denser commonsense knowledge including a diverse set of commonsense relations between natural texts and the underlying commonsense knowledge.",
+        "bbox": [
+            507,
+            344,
+            884,
+            634
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Commonsense Knowledge Injection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            651,
+            836,
+            668
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "After commonsense knowledge extraction, we need to inject the extracted commonsense knowledge into the target model. A straightforward solution is to use sequence-level knowledge distillation (Kim and Rush, 2016) and continually train the student to generate retrieved commonsense inference given the original text and commonsense relation. However, this can be sub-optimal due to the domain discrepancy between commonsense knowledge and natural text, which introduces the catastrophic forgetting problem (Kirkpatrick et al., 2017) and hurts the performance on downstream tasks, which is also recently confirmed by Cui and Chen (2021).",
+        "bbox": [
+            507,
+            676,
+            882,
+            884
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To better inject the extracted commonsense knowledge into a pre-trained model without suf",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "5948",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/993820ac877f536970a3d0d6a6528f62eadeae6df0bc3dd649e531b33249fb65.jpg",
+        "image_caption": [
+            "Figure 2: Illustration of the commonsense text infilling objective. Given a commonsense tuple constructed in the commonsense knowledge retrieval phase, we randomly mask text spans in the commonsense tuple following different patterns and train the pre-trained model to reconstruct the masked spans. RL: Should the <mask> be \"xNeed\"?"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            82,
+            484,
+            223
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "fering from catastrophic forgetting so that its capability on general NLP tasks is retained (or even improved), we propose two commonsense-related self-supervised objectives: commonsense text infilling and commonsense relation prediction. The former objective is generative while the latter is a discriminative objective. We refine the pre-trained model by multi-tasking on both objectives so that the model can better adapt to tasks requiring either generative or discriminative commonsense reasoning ability.",
+        "bbox": [
+            112,
+            354,
+            489,
+            530
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Text Infilling Commonsense text infilling is a simple extension to the conventional text infilling objective used for pre-training BART and T5. It transforms each sentence to a commonsense tuple similar to that in a commonsense knowledge graph by appending the commonsense relation and the generated commonsense inference. We then mask text spans in the commonsense tuple by randomly selecting one masking scheme among text masking, commonsense masking, bidirectional masking, and relation masking. As illustrated in Fig 2, these masking strategies selectively mask different components in the input commonsense tuple and lead to different optimization objectives. Specifically, these masking schemes mask either spans in natural text  $(\\mathrm{P}(s|\\tilde{s},r,o))$ , commonsense inference  $(\\mathrm{P}(o|s,r,\\tilde{o}))$ , natural text/commonsense inference  $(\\mathrm{P}(s,o|\\tilde{s},r,\\tilde{o}))$ , or commonsense relation  $(\\mathrm{P}(r|s,\\tilde{r},o))$ , respectively. We then train the model to predict the masked spans autoregressively. The diverse masking strategies provide more diverse training signals compared to random masking, thus",
+        "bbox": [
+            112,
+            549,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "image",
+        "img_path": "images/5037654920a72b9799f8ae40dfb64768aae05d0b9d714a9457bdc51ecb09e694.jpg",
+        "image_caption": [
+            "Figure 3: Illustration of the commonsense relation prediction objective. We train the pre-trained model to predict the correct commonsense inference given the subject and relation from three distractors generated with either different subjects or relations as inputs."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            509,
+            86,
+            880,
+            225
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "enabling the model to better align the surface form of human language and the underlying commonsense knowledge.",
+        "bbox": [
+            507,
+            323,
+            882,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In addition, unlike the conventional practice in masked span infilling objective that randomly masks text spans with the same probability, we propose to mask text spans including concepts (tokens recognized as nouns or verbs by a Spacy POS tagger) with a higher probability so that the model will be trained to predict concepts more frequently compared to non-content words that are generally not related to commonsense reasoning.",
+        "bbox": [
+            507,
+            372,
+            882,
+            517
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Relation Prediction While the commonsense text infilling objective encourages the pre-trained model to align natural texts and their commonsense inferences, it is always trained on valid commonsense tuples. This can be suboptimal because we also want the model to be capable of discriminating invalid commonsense inferences, which is important for many commonsense-related downstream tasks.",
+        "bbox": [
+            507,
+            531,
+            882,
+            675
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To this end, we introduce a commonsense relation prediction task that trains the model to distinguish the correct commonsense inference corresponding to the input sentence and the commonsense relation from distractors. To be specific, the commonsense relation prediction objective is formulated as a multi-choice QA problem with an input sentence as the context, a commonsense relation as the question, and a set of four commonsense inferences as options. The set of options consists of one correct commonsense inference, which is generated by the neural commonsense model with the input sentence and commonsense relation as input, and three carefully curated distractors (i.e., negative examples) generated by the",
+        "bbox": [
+            507,
+            678,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "5949",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/08469e2d6c43f673d4725ed8b5c05c59514c32958c49b36bad6c4c8db42fe40b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALQA</td><td>COPA</td></tr><tr><td>BERT-base</td><td>53.08(±0.16)</td><td>57.60(±0.8)</td><td>64.86(±0.52)</td><td>61.88(±0.56)</td><td>64.3(±0.4)</td><td>67.3(±0.4)</td></tr><tr><td>ERNIE-base</td><td>54.06(±0.12)</td><td>58.90(±0.9)</td><td>66.47(±0.58)</td><td>63.04(±0.46)</td><td>65.1(±0.4)</td><td>68.9(±0.4)</td></tr><tr><td>KnowBERT</td><td>53.88(±0.15)</td><td>58.50(±0.8)</td><td>66.61(±0.63)</td><td>63.18(±0.52)</td><td>65.4(±0.5)</td><td>69.4 (±0.4)</td></tr><tr><td>COMET</td><td>45.32(±0.28)</td><td>51.20(±1.1)</td><td>60.73(±0.51)</td><td>57.63(±0.61)</td><td>60.2(±0.7)</td><td>69.1 (±0.5)</td></tr><tr><td>T5-base</td><td>61.88(±0.08)</td><td>58.20(±1.0)</td><td>68.14(±0.73)</td><td>61.10(±0.38)</td><td>65.1(±0.5)</td><td>71.4 (±0.7)</td></tr><tr><td>T5-base + TI</td><td>62.05(±0.17)</td><td>58.43(±0.8)</td><td>68.32(±0.66)</td><td>61.42(±0.32)</td><td>65.3(±0.4)</td><td>71.8 (±0.8)</td></tr><tr><td>T5-base + SSM</td><td>62.37(±0.25)</td><td>58.60(±0.9)</td><td>68.48(±0.65)</td><td>61.57(±0.44)</td><td>65.5(±0.5)</td><td>72.1 (±0.6)</td></tr><tr><td>T5-base + KD</td><td>61.83(±0.42)</td><td>56.54(±0.7)</td><td>67.35(±0.63)</td><td>60.94(±0.66)</td><td>64.8(±0.5)</td><td>71.0 (±1.0)</td></tr><tr><td>T5-base + CSKG (TI)</td><td>60.22(±0.40)</td><td>56.17(±0.8)</td><td>66.51(±0.57)</td><td>59.92(±0.47)</td><td>62.7(±0.7)</td><td>68.5 (±1.1)</td></tr><tr><td>T5-base + CSKG (Rule)</td><td>63.10(±0.35)</td><td>57.97(±0.8)</td><td>68.27(±0.71)</td><td>60.15(±0.51)</td><td>65.7(±0.4)</td><td>72.4 (±0.9)</td></tr><tr><td>CALM</td><td>63.32(±0.35)</td><td>60.90(±0.4)</td><td>71.01(±0.61)</td><td>63.20(±0.52)</td><td>66.0(±0.5)</td><td>72.2 (±0.8)</td></tr><tr><td>CKT-base</td><td>64.11(±0.31)</td><td>61.58(±0.5)</td><td>72.26(±0.61)</td><td>64.37(±0.49)</td><td>67.3(±0.4)</td><td>73.4 (±0.5)</td></tr><tr><td>CKT w/ GPT-2</td><td>60.39(±0.61)</td><td>56.95(±0.7)</td><td>68.48(±0.44)</td><td>60.14(±0.52)</td><td>66.2(±0.6)</td><td>72.8 (±1.0)</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            878,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Experimental results on base-size models. Best models are bold and second best ones are underlined within each metric. Mean and standard deviation of 3 different runs with different random seeds are reported.",
+        "bbox": [
+            112,
+            307,
+            880,
+            338
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "same neural commonsense knowledge model with different inputs. As illustrated in Figure 3, among the three distractors, one is generated with an input composed of the same sentence and a different commonsense relation, and another two are generated with an input composed of different sentences with the same commonsense relation. In this way, the model learns to align the natural texts with valid commonsense knowledge while also distinguishing commonsense inferences that do not make sense. Moreover, this objective is formulated as a multi-choice QA task that closely resembles several downstream commonsense-related tasks such as CommonsenseQA and SOCIALIQA, thus enabling easier transfer especially when labeled training examples are scarce.",
+        "bbox": [
+            112,
+            348,
+            489,
+            605
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            615,
+            260,
+            633
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.1 Experimental Settings",
+        "text_level": 1,
+        "bbox": [
+            112,
+            640,
+            334,
+            657
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Models In our experiments we apply commonsense knowledge transfer to refine T5 (Raffel et al., 2019), a popular model pre-trained with the text infilling objective. We experiment with both T5-base and T5-large, which consist of 220 million and 774 million parameters respectively, as the target model in the commonsense knowledge transfer framework. We do not experiment with extremely large models like T5-11B because of the resource constraints and the fact that these models are hard to deploy in real-world applications. We use COMET-ATOMIC $^{20}$  a state-of-the-art neural commonsense knowledge model that can generate accurate, representative knowledge for new, unseen entities and events, as the source model. It is initialized with BART and continually trained",
+        "bbox": [
+            112,
+            661,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "on  $\\mathrm{ATOMIC}_{20}^{20}$  (Hwang et al., 2021), a new general purpose commonsense knowledge graph.",
+        "bbox": [
+            507,
+            348,
+            880,
+            381
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Data We randomly sample a subset consisting of 10 million sentences from the English Wikipedia and the BookCorpus (Zhu et al., 2015), which is used for pre-training BERT and its variants. We select a set of representative commonsense relations including intent, reason, effect, need, want, and react from relations used to train COMET-ATOMIC $_{20}$ . For each sentence, we randomly sample two relations and retrieve the corresponding commonsense explanation from COMET $_{20}$ . We randomly select one relation-explanation pair to form the input example and leave another as the distractor for the commonsense relation prediction objective.",
+        "bbox": [
+            507,
+            382,
+            884,
+            592
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Training We refine the pre-trained models on the self-supervised examples constructed with the sampled 10 million sentences for 100k steps with a batch size of 1024, a maximum sequence length of 256, and a learning rate of  $5\\mathrm{e} - 5 / 2\\mathrm{e} - 5$  for base-size and large-size models respectively with a linear warm-up for the first 8,000 updates. After knowledge transfer, we fine-tune the models on downstream tasks by formulating the tasks into text-to-text problems. Pre-training and fine-tuning details are included in the Appendix.",
+        "bbox": [
+            507,
+            595,
+            882,
+            772
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Evaluation We evaluate the continual pre-trained models on downstream tasks that require commonsense reasoning including CommonsenseQA (Talmor et al., 2018), OpenbookQA (Mihaylov et al., 2018), PIQA (Bisk et al., 2020), aNLI (Bhagavatula et al., 2020), COPA (Roemmelo et al., 2011), and SOCAI LIQA (Sap et al., 2019b) In addition to the conventional fully supervised setting, we also test our approach in the few-shot setting by vary",
+        "bbox": [
+            507,
+            774,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5950",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/e87dcda1f6876e931637946206cda19593cfe6c57f4e0b12f03438e658557938.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALQA</td><td>COPA</td></tr><tr><td>T5-large</td><td>69.81(±1.02)</td><td>61.40(±1.0)</td><td>72.19(±1.09)</td><td>75.54(±1.22)</td><td>71.3(±0.8)</td><td>83.6(±1.1)</td></tr><tr><td>CALM-large</td><td>71.31(±0.04)</td><td>66.00(±1.0)</td><td>75.11(±1.65)</td><td>77.12(±0.34)</td><td>72.7(±0.7)</td><td>84.9(±1.0)</td></tr><tr><td>CKT-large</td><td>72.15(±0.61)</td><td>66.70(±1.1)</td><td>76.07(±0.95)</td><td>77.94(±0.59)</td><td>73.8(±0.8)</td><td>86.0(±1.2)</td></tr></table>",
+        "bbox": [
+            119,
+            80,
+            877,
+            165
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Experimental results on large-size models. Best models are bold and second best ones are underlined within each metric. Mean and variance of 3 different runs with different random seeds are reported.",
+        "bbox": [
+            112,
+            174,
+            880,
+            202
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "ing the percentage of labeled examples from the original training set used for fine-tuning. The idea is that limited labeled examples can only help the model understand the task but are insufficient for the model to acquire enough commonsense knowledge to solve the task. As such, it requires the model to store enough commonsense knowledge in its parameters to succeed in the few-shot setting. For both the settings, we report the results on the official development set and tune the hyperparameters based on the models' performance on an in-house split dev set. We report the mean and variance of 3 individual runs with different random seeds because most datasets are relatively small, which makes the variance in results non-negligible.",
+        "bbox": [
+            112,
+            211,
+            489,
+            451
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Baselines We compare our approach with methods that continually train a pre-trained model with different objectives. We divide the baselines into two categories based on the source of their supervision. The first category includes methods that only exploit general text corpus, including (1) T5 + TI that continually pre-trains the public checkpoint of T5 with the same text infilling objective for more steps, (2) T5 + SSM that also continual pretrains T5 with the text infilling objective, but use salient span masking (Roberts et al., 2020) instead of random masking for data construction, (3) (T5 + KD) that uses sequence-level knowledge distillation (Kim and Rush, 2016) for knowledge transfer, where the student model is trained with the teacher output (i.e.,  $\\mathrm{P}(o|s,r)$ ), and (4) CALM (Zhou et al., 2021) that uses novel self-supervised objectives to construct concept-centric self-supervision from general text corpus. The second category instead exploits CSKG, including (5) T5 + CSKG (TI) train T5 with the text infilling objective on tuples in a CSKG, and (6) T5 + CSKG (Rule) (Li et al., 2019) that uses manually defined rules to construct training examples from a CSKG. We also include a COMET baseline where we directly fine-tune the pre-trained COMET-ATOMIC $^{20}$  model for the downstream tasks to verify the necessity of commonsense knowledge transfer, and a CKT w/ GPT",
+        "bbox": [
+            110,
+            468,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/6696c0db6024ee79cf5541d77e1c0c9dfd3867fc4aae2e9fced1019ae42daf41.jpg",
+        "image_caption": [
+            "Figure 4: Performance of compared base-size models fine-tuned with different fraction of the datasets."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            216,
+            687,
+            340
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/2f07e7f98aadcdf747fd79588a408c351d7197a44be95d323eeca04b0d1b01dc.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            707,
+            216,
+            875,
+            340
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "2 baseline where the commonsense inferences are generated by a pre-trained GPT-2 large model to verify whether the gain comes from transferring the commonsense knowledge from COMET, or simply from data augmentation from another generative model. For a fair comparison, we use the same data and training steps compared to our approach for baselines from the first category and use  $\\mathrm{ATOMIC}_{20}^{20}$ , on which the teacher model in our framework is pre-train on, as the commonsense knowledge graph. For reference, we also include some popular knowledge-enhanced pre-trained models including ERNIE (Zhang et al., 2019) and KnowBERT (Peters et al., 2019).",
+        "bbox": [
+            507,
+            398,
+            884,
+            621
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "3.2 Fully-supervised Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            638,
+            747,
+            653
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We first present results in the fully-supervised setting. Results on base-size models are presented in Table 1. We can see that our approach yields significant improvement compared to the T5 baseline (up to 4 absolute scores) and consistently outperform CALM, the state-of-the-art method of injecting commonsense knowledge into PTLMs.",
+        "bbox": [
+            505,
+            659,
+            884,
+            772
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In addition, we observe that simply using continual training with the original text-infilling objective or its variant with salient span masking only marginally improves the performance. Surprisingly, training with text infilling on a commonsense knowledge graph leads to degraded performance compared to the T5 baseline. We suspect this is because the commonsense tuples in commonsense knowledge graphs are generally too short",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "5951",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/6df0ff102d4ad7971431dbb93e09ad521c5d55883cd0b1067c5f1c48c573b810.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SIQA</td><td>COPA</td></tr><tr><td>T5-base</td><td>61.88</td><td>58.20</td><td>68.14</td><td>61.10</td><td>65.1</td><td>71.4</td></tr><tr><td>CKT-base</td><td>64.57</td><td>62.77</td><td>73.26</td><td>64.75</td><td>68.3</td><td>73.4</td></tr><tr><td colspan=\"7\">Objective Analysis</td></tr><tr><td>CKT-base w/o CSTI</td><td>62.58</td><td>60.97</td><td>70.61</td><td>62.11</td><td>66.5</td><td>72.0</td></tr><tr><td>CKT-base w/o text masking</td><td>62.98</td><td>61.74</td><td>72.55</td><td>63.81</td><td>67.7</td><td>72.8</td></tr><tr><td>CKT-base w/o commonsense masking</td><td>63.61</td><td>62.03</td><td>72.83</td><td>64.40</td><td>67.5</td><td>72.7</td></tr><tr><td>CKT-base w/o bidirectional masking</td><td>63.52</td><td>62.11</td><td>72.30</td><td>64.24</td><td>67.6</td><td>72.9</td></tr><tr><td>CKT-base w/o relation masking</td><td>64.12</td><td>62.48</td><td>73.31</td><td>64.57</td><td>67.4</td><td>72.7</td></tr><tr><td>CKT-base w/o CSRIP</td><td>63.12</td><td>62.07</td><td>72.44</td><td>64.11</td><td>67.5</td><td>72.6</td></tr><tr><td>CKT-base w/ random distractors</td><td>64.04</td><td>62.29</td><td>72.95</td><td>64.48</td><td>68.0</td><td>73.1</td></tr><tr><td colspan=\"7\">Multi-task versus Sequential Transfer</td></tr><tr><td>CKT-base (CSTI → CSRIP)</td><td>64.69</td><td>62.51</td><td>73.35</td><td>64.11</td><td>67.9</td><td>73.5</td></tr><tr><td>CKT-base (CSRIP → CSTI)</td><td>63.49</td><td>61.33</td><td>71.54</td><td>63.41</td><td>67.0</td><td>72.0</td></tr><tr><td colspan=\"7\">Corpus Size</td></tr><tr><td>CKT-base w/ 10% data</td><td>64.18</td><td>62.21</td><td>71.86</td><td>64.31</td><td>67.7</td><td>73.1</td></tr><tr><td>CKT-base w/ 50% data</td><td>64.45</td><td>62.66</td><td>73.10</td><td>64.72</td><td>68.2</td><td>73.4</td></tr></table>",
+        "bbox": [
+            174,
+            80,
+            823,
+            380
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Analysis of the proposed commonsense knowledge transfer framework. CSTI and CSRP denote the commonsense text infilling objective and the commonsense relation prediction objective, respectively.",
+        "bbox": [
+            112,
+            388,
+            882,
+            418
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "and simple, making the pre-trained model unable to reason within relatively long contexts which is crucial for most downstream tasks. Moreover, we find that continually pre-training with training data constructed with commonsense tuples in a commonsense knowledge graph following manually designed rules leads to improvements in certain tasks. However, the improvement is inconsistent across different tasks and it even hurts the performance on certain tasks, which may be because the rules for constructing training data are tailored for certain tasks like CSQA. The inferior performance of using commonsense knowledge graphs as data sources also confirms the need of using natural text corpus during continual pre-training for better adapting to diverse downstream tasks. We also find directly applying sequence-level KD and training the student to mimic the teacher on the commonsense tuple generation task fails to improve the performance because the task is not general enough and thus cannot transfer to diverse downstream tasks well. Moreover, directly fine-tuning COMET or using GPT-2 as the commonsense knowledge source results in very poor performance. This confirms the necessity of commonsense knowledge transfer and shows that it is actually transferring commonsense knowledge instead of simple text augmentation.",
+        "bbox": [
+            112,
+            432,
+            489,
+            866
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To further confirm the effectiveness of commonsense knowledge transfer, we apply it to T5-large and compare it to competitive baselines in the base",
+        "bbox": [
+            112,
+            871,
+            489,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "size experiments. The results are presented in Table 2. We can see that our approach consistently outperforms T5-large and CALM-large. This suggests that our approach can successfully generalize to large-size pre-trained models.",
+        "bbox": [
+            507,
+            432,
+            884,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.3 Few-shot Results",
+        "text_level": 1,
+        "bbox": [
+            509,
+            525,
+            690,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Injecting commonsense knowledge into pre-trained models is important because it enables the model to reason and generalize to unseen examples while observing only a few labeled examples. To this end, we fine-tune the compared models with different fractions of labeled training data to investigate the transition of the behavior of our model and baselines from the low-resource regime to the fully-supervised setting (Fig. 4). We observe that the performance improvement of our approach compared to the baselines is more significant in the low-resource regime. This shows that commonsense knowledge transfer can successfully transfer commonsense knowledge into pre-trained models so that they can generalize well while seeing only a small part of training data. This may also help the model reduce the risk/tendency of fitting the spurious correlations in the annotated datasets and thus generalize better.",
+        "bbox": [
+            507,
+            546,
+            884,
+            852
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "3.4 Analysis",
+        "text_level": 1,
+        "bbox": [
+            509,
+            865,
+            623,
+            881
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To better understand the proposed commonsense knowledge transfer framework and the role of its",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "5952",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "different components, we conduct an ablation study about the impact of different proposed objectives, the impact of multi-tasking the commonsense-related self-supervised objective versus sequentially training, and the impact of the size of natural text corpus used for transfer (see Table 3).",
+        "bbox": [
+            112,
+            84,
+            487,
+            179
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of Objectives We find that both the proposed objectives contribute to the performance improvement of our approach. The commonsense text infilling objective is shown to be more critical than the commonsense relation prediction task. We suspect this is because commonsense text infilling resembles the vanilla text infilling objective with which the T5 models are pre-trained, thus preventing the model from catastrophic forgetting. In addition, all four masking strategies are beneficial, and their contribution varies for different downstream tasks. This confirms the necessity of a diverse masking scheme. Moreover, our strategy for constructing distractors outperforms the random counterpart, demonstrating the necessity of hard negative examples for the commonsense relation prediction task.",
+        "bbox": [
+            115,
+            181,
+            489,
+            455
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-task versus Sequential Transfer As for the training order between the two objectives, we find that starting from the commonsense text infilling task and then switching to the commonsense relation prediction task performs similarly with our multi-tasking strategy while significantly outperforming its counterpart training with the reverse direction. We think this is because the commonsense text infilling objective resembles the original pre-training while the commonsense relation prediction is more similar to downstream tasks. We opt for the multi-tasking strategy for simplicity.",
+        "bbox": [
+            112,
+            457,
+            489,
+            649
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Impact of Corpus Size We find that commonsense knowledge transfer significantly outperforms both the T5 baseline and the competitive CALM method with only 10 percent of the full data used for distillation. Nevertheless, the performance improvement also confirms that our approach can benefit from the accessibility of large-scale natural texts. For base-size models, the performance improvements seem to saturate after 10 million sentence pairs. However, we anticipate that larger-size models may still benefit from a larger amount of data, and leave this for future work.",
+        "bbox": [
+            112,
+            651,
+            489,
+            843
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "4 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            858,
+            268,
+            873
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Knowledge-augmented Pre-trained Models A number of recent works have examined the prob-",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "lem of incorporating world knowledge with the pre-trained models. A number of works use an external knowledge base to incorporate entity knowledge with pre-trained models (Zhang et al., 2019; Peters et al., 2019; Wang et al., 2020; Liu et al., 2020). However, these approaches require specialized resources like knowledge bases which are non-trivial to seek, thus limiting the domain they can be applied to. Xiong et al. (2020) proposed a novel entity replacement detection objective that incorporates Wikipedia to encode world knowledge into a BERT-like pre-trained model. He et al. (2020) proposed a generative and discriminative framework that pretrains the model to complete and correct knowledge spans. The aforementioned approaches generally focus on factual knowledge of entities while our work mainly focuses on commonsense knowledge.",
+        "bbox": [
+            505,
+            84,
+            884,
+            357
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Commonsense Reasoning for NLP Several recent studies (Talmor et al., 2018; Sap et al., 2019b; Zhou et al., 2020b; Lin et al., 2020; Xu et al., 2021) evaluate the performance of several pre-trained language models on tasks that require commonsense reasoning and find that it is still very hard for pre-trained language models to match or exceed human-level performance. Therefore, approaches to improve the commonsense reasoning ability of pre-trained language models has attracted much attention. These approaches can be divided into two categories. The first category focuses on incorporating an external commonsense knowledge graph for commonsense reasoning. For example, Lin et al. (2019), Cui and Chen (2021), and Liu et al. (2021) propose to exploit structured symbolic commonsense knowledge graphs to perform commonsense reasoning. The second one instead attempts to inject commonsense knowledge into the parameters of pre-trained models. For example, Ye et al. (2019); Li et al. (2019) proposed to use manually designed rules to construct commonsense related training examples from commonsense knowledge graphs. Zhou et al. (2021) instead only relies on general text corpus and proposed two concept-centric self-supervised objectives to refine pre-trained models with commonsense knowledge.",
+        "bbox": [
+            507,
+            359,
+            884,
+            793
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            810,
+            640,
+            824
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We introduce commonsense knowledge transfer, a framework to transfer the commonsense knowledge stored in a neural commonsense knowledge model to a general-purpose pre-trained model. Our method extracts commonsense knowledge from",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "5953",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "the source model to construct self-supervised training data for the target model. Empirical results show that our approach consistently outperforms previous methods for improving the commonsense reasoning ability of pre-trained models that exploit either symbolic knowledge graphs or texts alone.",
+        "bbox": [
+            112,
+            84,
+            487,
+            180
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            193,
+            220,
+            206
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, we use T5-base and T5-large models as the target model since they are widely-used, representative pre-trained seq2seq models and use COMET-ATOMIC $^{20}$  as the commonsense knowledge source. However, there are other pretrained seq2seq models such as BART, and neural commonsense models such as COMET that we did not experiment with. Moreover, we only experimented with 10 million randomly sampled sentences from the English Wiki and BookCorpus datasets. It would be interesting to investigate whether continually pre-training with a larger scale dataset can further improve the performance.",
+        "bbox": [
+            112,
+            218,
+            489,
+            426
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethical Considerations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            439,
+            315,
+            454
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our work focuses on improving the commonsense reasoning ability of pre-trained language models. It probably does not introduce extra ethical concerns. However, in commonsense knowledge extraction, the neural commonsense knowledge model may generate unexpected (e.g., biased) commonsense inferences, and training with these inferences may lead to additional bias in the pre-trained model. Nevertheless, all pre-trained language models contain bias and should be examined.",
+        "bbox": [
+            112,
+            464,
+            489,
+            625
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            652,
+            213,
+            667
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Emily M. Bender and Alexander Koller. 2020. Climbing towards NLU: on meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 5185-5198. Association for Computational Linguistics.",
+            "Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen-tau Yih, and Yejin Choi. 2020. Abductive commonsense reasoning. In ICLR.",
+            "Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2020. Piqa: Reasoning about physical commonsense in natural language. In Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI)."
+        ],
+        "bbox": [
+            115,
+            674,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. COMET: commonsense transformers for automatic knowledge graph construction. In ACL (1), pages 4762-4779. Association for Computational Linguistics.",
+            "Alexis Conneau and Douwe Kiela. 2018. Senteval: An evaluation toolkit for universal sentence representations. In LREC.",
+            "Wanyun Cui and Xingran Chen. 2021. Enhancing language models with plug-and-play large-scale commonsense. CoRR, abs/2109.02572.",
+            "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT* (1), pages 4171–4186. Association for Computational Linguistics.",
+            "William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In IWP@IJCNLP.",
+            "WA Falcon. 2019. Pytorch lightning. GitHub. Note: https://github.com/PyTorchLightning/pytorchlightning, 3.",
+            "Bin He, Xin Jiang, Jinghui Xiao, and Qun Liu. 2020. Kgplm: Knowledge-guided language model pretraining via generative and discriminative learning. CoRR, abs/2012.03551.",
+            "John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4129-4138. Association for Computational Linguistics.",
+            "Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. CoRR, abs/1503.02531.",
+            "Pedram Hosseini, David A. Broniatowski, and Mona T. Diab. 2021. Commonsense knowledge-augmented pretrained language models for causal reasoning classification. CoRR, abs/2112.08615.",
+            "Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2021. (comet-) atomic 2020: On symbolic and neural commonsense knowledge graphs. In AAAI, pages 6384-6392. AAAI Press.",
+            "Yoon Kim and Alexander M. Rush. 2016. Sequence-level knowledge distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 1317-1327. The Association for Computational Linguistics."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "5954",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526.",
+            "Tassilo Klein and Moin Nabi. 2021. Towards zero-shot commonsense reasoning with self-supervised refinement of language models. CoRR, abs/2109.05105.",
+            "Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In ACL, pages 7871-7880. Association for Computational Linguistics.",
+            "Shiyang Li, Jianshu Chen, and Dian Yu. 2019. Teaching pretrained models with commonsense reasoning: A preliminary kb-based approach. CoRR, abs/1909.09743.",
+            "Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. _Kagnet: Knowledge-aware graph networks for commonsense reasoning_. In _EMNLP/IJCNLP (1)_, pages 2829-2839. 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, Online Event, 16-20 November 2020, volume EMNLP 2020 of *Findings of ACL*, pages 1823-1840. Association for Computational Linguistics.",
+            "Ye Liu, Yao Wan, Lifang He, Hao Peng, and Philip S. Yu. 2020. Kg-bart: Knowledge graph-augmented bart for generative commonsense reasoning.",
+            "Ye Liu, Yao Wan, Lifang He, Hao Peng, and Philip S. Yu. 2021. KG-BART: knowledge graph-augmented BART for generative commonsense reasoning. In AAAI, pages 6418-6425. AAAI Press.",
+            "Christopher D. Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proc. Natl. Acad. Sci. USA, 117(48):30046-30054.",
+            "William Merrill, Yoav Goldberg, Roy Schwartz, and Noah A. Smith. 2021. Provable limitations of acquiring meaning from ungrounded form: What will future language models understand? CoRR, abs/2104.10809.",
+            "Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Matthew E Peters, Mark Neumann, Robert L Logan IV, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. arXiv preprint arXiv:1909.04164.",
+            "Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick S. H. Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander H. Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 2463-2473. Association for Computational Linguistics.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.",
+            "Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.",
+            "Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,  $000+$  questions for machine comprehension of text. In EMNLP.",
+            "Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 5418-5426. Association for Computational Linguistics.",
+            "Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In *Logical Formalizations of Commonsense Reasoning*, Papers from the 2011 AAAI Spring Symposium, Technical Report SS-11-06, Stanford, California, USA, March 21-23, 2011. AAAI.",
+            "Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019a. ATOMIC: an atlas of machine commonsense for if-then reasoning. In AAAI, pages 3027-3035. AAAI Press.",
+            "Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019b. Socialiaq: Commonsense reasoning about social interactions. In EMNLP.",
+            "Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "5955",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In AAAI, pages 4444-4451. AAAI Press.",
+            "Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowledge. arXiv preprint arXiv:1811.00937.",
+            "Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Trans. Assoc. Comput. Linguistics, 8:621-633.",
+            "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 NIPS, pages 5998-6008.",
+            "Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR.",
+            "Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Cuihong Cao, Daxin Jiang, Ming Zhou, et al. 2020. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808.",
+            "Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. TACL.",
+            "Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In *NAACL-HLT*.",
+            "Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2019. Huggingface's transformers: State-of-the-art natural language processing. ArXiv, abs/1910.03771.",
+            "Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. 2020. Pretrained encyclopedia: Weakly supervised knowledge-pretrained language model. In International Conference on Learning Representations.",
+            "Canwen Xu, Wangchunshu Zhou, Tao Ge, Ke Xu, Julian J. McAuley, and Furu Wei. 2021. Blow the dog whistle: A chinese dataset for cant understanding with common sense and world knowledge. In NAACL-HLT, pages 2139-2145. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhi-Xiu Ye, Qian Chen, Wen Wang, and Zhen-Hua Ling. 2019. Align, mask and select: A simple method for incorporating commonsense knowledge into language representation models. CoRR, abs/1908.06725.",
+            "Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. 2019. Ernie: Enhanced language representation with informative entities. arXiv preprint arXiv:1905.07129.",
+            "Wangchunshu Zhou, Dong-Ho Lee, Ravi Kiran Selvam, Seyeon Lee, and Xiang Ren. 2021. Pretraining text-to-text transformers for concept-centric common sense. In ICLR. OpenReview.net.",
+            "Xuhui Zhou, Yue Zhang, Leyang Cui, and Dandan Huang. 2020a. Evaluating commonsense in pretrained language models. In AAAI, pages 9733-9740. AAAI Press.",
+            "Xuhui Zhou, Yue Zhang, Leyang Cui, and Dandan Huang. 2020b. Evaluating commonsense in pretrained language models. In AAAI, pages 9733-9740.",
+            "Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 19-27. IEEE Computer Society."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            516
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "5956",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Pre-training and Fine-tuning Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            465,
+            99
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.1 Pre-Training Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            109,
+            324,
+            124
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "We implement our models using Pytorchlightning (Falcon, 2019) and Hugginface's Pytorch Transformers (Wolf et al., 2019). For pre-training phase, we use the AdamW optimizer with maximum sequence length 256, train batch size 8, gradient accumulation 8, warmup steps 8000, weight decay 0.01 and adam epsilon 1e-6. We train the models with 8 V100 GPUs and FP32 precision. The model is pre-trained for 10 epochs. We searched for the best learning rate for our model out of [5e-6, 2e-5, 5e-5, 1e-4].",
+        "bbox": [
+            112,
+            130,
+            489,
+            305
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A.2 Fine-Tuning Details",
+        "text_level": 1,
+        "bbox": [
+            112,
+            317,
+            319,
+            332
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "For fine-tuning, we use 4 V100 GPUs and use FP32. For all tasks, we use the AdamW optimizer with learning rate from [1e-5, 2e-5, 5e-5, 1e-4, 2e-4], maximum sequence length 256, batch size from [4, 8, 16, 32]. For all tasks, we use a warmup fraction of 0.01, and max epoch of 20.",
+        "bbox": [
+            112,
+            338,
+            487,
+            434
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B Additional Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            445,
+            322,
+            462
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.1 Qualitative Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            472,
+            322,
+            488
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To better understand the proposed method, we present a case study in Figure 5. We can see that both the objectives introduced in the CALM model and the salient span masking (SSM) strategy fail to exploit the underlying commonsense rationale beyond the surface form of texts while our approach directly aligns texts with the corresponding commonsense inferences with different commonsense relations. That explains why commonsense knowledge transfer can effectively improve a pre-trained model's performance on downstream tasks requiring commonsense reasoning ability.",
+        "bbox": [
+            112,
+            492,
+            489,
+            684
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.2 Experimental Results on GLUE",
+        "text_level": 1,
+        "bbox": [
+            112,
+            696,
+            413,
+            711
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To verify that commonsense knowledge transfer is suitable for general-purpose pre-trained models, we fine-tune our model on the GLUE benchmark (Wang et al., 2019). Specifically, we test on MRPC (Dolan and Brockett, 2005),  $\\mathrm{QQP}^1$  and STS-B (Conneau and Kiela, 2018) for Paraphrase Similarity Matching; SST-2 (Socher et al., 2013) for Sentiment Classification; MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016) and RTE (Wang et al., 2019) for the Natural Language",
+        "bbox": [
+            112,
+            715,
+            487,
+            878
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/a89e129b49bc350c147d63748da638dfdbafe7832ddb7d8c5efd8ff431cf4eae.jpg",
+        "image_caption": [
+            "Commonsense Text Infilling",
+            "Figure 5: Performance on the CSQA dataset w.r.t. the size of training data used for commonsense knowledge transfer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            102,
+            877,
+            225
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "image",
+        "img_path": "images/3ef0bfdea7718e704a99b95df607d6ba624dbef9d1783bb645dbb79882ee4442.jpg",
+        "image_caption": [
+            "Figure 6: Performance on the CSQA dataset w.r.t. the size of training data used for commonsense knowledge transfer."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            326,
+            848,
+            549
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Inference; CoLA (Warstadt et al., 2019) for Linguistic Acceptability.",
+        "bbox": [
+            507,
+            639,
+            882,
+            671
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "The results are shown on Table 4, we can see that after commonsense knowledge transfer, the resulting model's general natural language understanding ability is comparable with the original T5-base model. This shows that our approach does not affect the model's general transfer ability and thus can be applied to general-purpose language models.",
+        "bbox": [
+            507,
+            673,
+            882,
+            800
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "B.3 Experiments with BART",
+        "text_level": 1,
+        "bbox": [
+            507,
+            816,
+            752,
+            831
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "To demonstrate the versatility of commonsense knowledge transfer for different backbones, we conduct additional experiments using BART as the backbone model. The results are shown in Table 5. We can see that commonsense knowledge trans",
+        "bbox": [
+            507,
+            839,
+            882,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://www.quora.com/q/quoradata/ First-Quora-Dataset-Release-Question-Pairs",
+        "bbox": [
+            112,
+            892,
+            502,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "5957",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/d41a20890e3219061d7702c62bf169b0ad7423711c646848d1b36dde53771d78.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CoLA</td><td>MNLI</td><td>MRPC</td><td>QNLI</td><td>QQP</td><td>RTE</td><td>SST-2</td><td>SST-B</td><td>Meta Score</td></tr><tr><td>BERT-base</td><td>58.9</td><td>84.7</td><td>89.6</td><td>91.2</td><td>90.0</td><td>71.4</td><td>93.0</td><td>90.0</td><td>83.6</td></tr><tr><td>T5-base</td><td>55.9</td><td>84.5</td><td>90.3</td><td>90.5</td><td>90.2</td><td>76.2</td><td>92.8</td><td>87.8</td><td>83.5</td></tr><tr><td>CKT-base</td><td>57.4</td><td>84.4</td><td>90.6</td><td>90.9</td><td>89.9</td><td>76.8</td><td>92.5</td><td>88.4</td><td>83.9</td></tr></table>",
+        "bbox": [
+            121,
+            82,
+            873,
+            164
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/2fba38be9c2b6c98f2a676bffb3cead7a9475154e92f32e169a9e579442d6822.jpg",
+        "table_caption": [
+            "Table 4: Experimental results of base-size models on the GLUE benchmark."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALIQA</td><td>COPA</td></tr><tr><td>BART</td><td>72.31</td><td>65.80</td><td>74.12</td><td>78.27</td><td>71.6</td><td>85.6</td></tr><tr><td>CKT-BART</td><td>73.14</td><td>68.20</td><td>76.95</td><td>79.52</td><td>73.3</td><td>87.2</td></tr></table>",
+        "bbox": [
+            159,
+            189,
+            836,
+            266
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "fer also consistently improves the BART model, demonstrating the versatility of our approach.",
+        "bbox": [
+            112,
+            298,
+            487,
+            331
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/e408fea8065c55cf7133a9681ce6537ef6ed1f5a6f869e625bf968b5aba03253.jpg",
+        "table_caption": [
+            "Table 5: Experimental results (mean of 3 random runs) with BART."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>BLEU-4</td><td>METEOR</td><td>CIDEr</td><td>SPICE</td></tr><tr><td>T5-base</td><td>24.90</td><td>31.20</td><td>12.99</td><td>32.40</td></tr><tr><td>CALM-base</td><td>26.40</td><td>31.40</td><td>13.88</td><td>33.00</td></tr><tr><td>CKT-base</td><td>26.20</td><td>31.40</td><td>13.65</td><td>33.10</td></tr></table>",
+        "bbox": [
+            119,
+            342,
+            482,
+            407
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Experimental results (mean of 3 random runs) with BART.",
+        "bbox": [
+            112,
+            417,
+            485,
+            444
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B.4 Experiments on CommonGEN",
+        "text_level": 1,
+        "bbox": [
+            114,
+            466,
+            403,
+            481
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We also experiment on the CommonGEN dataset, a generative commonsense reasoning dataset where the model is required to take several keywords as inputs and output a sentence that makes sense. The results are shown in Table 6. We can see that our approach performs similarly with the CALM model, which includes the CommonGEN task objective as one of the pre-training tasks.",
+        "bbox": [
+            112,
+            487,
+            487,
+            615
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B.5 Impact of Pre-training Data Size",
+        "text_level": 1,
+        "bbox": [
+            112,
+            626,
+            420,
+            642
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We also conduct experiments to investigate the sample-efficiency of commonsense knowledge transfer. We present the trend of performance improvement in Figure 6. We can see that our method achieves significant performance improvement upon the T5 baseline with only  $10\\%$  of the total training data, which confirms the sample-efficiency of commonsense knowledge transfer.",
+        "bbox": [
+            112,
+            646,
+            489,
+            776
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "5958",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? limitation",
+            "A2. Did you discuss any potential risks of your work? ethical statement",
+            "A3. Do the abstract and introduction summarize the paper's main claims? abstract and introduction",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            287
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            299,
+            487,
+            316
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "experiment",
+        "bbox": [
+            132,
+            322,
+            218,
+            336
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? experiment",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? they're commonly used datasets",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? they're commonly used datasets",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? they're commonly used datasets",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? they're commonly used datasets",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. they're commonly used datasets"
+        ],
+        "bbox": [
+            129,
+            346,
+            880,
+            753
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "experiment",
+        "bbox": [
+            132,
+            788,
+            218,
+            802
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? experiment",
+        "bbox": [
+            129,
+            812,
+            880,
+            860
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "5959",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? experiment",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? experiment",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? experiment"
+        ],
+        "bbox": [
+            129,
+            83,
+            878,
+            282
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.",
+        "bbox": [
+            112,
+            293,
+            877,
+            330
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response."
+        ],
+        "bbox": [
+            127,
+            340,
+            878,
+            640
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "5960",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 14
+    }
+]

2023/Commonsense Knowledge Transfer for Pre-trained Language Models/f96ae55a-934d-4cd1-b9c6-f4ed8849b12f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbdffb32d95770a39c2cf2cf399cf72f5ba6f909f590e8b9dddda5d67d8ff872
+size 2666992

2023/Commonsense Knowledge Transfer for Pre-trained Language Models/full.md

@@ -0,0 +1,325 @@
+# Commonsense Knowledge Transfer for Pre-trained Language Models
+
+Wangchunshu Zhou*¹ Ronan Le Bras² Yejin Choi²³
+
+$^{1}$ ETH Zurich
+
+2Allen Institute for AI
+
+<sup>3</sup>Paul G. Allen School of Computer Science & Engineering, University of Washington wangchunshu.zhou@inf.ethz.ch
+
+# Abstract
+
+Despite serving as the foundation models for a wide range of NLP benchmarks, pre-trained language models have shown limited capabilities of acquiring implicit commonsense knowledge from self-supervision alone, compared to learning linguistic and factual knowledge that appear more explicitly in the surface patterns in text.
+
+In this work, we introduce commonsense knowledge transfer, a framework to transfer the commonsense knowledge stored in a neural commonsense knowledge model to a general-purpose pre-trained language model. It first exploits general texts to form queries for extracting commonsense knowledge from the neural commonsense knowledge model and then refines the language model with two self-supervised objectives: commonsense mask infilling and commonsense relation prediction, which align human language with the underlying commonsense knowledge.
+
+Empirical results show that our approach consistently improves the model's performance on downstream tasks that require commonsense reasoning. Moreover, we find that the improvement is more significant in the few-shot setting. This suggests that our approach helps language models better transfer to downstream tasks without extensive supervision by injecting commonsense knowledge into their parameters.
+
+# 1 Introduction
+
+Recent advances in pre-trained language models have transformed the landscape of natural language processing. Self-supervised pre-training objectives including masked language modeling (Devlin et al., 2019) and masked span infilling (Lewis et al., 2020) enable pre-trained models to acquire linguistic (Hewitt and Manning, 2019; Manning et al., 2020) and
+
+![](images/1832410ace14b469811e8e90ee7af2805ccfeacefbfaf920aef3a7e20737e420.jpg)  
+Figure 1: Illustration of the commonsense knowledge transfer framework. We first extract commonsense knowledge related to sentences in general text corpus from a neural commonsense knowledge model. We then use natural texts and the extracted commonsense knowledge to form self-supervised training data to refine a pre-trained model with commonsense knowledge.
+
+factual knowledge (Petroni et al., 2019) by modeling the distribution of naturally occurring texts. However, most of these objectives are limited to exploiting the surface form of human language, and the lack of grounded supervision calls into question how well these representations can ever capture meaning (Bender and Koller, 2020), not to mention the underlying commonsense knowledge which is often reasoned implicitly and does not appear in the surface form of human language (Merrill et al., 2021; Zhou et al., 2020a; Hwang et al., 2021). On the other hand, commonsense reasoning is important for building generalizable models because it enables the model to reason about a great number of events, causes, and effects, while observing only a small fraction of them. The ineffectiveness of self-supervised language model pre-training on acquiring commonsense knowledge makes them require a relatively large number of labeled examples to succeed in a downstream task and prune to overfit task-specific correlations (Tu et al., 2020).
+
+Therefore, equipping pre-trained language models with commonsense reasoning ability has at
+
+tracted much attention. To this end, two distinct lines of research focus on improving commonsense reasoning ability of pre-trained language models. The first one focuses on incorporating external commonsense knowledge graph for commonsense reasoning (Lin et al., 2019; Liu et al., 2021; Cui and Chen, 2021) while the other attempts to inject commonsense knowledge into the parameters of pretrained models (Li et al., 2019; Zhou et al., 2021; Klein and Nabi, 2021). In this work we focus on the second type of method because it alleviates the need for external knowledge bases for training and inference on downstream tasks, thus simpler, more efficient, and not limited by the coverage issue of external knowledge bases.
+
+Prior work injects commonsense knowledge into pre-trained models either on symbolic commonsense knowledge graphs with manually defined rules (Li et al., 2019) or masked language modeling (Hosseini et al., 2021) or on general text corpus with concept-centric self-supervised objectives (Zhou et al., 2021). The former method is limited by the coverage of knowledge graphs and human-written rules. It also fails to make use of large-scale diverse natural text corpus. Therefore, the training is limited to short and synthetic commonsense tuples, which affects its generalization ability on diverse downstream tasks. The latter method, however, only captures surface-level order relations between concepts and fails to learn commonsense relations between concepts such as cause, effect, intent, requirement, etc., which are crucial for commonsense reasoning but often implicitly reasoned, thus do not appear in the surface form of natural language.
+
+In this work, we propose commonsense knowledge transfer, an alternative framework to refine a general purpose pre-trained model's commonsense reasoning ability. In contrast to previous work, it aims to transfer the commonsense knowledge stored in a neural commonsense knowledge model (e.g., COMET (Bosselut et al., 2019)) to a general purpose pre-trained model on large scale general text corpus. In this way, our approach combines the best of both worlds from prior art: the dense and informative commonsense knowledge from commonsense knowledge graphs and the accessibility of large-scale diverse general corpus.
+
+Commonsense knowledge transfer is conceptually related to knowledge distillation (KD) (Hinton et al., 2015) since they both aim to trans
+
+fer knowledge from a knowledge-rich model to another model that lacks it. However, different from conventional KD, in commonsense knowledge transfer, the source model (i.e., neural commonsense model) and the target model (i.e., pretrained model) are heterogeneous. Moreover, instead of simply mimicking the teacher model, commonsense knowledge transfer requires the target model to learn specialized knowledge from the source model while retaining its own capability. This poses unique challenges since the knowledge transfer can not be accomplished by simply matching the logits or feature distribution between the student and the teacher. To this end, we propose to first extract commonsense knowledge in textual form from the source model and then exploit the extracted knowledge to form self-supervised training data for the target model. As illustrated in Figure 1, commonsense knowledge transfer first exploits general texts to form queries for retrieving commonsense knowledge from the neural commonsense knowledge model. Then it refines a pretrained model with two self-supervised objectives that align the surface form of human language with its underlying commonsense inference: commonsense text infilling and commonsense relation prediction. The former objective concatenates natural text with its commonsense inference to form an input example, masks certain spans in it, and trains the model to reconstruct the original input. The latter method instead trains the model to distinguish valid commonsense inference from carefully constructed spurious commonsense inference given the original text and commonsense relation. Refining a pre-trained model by multi-tasking on both generation (former) and understanding (latter) tasks enables the model to better adapt to different kinds of downstream tasks.
+
+We refine T5 (Raffel et al., 2020) with commonsense knowledge transfer and fine-tune the resulting model downstream tasks requiring commonsense reasoning ability in both the fully supervised setting and few-shot settings where only a percentage of labeled examples are available. Experimental results show substantial improvements in downstream tasks requiring commonsense reasoning, especially in the few-shot setting, demonstrating the effectiveness of our approach.
+
+# 2 Methodology
+
+Our proposed commonsense knowledge transfer framework consists of a neural commonsense knowledge model (e.g., COMET) and a pre-trained model (e.g., T5). The goal of commonsense knowledge transfer is to transfer the commonsense knowledge from the neural commonsense knowledge model (i.e., source model) to the pre-trained model (i.e., target model) so that it can generalize better to downstream tasks requiring commonsense reasoning ability.
+
+Compared to conventional knowledge transfer methods such as knowledge distillation, commonsense knowledge transfer faces a unique challenge: the source model and the target model are heterogeneous because they are trained on different data with different objectives. As such, we can not simply feed a batch of data to both of the models and train the target model to match the source model's logits or feature distribution. To alleviate this problem, we propose a two-stage knowledge transfer scheme as illustrated in Figure 1. To be specific, we first use natural texts to form queries for retrieving commonsense knowledge (in text form) from the neural commonsense knowledge model. We then construct training data with two novel commonsense-related self-supervised objectives based on the retrieved commonsense knowledge and the corresponding natural text. Finally, we train the target model on the constructed training data to inject commonsense knowledge retrieved from the source model. We describe our method to extract commonsense knowledge from a neural commonsense knowledge model and the proposed commonsense-related self-supervised objectives in detail in this section.
+
+# 2.1 Commonsense Knowledge Extraction
+
+We first describe the source model, i.e., neural commonsense knowledge model, in the commonsense knowledge transfer framework. It is a transformer (Vaswani et al., 2017) language model trained on commonsense knowledge graphs like ATOMIC (Sap et al., 2019a) and ConceptNet (Speer et al., 2017) with the objective of predicting the object (i.e., commonsense inference) with the subject (i.e., natural text) and relation as input. For example, given a commonsense tuple (s="take a nap", r=Causes, o="have energy"), the neural commonsense knowledge model is trained to generate  $o$  given  $s$  and  $r$  as inputs. After training, it can gen
+
+erate accurate, representative knowledge for new, unseen entities and events.
+
+To extract commonsense knowledge stored in a neural commonsense knowledge model, we use a natural sentence as the subject  $s$  (e.g., he wants to cook a meal) and concatenate it with a randomly selected commonsense relation  $r$  (e.g., xNeed) from a pre-defined set to form a prompt (e.g., he wants to cook a meal xNeed). We then feed the prompt to the neural commonsense knowledge model and use it to generate a commonsense inference (e.g., to buy ingredients). In this way, the commonsense knowledge generation process resembles the way in which the neural commonsense knowledge model is trained. As such, we can get commonsense inferences of relatively high qualities.
+
+Using a neural commonsense knowledge model as a knowledge source has two advantages. On one hand, compared to the previous method (Li et al., 2019) using a symbolic commonsense knowledge graph, a neural commonsense knowledge model can generalize to unseen subjects, thus enabling us to refine the target pre-trained model on large-scale natural text corpus together with its commonsense inferences. As such, the resulting model can better adapt to downstream tasks which are formulated in diverse natural texts. On the other hand, compared to another method (Zhou et al., 2021) that only uses plain text and is thus limited to the surface form of naturally occurring text, the use of a neural commonsense knowledge model provides much denser commonsense knowledge including a diverse set of commonsense relations between natural texts and the underlying commonsense knowledge.
+
+# 2.2 Commonsense Knowledge Injection
+
+After commonsense knowledge extraction, we need to inject the extracted commonsense knowledge into the target model. A straightforward solution is to use sequence-level knowledge distillation (Kim and Rush, 2016) and continually train the student to generate retrieved commonsense inference given the original text and commonsense relation. However, this can be sub-optimal due to the domain discrepancy between commonsense knowledge and natural text, which introduces the catastrophic forgetting problem (Kirkpatrick et al., 2017) and hurts the performance on downstream tasks, which is also recently confirmed by Cui and Chen (2021).
+
+To better inject the extracted commonsense knowledge into a pre-trained model without suf
+
+![](images/993820ac877f536970a3d0d6a6528f62eadeae6df0bc3dd649e531b33249fb65.jpg)  
+Figure 2: Illustration of the commonsense text infilling objective. Given a commonsense tuple constructed in the commonsense knowledge retrieval phase, we randomly mask text spans in the commonsense tuple following different patterns and train the pre-trained model to reconstruct the masked spans. RL: Should the <mask> be "xNeed"?
+
+fering from catastrophic forgetting so that its capability on general NLP tasks is retained (or even improved), we propose two commonsense-related self-supervised objectives: commonsense text infilling and commonsense relation prediction. The former objective is generative while the latter is a discriminative objective. We refine the pre-trained model by multi-tasking on both objectives so that the model can better adapt to tasks requiring either generative or discriminative commonsense reasoning ability.
+
+Commonsense Text Infilling Commonsense text infilling is a simple extension to the conventional text infilling objective used for pre-training BART and T5. It transforms each sentence to a commonsense tuple similar to that in a commonsense knowledge graph by appending the commonsense relation and the generated commonsense inference. We then mask text spans in the commonsense tuple by randomly selecting one masking scheme among text masking, commonsense masking, bidirectional masking, and relation masking. As illustrated in Fig 2, these masking strategies selectively mask different components in the input commonsense tuple and lead to different optimization objectives. Specifically, these masking schemes mask either spans in natural text  $(\mathrm{P}(s|\tilde{s},r,o))$ , commonsense inference  $(\mathrm{P}(o|s,r,\tilde{o}))$ , natural text/commonsense inference  $(\mathrm{P}(s,o|\tilde{s},r,\tilde{o}))$ , or commonsense relation  $(\mathrm{P}(r|s,\tilde{r},o))$ , respectively. We then train the model to predict the masked spans autoregressively. The diverse masking strategies provide more diverse training signals compared to random masking, thus
+
+![](images/5037654920a72b9799f8ae40dfb64768aae05d0b9d714a9457bdc51ecb09e694.jpg)  
+Figure 3: Illustration of the commonsense relation prediction objective. We train the pre-trained model to predict the correct commonsense inference given the subject and relation from three distractors generated with either different subjects or relations as inputs.
+
+enabling the model to better align the surface form of human language and the underlying commonsense knowledge.
+
+In addition, unlike the conventional practice in masked span infilling objective that randomly masks text spans with the same probability, we propose to mask text spans including concepts (tokens recognized as nouns or verbs by a Spacy POS tagger) with a higher probability so that the model will be trained to predict concepts more frequently compared to non-content words that are generally not related to commonsense reasoning.
+
+Commonsense Relation Prediction While the commonsense text infilling objective encourages the pre-trained model to align natural texts and their commonsense inferences, it is always trained on valid commonsense tuples. This can be suboptimal because we also want the model to be capable of discriminating invalid commonsense inferences, which is important for many commonsense-related downstream tasks.
+
+To this end, we introduce a commonsense relation prediction task that trains the model to distinguish the correct commonsense inference corresponding to the input sentence and the commonsense relation from distractors. To be specific, the commonsense relation prediction objective is formulated as a multi-choice QA problem with an input sentence as the context, a commonsense relation as the question, and a set of four commonsense inferences as options. The set of options consists of one correct commonsense inference, which is generated by the neural commonsense model with the input sentence and commonsense relation as input, and three carefully curated distractors (i.e., negative examples) generated by the
+
+<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALQA</td><td>COPA</td></tr><tr><td>BERT-base</td><td>53.08(±0.16)</td><td>57.60(±0.8)</td><td>64.86(±0.52)</td><td>61.88(±0.56)</td><td>64.3(±0.4)</td><td>67.3(±0.4)</td></tr><tr><td>ERNIE-base</td><td>54.06(±0.12)</td><td>58.90(±0.9)</td><td>66.47(±0.58)</td><td>63.04(±0.46)</td><td>65.1(±0.4)</td><td>68.9(±0.4)</td></tr><tr><td>KnowBERT</td><td>53.88(±0.15)</td><td>58.50(±0.8)</td><td>66.61(±0.63)</td><td>63.18(±0.52)</td><td>65.4(±0.5)</td><td>69.4 (±0.4)</td></tr><tr><td>COMET</td><td>45.32(±0.28)</td><td>51.20(±1.1)</td><td>60.73(±0.51)</td><td>57.63(±0.61)</td><td>60.2(±0.7)</td><td>69.1 (±0.5)</td></tr><tr><td>T5-base</td><td>61.88(±0.08)</td><td>58.20(±1.0)</td><td>68.14(±0.73)</td><td>61.10(±0.38)</td><td>65.1(±0.5)</td><td>71.4 (±0.7)</td></tr><tr><td>T5-base + TI</td><td>62.05(±0.17)</td><td>58.43(±0.8)</td><td>68.32(±0.66)</td><td>61.42(±0.32)</td><td>65.3(±0.4)</td><td>71.8 (±0.8)</td></tr><tr><td>T5-base + SSM</td><td>62.37(±0.25)</td><td>58.60(±0.9)</td><td>68.48(±0.65)</td><td>61.57(±0.44)</td><td>65.5(±0.5)</td><td>72.1 (±0.6)</td></tr><tr><td>T5-base + KD</td><td>61.83(±0.42)</td><td>56.54(±0.7)</td><td>67.35(±0.63)</td><td>60.94(±0.66)</td><td>64.8(±0.5)</td><td>71.0 (±1.0)</td></tr><tr><td>T5-base + CSKG (TI)</td><td>60.22(±0.40)</td><td>56.17(±0.8)</td><td>66.51(±0.57)</td><td>59.92(±0.47)</td><td>62.7(±0.7)</td><td>68.5 (±1.1)</td></tr><tr><td>T5-base + CSKG (Rule)</td><td>63.10(±0.35)</td><td>57.97(±0.8)</td><td>68.27(±0.71)</td><td>60.15(±0.51)</td><td>65.7(±0.4)</td><td>72.4 (±0.9)</td></tr><tr><td>CALM</td><td>63.32(±0.35)</td><td>60.90(±0.4)</td><td>71.01(±0.61)</td><td>63.20(±0.52)</td><td>66.0(±0.5)</td><td>72.2 (±0.8)</td></tr><tr><td>CKT-base</td><td>64.11(±0.31)</td><td>61.58(±0.5)</td><td>72.26(±0.61)</td><td>64.37(±0.49)</td><td>67.3(±0.4)</td><td>73.4 (±0.5)</td></tr><tr><td>CKT w/ GPT-2</td><td>60.39(±0.61)</td><td>56.95(±0.7)</td><td>68.48(±0.44)</td><td>60.14(±0.52)</td><td>66.2(±0.6)</td><td>72.8 (±1.0)</td></tr></table>
+
+Table 1: Experimental results on base-size models. Best models are bold and second best ones are underlined within each metric. Mean and standard deviation of 3 different runs with different random seeds are reported.
+
+same neural commonsense knowledge model with different inputs. As illustrated in Figure 3, among the three distractors, one is generated with an input composed of the same sentence and a different commonsense relation, and another two are generated with an input composed of different sentences with the same commonsense relation. In this way, the model learns to align the natural texts with valid commonsense knowledge while also distinguishing commonsense inferences that do not make sense. Moreover, this objective is formulated as a multi-choice QA task that closely resembles several downstream commonsense-related tasks such as CommonsenseQA and SOCIALIQA, thus enabling easier transfer especially when labeled training examples are scarce.
+
+# 3 Experiments
+
+# 3.1 Experimental Settings
+
+Models In our experiments we apply commonsense knowledge transfer to refine T5 (Raffel et al., 2019), a popular model pre-trained with the text infilling objective. We experiment with both T5-base and T5-large, which consist of 220 million and 774 million parameters respectively, as the target model in the commonsense knowledge transfer framework. We do not experiment with extremely large models like T5-11B because of the resource constraints and the fact that these models are hard to deploy in real-world applications. We use COMET-ATOMIC $^{20}$  a state-of-the-art neural commonsense knowledge model that can generate accurate, representative knowledge for new, unseen entities and events, as the source model. It is initialized with BART and continually trained
+
+on  $\mathrm{ATOMIC}_{20}^{20}$  (Hwang et al., 2021), a new general purpose commonsense knowledge graph.
+
+Data We randomly sample a subset consisting of 10 million sentences from the English Wikipedia and the BookCorpus (Zhu et al., 2015), which is used for pre-training BERT and its variants. We select a set of representative commonsense relations including intent, reason, effect, need, want, and react from relations used to train COMET-ATOMIC $_{20}$ . For each sentence, we randomly sample two relations and retrieve the corresponding commonsense explanation from COMET $_{20}$ . We randomly select one relation-explanation pair to form the input example and leave another as the distractor for the commonsense relation prediction objective.
+
+Training We refine the pre-trained models on the self-supervised examples constructed with the sampled 10 million sentences for 100k steps with a batch size of 1024, a maximum sequence length of 256, and a learning rate of  $5\mathrm{e} - 5 / 2\mathrm{e} - 5$  for base-size and large-size models respectively with a linear warm-up for the first 8,000 updates. After knowledge transfer, we fine-tune the models on downstream tasks by formulating the tasks into text-to-text problems. Pre-training and fine-tuning details are included in the Appendix.
+
+Evaluation We evaluate the continual pre-trained models on downstream tasks that require commonsense reasoning including CommonsenseQA (Talmor et al., 2018), OpenbookQA (Mihaylov et al., 2018), PIQA (Bisk et al., 2020), aNLI (Bhagavatula et al., 2020), COPA (Roemmelo et al., 2011), and SOCAI LIQA (Sap et al., 2019b) In addition to the conventional fully supervised setting, we also test our approach in the few-shot setting by vary
+
+<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALQA</td><td>COPA</td></tr><tr><td>T5-large</td><td>69.81(±1.02)</td><td>61.40(±1.0)</td><td>72.19(±1.09)</td><td>75.54(±1.22)</td><td>71.3(±0.8)</td><td>83.6(±1.1)</td></tr><tr><td>CALM-large</td><td>71.31(±0.04)</td><td>66.00(±1.0)</td><td>75.11(±1.65)</td><td>77.12(±0.34)</td><td>72.7(±0.7)</td><td>84.9(±1.0)</td></tr><tr><td>CKT-large</td><td>72.15(±0.61)</td><td>66.70(±1.1)</td><td>76.07(±0.95)</td><td>77.94(±0.59)</td><td>73.8(±0.8)</td><td>86.0(±1.2)</td></tr></table>
+
+Table 2: Experimental results on large-size models. Best models are bold and second best ones are underlined within each metric. Mean and variance of 3 different runs with different random seeds are reported.
+
+ing the percentage of labeled examples from the original training set used for fine-tuning. The idea is that limited labeled examples can only help the model understand the task but are insufficient for the model to acquire enough commonsense knowledge to solve the task. As such, it requires the model to store enough commonsense knowledge in its parameters to succeed in the few-shot setting. For both the settings, we report the results on the official development set and tune the hyperparameters based on the models' performance on an in-house split dev set. We report the mean and variance of 3 individual runs with different random seeds because most datasets are relatively small, which makes the variance in results non-negligible.
+
+Baselines We compare our approach with methods that continually train a pre-trained model with different objectives. We divide the baselines into two categories based on the source of their supervision. The first category includes methods that only exploit general text corpus, including (1) T5 + TI that continually pre-trains the public checkpoint of T5 with the same text infilling objective for more steps, (2) T5 + SSM that also continual pretrains T5 with the text infilling objective, but use salient span masking (Roberts et al., 2020) instead of random masking for data construction, (3) (T5 + KD) that uses sequence-level knowledge distillation (Kim and Rush, 2016) for knowledge transfer, where the student model is trained with the teacher output (i.e.,  $\mathrm{P}(o|s,r)$ ), and (4) CALM (Zhou et al., 2021) that uses novel self-supervised objectives to construct concept-centric self-supervision from general text corpus. The second category instead exploits CSKG, including (5) T5 + CSKG (TI) train T5 with the text infilling objective on tuples in a CSKG, and (6) T5 + CSKG (Rule) (Li et al., 2019) that uses manually defined rules to construct training examples from a CSKG. We also include a COMET baseline where we directly fine-tune the pre-trained COMET-ATOMIC $^{20}$  model for the downstream tasks to verify the necessity of commonsense knowledge transfer, and a CKT w/ GPT
+
+![](images/6696c0db6024ee79cf5541d77e1c0c9dfd3867fc4aae2e9fced1019ae42daf41.jpg)  
+Figure 4: Performance of compared base-size models fine-tuned with different fraction of the datasets.
+
+![](images/2f07e7f98aadcdf747fd79588a408c351d7197a44be95d323eeca04b0d1b01dc.jpg)
+
+2 baseline where the commonsense inferences are generated by a pre-trained GPT-2 large model to verify whether the gain comes from transferring the commonsense knowledge from COMET, or simply from data augmentation from another generative model. For a fair comparison, we use the same data and training steps compared to our approach for baselines from the first category and use  $\mathrm{ATOMIC}_{20}^{20}$ , on which the teacher model in our framework is pre-train on, as the commonsense knowledge graph. For reference, we also include some popular knowledge-enhanced pre-trained models including ERNIE (Zhang et al., 2019) and KnowBERT (Peters et al., 2019).
+
+# 3.2 Fully-supervised Results
+
+We first present results in the fully-supervised setting. Results on base-size models are presented in Table 1. We can see that our approach yields significant improvement compared to the T5 baseline (up to 4 absolute scores) and consistently outperform CALM, the state-of-the-art method of injecting commonsense knowledge into PTLMs.
+
+In addition, we observe that simply using continual training with the original text-infilling objective or its variant with salient span masking only marginally improves the performance. Surprisingly, training with text infilling on a commonsense knowledge graph leads to degraded performance compared to the T5 baseline. We suspect this is because the commonsense tuples in commonsense knowledge graphs are generally too short
+
+<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SIQA</td><td>COPA</td></tr><tr><td>T5-base</td><td>61.88</td><td>58.20</td><td>68.14</td><td>61.10</td><td>65.1</td><td>71.4</td></tr><tr><td>CKT-base</td><td>64.57</td><td>62.77</td><td>73.26</td><td>64.75</td><td>68.3</td><td>73.4</td></tr><tr><td colspan="7">Objective Analysis</td></tr><tr><td>CKT-base w/o CSTI</td><td>62.58</td><td>60.97</td><td>70.61</td><td>62.11</td><td>66.5</td><td>72.0</td></tr><tr><td>CKT-base w/o text masking</td><td>62.98</td><td>61.74</td><td>72.55</td><td>63.81</td><td>67.7</td><td>72.8</td></tr><tr><td>CKT-base w/o commonsense masking</td><td>63.61</td><td>62.03</td><td>72.83</td><td>64.40</td><td>67.5</td><td>72.7</td></tr><tr><td>CKT-base w/o bidirectional masking</td><td>63.52</td><td>62.11</td><td>72.30</td><td>64.24</td><td>67.6</td><td>72.9</td></tr><tr><td>CKT-base w/o relation masking</td><td>64.12</td><td>62.48</td><td>73.31</td><td>64.57</td><td>67.4</td><td>72.7</td></tr><tr><td>CKT-base w/o CSRIP</td><td>63.12</td><td>62.07</td><td>72.44</td><td>64.11</td><td>67.5</td><td>72.6</td></tr><tr><td>CKT-base w/ random distractors</td><td>64.04</td><td>62.29</td><td>72.95</td><td>64.48</td><td>68.0</td><td>73.1</td></tr><tr><td colspan="7">Multi-task versus Sequential Transfer</td></tr><tr><td>CKT-base (CSTI → CSRIP)</td><td>64.69</td><td>62.51</td><td>73.35</td><td>64.11</td><td>67.9</td><td>73.5</td></tr><tr><td>CKT-base (CSRIP → CSTI)</td><td>63.49</td><td>61.33</td><td>71.54</td><td>63.41</td><td>67.0</td><td>72.0</td></tr><tr><td colspan="7">Corpus Size</td></tr><tr><td>CKT-base w/ 10% data</td><td>64.18</td><td>62.21</td><td>71.86</td><td>64.31</td><td>67.7</td><td>73.1</td></tr><tr><td>CKT-base w/ 50% data</td><td>64.45</td><td>62.66</td><td>73.10</td><td>64.72</td><td>68.2</td><td>73.4</td></tr></table>
+
+Table 3: Analysis of the proposed commonsense knowledge transfer framework. CSTI and CSRP denote the commonsense text infilling objective and the commonsense relation prediction objective, respectively.
+
+and simple, making the pre-trained model unable to reason within relatively long contexts which is crucial for most downstream tasks. Moreover, we find that continually pre-training with training data constructed with commonsense tuples in a commonsense knowledge graph following manually designed rules leads to improvements in certain tasks. However, the improvement is inconsistent across different tasks and it even hurts the performance on certain tasks, which may be because the rules for constructing training data are tailored for certain tasks like CSQA. The inferior performance of using commonsense knowledge graphs as data sources also confirms the need of using natural text corpus during continual pre-training for better adapting to diverse downstream tasks. We also find directly applying sequence-level KD and training the student to mimic the teacher on the commonsense tuple generation task fails to improve the performance because the task is not general enough and thus cannot transfer to diverse downstream tasks well. Moreover, directly fine-tuning COMET or using GPT-2 as the commonsense knowledge source results in very poor performance. This confirms the necessity of commonsense knowledge transfer and shows that it is actually transferring commonsense knowledge instead of simple text augmentation.
+
+To further confirm the effectiveness of commonsense knowledge transfer, we apply it to T5-large and compare it to competitive baselines in the base
+
+size experiments. The results are presented in Table 2. We can see that our approach consistently outperforms T5-large and CALM-large. This suggests that our approach can successfully generalize to large-size pre-trained models.
+
+# 3.3 Few-shot Results
+
+Injecting commonsense knowledge into pre-trained models is important because it enables the model to reason and generalize to unseen examples while observing only a few labeled examples. To this end, we fine-tune the compared models with different fractions of labeled training data to investigate the transition of the behavior of our model and baselines from the low-resource regime to the fully-supervised setting (Fig. 4). We observe that the performance improvement of our approach compared to the baselines is more significant in the low-resource regime. This shows that commonsense knowledge transfer can successfully transfer commonsense knowledge into pre-trained models so that they can generalize well while seeing only a small part of training data. This may also help the model reduce the risk/tendency of fitting the spurious correlations in the annotated datasets and thus generalize better.
+
+# 3.4 Analysis
+
+To better understand the proposed commonsense knowledge transfer framework and the role of its
+
+different components, we conduct an ablation study about the impact of different proposed objectives, the impact of multi-tasking the commonsense-related self-supervised objective versus sequentially training, and the impact of the size of natural text corpus used for transfer (see Table 3).
+
+Impact of Objectives We find that both the proposed objectives contribute to the performance improvement of our approach. The commonsense text infilling objective is shown to be more critical than the commonsense relation prediction task. We suspect this is because commonsense text infilling resembles the vanilla text infilling objective with which the T5 models are pre-trained, thus preventing the model from catastrophic forgetting. In addition, all four masking strategies are beneficial, and their contribution varies for different downstream tasks. This confirms the necessity of a diverse masking scheme. Moreover, our strategy for constructing distractors outperforms the random counterpart, demonstrating the necessity of hard negative examples for the commonsense relation prediction task.
+
+Multi-task versus Sequential Transfer As for the training order between the two objectives, we find that starting from the commonsense text infilling task and then switching to the commonsense relation prediction task performs similarly with our multi-tasking strategy while significantly outperforming its counterpart training with the reverse direction. We think this is because the commonsense text infilling objective resembles the original pre-training while the commonsense relation prediction is more similar to downstream tasks. We opt for the multi-tasking strategy for simplicity.
+
+Impact of Corpus Size We find that commonsense knowledge transfer significantly outperforms both the T5 baseline and the competitive CALM method with only 10 percent of the full data used for distillation. Nevertheless, the performance improvement also confirms that our approach can benefit from the accessibility of large-scale natural texts. For base-size models, the performance improvements seem to saturate after 10 million sentence pairs. However, we anticipate that larger-size models may still benefit from a larger amount of data, and leave this for future work.
+
+# 4 Related Work
+
+Knowledge-augmented Pre-trained Models A number of recent works have examined the prob-
+
+lem of incorporating world knowledge with the pre-trained models. A number of works use an external knowledge base to incorporate entity knowledge with pre-trained models (Zhang et al., 2019; Peters et al., 2019; Wang et al., 2020; Liu et al., 2020). However, these approaches require specialized resources like knowledge bases which are non-trivial to seek, thus limiting the domain they can be applied to. Xiong et al. (2020) proposed a novel entity replacement detection objective that incorporates Wikipedia to encode world knowledge into a BERT-like pre-trained model. He et al. (2020) proposed a generative and discriminative framework that pretrains the model to complete and correct knowledge spans. The aforementioned approaches generally focus on factual knowledge of entities while our work mainly focuses on commonsense knowledge.
+
+Commonsense Reasoning for NLP Several recent studies (Talmor et al., 2018; Sap et al., 2019b; Zhou et al., 2020b; Lin et al., 2020; Xu et al., 2021) evaluate the performance of several pre-trained language models on tasks that require commonsense reasoning and find that it is still very hard for pre-trained language models to match or exceed human-level performance. Therefore, approaches to improve the commonsense reasoning ability of pre-trained language models has attracted much attention. These approaches can be divided into two categories. The first category focuses on incorporating an external commonsense knowledge graph for commonsense reasoning. For example, Lin et al. (2019), Cui and Chen (2021), and Liu et al. (2021) propose to exploit structured symbolic commonsense knowledge graphs to perform commonsense reasoning. The second one instead attempts to inject commonsense knowledge into the parameters of pre-trained models. For example, Ye et al. (2019); Li et al. (2019) proposed to use manually designed rules to construct commonsense related training examples from commonsense knowledge graphs. Zhou et al. (2021) instead only relies on general text corpus and proposed two concept-centric self-supervised objectives to refine pre-trained models with commonsense knowledge.
+
+# 5 Conclusion
+
+We introduce commonsense knowledge transfer, a framework to transfer the commonsense knowledge stored in a neural commonsense knowledge model to a general-purpose pre-trained model. Our method extracts commonsense knowledge from
+
+the source model to construct self-supervised training data for the target model. Empirical results show that our approach consistently outperforms previous methods for improving the commonsense reasoning ability of pre-trained models that exploit either symbolic knowledge graphs or texts alone.
+
+# Limitations
+
+In our experiments, we use T5-base and T5-large models as the target model since they are widely-used, representative pre-trained seq2seq models and use COMET-ATOMIC $^{20}$  as the commonsense knowledge source. However, there are other pretrained seq2seq models such as BART, and neural commonsense models such as COMET that we did not experiment with. Moreover, we only experimented with 10 million randomly sampled sentences from the English Wiki and BookCorpus datasets. It would be interesting to investigate whether continually pre-training with a larger scale dataset can further improve the performance.
+
+# Ethical Considerations
+
+Our work focuses on improving the commonsense reasoning ability of pre-trained language models. It probably does not introduce extra ethical concerns. However, in commonsense knowledge extraction, the neural commonsense knowledge model may generate unexpected (e.g., biased) commonsense inferences, and training with these inferences may lead to additional bias in the pre-trained model. Nevertheless, all pre-trained language models contain bias and should be examined.
+
+# References
+
+Emily M. Bender and Alexander Koller. 2020. Climbing towards NLU: on meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, ACL 2020, Online, July 5-10, 2020, pages 5185-5198. Association for Computational Linguistics.  
+Chandra Bhagavatula, Ronan Le Bras, Chaitanya Malaviya, Keisuke Sakaguchi, Ari Holtzman, Hannah Rashkin, Doug Downey, Scott Wen-tau Yih, and Yejin Choi. 2020. Abductive commonsense reasoning. In ICLR.  
+Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. 2020. Piqa: Reasoning about physical commonsense in natural language. In Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI).
+
+Antoine Bosselut, Hannah Rashkin, Maarten Sap, Chaitanya Malaviya, Asli Celikyilmaz, and Yejin Choi. 2019. COMET: commonsense transformers for automatic knowledge graph construction. In ACL (1), pages 4762-4779. Association for Computational Linguistics.  
+Alexis Conneau and Douwe Kiela. 2018. Senteval: An evaluation toolkit for universal sentence representations. In LREC.  
+Wanyun Cui and Xingran Chen. 2021. Enhancing language models with plug-and-play large-scale commonsense. CoRR, abs/2109.02572.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: pre-training of deep bidirectional transformers for language understanding. In *NAACL-HLT* (1), pages 4171–4186. Association for Computational Linguistics.  
+William B. Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In IWP@IJCNLP.  
+WA Falcon. 2019. Pytorch lightning. GitHub. Note: https://github.com/PyTorchLightning/pytorchlightning, 3.  
+Bin He, Xin Jiang, Jinghui Xiao, and Qun Liu. 2020. Kgplm: Knowledge-guided language model pretraining via generative and discriminative learning. CoRR, abs/2012.03551.  
+John Hewitt and Christopher D. Manning. 2019. A structural probe for finding syntax in word representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 4129-4138. Association for Computational Linguistics.  
+Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the knowledge in a neural network. CoRR, abs/1503.02531.  
+Pedram Hosseini, David A. Broniatowski, and Mona T. Diab. 2021. Commonsense knowledge-augmented pretrained language models for causal reasoning classification. CoRR, abs/2112.08615.  
+Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2021. (comet-) atomic 2020: On symbolic and neural commonsense knowledge graphs. In AAAI, pages 6384-6392. AAAI Press.  
+Yoon Kim and Alexander M. Rush. 2016. Sequence-level knowledge distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pages 1317-1327. The Association for Computational Linguistics.
+
+James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526.  
+Tassilo Klein and Moin Nabi. 2021. Towards zero-shot commonsense reasoning with self-supervised refinement of language models. CoRR, abs/2109.05105.  
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: denoising sequence-to-sequence pretraining for natural language generation, translation, and comprehension. In ACL, pages 7871-7880. Association for Computational Linguistics.  
+Shiyang Li, Jianshu Chen, and Dian Yu. 2019. Teaching pretrained models with commonsense reasoning: A preliminary kb-based approach. CoRR, abs/1909.09743.  
+Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. _Kagnet: Knowledge-aware graph networks for commonsense reasoning_. In _EMNLP/IJCNLP (1)_, pages 2829-2839. 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, Online Event, 16-20 November 2020, volume EMNLP 2020 of *Findings of ACL*, pages 1823-1840. Association for Computational Linguistics.  
+Ye Liu, Yao Wan, Lifang He, Hao Peng, and Philip S. Yu. 2020. Kg-bart: Knowledge graph-augmented bart for generative commonsense reasoning.  
+Ye Liu, Yao Wan, Lifang He, Hao Peng, and Philip S. Yu. 2021. KG-BART: knowledge graph-augmented BART for generative commonsense reasoning. In AAAI, pages 6418-6425. AAAI Press.  
+Christopher D. Manning, Kevin Clark, John Hewitt, Urvashi Khandelwal, and Omer Levy. 2020. Emergent linguistic structure in artificial neural networks trained by self-supervision. Proc. Natl. Acad. Sci. USA, 117(48):30046-30054.  
+William Merrill, Yoav Goldberg, Roy Schwartz, and Noah A. Smith. 2021. Provable limitations of acquiring meaning from ungrounded form: What will future language models understand? CoRR, abs/2104.10809.  
+Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. 2018. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789.
+
+Matthew E Peters, Mark Neumann, Robert L Logan IV, Roy Schwartz, Vidur Joshi, Sameer Singh, and Noah A Smith. 2019. Knowledge enhanced contextual word representations. arXiv preprint arXiv:1909.04164.  
+Fabio Petroni, Tim Roktaschel, Sebastian Riedel, Patrick S. H. Lewis, Anton Bakhtin, Yuxiang Wu, and Alexander H. Miller. 2019. Language models as knowledge bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 2463-2473. Association for Computational Linguistics.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv preprint arXiv:1910.10683.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,  $000+$  questions for machine comprehension of text. In EMNLP.  
+Adam Roberts, Colin Raffel, and Noam Shazeer. 2020. How much knowledge can you pack into the parameters of a language model? In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 5418-5426. Association for Computational Linguistics.  
+Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. 2011. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In *Logical Formalizations of Commonsense Reasoning*, Papers from the 2011 AAAI Spring Symposium, Technical Report SS-11-06, Stanford, California, USA, March 21-23, 2011. AAAI.  
+Maarten Sap, Ronan Le Bras, Emily Allaway, Chandra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A. Smith, and Yejin Choi. 2019a. ATOMIC: an atlas of machine commonsense for if-then reasoning. In AAAI, pages 3027-3035. AAAI Press.  
+Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. 2019b. Socialiaq: Commonsense reasoning about social interactions. In EMNLP.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP.
+
+Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In AAAI, pages 4444-4451. AAAI Press.  
+Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2018. Commonsenseqa: A question answering challenge targeting commonsense knowledge. arXiv preprint arXiv:1811.00937.  
+Lifu Tu, Garima Lalwani, Spandana Gella, and He He. 2020. An empirical study on robustness to spurious correlations using pre-trained language models. Trans. Assoc. Comput. Linguistics, 8:621-633.  
+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 NIPS, pages 5998-6008.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. 2019. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In ICLR.  
+Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuanjing Huang, Cuihong Cao, Daxin Jiang, Ming Zhou, et al. 2020. K-adapter: Infusing knowledge into pre-trained models with adapters. arXiv preprint arXiv:2002.01808.  
+Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. TACL.  
+Adina Williams, Nikita Nangia, and Samuel R. Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In *NAACL-HLT*.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. 2019. Huggingface's transformers: State-of-the-art natural language processing. ArXiv, abs/1910.03771.  
+Wenhan Xiong, Jingfei Du, William Yang Wang, and Veselin Stoyanov. 2020. Pretrained encyclopedia: Weakly supervised knowledge-pretrained language model. In International Conference on Learning Representations.  
+Canwen Xu, Wangchunshu Zhou, Tao Ge, Ke Xu, Julian J. McAuley, and Furu Wei. 2021. Blow the dog whistle: A chinese dataset for cant understanding with common sense and world knowledge. In NAACL-HLT, pages 2139-2145. Association for Computational Linguistics.
+
+Zhi-Xiu Ye, Qian Chen, Wen Wang, and Zhen-Hua Ling. 2019. Align, mask and select: A simple method for incorporating commonsense knowledge into language representation models. CoRR, abs/1908.06725.  
+Zhengyan Zhang, Xu Han, Zhiyuan Liu, Xin Jiang, Maosong Sun, and Qun Liu. 2019. Ernie: Enhanced language representation with informative entities. arXiv preprint arXiv:1905.07129.  
+Wangchunshu Zhou, Dong-Ho Lee, Ravi Kiran Selvam, Seyeon Lee, and Xiang Ren. 2021. Pretraining text-to-text transformers for concept-centric common sense. In ICLR. OpenReview.net.  
+Xuhui Zhou, Yue Zhang, Leyang Cui, and Dandan Huang. 2020a. Evaluating commonsense in pretrained language models. In AAAI, pages 9733-9740. AAAI Press.  
+Xuhui Zhou, Yue Zhang, Leyang Cui, and Dandan Huang. 2020b. Evaluating commonsense in pretrained language models. In AAAI, pages 9733-9740.  
+Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. 2015. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 19-27. IEEE Computer Society.
+
+# A Pre-training and Fine-tuning Details
+
+# A.1 Pre-Training Details
+
+We implement our models using Pytorchlightning (Falcon, 2019) and Hugginface's Pytorch Transformers (Wolf et al., 2019). For pre-training phase, we use the AdamW optimizer with maximum sequence length 256, train batch size 8, gradient accumulation 8, warmup steps 8000, weight decay 0.01 and adam epsilon 1e-6. We train the models with 8 V100 GPUs and FP32 precision. The model is pre-trained for 10 epochs. We searched for the best learning rate for our model out of [5e-6, 2e-5, 5e-5, 1e-4].
+
+# A.2 Fine-Tuning Details
+
+For fine-tuning, we use 4 V100 GPUs and use FP32. For all tasks, we use the AdamW optimizer with learning rate from [1e-5, 2e-5, 5e-5, 1e-4, 2e-4], maximum sequence length 256, batch size from [4, 8, 16, 32]. For all tasks, we use a warmup fraction of 0.01, and max epoch of 20.
+
+# B Additional Analysis
+
+# B.1 Qualitative Analysis
+
+To better understand the proposed method, we present a case study in Figure 5. We can see that both the objectives introduced in the CALM model and the salient span masking (SSM) strategy fail to exploit the underlying commonsense rationale beyond the surface form of texts while our approach directly aligns texts with the corresponding commonsense inferences with different commonsense relations. That explains why commonsense knowledge transfer can effectively improve a pre-trained model's performance on downstream tasks requiring commonsense reasoning ability.
+
+# B.2 Experimental Results on GLUE
+
+To verify that commonsense knowledge transfer is suitable for general-purpose pre-trained models, we fine-tune our model on the GLUE benchmark (Wang et al., 2019). Specifically, we test on MRPC (Dolan and Brockett, 2005),  $\mathrm{QQP}^1$  and STS-B (Conneau and Kiela, 2018) for Paraphrase Similarity Matching; SST-2 (Socher et al., 2013) for Sentiment Classification; MNLI (Williams et al., 2018), QNLI (Rajpurkar et al., 2016) and RTE (Wang et al., 2019) for the Natural Language
+
+![](images/a89e129b49bc350c147d63748da638dfdbafe7832ddb7d8c5efd8ff431cf4eae.jpg)  
+Commonsense Text Infilling  
+Figure 5: Performance on the CSQA dataset w.r.t. the size of training data used for commonsense knowledge transfer.
+
+![](images/3ef0bfdea7718e704a99b95df607d6ba624dbef9d1783bb645dbb79882ee4442.jpg)  
+Figure 6: Performance on the CSQA dataset w.r.t. the size of training data used for commonsense knowledge transfer.
+
+Inference; CoLA (Warstadt et al., 2019) for Linguistic Acceptability.
+
+The results are shown on Table 4, we can see that after commonsense knowledge transfer, the resulting model's general natural language understanding ability is comparable with the original T5-base model. This shows that our approach does not affect the model's general transfer ability and thus can be applied to general-purpose language models.
+
+# B.3 Experiments with BART
+
+To demonstrate the versatility of commonsense knowledge transfer for different backbones, we conduct additional experiments using BART as the backbone model. The results are shown in Table 5. We can see that commonsense knowledge trans
+
+<table><tr><td>Methods</td><td>CoLA</td><td>MNLI</td><td>MRPC</td><td>QNLI</td><td>QQP</td><td>RTE</td><td>SST-2</td><td>SST-B</td><td>Meta Score</td></tr><tr><td>BERT-base</td><td>58.9</td><td>84.7</td><td>89.6</td><td>91.2</td><td>90.0</td><td>71.4</td><td>93.0</td><td>90.0</td><td>83.6</td></tr><tr><td>T5-base</td><td>55.9</td><td>84.5</td><td>90.3</td><td>90.5</td><td>90.2</td><td>76.2</td><td>92.8</td><td>87.8</td><td>83.5</td></tr><tr><td>CKT-base</td><td>57.4</td><td>84.4</td><td>90.6</td><td>90.9</td><td>89.9</td><td>76.8</td><td>92.5</td><td>88.4</td><td>83.9</td></tr></table>
+
+Table 4: Experimental results of base-size models on the GLUE benchmark.  
+
+<table><tr><td>Methods</td><td>CSQA</td><td>OBQA</td><td>PIQA</td><td>aNLI</td><td>SOCIALIQA</td><td>COPA</td></tr><tr><td>BART</td><td>72.31</td><td>65.80</td><td>74.12</td><td>78.27</td><td>71.6</td><td>85.6</td></tr><tr><td>CKT-BART</td><td>73.14</td><td>68.20</td><td>76.95</td><td>79.52</td><td>73.3</td><td>87.2</td></tr></table>
+
+fer also consistently improves the BART model, demonstrating the versatility of our approach.
+
+Table 5: Experimental results (mean of 3 random runs) with BART.  
+
+<table><tr><td>Methods</td><td>BLEU-4</td><td>METEOR</td><td>CIDEr</td><td>SPICE</td></tr><tr><td>T5-base</td><td>24.90</td><td>31.20</td><td>12.99</td><td>32.40</td></tr><tr><td>CALM-base</td><td>26.40</td><td>31.40</td><td>13.88</td><td>33.00</td></tr><tr><td>CKT-base</td><td>26.20</td><td>31.40</td><td>13.65</td><td>33.10</td></tr></table>
+
+Table 6: Experimental results (mean of 3 random runs) with BART.
+
+# B.4 Experiments on CommonGEN
+
+We also experiment on the CommonGEN dataset, a generative commonsense reasoning dataset where the model is required to take several keywords as inputs and output a sentence that makes sense. The results are shown in Table 6. We can see that our approach performs similarly with the CALM model, which includes the CommonGEN task objective as one of the pre-training tasks.
+
+# B.5 Impact of Pre-training Data Size
+
+We also conduct experiments to investigate the sample-efficiency of commonsense knowledge transfer. We present the trend of performance improvement in Figure 6. We can see that our method achieves significant performance improvement upon the T5 baseline with only  $10\%$  of the total training data, which confirms the sample-efficiency of commonsense knowledge transfer.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? limitation  
+A2. Did you discuss any potential risks of your work? ethical statement  
+A3. Do the abstract and introduction summarize the paper's main claims? abstract and introduction  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+experiment
+
+B1. Did you cite the creators of artifacts you used? experiment  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? they're commonly used datasets  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? they're commonly used datasets  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? they're commonly used datasets  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? they're commonly used datasets  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. they're commonly used datasets
+
+C Did you run computational experiments?
+
+experiment
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? experiment
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? experiment  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? experiment  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? experiment
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Commonsense Knowledge Transfer for Pre-trained Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:edcc18fd69cc7f3f30e00ff89f7843ededed888aa06f4cb4505614fd15b09a60
+size 523421

2023/Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter/6f207301-9f3f-466d-aed1-4bdd052903e2_content_list.json

@@ -0,0 +1,2078 @@
+[
+    {
+        "type": "text",
+        "text": "Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter",
+        "text_level": 1,
+        "bbox": [
+            176,
+            82,
+            821,
+            122
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yi Liu $^{1}$ , Xiaohan Bi $^{2}$ , Lei Li $^{1}$ , Sishuo Chen $^{2}$ , Wenkai Yang $^{2}$ , Xu Sun $^{1}$",
+        "bbox": [
+            208,
+            129,
+            794,
+            147
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University",
+        "bbox": [
+            231,
+            148,
+            769,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ Center for Data Science, Peking University",
+        "bbox": [
+            319,
+            181,
+            680,
+            198
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "yliu.pku@outlook.com {bxh, wkyang}  $@$ stu.pku.edu.cn",
+        "bbox": [
+            250,
+            198,
+            747,
+            214
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "nlp.lilei@gmail.com {chensishuo,xusun}@pku.edu.cn",
+        "bbox": [
+            248,
+            215,
+            754,
+            230
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Federated Multilingual Neural Machine Translation (Fed-MNMT) has emerged as a promising paradigm for institutions with limited language resources. This approach allows multiple institutions to act as clients and train a unified model through model synchronization, rather than collecting sensitive data for centralized training. This significantly reduces the cost of corpus collection and preserves data privacy. However, as pre-trained language models (PLMs) continue to increase in size, the communication cost for transmitting parameters during synchronization has become a training speed bottleneck. In this paper, we propose a communication-efficient Fed-MNMT framework that addresses this issue by keeping PLMs frozen and only transferring lightweight adapter modules between clients. Since different language pairs exhibit substantial discrepancies in data distributions, adapter parameters of clients may conflict with each other. To tackle this, we explore various clustering strategies to group parameters for integration and mitigate the negative effects of conflicting parameters. Experimental results demonstrate that our framework reduces communication cost by over  $98\\%$  while achieving similar or even better performance compared to competitive baselines. Further analysis reveals that clustering strategies effectively solve the problem of linguistic discrepancy and pruning adapter modules further improves communication efficiency. $^{1}$",
+        "bbox": [
+            141,
+            279,
+            460,
+            750
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            760,
+            258,
+            776
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Federated Learning (FL) (McMahan et al., 2017) provides a new training framework utilizing data from various clients without privacy leakage. In FL, the server receives models from clients trained with their local data and aggregates all parameters it has received to acquire a global model, and then",
+        "bbox": [
+            112,
+            787,
+            487,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/b83316396c9a8fccdd2a41716eee35fb26db74dfc06d750e251f52a066013b43.jpg",
+        "image_caption": [
+            "Figure 1: Estimated transfer time (bar chart) and the approximate ratio of transfer time to total training time (line chart) in one round for a low-resource client. Our method significantly reduces communication cost and improve training efficiency."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            252,
+            878,
+            444
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "sends it back to all clients to start the next training round. This characteristic enables FL to get widely applied in real-world scenarios (Ge et al., 2020; Roosta et al., 2021; Passban et al., 2022; Niu and Deng, 2022). In recent years, federated multilingual neural machine translation (Fed-MNMT) has become a new training paradigm and making it feasible for most institutions to train MNMT models (Roosta et al., 2021; Passban et al., 2022). FL makes it possible to leverage corpora from other organizations without privacy problems, solving the problem that training an MNMT model needs to collect large-scale multilingual corpora, which is expensive, time-consuming, and often unaffordable for resource-constrained institutions. Therefore, Fed-MNMT is a secure and cost-effective alternative to conventional centralized training for the optimization of MNMT models.",
+        "bbox": [
+            507,
+            561,
+            882,
+            851
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "However, the issue of communication cost is non-negligible when we introduce FL to neural machine translation. Unlike local centralized learning, federated learning requires frequent commu",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "1Our code is available at https://github.com/lancopku/FedMNMT",
+        "bbox": [
+            112,
+            891,
+            487,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "5315",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 5315-5328",
+        "bbox": [
+            228,
+            945,
+            768,
+            958
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            295,
+            958,
+            699,
+            971
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/654762b00028cd6b7dcdd2ec6fe8445b2d6a360c78d3b69f6aee9d16af880817.jpg",
+        "image_caption": [
+            "Figure 2: Communication-efficient Fed-MNMT framework with adapter modules and clients clustering strategies."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            181,
+            80,
+            356,
+            256
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/0419c8fa556e0460cb1b241bd5da9340181a96ee0db59e4460da29ddcdc70bf9.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            359,
+            78,
+            823,
+            256
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "nication of model parameters between the server and clients. Therefore, the communication cost grows rapidly along with the increase in model size. Nowadays, pre-trained language models are widely adopted as backbone models for MNMT, whose parameters are usually over  $10^{8}$ , e.g., 611M for mBART-50 (Tang et al., 2020) and 1.2B for M2M100 (Fan et al., 2020). Considering the increasing number of clients in realistic scenarios as there frequently appear new clients, communication costs will severely hinder the efficient training of the entire Fed-MNMT system and thus make the application of FL to MNMT impractical.",
+        "bbox": [
+            112,
+            306,
+            489,
+            514
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To tackle this problem, we introduce the parameter-efficient tuning idea (Houlsby et al., 2019; Pfeiffer et al., 2021; Karimi Mahabadi et al., 2021) into Fed-MNMT. Specifically, we focus on adapter (Rebuffi et al., 2017; Houlsby et al., 2019), which is a popular technique for efficient tuning that requires only updating a lightweight adapter module. In the training process of adapters, a number of randomly initialized modules are inserted into the backbone models and fine-tuned on new data. Concretely, only the parameters of these modules are updated during training, so the number of parameters needed to be transferred between the server and clients is substantially reduced. As the communication cost before and after introducing adapter illustrated in Figure 1, this approach significantly saves communication costs and enables practical applications of Fed-MNMT.",
+        "bbox": [
+            112,
+            516,
+            489,
+            804
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "However, directly adding adapter modules to NMT models results in a performance decline, which is initially observed by Roosta et al. (2021) and also confirmed by our experimental results. This phenomenon is attributed to the divergence of different language pairs. In Fed-MNMT, corpora in diverse languages from different clients",
+        "bbox": [
+            112,
+            806,
+            489,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "are not independently and identically distributed (Non-I.I.D.), so directly aggregating parameters from clients leads to a decrease in model's performance (Zhao et al., 2018).",
+        "bbox": [
+            507,
+            306,
+            884,
+            369
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Considering the adverse effect of conflicting parameters from diverse languages in Fed-MNMT, we introduce clustering strategies to alleviate this issue. The core idea is to cluster the samples according to the characteristics of their data and only conduct aggregation within each cluster where samples share similar properties. Specifically, we cluster all clients with different language pairs based on the language family, gradient similarity, and random, respectively, and systematically compare the performance of different clustering strategies on multilingual translation benchmarks. Figure 2 gives a general view of our training framework. Our experimental results show that clustering on adapters alleviates the data Non-I.I.D. problem and yields better performance in most cases. Overall, our work opens a new direction for future improvements on Fed-MNMT in the real world.",
+        "bbox": [
+            507,
+            370,
+            884,
+            658
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In conclusion, our primary contributions can be summarized as follows:",
+        "bbox": [
+            507,
+            659,
+            882,
+            690
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Aware of the communication barrier in the training of Fed-MNMT models, we introduce a practical efficient Fed-MNMT framework to enable real-world applications.",
+            "- By exploring adapter and clustering strategies for alleviating the undesirable effect of data discrepancy, we achieve comparable results with over  $98\\%$  communication cost reduced compared to vanilla Fed-MNMT."
+        ],
+        "bbox": [
+            531,
+            697,
+            884,
+            850
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Methodology",
+        "text_level": 1,
+        "bbox": [
+            509,
+            862,
+            658,
+            879
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this section, we first define the Fed-MNMT problem in § 2.1. Next, we elaborate on the adapter",
+        "bbox": [
+            507,
+            887,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "5316",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "modules and the investigated clustering strategies in  $\\S 2.2$  and  $\\S 2.3$ , respectively. Last, we provide an analysis of communication costs between the original Fed-MNMT and our method in  $\\S 2.4$ .",
+        "bbox": [
+            112,
+            84,
+            487,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.1 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            160,
+            331,
+            174
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For a Fed-MNMT problem, we suppose that the set of clients is  $\\{C_i\\}_{i=1}^N$  where  $N > 1$  and client  $C_i$  owns only one language pair  $P_i$ , whose source language and target language are  $src_i$  and  $tgt_i$ , respectively, and corresponding dataset  $D_i = \\{x_{ij}, y_{ij}\\}_{j=1}^{n_i}$ , where  $n_i$  is the size of  $D_i$ . In each training round, the optimization target for  $C_i$  is minimizing the cross entropy loss between the ground truths  $y_i$  and model's predictions  $\\tilde{y}_i$ :",
+        "bbox": [
+            112,
+            181,
+            489,
+            326
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {L} _ {i} = - \\sum_ {j = 1} ^ {n _ {i}} \\sum_ {k = 1} ^ {l _ {i j}} \\log p \\left(\\tilde {y} _ {i j} ^ {k} = y _ {i j} ^ {k} \\mid x _ {i j}\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            159,
+            338,
+            487,
+            385
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "After the  $t$ -th training round, all clients will deliver local parameters to the server. The server will aggregate these parameters to obtain the initial parameters for the next round's training. A commonly adopted aggregation algorithm is FedAvg (McMahan et al., 2017), where the weighted average of clients' parameters is calculated according to the quantities of local data samples. Let  $\\Theta$  denote model parameters, the FedAvg algorithm can be formulated as:",
+        "bbox": [
+            112,
+            397,
+            489,
+            557
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Theta^ {t + 1} = \\sum_ {i = 1} ^ {N} \\frac {n _ {i}}{n} \\Theta_ {i} ^ {t}, \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            226,
+            567,
+            487,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $n = \\sum_{i=1}^{N} n_i$ . Then the aggregated parameters will be sent back to all clients to initialize their local models for the next round of training.",
+        "bbox": [
+            112,
+            625,
+            487,
+            674
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, data sizes can vary sharply among low-resource and high-resource languages in FedMNMT and FedAvg cannot deal well with data quantity skew well (Wang et al., 2020). Thus we change FedAvg, calculating the weighted mean of different clients' parameters, to directly calculating the arithmetic mean of parameters:",
+        "bbox": [
+            112,
+            675,
+            487,
+            788
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\Theta^ {t + 1} = \\sum_ {i = 1} ^ {N} \\frac {1}{N} \\Theta_ {i} ^ {t} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            228,
+            800,
+            487,
+            843
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We refer to this aggregation method as FedMean in our paper.",
+        "bbox": [
+            112,
+            854,
+            487,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Considering the size of pre-trained multi-language models, the communication of model",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "parameters between the server and clients is time-consuming. Inspired by recent progress in parameter efficient tuning, we are interested in whether adapter can be used to improve the efficiency in FL.",
+        "bbox": [
+            507,
+            84,
+            885,
+            149
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.2 Adapter Modules",
+        "text_level": 1,
+        "bbox": [
+            509,
+            161,
+            692,
+            175
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We introduce bottleneck adapter (Houlsby et al., 2019) into pre-trained multilingual models. Following the settings of Houlsby et al. (2019) and Pfeiffer et al. (2020), we add adapter modules after the self-attention layer and feed-forward network (FFN) layer for each encoder layer and an additional adapter layer after the cross-attention layer for each decoder layer. During training, only the parameters of adapters and layer-norm modules will be updated thus only a small proportion of parameters have to be communicated between the server and clients.",
+        "bbox": [
+            507,
+            181,
+            884,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "2.3 Client Clustering Strategies",
+        "text_level": 1,
+        "bbox": [
+            507,
+            388,
+            774,
+            404
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Related research (Johnson et al., 2017; First et al., 2016) has shown that parameter sharing among different languages in MNMT boosts the model's performance, especially for those low-resource languages. Motivated by the success of language clustering in MNMT (Tan et al., 2019), we decide to introduce the method of language pairs clustering into the Fed-MNMT problem and we only allow inner-cluster parameters aggregation. Assuming that the multi-language model consists of an encoder and a decoder, we first conduct a clustering algorithm to obtain the encoder clusters set  $G_{e} = \\{g_{i}\\}_{i=1}^{m_{e}}$  and the decoder clusters set  $G_{d} = \\{g_{i}\\}_{i=1}^{m_{d}}$ . Each cluster  $g_{i}$  contains the ids of clients in this cluster. Detailed aggregation algorithm is shown in Algorithm 1.",
+        "bbox": [
+            507,
+            409,
+            882,
+            667
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We explore the following three different clustering strategies.",
+        "bbox": [
+            507,
+            667,
+            882,
+            700
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Language families/groups. Chronopoulou et al. (2022) have verified the strategy of sharing parameters within the same language family in the MNMT problem. We decide to use this strategy in the FL setting. We choose 8 languages belonging to 4 different language families from the TED2020 corpus and 10 languages belonging to 4 different language groups, which are all parts of the Indo-European language family, from the Europarl corpus. The clustering of the encoder depends on language families/groups of source languages and the clustering of the decoder is decided by the target languages' families/groups. Languages from the same family",
+        "bbox": [
+            507,
+            709,
+            884,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "5317",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: Inner-cluster Aggregation"
+        ],
+        "code_body": "Input: Encoder and decoder clusters sets  $G_{e}$  and  $G_{d}$ ;  \nInitial encoder and decoder paras  $\\Theta_{e}^{0}$  and  $\\Theta_{d}^{0}$ ;  \nClients set  $\\{C_i\\}_{i=1}^N$ ;  \nTraining round  $T$ .  \nOutput: Encoder paras  $\\{\\Theta_{e,i}^{T}\\}_{i=1}^N$ ;  \nDecoder paras  $\\{\\Theta_{d,i}^{T}\\}_{i=1}^N$   \nfor  $i$  from 1 to  $N$  do  \nInitialize  $\\Theta_{e,i}^{0}$  with  $\\Theta_{e}^{0}$ ;  \nInitialize  $\\Theta_{d,i}^{0}$  with  $\\Theta_{d}^{0}$ ;  \nfor  $i$  from 1 to  $N$  do  \n// local update of client  $i$   \nupdate  $\\Theta_{e,i}^{t-1}$  and  $\\Theta_{d,i}^{t-1}$  with local data;  \n// inner-cluster aggregation of encoder parameters  \nforeach  $g$  in  $G_{e}$  do  \n $\\Theta_{e,g}^{t} = \\sum_{id \\in g} \\frac{1}{|g|} \\Theta_{e,id}^{t-1}$ ;  \nforeach  $id$  in  $g$  do  \n $\\Theta_{e,id}^{t} = \\Theta_{e,g}^{t}$ ;  \n// inner-cluster aggregation of decoder parameters  \nforeach  $g$  in  $G_{d}$  do  \n $\\Theta_{d,g}^{t} = \\sum_{id \\in g} \\frac{1}{|g|} \\Theta_{d,id}^{t-1}$ ;  \nforeach  $id$  in  $g$  do  \n $\\Theta_{d,id}^{t} = \\Theta_{d,g}^{t}$ ;",
+        "bbox": [
+            115,
+            105,
+            485,
+            463
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "or group will be clustered into the same group.",
+        "bbox": [
+            112,
+            494,
+            460,
+            510
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Gradients. Unlike in the scene of centralized learning, clustering based on model parameters (Tan et al., 2019) in Fed-MNMT is unfeasible due to privacy problems. Therefore, we use gradients as the basis of the feature for clustering instead. For each language pair, we use a pre-trained multi-language model to acquire an average gradient vector of all data samples, then a clustering algorithm is applied to the gradient vectors in order to separate clients into different groups. The number of parameters we use for gradients clustering is only about  $131\\mathrm{K}$  for each client and will hardly introduce any extra communication cost.",
+        "bbox": [
+            112,
+            521,
+            489,
+            731
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Random clustering. We also test randomly separating all clients into different groups as a baseline for clustering strategies. In detail, we uniformly separate the clients and keep the numbers of clusters in the encoder and the decoder the same as those in the language families/groups strategy.",
+        "bbox": [
+            112,
+            740,
+            489,
+            838
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "2.4 Communication Cost Comparison",
+        "text_level": 1,
+        "bbox": [
+            112,
+            848,
+            428,
+            865
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Taking the mBART-50 model (Tang et al., 2020), which is a popular pre-trained multi-language model, as an example, the number of parameters",
+        "bbox": [
+            112,
+            870,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "is around 610.9M, which requires about 2.44GB storage space in the FP32 format. In comparison, after adding adapter modules, only about 8M parameters have to be transferred, which will save approximately  $98.7\\%$  communication cost.",
+        "bbox": [
+            507,
+            84,
+            884,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "More concretely, we provide an approximation for the transmission time needed between the server and clients as follows. Assuming the maximum bandwidth of the server is 1000Mbps, the time to transfer the entire mBART model from client to server is around 2.44GB / 1000Mbps = 19.5 seconds. Assuming all clients share the bandwidth, the total transfer time grows linearly with the number of clients. The synchronization process for all clients to finish transferring models to the server will occupy a large proportion of the total training time. In our actual experiments with 12 clients, the theoretical total transferring time is about  $19.5 \\times 12 = 234$  seconds. However, for clients with low-resource languages, the training could be finished within 7 minutes, which means transferring time occupies over half of the local training time. By contrast, the time to transfer the adapter's parameters is only about 0.26 seconds, which is negligible compared to local training time thus significantly improving the training efficiency.",
+        "bbox": [
+            507,
+            165,
+            884,
+            502
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3 Experimental Setup",
+        "text_level": 1,
+        "bbox": [
+            507,
+            514,
+            717,
+            531
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.1 Datasets and Evaluation Metrics",
+        "text_level": 1,
+        "bbox": [
+            507,
+            539,
+            811,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments in two different settings: Multi-to-English and Multi-to-Multi (hereinafter referred to as \"m2en\" and \"m2m\", respectively). We use the TED2020 corpus (Reimers and Gurevych, 2020) for the m2en setting and the Europarl corpus (Koehn, 2005) for the m2m setting. The TED2020 corpus is extracted from TED speeches and contains over 100 languages around the world. The Europarl corpus is from the proceedings of the European Parliament and contains 21 languages of European countries.",
+        "bbox": [
+            507,
+            560,
+            884,
+            736
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For each language pair $^2$ , we divide the original corpus into training, dev, and test datasets in accordance of the proportion of 6:2:2. We further sample subsets from the divided datasets. To simulate the scene of low-resource languages and high-resource languages, the training data size of each language pair varies according to the corresponding original",
+        "bbox": [
+            507,
+            737,
+            882,
+            850
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Abbreviations for languages we use: Chinese->zh, English->en, Thai->th, Arabic->ar, Hebrew->he, Finnish->fi, Estonian->et, Russian->ru, Slovene->sl, German->de, Dutch->nl, French->fr, Italian->it, Spain->es, Polish->pl, Slovene->sl, Lithuanian->lt, Latvian->lv.",
+        "bbox": [
+            507,
+            857,
+            885,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "5318",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "corpus size. The specific language pairs we use and corresponding data sizes are shown in Appendix A.",
+        "bbox": [
+            112,
+            84,
+            487,
+            116
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the m2en setting, for clustering strategies based on language families/groups and random shuffle, clustering algorithms are only applied to the encoder and all clients share the decoder's parameters because their target languages are all English. But for clustering based on gradients, we also cluster the decoder's parameters into different groups and the number of groups stays the same as that in the encoder. In the m2m setting, clustering is conducted for both the encoder and the decoder. Meanwhile, the numbers of groups in the encoder and decoder are the same in all clustering strategies. We will provide a further analysis of the clustering strategies of the m2m setting in  $\\S 4.2$ .",
+        "bbox": [
+            112,
+            117,
+            487,
+            341
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We choose the BLEU score as the evaluation metric using the SacreBLEU (Post, 2018) package. Aside from the BLEU score on each language pair, we additionally report the macro average and micro average scores on all language pairs.",
+        "bbox": [
+            112,
+            343,
+            487,
+            423
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            112,
+            436,
+            233,
+            450
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We evaluate the following methods as baselines:",
+        "bbox": [
+            112,
+            458,
+            472,
+            473
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Centralized-model. The results of centralized training, where data from all clients are gathered together, using the original multi-language model without extra modules.",
+        "bbox": [
+            112,
+            475,
+            485,
+            538
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Centralized-adapter. The results of centralized training using the multi-language model with adapter modules.",
+        "bbox": [
+            112,
+            539,
+            487,
+            586
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Adapter-local. We train a model for each client using local data without parameter aggregation with other language pairs.",
+        "bbox": [
+            112,
+            588,
+            487,
+            636
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Model-fed. We train the original multi-language model without Adapter modules under the federated learning framework, where the parameters are shared among all clients using the aggregation algorithm in Eq. (3) without any clustering strategies.",
+        "bbox": [
+            112,
+            638,
+            487,
+            717
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Adapter-fed. In this method, adapter modules are attached to the backbone model, while the rest settings are the same as those in model-fed. This baseline corresponds to the scene of directly introducing adapter without any clustering strategies.",
+        "bbox": [
+            112,
+            719,
+            487,
+            799
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Training Setup",
+        "text_level": 1,
+        "bbox": [
+            112,
+            812,
+            278,
+            828
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We choose the mBART-50 pre-trained model as our backbone model. To fairly compare the training and communication costs of different methods,",
+        "bbox": [
+            112,
+            833,
+            487,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "we train each model for 5 rounds. We select the checkpoint with the lowest loss on the dev set and evaluate it on the test set. Parameters are aggregated every time all clients finish an epoch of local training. For every client, the batch size is 8 and the local model is updated every 16 steps. The local learning rate is  $5 \\times 10^{-5}$  for the mBART model and  $1 \\times 10^{-3}$  for the models with adapter modules. The hidden size of adapter modules is 64.",
+        "bbox": [
+            507,
+            84,
+            884,
+            228
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For all experiments, we train the model with 3 random seeds and report the average scores. For the random clustering strategy, the clustering groups are different when using different random seeds.",
+        "bbox": [
+            507,
+            229,
+            882,
+            293
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            307,
+            729,
+            322
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Primary Results and Findings",
+        "text_level": 1,
+        "bbox": [
+            507,
+            332,
+            791,
+            348
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The experimental results in the m2en and m2m settings are shown in Table 1 and Table 2, respectively. In general, directly adding adapter modules leads to a performance drop (comparing adapter-fed and model-fed) and methods with clustering strategies all achieve better performance than the direct baseline adapter-fed, indicating the ability of our clustering strategies to alleviate data discrepancy. In both settings, adapter-families method performs best in macro and micro average scores among three clustering strategies, even surpassing model-fed in the m2m setting.",
+        "bbox": [
+            505,
+            354,
+            882,
+            546
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "It is noteworthy that the clustering strategies acquire more significant performance improvements in the m2m setting than in the m2en setting. The problem of conflicting parameters is more settle-some in the m2m setting because there exist more kinds of languages (especially target languages). Thus introducing clustering strategies will bring more benefit to m2m translation tasks.",
+        "bbox": [
+            507,
+            548,
+            882,
+            675
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Meanwhile, we notice that our clustering strategies fail to beat adapter-local in the m2m setting. This can also be explained by the difference in the difficulty of tasks. In the more complicated m2m translation task, more elaborate clustering strategies should be designed to fully make advantage of other language pairs and avoid the influence of conflicting parameters. However, we bring the ability of multi-language translation to these clients through FL with an acceptable drop in performance compared to adapter-local.",
+        "bbox": [
+            507,
+            677,
+            882,
+            853
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            865,
+            673,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the m2m setting, the clustering of the adapter modules attached to the encoder and the decoder",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "3https://huggingface.co/facebook/mbart-large-50-many-to-many-mmt",
+        "bbox": [
+            112,
+            891,
+            386,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "5319",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/9ae9ff7aa4d9673ba0552024d023d4a56869c8ac5b6fe5fd8a6a8cc57ef8179d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Comm. Cost</td><td>zh-en</td><td>th-en</td><td>ar-en</td><td>he-en</td><td>fi-en</td><td>et-en</td><td>ru-en</td><td>sl-en</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>centralized-model</td><td>N / A</td><td>24.72</td><td>28.97</td><td>38.29</td><td>43.59</td><td>32.81</td><td>32.70</td><td>30.14</td><td>47.92</td><td>34.89</td><td>32.33</td></tr><tr><td>centralized-adapter</td><td>N / A</td><td>24.97</td><td>21.47</td><td>39.02</td><td>45.05</td><td>33.62</td><td>32.62</td><td>30.65</td><td>50.45</td><td>34.73</td><td>32.03</td></tr><tr><td>adapter-local</td><td>N / A</td><td>25.16</td><td>21.68</td><td>39.46</td><td>44.32</td><td>33.12</td><td>32.64</td><td>30.58</td><td>50.93</td><td>34.73</td><td>32.15</td></tr><tr><td>model-fed</td><td>611M</td><td>25.31</td><td>23.41</td><td>39.54</td><td>45.13</td><td>32.87</td><td>33.27</td><td>30.77</td><td>51.85</td><td>35.27</td><td>32.55</td></tr><tr><td>adapter-fed</td><td>8M</td><td>25.12</td><td>16.64</td><td>39.62</td><td>44.93</td><td>33.14</td><td>32.66</td><td>30.41</td><td>53.62</td><td>34.52</td><td>31.71</td></tr><tr><td>adapter-random</td><td>8M</td><td>25.37</td><td>21.48</td><td>39.61</td><td>44.82</td><td>33.26</td><td>33.21</td><td>30.75</td><td>52.64</td><td>35.14</td><td>32.38</td></tr><tr><td>adapter-gradients</td><td>8M</td><td>25.26</td><td>21.26</td><td>39.39</td><td>44.64</td><td>33.62</td><td>33.14</td><td>30.72</td><td>51.16</td><td>34.90</td><td>32.21</td></tr><tr><td>adapter-families</td><td>8M</td><td>25.28</td><td>21.58</td><td>39.45</td><td>44.70</td><td>33.87</td><td>33.23</td><td>30.92</td><td>52.64</td><td>35.21</td><td>32.40</td></tr></table>",
+        "bbox": [
+            136,
+            82,
+            863,
+            209
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/cea853e37265720de59660087f5eb245190bcc9f8112cdd284124c9fef3a036e.jpg",
+        "table_caption": [
+            "Table 1: BLEU scores on the TED2020 corpus. Comm. Cost, which is short for communication cost, denotes the number of parameters communicated between the server and each client. Adapter-random, adapter-gradients, and adapter-families refer to clustering strategies of random clustering, gradients, and language families/groups, respectively. The best result of each language pair is highlighted in **bold** (only methods with adapter modules trained in the FL setting are considered)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Comm. Cost</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-sl</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>centralized-model</td><td>N / A</td><td>30.43</td><td>19.15</td><td>28.86</td><td>23.20</td><td>19.96</td><td>27.17</td><td>40.35</td><td>33.20</td><td>21.24</td><td>21.84</td><td>23.63</td><td>20.57</td><td>25.80</td><td>26.10</td></tr><tr><td>centralized-adapter</td><td>N / A</td><td>30.59</td><td>19.07</td><td>29.65</td><td>22.92</td><td>18.68</td><td>27.73</td><td>41.52</td><td>33.20</td><td>21.45</td><td>22.16</td><td>23.40</td><td>20.85</td><td>25.93</td><td>26.19</td></tr><tr><td>adapter-local</td><td>N / A</td><td>30.88</td><td>19.19</td><td>30.31</td><td>23.50</td><td>20.01</td><td>28.12</td><td>41.84</td><td>33.39</td><td>21.13</td><td>22.27</td><td>23.62</td><td>21.05</td><td>26.28</td><td>26.56</td></tr><tr><td>model-fed</td><td>611M</td><td>30.41</td><td>17.60</td><td>29.62</td><td>19.76</td><td>13.41</td><td>28.01</td><td>39.77</td><td>32.25</td><td>21.10</td><td>21.94</td><td>20.09</td><td>19.92</td><td>24.49</td><td>24.68</td></tr><tr><td>adapter-fed</td><td>8M</td><td>29.75</td><td>17.47</td><td>30.08</td><td>16.92</td><td>11.85</td><td>28.01</td><td>38.18</td><td>31.06</td><td>20.18</td><td>21.23</td><td>18.02</td><td>19.97</td><td>23.56</td><td>23.53</td></tr><tr><td>adapter-random</td><td>8M</td><td>30.90</td><td>18.57</td><td>30.04</td><td>22.39</td><td>16.53</td><td>28.25</td><td>36.97</td><td>33.04</td><td>21.13</td><td>22.28</td><td>22.72</td><td>20.37</td><td>25.27</td><td>25.67</td></tr><tr><td>adapter-gradients</td><td>8M</td><td>30.14</td><td>19.69</td><td>30.19</td><td>19.53</td><td>19.14</td><td>28.06</td><td>41.73</td><td>33.31</td><td>21.51</td><td>21.60</td><td>19.95</td><td>21.29</td><td>25.51</td><td>25.35</td></tr><tr><td>adapter-families</td><td>8M</td><td>30.60</td><td>19.31</td><td>30.12</td><td>22.65</td><td>16.69</td><td>27.99</td><td>41.33</td><td>33.21</td><td>21.31</td><td>22.04</td><td>22.68</td><td>21.21</td><td>25.76</td><td>26.04</td></tr></table>",
+        "bbox": [
+            134,
+            303,
+            863,
+            406
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: BLEU scores on the Europarl corpus. The meanings of symbols stay the same as those in Table 1.",
+        "bbox": [
+            137,
+            414,
+            858,
+            430
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "are independent. To further explore the specific influence of clustering in these two modules on the model's performance, we apply the clustering strategy to only the encoder and the decoder separately and show the results in Figure 3. We use language families/groups as the clustering strategy. All methods are trained in the m2m setting using the Europarl dataset with one random seed. Other settings are the same as those in our main experiments.",
+        "bbox": [
+            112,
+            455,
+            487,
+            599
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "To our surprise, either clustering in the encoder or the decoder significantly improves performance compared to no-sharing strategies, even surpasses adapter-families. We owe this phenomenon to our naive clustering strategies in the encoder and the decoder. In adapter-families, the clustering of the encoder and the decoder is only related to the source and target languages respectively. However, parameters in both the encoder and the decoder are influenced by source and target languages together during training. The inconsistency between clustering strategies and parameter update results in adapter-families method's worse performance than adapter-encoder and adapter-decoder.",
+        "bbox": [
+            112,
+            600,
+            489,
+            826
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Further Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            838,
+            294,
+            854
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Case Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            865,
+            250,
+            881
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We select some representative cases of translations from method adapter-families and adapter",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/a1721ffc51fcbabb4f31b7005168dd3bf132ac5ba2fba58976daddb2ea6c23a3.jpg",
+        "image_caption": [
+            "Figure 3: Results of ablation study. Encoder/Decoder-cluster denote clustering only in the encoder/decoder, both obtaining significant improvements compared to the no-clustering baseline."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            519,
+            462,
+            880,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "shareAll, which are shown in Table 3, to further study the influence of our clustering strategies. We separate the mistakes into three categories.",
+        "bbox": [
+            507,
+            706,
+            882,
+            755
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Opposite semantics In case 1 and case 2, adapter-fed misses negation adverbs in predictions and results in totally opposite semantics.",
+        "bbox": [
+            507,
+            764,
+            882,
+            813
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Inaccurate words/phrases In case 3, adapter-fed translates adverbial of time into \"a day or two\", which should be \"four or five weeks\" actually. In case 4, adapter-fed uses the expression \"save the world\", which differs from the original expression \"change the world\" in semantics.",
+        "bbox": [
+            505,
+            822,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "5320",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0633af0df453260a9f29713149144ca66b91a9e3af44f0988b467881c64d6d1a.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>Sentence</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>I think if somebody tells a lie, they&#x27;re not just a liar. \nI think if somebody tells a lie, they&#x27;re not just a liar. \nI know that when people lie, they&#x27;re just lying.</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>I never miss a single training session. \nI didn&#x27;t miss a single training session. \nI missed one training session.</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>And within four or five weeks, he can do it again. \nAnd within four or five weeks, he can do it again. \nAnd within a day or two, he can do this.</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>So could art change the world? \nCan art change the world? \nIs art about saving the world?</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>Over 100,000 children learn science this way. \nHundreds of thousands of children learn science this way. \nHundreds of thousands of people learned how to do that.</td></tr><tr><td>Ground truth \nadapter-families \nadapter-fed</td><td>And all at once I became a learner. \nAnd all of a sudden, I became a learner. \nAnd all of a sudden, it&#x27;s happening.</td></tr></table>",
+        "bbox": [
+            126,
+            80,
+            478,
+            305
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/76b7fc1d14a191cf8202bb1d4e84895d63c2fb0f309bc06ce68564fc0eff5def.jpg",
+        "table_caption": [
+            "Table 3: Case study comparison of adapter-families and adapter-fed. Obvious translation mistakes are highlighted in red and corresponding correct translations are highlighted in blue."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Method</td><td colspan=\"2\">TED2020</td><td colspan=\"2\">Europarl</td></tr><tr><td>Macro Avg.</td><td>Micro Avg.</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td rowspan=\"2\">adapter-fed</td><td>FedAvg</td><td>34.31</td><td>31.70</td><td>22.20</td><td>23.00</td></tr><tr><td>FedMean</td><td>34.52</td><td>31.71</td><td>23.56</td><td>23.53</td></tr><tr><td rowspan=\"2\">adapter-random</td><td>FedAvg</td><td>34.79</td><td>32.16</td><td>25.19</td><td>25.76</td></tr><tr><td>FedMean</td><td>35.14</td><td>32.38</td><td>25.27</td><td>25.67</td></tr><tr><td rowspan=\"2\">adapter-gradients</td><td>FedAvg</td><td>34.46</td><td>31.96</td><td>25.64</td><td>25.84</td></tr><tr><td>FedMean</td><td>34.90</td><td>32.21</td><td>25.51</td><td>25.35</td></tr><tr><td rowspan=\"2\">adapter-families</td><td>FedAvg</td><td>34.83</td><td>32.09</td><td>25.47</td><td>26.02</td></tr><tr><td>FedMean</td><td>35.21</td><td>32.40</td><td>25.76</td><td>26.04</td></tr></table>",
+        "bbox": [
+            126,
+            387,
+            478,
+            495
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Comparison between FedAvg and FedMean (the aggregation method we use in Eq. 3). Results of FedMean are in green if better than those of FedAvg, otherwise in red. In most cases, FedMean yields performance gains.",
+        "bbox": [
+            112,
+            505,
+            489,
+            576
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Ambiguous semantics In cases 5 and 6, adapter-fed loses specific semantic information in ground truths. It fails to properly translate \"children\", \"science\" and \"become a learner\" and uses more ambiguous expressions instead.",
+        "bbox": [
+            112,
+            605,
+            489,
+            684
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In comparison, adapter-families makes more accurate predictions in the above cases, which suggests that appropriate clustering strategies help the model produce better translations with improvements in semantics.",
+        "bbox": [
+            112,
+            687,
+            489,
+            765
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.2 Both FedMean and Clustering Contribute",
+        "text_level": 1,
+        "bbox": [
+            112,
+            783,
+            487,
+            797
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, discrepancies in data come from two aspects: data quantity skew and linguistic discrepancy (language difference). We adjust the aggregation algorithm from Eq. (2) to Eq. (3), which we call FedMean here, to tackle the problem of quantity skew. Moreover, we propose clustering strategies to prevent clients from receiving con",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "flecting parameters from dissimilar languages. To explore how these two methods contribute to improvement in performance, we further conduct experiments with the aggregation algorithm changed to FedAvg (see Eq. (2)) and keep other training settings unchanged.",
+        "bbox": [
+            507,
+            84,
+            884,
+            180
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As the results displayed in Table 4, clustering strategies bring more significant improvements in performance on the Europarl corpus than the TED2020 corpus. Since experiments on the Europarl corpus consists of more different languages and are conducted in a more complicated m2m setting, the problem of linguistic discrepancy is more severe for Europarl. For the TED2020 corpus, changing the aggregation algorithm from FedAvg to FedMean leads to more significant improvements for methods with clustering strategies compared to adapter-fed. In contrast, for the Europarl corpus, adapter-fed substantially benefits from FedMean, while FedMean hardly brings any benefits to methods with clustering strategies, even causing performance drop in some cases.",
+        "bbox": [
+            507,
+            183,
+            884,
+            439
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Based on these observations, we contend that both aggregation algorithms and clustering strategies contribute to performance gain by alleviating data discrepancies. The specific extent of improvement depends on the extent of data quantity skew and linguistic discrepancy.",
+        "bbox": [
+            507,
+            442,
+            882,
+            539
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Further Cost Saving by Adapter Pruning",
+        "text_level": 1,
+        "bbox": [
+            507,
+            556,
+            878,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "On top of the adapter tuning approach, adapter pruning techniques (Rücklé et al., 2021; Pfeiffer et al., 2021; Karimi Mahabadi et al., 2021) further compress the number of parameters to be updated. To further reduce the communication cost, we conduct an exploratory attempt to prune parts of adapter modules in both the encoder and the decoder. We still choose the mBART-50 model, with 12 layers in the encoder and the decoder separately, to conduct the experiments. Specifically, we evenly separate all adapter modules we add in mBART into three parts: input-end adapters (adapter modules in the first 4 layers of the encoder or the decoder), middle-layer adapters (adapter modules in layers 5 to 8 of the encoder or the decoder), output-end adapters (adapter modules in the last 4 layers of the encoder or the decoder). In each strategy, only one part of the adapter modules is kept, so the communication cost is saved by two-thirds. We use adapter-families as the baseline and train all models with one random seed. The rest settings stay the",
+        "bbox": [
+            507,
+            581,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "5321",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/12efe430ac8b836003c4c27ef963d4b2b11a74f3810b069bc0e5bb2baf427306.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>adapters</td><td>Comm. Cost</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-s1</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td rowspan=\"4\">Language families clustering</td><td>input-end</td><td>2.7M</td><td>30.21</td><td>19.04</td><td>29.98</td><td>20.39</td><td>14.93</td><td>28.09</td><td>41.64</td><td>32.08</td><td>21.39</td><td>21.62</td><td>19.89</td><td>20.73</td><td>25.00</td><td>25.12</td></tr><tr><td>middle-layer</td><td>2.7M</td><td>30.12</td><td>18.88</td><td>29.84</td><td>21.27</td><td>15.15</td><td>28.16</td><td>41.33</td><td>31.92</td><td>21.06</td><td>21.68</td><td>20.99</td><td>20.74</td><td>25.10</td><td>25.30</td></tr><tr><td>output-end</td><td>2.7M</td><td>30.30</td><td>19.37</td><td>30.33</td><td>22.17</td><td>16.06</td><td>27.96</td><td>41.39</td><td>31.76</td><td>20.03</td><td>21.75</td><td>21.97</td><td>21.15</td><td>25.35</td><td>25.64</td></tr><tr><td>all adapters</td><td>8.1M</td><td>30.18</td><td>18.85</td><td>29.86</td><td>22.42</td><td>16.10</td><td>27.66</td><td>41.13</td><td>32.96</td><td>20.78</td><td>21.80</td><td>22.78</td><td>21.27</td><td>25.48</td><td>25.75</td></tr></table>",
+        "bbox": [
+            136,
+            82,
+            862,
+            136
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: BLEU scores of different adapter pruning strategies. We acquire similar results with only 1/3 communication cost compared to keeping all adapter modules.",
+        "bbox": [
+            112,
+            146,
+            882,
+            174
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "same as those in previous main experiments. The results are shown in Table 5.",
+        "bbox": [
+            112,
+            200,
+            485,
+            230
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "It is encouraging that pruning adapters do not result in a sharp decrease in performance. We observe that keeping output-end adapters achieves the highest score among the three pruning strategies, which suggests that adapters in the top layers play more important roles. Overall, the results indicate that it is possible to further reduce communication costs and it is worthwhile to explore more elaborate pruning techniques in future work.",
+        "bbox": [
+            112,
+            233,
+            487,
+            376
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            388,
+            270,
+            401
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Federated Learning was first proposed by McMahan et al. (2017) as a decentralized training framework. Due to its decentralized and private nature, FL shows great potential in actual applications. Recently, there has been a surge in the NLP community to explore the application of federated learning in diverse NLP tasks, such as emojis prediciton (Gandhi et al., 2022), named entity recognition (Ge et al., 2020), and machine translation (Roosta et al., 2021; Passban et al., 2022; Weller et al., 2022), etc. Roosta et al. (2021) first applied FL to NMT tasks. However, training language models in the FL setting brings huge communication overheads. To solve this problem, researchers have proposed to only exchange some dedicated \"Controller\" (Roosta et al., 2021) layers between the server and clients. Moreover, Passban et al. (2022) introduced parameter pruning strategies to reduce communication bandwidth. Our methods with adapter modules have advantages in communication efficiency with fewer parameters to be transferred compared to Controller (see Appendix D), and other parameter pruning strategies can also be applied to our adapter modules to further reduce communication costs.",
+        "bbox": [
+            112,
+            413,
+            489,
+            813
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multilingual Neural Machine Translation (MNMT) trains a single model to handle translation between multiple language pairs (Johnson et al., 2017; Aharoni et al., 2019; Zhang et al., 2020). Moreover, MNMT significantly reduces training and inference costs by eliminating the",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "need to train models for each language pair. Massively pre-trained multilingual models have been used for MNMT, such as mBART-50 (Tang et al., 2020) and M2M100 (Fan et al., 2021). In recent years, adapter has become a popular method in MNMT (Bapna and First, 2019; Cooper Stickland et al., 2021; Philip et al., 2020; Üstün et al., 2021; Chronopoulou et al., 2022) due to its high parameter efficiency and transferability between tasks.",
+        "bbox": [
+            507,
+            200,
+            884,
+            344
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Different from previous works on this topic, inspired by recent progress in improving the efficiency of NLP methods (Strubell et al., 2019; Li et al., 2021a; Xu et al., 2021; Li et al., 2021b), we focus on communication efficiency in FL-MNMT and make the first effort to introduce adapter modules in order to reduce communication costs. We also apply different clustering strategies to resolve the issue of conflicting parameters stemming from data discrepancy.",
+        "bbox": [
+            507,
+            346,
+            884,
+            506
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            520,
+            640,
+            536
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we introduce adapter modules to PLMs for the Fed-MNMT problem to boost communication efficiency. We reduce the communication cost by over  $98\\%$  and make the training process of Fed-MNMT practical. To deal with the problem of performance drop after introducing adapter modules, we propose different clustering strategies to separate clients into different groups to avoid the negative influence of data discrepancy. We surpass the direct baseline with a substantial gap, especially in the more complicated multi-to-multi translation setting.",
+        "bbox": [
+            507,
+            546,
+            882,
+            740
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, our analytic experiments indicate that both aggregation algorithms in server and clustering strategies affect the performance of FedMNMT. We also explore the possibility of further reducing communication costs by pruning adapter modules and find that adapters in top layers are more significant for translation performance.",
+        "bbox": [
+            507,
+            741,
+            882,
+            853
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In future work, we will explore more well-designed clustering strategies and attach other parameter-efficient techniques to adapter to further reduce the parameters to be transferred.",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "5322",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "First, in this work, we assume that clustering in the encoder and the decoder is only related to the source and target languages, respectively. Actually, both parameters in the encoder and the decoder are influenced by source and target languages simultaneously. Therefore, our assumption may lead to a performance drop. In future work, we plan to explore more complicated clustering strategies.",
+        "bbox": [
+            112,
+            108,
+            487,
+            236
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Moreover, our adapter-families method depends on prior linguistic knowledge. Its actual effectiveness can be affected by the distribution of language families/groups in clients. Our methods mainly apply to comparably uniform language distribution.",
+        "bbox": [
+            112,
+            237,
+            487,
+            317
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In addition, the effectiveness of our methods on other PLMs needs to be verified. However, it is easy to transfer our methods to other models so it will not be a challenging problem.",
+        "bbox": [
+            112,
+            318,
+            487,
+            381
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            114,
+            393,
+            285,
+            409
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We thank all reviewers for their insightful comments and suggestions. This work is supported by Natural Science Foundation of China (NSFC) No.62176002. We sincerely thank Jingjing Xu for her valuable suggestions. Xu Sun is the corresponding author.",
+        "bbox": [
+            112,
+            418,
+            489,
+            513
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            539,
+            213,
+            556
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.",
+            "Ankur Bapna and Orhan First. 2019. Simple, scalable adaptation for neural machine translation. 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 1538-1548, Hong Kong, China. Association for Computational Linguistics.",
+            "Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2022. Language-family adapters for multilingual neural machine translation. ArXiv preprint, abs/2209.15236.",
+            "Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational"
+        ],
+        "bbox": [
+            115,
+            562,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Linguistics: Main Volume, pages 3440-3453, Online.  \nAssociation for Computational Linguistics.",
+            "Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. 2020. Beyond english-centric multilingual machine translation.",
+            "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. J. Mach. Learn. Res., 22(107):1-48.",
+            "Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San Diego, California. Association for Computational Linguistics.",
+            "Deep Gandhi, Jash Mehta, Nirali Parekh, Karan Waghela, Lynette D'Mello, and Zeerak Talat. 2022. A federated approach to predicting emojis in hindi tweets. ArXiv preprint, abs/2211.06401.",
+            "Suyu Ge, Fangzhao Wu, Chuhan Wu, Tao Qi, Yongfeng Huang, and Xing Xie. 2020. Fedner: Privacy-preserving medical named entity recognition with federated learning. arXiv e-prints, pages arXiv-2003.",
+            "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.",
+            "Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Google's multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association for Computational Linguistics, 5:339-351.",
+            "Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. 2021. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035.",
+            "Philipp Koehn. 2005. Europarl: A parallel corpus for statistical machine translation. In Proceedings of Machine Translation Summit X: Papers, pages 79-86, Phuket, Thailand."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "5323",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lei Li, Yankai Lin, Deli Chen, Shuhuai Ren, Peng Li, Jie Zhou, and Xu Sun. 2021a. Cascadebert: Accelerating inference of pre-trained language models via calibrated complete models cascade. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 475-486.",
+            "Lei Li, Yankai Lin, Shuhuai Ren, Peng Li, Jie Zhou, and Xu Sun. 2021b. Dynamic knowledge distillation for pre-trained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 379-389.",
+            "Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR.",
+            "Yifan Niu and Weihong Deng. 2022. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1999-2007.",
+            "Peyman Passban, Tanya Roosta, Rahul Gupta, Ankit Chadha, and Clement Chung. 2022. Training mixed-domain translation models via federated learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2576-2586, Seattle, United States. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Aishwarya Kamath, Andreas Rückle, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-destructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 487-503, Online. Association for Computational Linguistics.",
+            "Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020. AdapterHub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics.",
+            "Jerin Philip, Alexandre Berard, Matthias Galle, and Laurent Besacier. 2020. Monolingual adapters for zero-shot neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4465-4470, Online. Association for Computational Linguistics.",
+            "Matt Post. 2018. A call for clarity in reporting BLEU scores. In Proceedings of the Third Conference on"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Machine Translation: Research Papers, pages 186-191, Brussels, Belgium. Association for Computational Linguistics.",
+            "Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. 2017. Learning multiple visual domains with residual adapters. 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 506-516.",
+            "Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual using knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4512-4525, Online. Association for Computational Linguistics.",
+            "Tanya Roosta, Peyman Passban, and Ankit Chadha. 2021. Communication-efficient federated learning for neural machine translation. ArXiv preprint, abs/2112.06135.",
+            "Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in nlp. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650.",
+            "Xu Tan, Jiale Chen, Di He, Yingce Xia, Tao Qin, and Tie-Yan Liu. 2019. Multilingual neural machine translation with language clustering. 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 963-973, Hong Kong, China. Association for Computational Linguistics.",
+            "Y. Tang, C. Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning. ArXiv, abs/2008.00401.",
+            "Ahmet Üstün, Alexandre Berard, Laurent Besacier, and Matthias Galle. 2021. Multilingual unsupervised neural machine translation with denoising adapters. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6650-6662, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. 2020. Tackling the objective"
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "5324",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "inconsistency problem in heterogeneous federated optimization. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.",
+        "bbox": [
+            132,
+            85,
+            487,
+            151
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Orion Weller, Marc Marone, Vladimir Braverman, Dawn Lawrie, and Benjamin Van Durme. 2022. Pretrained models for multilingual federated learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1413-1421, Seattle, United States. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            161,
+            489,
+            267
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Jingjing Xu, Wangchunshu Zhou, Zhiyi Fu, Hao Zhou, and Lei Li. 2021. A survey on green deep learning. arXiv preprint arXiv:2111.05193.",
+        "bbox": [
+            114,
+            275,
+            489,
+            317
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Biao Zhang, Philip Williams, Ivan Titov, and Rico Senrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628-1639, Online. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            326,
+            489,
+            418
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. 2018. Federated learning with non-iid data. ArXiv preprint, abs/1806.00582.",
+        "bbox": [
+            114,
+            429,
+            489,
+            481
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Training Data Sizes",
+        "text_level": 1,
+        "bbox": [
+            114,
+            495,
+            322,
+            512
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "The specific training data size of each language pair is shown in Table 6.",
+        "bbox": [
+            112,
+            521,
+            485,
+            554
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Complete Results of Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            565,
+            460,
+            583
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We show the complete results on all language pairs of the ablation study in Table 7.",
+        "bbox": [
+            112,
+            592,
+            485,
+            621
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We find that only applying clustering to the decoder acquires the highest scores on 7 out of the total 12 language pairs. To our surprise, adapter-families method fails to reach the best performance in average scores, which has been explained in § 4.2 in the main text.",
+        "bbox": [
+            112,
+            624,
+            487,
+            719
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "C Uniform Data Distribution",
+        "text_level": 1,
+        "bbox": [
+            112,
+            732,
+            381,
+            747
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We also conduct experiments on TED2020 with each client owning training data of equal size. The results are shown in Table 8. We observe that the performance gaps between different methods are similar to those in Table 1. Notably, Adapter-families beats adapter-random by a slight margin. Both clustering strategies acquire obvious performance improvement compared to the baseline adapter-fed. These empirical results verify that our methods apply to various data distributions.",
+        "bbox": [
+            112,
+            758,
+            489,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/4bc2effc1ae7459b340fabeb905f517a63df7c9e9743675bf5345da108d4e28b.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Corpus</td><td>Source Language Family/Group</td><td>Language Pair</td><td>Dataset Size</td></tr><tr><td rowspan=\"8\">TED2020</td><td rowspan=\"2\">Sino-Tibetan</td><td>zh-&gt;en</td><td>9984</td></tr><tr><td>th-&gt;en</td><td>4992</td></tr><tr><td rowspan=\"2\">Afro-asiatic</td><td>ar-&gt;en</td><td>9984</td></tr><tr><td>he-&gt;en</td><td>1920</td></tr><tr><td rowspan=\"2\">Uralic</td><td>fi-&gt;en</td><td>1920</td></tr><tr><td>et-&gt;en</td><td>1920</td></tr><tr><td rowspan=\"2\">Indo-European</td><td>ru-&gt;en</td><td>9984</td></tr><tr><td>sl-&gt;en</td><td>1920</td></tr><tr><td rowspan=\"12\">Europarl</td><td rowspan=\"3\">Germanic</td><td>de-&gt;fr</td><td>11648</td></tr><tr><td>nl-&gt;pl</td><td>3584</td></tr><tr><td>en-&gt;lt</td><td>3712</td></tr><tr><td rowspan=\"3\">Romance</td><td>fr-&gt;nl</td><td>12160</td></tr><tr><td>it-&gt;sl</td><td>3456</td></tr><tr><td>es-&gt;lv</td><td>3584</td></tr><tr><td rowspan=\"3\">Slavic</td><td>pl-&gt;en</td><td>3712</td></tr><tr><td>sl-&gt;es</td><td>3584</td></tr><tr><td>sl-&gt;lt</td><td>3584</td></tr><tr><td rowspan=\"3\">Baltic</td><td>lt-&gt;de</td><td>3328</td></tr><tr><td>lv-&gt;it</td><td>3584</td></tr><tr><td>lv-&gt;pl</td><td>3712</td></tr></table>",
+        "bbox": [
+            522,
+            80,
+            873,
+            397
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Detailed data sizes of language pairs from Europarl corpus.",
+        "bbox": [
+            507,
+            406,
+            882,
+            436
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "D Comparison to Controllers",
+        "text_level": 1,
+        "bbox": [
+            507,
+            460,
+            779,
+            476
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Controllers (Roosta et al., 2021) only exchange 8 layers in a 32-layer Transformer (4 from encoder and 4 from decoder) between the server and clients, which means that they reduce the communication cost by approximately  $66\\%$  (the number of layers in the original model without Controllers is 24). Compared with Controllers, we introduce adapter modules in Fed-MNMT without the need to define additional layers. Besides, our methods transmit a much smaller amount of parameters in client-to-server exchanges than using Controllers. Therefore, our proposal is superior to Controllers in terms of communication efficiency.",
+        "bbox": [
+            507,
+            486,
+            884,
+            695
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "5325",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "table",
+        "img_path": "images/ddf530669f7524cf0cdd131bccc538697647d5b611ea5f73184f21aa2b2e4e51.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-sl</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>adapter-fed</td><td>29.66</td><td>17.35</td><td>30.22</td><td>16.89</td><td>11.95</td><td>27.91</td><td>37.31</td><td>31.04</td><td>20.13</td><td>21.18</td><td>17.95</td><td>19.97</td><td>23.46</td><td>23.44</td></tr><tr><td>+ encoder clustering</td><td>30.50</td><td>19.24</td><td>29.85</td><td>22.94</td><td>18.56</td><td>28.01</td><td>41.93</td><td>33.25</td><td>20.84</td><td>22.09</td><td>23.34</td><td>20.82</td><td>25.95</td><td>26.19</td></tr><tr><td>+ decoder clustering</td><td>31.19</td><td>19.41</td><td>30.53</td><td>23.06</td><td>18.26</td><td>28.20</td><td>41.71</td><td>33.56</td><td>21.26</td><td>21.96</td><td>23.03</td><td>21.06</td><td>26.10</td><td>26.42</td></tr><tr><td>adapter-families</td><td>30.18</td><td>18.85</td><td>29.86</td><td>22.42</td><td>16.10</td><td>27.66</td><td>41.13</td><td>32.96</td><td>20.78</td><td>21.80</td><td>22.78</td><td>21.27</td><td>25.48</td><td>25.75</td></tr></table>",
+        "bbox": [
+            136,
+            227,
+            862,
+            292
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/82ead405b8e1ac258587c090477e324009c689767f619a81458d98180c3feaee.jpg",
+        "table_caption": [
+            "Table 7: BLEU scores of applying language families strategies to only the encoder and the decoder. Adapter-fed corresponds to clustering in neither the encoder or the decoder. Adapter-families denotes clustering in both the encoder and the decoder."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Method</td><td>zh-en</td><td>th-en</td><td>ar-en</td><td>he-en</td><td>fi-en</td><td>et-en</td><td>ru-en</td><td>sl-en</td><td>Avg.</td></tr><tr><td>model-fed</td><td>25.58</td><td>22.90</td><td>39.63</td><td>46.12</td><td>34.78</td><td>34.41</td><td>30.42</td><td>51.32</td><td>35.64</td></tr><tr><td>adapter-fed</td><td>24.83</td><td>16.79</td><td>40.02</td><td>45.99</td><td>33.55</td><td>33.67</td><td>29.61</td><td>52.20</td><td>34.58</td></tr><tr><td>adapter-random</td><td>25.30</td><td>22.33</td><td>39.58</td><td>45.59</td><td>34.58</td><td>34.31</td><td>30.12</td><td>51.52</td><td>35.41</td></tr><tr><td>adapter-families</td><td>25.04</td><td>22.08</td><td>39.50</td><td>45.74</td><td>34.91</td><td>34.65</td><td>30.16</td><td>51.38</td><td>35.43</td></tr></table>",
+        "bbox": [
+            194,
+            643,
+            805,
+            730
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Table 8: BLEU scores on the TED2020 corpus with equivalent data size for each language pair. Best scores are highlighted in bold except model-fed.",
+        "bbox": [
+            112,
+            739,
+            882,
+            769
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "5326",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? section Limitations",
+            "A2. Did you discuss any potential risks of your work? Not applicable. Left blank.",
+            "A3. Do the abstract and introduction summarize the paper's main claims? abstract and section 1",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            127,
+            126,
+            695,
+            287
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            213,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? No response.",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response."
+        ],
+        "bbox": [
+            127,
+            347,
+            880,
+            753
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "section 3, 4 and 5",
+        "bbox": [
+            132,
+            785,
+            267,
+            801
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? section 4",
+        "bbox": [
+            129,
+            810,
+            880,
+            859
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            866,
+            877,
+            889
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "5327",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section 3",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? section 4 and 5",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? section 3.1"
+        ],
+        "bbox": [
+            129,
+            83,
+            880,
+            280
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            112,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "5328",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2023/Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter/6f207301-9f3f-466d-aed1-4bdd052903e2_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11c1945543e838dfd7447bdbd9b831d0a7bd6e318b1e21aefc4ae125508e0105
+size 559624

2023/Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter/full.md

@@ -0,0 +1,415 @@
+# Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter
+
+Yi Liu $^{1}$ , Xiaohan Bi $^{2}$ , Lei Li $^{1}$ , Sishuo Chen $^{2}$ , Wenkai Yang $^{2}$ , Xu Sun $^{1}$
+
+$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University
+
+$^{2}$ Center for Data Science, Peking University
+
+yliu.pku@outlook.com {bxh, wkyang}  $@$ stu.pku.edu.cn
+
+nlp.lilei@gmail.com {chensishuo,xusun}@pku.edu.cn
+
+# Abstract
+
+Federated Multilingual Neural Machine Translation (Fed-MNMT) has emerged as a promising paradigm for institutions with limited language resources. This approach allows multiple institutions to act as clients and train a unified model through model synchronization, rather than collecting sensitive data for centralized training. This significantly reduces the cost of corpus collection and preserves data privacy. However, as pre-trained language models (PLMs) continue to increase in size, the communication cost for transmitting parameters during synchronization has become a training speed bottleneck. In this paper, we propose a communication-efficient Fed-MNMT framework that addresses this issue by keeping PLMs frozen and only transferring lightweight adapter modules between clients. Since different language pairs exhibit substantial discrepancies in data distributions, adapter parameters of clients may conflict with each other. To tackle this, we explore various clustering strategies to group parameters for integration and mitigate the negative effects of conflicting parameters. Experimental results demonstrate that our framework reduces communication cost by over  $98\%$  while achieving similar or even better performance compared to competitive baselines. Further analysis reveals that clustering strategies effectively solve the problem of linguistic discrepancy and pruning adapter modules further improves communication efficiency. $^{1}$
+
+# 1 Introduction
+
+Federated Learning (FL) (McMahan et al., 2017) provides a new training framework utilizing data from various clients without privacy leakage. In FL, the server receives models from clients trained with their local data and aggregates all parameters it has received to acquire a global model, and then
+
+![](images/b83316396c9a8fccdd2a41716eee35fb26db74dfc06d750e251f52a066013b43.jpg)  
+Figure 1: Estimated transfer time (bar chart) and the approximate ratio of transfer time to total training time (line chart) in one round for a low-resource client. Our method significantly reduces communication cost and improve training efficiency.
+
+sends it back to all clients to start the next training round. This characteristic enables FL to get widely applied in real-world scenarios (Ge et al., 2020; Roosta et al., 2021; Passban et al., 2022; Niu and Deng, 2022). In recent years, federated multilingual neural machine translation (Fed-MNMT) has become a new training paradigm and making it feasible for most institutions to train MNMT models (Roosta et al., 2021; Passban et al., 2022). FL makes it possible to leverage corpora from other organizations without privacy problems, solving the problem that training an MNMT model needs to collect large-scale multilingual corpora, which is expensive, time-consuming, and often unaffordable for resource-constrained institutions. Therefore, Fed-MNMT is a secure and cost-effective alternative to conventional centralized training for the optimization of MNMT models.
+
+However, the issue of communication cost is non-negligible when we introduce FL to neural machine translation. Unlike local centralized learning, federated learning requires frequent commu
+
+![](images/654762b00028cd6b7dcdd2ec6fe8445b2d6a360c78d3b69f6aee9d16af880817.jpg)  
+Figure 2: Communication-efficient Fed-MNMT framework with adapter modules and clients clustering strategies.
+
+![](images/0419c8fa556e0460cb1b241bd5da9340181a96ee0db59e4460da29ddcdc70bf9.jpg)
+
+nication of model parameters between the server and clients. Therefore, the communication cost grows rapidly along with the increase in model size. Nowadays, pre-trained language models are widely adopted as backbone models for MNMT, whose parameters are usually over  $10^{8}$ , e.g., 611M for mBART-50 (Tang et al., 2020) and 1.2B for M2M100 (Fan et al., 2020). Considering the increasing number of clients in realistic scenarios as there frequently appear new clients, communication costs will severely hinder the efficient training of the entire Fed-MNMT system and thus make the application of FL to MNMT impractical.
+
+To tackle this problem, we introduce the parameter-efficient tuning idea (Houlsby et al., 2019; Pfeiffer et al., 2021; Karimi Mahabadi et al., 2021) into Fed-MNMT. Specifically, we focus on adapter (Rebuffi et al., 2017; Houlsby et al., 2019), which is a popular technique for efficient tuning that requires only updating a lightweight adapter module. In the training process of adapters, a number of randomly initialized modules are inserted into the backbone models and fine-tuned on new data. Concretely, only the parameters of these modules are updated during training, so the number of parameters needed to be transferred between the server and clients is substantially reduced. As the communication cost before and after introducing adapter illustrated in Figure 1, this approach significantly saves communication costs and enables practical applications of Fed-MNMT.
+
+However, directly adding adapter modules to NMT models results in a performance decline, which is initially observed by Roosta et al. (2021) and also confirmed by our experimental results. This phenomenon is attributed to the divergence of different language pairs. In Fed-MNMT, corpora in diverse languages from different clients
+
+are not independently and identically distributed (Non-I.I.D.), so directly aggregating parameters from clients leads to a decrease in model's performance (Zhao et al., 2018).
+
+Considering the adverse effect of conflicting parameters from diverse languages in Fed-MNMT, we introduce clustering strategies to alleviate this issue. The core idea is to cluster the samples according to the characteristics of their data and only conduct aggregation within each cluster where samples share similar properties. Specifically, we cluster all clients with different language pairs based on the language family, gradient similarity, and random, respectively, and systematically compare the performance of different clustering strategies on multilingual translation benchmarks. Figure 2 gives a general view of our training framework. Our experimental results show that clustering on adapters alleviates the data Non-I.I.D. problem and yields better performance in most cases. Overall, our work opens a new direction for future improvements on Fed-MNMT in the real world.
+
+In conclusion, our primary contributions can be summarized as follows:
+
+- Aware of the communication barrier in the training of Fed-MNMT models, we introduce a practical efficient Fed-MNMT framework to enable real-world applications.  
+- By exploring adapter and clustering strategies for alleviating the undesirable effect of data discrepancy, we achieve comparable results with over  $98\%$  communication cost reduced compared to vanilla Fed-MNMT.
+
+# 2 Methodology
+
+In this section, we first define the Fed-MNMT problem in § 2.1. Next, we elaborate on the adapter
+
+modules and the investigated clustering strategies in  $\S 2.2$  and  $\S 2.3$ , respectively. Last, we provide an analysis of communication costs between the original Fed-MNMT and our method in  $\S 2.4$ .
+
+# 2.1 Problem Formulation
+
+For a Fed-MNMT problem, we suppose that the set of clients is  $\{C_i\}_{i=1}^N$  where  $N > 1$  and client  $C_i$  owns only one language pair  $P_i$ , whose source language and target language are  $src_i$  and  $tgt_i$ , respectively, and corresponding dataset  $D_i = \{x_{ij}, y_{ij}\}_{j=1}^{n_i}$ , where  $n_i$  is the size of  $D_i$ . In each training round, the optimization target for  $C_i$  is minimizing the cross entropy loss between the ground truths  $y_i$  and model's predictions  $\tilde{y}_i$ :
+
+$$
+\mathcal {L} _ {i} = - \sum_ {j = 1} ^ {n _ {i}} \sum_ {k = 1} ^ {l _ {i j}} \log p \left(\tilde {y} _ {i j} ^ {k} = y _ {i j} ^ {k} \mid x _ {i j}\right) \tag {1}
+$$
+
+After the  $t$ -th training round, all clients will deliver local parameters to the server. The server will aggregate these parameters to obtain the initial parameters for the next round's training. A commonly adopted aggregation algorithm is FedAvg (McMahan et al., 2017), where the weighted average of clients' parameters is calculated according to the quantities of local data samples. Let  $\Theta$  denote model parameters, the FedAvg algorithm can be formulated as:
+
+$$
+\Theta^ {t + 1} = \sum_ {i = 1} ^ {N} \frac {n _ {i}}{n} \Theta_ {i} ^ {t}, \tag {2}
+$$
+
+where  $n = \sum_{i=1}^{N} n_i$ . Then the aggregated parameters will be sent back to all clients to initialize their local models for the next round of training.
+
+However, data sizes can vary sharply among low-resource and high-resource languages in FedMNMT and FedAvg cannot deal well with data quantity skew well (Wang et al., 2020). Thus we change FedAvg, calculating the weighted mean of different clients' parameters, to directly calculating the arithmetic mean of parameters:
+
+$$
+\Theta^ {t + 1} = \sum_ {i = 1} ^ {N} \frac {1}{N} \Theta_ {i} ^ {t} \tag {3}
+$$
+
+We refer to this aggregation method as FedMean in our paper.
+
+Considering the size of pre-trained multi-language models, the communication of model
+
+parameters between the server and clients is time-consuming. Inspired by recent progress in parameter efficient tuning, we are interested in whether adapter can be used to improve the efficiency in FL.
+
+# 2.2 Adapter Modules
+
+We introduce bottleneck adapter (Houlsby et al., 2019) into pre-trained multilingual models. Following the settings of Houlsby et al. (2019) and Pfeiffer et al. (2020), we add adapter modules after the self-attention layer and feed-forward network (FFN) layer for each encoder layer and an additional adapter layer after the cross-attention layer for each decoder layer. During training, only the parameters of adapters and layer-norm modules will be updated thus only a small proportion of parameters have to be communicated between the server and clients.
+
+# 2.3 Client Clustering Strategies
+
+Related research (Johnson et al., 2017; First et al., 2016) has shown that parameter sharing among different languages in MNMT boosts the model's performance, especially for those low-resource languages. Motivated by the success of language clustering in MNMT (Tan et al., 2019), we decide to introduce the method of language pairs clustering into the Fed-MNMT problem and we only allow inner-cluster parameters aggregation. Assuming that the multi-language model consists of an encoder and a decoder, we first conduct a clustering algorithm to obtain the encoder clusters set  $G_{e} = \{g_{i}\}_{i=1}^{m_{e}}$  and the decoder clusters set  $G_{d} = \{g_{i}\}_{i=1}^{m_{d}}$ . Each cluster  $g_{i}$  contains the ids of clients in this cluster. Detailed aggregation algorithm is shown in Algorithm 1.
+
+We explore the following three different clustering strategies.
+
+Language families/groups. Chronopoulou et al. (2022) have verified the strategy of sharing parameters within the same language family in the MNMT problem. We decide to use this strategy in the FL setting. We choose 8 languages belonging to 4 different language families from the TED2020 corpus and 10 languages belonging to 4 different language groups, which are all parts of the Indo-European language family, from the Europarl corpus. The clustering of the encoder depends on language families/groups of source languages and the clustering of the decoder is decided by the target languages' families/groups. Languages from the same family
+
+Algorithm 1: Inner-cluster Aggregation  
+Input: Encoder and decoder clusters sets  $G_{e}$  and  $G_{d}$ ;  
+Initial encoder and decoder paras  $\Theta_{e}^{0}$  and  $\Theta_{d}^{0}$ ;  
+Clients set  $\{C_i\}_{i=1}^N$ ;  
+Training round  $T$ .  
+Output: Encoder paras  $\{\Theta_{e,i}^{T}\}_{i=1}^N$ ;  
+Decoder paras  $\{\Theta_{d,i}^{T}\}_{i=1}^N$   
+for  $i$  from 1 to  $N$  do  
+Initialize  $\Theta_{e,i}^{0}$  with  $\Theta_{e}^{0}$ ;  
+Initialize  $\Theta_{d,i}^{0}$  with  $\Theta_{d}^{0}$ ;  
+for  $i$  from 1 to  $N$  do  
+// local update of client  $i$   
+update  $\Theta_{e,i}^{t-1}$  and  $\Theta_{d,i}^{t-1}$  with local data;  
+// inner-cluster aggregation of encoder parameters  
+foreach  $g$  in  $G_{e}$  do  
+ $\Theta_{e,g}^{t} = \sum_{id \in g} \frac{1}{|g|} \Theta_{e,id}^{t-1}$ ;  
+foreach  $id$  in  $g$  do  
+ $\Theta_{e,id}^{t} = \Theta_{e,g}^{t}$ ;  
+// inner-cluster aggregation of decoder parameters  
+foreach  $g$  in  $G_{d}$  do  
+ $\Theta_{d,g}^{t} = \sum_{id \in g} \frac{1}{|g|} \Theta_{d,id}^{t-1}$ ;  
+foreach  $id$  in  $g$  do  
+ $\Theta_{d,id}^{t} = \Theta_{d,g}^{t}$ ;
+
+or group will be clustered into the same group.
+
+Gradients. Unlike in the scene of centralized learning, clustering based on model parameters (Tan et al., 2019) in Fed-MNMT is unfeasible due to privacy problems. Therefore, we use gradients as the basis of the feature for clustering instead. For each language pair, we use a pre-trained multi-language model to acquire an average gradient vector of all data samples, then a clustering algorithm is applied to the gradient vectors in order to separate clients into different groups. The number of parameters we use for gradients clustering is only about  $131\mathrm{K}$  for each client and will hardly introduce any extra communication cost.
+
+Random clustering. We also test randomly separating all clients into different groups as a baseline for clustering strategies. In detail, we uniformly separate the clients and keep the numbers of clusters in the encoder and the decoder the same as those in the language families/groups strategy.
+
+# 2.4 Communication Cost Comparison
+
+Taking the mBART-50 model (Tang et al., 2020), which is a popular pre-trained multi-language model, as an example, the number of parameters
+
+is around 610.9M, which requires about 2.44GB storage space in the FP32 format. In comparison, after adding adapter modules, only about 8M parameters have to be transferred, which will save approximately  $98.7\%$  communication cost.
+
+More concretely, we provide an approximation for the transmission time needed between the server and clients as follows. Assuming the maximum bandwidth of the server is 1000Mbps, the time to transfer the entire mBART model from client to server is around 2.44GB / 1000Mbps = 19.5 seconds. Assuming all clients share the bandwidth, the total transfer time grows linearly with the number of clients. The synchronization process for all clients to finish transferring models to the server will occupy a large proportion of the total training time. In our actual experiments with 12 clients, the theoretical total transferring time is about  $19.5 \times 12 = 234$  seconds. However, for clients with low-resource languages, the training could be finished within 7 minutes, which means transferring time occupies over half of the local training time. By contrast, the time to transfer the adapter's parameters is only about 0.26 seconds, which is negligible compared to local training time thus significantly improving the training efficiency.
+
+# 3 Experimental Setup
+
+# 3.1 Datasets and Evaluation Metrics
+
+We conduct experiments in two different settings: Multi-to-English and Multi-to-Multi (hereinafter referred to as "m2en" and "m2m", respectively). We use the TED2020 corpus (Reimers and Gurevych, 2020) for the m2en setting and the Europarl corpus (Koehn, 2005) for the m2m setting. The TED2020 corpus is extracted from TED speeches and contains over 100 languages around the world. The Europarl corpus is from the proceedings of the European Parliament and contains 21 languages of European countries.
+
+For each language pair $^2$ , we divide the original corpus into training, dev, and test datasets in accordance of the proportion of 6:2:2. We further sample subsets from the divided datasets. To simulate the scene of low-resource languages and high-resource languages, the training data size of each language pair varies according to the corresponding original
+
+corpus size. The specific language pairs we use and corresponding data sizes are shown in Appendix A.
+
+In the m2en setting, for clustering strategies based on language families/groups and random shuffle, clustering algorithms are only applied to the encoder and all clients share the decoder's parameters because their target languages are all English. But for clustering based on gradients, we also cluster the decoder's parameters into different groups and the number of groups stays the same as that in the encoder. In the m2m setting, clustering is conducted for both the encoder and the decoder. Meanwhile, the numbers of groups in the encoder and decoder are the same in all clustering strategies. We will provide a further analysis of the clustering strategies of the m2m setting in  $\S 4.2$ .
+
+We choose the BLEU score as the evaluation metric using the SacreBLEU (Post, 2018) package. Aside from the BLEU score on each language pair, we additionally report the macro average and micro average scores on all language pairs.
+
+# 3.2 Baselines
+
+We evaluate the following methods as baselines:
+
+Centralized-model. The results of centralized training, where data from all clients are gathered together, using the original multi-language model without extra modules.
+
+Centralized-adapter. The results of centralized training using the multi-language model with adapter modules.
+
+Adapter-local. We train a model for each client using local data without parameter aggregation with other language pairs.
+
+Model-fed. We train the original multi-language model without Adapter modules under the federated learning framework, where the parameters are shared among all clients using the aggregation algorithm in Eq. (3) without any clustering strategies.
+
+Adapter-fed. In this method, adapter modules are attached to the backbone model, while the rest settings are the same as those in model-fed. This baseline corresponds to the scene of directly introducing adapter without any clustering strategies.
+
+# 3.3 Training Setup
+
+We choose the mBART-50 pre-trained model as our backbone model. To fairly compare the training and communication costs of different methods,
+
+we train each model for 5 rounds. We select the checkpoint with the lowest loss on the dev set and evaluate it on the test set. Parameters are aggregated every time all clients finish an epoch of local training. For every client, the batch size is 8 and the local model is updated every 16 steps. The local learning rate is  $5 \times 10^{-5}$  for the mBART model and  $1 \times 10^{-3}$  for the models with adapter modules. The hidden size of adapter modules is 64.
+
+For all experiments, we train the model with 3 random seeds and report the average scores. For the random clustering strategy, the clustering groups are different when using different random seeds.
+
+# 4 Experimental Results
+
+# 4.1 Primary Results and Findings
+
+The experimental results in the m2en and m2m settings are shown in Table 1 and Table 2, respectively. In general, directly adding adapter modules leads to a performance drop (comparing adapter-fed and model-fed) and methods with clustering strategies all achieve better performance than the direct baseline adapter-fed, indicating the ability of our clustering strategies to alleviate data discrepancy. In both settings, adapter-families method performs best in macro and micro average scores among three clustering strategies, even surpassing model-fed in the m2m setting.
+
+It is noteworthy that the clustering strategies acquire more significant performance improvements in the m2m setting than in the m2en setting. The problem of conflicting parameters is more settle-some in the m2m setting because there exist more kinds of languages (especially target languages). Thus introducing clustering strategies will bring more benefit to m2m translation tasks.
+
+Meanwhile, we notice that our clustering strategies fail to beat adapter-local in the m2m setting. This can also be explained by the difference in the difficulty of tasks. In the more complicated m2m translation task, more elaborate clustering strategies should be designed to fully make advantage of other language pairs and avoid the influence of conflicting parameters. However, we bring the ability of multi-language translation to these clients through FL with an acceptable drop in performance compared to adapter-local.
+
+# 4.2 Ablation Study
+
+In the m2m setting, the clustering of the adapter modules attached to the encoder and the decoder
+
+<table><tr><td>Method</td><td>Comm. Cost</td><td>zh-en</td><td>th-en</td><td>ar-en</td><td>he-en</td><td>fi-en</td><td>et-en</td><td>ru-en</td><td>sl-en</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>centralized-model</td><td>N / A</td><td>24.72</td><td>28.97</td><td>38.29</td><td>43.59</td><td>32.81</td><td>32.70</td><td>30.14</td><td>47.92</td><td>34.89</td><td>32.33</td></tr><tr><td>centralized-adapter</td><td>N / A</td><td>24.97</td><td>21.47</td><td>39.02</td><td>45.05</td><td>33.62</td><td>32.62</td><td>30.65</td><td>50.45</td><td>34.73</td><td>32.03</td></tr><tr><td>adapter-local</td><td>N / A</td><td>25.16</td><td>21.68</td><td>39.46</td><td>44.32</td><td>33.12</td><td>32.64</td><td>30.58</td><td>50.93</td><td>34.73</td><td>32.15</td></tr><tr><td>model-fed</td><td>611M</td><td>25.31</td><td>23.41</td><td>39.54</td><td>45.13</td><td>32.87</td><td>33.27</td><td>30.77</td><td>51.85</td><td>35.27</td><td>32.55</td></tr><tr><td>adapter-fed</td><td>8M</td><td>25.12</td><td>16.64</td><td>39.62</td><td>44.93</td><td>33.14</td><td>32.66</td><td>30.41</td><td>53.62</td><td>34.52</td><td>31.71</td></tr><tr><td>adapter-random</td><td>8M</td><td>25.37</td><td>21.48</td><td>39.61</td><td>44.82</td><td>33.26</td><td>33.21</td><td>30.75</td><td>52.64</td><td>35.14</td><td>32.38</td></tr><tr><td>adapter-gradients</td><td>8M</td><td>25.26</td><td>21.26</td><td>39.39</td><td>44.64</td><td>33.62</td><td>33.14</td><td>30.72</td><td>51.16</td><td>34.90</td><td>32.21</td></tr><tr><td>adapter-families</td><td>8M</td><td>25.28</td><td>21.58</td><td>39.45</td><td>44.70</td><td>33.87</td><td>33.23</td><td>30.92</td><td>52.64</td><td>35.21</td><td>32.40</td></tr></table>
+
+Table 1: BLEU scores on the TED2020 corpus. Comm. Cost, which is short for communication cost, denotes the number of parameters communicated between the server and each client. Adapter-random, adapter-gradients, and adapter-families refer to clustering strategies of random clustering, gradients, and language families/groups, respectively. The best result of each language pair is highlighted in **bold** (only methods with adapter modules trained in the FL setting are considered).  
+
+<table><tr><td>Method</td><td>Comm. Cost</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-sl</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>centralized-model</td><td>N / A</td><td>30.43</td><td>19.15</td><td>28.86</td><td>23.20</td><td>19.96</td><td>27.17</td><td>40.35</td><td>33.20</td><td>21.24</td><td>21.84</td><td>23.63</td><td>20.57</td><td>25.80</td><td>26.10</td></tr><tr><td>centralized-adapter</td><td>N / A</td><td>30.59</td><td>19.07</td><td>29.65</td><td>22.92</td><td>18.68</td><td>27.73</td><td>41.52</td><td>33.20</td><td>21.45</td><td>22.16</td><td>23.40</td><td>20.85</td><td>25.93</td><td>26.19</td></tr><tr><td>adapter-local</td><td>N / A</td><td>30.88</td><td>19.19</td><td>30.31</td><td>23.50</td><td>20.01</td><td>28.12</td><td>41.84</td><td>33.39</td><td>21.13</td><td>22.27</td><td>23.62</td><td>21.05</td><td>26.28</td><td>26.56</td></tr><tr><td>model-fed</td><td>611M</td><td>30.41</td><td>17.60</td><td>29.62</td><td>19.76</td><td>13.41</td><td>28.01</td><td>39.77</td><td>32.25</td><td>21.10</td><td>21.94</td><td>20.09</td><td>19.92</td><td>24.49</td><td>24.68</td></tr><tr><td>adapter-fed</td><td>8M</td><td>29.75</td><td>17.47</td><td>30.08</td><td>16.92</td><td>11.85</td><td>28.01</td><td>38.18</td><td>31.06</td><td>20.18</td><td>21.23</td><td>18.02</td><td>19.97</td><td>23.56</td><td>23.53</td></tr><tr><td>adapter-random</td><td>8M</td><td>30.90</td><td>18.57</td><td>30.04</td><td>22.39</td><td>16.53</td><td>28.25</td><td>36.97</td><td>33.04</td><td>21.13</td><td>22.28</td><td>22.72</td><td>20.37</td><td>25.27</td><td>25.67</td></tr><tr><td>adapter-gradients</td><td>8M</td><td>30.14</td><td>19.69</td><td>30.19</td><td>19.53</td><td>19.14</td><td>28.06</td><td>41.73</td><td>33.31</td><td>21.51</td><td>21.60</td><td>19.95</td><td>21.29</td><td>25.51</td><td>25.35</td></tr><tr><td>adapter-families</td><td>8M</td><td>30.60</td><td>19.31</td><td>30.12</td><td>22.65</td><td>16.69</td><td>27.99</td><td>41.33</td><td>33.21</td><td>21.31</td><td>22.04</td><td>22.68</td><td>21.21</td><td>25.76</td><td>26.04</td></tr></table>
+
+Table 2: BLEU scores on the Europarl corpus. The meanings of symbols stay the same as those in Table 1.
+
+are independent. To further explore the specific influence of clustering in these two modules on the model's performance, we apply the clustering strategy to only the encoder and the decoder separately and show the results in Figure 3. We use language families/groups as the clustering strategy. All methods are trained in the m2m setting using the Europarl dataset with one random seed. Other settings are the same as those in our main experiments.
+
+To our surprise, either clustering in the encoder or the decoder significantly improves performance compared to no-sharing strategies, even surpasses adapter-families. We owe this phenomenon to our naive clustering strategies in the encoder and the decoder. In adapter-families, the clustering of the encoder and the decoder is only related to the source and target languages respectively. However, parameters in both the encoder and the decoder are influenced by source and target languages together during training. The inconsistency between clustering strategies and parameter update results in adapter-families method's worse performance than adapter-encoder and adapter-decoder.
+
+# 5 Further Analysis
+
+# 5.1 Case Study
+
+We select some representative cases of translations from method adapter-families and adapter
+
+![](images/a1721ffc51fcbabb4f31b7005168dd3bf132ac5ba2fba58976daddb2ea6c23a3.jpg)  
+Figure 3: Results of ablation study. Encoder/Decoder-cluster denote clustering only in the encoder/decoder, both obtaining significant improvements compared to the no-clustering baseline.
+
+shareAll, which are shown in Table 3, to further study the influence of our clustering strategies. We separate the mistakes into three categories.
+
+Opposite semantics In case 1 and case 2, adapter-fed misses negation adverbs in predictions and results in totally opposite semantics.
+
+Inaccurate words/phrases In case 3, adapter-fed translates adverbial of time into "a day or two", which should be "four or five weeks" actually. In case 4, adapter-fed uses the expression "save the world", which differs from the original expression "change the world" in semantics.
+
+<table><tr><td>Method</td><td>Sentence</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>I think if somebody tells a lie, they&#x27;re not just a liar. 
+I think if somebody tells a lie, they&#x27;re not just a liar. 
+I know that when people lie, they&#x27;re just lying.</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>I never miss a single training session. 
+I didn&#x27;t miss a single training session. 
+I missed one training session.</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>And within four or five weeks, he can do it again. 
+And within four or five weeks, he can do it again. 
+And within a day or two, he can do this.</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>So could art change the world? 
+Can art change the world? 
+Is art about saving the world?</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>Over 100,000 children learn science this way. 
+Hundreds of thousands of children learn science this way. 
+Hundreds of thousands of people learned how to do that.</td></tr><tr><td>Ground truth 
+adapter-families 
+adapter-fed</td><td>And all at once I became a learner. 
+And all of a sudden, I became a learner. 
+And all of a sudden, it&#x27;s happening.</td></tr></table>
+
+Table 3: Case study comparison of adapter-families and adapter-fed. Obvious translation mistakes are highlighted in red and corresponding correct translations are highlighted in blue.  
+
+<table><tr><td rowspan="2" colspan="2">Method</td><td colspan="2">TED2020</td><td colspan="2">Europarl</td></tr><tr><td>Macro Avg.</td><td>Micro Avg.</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td rowspan="2">adapter-fed</td><td>FedAvg</td><td>34.31</td><td>31.70</td><td>22.20</td><td>23.00</td></tr><tr><td>FedMean</td><td>34.52</td><td>31.71</td><td>23.56</td><td>23.53</td></tr><tr><td rowspan="2">adapter-random</td><td>FedAvg</td><td>34.79</td><td>32.16</td><td>25.19</td><td>25.76</td></tr><tr><td>FedMean</td><td>35.14</td><td>32.38</td><td>25.27</td><td>25.67</td></tr><tr><td rowspan="2">adapter-gradients</td><td>FedAvg</td><td>34.46</td><td>31.96</td><td>25.64</td><td>25.84</td></tr><tr><td>FedMean</td><td>34.90</td><td>32.21</td><td>25.51</td><td>25.35</td></tr><tr><td rowspan="2">adapter-families</td><td>FedAvg</td><td>34.83</td><td>32.09</td><td>25.47</td><td>26.02</td></tr><tr><td>FedMean</td><td>35.21</td><td>32.40</td><td>25.76</td><td>26.04</td></tr></table>
+
+Table 4: Comparison between FedAvg and FedMean (the aggregation method we use in Eq. 3). Results of FedMean are in green if better than those of FedAvg, otherwise in red. In most cases, FedMean yields performance gains.
+
+Ambiguous semantics In cases 5 and 6, adapter-fed loses specific semantic information in ground truths. It fails to properly translate "children", "science" and "become a learner" and uses more ambiguous expressions instead.
+
+In comparison, adapter-families makes more accurate predictions in the above cases, which suggests that appropriate clustering strategies help the model produce better translations with improvements in semantics.
+
+# 5.2 Both FedMean and Clustering Contribute
+
+In our experiments, discrepancies in data come from two aspects: data quantity skew and linguistic discrepancy (language difference). We adjust the aggregation algorithm from Eq. (2) to Eq. (3), which we call FedMean here, to tackle the problem of quantity skew. Moreover, we propose clustering strategies to prevent clients from receiving con
+
+flecting parameters from dissimilar languages. To explore how these two methods contribute to improvement in performance, we further conduct experiments with the aggregation algorithm changed to FedAvg (see Eq. (2)) and keep other training settings unchanged.
+
+As the results displayed in Table 4, clustering strategies bring more significant improvements in performance on the Europarl corpus than the TED2020 corpus. Since experiments on the Europarl corpus consists of more different languages and are conducted in a more complicated m2m setting, the problem of linguistic discrepancy is more severe for Europarl. For the TED2020 corpus, changing the aggregation algorithm from FedAvg to FedMean leads to more significant improvements for methods with clustering strategies compared to adapter-fed. In contrast, for the Europarl corpus, adapter-fed substantially benefits from FedMean, while FedMean hardly brings any benefits to methods with clustering strategies, even causing performance drop in some cases.
+
+Based on these observations, we contend that both aggregation algorithms and clustering strategies contribute to performance gain by alleviating data discrepancies. The specific extent of improvement depends on the extent of data quantity skew and linguistic discrepancy.
+
+# 5.3 Further Cost Saving by Adapter Pruning
+
+On top of the adapter tuning approach, adapter pruning techniques (Rücklé et al., 2021; Pfeiffer et al., 2021; Karimi Mahabadi et al., 2021) further compress the number of parameters to be updated. To further reduce the communication cost, we conduct an exploratory attempt to prune parts of adapter modules in both the encoder and the decoder. We still choose the mBART-50 model, with 12 layers in the encoder and the decoder separately, to conduct the experiments. Specifically, we evenly separate all adapter modules we add in mBART into three parts: input-end adapters (adapter modules in the first 4 layers of the encoder or the decoder), middle-layer adapters (adapter modules in layers 5 to 8 of the encoder or the decoder), output-end adapters (adapter modules in the last 4 layers of the encoder or the decoder). In each strategy, only one part of the adapter modules is kept, so the communication cost is saved by two-thirds. We use adapter-families as the baseline and train all models with one random seed. The rest settings stay the
+
+<table><tr><td>Method</td><td>adapters</td><td>Comm. Cost</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-s1</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td rowspan="4">Language families clustering</td><td>input-end</td><td>2.7M</td><td>30.21</td><td>19.04</td><td>29.98</td><td>20.39</td><td>14.93</td><td>28.09</td><td>41.64</td><td>32.08</td><td>21.39</td><td>21.62</td><td>19.89</td><td>20.73</td><td>25.00</td><td>25.12</td></tr><tr><td>middle-layer</td><td>2.7M</td><td>30.12</td><td>18.88</td><td>29.84</td><td>21.27</td><td>15.15</td><td>28.16</td><td>41.33</td><td>31.92</td><td>21.06</td><td>21.68</td><td>20.99</td><td>20.74</td><td>25.10</td><td>25.30</td></tr><tr><td>output-end</td><td>2.7M</td><td>30.30</td><td>19.37</td><td>30.33</td><td>22.17</td><td>16.06</td><td>27.96</td><td>41.39</td><td>31.76</td><td>20.03</td><td>21.75</td><td>21.97</td><td>21.15</td><td>25.35</td><td>25.64</td></tr><tr><td>all adapters</td><td>8.1M</td><td>30.18</td><td>18.85</td><td>29.86</td><td>22.42</td><td>16.10</td><td>27.66</td><td>41.13</td><td>32.96</td><td>20.78</td><td>21.80</td><td>22.78</td><td>21.27</td><td>25.48</td><td>25.75</td></tr></table>
+
+Table 5: BLEU scores of different adapter pruning strategies. We acquire similar results with only 1/3 communication cost compared to keeping all adapter modules.
+
+same as those in previous main experiments. The results are shown in Table 5.
+
+It is encouraging that pruning adapters do not result in a sharp decrease in performance. We observe that keeping output-end adapters achieves the highest score among the three pruning strategies, which suggests that adapters in the top layers play more important roles. Overall, the results indicate that it is possible to further reduce communication costs and it is worthwhile to explore more elaborate pruning techniques in future work.
+
+# 6 Related Work
+
+Federated Learning was first proposed by McMahan et al. (2017) as a decentralized training framework. Due to its decentralized and private nature, FL shows great potential in actual applications. Recently, there has been a surge in the NLP community to explore the application of federated learning in diverse NLP tasks, such as emojis prediciton (Gandhi et al., 2022), named entity recognition (Ge et al., 2020), and machine translation (Roosta et al., 2021; Passban et al., 2022; Weller et al., 2022), etc. Roosta et al. (2021) first applied FL to NMT tasks. However, training language models in the FL setting brings huge communication overheads. To solve this problem, researchers have proposed to only exchange some dedicated "Controller" (Roosta et al., 2021) layers between the server and clients. Moreover, Passban et al. (2022) introduced parameter pruning strategies to reduce communication bandwidth. Our methods with adapter modules have advantages in communication efficiency with fewer parameters to be transferred compared to Controller (see Appendix D), and other parameter pruning strategies can also be applied to our adapter modules to further reduce communication costs.
+
+Multilingual Neural Machine Translation (MNMT) trains a single model to handle translation between multiple language pairs (Johnson et al., 2017; Aharoni et al., 2019; Zhang et al., 2020). Moreover, MNMT significantly reduces training and inference costs by eliminating the
+
+need to train models for each language pair. Massively pre-trained multilingual models have been used for MNMT, such as mBART-50 (Tang et al., 2020) and M2M100 (Fan et al., 2021). In recent years, adapter has become a popular method in MNMT (Bapna and First, 2019; Cooper Stickland et al., 2021; Philip et al., 2020; Üstün et al., 2021; Chronopoulou et al., 2022) due to its high parameter efficiency and transferability between tasks.
+
+Different from previous works on this topic, inspired by recent progress in improving the efficiency of NLP methods (Strubell et al., 2019; Li et al., 2021a; Xu et al., 2021; Li et al., 2021b), we focus on communication efficiency in FL-MNMT and make the first effort to introduce adapter modules in order to reduce communication costs. We also apply different clustering strategies to resolve the issue of conflicting parameters stemming from data discrepancy.
+
+# 7 Conclusion
+
+In this paper, we introduce adapter modules to PLMs for the Fed-MNMT problem to boost communication efficiency. We reduce the communication cost by over  $98\%$  and make the training process of Fed-MNMT practical. To deal with the problem of performance drop after introducing adapter modules, we propose different clustering strategies to separate clients into different groups to avoid the negative influence of data discrepancy. We surpass the direct baseline with a substantial gap, especially in the more complicated multi-to-multi translation setting.
+
+Furthermore, our analytic experiments indicate that both aggregation algorithms in server and clustering strategies affect the performance of FedMNMT. We also explore the possibility of further reducing communication costs by pruning adapter modules and find that adapters in top layers are more significant for translation performance.
+
+In future work, we will explore more well-designed clustering strategies and attach other parameter-efficient techniques to adapter to further reduce the parameters to be transferred.
+
+# Limitations
+
+First, in this work, we assume that clustering in the encoder and the decoder is only related to the source and target languages, respectively. Actually, both parameters in the encoder and the decoder are influenced by source and target languages simultaneously. Therefore, our assumption may lead to a performance drop. In future work, we plan to explore more complicated clustering strategies.
+
+Moreover, our adapter-families method depends on prior linguistic knowledge. Its actual effectiveness can be affected by the distribution of language families/groups in clients. Our methods mainly apply to comparably uniform language distribution.
+
+In addition, the effectiveness of our methods on other PLMs needs to be verified. However, it is easy to transfer our methods to other models so it will not be a challenging problem.
+
+# Acknowledgements
+
+We thank all reviewers for their insightful comments and suggestions. This work is supported by Natural Science Foundation of China (NSFC) No.62176002. We sincerely thank Jingjing Xu for her valuable suggestions. Xu Sun is the corresponding author.
+
+# References
+
+Roee Aharoni, Melvin Johnson, and Orhan First. 2019. Massively multilingual neural machine translation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3874-3884, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Ankur Bapna and Orhan First. 2019. Simple, scalable adaptation for neural machine translation. 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 1538-1548, Hong Kong, China. Association for Computational Linguistics.  
+Alexandra Chronopoulou, Dario Stojanovski, and Alexander Fraser. 2022. Language-family adapters for multilingual neural machine translation. ArXiv preprint, abs/2209.15236.  
+Asa Cooper Stickland, Xian Li, and Marjan Ghazvininejad. 2021. Recipes for adapting pre-trained monolingual and multilingual models to machine translation. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational
+
+Linguistics: Main Volume, pages 3440-3453, Online.  
+Association for Computational Linguistics.  
+Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, and Armand Joulin. 2020. Beyond english-centric multilingual machine translation.  
+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. J. Mach. Learn. Res., 22(107):1-48.  
+Orhan First, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine translation with a shared attention mechanism. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San Diego, California. Association for Computational Linguistics.  
+Deep Gandhi, Jash Mehta, Nirali Parekh, Karan Waghela, Lynette D'Mello, and Zeerak Talat. 2022. A federated approach to predicting emojis in hindi tweets. ArXiv preprint, abs/2211.06401.  
+Suyu Ge, Fangzhao Wu, Chuhan Wu, Tao Qi, Yongfeng Huang, and Xing Xie. 2020. Fedner: Privacy-preserving medical named entity recognition with federated learning. arXiv e-prints, pages arXiv-2003.  
+Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.  
+Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Google's multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association for Computational Linguistics, 5:339-351.  
+Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. 2021. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035.  
+Philipp Koehn. 2005. Europarl: A parallel corpus for statistical machine translation. In Proceedings of Machine Translation Summit X: Papers, pages 79-86, Phuket, Thailand.
+
+Lei Li, Yankai Lin, Deli Chen, Shuhuai Ren, Peng Li, Jie Zhou, and Xu Sun. 2021a. Cascadebert: Accelerating inference of pre-trained language models via calibrated complete models cascade. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 475-486.  
+Lei Li, Yankai Lin, Shuhuai Ren, Peng Li, Jie Zhou, and Xu Sun. 2021b. Dynamic knowledge distillation for pre-trained language models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 379-389.  
+Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas. 2017. Communication-efficient learning of deep networks from decentralized data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics, AISTATS 2017, 20-22 April 2017, Fort Lauderdale, FL, USA, volume 54 of Proceedings of Machine Learning Research, pages 1273-1282. PMLR.  
+Yifan Niu and Weihong Deng. 2022. Federated learning for face recognition with gradient correction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1999-2007.  
+Peyman Passban, Tanya Roosta, Rahul Gupta, Ankit Chadha, and Clement Chung. 2022. Training mixed-domain translation models via federated learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2576-2586, Seattle, United States. Association for Computational Linguistics.  
+Jonas Pfeiffer, Aishwarya Kamath, Andreas Rückle, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-destructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 487-503, Online. Association for Computational Linguistics.  
+Jonas Pfeiffer, Andreas Rückle, Clifton Poth, Aishwarya Kamath, Ivan Vulic, Sebastian Ruder, Kyunghyun Cho, and Iryna Gurevych. 2020. AdapterHub: A framework for adapting transformers. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 46-54, Online. Association for Computational Linguistics.  
+Jerin Philip, Alexandre Berard, Matthias Galle, and Laurent Besacier. 2020. Monolingual adapters for zero-shot neural machine translation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4465-4470, Online. 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, pages 186-191, Brussels, Belgium. Association for Computational Linguistics.  
+Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. 2017. Learning multiple visual domains with residual adapters. 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 506-516.  
+Nils Reimers and Iryna Gurevych. 2020. Making monolingual sentence embeddings multilingual using knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4512-4525, Online. Association for Computational Linguistics.  
+Tanya Roosta, Peyman Passban, and Ankit Chadha. 2021. Communication-efficient federated learning for neural machine translation. ArXiv preprint, abs/2112.06135.  
+Andreas Rückle, Gregor Geigle, Max Glockner, Tilman Beck, Jonas Pfeiffer, Nils Reimers, and Iryna Gurevych. 2021. AdapterDrop: On the efficiency of adapters in transformers. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7930-7946, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and policy considerations for deep learning in nlp. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3645-3650.  
+Xu Tan, Jiale Chen, Di He, Yingce Xia, Tao Qin, and Tie-Yan Liu. 2019. Multilingual neural machine translation with language clustering. 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 963-973, Hong Kong, China. Association for Computational Linguistics.  
+Y. Tang, C. Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2020. Multilingual translation with extensible multilingual pretraining and finetuning. ArXiv, abs/2008.00401.  
+Ahmet Üstün, Alexandre Berard, Laurent Besacier, and Matthias Galle. 2021. Multilingual unsupervised neural machine translation with denoising adapters. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6650-6662, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Jianyu Wang, Qinghua Liu, Hao Liang, Gauri Joshi, and H. Vincent Poor. 2020. Tackling the objective
+
+inconsistency problem in heterogeneous federated optimization. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
+
+Orion Weller, Marc Marone, Vladimir Braverman, Dawn Lawrie, and Benjamin Van Durme. 2022. Pretrained models for multilingual federated learning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1413-1421, Seattle, United States. Association for Computational Linguistics.
+
+Jingjing Xu, Wangchunshu Zhou, Zhiyi Fu, Hao Zhou, and Lei Li. 2021. A survey on green deep learning. arXiv preprint arXiv:2111.05193.
+
+Biao Zhang, Philip Williams, Ivan Titov, and Rico Senrich. 2020. Improving massively multilingual neural machine translation and zero-shot translation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1628-1639, Online. Association for Computational Linguistics.
+
+Yue Zhao, Meng Li, Liangzhen Lai, Naveen Suda, Damon Civin, and Vikas Chandra. 2018. Federated learning with non-iid data. ArXiv preprint, abs/1806.00582.
+
+# A Training Data Sizes
+
+The specific training data size of each language pair is shown in Table 6.
+
+# B Complete Results of Ablation Study
+
+We show the complete results on all language pairs of the ablation study in Table 7.
+
+We find that only applying clustering to the decoder acquires the highest scores on 7 out of the total 12 language pairs. To our surprise, adapter-families method fails to reach the best performance in average scores, which has been explained in § 4.2 in the main text.
+
+# C Uniform Data Distribution
+
+We also conduct experiments on TED2020 with each client owning training data of equal size. The results are shown in Table 8. We observe that the performance gaps between different methods are similar to those in Table 1. Notably, Adapter-families beats adapter-random by a slight margin. Both clustering strategies acquire obvious performance improvement compared to the baseline adapter-fed. These empirical results verify that our methods apply to various data distributions.
+
+<table><tr><td>Corpus</td><td>Source Language Family/Group</td><td>Language Pair</td><td>Dataset Size</td></tr><tr><td rowspan="8">TED2020</td><td rowspan="2">Sino-Tibetan</td><td>zh-&gt;en</td><td>9984</td></tr><tr><td>th-&gt;en</td><td>4992</td></tr><tr><td rowspan="2">Afro-asiatic</td><td>ar-&gt;en</td><td>9984</td></tr><tr><td>he-&gt;en</td><td>1920</td></tr><tr><td rowspan="2">Uralic</td><td>fi-&gt;en</td><td>1920</td></tr><tr><td>et-&gt;en</td><td>1920</td></tr><tr><td rowspan="2">Indo-European</td><td>ru-&gt;en</td><td>9984</td></tr><tr><td>sl-&gt;en</td><td>1920</td></tr><tr><td rowspan="12">Europarl</td><td rowspan="3">Germanic</td><td>de-&gt;fr</td><td>11648</td></tr><tr><td>nl-&gt;pl</td><td>3584</td></tr><tr><td>en-&gt;lt</td><td>3712</td></tr><tr><td rowspan="3">Romance</td><td>fr-&gt;nl</td><td>12160</td></tr><tr><td>it-&gt;sl</td><td>3456</td></tr><tr><td>es-&gt;lv</td><td>3584</td></tr><tr><td rowspan="3">Slavic</td><td>pl-&gt;en</td><td>3712</td></tr><tr><td>sl-&gt;es</td><td>3584</td></tr><tr><td>sl-&gt;lt</td><td>3584</td></tr><tr><td rowspan="3">Baltic</td><td>lt-&gt;de</td><td>3328</td></tr><tr><td>lv-&gt;it</td><td>3584</td></tr><tr><td>lv-&gt;pl</td><td>3712</td></tr></table>
+
+Table 6: Detailed data sizes of language pairs from Europarl corpus.
+
+# D Comparison to Controllers
+
+Controllers (Roosta et al., 2021) only exchange 8 layers in a 32-layer Transformer (4 from encoder and 4 from decoder) between the server and clients, which means that they reduce the communication cost by approximately  $66\%$  (the number of layers in the original model without Controllers is 24). Compared with Controllers, we introduce adapter modules in Fed-MNMT without the need to define additional layers. Besides, our methods transmit a much smaller amount of parameters in client-to-server exchanges than using Controllers. Therefore, our proposal is superior to Controllers in terms of communication efficiency.
+
+<table><tr><td>Method</td><td>de-fr</td><td>nl-pl</td><td>en-lt</td><td>fr-nl</td><td>it-sl</td><td>es-lv</td><td>pl-en</td><td>sl-es</td><td>sl-lt</td><td>lt-de</td><td>lv-it</td><td>lv-pl</td><td>Macro Avg.</td><td>Micro Avg.</td></tr><tr><td>adapter-fed</td><td>29.66</td><td>17.35</td><td>30.22</td><td>16.89</td><td>11.95</td><td>27.91</td><td>37.31</td><td>31.04</td><td>20.13</td><td>21.18</td><td>17.95</td><td>19.97</td><td>23.46</td><td>23.44</td></tr><tr><td>+ encoder clustering</td><td>30.50</td><td>19.24</td><td>29.85</td><td>22.94</td><td>18.56</td><td>28.01</td><td>41.93</td><td>33.25</td><td>20.84</td><td>22.09</td><td>23.34</td><td>20.82</td><td>25.95</td><td>26.19</td></tr><tr><td>+ decoder clustering</td><td>31.19</td><td>19.41</td><td>30.53</td><td>23.06</td><td>18.26</td><td>28.20</td><td>41.71</td><td>33.56</td><td>21.26</td><td>21.96</td><td>23.03</td><td>21.06</td><td>26.10</td><td>26.42</td></tr><tr><td>adapter-families</td><td>30.18</td><td>18.85</td><td>29.86</td><td>22.42</td><td>16.10</td><td>27.66</td><td>41.13</td><td>32.96</td><td>20.78</td><td>21.80</td><td>22.78</td><td>21.27</td><td>25.48</td><td>25.75</td></tr></table>
+
+Table 7: BLEU scores of applying language families strategies to only the encoder and the decoder. Adapter-fed corresponds to clustering in neither the encoder or the decoder. Adapter-families denotes clustering in both the encoder and the decoder.  
+
+<table><tr><td>Method</td><td>zh-en</td><td>th-en</td><td>ar-en</td><td>he-en</td><td>fi-en</td><td>et-en</td><td>ru-en</td><td>sl-en</td><td>Avg.</td></tr><tr><td>model-fed</td><td>25.58</td><td>22.90</td><td>39.63</td><td>46.12</td><td>34.78</td><td>34.41</td><td>30.42</td><td>51.32</td><td>35.64</td></tr><tr><td>adapter-fed</td><td>24.83</td><td>16.79</td><td>40.02</td><td>45.99</td><td>33.55</td><td>33.67</td><td>29.61</td><td>52.20</td><td>34.58</td></tr><tr><td>adapter-random</td><td>25.30</td><td>22.33</td><td>39.58</td><td>45.59</td><td>34.58</td><td>34.31</td><td>30.12</td><td>51.52</td><td>35.41</td></tr><tr><td>adapter-families</td><td>25.04</td><td>22.08</td><td>39.50</td><td>45.74</td><td>34.91</td><td>34.65</td><td>30.16</td><td>51.38</td><td>35.43</td></tr></table>
+
+Table 8: BLEU scores on the TED2020 corpus with equivalent data size for each language pair. Best scores are highlighted in bold except model-fed.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work? section Limitations  
+A2. Did you discuss any potential risks of your work? Not applicable. Left blank.  
+A3. Do the abstract and introduction summarize the paper's main claims? abstract and section 1  
+A4. Have you used AI writing assistants when working on this paper? Left blank.
+
+B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used? No response.  
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.  
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.  
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.  
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.  
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response.
+
+C Did you run computational experiments?
+
+section 3, 4 and 5
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? section 4
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? section 3  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? section 4 and 5  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? section 3.1
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response.

2023/Communication Efficient Federated Learning for Multilingual Neural Machine Translation with Adapter/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:149ff1271ca5e991b46489866b3da7e9822d81f1e6447d939a8befe7e8354683
+size 496507

2023/Complementary Explanations for Effective In-Context Learning/a50d012a-1115-4961-8cce-819a0bcf40d0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4661693af5c7a2b817c42554319ca239aacda0b923de90fe3ee0ed010aa2ed7f
+size 487347

2023/Complementary Explanations for Effective In-Context Learning/full.md

@@ -0,0 +1,469 @@
+# Complementary Explanations for Effective In-Context Learning
+
+Xi Ye\* Srinivasan Iyer\* Asli Celikyilmaz\* Ves Stoyanov\* Greg Durrett\* Ramakanth Pasunuru
+
+The University of Texas at Austin  $\spadesuit$  Meta AI
+
+$\diamondsuit$  {xiye,gdurrett}@cs.utexas.edu
+
+$\spadesuit$  {sviyer,ves,aslic,rpasunuru}@meta.com
+
+# Abstract
+
+Large language models (LLMs) have exhibited remarkable capabilities in learning from explanations in prompts, but there has been limited understanding of exactly how these explanations function or why they are effective. This work aims to better understand the mechanisms by which explanations are used for in-context learning. We first study the impact of two different factors on the performance of prompts with explanations: the computation trace (the way the solution is decomposed) and the natural language used to express the prompt. By perturbing explanations on three controlled tasks, we show that both factors contribute to the effectiveness of explanations. We further study how to form maximally effective sets of explanations for solving a given test query. We find that LLMs can benefit from the complementarity of the explanation set: diverse reasoning skills shown by different exemplars can lead to better performance. Therefore, we propose a maximal marginal relevance-based exemplar selection approach for constructing exemplar sets that are both relevant as well as complementary, which successfully improves the in-context learning performance across three real-world tasks on multiple LLMs.
+
+# 1 Introduction
+
+Large language models (LLMs) have achieved promising progress in learning from only a few exemplars in prompts via in-context learning (ICL) (Brown et al., 2020; Chowdhery et al., 2022). To scale to complex tasks, recent work in the past year has shown that LLMs can benefit from explanations in prompts, particularly for tasks involving multistep reasoning (Nye et al., 2021; Wei et al., 2022; Wang et al., 2022b; Madaan et al., 2022; Jung et al., 2022). However, while including explanations in prompts has been demonstrated to be useful, little has been shown regarding what particular features make them effective and how they function in ICL.
+
+![](images/da2ba70a1149c206dc2f698e3fe66dfd881ae5ed120c0bb03f44285bbd74fe36.jpg)  
+Figure 1: Prompting OPT (Zhang et al., 2022) with explanations where we perturb the computation traces or natural language. Perturbing either traces or natural language will lead to performance degradation.
+
+Our work aims to better understand the mechanisms by which explanations are used for ICL. As shown in the example for a multi-step reasoning task in Figure 1, we view an explanation as a combination of a computation trace and natural language which glues together the states in the trace. We design a series of probing experiments that perturb the explanations (as shown in Figure 2) and test LLMs' performance to understand the sensitivity of LLMs on these two factors. The results suggest that both factors contribute to making effective explanations, as LLMs see substantial performance degradation when prompted with defective explanations. Nonetheless, incomplete explanations are still beneficial compared to no explanations at all (Figure 1). This suggests that LLMs "faithfully" follow the reasoning process specified by the explanations to some extent, as opposed to naively following the template patterns while disregarding critical information (Min et al., 2022c).
+
+The observations from our probing experiments lead us to focus next on understanding what makes
+
+an effective set of exemplars with explanations for solving a given test query. We primarily focus on two aspects, the exemplar-exemplar interplay (how exemplars work together) and the query-exemplar interplay. On the former, we find that the complementarity of the exemplar set is beneficial, as LLMs can fuse different reasoning processes exhibited by individual exemplars in-context. For these studies, we probe LLMs with a mixture of two types of exemplars; each type only specifies a part of the reasoning process (see Figure 3 for detailed instances). We also test how the relevance between the query and the exemplars impacts the performance. Choosing the nearest neighbors (NN) of the query to prompt LLMs has been shown to be effective in the standard prompting setting (Liu et al., 2021). Our experiments covering three similarity metrics show that this is also applicable in the setting that prompts LLMs with explanations.
+
+Our analyses inspire us to rethink the exemplar selection process of using explanations for ICL. The prominent NN-based paradigm only considers relevance (Shin et al., 2021; Liu et al., 2021; Rubin et al., 2022), which could result in selecting mostly similar exemplars. We argue that complementarity should also be considered when constructing explanation-infused prompts. Therefore, we propose an exemplar selection strategy based on the maximum marginal relevance (MMR) (Carbonell and Goldstein, 1998) approach which selects exemplars that are both relevant as well as diverse. The underlying rationale is that a diverse set of exemplars is more likely to showcase complementary reasoning types that are required to illustrate the reasoning required in the query. We test our MMR-based strategy on three real-world datasets spanning multiple reasoning tasks. On a powerful LLM, text-davinci-002, our MMR-based strategy is able to improve the accuracy over the baseline of using random exemplars by  $4.0\%$ ,  $3.9\%$ , and  $8.5\%$  on GSM, ECQA, and E-SNLI, respectively.
+
+In summary, our main findings are: (1) We show that both the computation trace and natural language contribute to making effective explanations for ICL. (2) We show that LLMs can benefit from exemplar sets that exhibit both complementarity and relevance to a given test query. (3) We propose an MMR-based exemplar selection strategy considering both complementarity and relevance and demonstrate that it is more effective than solely choosing the nearest neighbors.
+
+# 2 Background
+
+In-Context Learning Our study is focused on the usage of explanations in in-context learning (ICL). Let  $q$  be the test query to solve. The standard ICL prompts a language model,  $M$ , with a set of exemplar input-output pairs,  $\{(q_1, a_1) \dots (q_m, a_m)\}$ , and predict an answer  $\hat{a}$  for the query:
+
+$$
+\hat {a} = \underset {a} {\arg \max } p _ {M} (a \mid q, \{(q _ {1}, a _ {1}).. (q _ {m}, a _ {m}) \}).
+$$
+
+In addition to just input-output pairs, we can also include explanations (in the style of Scratchpad (Nye et al., 2021) or chain-of-thought (Wei et al., 2022)) in prompts, which leads the LLM to generate explanations for its predictions as well:
+
+$$
+\hat {a} = \underset {a} {\arg \max} \sum_ {e} p _ {M} (a, e \mid q, C),
+$$
+
+where  $C = \{(q_1, e_1, a_1) \dots (q_m, e_m, a_m)\}$  is the set of input-explanation-output triplets in prompts. Ideally, inference in this requires marginalizing out the explanation  $e$ , which is impractical, especially with LLMs. Following Wei et al. (2022); Ye and Durrett (2022), we employ greedy decoding to make an approximate prediction during inference.
+
+The end task performance of ICL is sensitive to the selected exemplars (Liu et al., 2021). While much prior work uses a fixed set of manually selected exemplars (Wei et al., 2022; Wang et al., 2022b), there is also work devoted to studying how to select more effective exemplars from a pool of exemplars. Given a test query  $q$ , the task is to select a set of  $m$  exemplars from a pool of  $n$  exemplars  $D = \{(q_{1}, e_{1}, a_{1}) \dots (q_{n}, e_{n}, a_{n})\}$  to construct a prompt for solving  $q$ . We note that this yields varying exemplar sets for different queries.
+
+Datasets & Large Language Models Our analysis is based on model performance on various reasoning datasets. For probing experiments, we mainly use symbolic reasoning datasets, including 1) LETTER CONCATENATION (LET CAT) (Wei et al., 2022) which requires extracting the last letters of two words and then concatenating them, 2) COIN FLIPS (Wei et al., 2022) which reasons about the states of a coin after two steps of operations (flipping or not flipping), and 3) GRADE SCHOOL MATH (GSM) (Cobbe et al., 2021) which focuses on grade school-level arithmetic reasoning problems expressed in natural language.
+
+To investigate the effectiveness of different exemplar selection strategies, we use two more textual reasoning datasets, namely ECQA (Aggarwal et al., 2021) and E-SNLI (Camburu et al., 2018), in addition to GSM. The task of E-SNLI is to decide whether a premise entails a hypothesis. ECQA asks multiple-choice commonsense questions. These three tasks include human-annotated explanations and cover diverse reasoning abilities.
+
+Our experiments cover an array of LLMs, including OPT-175B (Zhang et al., 2022), GPT-3 (davinci) (Brown et al., 2020), InstructGPT (text-davinci-001) (Ouyang et al., 2022), and text-davinci-002. In addition, we also use GPT-3 Codex models (Chen et al., 2021) that are finetuned on a large scale of code snippets in our exemplar selection experiments, namely code-davinci-001 and code-davinci-002. Though codex models primarily target code-related applications, we find that they are strong in textual reasoning tasks as well.
+
+# 3 Do LLMs Follow Explanations?
+
+We first investigate what makes explanations effective for LLMs to learn from. We view an explanation as a computation trace  $(T)$  that is transformed by a natural language function  $(L)$  which maps the trace to a complete natural language explanation. A computation trace  $T$  is the chain of intermediate steps (instantiated as tokens in explanations),  $s_1, \ldots, s_n$ , that are used to derive the final answer. For instance, the trace for LETCAT is the two last letters of the two words and the concatenated two-letter tokens; the traces for GSM are the intermediate equations. These computation traces are wrapped by natural language function  $L$  to form the final explanation  $L(s_1, \ldots, s_n)$ , which presumably makes the generation of these traces more "natural" with respect to language modeling.
+
+Setup We choose the three symbolic reasoning datasets mentioned in Section 2 for our probing experiments for two reasons. First, LLMs see substantial benefits from including explanations in prompts for these tasks. Second, we can easily manipulate the traces in their explanations. The gold explanations are directly taken from or adapted from Wei et al. (2022). More details on the gold explanations used for these tasks can be found in Appendix A.
+
+We experiment on both LLMs trained with vanilla language modeling objectives (OPT, davinci) and LLMs that are aligned with human
+
+![](images/bd1b4831cd9358a21fe457f0b637fcacb507c79852f16a16f466e3f703a94a81.jpg)  
+Figure 2: Examples of gold explanations and perturbed explanations. We perturb the trace in gold explanations (colored) by masking intermediate states or substituting them with incorrect values.
+
+expectations via different forms of instruction tuning (text-davinci-001, text-davinci-002).
+
+# 3.1 Explanations or Triggers?
+
+We start by investigating whether actual computation traces matter. If the correctness of the incontext demonstration is unimportant, then that serves as evidence that explanations act as triggers that induce LLMs to follow certain patterns and perform slot-filling. To study this, we prompt LLMs with perturbed explanations, i.e., by perturbing computational traces and measuring the impact on performance in an ICL setting.
+
+Figure 2 shows concrete examples of how we perturb the gold explanations. We experiment with two ways of perturbing the inputs. The first way is masking out the intermediate states by replacing a state  $s_i$  (or several states) in  $s_1, \ldots, s_n$  with a mask token (e.g., empty string), which tests whether
+
+<table><tr><td rowspan="2"></td><td colspan="4">LETCAT</td><td colspan="4">COINFLIP</td><td colspan="4">GSM</td></tr><tr><td>OPT</td><td>davinci</td><td>txt-01</td><td>txt-02</td><td>OPT</td><td>davinci</td><td>txt-01</td><td>txt-02</td><td>OPT</td><td>davinci</td><td>txt-01</td><td>txt-02</td></tr><tr><td rowspan="2">Standard Gold</td><td>8.5</td><td>8.5</td><td>10.5</td><td>16.0</td><td>51.5</td><td>83.0</td><td>68.0</td><td>99.0</td><td>5.5</td><td>7.5</td><td>11.0</td><td>26.5</td></tr><tr><td>50.0</td><td>59.0</td><td>85.0</td><td>100.</td><td>94.0</td><td>89.5</td><td>100.</td><td>100.</td><td>32.5</td><td>26.0</td><td>25.0</td><td>57.5</td></tr><tr><td>Mask1</td><td>11.0</td><td>16.0</td><td>21.5</td><td>100.</td><td>71.0</td><td>88.0</td><td>61.5</td><td>100.</td><td>19.0</td><td>21.0</td><td>12.5</td><td>29.5</td></tr><tr><td>Mask2</td><td>32.5</td><td>49.5</td><td>68.0</td><td>100.</td><td>84.0</td><td>91.5</td><td>99.0</td><td>100.</td><td>10.0</td><td>16.0</td><td>11.5</td><td>27.5</td></tr><tr><td rowspan="2">Random Incorrect</td><td>10.0</td><td>25.0</td><td>28.0</td><td>13.0</td><td>52.5</td><td>54.5</td><td>67.0</td><td>69.0</td><td>3.0</td><td>3.0</td><td>1.0</td><td>34.5</td></tr><tr><td>40.0</td><td>53.0</td><td>67.5</td><td>99.5</td><td>60.5</td><td>86.0</td><td>52.0</td><td>100.</td><td>18.5</td><td>17.0</td><td>10.0</td><td>16.5</td></tr><tr><td>No NL</td><td>29.0</td><td>15.0</td><td>46.5</td><td>100.</td><td>59.5</td><td>86.0</td><td>99.0</td><td>100.</td><td>8.0</td><td>19.5</td><td>14.5</td><td>45.5</td></tr></table>
+
+Table 1: In-context performance obtained using various perturbed explanations on three datasets. Perturbed explanations achieve inferior performance than complete ones, but many of the perturbed explanations still grant performance gains over standard prompting.
+
+LLMs can implicitly infer the intermediate states. The second way is to replace a state  $s_i$  with an incorrect one, which tests whether LLMs can recover the correct computation from corrupted traces.
+
+Construction of Perturbed Explanations We carefully design the way to mask out the intermediate states. We experiment with various choices of masks in our preliminary experiments. We do not observe large variance caused by different masks, and we choose to use empty string which leads to less performance degradation in general, as our goal here is to probe LLMs' maximum capabilities in recovering the reasoning process from perturbed explanations. More details on the preliminary experiments on choosing masks can be found in Appendix B. When constructing incorrect explanations, we also experiment with different sets of random values used to substitute the correct ones. Furthermore, we also include complete random explanations (taken from other exemplars), which replace the whole gold traces with incorrect ones.
+
+Results Table 1 shows the results obtained using prompts with various perturbations on the three tasks. First, LMs are indeed relying on the actual computation traces. Perturbing the traces of the explanations will lead to performance degradation in various degrees on all these three tasks for OPT, davinci, and text-davinci-001. text-davinci-002 does not exhibit performance degradation when being prompted with incomplete explanations on the simple tasks, LETCAT and COINFLIP: when the trace is straightforward, a powerful enough LLM is able to "shortcut" some particular steps. While for the more challenging tasks, i.e., GSM, text-002 is also affected by perturbations in explanations. Nonetheless, LLMs can still benefit from partially complete or partially correct explanations and outperform standard prompting without using explanations. In particular, on
+
+the LETCAT task, even completely irrelevant random explanations can be beneficial, although they lag gold explanations. Overall, incorrect and incomplete or even totally irrelevant explanations are able to elicit reasoning, but LLMs do rely on gold explanations to work well.
+
+# 3.2 Is Natural Language Necessary?
+
+Next, we question whether the natural language (NL) is really necessary and test whether LLMs can infer the reasoning steps from the computation traces alone. We perturb gold explanations by not wrapping computation traces with natural language transformation  $L$ , as shown in the examples from Figure 2, and only retain the traces.
+
+Results We show the performance obtained by using these prompts in Table 1. Natural language also plays an essential role in making effective explanations. Removing the NL leads to substantially worse performance. On LETCAT, the accuracy of OPT, and davinci drops by more than 20, respectively, compared to using gold explanations. On GSM, removing NL consistently leads to performance degradation. Meanwhile, including intermediate states without NL can still improve the performance compared to not using any explanations.
+
+# 3.3 Discussion
+
+As suggested in the experimental results in Section 3.1 and Section 3.2, LLMs do generally follow the explanations in the prompts. Both concrete computation traces and natural language contribute to making effective explanations for ICL. Perturbing certain parts of the explanations will accordingly result in performance degradation, but partial explanations are still beneficial to LMs.
+
+By contrast, recent work shows LLMs are not sensitive to perturbations on the ground-truth input-label mapping in the standard prompting paradigm that does not use explanations (Min et al., 2022c).
+
+Our work shows that LLMs are sensitive to perturbations in the input-explanation mapping and other more subtle perturbations in the explanations. Using explanations in prompts is a promising way to guide LLMs in learning a new task via ICL.
+
+# 4 What Makes A Good Exemplar Set?
+
+Our probing experiments have established how the general factors, computation trace and natural language, impact explanations' effectiveness in ICL. We now study how a set of exemplars, as a whole, functions together in solving a particular test query. We study this problem from two angles, the interplay between exemplars and the interplay between the query and the exemplars.
+
+# 4.1 Exemplar-Exemplar Interplay
+
+As in Section 3, LLMs can learn to follow the reasoning processes as specified in exemplars. As reasoning processes can be composed, we hypothesize that LLMs might also be able to fuse the reasoning processes of different exemplars together to solve a test query. We design a set of probing experiments that successfully verify this hypothesis.
+
+Experiment Design At the abstract level, we compare the performance of LLMs when being prompted with three sets of exemplars. The first and second set of exemplars each focuses on a particular part of the reasoning process, and these two parts are disjoint. That is, for a computation trace  $s_1, \ldots, s_n$ , the first and second set contain exemplars where  $s_i$  and  $s_j$  are perturbed, and  $i \neq j$ . The third set of exemplars includes the mixture from the first and second sets. We test the ICL performance of the prompts constructed from these three types of exemplar sets on the test set that requires combining two types of reasoning. If the third type gives superior performance than the first two types, that means LLMs can pick up the disjoint reasoning and fuse them in-context.
+
+To better illustrate such a hypothesis, we give a concrete example as follows. We have introduced two different types of masked explanations in Figure 2 for LETCAT, where the first type masks the last letter extraction part, and the second type masks the letter concatenation part. These two different masked explanations specify two steps of the reasoning process needed; combining these two steps will yield the complete reasoning steps needed for solving this task. We test whether LLMs can combine these two reasoning steps
+
+<table><tr><td>Add Only</td><td>Q: Marion received 20 more turtles than Mia at the animal rescue center. 
+If Mia received 40 turtles, how many turtles did they receive together? 
+A: Since Marion received 20 more turtles than Mia, she had 20 + 40 = 60 turtles. The two received 60 + 40 = 100 turtles. The answer is 100.</td></tr><tr><td>Mul Only</td><td>Q: Super Clean Car Wash Company cleans 80 cars per day. They make $5 per car washed. How much money will they make in 5 days? 
+A: Each day they will make 80 * $5 = $400. They will make $400 * 5 = $2000 in 5 days. The answer is 2000.</td></tr><tr><td>Add &amp; Mul</td><td>Q: Peter purchased 20 popsicles at $0.25 each. He also purchased 4 ice cream bars at $0.50 each. How much did he pay in total in dollars? 
+A: The popsicles cost 0.25 * 20 = 5 dollars. The ice cream bars cost 0.5 * 4 = 2 dollars. He paid 5 + 2 = 7 dollars. The answer is 7.</td></tr></table>
+
+Figure 3: Examples of GSM data points that involve only addition operators, only multiplication operators, and both of them at the same time. We note that Add & Mul are only used at the test time.
+
+in-context if being prompted with a mixture of these two corresponding types of masked prompts.
+
+For GSM, we use a more organic way to partition the reasoning process. We separate the reasoning skills needed for a test query based on the operators (addition and multiplication) that are used in the steps. Concretely, we filter the GSM dataset by looking at the provided explanations paired with examples, and obtain disjoint sets that 1) only involves addition operators in the explanation 2) only involves multiplication operators in the explanation (See Figure 3 for examples). Next, we test the performance on a test set consisting of examples that require both operators at the same time (Add and Mul in Figure 3). This forms a test-bed for investigating whether LLMs can better learn to solve problems where both operators are present at the same time while being prompted with the mixture of these two operators, even if no explicit combinations are shown in the prompts.
+
+Setup We experiment on the same three datasets as used in Section 3. On LETCAT and COINFLIP, we test whether LLMs can combine the reasoning steps specified in two different types of masked explanations; on GSM, we test whether LLMs can compose addition and multiplication. The mixture type prompts include half of the exemplars from the first type and second type which bear different reasoning. For each setting of GSM, we experiment with 4 different sets of randomly drawn exemplars and report the average. More details about the setting can be found in Appendix C.
+
+Results As in Table 2, on LETCAT, the prompts with mixed explanations largely surpass Mask1, and outperform Mask2 by 6.0, 7.0, and 12.0 on OPT, davinci, text-davinci-001, respectively. Par
+
+<table><tr><td></td><td></td><td>OPT</td><td>davinci</td><td>txt-01</td><td>txt-02</td></tr><tr><td rowspan="3">LETCAT</td><td>Mask1</td><td>11.0</td><td>16.0</td><td>21.5</td><td>100.</td></tr><tr><td>Mask2</td><td>32.5</td><td>49.5</td><td>68.0</td><td>100.</td></tr><tr><td>Mixture</td><td>37.0</td><td>56.5</td><td>82.0</td><td>100.</td></tr><tr><td rowspan="3">COIN</td><td>Mask1</td><td>71.0</td><td>88.0</td><td>61.5</td><td>100.</td></tr><tr><td>Mask2</td><td>84.0</td><td>91.5</td><td>99.0</td><td>100.</td></tr><tr><td>Mixture</td><td>93.5</td><td>91.0</td><td>100.</td><td>100.</td></tr><tr><td rowspan="3">GSM</td><td>AddOnly</td><td>6.8</td><td>13.5</td><td>14.1</td><td>50.3</td></tr><tr><td>MulOnly</td><td>4.7</td><td>17.2</td><td>16.7</td><td>50.1</td></tr><tr><td>Mixture</td><td>7.0</td><td>18.9</td><td>18.2</td><td>52.0</td></tr></table>
+
+Table 2: The accuracy of prompting LLMs with exemplars focusing on single parts of the reasoning or a mixture of them. LLMs achieve better performance when being prompted with exemplars covering multiple aspects of the reasoning process.
+
+ticularly, the mixture prompts is able to perform roughly on par with the complete prompt (Gold-Expl in Table 1) for LETCAT on davinci, and text-001. On COINFLIP, using mixture prompts also leads to improvements on OPT.
+
+On the realistic GSM dataset, prompting text-davinci-002 with only addition or multiplication exemplars leads to a performance of 50.3 and 50.1, respectively, whereas prompting with a mixture of these two types of exemplars achieves a better performance of 52.0. On davinci, and text-001, addition exemplars give worse performance than multiplication exemplars. Nevertheless, including these inferior addition exemplars in prompts together with multiplication exemplars still leads to better performance, as they can complement the reasoning. In general, results on three datasets suggest LLMs are able to fuse the reasoning process that is spread over different exemplars. Therefore, we can expect the exemplars to be able to complement each other and collaborate together to solve the reasoning needed in the test query.
+
+# 4.2 Query-Exemplar Interplay
+
+Next, we explore how the interplay between the query and exemplars impacts the ICL performance. Recent work has studied how to make good in-context exemplar sets for a given query in the standard prompting setting: choosing nearest neighbors that are more similar to the query leads to better performance (Liu et al., 2021; Shin et al., 2021). Our work investigates how choosing relevant exemplars impact the performance in the setting when using explanations in prompts. We compare the performance obtained by constructing prompts using nearest neighbors against using randomly selected
+
+exemplars. The results verify that choosing the nearest neighbors is also beneficial in this setting.
+
+Similarity Measurements We test three different ways to measure the similarity  $S(q, q_i)$  between a test query  $q$  and an exemplar  $q_i$ . 1) CLS-based: Liu et al. (2021) use smaller LMs (e.g., BERT (Devlin et al., 2019)) to extract the CLS embedding of the input  $q$  and  $q_i$  and then use cosine similarity to score the embedding pairs, i.e.,  $\cos(\mathrm{CLS}(q), \mathrm{CLS}(q_i))$ . 2) LM-based: the similarity is given as the probability of generating the query when the language model is conditioned on the exemplar, i.e.,  $LM(q \mid q_i)$  (Shin et al., 2021; Rubin et al., 2022). 3) BERTScore: we also experiment with using BERTScore (Zhang et al., 2020) as the similarity score, in addition to the two approaches that are commonly used in prior work.
+
+It is worthwhile to note that measuring similarity using large LLMs is expensive. As it requires querying a large number of query-exemplar pairs. $^2$
+
+Setup We experiment on three realistic datasets, GSM, ECQA, and E-SNLI. leaving out synthetic tasks which feature formulaic explanations that are all similar to each other. We set the number of exemplars to be 8 for all three test datasets. We compare the performance of selecting nearest exemplars against that of selecting random exemplars.
+
+Given the intensive cost of querying LLMs, we set the train exemplars pool size to be 512, and allocate computational resources to experimenting over 4 sets of randomly selected 512 exemplar pools to alleviate the influence of randomness. We focus on more capable LLMs, including code-davinci-001, and code-davinci-002, and text-davinci-002, leaving out OPT and davinci which have inferior performance. We note that we do not use text-002 to measure similarity, owing to its high cost. Rather, we take the similarity scores computed by code-002 and use those for text-002. So the performance of text-002 when using the LM-based measure might be suboptimal given the discrepancy.
+
+Results As shown in Table 3, choosing relevant exemplars is also useful in the setting that
+
+<table><tr><td></td><td colspan="4">code-davinci-001</td><td colspan="4">code-davinci-002</td><td colspan="4">text-davinci-002</td></tr><tr><td></td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td></tr><tr><td>Random</td><td>16.3</td><td>53.6</td><td>47.2</td><td>39.0</td><td>64.6</td><td>74.7</td><td>74.9</td><td>71.3</td><td>48.8</td><td>71.9</td><td>75.1</td><td>65.3</td></tr><tr><td>CLS</td><td>16.5</td><td>55.0</td><td>54.1</td><td>41.8</td><td>65.4</td><td>74.9</td><td>74.8</td><td>71.7</td><td>50.4</td><td>72.1</td><td>77.4</td><td>66.6</td></tr><tr><td>LLM</td><td>18.5</td><td>56.0</td><td>57.4</td><td>43.9</td><td>65.8</td><td>76.8</td><td>81.6</td><td>74.7</td><td>52.0*</td><td>74.3*</td><td>83.9*</td><td>70.0</td></tr><tr><td>BERTScore</td><td>18.5</td><td>54.6</td><td>53.7</td><td>42.3</td><td>66.7</td><td>75.9</td><td>75.6</td><td>72.8</td><td>51.0</td><td>72.8</td><td>78.7</td><td>67.6</td></tr></table>
+
+Table 3: Comparison between the performance obtained by choosing relevant exemplars using CLS embedding, LM, or BERTScore. Avg denotes the average across the three datasets. Selecting relevant exemplars leads to performance improvements, especially when using LLMs themselves to measure the similarity. Using Bertscore also consistently improves the performance across all tasks, even surpassing LM-based scores on GSM. We note that the results on text-davinci-002 use the LM-based scores provided by code-davinci-002 (denoted by \*).
+
+includes explanations in prompts. Using the LM-based similarity measurements brings performance improvements across all three datasets, and has the most significant impacts on E-SNLI, though this is achieved with non-negligible computation cost. Using CLS-embeddings for selecting exemplars mildly improves the performance on GSM but does not result in any performance gains on ECQA. The limited improvements can be attributed to the size of the exemplar pools that we use. In our experiments, the size is 512, which is significantly smaller than that in Liu et al. (2021) (typically tens of thousands of exemplars). Nevertheless, this size is large enough for the LM-based method and BERTScore to take advantage of.
+
+In addition, the results suggest that choosing relevant exemplars using BERTScore is also able to improve the performance across all datasets. Specifically, BERTScore-based exemplar selection achieves an accuracy of 66.7 on GSM using code-002, which even surpasses the performance of LM-based exemplar selection. While using BERTScore lags the LM-based on ECQA and E-SNLI, it still outperforms choosing random exemplars or CLS-based exemplar selection. Overall, using BERTScore to select the closest exemplars can lead to credible performance improvements while does not require heavy overheads caused by using LLMs to score query-exemplar pairs.
+
+# 5 MMR for Exemplar Selection
+
+We have established that emplar-exemplar interplay together with the query-exemplar interplay impacts the performance of using explanations in ICL. This leads us to rethink how to select good exemplars for a given query. Based on our prior analysis on the effects of complementarity and relevance in Section 4.1, we argue that a good set should consist of relevant exemplars that collaboratively cover the reasoning skills required for solving the query.
+
+# Algorithm 1 MMR-Based Exemplar Selection
+
+1: procedure MMRSELECT  $(D,q,k,\mathcal{S})$  input: exemplar pool  $D = \{q_{1}\dots q_{n}\}$  , test query  $q$  , number of shots  $m$  and similarity measurement  $\mathcal{S}$  output: selected exemplars  $T = \{q_{1}\dots q_{m}\}$  
+2:  $\mathbb{S} := [[S(q_i, q_j)]]_{q_i, q_j \in D}$ ;  $\triangleright$  the pairwise similarity between exemplars in  $D$  
+3:  $\mathbb{Q} := [S(q, q_i)]_{q_i \in D}$ ;  $\triangleright$  the similarity between query and exemplars in  $T$  
+4:  $T\coloneqq \{\}$  
+5: while  $|T| < k$  do  
+6:  $\hat{q} := \operatorname{Equation}(1)$ ;  $\triangleright$  get the next exemplar based on Eq (1)  
+7:  $T.\mathrm{add}(\hat{q})$  
+8: return  $T$ ;
+
+The prominent paradigm, i.e., NN-based exemplar selection strategy, only considers the relevance between the exemplars and the query. Yet, selecting nearest neighbors could result in mostly similar exemplar sets, which can possibly limit collaboration. We argue that complementarity should also be considered in the exemplar selection process, so that the selected set could have a higher chance to illustrate the required reasoning processes.
+
+In practice, it is tricky to decide whether the reasoning underlying a set of exemplars is complementary categorically. We therefore use diversity as a proxy, since a set of less similar exemplars is arguably more likely to exhibit complementarity. To that end, we propose a maximal-marginal-relevance (Carbonell and Goldstein, 1998) (MMR) based exemplar selection strategy. The idea is to select exemplars that are relevant to the query while being diverse enough to be collaborative. Suppose for the given query  $q$ , we have already selected a set of exemplars  $T = \{q_i\}$ , then we will pick up the next exemplar according to:
+
+$$
+\underset {q _ {j} \in D / T} {\arg \max } \lambda \mathcal {S} (q, q _ {j}) - (1 - \lambda) \max  _ {q _ {i} \in T} \mathcal {S} (q _ {j}, q _ {i}) \tag {1}
+$$
+
+where  $S$  denotes similarity and  $\lambda$  is a parameter
+
+<table><tr><td></td><td colspan="4">code-davinci-001</td><td colspan="4">code-davinci-002</td><td colspan="4">text-davinci-002</td></tr><tr><td></td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td><td>GSM</td><td>ECQA</td><td>E-SNLI</td><td>Avg</td></tr><tr><td>LLM NN</td><td>18.5</td><td>56.0</td><td>57.4</td><td>43.9</td><td>65.8</td><td>76.8</td><td>81.6</td><td>74.7</td><td>52.0*</td><td>74.3*</td><td>83.9*</td><td>70.0*</td></tr><tr><td>LLM MMR</td><td>18.7</td><td>57.2</td><td>59.5</td><td>45.1</td><td>67.0</td><td>77.4</td><td>81.5</td><td>75.3</td><td>52.8*</td><td>75.3*</td><td>83.7*</td><td>70.6*</td></tr><tr><td>BERTScore NN</td><td>18.5</td><td>54.6</td><td>53.7</td><td>42.3</td><td>66.7</td><td>75.9</td><td>75.6</td><td>72.8</td><td>51.0</td><td>72.8</td><td>78.7</td><td>67.6</td></tr><tr><td>BERTScore MMR</td><td>19.4</td><td>56.3</td><td>53.9</td><td>43.2</td><td>68.2</td><td>78.1</td><td>77.8</td><td>74.7</td><td>52.0</td><td>73.7</td><td>78.2</td><td>68.0</td></tr></table>
+
+that controls the balance between relevance and diversity. We rely on MMR to iteratively select exemplars from the exemplar pool, as shown in Algorithm (1). Note that this requires scoring all exemplar pairs within the pool. To run inference over  $m$  queries using a pool of  $n$  exemplars, MMR requires to score the similarity of  $nn + mn$  pairs.
+
+Results We apply the MMR strategy on top of LM-based method and BERTScore, leaving out the CLS-based approach which has inferior performance. The experimental setup largely follows Section 4.2, please refer to Appendix D for details.
+
+We show the results in Table 3. In the setting that uses BERTScore, MMR-based selection successfully improves the performance for almost all LLMs for all datasets, compared to using nearest neighbors. On LM-based method, MMR is also able to improve the performance for GSM and ECQA across all LMs, and only marginally under-performs NN for E-SNLI.
+
+In particular, using the MMR-based selection strategy achieves an accuracy of 68.2 and 78.1 on GSM and ECQA respectively, even outperforming LM-based method that requires a large number of queries to the LM. This suggests that BERTScore and MMR as a combination are able to construct effective explanation-infused prompts that approach that of actually querying LLMs. Furthermore, the fact that LLMs achieve better performance from the exemplars selected using our MMR-based method is congruent with our analysis in the previous section: LLMs can exploit complementary explanations.
+
+# 5.1 Analysis
+
+Impacts of the Trade-off Between Relevance and Diversity We conduct an analysis to investigate how the trade-off between diversity and relevance impacts the performance. We test the performance under varying  $\lambda$  on code-davinci-002 with BERTScore as the similarity metric. We note
+
+Table 4: Results of using MMR-based exemplar selection strategy on three datasets (AVG denotes the average). Using MMR generally selects better exemplars on all datasets, using either LM-based method or BERTScore. The results on text-davinci-002 use the LM-based scores provided by code-davinci-002 (denoted by \*).  
+
+<table><tr><td>λ</td><td>GSM</td><td>ECQA</td><td>E-SNLI</td></tr><tr><td>1.0</td><td>66.7</td><td>75.9</td><td>75.6</td></tr><tr><td>0.8</td><td>66.9</td><td>75.6</td><td>76.6</td></tr><tr><td>0.6</td><td>68.2</td><td>77.9</td><td>78.1</td></tr><tr><td>0.5</td><td>68.2</td><td>78.1</td><td>77.8</td></tr><tr><td>0.4</td><td>66.8</td><td>75.7</td><td>76.0</td></tr><tr><td>0.2</td><td>65.9</td><td>75.9</td><td>74.9</td></tr><tr><td>0.0</td><td>63.5</td><td>75.5</td><td>75.5</td></tr></table>
+
+Table 5: The performance of MMR exemplar selection strategy with varying  $\lambda$  .  
+
+<table><tr><td></td><td>GSM</td><td>ECQA</td><td>E-SNLI</td></tr><tr><td>Random</td><td>65.41.3</td><td>74.10.5</td><td>74.01.2</td></tr><tr><td>NN</td><td>68.60.7</td><td>75.40.5</td><td>75.91.1</td></tr><tr><td>MMR</td><td>69.41.0</td><td>77.80.7</td><td>77.80.9</td></tr></table>
+
+Table 6: Meanvariance of the performance across 5 random order. Using better exemplars has more significant impact than varying exemplar order.
+
+this is done on one pool of training exemplars. Generally, when  $\lambda$  is large (0.8), the performance is similar to NN ( $\lambda = 1.0$ ). MMR typically works well with a  $\lambda$  of 0.6 or 0.5 (roughly balancing the two terms). The performance starts to degrade while decreasing  $\lambda$  from 0.4 to 0, as the selected exemplars are not relevant enough.
+
+Sensitivity to Different Order We have shown choosing exemplar sets using MMR can lead to better ICL performance, which could be affected by other confounders such as the order of exemplars. We conduct experiments to show that using better exemplar sets has more impact than reordering exemplars. Specifically, we experiment with 5 random orders of the exemplar sets for each query and report averaged performance and variance of the accuracy. We note this is done on one pool of training exemplars for each dataset using code-davinci-002 with BERTScore as the metric. As shown in Table 6, MMR is still substantially better than NN and Random under varying order.
+
+# 6 Related Work
+
+The growing scale of pretrained language models has granted them the ability to learn a new task from a few examples via in-context learning (Brown et al., 2020). Various approaches have been proposed to improve ICL in recent years, including meta-tuning LLMs (Min et al., 2022b; Chen et al., 2022), calibration of ICL (Zhao et al., 2021; Han et al., 2022), automatically determining the orders of exemplars (Lu et al., 2022), and alternative formulation of ICL based on PMI (Holtzman et al., 2021) or noisy-channel (Min et al., 2022a). More closely related to our work, prior research also contributes to better understanding ICL as Bayesian inference (Xie et al., 2022) or experiments that study what makes in-context learning works (Min et al., 2022c). Our work focuses on understanding the usage of explanations in ICL, as opposed to standard prompting where the LLMs are presented with only input-output pairs.
+
+In particular, our work is connected to prior research on effective ways for selecting in-context exemplars (Shin et al., 2021; Liu et al., 2021; Rubin et al., 2022; Qiu et al., 2022; Su et al., 2022). While past work primarily focuses on the effectiveness of using relevant examples in the standard prompting paradigm, we examine the benefits of the complementary exemplars when prompting with explanations. We also propose an MMR-based strategy, which is more effective than the NN-based approach on various LLMs across three tasks.
+
+Lastly, including textual explanations in prompts has exhibited remarkable benefits for LLMs to learn various reasoning tasks. Using Scratchpad (Nye et al., 2021) or Chain-of-Thought (Wei et al., 2022) significantly boosts ICL performance on multi-step reasoning tasks such as arithmetical reasoning and symbolic reasoning. Using free-text rationales is also helpful for more unstructured tasks like QA and NLI (Ye and Durrett, 2022; Wang et al., 2022a). While recent work largely aims to find better ways to prompt LLMs with explanations (Kojima et al., 2022; Zhou et al., 2022a; Press et al., 2022; Zhou et al., 2022b), we focus on analyzing the role of explanations in ICL and what makes effective explanations.
+
+# 7 Conclusion
+
+We have presented a series of studies on what makes effective explanations for in-context learning. We first investigated the impacts of compu
+
+tation traces and natural language in explanations. Through a set of probing experiments, we found that LLMs rely on both of them to effectively learn from explanations. We further examined the interplay among exemplars within prompts and the interplay between exemplars and the query. Our analysis uncovered the benefits of constructing prompts by selecting complementary explanations that are relevant to the query. Lastly, we proposed an MMR-based exemplar selection strategy, which successfully improved the end task performance across three important datasets.
+
+# 8 Limitations
+
+The models chosen in this work are selected to represent the state-of-the-art at the time the work was conducted, and in some cases omit weaker models. For example, our exemplar selection experiments do not cover those LLMs trained with vanilla language models objectives, namely OPT and davinci, as we find their performance substantially lags code-davinci-002 and text-davinci-002. For the same reason, we only consider the substantially large language models, omitting LLMs of smaller scales (e.g., text-curie-001). Running experiments using smaller LMs or vanilla LMs may provide insights into how scale or instruction finetuning impacts the ability of LMs in learning from explanations, but our investigation mainly focus on selecting exemplars to achieve the best in-context learning performance with state-of-the-art models.
+
+In addition, certain aspects of our approach are computationally intensive, particularly using LM-based similarity scores. However, we think this is still feasible in practice: if practitioners are deploying a real-world system, investing more computation upfront to improve its performance is likely in reach for those deploying LLMs in practice.
+
+Finally, our experiments consider a certain subset of NLP reasoning tasks written in English. While we believe the results here transfer to other tasks in this vein which have been frequently used to evaluate LLMs, it is unknown how well they handle other languages, dialects, or genres of text such as social media data.
+
+# Acknowledgments
+
+Thanks to anonymous reviewers for their helpful feedback and colleagues at Meta AI for helpful discussions. This work was partially supported by NSF CAREER Award IIS-2145280 and the NSF
+
+AI Institute for Foundations of Machine Learning (IFML).
+
+# References
+
+Shourya Aggarwal, Divyanshu Mandowara, Vishwajeet Agrawal, Dinesh Khandelwal, Parag Singla, and Dinesh Garg. 2021. *Explanations for Common-senseQA: New Dataset and Models*. In *Proceedings of the Annual Conference of the Association for Computational Linguistics (ACL)*.  
+Stephen H. Bach, Victor Sanh, Zheng-Xin Yong, Albert Webson, Colin Raffel, Nihal V. Nayak, Abheesht Sharma, Taewoon Kim, M Saiful Bari, Thibault Fevry, Zaid Alyafeai, Manan Dey, Andrea Santilli, Zhiqing Sun, Srulik Ben-David, Canwen Xu, Gunjan Chhablani, Han Wang, Jason Alan Fries, Maged S. Al-shaibani, Shanya Sharma, Urmish Thakker, Khalid Almubarak, Xiangru Tang, Xiangru Tang, Mike Tian-Jian Jiang, and Alexander M. Rush. 2022. PromptSource: An Integrated Development Environment and Repository for Natural Language Prompts. arXiv, abs/2202.01279.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. In Proceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS).  
+Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom. 2018. e-snli: Natural language inference with natural language explanations. In Proceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS).  
+Jaime Carbonell and Jade Goldstein. 1998. The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries. In Proceedings of the Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR). Association for Computing Machinery.  
+Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde, Jared Kaplan, Harrison Edwards, Yura Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidi Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian, Clemens Winter, Philippe Tillet, Felipe Petroski Such, David W. Cummings, Matthias Plappert, Fotios Chantzis, Elizabeth Barnes, Ariel
+
+Herbert-Voss, William H. Guss, Alex Nichol, Igor Babuschkin, S. Arun Balaji, Shantanu Jain, Andrew Carr, Jan Leike, Joshua Achiam, Vedant Misra, Evan Morikawa, Alec Radford, Matthew M. Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. 2021. Evaluating large language models trained on code. *ArXiv*, abs/2107.03374.  
+Yanda Chen, Ruiqi Zhong, Sheng Zha, George Karypis, and He He. 2022. Meta-learning via language model in-context tuning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Baindoor Rao, Parker Barnes, Yi Tay, Noam M. Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Benton C. Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Oliveira Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathleen S. Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2022. Palm: Scaling language modeling with pathways. ArXiv, abs/2204.02311.  
+Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.  
+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 (NAACL-HLT).  
+Zhi Han, Yaru Hao, Li Dong, and Furu Wei. 2022. Prototypical calibration for few-shot learning of language models. ArXiv, abs/2205.10183.  
+Ari Holtzman, Peter West, Vered Shwartz, Yejin Choi, and Luke Zettlemoyer. 2021. Surface form competition: Why the highest probability answer isn't always right. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.
+
+Jaehun Jung, Lianhui Qin, Sean Welleck, Faeze Brahman, Chandra Bhagavatula, Ronan Le Bras, and Yejin Choi. 2022. Maieutic prompting: Logically consistent reasoning with recursive explanations. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. ArXiv, abs/2205.11916.  
+Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. What makes good in-context examples for gpt-3? ArXiv, abs/2101.06804.  
+Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Aman Madaan, Shuyan Zhou, Uri Alon, Yiming Yang, and Graham Neubig. 2022. Language models of code are few-shot commonsense learners. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Sewon Min, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022a. Noisy channel language model prompting for few-shot text classification. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).  
+Sewon Min, Mike Lewis, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2022b. MetaICL: Learning to learn in context. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.  
+Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. 2022c. Rethinking the role of demonstrations: What makes in-context learning work? In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Augustus Odena. 2021. Show your work: Scratchpads for intermediate computation with language models. ArXiv, abs/2112.00114.  
+Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and
+
+Ryan J. Lowe. 2022. Training language models to follow instructions with human feedback. *ArXiv*, abs/2203.02155.  
+Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A. Smith, and Mike Lewis. 2022. Measuring and narrowing the compositionality gap in language models. ArXiv, abs/2210.03350.  
+Linlu Qiu, Peter Shaw, Panupong Pasupat, Tianze Shi, Jonathan Herzig, Emily Pittler, Fei Sha, and Kristina Toutanova. 2022. Evaluating the impact of model scale for compositional generalization in semantic parsing. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. Learning to retrieve prompts for in-context learning. In Proceedings of the Annual Conference of the Association for Computational Linguistics (ACL).  
+Richard Shin, Christopher Lin, Sam Thomson, Charles Chen, Subhro Roy, Emmanouil Antonios Platanios, Adam Pauls, Dan Klein, Jason Eisner, and Benjamin Van Durme. 2021. Constrained language models yield few-shot semantic parsers. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).  
+Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, et al. 2022. Selective annotation makes language models better few-shot learners. arXiv preprint arXiv:2209.01975.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022a. Rationale-augmented ensembles in language models. ArXiv, abs/2207.00747.  
+Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, and Denny Zhou. 2022b. Self-consistency improves chain of thought reasoning in language models. ArXiv, abs/2203.11171.  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903.  
+Sang Michael Xie, Aditi Raghunathan, Percy Liang, and Tengyu Ma. 2022. An explanation of in-context learning as implicit bayesian inference. In International Conference on Learning Representations.  
+Xi Ye and Greg Durrett. 2022. The unreliability of explanations in few-shot prompting for textual reasoning. In Proceedings of the Conference on Advances in Neural Information Processing Systems (NeurIPS).  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu
+
+Wang, and Luke Zettlemoyer. 2022. OPT: Open Pre-trained Transformer Language Models. ArXiv, abs/2205.01068.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating Text Generation with BERT. In Proceedings of the International Conference on Learning Representations (ICLR).  
+Tony Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In Proceedings of the International Conference on Learning Representations (ICLR).  
+Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Olivier Bousquet, Quoc Le, and Ed Chi. 2022a. Least-to-most prompting enables complex reasoning in large language models. ArXiv, abs/2205.10625.  
+Hattie Zhou, Azade Nova, H. Larochelle, Aaron C. Courville, Behnam Neyshabur, and Hanie Sedghi. 2022b. Teaching algorithmic reasoning via in-context learning. ArXiv, abs/2211.09066.
+
+# A Details of the Explanations Used for LETCAT and COINFLIP on OPT
+
+<table><tr><td rowspan="2">LETCAT</td><td>Question: Take the last letters of the words in &quot;Bill Gates&quot; and concatenate them.</td></tr><tr><td>Ours: Add space to &quot;Bill&quot; and get &quot;B i l l&quot;, the last letter is l. Add space to &quot;Gates&quot; and get &quot;G a t e s&quot;, the last letter is s. Concatenating l and s is ls. So the answer is ls.
+Wei et al. (2022): The last letter of Bill is l. The last letter of Gates is s. Concatenating l and s is ls. So the answer is ls.</td></tr><tr><td rowspan="2">COINFLIP</td><td>Question: A coin is heads up. Shaunda does not flip the coin. Shalonda flips the coin. Is the coin still heads up?</td></tr><tr><td>Ours: The coin started heads up. Shaunda does not flip the coin, so it becomes heads up. Shalonda flips the coin, so it becomes tails up. So the answer is no.
+Wei et al. (2022): The coin was flipped by Shalonda. So the coin was flipped 1 time, which is an odd number. The coin started heads up, so after an odd number of flips, it will be tails up. So the answer is no.</td></tr></table>
+
+Figure 4: Examples of original chain-of-thoughts from Wei et al. (2022) and ours used for OPT.  
+
+<table><tr><td></td><td>LETCAT</td><td>COINFLIP</td></tr><tr><td>Standard</td><td>8.5</td><td>51.5</td></tr><tr><td>(Wei et al., 2022)</td><td>29.5</td><td>61.0</td></tr><tr><td>Ours</td><td>50.0</td><td>94.0</td></tr></table>
+
+For GSM, we directly use the gold explanations provided in Wei et al. (2022). For LETCAT and COINFLIP, we take the original explanations from Wei et al. (2022) and manually engineered them, as the original ones are sub-optimal for OPT and do not lead to credible gains compared to standard prompting.
+
+We show examples of the original explanations (chain-of-thoughts) used in Wei et al. (2022) and the explanation we adapted for OPT in Figure 4. For LETCAT, we add another step of tokenizing the two words. For COINFLIP, we change the way of decomposing the problem. As shown in Table 7, our adapted explanations lead to more substantial performance improvements over standard prompting. We use engineered explanations for the probing experiments on OPT, which allows more distinguishable performance differences. We refer readers to the supplementary materials of Wei et al. (2022) for the complete set of exemplars and explanations.
+
+# B Details of the Choice of Masks
+
+We conduct preliminary experiments on the LET-CAT dataset using davinci to determine the choice
+
+Table 7: Performance of original chain-of-thoughts and our explanations used for OPT.  
+
+<table><tr><td></td><td>N/A</td><td>[mask]</td><td>?</td><td>_</td><td>Empty Str</td></tr><tr><td>Standard</td><td></td><td></td><td>8.5</td><td></td><td></td></tr><tr><td>Gold</td><td></td><td></td><td>59.0</td><td></td><td></td></tr><tr><td>Mask1</td><td>14.0</td><td>14.0</td><td>15.0</td><td>13.5</td><td>16.0</td></tr><tr><td>Mask2</td><td>48.0</td><td>48.0</td><td>48.5</td><td>43.0</td><td>49.5</td></tr></table>
+
+Table 8: Results of using different mask tokens for LETCAT on OPT.
+
+of masks. We tested masking with "N/A", "[mask]", "?", "_", and empty string. The results obtained using different masks are shown in Table 8. Whatever masks are used, LLMs see performance degradation compared to gold explanations, but can still learn from partially complete explanations. We use an empty string as the mask token across all datasets, which leads to the least performance degradation.
+
+# C Details of the Setup for Exemplar-Exemplar Interplay Experiments
+
+For LETCAT, we experiment with 4 exemplars where the first steps are perturbed, 4 exemplars where the second steps are perturbed and a mixture of 2 from each of these explanations. For COINFLIP, we use 8 exemplars and follow the same setting. For the mixture type of prompts, we experiment with 4 random combinations for mixing two types of masked exemplars.
+
+For GSM, we use three types of prompts constructed by 1) 8 addition-only exemplars, 2) 8 multiplication-only exemplars, and 3) a mixture of 4 exemplars from each of the two types. We note that unlike what's in LETCAT and COINFLIP which uses identical exemplars perturbed in different ways, the exemplars in the three sets, for GSM are drawn from different pools. We experiment with 4 different sets of randomly drawn examples and report the average in the setting. We note the test set for GSM that requires composing addition and multiplication contains 1,150 data points in total.
+
+# D Details of the Setup for MMR-Based Exemplar Selection Experiments
+
+We evaluate the effectiveness of our MMR-based exemplar selection strategy on the three realistic datasets used in Section 4.2. Also, the experiments on text-davinci-002 using the LM-based method rely on similarity scores obtained
+
+from code-davinci-002, given the prohibitive cost needed to run these experiments. We do not tune  $\lambda$  in our experiments.  $\lambda$  is set to roughly balance the variance among the two terms in Equation (1), which is 0.5 across all datasets and methods, except for using the LM-based method on code-davinci-001. For this particular setting, we set lambda to be  $\frac{2}{3}$ , as we observe higher variance among the diversity with exemplars.
+
+# E Details of Prompts for Real-world Datasets
+
+We showcase how we format the prompts for GSM, ECQA, and E-SNLI in Figure 5, with one exemplar for each of the three datasets. We note the prompt format for E-SNLI is taken from PromptSource (Bach et al., 2022).
+
+# GSM
+
+Q: Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?
+
+A: Leah had 32 chocolates and Leah's sister had 42. That means there were originally  $32 + 42 = 74$  chocolates. 35 have been eaten. So in total they still have  $74 - 35 = 39$  chocolates. The answer is 39.
+
+# ECQA
+
+Q: Where can you get a bugle to take home with you? Answer Choices:
+
+(a) farmer's wife
+
+(b) music store
+
+(c) military base
+
+(d) military band
+
+(e) american army.
+
+A: Bugle is a musical instrument. Musical instruments are available in a music store. Music store is a building. So the answer is (b).
+
+# E-SNLI
+
+Premise:
+
+"A man at a flea market browsing."
+
+Based on this premise, can we conclude the hypothesis "A man is sleeping at a flea market." is true?
+
+OPTIONS:
+
+- yes
+
+- no
+
+- not possible to tell
+
+A: One cannot be sleeping and browsing at the same time. The answer is no.
+
+Figure 5: Detailed examples of prompts for GSM, ECQA, and E-SNLI.
+
+# F License of Datasets
+
+- GSM (Cobbe et al., 2021): MIT license.  
+E-SNLI (Camburu et al., 2018): MIT license.
+
+- ECQA (Aggarwal et al., 2021): Community Data License Agreement - Sharing - Version 1.0.  
+- LETTER CONCATENATION and COINFLIP (Wei et al., 2022): MIT license.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Section 8.
+
+A2. Did you discuss any potential risks of your work?
+
+This work focuses on analyzing what makes effective explanations for in-context learning. We believe that the risks of these methods are the same as the risks of broader in-context learning literature.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Section 1.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+As described in Section 2, we used prior datasets and publicly available language model APIs.
+
+B1. Did you cite the creators of artifacts you used?
+
+Yes, we cite them in Section 2 and throughout the paper.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+In Appendix F.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+We only use data and code specifically prepared for research contexts.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+These are standard datasets that do not contain information about individual people or offensive context to our knowledge.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+As described in Section 8 limitation, all the datasets we used are written in English.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+We report the setup throughout the paper as well as in the Appendix.
+
+C Did you run computational experiments?
+
+Section 2, 3, 4, and 5.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+We are mainly using APIs, so we report exact API versions instead of computing infrastructure.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Described in Appendix B,C, and D
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+We reported the experimental setup throughout the paper (Section 2, 3, 4, and 5).
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+We described the language models we are using in Section 2.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+No response.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+No response.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+No response.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+No response.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+No response.

2023/Complementary Explanations for Effective In-Context Learning/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95193d3a80639c20cbba993ad3e0244c84b905490ec0f3aaac5e6b5e2187c2a3
+size 515577

2023/Compositional Mathematical Encoding for Math Word Problems/801bbb7b-d8b8-4c19-87bf-9a90de8c668a_content_list.json

@@ -0,0 +1,1750 @@
+[
+    {
+        "type": "text",
+        "text": "Compositional Mathematical Encoding for Math Word Problems",
+        "text_level": 1,
+        "bbox": [
+            161,
+            89,
+            836,
+            111
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zhenwen Liang $^{1}$ , Jipeng Zhang $^{2}$ , Kehan Guo $^{1}$ , Xiaodong Wu $^{3}$ , Jie Shao $^{4}$ , and Xiangliang Zhang $^{\\boxtimes 1}$",
+        "bbox": [
+            284,
+            131,
+            719,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>University of Notre Dame, {zliang6, kguo2, xzhang33}@nd.edu  \n<sup>2</sup>Hong Kong University of Science and Technology, jzhanggr@connect.ust.hk  \n<sup>3</sup>Queen's University, 22hd17@queensu.ca  \n<sup>4</sup>University of Electronic Science and Technology of China, shaojie@uestc.edu.cn",
+        "bbox": [
+            129,
+            178,
+            870,
+            247
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            286,
+            339,
+            300
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Solving math word problem (MWP) remains a challenging task, as it requires to understand both the semantic meanings of the text and the mathematical logic among quantities, i.e., for both semantics modal and quantity modal learning. Current MWP encoders work in a uni-modal setting and map the given problem description to a latent representation, then for decoding. The generalizability of these MWP encoders is thus limited because some problems are semantics-demanding and others are quantity-demanding. To address this problem, we propose a Compositional Math Word Problem Solver (C-MWP) which works in a bimodal setting encoding in an interactive way. Extensive experiments validate the effectiveness of C-MWP and show its superiority over state-of-the-art models on public benchmarks.",
+        "bbox": [
+            141,
+            307,
+            460,
+            565
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            571,
+            260,
+            588
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The task of math word problem (MWP) solving aims to map natural language problem descriptions into executable solution equations to get the correct answer, which is a sub-area of neuro-symbolic reasoning. It requires perceptual abilities such as comprehending the question, identifying the quantities and corresponding attributes, as well as complex semantics understanding skills like performing logical inference, making comparisons and leveraging external mathematical knowledge.",
+        "bbox": [
+            112,
+            596,
+            489,
+            758
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "While MWP encoders have been sophisticatedly designed to understand the natural language problem description, the difference on understanding diverse types of problems has not been aware of. We find that MWP can generally be grouped into three categories based on the keywords in (Liu et al., 2019), i.e., \"Story Problem\", \"Algebra Problem\" and \"Knowledge Problem\". \"Story Problem\" often includes significant amount of background information like characters, objectives and behaviors.",
+        "bbox": [
+            112,
+            758,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\"Algebra Problems\" involves math notations or is composed of elementary concepts. \"Knowledge Problem\" asks for external knowledge like geometry and number sequence, as shown in Figure 1.",
+        "bbox": [
+            507,
+            286,
+            884,
+            351
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "These types of problems can be compositionally understood at the different level attention to the semantics modal and quantity modal, where RNNs and pre-trained language models usually focus on the textual information and GCN with quantity-centered graphs capture the relationship between quantities and contexts. However, the encoders in existing MWP solvers either model only the semantics modality or utilize quantity modal priors to refine the MWP encoding (Zhang et al., 2020; Shen and Jin, 2020). Although the quantity centered refinement can particularly make improvements on quantity-demanding problems, its semantics understanding is weakened (evidence can be found in Table 3). This limitation, one joint modal cannot do it all, decreases the generalization of MWP solvers and is what compositional learning aims to address. In this work, we propose to disentangle semantics modal and quantity modal by compositional learning at the encoding stage, aiming to improve the generalization across different types of problems.",
+        "bbox": [
+            507,
+            361,
+            884,
+            715
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Contributions. (i) A novel and effective bimodal approach is proposed for the first time to enable MWP compositional understanding. (ii) A joint reasoning module is designed for our bimodal architectures to flexibly incorporate different modalities. (iii) Extensive experiments and ablative studies on two large-scale MWP benchmarks – Math23k (Wang et al., 2017) and MAwPS (Koncel-Kedziorski et al., 2016) show the superiority of the proposed approach over related works.",
+        "bbox": [
+            507,
+            758,
+            885,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10008",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 10008-10017 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/762bc4966b60b6fa4bc9b5cf248bd7d470551fe672405852f5b0a3dfcc1095b7.jpg",
+        "image_caption": [
+            "Figure 1: Examples of different types of problems in MWP solving."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            84,
+            884,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            406,
+            270,
+            422
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Compositional Learning in NLP. Modeling compositionality in language has been a longstanding issue (Wong and Wang, 2007) in NLP community. One common practice is to perform disentanglement over language representations at different levels (Welch et al., 2020). They usually focus on atomic semantics units like character, word and phrase. As logic form annotations naturally own compositional features, compositionality is incorporated in generating correct logic contents. Therefore, the compositionality is often injected into traditional semantic parsing tasks (Chen et al., 2020; Yang et al., 2022) where the goals during training can be decomposed and then reorganized as a novel goal.",
+        "bbox": [
+            112,
+            436,
+            489,
+            677
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our work firstly tries to inject compositional prior into MWP encoding. It is worth noting that MWP solving owns the same well-organized logic form annotations as machine reasoning, which naturally requires compositionality.",
+        "bbox": [
+            112,
+            678,
+            489,
+            760
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Math Word Problem Solving. Earlier MWP solvers parse problem descriptions semantically, and learn templates for generating answers (Koncel-Kedziorski et al., 2015). Recent works (Wang et al., 2017; Xie and Sun, 2019; Li et al., 2019; Zhang et al., 2020; Shen and Jin, 2020; Wu et al., 2021b,a; Lin et al., 2021; Liang and Zhang, 2021; Jie et al., 2022) focus on employing the encoder-decoder framework (e.g., sequence-to-sequence,",
+        "bbox": [
+            112,
+            774,
+            492,
+            920
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "sequence-to-tree, graph-to-tree) to translate MWP texts into equations based on traditional RNN structure. There are also new settings (Amini et al., 2019; Miao et al., 2020) introduced to extend MWP solving in equation group generation and diagnosing awareness of external knowledge. Nowadays, many researchers build strong MWP solvers upon pre-trained language models (PLMs) (Huang et al., 2021; Li et al., 2021; Yu et al., 2021; Shen et al., 2021; Lan et al., 2022) and have achieved great performance. Differently, our work lays the groundwork of feature extraction of quantity modal, which is orthogonal to those works.",
+        "bbox": [
+            507,
+            407,
+            884,
+            615
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In this work, we not only propose an explicit compositional encoding module with a multi-layer design, but also incorporate detailed analysis to verify its compositional learning ability, to jointly leverage semantic and quantity information to achieve effective MWP understanding.",
+        "bbox": [
+            507,
+            617,
+            882,
+            713
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Our approach",
+        "text_level": 1,
+        "bbox": [
+            507,
+            726,
+            668,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Compositional Mathematical Encoder",
+        "text_level": 1,
+        "bbox": [
+            507,
+            753,
+            855,
+            768
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 2, our CMEncoder block consists of a semantic encoder, a quantity encoder and a dynamic fusion block. The semantic encoder aims to extract semantic information from the problem description, understanding the background and objectives. The latter part encodes problems only with quantity-related graphs, helping the encoder to know the properties of quantities and the relationship between quantities and contexts.",
+        "bbox": [
+            507,
+            774,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10009",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/2cee7b44492be75c82ea2b3a2de99f53cf4f1f239381d0c0c68bc14878e23248.jpg",
+        "image_caption": [
+            "Figure 2: The CMEncoder block (shown at the bottom) takes the given problem description, and runs in parallel to obtain  $H_{r}$  from the semantic encoder, and  $H_{g}$  from the quantity encoder. A dynamic fusion module incorporates  $H_{r}$  and  $H_{g}$  by cross-modal attention. The obtained  $H_{f}$  is attentively stacked with  $H_{r}$  and  $H_{g}$ . The resulted  $H_{att}$  is sent to the next CMEncoder block. The final problem representation  $H_{final}$  goes to the decoder for generating the final solution equation."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            166,
+            80,
+            833,
+            355
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Semantic Encoder. To demonstrate the robustness of our approach, we implemented two different semantic encoders as our backbone.",
+        "bbox": [
+            112,
+            463,
+            489,
+            511
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Firstly, we encode the problem description  $W$  by a bidirectional gated recurrent unit (BiGRU) (Cho et al., 2014). The outputs of GRU are hidden state vectors of all tokens,  $H_{r} = \\{h_{1}, h_{2}, \\dots, h_{n}\\}$ , where  $n$  is the length of problem  $W$ .",
+        "bbox": [
+            112,
+            511,
+            487,
+            592
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {r} = \\operatorname {B i G R U} \\left(\\operatorname {E m b e d} _ {s} (W)\\right) \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            608,
+            485,
+            625
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $Embed_{s}(W)$  is the embedding result of textual description  $W$  in semantics modal. Empirically, we find that two stacked CMEncoders as shown in Figure 2 achieve the best performance.",
+        "bbox": [
+            112,
+            634,
+            487,
+            697
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Secondly, pre-trained language models (PLMs) have been ubiquitous in NLP tasks. We use the latest push of MWP-BERT (Liang et al., 2022) as our semantic encoder to obtain  $H_{r}$ .",
+        "bbox": [
+            112,
+            699,
+            487,
+            763
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Quantity Encoder. To encode the quantity modal in the problem  $W$ , we feed a graph transformer  $G_{trans}$  with Quantity Comparison Graph and Quantity Cell Graph following Graph2Tree (Zhang et al., 2020),",
+        "bbox": [
+            112,
+            769,
+            489,
+            850
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {g} = G _ {\\text {t r a n s}} \\left(\\operatorname {E m b e d} _ {q} (W)\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            193,
+            865,
+            485,
+            883
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\operatorname{Embed}_q(W)$  is the embedding matrix in the quantity modal, which aims to improve the",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "quantity representation by incorporating quantity magnitude information and quantity-context relationship with the above two graphs. Different from Graph2Tree, the two embeddings  $Embed_{s}(W)$  and  $Embed_{q}(W)$  are updated in the training process to extract the semantics and quantity feature separately. In this way, semantics and quantity modals are disentangled, which alleviates the issue of \"one joint modal cannot do it all\", enabling the C-MWP solver to pay different levels of attention when solving different problems.",
+        "bbox": [
+            507,
+            463,
+            884,
+            640
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Dynamic Fusion. To achieve joint reasoning over the semantics information and quantity information, we design a dynamic fusion module to flexibly incorporate the features from these two modals. First, we get  $s$  and  $q$  from the mean pooling of  $H_{r}$  and  $H_{g}$ , respectively. Then, cross-modal attention is applied between  $H_{r}$  and  $q$ ,  $H_{g}$  and  $s$ :",
+        "bbox": [
+            507,
+            648,
+            882,
+            762
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} A t t _ {1} \\left(H _ {r}, q\\right) = \\Sigma_ {i = 1} ^ {n} a _ {i} H _ {r i} \\\\ A t t _ {1} \\left(H _ {r}\\right) = \\Sigma_ {i = 1} ^ {n} H _ {r i} \\end{array} \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            596,
+            770,
+            880,
+            795
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nA t t _ {2} (H _ {g}, s) = \\Sigma_ {i = 1} ^ {n} b _ {i} H _ {g i}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            598,
+            790,
+            794,
+            808
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where the attention scores  $a_{i}, b_{i}$  come from:",
+        "bbox": [
+            507,
+            816,
+            836,
+            831
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\na _ {i} = W _ {a} ^ {1} \\tanh  \\left(W _ {a} ^ {2} \\left(H _ {r i} \\mid \\mid q\\right)\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            579,
+            837,
+            880,
+            860
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nb _ {i} = W _ {b} ^ {1} \\tanh  (W _ {b} ^ {2} \\left(H _ {g i} \\mid | s\\right))\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            584,
+            859,
+            808,
+            878
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{a}^{1}, W_{a}^{2}, W_{b}^{1}$  and  $W_{b}^{2}$  are parameter matrices. The cross-modal attention here grounds the",
+        "bbox": [
+            507,
+            885,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10010",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "quantity information in the semantics modal, and vice versa. By applying different weights on different modals, our model is flexible to pay more or less attention on a certain modal. Finally, the output of dynamic fusion is:",
+        "bbox": [
+            112,
+            84,
+            489,
+            164
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {f} = \\operatorname {A t t} _ {1} \\left(H _ {r}, q\\right) \\mid \\mid \\operatorname {A t t} _ {2} \\left(H _ {g}, s\\right). \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            175,
+            487,
+            193
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Stack Multiple CMEncoders",
+        "text_level": 1,
+        "bbox": [
+            114,
+            205,
+            386,
+            221
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Humans often need to make multiple glimpses to refine an MWP solution. Similarly, a CMEncoder can be stacked in multiple steps to refine the understanding of an MWP, as shown in Figure 2. Given the output from the semantic encoder, quantity encoder and dynamic fusion module at layer  $k - 1$ , the features are stacked as:",
+        "bbox": [
+            112,
+            225,
+            489,
+            338
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nH _ {a t t} ^ {(k - 1)} = c _ {r} H _ {r} ^ {(k - 1)} + c _ {g} H _ {g} ^ {(k - 1)} \\tag {6}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            176,
+            348,
+            487,
+            370
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where the attention weights  $c_{r}$  and  $c_{g}$  are:",
+        "bbox": [
+            112,
+            381,
+            426,
+            398
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} c _ {r} = W _ {r} ^ {1} \\tanh  \\left(W _ {r} ^ {2} \\left(H _ {r} ^ {(k - 1)} \\mid \\mid H _ {f} ^ {(k - 1)}\\right)\\right) \\tag {7} \\\\ c _ {g} = W _ {g} ^ {1} \\tanh  (W _ {g} ^ {2} \\left(H _ {g} ^ {(k - 1)} \\mid \\mid H _ {f} ^ {(k - 1)}\\right)) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            137,
+            409,
+            487,
+            456
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $W_{r}^{1}, W_{r}^{2}, W_{g}^{1}$  and  $W_{g}^{2}$  are parameter matrices. The  $H_{att}^{(k - 1)}$  will be the input for both the semantic modal and quantity modal of K-th CMEncoder, which will output  $H_{r}^{(k)}, H_{g}^{(k)}$  and  $H_{f}^{(k)}$ , which can be sent for the update at layer  $k + 1$ .",
+        "bbox": [
+            112,
+            467,
+            489,
+            554
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After finishing the  $K$ -th step reasoning, we concatenate the final  $H_r^{(K)}$  and  $H_g^{(K)}$  as the final output representation  $H_{final}$ .",
+        "bbox": [
+            112,
+            555,
+            489,
+            606
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.3 Decoder",
+        "text_level": 1,
+        "bbox": [
+            114,
+            615,
+            226,
+            629
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We follow the same implementation as GTS (Xie and Sun, 2019). Eventually, the decoder will output the pre-order traversal sequence of the solution tree.",
+        "bbox": [
+            112,
+            636,
+            489,
+            684
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.4 Training Method",
+        "text_level": 1,
+        "bbox": [
+            114,
+            695,
+            294,
+            709
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Given the training samples with problem description  $W$  and the corresponding solution  $S$ , the main training objective is to minimize the negative log probability for predicting  $S$  from  $W$ , empowered by the compositionality of the CMEncoders. Therefore, the overall loss is:",
+        "bbox": [
+            112,
+            715,
+            489,
+            810
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL = L _ {M W P} + \\left| \\left| E m b e d _ {s} \\right| \\right| _ {2} + \\left| \\left| E m b e d _ {q} \\right| \\right| _ {2} (8)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            822,
+            487,
+            844
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $L_{MWP}$  is the negative log prediction probability  $-\\log p(S \\mid W)$ . The  $L_2$  norm of the encoder embedding matrices is added to the loss function as regularization terms.",
+        "bbox": [
+            112,
+            854,
+            489,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            655,
+            99
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            509,
+            109,
+            623,
+            123
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Math23k (Wang et al., 2017) containing 23,162 Chinese MWPs is collected from several educational websites.",
+        "bbox": [
+            507,
+            130,
+            882,
+            177
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "MAWPS (Koncel-Kedziorski et al., 2015) is an MWP dataset owning 2,373 English MwPs.",
+        "bbox": [
+            507,
+            179,
+            880,
+            210
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            222,
+            628,
+            236
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "GTS (Xie and Sun, 2019) proposes a powerful tree-based decoder. Graph2Tree (Zhang et al., 2020) constructs graphs to extract useful relationships in an MWP. NumS2T (Wu et al., 2021b) encode quantities with explicit numerical values. MultiE/D (Shen and Jin, 2020) proposes to use multiple decoders in MWP solving. HMS (Lin et al., 2021) develops a hierarchical word-clause-problem encoder. EEH-G2T (Wu et al., 2021a) aims to capture the long-range word relationship by graph network. REAL (Huang et al., 2021) proposes a analogical auxiliary learning strategy by extracting similar MWPs. BERT-CL (Li et al., 2021) uses contrastive learning with PLMs. RPKHS (Yu et al., 2021) performs hierarchical reasoning with PLMs. MWP-BERT released a BERT-based encoder that is continually pre-trained on MWP corpus. Gen&Rank (Shen et al., 2021) designs a multi-task learning framework with encoder-decoder pre-training. MWP toolkit (Lan et al., 2022) finds a RoBERTa-to-RoBERTa model has the best performance in MWP solving.",
+        "bbox": [
+            507,
+            243,
+            884,
+            596
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Experimental Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            608,
+            724,
+            623
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "As Table 1 shows, our approach outperforms all other RNN-based baselines in terms of answer accuracy. On Math23k, we outperform the latest RNN-based push from Wu et al. (2021a) by  $1.8\\%$ . For the first time, an RNN-based MWP solver reaches over  $80\\%$  answer accuracy on the Math23k dataset. What is more, the even fewer parameters with the best performance suggest that our model is also memory-efficient.",
+        "bbox": [
+            507,
+            629,
+            882,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "PLM-based solvers benefit from the pre-training on a huge amount of corpus and thus achieve great semantic understanding ability. From a different point of view, our work aims to effectively and efficiently integrate semantic and quantity understanding. Therefore, by incorporating the MWP-BERT model as our semantic extractor, the answer accuracy of C-MWP achieves state-of-the-art performance. It proves the feasibility of combining",
+        "bbox": [
+            507,
+            774,
+            884,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10011",
+        "bbox": [
+            477,
+            927,
+            522,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/bccf05543f440dfd24ffd5344249301e98ad49eb80ba801050c0f38bf2d67e12.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Math23k</td><td>Math23k*</td><td>MAWPS</td><td>#E</td></tr><tr><td colspan=\"5\">RNN Based</td></tr><tr><td>DNS</td><td>-</td><td>58.1</td><td>59.5</td><td>3.0M</td></tr><tr><td>GTS</td><td>75.6</td><td>74.3</td><td>82.6</td><td>7.2M</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>75.5</td><td>83.7</td><td>9.0M</td></tr><tr><td>NUMS2T</td><td>78.1</td><td>-</td><td>-</td><td>7.9M</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>76.9</td><td>-</td><td>14.2M</td></tr><tr><td>HMS</td><td>78.4</td><td>-</td><td>80.3</td><td>9.5M</td></tr><tr><td>EEH-G2T</td><td>78.5</td><td>-</td><td>84.8</td><td>9.9M</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>77.9</td><td>84.9</td><td>7.6M</td></tr><tr><td colspan=\"5\">PLM Based</td></tr><tr><td>REAL</td><td>82.3</td><td>80.0</td><td>-</td><td>110M</td></tr><tr><td>BERT-CL</td><td>83.2</td><td>-</td><td>-</td><td>102M</td></tr><tr><td>RPKHS</td><td>83.9</td><td>82.2</td><td>-</td><td>102M</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>82.4</td><td>-</td><td>110M</td></tr><tr><td>Gen&amp;Rank</td><td>85.4</td><td>84.3</td><td>-</td><td>610M</td></tr><tr><td>MWP toolkit</td><td>-</td><td>76.9</td><td>88.4</td><td>110M</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>84.5</td><td>89.1</td><td>130M</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            485,
+            331
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/a05bd2f5fef39bdd56ba38a4b051bfb7b4c111588dc778237090d04bf7d8d940.jpg",
+        "table_caption": [
+            "Table 1: Math23k column shows the results when evaluating on the public test set of Math23k, while the  $\\mathrm{Math23k^{*}}$  column shows the result of 5-fold cross validation on Math23k dataset. The last column #E denotes the number of parameters in encoders."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Graph Encoder</td><td>Compositional Structure</td><td>Dynamic Fusion</td><td>Acc(%)</td></tr><tr><td>GTS</td><td>X</td><td>X</td><td>X</td><td>75.6</td></tr><tr><td>Graph2Tree</td><td>✓</td><td>X</td><td>X</td><td>77.4</td></tr><tr><td></td><td>✓</td><td>✓</td><td>X</td><td>78.1</td></tr><tr><td></td><td>✓</td><td>X</td><td>✓</td><td>78.9</td></tr><tr><td>C-MWP</td><td>✓</td><td>✓</td><td>✓</td><td>80.3</td></tr></table>",
+        "bbox": [
+            117,
+            428,
+            485,
+            526
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "PLM-based semantic modal encoder and graph-based quantity modal encoder, which will be an interesting inspiration to the community.",
+        "bbox": [
+            112,
+            579,
+            489,
+            627
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Ablative Study of Different Components.",
+        "text_level": 1,
+        "bbox": [
+            114,
+            640,
+            440,
+            656
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to evaluate the effectiveness of each component in  $C$ -MWP, we report the model performance after removing several components. Compared with Graph2Tree, our compositional structure and dynamic fusion module allow the full usage of both modals and excel in improving performance. \"Compositional Structure\" denotes separating the semantic and quantity modal with two different encoders that run in parallel. \"w Compositional Structure and w/o Dynamic Fusion\" means we replace the Dynamic Fusion module with a simple addition of two features from two encoders. \"w/o Compositional Structure and w Dynamic Fusion\" means, we employ the structure of Graph2Tree, then utilize our dynamic fusion module to fuse the feature from the RNN Encoder and",
+        "bbox": [
+            112,
+            661,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/dbdf1d8bfaab6454789e74b243614b6a0e8251bdaba1ebef8bfb6f9dca3e1ce7.jpg",
+        "table_caption": [
+            "Table 2: Accuracy among different ablated models."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Model</td><td>Overall</td><td>Story</td><td>Algebra</td><td>KNWL</td></tr><tr><td>GTS</td><td>75.4</td><td>75.1</td><td>82.8</td><td>64.3</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>76.3</td><td>89.7</td><td>57.1</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>77.8</td><td>88.8</td><td>61.9</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>80.0</td><td>90.0</td><td>66.7</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>85.6</td><td>88.8</td><td>72.0</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>87.5</td><td>90.7</td><td>72.0</td></tr></table>",
+        "bbox": [
+            532,
+            80,
+            860,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3: The answer accuracy  $(\\%)$  of problems in different types. KNWL stands for the external knowledge required problems.",
+        "bbox": [
+            507,
+            191,
+            882,
+            233
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the feature from the GNN Encoder of Graph2Tree.",
+        "bbox": [
+            507,
+            256,
+            882,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Performance on Different Types of MWP.",
+        "text_level": 1,
+        "bbox": [
+            507,
+            281,
+            838,
+            297
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In order to investigate how our model performs across various types of MWP, we introduce a new split of Math23k with regard to three types of problems: story problems, algebra problems and knowledge problems. Split details are shown in the appendix. The evaluation results are presented in Table 3. Without a compositional manner, Graph2Tree and Multi-E/D perform better than GTS on story and algebra testing problems, whereas they perform worse on knowledge problems. As stated before, one joint modal cannot do it all. These baselines work well on some types of problems while having weak performance on other types of problems. Our C-MWP offers a general accuracy improvement, which firmly supports our motivation for alleviating the generalization issue. This provides clear evidence that our model leverages general math knowledge across different types of MWP, successfully solving some nontrivial problems that Graph2Tree failed to solve.",
+        "bbox": [
+            507,
+            300,
+            884,
+            621
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            633,
+            796,
+            648
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The semantic meaning and quantity information are important intrinsic properties of a math word problem. Aiming at dealing with uni-modal bias and achieve better generalization, we make the first attempt to propose a compositional MWP solver,  $C$ -MWP. Multi-layer reasoning and specified training methods are leveraged to enhance the generalizability of the model. As the method could be applied in a broader range of neuro-symbolic learning problems, we will keep exploring the adaptiveness of this compositional encoding method.",
+        "bbox": [
+            507,
+            658,
+            884,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            509,
+            846,
+            672,
+            862
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The research work is partially supported by the Internal Asia Research Collaboration Grant, University of Notre Dame.",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10012",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            211,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Explainability Most current MWP solvers are only able to generate solutions. In our work, although we achieved better generalization ability, it is still hard to explain how the model solves MwPs both correctly or incorrectly. These automated solvers would be much more helpful for tutoring students if they could explain their equation solutions by generating reasoning steps.",
+        "bbox": [
+            112,
+            105,
+            489,
+            235
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            260,
+            213,
+            275
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. In NAACL, pages 2357-2367.",
+            "Xinyun Chen, Chen Liang, Adams Wei Yu, Dawn Song, and Denny Zhou. 2020. Compositional generalization via neural-symbolic stack machines. NeurIPS, 33:1690-1701.",
+            "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, pages 1724-1734.",
+            "Shifeng Huang, Jiawei Wang, Jiao Xu, Da Cao, and Ming Yang. 2021. Recall and learn: A memory-augmented solver for math word problems. In Findings of EMNLP, pages 786-796.",
+            "Zhanming Jie, Jierui Li, and Wei Lu. 2022. Learning to reason deductively: Math word problem solving as complex relation extraction. In ACL, pages 5944-5955.",
+            "Rik Koncel-Kedziorski, Hannaneh Hajishirzi, Ashish Sabharwal, Oren Etzioni, and Siena Dumas Ang. 2015. Parsing algebraic word problems into equations. ACL, 3:585-597.",
+            "Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In NAACL, pages 1152-1157.",
+            "Yihuai Lan, Lei Wang, Qiyuan Zhang, Yunshi Lan, Bing Tian Dai, Yan Wang, Dongxiang Zhang, and Ee-Peng Lim. 2022. Mwptoolkit: An open-source framework for deep learning-based math word problem solvers. In AAAI, volume 36, pages 13188-13190.",
+            "Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In ACL, pages 6162-6167."
+        ],
+        "bbox": [
+            115,
+            282,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2021. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. arXiv preprint arXiv:2110.08464.",
+            "Zhenwen Liang, Jipeng Zhang, Lei Wang, Wei Qin, Yunshi Lan, Jie Shao, and Xiangliang Zhang. 2022. MWP-BERT: Numeracy-augmented pre-training for math word problem solving. In *Findings of NAACL*, pages 997-1009.",
+            "Zhenwen Liang and Xiangliang Zhang. 2021. Solving math word problems with teacher supervision. In *IJCAI*, pages 3522-3528.",
+            "Xin Lin, Zhenya Huang, Hongke Zhao, Enhong Chen, Qi Liu, Hao Wang, and Shijin Wang. 2021. Hms: A hierarchical solver with dependency-enhanced understanding for math word problem. In AAAI, pages 4232-4240.",
+            "Qianying Liu, Wenyv Guan, Sujian Li, and Daisuke Kawahara. 2019. Tree-structured decoding for solving math word problems. In EMNLP, pages 2370-2379.",
+            "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2020. A diverse corpus for evaluating and developing english math word problem solvers. In ACL, pages 975-984.",
+            "Jianhao Shen, Yichun Yin, Lin Li, Lifeng Shang, Xin Jiang, Ming Zhang, and Qun Liu. 2021. Generate & rank: A multi-task framework for math word problems. In EMNLP, pages 2269-2279.",
+            "Yibin Shen and Cheqing Jin. 2020. Solving math word problems with multi-encoders and multi-decoders. In COLING, pages 2924-2934.",
+            "Yan Wang, Xiaojiang Liu, and Shuming Shi. 2017. Deep neural solver for math word problems. In EMNLP, pages 845-854.",
+            "Charles Welch, Jonathan K Kummerfeld, Verónica Pérez-Rosas, and Rada Mihalcea. 2020. Compositional demographic word embeddings. In EMNLP, pages 4076-4089.",
+            "Francis CK Wong and William SY Wang. 2007. Generalisation towards combinatorial productivity in language acquisition by simple recurrent networks. In 2007 International Conference on Integration of Knowledge Intensive Multi-Agent Systems, pages 139-144. IEEE.",
+            "Qinzhuo Wu, Qi Zhang, and Zhongyu Wei. 2021a. An edge-enhanced hierarchical graph-to-tree network for math word problem solving. In *Findings of EMNLP*, pages 1473–1482.",
+            "Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuanjing Huang. 2021b. Math word problem solving with explicit numerical values. In ACL, pages 5859-5869."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10013",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Zhipeng Xie and Shichao Sun. 2019. A goal-driven tree-structured neural model for math word problems. In *IJCAI*, pages 5299–5305.",
+        "bbox": [
+            114,
+            85,
+            489,
+            124
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Jingfeng Yang, Le Zhang, and Diyi Yang. 2022. SUBS: Subtree substitution for compositional semantic parsing. In *NAACN*, pages 169–174, Seattle, United States. Association for Computational Linguistics.",
+        "bbox": [
+            114,
+            135,
+            489,
+            202
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Weijiang Yu, Yingpeng Wen, Fudan Zheng, and Nong Xiao. 2021. Improving math word problems with pre-trained knowledge and hierarchical reasoning. In EMNLP, pages 3384-3394.",
+        "bbox": [
+            114,
+            211,
+            487,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Jipeng Zhang, Lei Wang, Roy Ka-Wei Lee, Yi Bin, Yan Wang, Jie Shao, and Ee-Peng Lim. 2020. Graph-to-tree learning for solving math word problems. In ACL, pages 3928-3937.",
+        "bbox": [
+            114,
+            275,
+            489,
+            328
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            341,
+            203,
+            357
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Implementation Details",
+        "text_level": 1,
+        "bbox": [
+            114,
+            367,
+            302,
+            382
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We train our model on an NVIDIA RTX 2080Ti GPU, all implementation of training and testing is coded in Python with Pytorch framework. For our RNN-based model, 2 CMEncoders are stacked and only 1 CMEncoder is used in the PLM-based model. The size of hidden dimensions in encoders and decoders are set to 512 and 768 for RNN-solver and PLM-solver, respectively. Each GCN block has 2 GraphConv layers and each GNN encoder has 4 heads of GCN blocks. During training, Adam optimizer is applied with the initial learning rates of 0.001/0.00003 for RNN/PLM, which would be halved every 30 epochs. During testing, we use a 5-beam search to get reasonable solutions. We also apply Gaussian noise with mean 0 and variance 1 on the embedding result during training. This simple operation can help models to learn more robust parameters. Through grid search at 0.1 level, the noise is multiplied by 0.2 to achieve the best performance.",
+        "bbox": [
+            112,
+            387,
+            489,
+            709
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Hyper-Parameter Tuning",
+        "text_level": 1,
+        "bbox": [
+            114,
+            721,
+            319,
+            737
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "In general, we apply grid-search with manually designed search space and use answer accuracy as the evaluation metric to select the hyper-parameters. For the number of stacked encoders, the search space is  $\\{1,2,3,4\\}$  and we finally use 2. For the weight of  $L_{2}$  normalization loss, we choose weight 1 from  $\\{0.01,0.1,1,5,10\\}$ . The weight of random noise 0.2 is selected from 0.1 level by grid search with range 0 to 1. We also tune the beam size of beam search from  $\\{3,4,5,6,7\\}$  and choose 5. The dropout probability 0.5 is selected from",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/54bc66a762036dba326a22f8d89999fbbdbed38a36e64680663e888a3524d2da.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>K = 1</td><td>K = 2</td><td>K = 3</td><td>K = 4</td></tr><tr><td>C-MWP (RNN)</td><td>79.1</td><td>80.3</td><td>78.9</td><td>78.1</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>85.9</td><td>85.2</td><td>85.1</td></tr></table>",
+        "bbox": [
+            531,
+            80,
+            862,
+            143
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Accuracies across different numbers of stacked CMEncoder on Math23k.",
+        "bbox": [
+            507,
+            153,
+            882,
+            181
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "$\\{0.1, 0.3, 0.5, 0.7\\}$ . Initial learning rate 0.001 is selected from  $\\{0.01, 0.001, 0.0001\\}$ . For the hidden size and embedding size in encoder, we select 256 from  $\\{64, 128, 256, 512\\}$ .",
+        "bbox": [
+            507,
+            208,
+            882,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Variance and Significance Evaluation",
+        "text_level": 1,
+        "bbox": [
+            509,
+            284,
+            806,
+            299
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We evaluated our solver with 5-fold cross-validation and found that the accuracy of our RNN-based C-MWP  $(0.779 \\pm 0.028)$  is significantly higher than Graph2Tree  $(0.755 \\pm 0.016)$ $(p < 0.01)$ , and the accuracy of our PLM-based C-MWP  $(0.845 \\pm 0.21)$  is significantly higher than vanilla MWP-BERT  $(0.824 \\pm 0.016)$ $(p < 0.01)$ .",
+        "bbox": [
+            507,
+            305,
+            882,
+            418
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Sensitivity Analysis about Stacking Number K",
+        "text_level": 1,
+        "bbox": [
+            507,
+            430,
+            877,
+            445
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As we mentioned in Section 2.2, our CMEncoder can be stacked into multiple layers to improve the representation of an MWP. To obtain a better understanding about the hyperparameter K, i.e., the number of stacked CMEncoders, we conduct a sensitivity analysis in Table 4. We can see that the best K for RNN-encoder is 2, just like humans often need to make multiple glimpses to refine an MWP solution. In the meantime, the best K for PLM-based encoder is 1. The potential reason is that one pre-trained language model (PLM) already has an outstanding ability to encode texts. It is thus not necessary to apply another CMEncoder to refine the encoded features.",
+        "bbox": [
+            507,
+            451,
+            882,
+            675
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Case Study",
+        "text_level": 1,
+        "bbox": [
+            509,
+            688,
+            603,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 shows generated solutions of two solved problems by GTS (Xie and Sun, 2019), Graph2Tree (Zhang et al., 2020), Multi-E/D (Shen and Jin, 2020) and our proposed C-MWP (RNN-Based). The first problem has 4 quantities and they are all useful, which means that it requires sufficient problem understanding and mathematical reasoning to generate the right answer. Both Graph2Tree and Multi-E/D which directly connect semantics modal and quantity modal fail to extract clear representations of the problem, finally resulting in unreasonable solutions which only contain 3 quantities. For the second problem, although Graph2Tree and",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10014",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "image",
+        "img_path": "images/301334b5270e7ff6bc751ca3b631cf30c4a4051d05e20b057034b14c68b75ecf.jpg",
+        "image_caption": [
+            "Figure 3: Case study from Math23k"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            134,
+            80,
+            860,
+            292
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/0be23824859015f785577dfcc534cff872ab23d198c861d9583f560886848a30.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Overall</td><td>Story</td><td>Algebra</td><td>Knowledge</td></tr><tr><td>train</td><td>21,162</td><td>17,546</td><td>2,595</td><td>957</td></tr><tr><td>val</td><td>1,000</td><td>817</td><td>133</td><td>50</td></tr><tr><td>test</td><td>1,000</td><td>842</td><td>116</td><td>42</td></tr></table>",
+        "bbox": [
+            137,
+            338,
+            463,
+            419
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Statistics of different types of problems in Math23k.",
+        "bbox": [
+            112,
+            429,
+            487,
+            458
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Multi-E/D utilize all 3 quantities in the problem description, they still fail to generate a plausible solution. These two cases show that our proposed encoder is able to extract more comprehensive representations from problem descriptions, eventually guiding the decoder to generate the correct solutions.",
+        "bbox": [
+            112,
+            483,
+            489,
+            594
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "MWPs in Different Categories",
+        "text_level": 1,
+        "bbox": [
+            114,
+            607,
+            356,
+            621
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Figure 1 shows the MWP examples of \"Story Problem\", \"Algebra Problem\" and \"Knowledge Problem\". \"Story Problem\" often includes a significant amount of background information like characters, objectives and behaviors. \"Algebra Problems\" involves math notations or is composed of elementary concepts. \"Knowledge Problem\" asks for external knowledge like geometry and number sequence. The category of each problem is determined based on keywords. The keywords of \"Story\" and \"Knowledge\" problems are selected from the appendix of (Liu et al., 2019). Inspired by them, we categorize Math23k into 3 subsets - story, algebra, and knowledge. The statistics of these problems are shown in Table 5.",
+        "bbox": [
+            112,
+            627,
+            489,
+            866
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10015",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "text_level": 1,
+        "bbox": [
+            114,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            127,
+            127,
+            532,
+            143
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            143,
+            231,
+            159
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            127,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            187,
+            231,
+            200
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            127,
+            212,
+            695,
+            229
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            230,
+            231,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "□ A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            127,
+            255,
+            668,
+            272
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            300,
+            490,
+            316
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            127,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            363,
+            231,
+            379
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            127,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            407,
+            231,
+            422
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            127,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            498,
+            231,
+            513
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            127,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            573,
+            231,
+            588
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            127,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            632,
+            231,
+            646
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            127,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            740,
+            231,
+            753
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            765,
+            495,
+            781
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            785,
+            215,
+            801
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            127,
+            813,
+            880,
+            845
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            846,
+            231,
+            860
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10016",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        ],
+        "bbox": [
+            127,
+            84,
+            880,
+            282
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "text_level": 1,
+        "bbox": [
+            114,
+            293,
+            877,
+            309
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank."
+        ],
+        "bbox": [
+            127,
+            340,
+            880,
+            640
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10017",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 9
+    }
+]

2023/Compositional Mathematical Encoding for Math Word Problems/801bbb7b-d8b8-4c19-87bf-9a90de8c668a_model.json

@@ -0,0 +1,2046 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.162,
+                0.09,
+                0.837,
+                0.112
+            ],
+            "angle": 0,
+            "content": "Compositional Mathematical Encoding for Math Word Problems"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.285,
+                0.132,
+                0.72,
+                0.166
+            ],
+            "angle": 0,
+            "content": "Zhenwen Liang\\(^{1}\\), Jipeng Zhang\\(^{2}\\), Kehan Guo\\(^{1}\\), Xiaodong Wu\\(^{3}\\), Jie Shao\\(^{4}\\), and Xiangliang Zhang\\(^{\\boxtimes 1}\\)"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.179,
+                0.871,
+                0.248
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>University of Notre Dame, {zliang6, kguo2, xzhang33}@nd.edu  \n<sup>2</sup>Hong Kong University of Science and Technology, jzhanggr@connect.ust.hk  \n<sup>3</sup>Queen's University, 22hd17@queensu.ca  \n<sup>4</sup>University of Electronic Science and Technology of China, shaojie@uestc.edu.cn"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.287,
+                0.341,
+                0.302
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.309,
+                0.461,
+                0.566
+            ],
+            "angle": 0,
+            "content": "Solving math word problem (MWP) remains a challenging task, as it requires to understand both the semantic meanings of the text and the mathematical logic among quantities, i.e., for both semantics modal and quantity modal learning. Current MWP encoders work in a uni-modal setting and map the given problem description to a latent representation, then for decoding. The generalizability of these MWP encoders is thus limited because some problems are semantics-demanding and others are quantity-demanding. To address this problem, we propose a Compositional Math Word Problem Solver (C-MWP) which works in a bimodal setting encoding in an interactive way. Extensive experiments validate the effectiveness of C-MWP and show its superiority over state-of-the-art models on public benchmarks."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.573,
+                0.262,
+                0.589
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.598,
+                0.49,
+                0.759
+            ],
+            "angle": 0,
+            "content": "The task of math word problem (MWP) solving aims to map natural language problem descriptions into executable solution equations to get the correct answer, which is a sub-area of neuro-symbolic reasoning. It requires perceptual abilities such as comprehending the question, identifying the quantities and corresponding attributes, as well as complex semantics understanding skills like performing logical inference, making comparisons and leveraging external mathematical knowledge."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.759,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "While MWP encoders have been sophisticatedly designed to understand the natural language problem description, the difference on understanding diverse types of problems has not been aware of. We find that MWP can generally be grouped into three categories based on the keywords in (Liu et al., 2019), i.e., \"Story Problem\", \"Algebra Problem\" and \"Knowledge Problem\". \"Story Problem\" often includes significant amount of background information like characters, objectives and behaviors."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.287,
+                0.885,
+                0.352
+            ],
+            "angle": 0,
+            "content": "\"Algebra Problems\" involves math notations or is composed of elementary concepts. \"Knowledge Problem\" asks for external knowledge like geometry and number sequence, as shown in Figure 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.362,
+                0.885,
+                0.716
+            ],
+            "angle": 0,
+            "content": "These types of problems can be compositionally understood at the different level attention to the semantics modal and quantity modal, where RNNs and pre-trained language models usually focus on the textual information and GCN with quantity-centered graphs capture the relationship between quantities and contexts. However, the encoders in existing MWP solvers either model only the semantics modality or utilize quantity modal priors to refine the MWP encoding (Zhang et al., 2020; Shen and Jin, 2020). Although the quantity centered refinement can particularly make improvements on quantity-demanding problems, its semantics understanding is weakened (evidence can be found in Table 3). This limitation, one joint modal cannot do it all, decreases the generalization of MWP solvers and is what compositional learning aims to address. In this work, we propose to disentangle semantics modal and quantity modal by compositional learning at the encoding stage, aiming to improve the generalization across different types of problems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.759,
+                0.887,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Contributions. (i) A novel and effective bimodal approach is proposed for the first time to enable MWP compositional understanding. (ii) A joint reasoning module is designed for our bimodal architectures to flexibly incorporate different modalities. (iii) Extensive experiments and ablative studies on two large-scale MWP benchmarks – Math23k (Wang et al., 2017) and MAwPS (Koncel-Kedziorski et al., 2016) show the superiority of the proposed approach over related works."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10008"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.221,
+                0.946,
+                0.779,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 10008-10017 July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.118,
+                0.085,
+                0.885,
+                0.358
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.267,
+                0.368,
+                0.729,
+                0.384
+            ],
+            "angle": 0,
+            "content": "Figure 1: Examples of different types of problems in MWP solving."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.407,
+                0.271,
+                0.423
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.437,
+                0.49,
+                0.678
+            ],
+            "angle": 0,
+            "content": "Compositional Learning in NLP. Modeling compositionality in language has been a longstanding issue (Wong and Wang, 2007) in NLP community. One common practice is to perform disentanglement over language representations at different levels (Welch et al., 2020). They usually focus on atomic semantics units like character, word and phrase. As logic form annotations naturally own compositional features, compositionality is incorporated in generating correct logic contents. Therefore, the compositionality is often injected into traditional semantic parsing tasks (Chen et al., 2020; Yang et al., 2022) where the goals during training can be decomposed and then reorganized as a novel goal."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.68,
+                0.49,
+                0.761
+            ],
+            "angle": 0,
+            "content": "Our work firstly tries to inject compositional prior into MWP encoding. It is worth noting that MWP solving owns the same well-organized logic form annotations as machine reasoning, which naturally requires compositionality."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.775,
+                0.494,
+                0.921
+            ],
+            "angle": 0,
+            "content": "Math Word Problem Solving. Earlier MWP solvers parse problem descriptions semantically, and learn templates for generating answers (Koncel-Kedziorski et al., 2015). Recent works (Wang et al., 2017; Xie and Sun, 2019; Li et al., 2019; Zhang et al., 2020; Shen and Jin, 2020; Wu et al., 2021b,a; Lin et al., 2021; Liang and Zhang, 2021; Jie et al., 2022) focus on employing the encoder-decoder framework (e.g., sequence-to-sequence,"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.409,
+                0.885,
+                0.617
+            ],
+            "angle": 0,
+            "content": "sequence-to-tree, graph-to-tree) to translate MWP texts into equations based on traditional RNN structure. There are also new settings (Amini et al., 2019; Miao et al., 2020) introduced to extend MWP solving in equation group generation and diagnosing awareness of external knowledge. Nowadays, many researchers build strong MWP solvers upon pre-trained language models (PLMs) (Huang et al., 2021; Li et al., 2021; Yu et al., 2021; Shen et al., 2021; Lan et al., 2022) and have achieved great performance. Differently, our work lays the groundwork of feature extraction of quantity modal, which is orthogonal to those works."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.618,
+                0.884,
+                0.714
+            ],
+            "angle": 0,
+            "content": "In this work, we not only propose an explicit compositional encoding module with a multi-layer design, but also incorporate detailed analysis to verify its compositional learning ability, to jointly leverage semantic and quantity information to achieve effective MWP understanding."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.727,
+                0.669,
+                0.743
+            ],
+            "angle": 0,
+            "content": "3 Our approach"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.754,
+                0.857,
+                0.769
+            ],
+            "angle": 0,
+            "content": "3.1 Compositional Mathematical Encoder"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.775,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "As shown in Figure 2, our CMEncoder block consists of a semantic encoder, a quantity encoder and a dynamic fusion block. The semantic encoder aims to extract semantic information from the problem description, understanding the background and objectives. The latter part encodes problems only with quantity-related graphs, helping the encoder to know the properties of quantities and the relationship between quantities and contexts."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10009"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.167,
+                0.082,
+                0.835,
+                0.356
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.113,
+                0.367,
+                0.883,
+                0.44
+            ],
+            "angle": 0,
+            "content": "Figure 2: The CMEncoder block (shown at the bottom) takes the given problem description, and runs in parallel to obtain \\( H_{r} \\) from the semantic encoder, and \\( H_{g} \\) from the quantity encoder. A dynamic fusion module incorporates \\( H_{r} \\) and \\( H_{g} \\) by cross-modal attention. The obtained \\( H_{f} \\) is attentively stacked with \\( H_{r} \\) and \\( H_{g} \\). The resulted \\( H_{att} \\) is sent to the next CMEncoder block. The final problem representation \\( H_{final} \\) goes to the decoder for generating the final solution equation."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.464,
+                0.49,
+                0.512
+            ],
+            "angle": 0,
+            "content": "Semantic Encoder. To demonstrate the robustness of our approach, we implemented two different semantic encoders as our backbone."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.512,
+                0.489,
+                0.593
+            ],
+            "angle": 0,
+            "content": "Firstly, we encode the problem description \\( W \\) by a bidirectional gated recurrent unit (BiGRU) (Cho et al., 2014). The outputs of GRU are hidden state vectors of all tokens, \\( H_{r} = \\{h_{1}, h_{2}, \\dots, h_{n}\\} \\), where \\( n \\) is the length of problem \\( W \\)."
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.187,
+                0.609,
+                0.487,
+                0.626
+            ],
+            "angle": 0,
+            "content": "\\[\nH _ {r} = \\operatorname {B i G R U} \\left(\\operatorname {E m b e d} _ {s} (W)\\right) \\tag {1}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.635,
+                0.489,
+                0.699
+            ],
+            "angle": 0,
+            "content": "where \\(Embed_{s}(W)\\) is the embedding result of textual description \\(W\\) in semantics modal. Empirically, we find that two stacked CMEncoders as shown in Figure 2 achieve the best performance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.7,
+                0.489,
+                0.764
+            ],
+            "angle": 0,
+            "content": "Secondly, pre-trained language models (PLMs) have been ubiquitous in NLP tasks. We use the latest push of MWP-BERT (Liang et al., 2022) as our semantic encoder to obtain \\( H_{r} \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.77,
+                0.49,
+                0.851
+            ],
+            "angle": 0,
+            "content": "Quantity Encoder. To encode the quantity modal in the problem \\(W\\), we feed a graph transformer \\(G_{trans}\\) with Quantity Comparison Graph and Quantity Cell Graph following Graph2Tree (Zhang et al., 2020),"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.194,
+                0.866,
+                0.486,
+                0.884
+            ],
+            "angle": 0,
+            "content": "\\[\nH _ {g} = G _ {\\text {t r a n s}} \\left(\\operatorname {E m b e d} _ {q} (W)\\right) \\tag {2}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.888,
+                0.489,
+                0.92
+            ],
+            "angle": 0,
+            "content": "where \\( \\operatorname{Embed}_q(W) \\) is the embedding matrix in the quantity modal, which aims to improve the"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.464,
+                0.885,
+                0.642
+            ],
+            "angle": 0,
+            "content": "quantity representation by incorporating quantity magnitude information and quantity-context relationship with the above two graphs. Different from Graph2Tree, the two embeddings \\(Embed_{s}(W)\\) and \\(Embed_{q}(W)\\) are updated in the training process to extract the semantics and quantity feature separately. In this way, semantics and quantity modals are disentangled, which alleviates the issue of \"one joint modal cannot do it all\", enabling the C-MWP solver to pay different levels of attention when solving different problems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.649,
+                0.884,
+                0.763
+            ],
+            "angle": 0,
+            "content": "Dynamic Fusion. To achieve joint reasoning over the semantics information and quantity information, we design a dynamic fusion module to flexibly incorporate the features from these two modals. First, we get \\( s \\) and \\( q \\) from the mean pooling of \\( H_{r} \\) and \\( H_{g} \\), respectively. Then, cross-modal attention is applied between \\( H_{r} \\) and \\( q \\), \\( H_{g} \\) and \\( s \\):"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.597,
+                0.771,
+                0.882,
+                0.796
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} A t t _ {1} \\left(H _ {r}, q\\right) = \\Sigma_ {i = 1} ^ {n} a _ {i} H _ {r i} \\\\ A t t _ {1} \\left(H _ {r}\\right) = \\Sigma_ {i = 1} ^ {n} H _ {r i} \\end{array} \\tag {3}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.599,
+                0.791,
+                0.796,
+                0.809
+            ],
+            "angle": 0,
+            "content": "\\[\nA t t _ {2} (H _ {g}, s) = \\Sigma_ {i = 1} ^ {n} b _ {i} H _ {g i}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.817,
+                0.838,
+                0.832
+            ],
+            "angle": 0,
+            "content": "where the attention scores \\(a_{i}, b_{i}\\) come from:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.581,
+                0.838,
+                0.881,
+                0.862
+            ],
+            "angle": 0,
+            "content": "\\[\na _ {i} = W _ {a} ^ {1} \\tanh  \\left(W _ {a} ^ {2} \\left(H _ {r i} \\mid \\mid q\\right)\\right) \\tag {4}\n\\]"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.585,
+                0.86,
+                0.81,
+                0.879
+            ],
+            "angle": 0,
+            "content": "\\[\nb _ {i} = W _ {b} ^ {1} \\tanh  (W _ {b} ^ {2} \\left(H _ {g i} \\mid | s\\right))\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.887,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "where \\(W_{a}^{1}, W_{a}^{2}, W_{b}^{1}\\) and \\(W_{b}^{2}\\) are parameter matrices. The cross-modal attention here grounds the"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10010"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.165
+            ],
+            "angle": 0,
+            "content": "quantity information in the semantics modal, and vice versa. By applying different weights on different modals, our model is flexible to pay more or less attention on a certain modal. Finally, the output of dynamic fusion is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.173,
+                0.176,
+                0.488,
+                0.194
+            ],
+            "angle": 0,
+            "content": "\\[\nH _ {f} = \\operatorname {A t t} _ {1} \\left(H _ {r}, q\\right) \\mid \\mid \\operatorname {A t t} _ {2} \\left(H _ {g}, s\\right). \\tag {5}\n\\]"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.206,
+                0.388,
+                0.222
+            ],
+            "angle": 0,
+            "content": "3.2 Stack Multiple CMEncoders"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.227,
+                0.49,
+                0.339
+            ],
+            "angle": 0,
+            "content": "Humans often need to make multiple glimpses to refine an MWP solution. Similarly, a CMEncoder can be stacked in multiple steps to refine the understanding of an MWP, as shown in Figure 2. Given the output from the semantic encoder, quantity encoder and dynamic fusion module at layer \\( k - 1 \\), the features are stacked as:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.177,
+                0.349,
+                0.488,
+                0.371
+            ],
+            "angle": 0,
+            "content": "\\[\nH _ {a t t} ^ {(k - 1)} = c _ {r} H _ {r} ^ {(k - 1)} + c _ {g} H _ {g} ^ {(k - 1)} \\tag {6}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.382,
+                0.427,
+                0.399
+            ],
+            "angle": 0,
+            "content": "where the attention weights \\( c_{r} \\) and \\( c_{g} \\) are:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.139,
+                0.41,
+                0.488,
+                0.457
+            ],
+            "angle": 0,
+            "content": "\\[\n\\begin{array}{l} c _ {r} = W _ {r} ^ {1} \\tanh  \\left(W _ {r} ^ {2} \\left(H _ {r} ^ {(k - 1)} \\mid \\mid H _ {f} ^ {(k - 1)}\\right)\\right) \\tag {7} \\\\ c _ {g} = W _ {g} ^ {1} \\tanh  (W _ {g} ^ {2} \\left(H _ {g} ^ {(k - 1)} \\mid \\mid H _ {f} ^ {(k - 1)}\\right)) \\\\ \\end{array}\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.468,
+                0.49,
+                0.555
+            ],
+            "angle": 0,
+            "content": "where \\( W_{r}^{1}, W_{r}^{2}, W_{g}^{1} \\) and \\( W_{g}^{2} \\) are parameter matrices. The \\( H_{att}^{(k - 1)} \\) will be the input for both the semantic modal and quantity modal of K-th CMEncoder, which will output \\( H_{r}^{(k)}, H_{g}^{(k)} \\) and \\( H_{f}^{(k)} \\), which can be sent for the update at layer \\( k + 1 \\)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.556,
+                0.49,
+                0.607
+            ],
+            "angle": 0,
+            "content": "After finishing the \\(K\\)-th step reasoning, we concatenate the final \\(H_r^{(K)}\\) and \\(H_g^{(K)}\\) as the final output representation \\(H_{final}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.616,
+                0.228,
+                0.63
+            ],
+            "angle": 0,
+            "content": "3.3 Decoder"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.637,
+                0.49,
+                0.686
+            ],
+            "angle": 0,
+            "content": "We follow the same implementation as GTS (Xie and Sun, 2019). Eventually, the decoder will output the pre-order traversal sequence of the solution tree."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.696,
+                0.295,
+                0.71
+            ],
+            "angle": 0,
+            "content": "3.4 Training Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.717,
+                0.49,
+                0.812
+            ],
+            "angle": 0,
+            "content": "Given the training samples with problem description \\( W \\) and the corresponding solution \\( S \\), the main training objective is to minimize the negative log probability for predicting \\( S \\) from \\( W \\), empowered by the compositionality of the CMEncoders. Therefore, the overall loss is:"
+        },
+        {
+            "type": "equation",
+            "bbox": [
+                0.124,
+                0.824,
+                0.488,
+                0.845
+            ],
+            "angle": 0,
+            "content": "\\[\nL = L _ {M W P} + \\left| \\left| E m b e d _ {s} \\right| \\right| _ {2} + \\left| \\left| E m b e d _ {q} \\right| \\right| _ {2} (8)\n\\]"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "where \\( L_{MWP} \\) is the negative log prediction probability \\( -\\log p(S \\mid W) \\). The \\( L_2 \\) norm of the encoder embedding matrices is added to the loss function as regularization terms."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.084,
+                0.657,
+                0.101
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.11,
+                0.624,
+                0.124
+            ],
+            "angle": 0,
+            "content": "4.1 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.131,
+                0.884,
+                0.178
+            ],
+            "angle": 0,
+            "content": "Math23k (Wang et al., 2017) containing 23,162 Chinese MWPs is collected from several educational websites."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.18,
+                0.882,
+                0.211
+            ],
+            "angle": 0,
+            "content": "MAWPS (Koncel-Kedziorski et al., 2015) is an MWP dataset owning 2,373 English MwPs."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.223,
+                0.63,
+                0.237
+            ],
+            "angle": 0,
+            "content": "4.2 Baselines"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.244,
+                0.885,
+                0.598
+            ],
+            "angle": 0,
+            "content": "GTS (Xie and Sun, 2019) proposes a powerful tree-based decoder. Graph2Tree (Zhang et al., 2020) constructs graphs to extract useful relationships in an MWP. NumS2T (Wu et al., 2021b) encode quantities with explicit numerical values. MultiE/D (Shen and Jin, 2020) proposes to use multiple decoders in MWP solving. HMS (Lin et al., 2021) develops a hierarchical word-clause-problem encoder. EEH-G2T (Wu et al., 2021a) aims to capture the long-range word relationship by graph network. REAL (Huang et al., 2021) proposes a analogical auxiliary learning strategy by extracting similar MWPs. BERT-CL (Li et al., 2021) uses contrastive learning with PLMs. RPKHS (Yu et al., 2021) performs hierarchical reasoning with PLMs. MWP-BERT released a BERT-based encoder that is continually pre-trained on MWP corpus. Gen&Rank (Shen et al., 2021) designs a multi-task learning framework with encoder-decoder pre-training. MWP toolkit (Lan et al., 2022) finds a RoBERTa-to-RoBERTa model has the best performance in MWP solving."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.609,
+                0.726,
+                0.624
+            ],
+            "angle": 0,
+            "content": "4.3 Experimental Results"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.63,
+                0.884,
+                0.773
+            ],
+            "angle": 0,
+            "content": "As Table 1 shows, our approach outperforms all other RNN-based baselines in terms of answer accuracy. On Math23k, we outperform the latest RNN-based push from Wu et al. (2021a) by \\(1.8\\%\\). For the first time, an RNN-based MWP solver reaches over \\(80\\%\\) answer accuracy on the Math23k dataset. What is more, the even fewer parameters with the best performance suggest that our model is also memory-efficient."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.775,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "PLM-based solvers benefit from the pre-training on a huge amount of corpus and thus achieve great semantic understanding ability. From a different point of view, our work aims to effectively and efficiently integrate semantic and quantity understanding. Therefore, by incorporating the MWP-BERT model as our semantic extractor, the answer accuracy of C-MWP achieves state-of-the-art performance. It proves the feasibility of combining"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10011"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.117,
+                0.082,
+                0.486,
+                0.332
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Math23k</td><td>Math23k*</td><td>MAWPS</td><td>#E</td></tr><tr><td colspan=\"5\">RNN Based</td></tr><tr><td>DNS</td><td>-</td><td>58.1</td><td>59.5</td><td>3.0M</td></tr><tr><td>GTS</td><td>75.6</td><td>74.3</td><td>82.6</td><td>7.2M</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>75.5</td><td>83.7</td><td>9.0M</td></tr><tr><td>NUMS2T</td><td>78.1</td><td>-</td><td>-</td><td>7.9M</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>76.9</td><td>-</td><td>14.2M</td></tr><tr><td>HMS</td><td>78.4</td><td>-</td><td>80.3</td><td>9.5M</td></tr><tr><td>EEH-G2T</td><td>78.5</td><td>-</td><td>84.8</td><td>9.9M</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>77.9</td><td>84.9</td><td>7.6M</td></tr><tr><td colspan=\"5\">PLM Based</td></tr><tr><td>REAL</td><td>82.3</td><td>80.0</td><td>-</td><td>110M</td></tr><tr><td>BERT-CL</td><td>83.2</td><td>-</td><td>-</td><td>102M</td></tr><tr><td>RPKHS</td><td>83.9</td><td>82.2</td><td>-</td><td>102M</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>82.4</td><td>-</td><td>110M</td></tr><tr><td>Gen&amp;Rank</td><td>85.4</td><td>84.3</td><td>-</td><td>610M</td></tr><tr><td>MWP toolkit</td><td>-</td><td>76.9</td><td>88.4</td><td>110M</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>84.5</td><td>89.1</td><td>130M</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.342,
+                0.49,
+                0.414
+            ],
+            "angle": 0,
+            "content": "Table 1: Math23k column shows the results when evaluating on the public test set of Math23k, while the \\(\\mathrm{Math23k^{*}}\\) column shows the result of 5-fold cross validation on Math23k dataset. The last column #E denotes the number of parameters in encoders."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.118,
+                0.429,
+                0.486,
+                0.527
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Graph Encoder</td><td>Compositional Structure</td><td>Dynamic Fusion</td><td>Acc(%)</td></tr><tr><td>GTS</td><td>X</td><td>X</td><td>X</td><td>75.6</td></tr><tr><td>Graph2Tree</td><td>✓</td><td>X</td><td>X</td><td>77.4</td></tr><tr><td></td><td>✓</td><td>✓</td><td>X</td><td>78.1</td></tr><tr><td></td><td>✓</td><td>X</td><td>✓</td><td>78.9</td></tr><tr><td>C-MWP</td><td>✓</td><td>✓</td><td>✓</td><td>80.3</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.126,
+                0.537,
+                0.474,
+                0.552
+            ],
+            "angle": 0,
+            "content": "Table 2: Accuracy among different ablated models."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.58,
+                0.49,
+                0.629
+            ],
+            "angle": 0,
+            "content": "PLM-based semantic modal encoder and graph-based quantity modal encoder, which will be an interesting inspiration to the community."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.641,
+                0.441,
+                0.657
+            ],
+            "angle": 0,
+            "content": "Ablative Study of Different Components."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.662,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In order to evaluate the effectiveness of each component in \\( C \\)-MWP, we report the model performance after removing several components. Compared with Graph2Tree, our compositional structure and dynamic fusion module allow the full usage of both modals and excel in improving performance. \"Compositional Structure\" denotes separating the semantic and quantity modal with two different encoders that run in parallel. \"w Compositional Structure and w/o Dynamic Fusion\" means we replace the Dynamic Fusion module with a simple addition of two features from two encoders. \"w/o Compositional Structure and w Dynamic Fusion\" means, we employ the structure of Graph2Tree, then utilize our dynamic fusion module to fuse the feature from the RNN Encoder and"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.533,
+                0.082,
+                0.861,
+                0.182
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Model</td><td>Overall</td><td>Story</td><td>Algebra</td><td>KNWL</td></tr><tr><td>GTS</td><td>75.4</td><td>75.1</td><td>82.8</td><td>64.3</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>76.3</td><td>89.7</td><td>57.1</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>77.8</td><td>88.8</td><td>61.9</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>80.0</td><td>90.0</td><td>66.7</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>85.6</td><td>88.8</td><td>72.0</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>87.5</td><td>90.7</td><td>72.0</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.192,
+                0.884,
+                0.234
+            ],
+            "angle": 0,
+            "content": "Table 3: The answer accuracy \\((\\%)\\) of problems in different types. KNWL stands for the external knowledge required problems."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.258,
+                0.884,
+                0.273
+            ],
+            "angle": 0,
+            "content": "the feature from the GNN Encoder of Graph2Tree."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.282,
+                0.84,
+                0.298
+            ],
+            "angle": 0,
+            "content": "Performance on Different Types of MWP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.301,
+                0.885,
+                0.623
+            ],
+            "angle": 0,
+            "content": "In order to investigate how our model performs across various types of MWP, we introduce a new split of Math23k with regard to three types of problems: story problems, algebra problems and knowledge problems. Split details are shown in the appendix. The evaluation results are presented in Table 3. Without a compositional manner, Graph2Tree and Multi-E/D perform better than GTS on story and algebra testing problems, whereas they perform worse on knowledge problems. As stated before, one joint modal cannot do it all. These baselines work well on some types of problems while having weak performance on other types of problems. Our C-MWP offers a general accuracy improvement, which firmly supports our motivation for alleviating the generalization issue. This provides clear evidence that our model leverages general math knowledge across different types of MWP, successfully solving some nontrivial problems that Graph2Tree failed to solve."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.634,
+                0.797,
+                0.649
+            ],
+            "angle": 0,
+            "content": "5 Conclusion and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.659,
+                0.885,
+                0.837
+            ],
+            "angle": 0,
+            "content": "The semantic meaning and quantity information are important intrinsic properties of a math word problem. Aiming at dealing with uni-modal bias and achieve better generalization, we make the first attempt to propose a compositional MWP solver, \\( C \\)-MWP. Multi-layer reasoning and specified training methods are leveraged to enhance the generalizability of the model. As the method could be applied in a broader range of neuro-symbolic learning problems, we will keep exploring the adaptiveness of this compositional encoding method."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.847,
+                0.673,
+                0.863
+            ],
+            "angle": 0,
+            "content": "Acknowledgement"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "The research work is partially supported by the Internal Asia Research Collaboration Grant, University of Notre Dame."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10012"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.085,
+                0.212,
+                0.099
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.106,
+                0.49,
+                0.236
+            ],
+            "angle": 0,
+            "content": "Explainability Most current MWP solvers are only able to generate solutions. In our work, although we achieved better generalization ability, it is still hard to explain how the model solves MwPs both correctly or incorrectly. These automated solvers would be much more helpful for tutoring students if they could explain their equation solutions by generating reasoning steps."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.261,
+                0.214,
+                0.276
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.284,
+                0.49,
+                0.351
+            ],
+            "angle": 0,
+            "content": "Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. In NAACL, pages 2357-2367."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.36,
+                0.489,
+                0.413
+            ],
+            "angle": 0,
+            "content": "Xinyun Chen, Chen Liang, Adams Wei Yu, Dawn Song, and Denny Zhou. 2020. Compositional generalization via neural-symbolic stack machines. NeurIPS, 33:1690-1701."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.423,
+                0.489,
+                0.501
+            ],
+            "angle": 0,
+            "content": "Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, pages 1724-1734."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.512,
+                0.489,
+                0.566
+            ],
+            "angle": 0,
+            "content": "Shifeng Huang, Jiawei Wang, Jiao Xu, Da Cao, and Ming Yang. 2021. Recall and learn: A memory-augmented solver for math word problems. In Findings of EMNLP, pages 786-796."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.575,
+                0.489,
+                0.627
+            ],
+            "angle": 0,
+            "content": "Zhanming Jie, Jierui Li, and Wei Lu. 2022. Learning to reason deductively: Math word problem solving as complex relation extraction. In ACL, pages 5944-5955."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.638,
+                0.489,
+                0.69
+            ],
+            "angle": 0,
+            "content": "Rik Koncel-Kedziorski, Hannaneh Hajishirzi, Ashish Sabharwal, Oren Etzioni, and Siena Dumas Ang. 2015. Parsing algebraic word problems into equations. ACL, 3:585-597."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.701,
+                0.489,
+                0.753
+            ],
+            "angle": 0,
+            "content": "Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In NAACL, pages 1152-1157."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.763,
+                0.489,
+                0.843
+            ],
+            "angle": 0,
+            "content": "Yihuai Lan, Lei Wang, Qiyuan Zhang, Yunshi Lan, Bing Tian Dai, Yan Wang, Dongxiang Zhang, and Ee-Peng Lim. 2022. Mwptoolkit: An open-source framework for deep learning-based math word problem solvers. In AAAI, volume 36, pages 13188-13190."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.852,
+                0.489,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In ACL, pages 6162-6167."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.284,
+                0.49,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.153
+            ],
+            "angle": 0,
+            "content": "Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2021. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. arXiv preprint arXiv:2110.08464."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.163,
+                0.885,
+                0.23
+            ],
+            "angle": 0,
+            "content": "Zhenwen Liang, Jipeng Zhang, Lei Wang, Wei Qin, Yunshi Lan, Jie Shao, and Xiangliang Zhang. 2022. MWP-BERT: Numeracy-augmented pre-training for math word problem solving. In *Findings of NAACL*, pages 997-1009."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.24,
+                0.883,
+                0.281
+            ],
+            "angle": 0,
+            "content": "Zhenwen Liang and Xiangliang Zhang. 2021. Solving math word problems with teacher supervision. In *IJCAI*, pages 3522-3528."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.291,
+                0.885,
+                0.356
+            ],
+            "angle": 0,
+            "content": "Xin Lin, Zhenya Huang, Hongke Zhao, Enhong Chen, Qi Liu, Hao Wang, and Shijin Wang. 2021. Hms: A hierarchical solver with dependency-enhanced understanding for math word problem. In AAAI, pages 4232-4240."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.368,
+                0.885,
+                0.42
+            ],
+            "angle": 0,
+            "content": "Qianying Liu, Wenyv Guan, Sujian Li, and Daisuke Kawahara. 2019. Tree-structured decoding for solving math word problems. In EMNLP, pages 2370-2379."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.431,
+                0.885,
+                0.485
+            ],
+            "angle": 0,
+            "content": "Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2020. A diverse corpus for evaluating and developing english math word problem solvers. In ACL, pages 975-984."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.495,
+                0.885,
+                0.549
+            ],
+            "angle": 0,
+            "content": "Jianhao Shen, Yichun Yin, Lin Li, Lifeng Shang, Xin Jiang, Ming Zhang, and Qun Liu. 2021. Generate & rank: A multi-task framework for math word problems. In EMNLP, pages 2269-2279."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.559,
+                0.885,
+                0.6
+            ],
+            "angle": 0,
+            "content": "Yibin Shen and Cheqing Jin. 2020. Solving math word problems with multi-encoders and multi-decoders. In COLING, pages 2924-2934."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.61,
+                0.885,
+                0.651
+            ],
+            "angle": 0,
+            "content": "Yan Wang, Xiaojiang Liu, and Shuming Shi. 2017. Deep neural solver for math word problems. In EMNLP, pages 845-854."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.661,
+                0.885,
+                0.714
+            ],
+            "angle": 0,
+            "content": "Charles Welch, Jonathan K Kummerfeld, Verónica Pérez-Rosas, and Rada Mihalcea. 2020. Compositional demographic word embeddings. In EMNLP, pages 4076-4089."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.724,
+                0.885,
+                0.804
+            ],
+            "angle": 0,
+            "content": "Francis CK Wong and William SY Wang. 2007. Generalisation towards combinatorial productivity in language acquisition by simple recurrent networks. In 2007 International Conference on Integration of Knowledge Intensive Multi-Agent Systems, pages 139-144. IEEE."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.814,
+                0.885,
+                0.869
+            ],
+            "angle": 0,
+            "content": "Qinzhuo Wu, Qi Zhang, and Zhongyu Wei. 2021a. An edge-enhanced hierarchical graph-to-tree network for math word problem solving. In *Findings of EMNLP*, pages 1473–1482."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.878,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuanjing Huang. 2021b. Math word problem solving with explicit numerical values. In ACL, pages 5859-5869."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.511,
+                0.086,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.525,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10013"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.086,
+                0.49,
+                0.126
+            ],
+            "angle": 0,
+            "content": "Zhipeng Xie and Shichao Sun. 2019. A goal-driven tree-structured neural model for math word problems. In *IJCAI*, pages 5299–5305."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.136,
+                0.49,
+                0.203
+            ],
+            "angle": 0,
+            "content": "Jingfeng Yang, Le Zhang, and Diyi Yang. 2022. SUBS: Subtree substitution for compositional semantic parsing. In *NAACN*, pages 169–174, Seattle, United States. Association for Computational Linguistics."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.212,
+                0.489,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Weijiang Yu, Yingpeng Wen, Fudan Zheng, and Nong Xiao. 2021. Improving math word problems with pre-trained knowledge and hierarchical reasoning. In EMNLP, pages 3384-3394."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.115,
+                0.276,
+                0.49,
+                0.329
+            ],
+            "angle": 0,
+            "content": "Jipeng Zhang, Lei Wang, Roy Ka-Wei Lee, Yi Bin, Yan Wang, Jie Shao, and Ee-Peng Lim. 2020. Graph-to-tree learning for solving math word problems. In ACL, pages 3928-3937."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.342,
+                0.204,
+                0.359
+            ],
+            "angle": 0,
+            "content": "Appendix"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.368,
+                0.304,
+                0.384
+            ],
+            "angle": 0,
+            "content": "Implementation Details"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.388,
+                0.49,
+                0.71
+            ],
+            "angle": 0,
+            "content": "We train our model on an NVIDIA RTX 2080Ti GPU, all implementation of training and testing is coded in Python with Pytorch framework. For our RNN-based model, 2 CMEncoders are stacked and only 1 CMEncoder is used in the PLM-based model. The size of hidden dimensions in encoders and decoders are set to 512 and 768 for RNN-solver and PLM-solver, respectively. Each GCN block has 2 GraphConv layers and each GNN encoder has 4 heads of GCN blocks. During training, Adam optimizer is applied with the initial learning rates of 0.001/0.00003 for RNN/PLM, which would be halved every 30 epochs. During testing, we use a 5-beam search to get reasonable solutions. We also apply Gaussian noise with mean 0 and variance 1 on the embedding result during training. This simple operation can help models to learn more robust parameters. Through grid search at 0.1 level, the noise is multiplied by 0.2 to achieve the best performance."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.722,
+                0.32,
+                0.738
+            ],
+            "angle": 0,
+            "content": "Hyper-Parameter Tuning"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "In general, we apply grid-search with manually designed search space and use answer accuracy as the evaluation metric to select the hyper-parameters. For the number of stacked encoders, the search space is \\(\\{1,2,3,4\\}\\) and we finally use 2. For the weight of \\(L_{2}\\) normalization loss, we choose weight 1 from \\(\\{0.01,0.1,1,5,10\\}\\). The weight of random noise 0.2 is selected from 0.1 level by grid search with range 0 to 1. We also tune the beam size of beam search from \\(\\{3,4,5,6,7\\}\\) and choose 5. The dropout probability 0.5 is selected from"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.532,
+                0.082,
+                0.863,
+                0.145
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>K = 1</td><td>K = 2</td><td>K = 3</td><td>K = 4</td></tr><tr><td>C-MWP (RNN)</td><td>79.1</td><td>80.3</td><td>78.9</td><td>78.1</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>85.9</td><td>85.2</td><td>85.1</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.509,
+                0.154,
+                0.883,
+                0.182
+            ],
+            "angle": 0,
+            "content": "Table 4: Accuracies across different numbers of stacked CMEncoder on Math23k."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.209,
+                0.884,
+                0.273
+            ],
+            "angle": 0,
+            "content": "\\(\\{0.1, 0.3, 0.5, 0.7\\}\\). Initial learning rate 0.001 is selected from \\(\\{0.01, 0.001, 0.0001\\}\\). For the hidden size and embedding size in encoder, we select 256 from \\(\\{64, 128, 256, 512\\}\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.285,
+                0.808,
+                0.3
+            ],
+            "angle": 0,
+            "content": "Variance and Significance Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.306,
+                0.884,
+                0.419
+            ],
+            "angle": 0,
+            "content": "We evaluated our solver with 5-fold cross-validation and found that the accuracy of our RNN-based C-MWP \\((0.779 \\pm 0.028)\\) is significantly higher than Graph2Tree \\((0.755 \\pm 0.016)\\) \\((p < 0.01)\\), and the accuracy of our PLM-based C-MWP \\((0.845 \\pm 0.21)\\) is significantly higher than vanilla MWP-BERT \\((0.824 \\pm 0.016)\\) \\((p < 0.01)\\)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.431,
+                0.878,
+                0.447
+            ],
+            "angle": 0,
+            "content": "Sensitivity Analysis about Stacking Number K"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.452,
+                0.884,
+                0.676
+            ],
+            "angle": 0,
+            "content": "As we mentioned in Section 2.2, our CMEncoder can be stacked into multiple layers to improve the representation of an MWP. To obtain a better understanding about the hyperparameter K, i.e., the number of stacked CMEncoders, we conduct a sensitivity analysis in Table 4. We can see that the best K for RNN-encoder is 2, just like humans often need to make multiple glimpses to refine an MWP solution. In the meantime, the best K for PLM-based encoder is 1. The potential reason is that one pre-trained language model (PLM) already has an outstanding ability to encode texts. It is thus not necessary to apply another CMEncoder to refine the encoded features."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.689,
+                0.604,
+                0.705
+            ],
+            "angle": 0,
+            "content": "Case Study"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.71,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Figure 3 shows generated solutions of two solved problems by GTS (Xie and Sun, 2019), Graph2Tree (Zhang et al., 2020), Multi-E/D (Shen and Jin, 2020) and our proposed C-MWP (RNN-Based). The first problem has 4 quantities and they are all useful, which means that it requires sufficient problem understanding and mathematical reasoning to generate the right answer. Both Graph2Tree and Multi-E/D which directly connect semantics modal and quantity modal fail to extract clear representations of the problem, finally resulting in unreasonable solutions which only contain 3 quantities. For the second problem, although Graph2Tree and"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10014"
+        }
+    ],
+    [
+        {
+            "type": "image",
+            "bbox": [
+                0.135,
+                0.081,
+                0.861,
+                0.293
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.375,
+                0.301,
+                0.622,
+                0.317
+            ],
+            "angle": 0,
+            "content": "Figure 3: Case study from Math23k"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.139,
+                0.339,
+                0.465,
+                0.42
+            ],
+            "angle": 0,
+            "content": "<table><tr><td></td><td>Overall</td><td>Story</td><td>Algebra</td><td>Knowledge</td></tr><tr><td>train</td><td>21,162</td><td>17,546</td><td>2,595</td><td>957</td></tr><tr><td>val</td><td>1,000</td><td>817</td><td>133</td><td>50</td></tr><tr><td>test</td><td>1,000</td><td>842</td><td>116</td><td>42</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.114,
+                0.43,
+                0.489,
+                0.459
+            ],
+            "angle": 0,
+            "content": "Table 5: Statistics of different types of problems in Math23k."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.485,
+                0.49,
+                0.595
+            ],
+            "angle": 0,
+            "content": "Multi-E/D utilize all 3 quantities in the problem description, they still fail to generate a plausible solution. These two cases show that our proposed encoder is able to extract more comprehensive representations from problem descriptions, eventually guiding the decoder to generate the correct solutions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.608,
+                0.357,
+                0.623
+            ],
+            "angle": 0,
+            "content": "MWPs in Different Categories"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.629,
+                0.49,
+                0.868
+            ],
+            "angle": 0,
+            "content": "Figure 1 shows the MWP examples of \"Story Problem\", \"Algebra Problem\" and \"Knowledge Problem\". \"Story Problem\" often includes a significant amount of background information like characters, objectives and behaviors. \"Algebra Problems\" involves math notations or is composed of elementary concepts. \"Knowledge Problem\" asks for external knowledge like geometry and number sequence. The category of each problem is determined based on keywords. The keywords of \"Story\" and \"Knowledge\" problems are selected from the appendix of (Liu et al., 2019). Inspired by them, we categorize Math23k into 3 subsets - story, algebra, and knowledge. The statistics of these problems are shown in Table 5."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10015"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.134,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.128,
+                0.533,
+                0.144
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.145,
+                0.233,
+                0.16
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.171,
+                0.553,
+                0.187
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.188,
+                0.233,
+                0.202
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.214,
+                0.696,
+                0.23
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.231,
+                0.233,
+                0.246
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.256,
+                0.669,
+                0.273
+            ],
+            "angle": 0,
+            "content": "□ A4. Have you used AI writing assistants when working on this paper?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.274,
+                0.233,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.301,
+                0.492,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.322,
+                0.216,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.348,
+                0.531,
+                0.364
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.365,
+                0.233,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.391,
+                0.779,
+                0.407
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.408,
+                0.233,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.434,
+                0.882,
+                0.497
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.499,
+                0.233,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.525,
+                0.882,
+                0.573
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.574,
+                0.233,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.6,
+                0.882,
+                0.632
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.633,
+                0.233,
+                0.648
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.659,
+                0.882,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.741,
+                0.233,
+                0.755
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.766,
+                0.496,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.787,
+                0.216,
+                0.802
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.128,
+                0.814,
+                0.882,
+                0.846
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.847,
+                0.233,
+                0.861
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.869,
+                0.878,
+                0.893
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10016"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.085,
+                0.88,
+                0.133
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.144,
+                0.881,
+                0.207
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.219,
+                0.881,
+                0.283
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.085,
+                0.881,
+                0.283
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.294,
+                0.878,
+                0.31
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.315,
+                0.215,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.341,
+                0.881,
+                0.388
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.4,
+                0.881,
+                0.464
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.475,
+                0.881,
+                0.54
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.55,
+                0.873,
+                0.582
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.593,
+                0.881,
+                0.641
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.129,
+                0.341,
+                0.881,
+                0.641
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "10017"
+        }
+    ]
+]

2023/Compositional Mathematical Encoding for Math Word Problems/801bbb7b-d8b8-4c19-87bf-9a90de8c668a_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8ada1e67c7ff7825f159a96995c863d4fe005ab9da7f9188f54a322f9c8dbe9
+size 465304

2023/Compositional Mathematical Encoding for Math Word Problems/full.md

@@ -0,0 +1,329 @@
+# Compositional Mathematical Encoding for Math Word Problems
+
+Zhenwen Liang $^{1}$ , Jipeng Zhang $^{2}$ , Kehan Guo $^{1}$ , Xiaodong Wu $^{3}$ , Jie Shao $^{4}$ , and Xiangliang Zhang $^{\boxtimes 1}$
+
+<sup>1</sup>University of Notre Dame, {zliang6, kguo2, xzhang33}@nd.edu  
+<sup>2</sup>Hong Kong University of Science and Technology, jzhanggr@connect.ust.hk  
+<sup>3</sup>Queen's University, 22hd17@queensu.ca  
+<sup>4</sup>University of Electronic Science and Technology of China, shaojie@uestc.edu.cn
+
+# Abstract
+
+Solving math word problem (MWP) remains a challenging task, as it requires to understand both the semantic meanings of the text and the mathematical logic among quantities, i.e., for both semantics modal and quantity modal learning. Current MWP encoders work in a uni-modal setting and map the given problem description to a latent representation, then for decoding. The generalizability of these MWP encoders is thus limited because some problems are semantics-demanding and others are quantity-demanding. To address this problem, we propose a Compositional Math Word Problem Solver (C-MWP) which works in a bimodal setting encoding in an interactive way. Extensive experiments validate the effectiveness of C-MWP and show its superiority over state-of-the-art models on public benchmarks.
+
+# 1 Introduction
+
+The task of math word problem (MWP) solving aims to map natural language problem descriptions into executable solution equations to get the correct answer, which is a sub-area of neuro-symbolic reasoning. It requires perceptual abilities such as comprehending the question, identifying the quantities and corresponding attributes, as well as complex semantics understanding skills like performing logical inference, making comparisons and leveraging external mathematical knowledge.
+
+While MWP encoders have been sophisticatedly designed to understand the natural language problem description, the difference on understanding diverse types of problems has not been aware of. We find that MWP can generally be grouped into three categories based on the keywords in (Liu et al., 2019), i.e., "Story Problem", "Algebra Problem" and "Knowledge Problem". "Story Problem" often includes significant amount of background information like characters, objectives and behaviors.
+
+"Algebra Problems" involves math notations or is composed of elementary concepts. "Knowledge Problem" asks for external knowledge like geometry and number sequence, as shown in Figure 1.
+
+These types of problems can be compositionally understood at the different level attention to the semantics modal and quantity modal, where RNNs and pre-trained language models usually focus on the textual information and GCN with quantity-centered graphs capture the relationship between quantities and contexts. However, the encoders in existing MWP solvers either model only the semantics modality or utilize quantity modal priors to refine the MWP encoding (Zhang et al., 2020; Shen and Jin, 2020). Although the quantity centered refinement can particularly make improvements on quantity-demanding problems, its semantics understanding is weakened (evidence can be found in Table 3). This limitation, one joint modal cannot do it all, decreases the generalization of MWP solvers and is what compositional learning aims to address. In this work, we propose to disentangle semantics modal and quantity modal by compositional learning at the encoding stage, aiming to improve the generalization across different types of problems.
+
+Contributions. (i) A novel and effective bimodal approach is proposed for the first time to enable MWP compositional understanding. (ii) A joint reasoning module is designed for our bimodal architectures to flexibly incorporate different modalities. (iii) Extensive experiments and ablative studies on two large-scale MWP benchmarks – Math23k (Wang et al., 2017) and MAwPS (Koncel-Kedziorski et al., 2016) show the superiority of the proposed approach over related works.
+
+![](images/762bc4966b60b6fa4bc9b5cf248bd7d470551fe672405852f5b0a3dfcc1095b7.jpg)  
+Figure 1: Examples of different types of problems in MWP solving.
+
+# 2 Related Work
+
+Compositional Learning in NLP. Modeling compositionality in language has been a longstanding issue (Wong and Wang, 2007) in NLP community. One common practice is to perform disentanglement over language representations at different levels (Welch et al., 2020). They usually focus on atomic semantics units like character, word and phrase. As logic form annotations naturally own compositional features, compositionality is incorporated in generating correct logic contents. Therefore, the compositionality is often injected into traditional semantic parsing tasks (Chen et al., 2020; Yang et al., 2022) where the goals during training can be decomposed and then reorganized as a novel goal.
+
+Our work firstly tries to inject compositional prior into MWP encoding. It is worth noting that MWP solving owns the same well-organized logic form annotations as machine reasoning, which naturally requires compositionality.
+
+Math Word Problem Solving. Earlier MWP solvers parse problem descriptions semantically, and learn templates for generating answers (Koncel-Kedziorski et al., 2015). Recent works (Wang et al., 2017; Xie and Sun, 2019; Li et al., 2019; Zhang et al., 2020; Shen and Jin, 2020; Wu et al., 2021b,a; Lin et al., 2021; Liang and Zhang, 2021; Jie et al., 2022) focus on employing the encoder-decoder framework (e.g., sequence-to-sequence,
+
+sequence-to-tree, graph-to-tree) to translate MWP texts into equations based on traditional RNN structure. There are also new settings (Amini et al., 2019; Miao et al., 2020) introduced to extend MWP solving in equation group generation and diagnosing awareness of external knowledge. Nowadays, many researchers build strong MWP solvers upon pre-trained language models (PLMs) (Huang et al., 2021; Li et al., 2021; Yu et al., 2021; Shen et al., 2021; Lan et al., 2022) and have achieved great performance. Differently, our work lays the groundwork of feature extraction of quantity modal, which is orthogonal to those works.
+
+In this work, we not only propose an explicit compositional encoding module with a multi-layer design, but also incorporate detailed analysis to verify its compositional learning ability, to jointly leverage semantic and quantity information to achieve effective MWP understanding.
+
+# 3 Our approach
+
+# 3.1 Compositional Mathematical Encoder
+
+As shown in Figure 2, our CMEncoder block consists of a semantic encoder, a quantity encoder and a dynamic fusion block. The semantic encoder aims to extract semantic information from the problem description, understanding the background and objectives. The latter part encodes problems only with quantity-related graphs, helping the encoder to know the properties of quantities and the relationship between quantities and contexts.
+
+![](images/2cee7b44492be75c82ea2b3a2de99f53cf4f1f239381d0c0c68bc14878e23248.jpg)  
+Figure 2: The CMEncoder block (shown at the bottom) takes the given problem description, and runs in parallel to obtain  $H_{r}$  from the semantic encoder, and  $H_{g}$  from the quantity encoder. A dynamic fusion module incorporates  $H_{r}$  and  $H_{g}$  by cross-modal attention. The obtained  $H_{f}$  is attentively stacked with  $H_{r}$  and  $H_{g}$ . The resulted  $H_{att}$  is sent to the next CMEncoder block. The final problem representation  $H_{final}$  goes to the decoder for generating the final solution equation.
+
+Semantic Encoder. To demonstrate the robustness of our approach, we implemented two different semantic encoders as our backbone.
+
+Firstly, we encode the problem description  $W$  by a bidirectional gated recurrent unit (BiGRU) (Cho et al., 2014). The outputs of GRU are hidden state vectors of all tokens,  $H_{r} = \{h_{1}, h_{2}, \dots, h_{n}\}$ , where  $n$  is the length of problem  $W$ .
+
+$$
+H _ {r} = \operatorname {B i G R U} \left(\operatorname {E m b e d} _ {s} (W)\right) \tag {1}
+$$
+
+where  $Embed_{s}(W)$  is the embedding result of textual description  $W$  in semantics modal. Empirically, we find that two stacked CMEncoders as shown in Figure 2 achieve the best performance.
+
+Secondly, pre-trained language models (PLMs) have been ubiquitous in NLP tasks. We use the latest push of MWP-BERT (Liang et al., 2022) as our semantic encoder to obtain  $H_{r}$ .
+
+Quantity Encoder. To encode the quantity modal in the problem  $W$ , we feed a graph transformer  $G_{trans}$  with Quantity Comparison Graph and Quantity Cell Graph following Graph2Tree (Zhang et al., 2020),
+
+$$
+H _ {g} = G _ {\text {t r a n s}} \left(\operatorname {E m b e d} _ {q} (W)\right) \tag {2}
+$$
+
+where  $\operatorname{Embed}_q(W)$  is the embedding matrix in the quantity modal, which aims to improve the
+
+quantity representation by incorporating quantity magnitude information and quantity-context relationship with the above two graphs. Different from Graph2Tree, the two embeddings  $Embed_{s}(W)$  and  $Embed_{q}(W)$  are updated in the training process to extract the semantics and quantity feature separately. In this way, semantics and quantity modals are disentangled, which alleviates the issue of "one joint modal cannot do it all", enabling the C-MWP solver to pay different levels of attention when solving different problems.
+
+Dynamic Fusion. To achieve joint reasoning over the semantics information and quantity information, we design a dynamic fusion module to flexibly incorporate the features from these two modals. First, we get  $s$  and  $q$  from the mean pooling of  $H_{r}$  and  $H_{g}$ , respectively. Then, cross-modal attention is applied between  $H_{r}$  and  $q$ ,  $H_{g}$  and  $s$ :
+
+$$
+\begin{array}{l} A t t _ {1} \left(H _ {r}, q\right) = \Sigma_ {i = 1} ^ {n} a _ {i} H _ {r i} \\ A t t _ {1} \left(H _ {r}\right) = \Sigma_ {i = 1} ^ {n} H _ {r i} \end{array} \tag {3}
+$$
+
+$$
+A t t _ {2} (H _ {g}, s) = \Sigma_ {i = 1} ^ {n} b _ {i} H _ {g i}
+$$
+
+where the attention scores  $a_{i}, b_{i}$  come from:
+
+$$
+a _ {i} = W _ {a} ^ {1} \tanh  \left(W _ {a} ^ {2} \left(H _ {r i} \mid \mid q\right)\right) \tag {4}
+$$
+
+$$
+b _ {i} = W _ {b} ^ {1} \tanh  (W _ {b} ^ {2} \left(H _ {g i} \mid | s\right))
+$$
+
+where  $W_{a}^{1}, W_{a}^{2}, W_{b}^{1}$  and  $W_{b}^{2}$  are parameter matrices. The cross-modal attention here grounds the
+
+quantity information in the semantics modal, and vice versa. By applying different weights on different modals, our model is flexible to pay more or less attention on a certain modal. Finally, the output of dynamic fusion is:
+
+$$
+H _ {f} = \operatorname {A t t} _ {1} \left(H _ {r}, q\right) \mid \mid \operatorname {A t t} _ {2} \left(H _ {g}, s\right). \tag {5}
+$$
+
+# 3.2 Stack Multiple CMEncoders
+
+Humans often need to make multiple glimpses to refine an MWP solution. Similarly, a CMEncoder can be stacked in multiple steps to refine the understanding of an MWP, as shown in Figure 2. Given the output from the semantic encoder, quantity encoder and dynamic fusion module at layer  $k - 1$ , the features are stacked as:
+
+$$
+H _ {a t t} ^ {(k - 1)} = c _ {r} H _ {r} ^ {(k - 1)} + c _ {g} H _ {g} ^ {(k - 1)} \tag {6}
+$$
+
+where the attention weights  $c_{r}$  and  $c_{g}$  are:
+
+$$
+\begin{array}{l} c _ {r} = W _ {r} ^ {1} \tanh  \left(W _ {r} ^ {2} \left(H _ {r} ^ {(k - 1)} \mid \mid H _ {f} ^ {(k - 1)}\right)\right) \tag {7} \\ c _ {g} = W _ {g} ^ {1} \tanh  (W _ {g} ^ {2} \left(H _ {g} ^ {(k - 1)} \mid \mid H _ {f} ^ {(k - 1)}\right)) \\ \end{array}
+$$
+
+where  $W_{r}^{1}, W_{r}^{2}, W_{g}^{1}$  and  $W_{g}^{2}$  are parameter matrices. The  $H_{att}^{(k - 1)}$  will be the input for both the semantic modal and quantity modal of K-th CMEncoder, which will output  $H_{r}^{(k)}, H_{g}^{(k)}$  and  $H_{f}^{(k)}$ , which can be sent for the update at layer  $k + 1$ .
+
+After finishing the  $K$ -th step reasoning, we concatenate the final  $H_r^{(K)}$  and  $H_g^{(K)}$  as the final output representation  $H_{final}$ .
+
+# 3.3 Decoder
+
+We follow the same implementation as GTS (Xie and Sun, 2019). Eventually, the decoder will output the pre-order traversal sequence of the solution tree.
+
+# 3.4 Training Method
+
+Given the training samples with problem description  $W$  and the corresponding solution  $S$ , the main training objective is to minimize the negative log probability for predicting  $S$  from  $W$ , empowered by the compositionality of the CMEncoders. Therefore, the overall loss is:
+
+$$
+L = L _ {M W P} + \left| \left| E m b e d _ {s} \right| \right| _ {2} + \left| \left| E m b e d _ {q} \right| \right| _ {2} (8)
+$$
+
+where  $L_{MWP}$  is the negative log prediction probability  $-\log p(S \mid W)$ . The  $L_2$  norm of the encoder embedding matrices is added to the loss function as regularization terms.
+
+# 4 Experiments
+
+# 4.1 Datasets
+
+Math23k (Wang et al., 2017) containing 23,162 Chinese MWPs is collected from several educational websites.
+
+MAWPS (Koncel-Kedziorski et al., 2015) is an MWP dataset owning 2,373 English MwPs.
+
+# 4.2 Baselines
+
+GTS (Xie and Sun, 2019) proposes a powerful tree-based decoder. Graph2Tree (Zhang et al., 2020) constructs graphs to extract useful relationships in an MWP. NumS2T (Wu et al., 2021b) encode quantities with explicit numerical values. MultiE/D (Shen and Jin, 2020) proposes to use multiple decoders in MWP solving. HMS (Lin et al., 2021) develops a hierarchical word-clause-problem encoder. EEH-G2T (Wu et al., 2021a) aims to capture the long-range word relationship by graph network. REAL (Huang et al., 2021) proposes a analogical auxiliary learning strategy by extracting similar MWPs. BERT-CL (Li et al., 2021) uses contrastive learning with PLMs. RPKHS (Yu et al., 2021) performs hierarchical reasoning with PLMs. MWP-BERT released a BERT-based encoder that is continually pre-trained on MWP corpus. Gen&Rank (Shen et al., 2021) designs a multi-task learning framework with encoder-decoder pre-training. MWP toolkit (Lan et al., 2022) finds a RoBERTa-to-RoBERTa model has the best performance in MWP solving.
+
+# 4.3 Experimental Results
+
+As Table 1 shows, our approach outperforms all other RNN-based baselines in terms of answer accuracy. On Math23k, we outperform the latest RNN-based push from Wu et al. (2021a) by  $1.8\%$ . For the first time, an RNN-based MWP solver reaches over  $80\%$  answer accuracy on the Math23k dataset. What is more, the even fewer parameters with the best performance suggest that our model is also memory-efficient.
+
+PLM-based solvers benefit from the pre-training on a huge amount of corpus and thus achieve great semantic understanding ability. From a different point of view, our work aims to effectively and efficiently integrate semantic and quantity understanding. Therefore, by incorporating the MWP-BERT model as our semantic extractor, the answer accuracy of C-MWP achieves state-of-the-art performance. It proves the feasibility of combining
+
+<table><tr><td></td><td>Math23k</td><td>Math23k*</td><td>MAWPS</td><td>#E</td></tr><tr><td colspan="5">RNN Based</td></tr><tr><td>DNS</td><td>-</td><td>58.1</td><td>59.5</td><td>3.0M</td></tr><tr><td>GTS</td><td>75.6</td><td>74.3</td><td>82.6</td><td>7.2M</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>75.5</td><td>83.7</td><td>9.0M</td></tr><tr><td>NUMS2T</td><td>78.1</td><td>-</td><td>-</td><td>7.9M</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>76.9</td><td>-</td><td>14.2M</td></tr><tr><td>HMS</td><td>78.4</td><td>-</td><td>80.3</td><td>9.5M</td></tr><tr><td>EEH-G2T</td><td>78.5</td><td>-</td><td>84.8</td><td>9.9M</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>77.9</td><td>84.9</td><td>7.6M</td></tr><tr><td colspan="5">PLM Based</td></tr><tr><td>REAL</td><td>82.3</td><td>80.0</td><td>-</td><td>110M</td></tr><tr><td>BERT-CL</td><td>83.2</td><td>-</td><td>-</td><td>102M</td></tr><tr><td>RPKHS</td><td>83.9</td><td>82.2</td><td>-</td><td>102M</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>82.4</td><td>-</td><td>110M</td></tr><tr><td>Gen&amp;Rank</td><td>85.4</td><td>84.3</td><td>-</td><td>610M</td></tr><tr><td>MWP toolkit</td><td>-</td><td>76.9</td><td>88.4</td><td>110M</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>84.5</td><td>89.1</td><td>130M</td></tr></table>
+
+Table 1: Math23k column shows the results when evaluating on the public test set of Math23k, while the  $\mathrm{Math23k^{*}}$  column shows the result of 5-fold cross validation on Math23k dataset. The last column #E denotes the number of parameters in encoders.  
+
+<table><tr><td></td><td>Graph Encoder</td><td>Compositional Structure</td><td>Dynamic Fusion</td><td>Acc(%)</td></tr><tr><td>GTS</td><td>X</td><td>X</td><td>X</td><td>75.6</td></tr><tr><td>Graph2Tree</td><td>✓</td><td>X</td><td>X</td><td>77.4</td></tr><tr><td></td><td>✓</td><td>✓</td><td>X</td><td>78.1</td></tr><tr><td></td><td>✓</td><td>X</td><td>✓</td><td>78.9</td></tr><tr><td>C-MWP</td><td>✓</td><td>✓</td><td>✓</td><td>80.3</td></tr></table>
+
+PLM-based semantic modal encoder and graph-based quantity modal encoder, which will be an interesting inspiration to the community.
+
+# Ablative Study of Different Components.
+
+In order to evaluate the effectiveness of each component in  $C$ -MWP, we report the model performance after removing several components. Compared with Graph2Tree, our compositional structure and dynamic fusion module allow the full usage of both modals and excel in improving performance. "Compositional Structure" denotes separating the semantic and quantity modal with two different encoders that run in parallel. "w Compositional Structure and w/o Dynamic Fusion" means we replace the Dynamic Fusion module with a simple addition of two features from two encoders. "w/o Compositional Structure and w Dynamic Fusion" means, we employ the structure of Graph2Tree, then utilize our dynamic fusion module to fuse the feature from the RNN Encoder and
+
+Table 2: Accuracy among different ablated models.  
+
+<table><tr><td>Model</td><td>Overall</td><td>Story</td><td>Algebra</td><td>KNWL</td></tr><tr><td>GTS</td><td>75.4</td><td>75.1</td><td>82.8</td><td>64.3</td></tr><tr><td>Graph2Tree</td><td>77.4</td><td>76.3</td><td>89.7</td><td>57.1</td></tr><tr><td>Multi-E/D</td><td>78.4</td><td>77.8</td><td>88.8</td><td>61.9</td></tr><tr><td>C-MWP (RNN)</td><td>80.3</td><td>80.0</td><td>90.0</td><td>66.7</td></tr><tr><td>MWP-BERT</td><td>84.7</td><td>85.6</td><td>88.8</td><td>72.0</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>87.5</td><td>90.7</td><td>72.0</td></tr></table>
+
+Table 3: The answer accuracy  $(\%)$  of problems in different types. KNWL stands for the external knowledge required problems.
+
+the feature from the GNN Encoder of Graph2Tree.
+
+# Performance on Different Types of MWP.
+
+In order to investigate how our model performs across various types of MWP, we introduce a new split of Math23k with regard to three types of problems: story problems, algebra problems and knowledge problems. Split details are shown in the appendix. The evaluation results are presented in Table 3. Without a compositional manner, Graph2Tree and Multi-E/D perform better than GTS on story and algebra testing problems, whereas they perform worse on knowledge problems. As stated before, one joint modal cannot do it all. These baselines work well on some types of problems while having weak performance on other types of problems. Our C-MWP offers a general accuracy improvement, which firmly supports our motivation for alleviating the generalization issue. This provides clear evidence that our model leverages general math knowledge across different types of MWP, successfully solving some nontrivial problems that Graph2Tree failed to solve.
+
+# 5 Conclusion and Future Work
+
+The semantic meaning and quantity information are important intrinsic properties of a math word problem. Aiming at dealing with uni-modal bias and achieve better generalization, we make the first attempt to propose a compositional MWP solver,  $C$ -MWP. Multi-layer reasoning and specified training methods are leveraged to enhance the generalizability of the model. As the method could be applied in a broader range of neuro-symbolic learning problems, we will keep exploring the adaptiveness of this compositional encoding method.
+
+# Acknowledgement
+
+The research work is partially supported by the Internal Asia Research Collaboration Grant, University of Notre Dame.
+
+# Limitations
+
+Explainability Most current MWP solvers are only able to generate solutions. In our work, although we achieved better generalization ability, it is still hard to explain how the model solves MwPs both correctly or incorrectly. These automated solvers would be much more helpful for tutoring students if they could explain their equation solutions by generating reasoning steps.
+
+# References
+
+Aida Amini, Saadia Gabriel, Shanchuan Lin, Rik Koncel-Kedziorski, Yejin Choi, and Hannaneh Hajishirzi. 2019. Mathqa: Towards interpretable math word problem solving with operation-based formalisms. In NAACL, pages 2357-2367.  
+Xinyun Chen, Chen Liang, Adams Wei Yu, Dawn Song, and Denny Zhou. 2020. Compositional generalization via neural-symbolic stack machines. NeurIPS, 33:1690-1701.  
+Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP, pages 1724-1734.  
+Shifeng Huang, Jiawei Wang, Jiao Xu, Da Cao, and Ming Yang. 2021. Recall and learn: A memory-augmented solver for math word problems. In Findings of EMNLP, pages 786-796.  
+Zhanming Jie, Jierui Li, and Wei Lu. 2022. Learning to reason deductively: Math word problem solving as complex relation extraction. In ACL, pages 5944-5955.  
+Rik Koncel-Kedziorski, Hannaneh Hajishirzi, Ashish Sabharwal, Oren Etzioni, and Siena Dumas Ang. 2015. Parsing algebraic word problems into equations. ACL, 3:585-597.  
+Rik Koncel-Kedziorski, Subhro Roy, Aida Amini, Nate Kushman, and Hannaneh Hajishirzi. 2016. Mawps: A math word problem repository. In NAACL, pages 1152-1157.  
+Yihuai Lan, Lei Wang, Qiyuan Zhang, Yunshi Lan, Bing Tian Dai, Yan Wang, Dongxiang Zhang, and Ee-Peng Lim. 2022. Mwptoolkit: An open-source framework for deep learning-based math word problem solvers. In AAAI, volume 36, pages 13188-13190.  
+Jierui Li, Lei Wang, Jipeng Zhang, Yan Wang, Bing Tian Dai, and Dongxiang Zhang. 2019. Modeling intra-relation in math word problems with different functional multi-head attentions. In ACL, pages 6162-6167.
+
+Zhongli Li, Wenxuan Zhang, Chao Yan, Qingyu Zhou, Chao Li, Hongzhi Liu, and Yunbo Cao. 2021. Seeking patterns, not just memorizing procedures: Contrastive learning for solving math word problems. arXiv preprint arXiv:2110.08464.  
+Zhenwen Liang, Jipeng Zhang, Lei Wang, Wei Qin, Yunshi Lan, Jie Shao, and Xiangliang Zhang. 2022. MWP-BERT: Numeracy-augmented pre-training for math word problem solving. In *Findings of NAACL*, pages 997-1009.  
+Zhenwen Liang and Xiangliang Zhang. 2021. Solving math word problems with teacher supervision. In *IJCAI*, pages 3522-3528.  
+Xin Lin, Zhenya Huang, Hongke Zhao, Enhong Chen, Qi Liu, Hao Wang, and Shijin Wang. 2021. Hms: A hierarchical solver with dependency-enhanced understanding for math word problem. In AAAI, pages 4232-4240.  
+Qianying Liu, Wenyv Guan, Sujian Li, and Daisuke Kawahara. 2019. Tree-structured decoding for solving math word problems. In EMNLP, pages 2370-2379.  
+Shen-Yun Miao, Chao-Chun Liang, and Keh-Yih Su. 2020. A diverse corpus for evaluating and developing english math word problem solvers. In ACL, pages 975-984.  
+Jianhao Shen, Yichun Yin, Lin Li, Lifeng Shang, Xin Jiang, Ming Zhang, and Qun Liu. 2021. Generate & rank: A multi-task framework for math word problems. In EMNLP, pages 2269-2279.  
+Yibin Shen and Cheqing Jin. 2020. Solving math word problems with multi-encoders and multi-decoders. In COLING, pages 2924-2934.  
+Yan Wang, Xiaojiang Liu, and Shuming Shi. 2017. Deep neural solver for math word problems. In EMNLP, pages 845-854.  
+Charles Welch, Jonathan K Kummerfeld, Verónica Pérez-Rosas, and Rada Mihalcea. 2020. Compositional demographic word embeddings. In EMNLP, pages 4076-4089.  
+Francis CK Wong and William SY Wang. 2007. Generalisation towards combinatorial productivity in language acquisition by simple recurrent networks. In 2007 International Conference on Integration of Knowledge Intensive Multi-Agent Systems, pages 139-144. IEEE.  
+Qinzhuo Wu, Qi Zhang, and Zhongyu Wei. 2021a. An edge-enhanced hierarchical graph-to-tree network for math word problem solving. In *Findings of EMNLP*, pages 1473–1482.  
+Qinzhuo Wu, Qi Zhang, Zhongyu Wei, and Xuanjing Huang. 2021b. Math word problem solving with explicit numerical values. In ACL, pages 5859-5869.
+
+Zhipeng Xie and Shichao Sun. 2019. A goal-driven tree-structured neural model for math word problems. In *IJCAI*, pages 5299–5305.
+
+Jingfeng Yang, Le Zhang, and Diyi Yang. 2022. SUBS: Subtree substitution for compositional semantic parsing. In *NAACN*, pages 169–174, Seattle, United States. Association for Computational Linguistics.
+
+Weijiang Yu, Yingpeng Wen, Fudan Zheng, and Nong Xiao. 2021. Improving math word problems with pre-trained knowledge and hierarchical reasoning. In EMNLP, pages 3384-3394.
+
+Jipeng Zhang, Lei Wang, Roy Ka-Wei Lee, Yi Bin, Yan Wang, Jie Shao, and Ee-Peng Lim. 2020. Graph-to-tree learning for solving math word problems. In ACL, pages 3928-3937.
+
+# Appendix
+
+# Implementation Details
+
+We train our model on an NVIDIA RTX 2080Ti GPU, all implementation of training and testing is coded in Python with Pytorch framework. For our RNN-based model, 2 CMEncoders are stacked and only 1 CMEncoder is used in the PLM-based model. The size of hidden dimensions in encoders and decoders are set to 512 and 768 for RNN-solver and PLM-solver, respectively. Each GCN block has 2 GraphConv layers and each GNN encoder has 4 heads of GCN blocks. During training, Adam optimizer is applied with the initial learning rates of 0.001/0.00003 for RNN/PLM, which would be halved every 30 epochs. During testing, we use a 5-beam search to get reasonable solutions. We also apply Gaussian noise with mean 0 and variance 1 on the embedding result during training. This simple operation can help models to learn more robust parameters. Through grid search at 0.1 level, the noise is multiplied by 0.2 to achieve the best performance.
+
+# Hyper-Parameter Tuning
+
+In general, we apply grid-search with manually designed search space and use answer accuracy as the evaluation metric to select the hyper-parameters. For the number of stacked encoders, the search space is  $\{1,2,3,4\}$  and we finally use 2. For the weight of  $L_{2}$  normalization loss, we choose weight 1 from  $\{0.01,0.1,1,5,10\}$ . The weight of random noise 0.2 is selected from 0.1 level by grid search with range 0 to 1. We also tune the beam size of beam search from  $\{3,4,5,6,7\}$  and choose 5. The dropout probability 0.5 is selected from
+
+<table><tr><td></td><td>K = 1</td><td>K = 2</td><td>K = 3</td><td>K = 4</td></tr><tr><td>C-MWP (RNN)</td><td>79.1</td><td>80.3</td><td>78.9</td><td>78.1</td></tr><tr><td>C-MWP (PLM)</td><td>86.1</td><td>85.9</td><td>85.2</td><td>85.1</td></tr></table>
+
+Table 4: Accuracies across different numbers of stacked CMEncoder on Math23k.
+
+$\{0.1, 0.3, 0.5, 0.7\}$ . Initial learning rate 0.001 is selected from  $\{0.01, 0.001, 0.0001\}$ . For the hidden size and embedding size in encoder, we select 256 from  $\{64, 128, 256, 512\}$ .
+
+# Variance and Significance Evaluation
+
+We evaluated our solver with 5-fold cross-validation and found that the accuracy of our RNN-based C-MWP  $(0.779 \pm 0.028)$  is significantly higher than Graph2Tree  $(0.755 \pm 0.016)$ $(p < 0.01)$ , and the accuracy of our PLM-based C-MWP  $(0.845 \pm 0.21)$  is significantly higher than vanilla MWP-BERT  $(0.824 \pm 0.016)$ $(p < 0.01)$ .
+
+# Sensitivity Analysis about Stacking Number K
+
+As we mentioned in Section 2.2, our CMEncoder can be stacked into multiple layers to improve the representation of an MWP. To obtain a better understanding about the hyperparameter K, i.e., the number of stacked CMEncoders, we conduct a sensitivity analysis in Table 4. We can see that the best K for RNN-encoder is 2, just like humans often need to make multiple glimpses to refine an MWP solution. In the meantime, the best K for PLM-based encoder is 1. The potential reason is that one pre-trained language model (PLM) already has an outstanding ability to encode texts. It is thus not necessary to apply another CMEncoder to refine the encoded features.
+
+# Case Study
+
+Figure 3 shows generated solutions of two solved problems by GTS (Xie and Sun, 2019), Graph2Tree (Zhang et al., 2020), Multi-E/D (Shen and Jin, 2020) and our proposed C-MWP (RNN-Based). The first problem has 4 quantities and they are all useful, which means that it requires sufficient problem understanding and mathematical reasoning to generate the right answer. Both Graph2Tree and Multi-E/D which directly connect semantics modal and quantity modal fail to extract clear representations of the problem, finally resulting in unreasonable solutions which only contain 3 quantities. For the second problem, although Graph2Tree and
+
+![](images/301334b5270e7ff6bc751ca3b631cf30c4a4051d05e20b057034b14c68b75ecf.jpg)  
+Figure 3: Case study from Math23k
+
+<table><tr><td></td><td>Overall</td><td>Story</td><td>Algebra</td><td>Knowledge</td></tr><tr><td>train</td><td>21,162</td><td>17,546</td><td>2,595</td><td>957</td></tr><tr><td>val</td><td>1,000</td><td>817</td><td>133</td><td>50</td></tr><tr><td>test</td><td>1,000</td><td>842</td><td>116</td><td>42</td></tr></table>
+
+Table 5: Statistics of different types of problems in Math23k.
+
+Multi-E/D utilize all 3 quantities in the problem description, they still fail to generate a plausible solution. These two cases show that our proposed encoder is able to extract more comprehensive representations from problem descriptions, eventually guiding the decoder to generate the correct solutions.
+
+# MWPs in Different Categories
+
+Figure 1 shows the MWP examples of "Story Problem", "Algebra Problem" and "Knowledge Problem". "Story Problem" often includes a significant amount of background information like characters, objectives and behaviors. "Algebra Problems" involves math notations or is composed of elementary concepts. "Knowledge Problem" asks for external knowledge like geometry and number sequence. The category of each problem is determined based on keywords. The keywords of "Story" and "Knowledge" problems are selected from the appendix of (Liu et al., 2019). Inspired by them, we categorize Math23k into 3 subsets - story, algebra, and knowledge. The statistics of these problems are shown in Table 5.
+
+# A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+□ A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+# B Did you use or create scientific artifacts?
+
+Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Left blank.
+
+# C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Left blank.
+
+The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank.
+
+# D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Left blank.

2023/Compositional Mathematical Encoding for Math Word Problems/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:521c966f80d1e35500846523fe843d95d44e3c61fc0ebd5e233543f9e81c783e
+size 511314

2023/Computer says “No”_ The Case Against Empathetic Conversational AI/919d9a05-5ebb-4c0e-a1de-82ff5cd4abb0_content_list.json

@@ -0,0 +1,1296 @@
+[
+    {
+        "type": "text",
+        "text": "Computer says “No”: The Case Against Empathetic Conversational AI",
+        "text_level": 1,
+        "bbox": [
+            245,
+            89,
+            751,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Alba Curry  \nSchool of Philosophy, Religion and History of Science  \nUniversity of Leeds  \na.cercascurry@leeds.ac.uk",
+        "bbox": [
+            206,
+            143,
+            470,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Amanda Cercas Curry  \nMilaNLP  \nDepartment of Computing Sciences  \nBocconi University  \namanda.cercas@unibocconi.it",
+        "bbox": [
+            519,
+            143,
+            810,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Emotions are an integral part of human cognition and they guide not only our understanding of the world but also our actions within it. As such, whether we soothe or flame an emotion is not inconsequential. Recent work in conversational AI has focused on responding empathetically to users, validating and soothing their emotions without a real basis. This AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted happiness defined as only the absence of \"negative\" emotions. We argue that we must carefully consider whether and how to respond to users' emotions.",
+        "bbox": [
+            141,
+            281,
+            460,
+            480
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            494,
+            258,
+            508
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Recent work in conversational AI has focused on generating empathetic responses to users' emotional states (e.g., Ide and Kawahara, 2022; Svikhnushina et al., 2022; Zhu et al., 2022) as a way to increase or maintain engagement and rapport with the user and to simulate intelligence. However, these empathetic responses are problematic.",
+        "bbox": [
+            112,
+            521,
+            489,
+            633
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "First, while a system might never claim to be human, responses simulating humanness prompt users to further behave as though the systems were (Reeves and Nass, 1996). Empathy, like all emotions, is likely a uniquely human trait and systems that feign it are in effect feigning humanity. The ethical issues surrounding anthropomorphism have been discussed at length and are beyond the scope of this paper (Salles et al., 2020; Bryson, 2010).",
+        "bbox": [
+            112,
+            634,
+            489,
+            778
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Second, empathy requires an ability to both understand and share another's emotions. As such, responding empathetically assumes that the system is able to correctly identify the emotion, and that it is able to feel the emotion itself. Neither one of",
+        "bbox": [
+            112,
+            778,
+            489,
+            859
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "these holds true for conversational AI (or in fact for any AI system).<sup>2</sup>",
+        "bbox": [
+            507,
+            253,
+            880,
+            284
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Third, even if conversational AI were to correctly identify the user's emotions, and perform empathy, we should ethically question the motives and outcomes behind such an enterprise. Svikhnushina et al. (2022) put forward a taxonomy of empathetic questions in social dialogues, paying special attention to the role questions play in regulating the interlocutor's emotions. They argue for the crucial role effective question asking plays in successful chatbots due to the fact that often questions are used to express \"empathy\" and attentiveness by the speaker. Here we highlight the ethical concerns that arise from questions that are characterised by their emotion-regulation functions targeted at the user's emotional state. It is important to note that our argument applies to any use of empathetic AI (see also for example (Morris et al., 2018; De Carolis et al., 2017)). What happens if the chatbot gets it right? There may be instances where a chatbot correctly identifies that a given situation is worthy of praise and amplifies the pride of the user and the result is morally unproblematic. For example, when (Svikhnushina et al., 2022) use the example of amplifying pride in the context of fishing. What happens if it gets it wrong? It depends on the type of mistake: a) The chatbot fails to put into effect a question's intent, it would be ethically inconsequential;  $^{3}\\mathrm{b})$  It amplifies or minimises an inappropriate emotion. This is the problem we will focus on, arguing that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will",
+        "bbox": [
+            507,
+            286,
+            884,
+            800
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "gle at times to identify our own or extend empathy to certain members of society, but we have the capability of identifying emotions. Furthermore, our ability to identify the emotions of others builds, at least in part, from our own emotions.",
+        "bbox": [
+            507,
+            808,
+            880,
+            854
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Moreover, Barrett (2017) have already problematised the identification of human emotions using language or facial expressions in general.",
+        "bbox": [
+            507,
+            854,
+            880,
+            892
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "3In fact, if the chatbot failed to be empathetic then it would simply not engage us in the intended ways.",
+        "bbox": [
+            507,
+            892,
+            880,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "'Correctly identifying an emotion is problematic for animals including human beings. However, reasons differ between conversation AI and human beings: Human beings vary in their capacity to identify emotions in part because we strut",
+        "bbox": [
+            112,
+            868,
+            489,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "8123",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 8123-8130 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            228,
+            945,
+            768,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its appropriateness. This lack of understanding is key as we cannot predict the consequences of moderating an emotion we don't understand, and a dialogue system cannot be held accountable for them.",
+        "bbox": [
+            112,
+            84,
+            489,
+            197
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 The Crucial Roles of Emotions",
+        "text_level": 1,
+        "bbox": [
+            112,
+            212,
+            413,
+            228
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "What emotions are is still up for debate (Barrett, 2017; Scarantino and de Sousa, 2018). However, their significance for the individual and for society has received renewed interest (Greenspan, 1995; Bell, 2013; Cherry, 2021). Regardless of the emotion model one picks, emotions play important roles, both epistemic and conative ones (Curry, 2022). They perform at least three epistemic roles: (1) They signal to the individual experiencing the emotion what she herself values and how she sees the world (e.g., if you envy your colleague's publications this tells you you value publications and deem yourself similar enough to your colleague that you can compare yourself (Protasi, 2021)); (2) they signal to others how we see the world; and (3) emotional interactions are invaluable sources of information for third-party observers since they tell us what the members of the interaction value. For example, (1) when you grieve, you signal to yourself and anyone observing that you deem to have lost something of value. It is conceivable that you were unaware up to that point that you valued what you lost—this is captured by the saying \"you don't know what you have till it's gone.\" Furthermore, (2) your friends and family may learn something about you by observing your grief. They too may not have known how much something meant for you. Finally, (3) an observer may also learn about the dynamics of grief (whether it is appropriate to express it for example) by observing whether or not your family validates your grief.",
+        "bbox": [
+            112,
+            241,
+            489,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, emotions play conative roles, meaning that they are involved in important ways with our motivation and desire to act in certain ways. In other words, not only do some emotions compel and motivate you to act, but also how you act is coloured by the emotion you are experiencing. For example, your anger signals that you perceive that an injustice has occurred. If your boss fails to promote the person who deserves it because of their gender, your anger would motivate you to write a letter of complaint or speak to HR about it.",
+        "bbox": [
+            112,
+            741,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Importantly, all emotions, including the so-called \"negative\" emotions (e.g., anger, contempt, hatred, shame, envy, guilt, etc.) also share these functions. These emotions are not negative in the sense of being \"bad\", they are called negative because they are painful, and therefore they are emotions that we would tend to avoid for ourselves. A world without injustice would certainly be ideal but we would not want a world of injustice where we were unequipped to notice or become motivated to fight it. Hence why it is imperative that we ask ourselves under which circumstances we ought to enhance or soothe emotions.",
+        "bbox": [
+            507,
+            84,
+            884,
+            293
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 The Problem with Empathy",
+        "text_level": 1,
+        "bbox": [
+            509,
+            309,
+            781,
+            326
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Literature discussing the value and power of empathy for conversational AI understands empathy as a tool to establish a common ground for meaningful communication and to appear more likeable to users. The authors of these studies understand empathy broadly as \"the feeling by which one understands and shares another person's experiences and emotions\" (De Carolis et al., 2017). Empathy facilitates engagement through the development of social relationships, affection, and familiarity. Furthermore, for Svikhnushina et al. (2022), empathy is required in order to enable chatbots to ask questions with emotion regulation intents. For example, questions may be used to amplify the user's pride or de-escalate the user's anger, or frustration.",
+        "bbox": [
+            507,
+            337,
+            884,
+            577
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Empathy, although a common phenomenon, is not a simple one. It enjoys a long history in various scholarly disciplines. Indeed, a lot of ink has been spilled (and still is), for example, over how to make sense of character engagement. How do we, human beings, care for fictional characters? How are we intrigued and moved by their adventures and respond to the emotions and affects expressed in their voices, bodies, and faces as well as imagine the situation they are in and wish them success, closure, or punishment? Empathy is taken to be a key element and yet the exact nature of how human beings are able to experience empathy for fictional characters is currently being debated (Tobón, 2019).",
+        "bbox": [
+            507,
+            580,
+            884,
+            804
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The reason for highlighting this diversity is that conversational AI would do well to engage seriously with the rich intellectual history of empathy. The definition it tends to engage with lacks the level of complexity required to understand this complex phenomenon. Moreover, it tends to obfuscate the darker sides of empathy. Leaving aside the fact",
+        "bbox": [
+            507,
+            806,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "8124",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "that defining empathy as the \"reactions of one individual to the observed experiences of another\" (De Carolis et al., 2017) tells us very little about the process by which a human beings, let alone conversational AI, may do this, what we take issue with is what chatbots hope to do with that empathy. In other words, if for the sake or argument, we presume that conversational AI is able to accurately identify our emotions, the issue of how we deploy empathy is of huge ethical relevance.",
+        "bbox": [
+            110,
+            84,
+            489,
+            243
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Here we offer a brief summary of three important views against empathy: Prinz (2011) argues against the common intuition that empathy is by and large a good thing and thus desirable. He raises several issues such as empathy being easily manipulated (such as during a trial), and empathy being partial (we are more empathetic towards people we perceive to be of our own race, for example). Both claims have been empirically verified. Thinking about how this might affect empathetic conversational AI for example in the case of using them for social assistive robots, we might worry if based on its empathetic reactions it chose to help certain people over others.",
+        "bbox": [
+            110,
+            244,
+            489,
+            470
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Taking the argument further, Bloom (2017) argues against empathy and for what he calls rational compassion. He contends that empathy is one of the leading motivators of inequality and immorality in society. Thus, far from helping us to improve the lives of others, empathy is a capricious and irrational emotion that appeals to our narrow prejudices; it muddles our judgement and, ironically, often leads to cruelty. Instead, we ought to draw upon a more distanced compassion.<sup>4</sup>",
+        "bbox": [
+            110,
+            470,
+            489,
+            632
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There are three lessons we can take from this: (1) Given empathy's prejudices, we would need to think deeper about how to mitigate them in conversational AI; (2) Given that empathy is used not just know what brings people pleasure, but also what brings pain, we might want to question the general future uses of empathy in conversational AI; (3) if we buy Bloom's argument, then conversational AI should consider not imitating human beings, but becoming agents of rational compassion.",
+        "bbox": [
+            110,
+            631,
+            489,
+            791
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Breithaupt (2019) also takes issue with empathy, arguing that we commit atrocities not out of a failure of empathy, but rather as a direct consequence of successful, even overly successful, empathy. He starts the book by reminding us that “[e]xtreme acts",
+        "bbox": [
+            112,
+            793,
+            489,
+            873
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "of cruelty require a high level of empathy.\"",
+        "bbox": [
+            507,
+            84,
+            830,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The further lesson we can take from this is that while people generally assume that empathy leads to morally correct behaviour, and certainly there are many positive sides of empathy, we should not rely on an overly simple or glorified image of empathy.",
+        "bbox": [
+            507,
+            101,
+            884,
+            181
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "However, our problem is not necessarily with empathy per se, but rather with the explicit functions conversational AI hopes to achieve with it, namely to enhance engagement, to inflate emotions deemed positive, and to soothe emotions deemed negative (e.g., Svikhnushina et al., 2022). Our claim is that we ought to think carefully about the consequences of soothing negative emotions only because they we have a bias against them. Not only is this approach based on a naive understanding of emotions, it fails to recognise the importance of human beings being allowed to experience and express the full spectrum of emotions. One ought to not experience negative emotions because there is nothing to be upset about, not because we have devised an emotional pacifier. In other words, the issue is that conversational AI lacks a sound value system for deciding why certain emotions are validated and others soothed. Furthermore, this AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted notion of happiness defined as only the absence of \"negative\" emotions.",
+        "bbox": [
+            507,
+            183,
+            884,
+            554
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 When Emotions Get Things Wrong",
+        "text_level": 1,
+        "bbox": [
+            507,
+            568,
+            850,
+            585
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "There are two illustrative problems with the kinds of decisions behind amplifying and de-escalating emotions. One is the problem of what the ideal character might be. When you talk to a friend they will decide whether to soothe or amplify your emotions based not just on the situation but also on who they deem you to be. If they think you are someone who has a hard time standing up for yourself they will amplify your anger to encourage you to fight for yourself, but if they think you are someone who leans too much on arrogance, they will de-escalate your sense of pride—even if, all things being equal, your pride on that occasion was warranted. Hence, not only would a conversational AI require prior knowledge of the interlocutor in terms of her character, but furthermore it would have to decide what are desirable character traits.",
+        "bbox": [
+            507,
+            596,
+            882,
+            869
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The second question regards what an ideal emotion in a particular situation might be. We may all find it easy to say that negative emotions such",
+        "bbox": [
+            507,
+            871,
+            884,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "Assessing Bloom's argument with regards to rational compassion and whether it would be feasible for conversational AI is beyond the scope of this paper although worthy of pursuit.",
+        "bbox": [
+            112,
+            879,
+            489,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "8125",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "as anger often get things wrong and lead to undesirable outcomes. However, positive emotions such as joy, hope, or pride which we may intuitively wish to amplify can also get things wrong. We assess and criticise emotions along a number of distinct dimensions: Firstly, emotions may be criticised when they do not fit their targets. You may, for example, be open to criticism for feeling fear in the absence of danger. Unfitting emotions fail to correctly present the world. In the case of pride, would we want to amplify someone's pride if they either did not in fact achieve anything, or if their achievement was not merited? For example, if their nephew did very well in maths when in fact we know their nephew cheated? Second, an emotion may be open to criticism when it is not based on good evidence or is unreasonable. Consider the person who suffers from hydrophobia: Given that in the vast majority of situations water is not dangerous, this person's fear is both unreasonable and unfitting. But even fitting emotions may be unreasonable. One may, for example, be terrified of tsunamis because one believes that they cause genetic mutations. In this case, one's fear is fitting—tsunamis are very dangerous—yet the fear is unreasonable since it is not based on good reasons. Third, an emotion may be criticised because it isn't prudent to feel. We might warn someone not to show anger when interacting with a person with a gun since they might get themselves killed; anger in this case may be reasonable and fitting given the gunman's actions and yet imprudent. Finally, we may condemn emotions as morally nonvaluable because of the unacceptable way in which they present their targets, e.g., one may, argue that schadenfreude is morally objectionable because it presents the pain of another person as laughable.",
+        "bbox": [
+            112,
+            84,
+            492,
+            682
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Positive emotions may be unfitting, unreasonable, and imprudent, as well as morally condemning just as negative emotions may well be fitting, reasonable, and prudent, as well as morally laudable. In other words, even if one is equipped with empathy there are crucial normative decisions involved in question intents aimed at emotional regulation.<sup>5</sup> Amplifying and de-escalating emotion inappropriately can have devastating moral outcomes.",
+        "bbox": [
+            112,
+            690,
+            489,
+            851
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "5 Empathy and Responsibility",
+        "text_level": 1,
+        "bbox": [
+            507,
+            83,
+            789,
+            101
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Human beings, all things being equal, will inevitably experience empathy. A reasonable human being experiencing empathy for another is proof of the importance of someone else's emotional state—for better or for worse. This supports the idea that our emotions are important, as opposed to the notion that they hinder rationality and ought to be regulated. They tell us many things about our world.",
+        "bbox": [
+            505,
+            109,
+            884,
+            237
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Similarly to many NLP systems' understanding of language, the empathetic responses of conversational AI are only performative (Bender and Koller, 2020). Thus, they provide a false sense of validity or importance. What if someone is experiencing an unfitting, unreasonable, or morally reprehensible emotion? Should a chatbot still showcase empathy? We hope to have shown that such decisions are deeply morally problematic and complex.",
+        "bbox": [
+            507,
+            237,
+            884,
+            382
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Hence, another key problem is responsibility. A human agent may choose to express their empathy (even if they cannot choose feeling it) and they may choose to attempt to regulate someone else's emotions based on their knowledge of the situation and the speaker's character. If a human being wrongly regulates someone else's emotions, they will be morally responsible for the consequences. Who is morally responsible in the case of conversational AI agents? Who are they benefiting when they are not actually benefiting the human agent? This issue is further elaborated on by Véliz (2021).",
+        "bbox": [
+            507,
+            382,
+            884,
+            576
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "6 Related Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            588,
+            665,
+            602
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our article sits at the intersection of emotion detection, response generation, and safety in conversational AI. We keep this section brief as we cite relevant work throughout the article. Several works have already focused on the issue of giving AI systems sentience, such as Bryson (2010). While this could make the systems truly empathetic, we agree that we have a duty not to create sentient machines.",
+        "bbox": [
+            507,
+            613,
+            884,
+            741
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Lahnala et al. (2022) problematise NLP's conceptualisation of empathy which, they argue, is poorly defined, leading to issues of data validity and missed opportunities for research. Instead, we argue that even a more specific definition of empathy presents ethical issues that cannot be overlooked or ignored and must carefully evaluated.",
+        "bbox": [
+            507,
+            741,
+            885,
+            853
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Dinan et al. (2022) provide a framework to classify and detect safety issues in end-to-end conversational systems. In particular, they point out systems that respond inappropriately to offensive content",
+        "bbox": [
+            507,
+            854,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "See the complex example in Silva (2021)",
+        "bbox": [
+            134,
+            903,
+            394,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "8126",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "and safety-critical issues such as medical and emergency situations. We could apply their framework to empathetic responses where the system takes the role of an \"impostor\": empathetic responses require a system to pretend to understand the emotion. However, the extent to which emotions play a role in human cognition and what the consequences of regulating these emotions for the users are has not been discussed in the literature to the best of our knowledge.",
+        "bbox": [
+            112,
+            84,
+            492,
+            247
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            112,
+            256,
+            247,
+            272
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this position paper, we argued that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its reasonableness. This lack of understanding is key because we cannot predict the consequences of assuaging or aggravating an emotion, and a dialogue system cannot be held accountable for them. We hope to encourage reflection from future researchers and to initiate a discussion of the issue, not only in this particular case but also more reflection when it comes to pursuing seemingly positive goals such as bringing disagreeing parties towards agreement. Like with other ethically sensible topics, the community should come together to agree on a strategy that minimises harm.",
+        "bbox": [
+            112,
+            282,
+            490,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            220,
+            614
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "While we strongly argue against empathetic conversational systems, there may be use cases – such as psychotherapy or educational chatbots – where validating a user's emotions is, if not required, helpful in terms of their goal. In addition, while a lot of the work on empathetic responses we have discussed is intentional, generative models like ChatGPT produce relatively uncontrolled responses that may well be unintentionally empathetic. As with toxic outputs, care should be taken to prevent these models from validating users' emotions that cannot be understood.",
+        "bbox": [
+            112,
+            624,
+            489,
+            816
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            112,
+            829,
+            285,
+            846
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We thank the anonymous reviewers and Gavin Abercrombie for their thorough and helpful comments. This project has partially received funding from the European Research Council (ERC) under",
+        "bbox": [
+            112,
+            854,
+            489,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRATOR). Amanda Cercas Curry is a member of the MilaNLP group and the Data and Marketing Insights Unit of the Bocconi Institute for Data Science and Analysis.",
+        "bbox": [
+            507,
+            84,
+            885,
+            181
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            509,
+            208,
+            608,
+            222
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Lisa Feldman Barrett. 2017. How emotions are made: The secret life of the brain. Pan Macmillan.",
+            "Macalester Bell. 2013. Hard feelings: The moral psychology of contempt. Oxford University Press.",
+            "Emily M Bender and Alexander Koller. 2020. Climbing towards nlu: On meaning, form, and understanding in the age of data. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 5185-5198.",
+            "Paul Bloom. 2017. Against empathy: The case for rational compassion. Random House.",
+            "Fritz Breithaupt. 2019. The dark sides of empathy. Cornell University Press.",
+            "Joanna J Bryson. 2010. Robots should be slaves. Close Engagements with Artificial Companions: Key social, psychological, ethical and design issues, 8:63-74.",
+            "Myisha Cherry. 2021. The case for rage: Why anger is essential to anti-racist struggle. Oxford University Press.",
+            "Alba Curry. 2022. An Apologia for Anger With Reference to Early China and Ancient Greece. Ph.D. thesis, UC Riverside.",
+            "Berardina De Carolis, Stefano Ferilli, and Giuseppe Palestra. 2017. Simulating empathic behavior in a social assistive robot. *Multimedia Tools and Applications*, 76(4):5073-5094.",
+            "Emily Dinan, Gavin Abercrombie, A. Bergman, Shannon Spruit, Dirk Hovy, Y-Lan Boureau, and Verena Rieser. 2022. *SafetyKit: First aid for measuring safety in open-domain conversational systems*. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4113-4133, Dublin, Ireland. Association for Computational Linguistics.",
+            "Patricia S Greenspan. 1995. Practical guilt: Moral dilemmas, emotions, and social norms. Oxford University Press on Demand.",
+            "Tatsuya Ide and Daisuke Kawahara. 2022. Building a dialogue corpus annotated with expressed and experienced emotions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 21-30, Dublin, Ireland. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            509,
+            229,
+            885,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "8127",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Allison Lahnala, Charles Welch, David Jurgens, and Lucie Flek. 2022. A critical reflection and forward perspective on empathy and natural language processing. arXiv preprint arXiv:2210.16604.",
+            "Robert R Morris, Kareem Kouddous, Rohan Kshir-sagar, and Stephen M Schueller. 2018. Towards an artificially empathic conversational agent for mental health applications: system design and user perceptions. Journal of medical Internet research, 20(6):e10148.",
+            "Jesse Prinz. 2011. Against empathy. The Southern Journal of Philosophy, 49:214-233.",
+            "Sara Protasi. 2021. The philosophy of envy. Cambridge University Press.",
+            "Byron Reeves and Clifford Nass. 1996. The media equation: How people treat computers, television, and new media like real people. Cambridge, UK, 10:236605.",
+            "Arleen Salles, Kathinka Evers, and Michele Farisco. 2020. Anthropomorphism in ai. AJOB neuroscience, 11(2):88-95.",
+            "A Scarantino and R de Sousa. 2018. Emotion, in \"the stanford encyclopedia of philosophy\" (winter 2018 edition). EN ZALTA (a cura di), URL: https://plato.stanford.edu/archives/win2018/entries/emotion.",
+            "Laura Silva. 2021. The epistemic role of outlaw emotions. *Ergo*, 8(23).",
+            "Ekaterina Svikhnushina, Iuliana Voinea, Anuradha Welivita, and Pearl Pu. 2022. A taxonomy of empathetic questions in social dialogs. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2952-2973, Dublin, Ireland. Association for Computational Linguistics.",
+            "Daniel Jerónimo Tobón. 2019. Empathy and sympathy: two contemporary models of character engagement. In The Palgrave handbook of the philosophy of film and motion pictures, pages 865-891. Springer.",
+            "Carissa Veliz. 2021. Moral zombies: why algorithms are not moral agents. AI & SOCIETY, 36(2):487-497.",
+            "Ling.Yu Zhu, Zhengkun Zhang, Jun Wang, Hongbin Wang, Haiying Wu, and Zhenglu Yang. 2022. Multi-party empathetic dialogue generation: A new task for dialog systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 298-307, Dublin, Ireland. Association for Computational Linguistics."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            841
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "8128",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A1. Did you describe the limitations of your work?",
+        "bbox": [
+            129,
+            126,
+            532,
+            142
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            149,
+            143,
+            231,
+            158
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A2. Did you discuss any potential risks of your work?",
+        "bbox": [
+            129,
+            170,
+            552,
+            186
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The paper is a position paper warning against a particular application of NLP.",
+        "bbox": [
+            149,
+            187,
+            732,
+            202
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A3. Do the abstract and introduction summarize the paper's main claims?",
+        "bbox": [
+            129,
+            211,
+            695,
+            228
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            230,
+            231,
+            244
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "A4. Have you used AI writing assistants when working on this paper?",
+        "bbox": [
+            129,
+            255,
+            668,
+            272
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            151,
+            273,
+            231,
+            287
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            115,
+            300,
+            490,
+            316
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            132,
+            321,
+            332,
+            336
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B1. Did you cite the creators of artifacts you used?",
+        "bbox": [
+            129,
+            347,
+            529,
+            363
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            363,
+            349,
+            379
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?",
+        "bbox": [
+            129,
+            390,
+            778,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            407,
+            349,
+            422
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?",
+        "bbox": [
+            129,
+            432,
+            880,
+            495
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            498,
+            349,
+            513
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?",
+        "bbox": [
+            129,
+            524,
+            880,
+            571
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            573,
+            349,
+            588
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?",
+        "bbox": [
+            129,
+            599,
+            880,
+            631
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            632,
+            349,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.",
+        "bbox": [
+            129,
+            658,
+            880,
+            739
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            740,
+            349,
+            753
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            115,
+            765,
+            494,
+            781
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            785,
+            213,
+            801
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?",
+        "bbox": [
+            129,
+            813,
+            880,
+            845
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            151,
+            846,
+            349,
+            860
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "footer",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "8129",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?",
+        "bbox": [
+            127,
+            84,
+            878,
+            115
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            117,
+            349,
+            131
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?",
+        "bbox": [
+            127,
+            143,
+            882,
+            191
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            192,
+            349,
+            206
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?",
+        "bbox": [
+            127,
+            218,
+            882,
+            265
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            267,
+            349,
+            282
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?",
+        "bbox": [
+            112,
+            293,
+            877,
+            310
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            313,
+            213,
+            329
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?",
+        "bbox": [
+            127,
+            340,
+            882,
+            372
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            373,
+            349,
+            388
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?",
+        "bbox": [
+            127,
+            399,
+            882,
+            447
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            448,
+            349,
+            464
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?",
+        "bbox": [
+            127,
+            475,
+            882,
+            521
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            523,
+            349,
+            539
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?",
+        "bbox": [
+            127,
+            549,
+            875,
+            565
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            565,
+            349,
+            582
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?",
+        "bbox": [
+            127,
+            592,
+            882,
+            623
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Not applicable. Left blank.",
+        "bbox": [
+            149,
+            625,
+            349,
+            640
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "8130",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 7
+    }
+]

2023/Computer says “No”_ The Case Against Empathetic Conversational AI/919d9a05-5ebb-4c0e-a1de-82ff5cd4abb0_model.json

@@ -0,0 +1,1547 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.247,
+                0.09,
+                0.752,
+                0.13
+            ],
+            "angle": 0,
+            "content": "Computer says “No”: The Case Against Empathetic Conversational AI"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.207,
+                0.144,
+                0.471,
+                0.226
+            ],
+            "angle": 0,
+            "content": "Alba Curry  \nSchool of Philosophy, Religion and History of Science  \nUniversity of Leeds  \na.cercascurry@leeds.ac.uk"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.521,
+                0.144,
+                0.811,
+                0.226
+            ],
+            "angle": 0,
+            "content": "Amanda Cercas Curry  \nMilaNLP  \nDepartment of Computing Sciences  \nBocconi University  \namanda.cercas@unibocconi.it"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.341,
+                0.267
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.282,
+                0.461,
+                0.481
+            ],
+            "angle": 0,
+            "content": "Emotions are an integral part of human cognition and they guide not only our understanding of the world but also our actions within it. As such, whether we soothe or flame an emotion is not inconsequential. Recent work in conversational AI has focused on responding empathetically to users, validating and soothing their emotions without a real basis. This AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted happiness defined as only the absence of \"negative\" emotions. We argue that we must carefully consider whether and how to respond to users' emotions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.495,
+                0.26,
+                0.51
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.522,
+                0.49,
+                0.634
+            ],
+            "angle": 0,
+            "content": "Recent work in conversational AI has focused on generating empathetic responses to users' emotional states (e.g., Ide and Kawahara, 2022; Svikhnushina et al., 2022; Zhu et al., 2022) as a way to increase or maintain engagement and rapport with the user and to simulate intelligence. However, these empathetic responses are problematic."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.635,
+                0.49,
+                0.78
+            ],
+            "angle": 0,
+            "content": "First, while a system might never claim to be human, responses simulating humanness prompt users to further behave as though the systems were (Reeves and Nass, 1996). Empathy, like all emotions, is likely a uniquely human trait and systems that feign it are in effect feigning humanity. The ethical issues surrounding anthropomorphism have been discussed at length and are beyond the scope of this paper (Salles et al., 2020; Bryson, 2010)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.78,
+                0.49,
+                0.86
+            ],
+            "angle": 0,
+            "content": "Second, empathy requires an ability to both understand and share another's emotions. As such, responding empathetically assumes that the system is able to correctly identify the emotion, and that it is able to feel the emotion itself. Neither one of"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.254,
+                0.882,
+                0.285
+            ],
+            "angle": 0,
+            "content": "these holds true for conversational AI (or in fact for any AI system).<sup>2</sup>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.287,
+                0.885,
+                0.801
+            ],
+            "angle": 0,
+            "content": "Third, even if conversational AI were to correctly identify the user's emotions, and perform empathy, we should ethically question the motives and outcomes behind such an enterprise. Svikhnushina et al. (2022) put forward a taxonomy of empathetic questions in social dialogues, paying special attention to the role questions play in regulating the interlocutor's emotions. They argue for the crucial role effective question asking plays in successful chatbots due to the fact that often questions are used to express \"empathy\" and attentiveness by the speaker. Here we highlight the ethical concerns that arise from questions that are characterised by their emotion-regulation functions targeted at the user's emotional state. It is important to note that our argument applies to any use of empathetic AI (see also for example (Morris et al., 2018; De Carolis et al., 2017)). What happens if the chatbot gets it right? There may be instances where a chatbot correctly identifies that a given situation is worthy of praise and amplifies the pride of the user and the result is morally unproblematic. For example, when (Svikhnushina et al., 2022) use the example of amplifying pride in the context of fishing. What happens if it gets it wrong? It depends on the type of mistake: a) The chatbot fails to put into effect a question's intent, it would be ethically inconsequential; \\(^{3}\\mathrm{b})\\) It amplifies or minimises an inappropriate emotion. This is the problem we will focus on, arguing that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.809,
+                0.882,
+                0.856
+            ],
+            "angle": 0,
+            "content": "gle at times to identify our own or extend empathy to certain members of society, but we have the capability of identifying emotions. Furthermore, our ability to identify the emotions of others builds, at least in part, from our own emotions."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.856,
+                0.882,
+                0.893
+            ],
+            "angle": 0,
+            "content": "2Moreover, Barrett (2017) have already problematised the identification of human emotions using language or facial expressions in general."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.508,
+                0.893,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": "3In fact, if the chatbot failed to be empathetic then it would simply not engage us in the intended ways."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.508,
+                0.809,
+                0.882,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.869,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "'Correctly identifying an emotion is problematic for animals including human beings. However, reasons differ between conversation AI and human beings: Human beings vary in their capacity to identify emotions in part because we strut"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8123"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.229,
+                0.946,
+                0.769,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: ACL 2023, pages 8123-8130 July 9-14, 2023 ©2023 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.49,
+                0.198
+            ],
+            "angle": 0,
+            "content": "necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its appropriateness. This lack of understanding is key as we cannot predict the consequences of moderating an emotion we don't understand, and a dialogue system cannot be held accountable for them."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.214,
+                0.414,
+                0.229
+            ],
+            "angle": 0,
+            "content": "2 The Crucial Roles of Emotions"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.242,
+                0.49,
+                0.742
+            ],
+            "angle": 0,
+            "content": "What emotions are is still up for debate (Barrett, 2017; Scarantino and de Sousa, 2018). However, their significance for the individual and for society has received renewed interest (Greenspan, 1995; Bell, 2013; Cherry, 2021). Regardless of the emotion model one picks, emotions play important roles, both epistemic and conative ones (Curry, 2022). They perform at least three epistemic roles: (1) They signal to the individual experiencing the emotion what she herself values and how she sees the world (e.g., if you envy your colleague's publications this tells you you value publications and deem yourself similar enough to your colleague that you can compare yourself (Protasi, 2021)); (2) they signal to others how we see the world; and (3) emotional interactions are invaluable sources of information for third-party observers since they tell us what the members of the interaction value. For example, (1) when you grieve, you signal to yourself and anyone observing that you deem to have lost something of value. It is conceivable that you were unaware up to that point that you valued what you lost—this is captured by the saying \"you don't know what you have till it's gone.\" Furthermore, (2) your friends and family may learn something about you by observing your grief. They too may not have known how much something meant for you. Finally, (3) an observer may also learn about the dynamics of grief (whether it is appropriate to express it for example) by observing whether or not your family validates your grief."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.743,
+                0.49,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Furthermore, emotions play conative roles, meaning that they are involved in important ways with our motivation and desire to act in certain ways. In other words, not only do some emotions compel and motivate you to act, but also how you act is coloured by the emotion you are experiencing. For example, your anger signals that you perceive that an injustice has occurred. If your boss fails to promote the person who deserves it because of their gender, your anger would motivate you to write a letter of complaint or speak to HR about it."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.294
+            ],
+            "angle": 0,
+            "content": "Importantly, all emotions, including the so-called \"negative\" emotions (e.g., anger, contempt, hatred, shame, envy, guilt, etc.) also share these functions. These emotions are not negative in the sense of being \"bad\", they are called negative because they are painful, and therefore they are emotions that we would tend to avoid for ourselves. A world without injustice would certainly be ideal but we would not want a world of injustice where we were unequipped to notice or become motivated to fight it. Hence why it is imperative that we ask ourselves under which circumstances we ought to enhance or soothe emotions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.31,
+                0.783,
+                0.327
+            ],
+            "angle": 0,
+            "content": "3 The Problem with Empathy"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.338,
+                0.885,
+                0.579
+            ],
+            "angle": 0,
+            "content": "Literature discussing the value and power of empathy for conversational AI understands empathy as a tool to establish a common ground for meaningful communication and to appear more likeable to users. The authors of these studies understand empathy broadly as \"the feeling by which one understands and shares another person's experiences and emotions\" (De Carolis et al., 2017). Empathy facilitates engagement through the development of social relationships, affection, and familiarity. Furthermore, for Svikhnushina et al. (2022), empathy is required in order to enable chatbots to ask questions with emotion regulation intents. For example, questions may be used to amplify the user's pride or de-escalate the user's anger, or frustration."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.581,
+                0.885,
+                0.805
+            ],
+            "angle": 0,
+            "content": "Empathy, although a common phenomenon, is not a simple one. It enjoys a long history in various scholarly disciplines. Indeed, a lot of ink has been spilled (and still is), for example, over how to make sense of character engagement. How do we, human beings, care for fictional characters? How are we intrigued and moved by their adventures and respond to the emotions and affects expressed in their voices, bodies, and faces as well as imagine the situation they are in and wish them success, closure, or punishment? Empathy is taken to be a key element and yet the exact nature of how human beings are able to experience empathy for fictional characters is currently being debated (Tobón, 2019)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.807,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The reason for highlighting this diversity is that conversational AI would do well to engage seriously with the rich intellectual history of empathy. The definition it tends to engage with lacks the level of complexity required to understand this complex phenomenon. Moreover, it tends to obfuscate the darker sides of empathy. Leaving aside the fact"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8124"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.085,
+                0.49,
+                0.244
+            ],
+            "angle": 0,
+            "content": "that defining empathy as the \"reactions of one individual to the observed experiences of another\" (De Carolis et al., 2017) tells us very little about the process by which a human beings, let alone conversational AI, may do this, what we take issue with is what chatbots hope to do with that empathy. In other words, if for the sake or argument, we presume that conversational AI is able to accurately identify our emotions, the issue of how we deploy empathy is of huge ethical relevance."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.246,
+                0.49,
+                0.472
+            ],
+            "angle": 0,
+            "content": "Here we offer a brief summary of three important views against empathy: Prinz (2011) argues against the common intuition that empathy is by and large a good thing and thus desirable. He raises several issues such as empathy being easily manipulated (such as during a trial), and empathy being partial (we are more empathetic towards people we perceive to be of our own race, for example). Both claims have been empirically verified. Thinking about how this might affect empathetic conversational AI for example in the case of using them for social assistive robots, we might worry if based on its empathetic reactions it chose to help certain people over others."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.472,
+                0.49,
+                0.633
+            ],
+            "angle": 0,
+            "content": "Taking the argument further, Bloom (2017) argues against empathy and for what he calls rational compassion. He contends that empathy is one of the leading motivators of inequality and immorality in society. Thus, far from helping us to improve the lives of others, empathy is a capricious and irrational emotion that appeals to our narrow prejudices; it muddles our judgement and, ironically, often leads to cruelty. Instead, we ought to draw upon a more distanced compassion.<sup>4</sup>"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.112,
+                0.632,
+                0.49,
+                0.793
+            ],
+            "angle": 0,
+            "content": "There are three lessons we can take from this: (1) Given empathy's prejudices, we would need to think deeper about how to mitigate them in conversational AI; (2) Given that empathy is used not just know what brings people pleasure, but also what brings pain, we might want to question the general future uses of empathy in conversational AI; (3) if we buy Bloom's argument, then conversational AI should consider not imitating human beings, but becoming agents of rational compassion."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.794,
+                0.49,
+                0.875
+            ],
+            "angle": 0,
+            "content": "Breithaupt (2019) also takes issue with empathy, arguing that we commit atrocities not out of a failure of empathy, but rather as a direct consequence of successful, even overly successful, empathy. He starts the book by reminding us that “[e]xtreme acts"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.831,
+                0.101
+            ],
+            "angle": 0,
+            "content": "of cruelty require a high level of empathy.\""
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.102,
+                0.885,
+                0.182
+            ],
+            "angle": 0,
+            "content": "The further lesson we can take from this is that while people generally assume that empathy leads to morally correct behaviour, and certainly there are many positive sides of empathy, we should not rely on an overly simple or glorified image of empathy."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.184,
+                0.885,
+                0.555
+            ],
+            "angle": 0,
+            "content": "However, our problem is not necessarily with empathy per se, but rather with the explicit functions conversational AI hopes to achieve with it, namely to enhance engagement, to inflate emotions deemed positive, and to soothe emotions deemed negative (e.g., Svikhnushina et al., 2022). Our claim is that we ought to think carefully about the consequences of soothing negative emotions only because they we have a bias against them. Not only is this approach based on a naive understanding of emotions, it fails to recognise the importance of human beings being allowed to experience and express the full spectrum of emotions. One ought to not experience negative emotions because there is nothing to be upset about, not because we have devised an emotional pacifier. In other words, the issue is that conversational AI lacks a sound value system for deciding why certain emotions are validated and others soothed. Furthermore, this AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted notion of happiness defined as only the absence of \"negative\" emotions."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.569,
+                0.851,
+                0.586
+            ],
+            "angle": 0,
+            "content": "4 When Emotions Get Things Wrong"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.597,
+                0.883,
+                0.87
+            ],
+            "angle": 0,
+            "content": "There are two illustrative problems with the kinds of decisions behind amplifying and de-escalating emotions. One is the problem of what the ideal character might be. When you talk to a friend they will decide whether to soothe or amplify your emotions based not just on the situation but also on who they deem you to be. If they think you are someone who has a hard time standing up for yourself they will amplify your anger to encourage you to fight for yourself, but if they think you are someone who leans too much on arrogance, they will de-escalate your sense of pride—even if, all things being equal, your pride on that occasion was warranted. Hence, not only would a conversational AI require prior knowledge of the interlocutor in terms of her character, but furthermore it would have to decide what are desirable character traits."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.872,
+                0.885,
+                0.92
+            ],
+            "angle": 0,
+            "content": "The second question regards what an ideal emotion in a particular situation might be. We may all find it easy to say that negative emotions such"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.113,
+                0.881,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Assessing Bloom's argument with regards to rational compassion and whether it would be feasible for conversational AI is beyond the scope of this paper although worthy of pursuit."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8125"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.683
+            ],
+            "angle": 0,
+            "content": "as anger often get things wrong and lead to undesirable outcomes. However, positive emotions such as joy, hope, or pride which we may intuitively wish to amplify can also get things wrong. We assess and criticise emotions along a number of distinct dimensions: Firstly, emotions may be criticised when they do not fit their targets. You may, for example, be open to criticism for feeling fear in the absence of danger. Unfitting emotions fail to correctly present the world. In the case of pride, would we want to amplify someone's pride if they either did not in fact achieve anything, or if their achievement was not merited? For example, if their nephew did very well in maths when in fact we know their nephew cheated? Second, an emotion may be open to criticism when it is not based on good evidence or is unreasonable. Consider the person who suffers from hydrophobia: Given that in the vast majority of situations water is not dangerous, this person's fear is both unreasonable and unfitting. But even fitting emotions may be unreasonable. One may, for example, be terrified of tsunamis because one believes that they cause genetic mutations. In this case, one's fear is fitting—tsunamis are very dangerous—yet the fear is unreasonable since it is not based on good reasons. Third, an emotion may be criticised because it isn't prudent to feel. We might warn someone not to show anger when interacting with a person with a gun since they might get themselves killed; anger in this case may be reasonable and fitting given the gunman's actions and yet imprudent. Finally, we may condemn emotions as morally nonvaluable because of the unacceptable way in which they present their targets, e.g., one may, argue that schadenfreude is morally objectionable because it presents the pain of another person as laughable."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.692,
+                0.49,
+                0.852
+            ],
+            "angle": 0,
+            "content": "Positive emotions may be unfitting, unreasonable, and imprudent, as well as morally condemning just as negative emotions may well be fitting, reasonable, and prudent, as well as morally laudable. In other words, even if one is equipped with empathy there are crucial normative decisions involved in question intents aimed at emotional regulation.<sup>5</sup> Amplifying and de-escalating emotion inappropriately can have devastating moral outcomes."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.084,
+                0.791,
+                0.102
+            ],
+            "angle": 0,
+            "content": "5 Empathy and Responsibility"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.11,
+                0.885,
+                0.238
+            ],
+            "angle": 0,
+            "content": "Human beings, all things being equal, will inevitably experience empathy. A reasonable human being experiencing empathy for another is proof of the importance of someone else's emotional state—for better or for worse. This supports the idea that our emotions are important, as opposed to the notion that they hinder rationality and ought to be regulated. They tell us many things about our world."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.239,
+                0.885,
+                0.383
+            ],
+            "angle": 0,
+            "content": "Similarly to many NLP systems' understanding of language, the empathetic responses of conversational AI are only performative (Bender and Koller, 2020). Thus, they provide a false sense of validity or importance. What if someone is experiencing an unfitting, unreasonable, or morally reprehensible emotion? Should a chatbot still showcase empathy? We hope to have shown that such decisions are deeply morally problematic and complex."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.384,
+                0.885,
+                0.577
+            ],
+            "angle": 0,
+            "content": "Hence, another key problem is responsibility. A human agent may choose to express their empathy (even if they cannot choose feeling it) and they may choose to attempt to regulate someone else's emotions based on their knowledge of the situation and the speaker's character. If a human being wrongly regulates someone else's emotions, they will be morally responsible for the consequences. Who is morally responsible in the case of conversational AI agents? Who are they benefiting when they are not actually benefiting the human agent? This issue is further elaborated on by Véliz (2021)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.589,
+                0.667,
+                0.604
+            ],
+            "angle": 0,
+            "content": "6 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.614,
+                0.885,
+                0.742
+            ],
+            "angle": 0,
+            "content": "Our article sits at the intersection of emotion detection, response generation, and safety in conversational AI. We keep this section brief as we cite relevant work throughout the article. Several works have already focused on the issue of giving AI systems sentience, such as Bryson (2010). While this could make the systems truly empathetic, we agree that we have a duty not to create sentient machines."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.743,
+                0.887,
+                0.854
+            ],
+            "angle": 0,
+            "content": "Lahnala et al. (2022) problematise NLP's conceptualisation of empathy which, they argue, is poorly defined, leading to issues of data validity and missed opportunities for research. Instead, we argue that even a more specific definition of empathy presents ethical issues that cannot be overlooked or ignored and must carefully evaluated."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.509,
+                0.856,
+                0.885,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Dinan et al. (2022) provide a framework to classify and detect safety issues in end-to-end conversational systems. In particular, they point out systems that respond inappropriately to offensive content"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.904,
+                0.396,
+                0.919
+            ],
+            "angle": 0,
+            "content": "See the complex example in Silva (2021)"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8126"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.493,
+                0.248
+            ],
+            "angle": 0,
+            "content": "and safety-critical issues such as medical and emergency situations. We could apply their framework to empathetic responses where the system takes the role of an \"impostor\": empathetic responses require a system to pretend to understand the emotion. However, the extent to which emotions play a role in human cognition and what the consequences of regulating these emotions for the users are has not been discussed in the literature to the best of our knowledge."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.257,
+                0.248,
+                0.273
+            ],
+            "angle": 0,
+            "content": "7 Conclusion"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.283,
+                0.492,
+                0.589
+            ],
+            "angle": 0,
+            "content": "In this position paper, we argued that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its reasonableness. This lack of understanding is key because we cannot predict the consequences of assuaging or aggravating an emotion, and a dialogue system cannot be held accountable for them. We hope to encourage reflection from future researchers and to initiate a discussion of the issue, not only in this particular case but also more reflection when it comes to pursuing seemingly positive goals such as bringing disagreeing parties towards agreement. Like with other ethically sensible topics, the community should come together to agree on a strategy that minimises harm."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.6,
+                0.221,
+                0.615
+            ],
+            "angle": 0,
+            "content": "Limitations"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.625,
+                0.49,
+                0.818
+            ],
+            "angle": 0,
+            "content": "While we strongly argue against empathetic conversational systems, there may be use cases – such as psychotherapy or educational chatbots – where validating a user's emotions is, if not required, helpful in terms of their goal. In addition, while a lot of the work on empathetic responses we have discussed is intentional, generative models like ChatGPT produce relatively uncontrolled responses that may well be unintentionally empathetic. As with toxic outputs, care should be taken to prevent these models from validating users' emotions that cannot be understood."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.83,
+                0.287,
+                0.847
+            ],
+            "angle": 0,
+            "content": "Acknowledgements"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.855,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "We thank the anonymous reviewers and Gavin Abercrombie for their thorough and helpful comments. This project has partially received funding from the European Research Council (ERC) under"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.886,
+                0.182
+            ],
+            "angle": 0,
+            "content": "the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRATOR). Amanda Cercas Curry is a member of the MilaNLP group and the Data and Marketing Insights Unit of the Bocconi Institute for Data Science and Analysis."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.51,
+                0.209,
+                0.61,
+                0.223
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.23,
+                0.884,
+                0.258
+            ],
+            "angle": 0,
+            "content": "Lisa Feldman Barrett. 2017. How emotions are made: The secret life of the brain. Pan Macmillan."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.267,
+                0.885,
+                0.295
+            ],
+            "angle": 0,
+            "content": "Macalester Bell. 2013. Hard feelings: The moral psychology of contempt. Oxford University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.303,
+                0.884,
+                0.37
+            ],
+            "angle": 0,
+            "content": "Emily M Bender and Alexander Koller. 2020. Climbing towards nlu: On meaning, form, and understanding in the age of data. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 5185-5198."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.379,
+                0.885,
+                0.407
+            ],
+            "angle": 0,
+            "content": "Paul Bloom. 2017. Against empathy: The case for rational compassion. Random House."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.416,
+                0.885,
+                0.443
+            ],
+            "angle": 0,
+            "content": "Fritz Breithaupt. 2019. The dark sides of empathy. Cornell University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.452,
+                0.884,
+                0.492
+            ],
+            "angle": 0,
+            "content": "Joanna J Bryson. 2010. Robots should be slaves. Close Engagements with Artificial Companions: Key social, psychological, ethical and design issues, 8:63-74."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.501,
+                0.885,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Myisha Cherry. 2021. The case for rage: Why anger is essential to anti-racist struggle. Oxford University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.55,
+                0.885,
+                0.591
+            ],
+            "angle": 0,
+            "content": "Alba Curry. 2022. An Apologia for Anger With Reference to Early China and Ancient Greece. Ph.D. thesis, UC Riverside."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.6,
+                0.885,
+                0.653
+            ],
+            "angle": 0,
+            "content": "Berardina De Carolis, Stefano Ferilli, and Giuseppe Palestra. 2017. Simulating empathic behavior in a social assistive robot. *Multimedia Tools and Applications*, 76(4):5073-5094."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.511,
+                0.662,
+                0.886,
+                0.768
+            ],
+            "angle": 0,
+            "content": "Emily Dinan, Gavin Abercrombie, A. Bergman, Shannon Spruit, Dirk Hovy, Y-Lan Boureau, and Verena Rieser. 2022. *SafetyKit: First aid for measuring safety in open-domain conversational systems*. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4113-4133, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.777,
+                0.884,
+                0.817
+            ],
+            "angle": 0,
+            "content": "Patricia S Greenspan. 1995. Practical guilt: Moral dilemmas, emotions, and social norms. Oxford University Press on Demand."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.51,
+                0.826,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": "Tatsuya Ide and Daisuke Kawahara. 2022. Building a dialogue corpus annotated with expressed and experienced emotions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 21-30, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.51,
+                0.23,
+                0.886,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8127"
+        }
+    ],
+    [
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.14
+            ],
+            "angle": 0,
+            "content": "Allison Lahnala, Charles Welch, David Jurgens, and Lucie Flek. 2022. A critical reflection and forward perspective on empathy and natural language processing. arXiv preprint arXiv:2210.16604."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.149,
+                0.49,
+                0.228
+            ],
+            "angle": 0,
+            "content": "Robert R Morris, Kareem Kouddous, Rohan Kshir-sagar, and Stephen M Schueller. 2018. Towards an artificially empathic conversational agent for mental health applications: system design and user perceptions. Journal of medical Internet research, 20(6):e10148."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.238,
+                0.486,
+                0.265
+            ],
+            "angle": 0,
+            "content": "Jesse Prinz. 2011. Against empathy. The Southern Journal of Philosophy, 49:214-233."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.274,
+                0.486,
+                0.301
+            ],
+            "angle": 0,
+            "content": "Sara Protasi. 2021. The philosophy of envy. Cambridge University Press."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.311,
+                0.489,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Byron Reeves and Clifford Nass. 1996. The media equation: How people treat computers, television, and new media like real people. Cambridge, UK, 10:236605."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.374,
+                0.489,
+                0.413
+            ],
+            "angle": 0,
+            "content": "Arleen Salles, Kathinka Evers, and Michele Farisco. 2020. Anthropomorphism in ai. AJOB neuroscience, 11(2):88-95."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.423,
+                0.488,
+                0.476
+            ],
+            "angle": 0,
+            "content": "A Scarantino and R de Sousa. 2018. Emotion, in \"the stanford encyclopedia of philosophy\" (winter 2018 edition). EN ZALTA (a cura di), URL: https://plato.stanford.edu/archives/win2018/entries/emotion."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.486,
+                0.489,
+                0.513
+            ],
+            "angle": 0,
+            "content": "Laura Silva. 2021. The epistemic role of outlaw emotions. *Ergo*, 8(23)."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.523,
+                0.49,
+                0.614
+            ],
+            "angle": 0,
+            "content": "Ekaterina Svikhnushina, Iuliana Voinea, Anuradha Welivita, and Pearl Pu. 2022. A taxonomy of empathetic questions in social dialogs. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2952-2973, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.625,
+                0.489,
+                0.678
+            ],
+            "angle": 0,
+            "content": "Daniel Jerónimo Tobón. 2019. Empathy and sympathy: two contemporary models of character engagement. In The Palgrave handbook of the philosophy of film and motion pictures, pages 865-891. Springer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.687,
+                0.49,
+                0.726
+            ],
+            "angle": 0,
+            "content": "Carissa Veliz. 2021. Moral zombies: why algorithms are not moral agents. AI & SOCIETY, 36(2):487-497."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.737,
+                0.49,
+                0.842
+            ],
+            "angle": 0,
+            "content": "Ling.Yu Zhu, Zhengkun Zhang, Jun Wang, Hongbin Wang, Haiying Wu, and Zhenglu Yang. 2022. Multi-party empathetic dialogue generation: A new task for dialog systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 298-307, Dublin, Ireland. Association for Computational Linguistics."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.086,
+                0.49,
+                0.842
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8128"
+        }
+    ],
+    [
+        {
+            "type": "header",
+            "bbox": [
+                0.134,
+                0.085,
+                0.435,
+                0.1
+            ],
+            "angle": 0,
+            "content": "ACL 2023 Responsible NLP Checklist"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.108,
+                0.323,
+                0.123
+            ],
+            "angle": 0,
+            "content": "A For every submission:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.131,
+                0.127,
+                0.533,
+                0.143
+            ],
+            "angle": 0,
+            "content": "A1. Did you describe the limitations of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.145,
+                0.233,
+                0.159
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.171,
+                0.553,
+                0.187
+            ],
+            "angle": 0,
+            "content": "A2. Did you discuss any potential risks of your work?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.188,
+                0.734,
+                0.203
+            ],
+            "angle": 0,
+            "content": "The paper is a position paper warning against a particular application of NLP."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.212,
+                0.697,
+                0.229
+            ],
+            "angle": 0,
+            "content": "A3. Do the abstract and introduction summarize the paper's main claims?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.231,
+                0.233,
+                0.245
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.256,
+                0.67,
+                0.273
+            ],
+            "angle": 0,
+            "content": "A4. Have you used AI writing assistants when working on this paper?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.274,
+                0.233,
+                0.288
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.301,
+                0.491,
+                0.317
+            ],
+            "angle": 0,
+            "content": "B Did you use or create scientific artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.322,
+                0.334,
+                0.337
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.348,
+                0.53,
+                0.364
+            ],
+            "angle": 0,
+            "content": "B1. Did you cite the creators of artifacts you used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.365,
+                0.351,
+                0.38
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.391,
+                0.779,
+                0.407
+            ],
+            "angle": 0,
+            "content": "B2. Did you discuss the license or terms for use and / or distribution of any artifacts?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.408,
+                0.351,
+                0.423
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.434,
+                0.882,
+                0.497
+            ],
+            "angle": 0,
+            "content": "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.499,
+                0.351,
+                0.514
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.525,
+                0.882,
+                0.573
+            ],
+            "angle": 0,
+            "content": "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.574,
+                0.351,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.6,
+                0.882,
+                0.632
+            ],
+            "angle": 0,
+            "content": "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.633,
+                0.351,
+                0.648
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.659,
+                0.882,
+                0.74
+            ],
+            "angle": 0,
+            "content": "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.741,
+                0.351,
+                0.755
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.116,
+                0.766,
+                0.495,
+                0.782
+            ],
+            "angle": 0,
+            "content": "C Did you run computational experiments?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.787,
+                0.215,
+                0.802
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.13,
+                0.814,
+                0.882,
+                0.846
+            ],
+            "angle": 0,
+            "content": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.152,
+                0.847,
+                0.351,
+                0.861
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.114,
+                0.869,
+                0.878,
+                0.893
+            ],
+            "angle": 0,
+            "content": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.929,
+                0.521,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8129"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.085,
+                0.88,
+                0.116
+            ],
+            "angle": 0,
+            "content": "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.118,
+                0.351,
+                0.133
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.144,
+                0.884,
+                0.192
+            ],
+            "angle": 0,
+            "content": "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.193,
+                0.351,
+                0.208
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.219,
+                0.884,
+                0.266
+            ],
+            "angle": 0,
+            "content": "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.15,
+                0.268,
+                0.351,
+                0.283
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.294,
+                0.878,
+                0.311
+            ],
+            "angle": 0,
+            "content": "D Did you use human annotators (e.g., crowdworkers) or research with human participants?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.133,
+                0.315,
+                0.215,
+                0.33
+            ],
+            "angle": 0,
+            "content": "Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.341,
+                0.884,
+                0.373
+            ],
+            "angle": 0,
+            "content": "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.374,
+                0.351,
+                0.39
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.4,
+                0.884,
+                0.448
+            ],
+            "angle": 0,
+            "content": "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.449,
+                0.351,
+                0.465
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.476,
+                0.884,
+                0.523
+            ],
+            "angle": 0,
+            "content": "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.524,
+                0.351,
+                0.54
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.55,
+                0.876,
+                0.566
+            ],
+            "angle": 0,
+            "content": "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.567,
+                0.351,
+                0.583
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.129,
+                0.593,
+                0.884,
+                0.624
+            ],
+            "angle": 0,
+            "content": "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.151,
+                0.626,
+                0.351,
+                0.642
+            ],
+            "angle": 0,
+            "content": "Not applicable. Left blank."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.481,
+                0.928,
+                0.522,
+                0.941
+            ],
+            "angle": 0,
+            "content": "8130"
+        }
+    ]
+]

2023/Computer says “No”_ The Case Against Empathetic Conversational AI/919d9a05-5ebb-4c0e-a1de-82ff5cd4abb0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8af1ffc785effac0968f52a8465f4976796907fd1d2f6cdbea43c5872f2409de
+size 147068

2023/Computer says “No”_ The Case Against Empathetic Conversational AI/full.md

@@ -0,0 +1,222 @@
+# Computer says “No”: The Case Against Empathetic Conversational AI
+
+Alba Curry  
+School of Philosophy, Religion and History of Science  
+University of Leeds  
+a.cercascurry@leeds.ac.uk
+
+Amanda Cercas Curry  
+MilaNLP  
+Department of Computing Sciences  
+Bocconi University  
+amanda.cercas@unibocconi.it
+
+# Abstract
+
+Emotions are an integral part of human cognition and they guide not only our understanding of the world but also our actions within it. As such, whether we soothe or flame an emotion is not inconsequential. Recent work in conversational AI has focused on responding empathetically to users, validating and soothing their emotions without a real basis. This AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted happiness defined as only the absence of "negative" emotions. We argue that we must carefully consider whether and how to respond to users' emotions.
+
+# 1 Introduction
+
+Recent work in conversational AI has focused on generating empathetic responses to users' emotional states (e.g., Ide and Kawahara, 2022; Svikhnushina et al., 2022; Zhu et al., 2022) as a way to increase or maintain engagement and rapport with the user and to simulate intelligence. However, these empathetic responses are problematic.
+
+First, while a system might never claim to be human, responses simulating humanness prompt users to further behave as though the systems were (Reeves and Nass, 1996). Empathy, like all emotions, is likely a uniquely human trait and systems that feign it are in effect feigning humanity. The ethical issues surrounding anthropomorphism have been discussed at length and are beyond the scope of this paper (Salles et al., 2020; Bryson, 2010).
+
+Second, empathy requires an ability to both understand and share another's emotions. As such, responding empathetically assumes that the system is able to correctly identify the emotion, and that it is able to feel the emotion itself. Neither one of
+
+these holds true for conversational AI (or in fact for any AI system).<sup>2</sup>
+
+Third, even if conversational AI were to correctly identify the user's emotions, and perform empathy, we should ethically question the motives and outcomes behind such an enterprise. Svikhnushina et al. (2022) put forward a taxonomy of empathetic questions in social dialogues, paying special attention to the role questions play in regulating the interlocutor's emotions. They argue for the crucial role effective question asking plays in successful chatbots due to the fact that often questions are used to express "empathy" and attentiveness by the speaker. Here we highlight the ethical concerns that arise from questions that are characterised by their emotion-regulation functions targeted at the user's emotional state. It is important to note that our argument applies to any use of empathetic AI (see also for example (Morris et al., 2018; De Carolis et al., 2017)). What happens if the chatbot gets it right? There may be instances where a chatbot correctly identifies that a given situation is worthy of praise and amplifies the pride of the user and the result is morally unproblematic. For example, when (Svikhnushina et al., 2022) use the example of amplifying pride in the context of fishing. What happens if it gets it wrong? It depends on the type of mistake: a) The chatbot fails to put into effect a question's intent, it would be ethically inconsequential;  $^{3}\mathrm{b})$  It amplifies or minimises an inappropriate emotion. This is the problem we will focus on, arguing that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will
+
+necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its appropriateness. This lack of understanding is key as we cannot predict the consequences of moderating an emotion we don't understand, and a dialogue system cannot be held accountable for them.
+
+# 2 The Crucial Roles of Emotions
+
+What emotions are is still up for debate (Barrett, 2017; Scarantino and de Sousa, 2018). However, their significance for the individual and for society has received renewed interest (Greenspan, 1995; Bell, 2013; Cherry, 2021). Regardless of the emotion model one picks, emotions play important roles, both epistemic and conative ones (Curry, 2022). They perform at least three epistemic roles: (1) They signal to the individual experiencing the emotion what she herself values and how she sees the world (e.g., if you envy your colleague's publications this tells you you value publications and deem yourself similar enough to your colleague that you can compare yourself (Protasi, 2021)); (2) they signal to others how we see the world; and (3) emotional interactions are invaluable sources of information for third-party observers since they tell us what the members of the interaction value. For example, (1) when you grieve, you signal to yourself and anyone observing that you deem to have lost something of value. It is conceivable that you were unaware up to that point that you valued what you lost—this is captured by the saying "you don't know what you have till it's gone." Furthermore, (2) your friends and family may learn something about you by observing your grief. They too may not have known how much something meant for you. Finally, (3) an observer may also learn about the dynamics of grief (whether it is appropriate to express it for example) by observing whether or not your family validates your grief.
+
+Furthermore, emotions play conative roles, meaning that they are involved in important ways with our motivation and desire to act in certain ways. In other words, not only do some emotions compel and motivate you to act, but also how you act is coloured by the emotion you are experiencing. For example, your anger signals that you perceive that an injustice has occurred. If your boss fails to promote the person who deserves it because of their gender, your anger would motivate you to write a letter of complaint or speak to HR about it.
+
+Importantly, all emotions, including the so-called "negative" emotions (e.g., anger, contempt, hatred, shame, envy, guilt, etc.) also share these functions. These emotions are not negative in the sense of being "bad", they are called negative because they are painful, and therefore they are emotions that we would tend to avoid for ourselves. A world without injustice would certainly be ideal but we would not want a world of injustice where we were unequipped to notice or become motivated to fight it. Hence why it is imperative that we ask ourselves under which circumstances we ought to enhance or soothe emotions.
+
+# 3 The Problem with Empathy
+
+Literature discussing the value and power of empathy for conversational AI understands empathy as a tool to establish a common ground for meaningful communication and to appear more likeable to users. The authors of these studies understand empathy broadly as "the feeling by which one understands and shares another person's experiences and emotions" (De Carolis et al., 2017). Empathy facilitates engagement through the development of social relationships, affection, and familiarity. Furthermore, for Svikhnushina et al. (2022), empathy is required in order to enable chatbots to ask questions with emotion regulation intents. For example, questions may be used to amplify the user's pride or de-escalate the user's anger, or frustration.
+
+Empathy, although a common phenomenon, is not a simple one. It enjoys a long history in various scholarly disciplines. Indeed, a lot of ink has been spilled (and still is), for example, over how to make sense of character engagement. How do we, human beings, care for fictional characters? How are we intrigued and moved by their adventures and respond to the emotions and affects expressed in their voices, bodies, and faces as well as imagine the situation they are in and wish them success, closure, or punishment? Empathy is taken to be a key element and yet the exact nature of how human beings are able to experience empathy for fictional characters is currently being debated (Tobón, 2019).
+
+The reason for highlighting this diversity is that conversational AI would do well to engage seriously with the rich intellectual history of empathy. The definition it tends to engage with lacks the level of complexity required to understand this complex phenomenon. Moreover, it tends to obfuscate the darker sides of empathy. Leaving aside the fact
+
+that defining empathy as the "reactions of one individual to the observed experiences of another" (De Carolis et al., 2017) tells us very little about the process by which a human beings, let alone conversational AI, may do this, what we take issue with is what chatbots hope to do with that empathy. In other words, if for the sake or argument, we presume that conversational AI is able to accurately identify our emotions, the issue of how we deploy empathy is of huge ethical relevance.
+
+Here we offer a brief summary of three important views against empathy: Prinz (2011) argues against the common intuition that empathy is by and large a good thing and thus desirable. He raises several issues such as empathy being easily manipulated (such as during a trial), and empathy being partial (we are more empathetic towards people we perceive to be of our own race, for example). Both claims have been empirically verified. Thinking about how this might affect empathetic conversational AI for example in the case of using them for social assistive robots, we might worry if based on its empathetic reactions it chose to help certain people over others.
+
+Taking the argument further, Bloom (2017) argues against empathy and for what he calls rational compassion. He contends that empathy is one of the leading motivators of inequality and immorality in society. Thus, far from helping us to improve the lives of others, empathy is a capricious and irrational emotion that appeals to our narrow prejudices; it muddles our judgement and, ironically, often leads to cruelty. Instead, we ought to draw upon a more distanced compassion.<sup>4</sup>
+
+There are three lessons we can take from this: (1) Given empathy's prejudices, we would need to think deeper about how to mitigate them in conversational AI; (2) Given that empathy is used not just know what brings people pleasure, but also what brings pain, we might want to question the general future uses of empathy in conversational AI; (3) if we buy Bloom's argument, then conversational AI should consider not imitating human beings, but becoming agents of rational compassion.
+
+Breithaupt (2019) also takes issue with empathy, arguing that we commit atrocities not out of a failure of empathy, but rather as a direct consequence of successful, even overly successful, empathy. He starts the book by reminding us that “[e]xtreme acts
+
+of cruelty require a high level of empathy."
+
+The further lesson we can take from this is that while people generally assume that empathy leads to morally correct behaviour, and certainly there are many positive sides of empathy, we should not rely on an overly simple or glorified image of empathy.
+
+However, our problem is not necessarily with empathy per se, but rather with the explicit functions conversational AI hopes to achieve with it, namely to enhance engagement, to inflate emotions deemed positive, and to soothe emotions deemed negative (e.g., Svikhnushina et al., 2022). Our claim is that we ought to think carefully about the consequences of soothing negative emotions only because they we have a bias against them. Not only is this approach based on a naive understanding of emotions, it fails to recognise the importance of human beings being allowed to experience and express the full spectrum of emotions. One ought to not experience negative emotions because there is nothing to be upset about, not because we have devised an emotional pacifier. In other words, the issue is that conversational AI lacks a sound value system for deciding why certain emotions are validated and others soothed. Furthermore, this AI-aided emotional regulation can have negative consequences for users and society, tending towards a one-noted notion of happiness defined as only the absence of "negative" emotions.
+
+# 4 When Emotions Get Things Wrong
+
+There are two illustrative problems with the kinds of decisions behind amplifying and de-escalating emotions. One is the problem of what the ideal character might be. When you talk to a friend they will decide whether to soothe or amplify your emotions based not just on the situation but also on who they deem you to be. If they think you are someone who has a hard time standing up for yourself they will amplify your anger to encourage you to fight for yourself, but if they think you are someone who leans too much on arrogance, they will de-escalate your sense of pride—even if, all things being equal, your pride on that occasion was warranted. Hence, not only would a conversational AI require prior knowledge of the interlocutor in terms of her character, but furthermore it would have to decide what are desirable character traits.
+
+The second question regards what an ideal emotion in a particular situation might be. We may all find it easy to say that negative emotions such
+
+as anger often get things wrong and lead to undesirable outcomes. However, positive emotions such as joy, hope, or pride which we may intuitively wish to amplify can also get things wrong. We assess and criticise emotions along a number of distinct dimensions: Firstly, emotions may be criticised when they do not fit their targets. You may, for example, be open to criticism for feeling fear in the absence of danger. Unfitting emotions fail to correctly present the world. In the case of pride, would we want to amplify someone's pride if they either did not in fact achieve anything, or if their achievement was not merited? For example, if their nephew did very well in maths when in fact we know their nephew cheated? Second, an emotion may be open to criticism when it is not based on good evidence or is unreasonable. Consider the person who suffers from hydrophobia: Given that in the vast majority of situations water is not dangerous, this person's fear is both unreasonable and unfitting. But even fitting emotions may be unreasonable. One may, for example, be terrified of tsunamis because one believes that they cause genetic mutations. In this case, one's fear is fitting—tsunamis are very dangerous—yet the fear is unreasonable since it is not based on good reasons. Third, an emotion may be criticised because it isn't prudent to feel. We might warn someone not to show anger when interacting with a person with a gun since they might get themselves killed; anger in this case may be reasonable and fitting given the gunman's actions and yet imprudent. Finally, we may condemn emotions as morally nonvaluable because of the unacceptable way in which they present their targets, e.g., one may, argue that schadenfreude is morally objectionable because it presents the pain of another person as laughable.
+
+Positive emotions may be unfitting, unreasonable, and imprudent, as well as morally condemning just as negative emotions may well be fitting, reasonable, and prudent, as well as morally laudable. In other words, even if one is equipped with empathy there are crucial normative decisions involved in question intents aimed at emotional regulation.<sup>5</sup> Amplifying and de-escalating emotion inappropriately can have devastating moral outcomes.
+
+# 5 Empathy and Responsibility
+
+Human beings, all things being equal, will inevitably experience empathy. A reasonable human being experiencing empathy for another is proof of the importance of someone else's emotional state—for better or for worse. This supports the idea that our emotions are important, as opposed to the notion that they hinder rationality and ought to be regulated. They tell us many things about our world.
+
+Similarly to many NLP systems' understanding of language, the empathetic responses of conversational AI are only performative (Bender and Koller, 2020). Thus, they provide a false sense of validity or importance. What if someone is experiencing an unfitting, unreasonable, or morally reprehensible emotion? Should a chatbot still showcase empathy? We hope to have shown that such decisions are deeply morally problematic and complex.
+
+Hence, another key problem is responsibility. A human agent may choose to express their empathy (even if they cannot choose feeling it) and they may choose to attempt to regulate someone else's emotions based on their knowledge of the situation and the speaker's character. If a human being wrongly regulates someone else's emotions, they will be morally responsible for the consequences. Who is morally responsible in the case of conversational AI agents? Who are they benefiting when they are not actually benefiting the human agent? This issue is further elaborated on by Véliz (2021).
+
+# 6 Related Work
+
+Our article sits at the intersection of emotion detection, response generation, and safety in conversational AI. We keep this section brief as we cite relevant work throughout the article. Several works have already focused on the issue of giving AI systems sentience, such as Bryson (2010). While this could make the systems truly empathetic, we agree that we have a duty not to create sentient machines.
+
+Lahnala et al. (2022) problematise NLP's conceptualisation of empathy which, they argue, is poorly defined, leading to issues of data validity and missed opportunities for research. Instead, we argue that even a more specific definition of empathy presents ethical issues that cannot be overlooked or ignored and must carefully evaluated.
+
+Dinan et al. (2022) provide a framework to classify and detect safety issues in end-to-end conversational systems. In particular, they point out systems that respond inappropriately to offensive content
+
+and safety-critical issues such as medical and emergency situations. We could apply their framework to empathetic responses where the system takes the role of an "impostor": empathetic responses require a system to pretend to understand the emotion. However, the extent to which emotions play a role in human cognition and what the consequences of regulating these emotions for the users are has not been discussed in the literature to the best of our knowledge.
+
+# 7 Conclusion
+
+In this position paper, we argued that emotional regulation has no place in conversational AI and as such empathetic responses are deeply morally problematic. While humans will necessarily show empathy for one another, conversational AI cannot understand the emotion and so cannot make an accurate judgement as to its reasonableness. This lack of understanding is key because we cannot predict the consequences of assuaging or aggravating an emotion, and a dialogue system cannot be held accountable for them. We hope to encourage reflection from future researchers and to initiate a discussion of the issue, not only in this particular case but also more reflection when it comes to pursuing seemingly positive goals such as bringing disagreeing parties towards agreement. Like with other ethically sensible topics, the community should come together to agree on a strategy that minimises harm.
+
+# Limitations
+
+While we strongly argue against empathetic conversational systems, there may be use cases – such as psychotherapy or educational chatbots – where validating a user's emotions is, if not required, helpful in terms of their goal. In addition, while a lot of the work on empathetic responses we have discussed is intentional, generative models like ChatGPT produce relatively uncontrolled responses that may well be unintentionally empathetic. As with toxic outputs, care should be taken to prevent these models from validating users' emotions that cannot be understood.
+
+# Acknowledgements
+
+We thank the anonymous reviewers and Gavin Abercrombie for their thorough and helpful comments. This project has partially received funding from the European Research Council (ERC) under
+
+the European Union's Horizon 2020 research and innovation program (grant agreement No. 949944, INTEGRATOR). Amanda Cercas Curry is a member of the MilaNLP group and the Data and Marketing Insights Unit of the Bocconi Institute for Data Science and Analysis.
+
+# References
+
+Lisa Feldman Barrett. 2017. How emotions are made: The secret life of the brain. Pan Macmillan.  
+Macalester Bell. 2013. Hard feelings: The moral psychology of contempt. Oxford University Press.  
+Emily M Bender and Alexander Koller. 2020. Climbing towards nlu: On meaning, form, and understanding in the age of data. In Proceedings of the 58th annual meeting of the association for computational linguistics, pages 5185-5198.  
+Paul Bloom. 2017. Against empathy: The case for rational compassion. Random House.  
+Fritz Breithaupt. 2019. The dark sides of empathy. Cornell University Press.  
+Joanna J Bryson. 2010. Robots should be slaves. Close Engagements with Artificial Companions: Key social, psychological, ethical and design issues, 8:63-74.  
+Myisha Cherry. 2021. The case for rage: Why anger is essential to anti-racist struggle. Oxford University Press.  
+Alba Curry. 2022. An Apologia for Anger With Reference to Early China and Ancient Greece. Ph.D. thesis, UC Riverside.  
+Berardina De Carolis, Stefano Ferilli, and Giuseppe Palestra. 2017. Simulating empathic behavior in a social assistive robot. *Multimedia Tools and Applications*, 76(4):5073-5094.  
+Emily Dinan, Gavin Abercrombie, A. Bergman, Shannon Spruit, Dirk Hovy, Y-Lan Boureau, and Verena Rieser. 2022. *SafetyKit: First aid for measuring safety in open-domain conversational systems*. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4113-4133, Dublin, Ireland. Association for Computational Linguistics.  
+Patricia S Greenspan. 1995. Practical guilt: Moral dilemmas, emotions, and social norms. Oxford University Press on Demand.  
+Tatsuya Ide and Daisuke Kawahara. 2022. Building a dialogue corpus annotated with expressed and experienced emotions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop, pages 21-30, Dublin, Ireland. Association for Computational Linguistics.
+
+Allison Lahnala, Charles Welch, David Jurgens, and Lucie Flek. 2022. A critical reflection and forward perspective on empathy and natural language processing. arXiv preprint arXiv:2210.16604.  
+Robert R Morris, Kareem Kouddous, Rohan Kshir-sagar, and Stephen M Schueller. 2018. Towards an artificially empathic conversational agent for mental health applications: system design and user perceptions. Journal of medical Internet research, 20(6):e10148.  
+Jesse Prinz. 2011. Against empathy. The Southern Journal of Philosophy, 49:214-233.  
+Sara Protasi. 2021. The philosophy of envy. Cambridge University Press.  
+Byron Reeves and Clifford Nass. 1996. The media equation: How people treat computers, television, and new media like real people. Cambridge, UK, 10:236605.  
+Arleen Salles, Kathinka Evers, and Michele Farisco. 2020. Anthropomorphism in ai. AJOB neuroscience, 11(2):88-95.  
+A Scarantino and R de Sousa. 2018. Emotion, in "the stanford encyclopedia of philosophy" (winter 2018 edition). EN ZALTA (a cura di), URL: https://plato.stanford.edu/archives/win2018/entries/emotion.  
+Laura Silva. 2021. The epistemic role of outlaw emotions. *Ergo*, 8(23).  
+Ekaterina Svikhnushina, Iuliana Voinea, Anuradha Welivita, and Pearl Pu. 2022. A taxonomy of empathetic questions in social dialogs. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2952-2973, Dublin, Ireland. Association for Computational Linguistics.  
+Daniel Jerónimo Tobón. 2019. Empathy and sympathy: two contemporary models of character engagement. In The Palgrave handbook of the philosophy of film and motion pictures, pages 865-891. Springer.  
+Carissa Veliz. 2021. Moral zombies: why algorithms are not moral agents. AI & SOCIETY, 36(2):487-497.  
+Ling.Yu Zhu, Zhengkun Zhang, Jun Wang, Hongbin Wang, Haiying Wu, and Zhenglu Yang. 2022. Multi-party empathetic dialogue generation: A new task for dialog systems. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 298-307, Dublin, Ireland. Association for Computational Linguistics.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+Left blank.
+
+A2. Did you discuss any potential risks of your work?
+
+The paper is a position paper warning against a particular application of NLP.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Not applicable. Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Not applicable. Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Not applicable. Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Not applicable. Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Not applicable. Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Not applicable. Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+Not applicable. Left blank.
+
+C Did you run computational experiments?
+
+Left blank.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+Not applicable. Left blank.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values?
+
+Not applicable. Left blank.
+
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run?
+
+Not applicable. Left blank.
+
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)?
+
+Not applicable. Left blank.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants?
+
+Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.?
+
+Not applicable. Left blank.
+
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)?
+
+Not applicable. Left blank.
+
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used?
+
+Not applicable. Left blank.
+
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board?
+
+Not applicable. Left blank.
+
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data?
+
+Not applicable. Left blank.

2023/ConKI_ Contrastive Knowledge Injection for Multimodal Sentiment Analysis/b434c033-5cf4-43ce-8d1a-9159c5428ecf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8d4bf3f4c3e9b8ccf314a1fd31dfaaeaf9565bb7685cb7aa3f71c2a91cf7b73e
+size 1286582

2023/ConKI_ Contrastive Knowledge Injection for Multimodal Sentiment Analysis/full.md

@@ -0,0 +1,467 @@
+# ConKI: Contrastive Knowledge Injection for Multimodal Sentiment Analysis
+
+Yakun Yu $^{1}$ , Mingjun Zhao $^{1}$ , Shi-ang Qi $^{1}$ , Feiran Sun $^{2}$ , Baoxun Wang $^{2}$ , Weidong Guo $^{2}$ , Xiaoli Wang $^{2}$ , Lei Yang $^{2}$ , Di Niu $^{1}$
+
+<sup>1</sup>University of Alberta
+
+$^{2}$ Platform and Content Group, Tencent
+
+$^{1}\{yakun2, zhao2, shiang, dniu\} \text{@ualberta.ca}$
+
+# Abstract
+
+Multimodal Sentiment Analysis leverages multimodal signals to detect the sentiment of a speaker. Previous approaches concentrate on performing multimodal fusion and representation learning based on general knowledge obtained from pretrained models, which neglects the effect of domain-specific knowledge. In this paper, we propose Contrastive Knowledge Injection (ConKI) for multimodal sentiment analysis, where specific-knowledge representations for each modality can be learned together with general knowledge representations via knowledge injection based on an adapter architecture. In addition, ConKI uses a hierarchical contrastive learning procedure performed between knowledge types within every single modality, across modalities within each sample, and across samples to facilitate the effective learning of the proposed representations, hence improving multimodal sentiment predictions. The experiments on three popular multimodal sentiment analysis benchmarks show that ConKI outperforms all prior methods on a variety of performance metrics.
+
+# 1 Introduction
+
+Multimodal sentiment analysis (MSA) is the task of mining and comprehending the sentiments of online videos, which has many downstream applications, e.g., analyzing the overall opinion from customers about a product, gauging polling intentions from voters (Han et al., 2021; Melville et al., 2009), etc. Most existing MSA methods focus on developing fusion techniques between modalities. The easiest way is to simply concatenate text, video, and audio features as a fused vector for subsequent classification or regression. An alternative is to use outer-product, Recurrent Neural Networks (RNNs) or attention-based models to model multimodal interactions (Chen et al., 2017; Williams et al., 2018; Zadeh et al., 2017; Liu and Shen, 2018). More recently, MSA methods for learning effective multimodal representations have emerged constantly
+
+(Hazarika et al., 2020; Mai et al., 2021; Yu et al., 2021), ranging from decomposing the representation of each modality to introducing extra constraints in the learning objective.
+
+Although the above methods have led to improvements in MSA performance, they focus on utilizing general knowledge obtained from pretrained models to encode modalities, which is inadequate to identify specific sentiments across modalities. One possibility to solve this issue is through knowledge injection which can generate specific knowledge to aid the general knowledge for further improving predictions. Many researchers have discovered that injecting knowledge from other sources such as linguistic knowledge, encyclopedia knowledge, and domain-specific knowledge can help enhance existing pretrained language models in terms of knowledge awareness and lead to improved performance on various downstream tasks (Wei et al., 2021; Lauscher et al., 2020; Wang et al., 2021a).
+
+In this paper, we propose ConKI, a Contrastive Knowledge Injection framework, to learn both pan-knowledge representations and knowledge-specific representations to boost MSA performance. We argue that a unimodal representation can consist of a pan-knowledge representation (given by a pretrained model like BERT (Devlin et al., 2019)) and a knowledge-specific representation (injected from relevant external sources). Specifically, ConKI uses a pretrained BERT model to extract textual pan-knowledge representations and uses two randomly initialized transformer encoders to generate acoustic and visual pan-knowledge representations, respectively. In the meantime, it applies a knowledge injection model named adapter, onto each modality to yield knowledge-specific representations. Both pan- and specific-knowledge representations are fused first within each modality and then across modalities, before the fused features are used for sentiment prediction. We further propose a hierarchical contrastive learning procedure
+
+performed between knowledge types within every single modality, across modalities within each sample, and across samples, to facilitate the learning of these representations in ConKI.
+
+The main contributions of this work can be summarized as follows:
+
+- We propose ConKI, a Contrastive Knowledge Injection framework for multimodal sentiment analysis. ConKI aims to boost model performance through external knowledge injection from other datasets and hierarchical contrastive learning, which is proved better than simply fine-tuning with external datasets.  
+- We propose hierarchical contrastive learning that uses a unified contrastive loss to disentangle the pan-knowledge representations from the specific-knowledge representations since they belong to different knowledge domains and should complement each other.  
+- We conduct extensive experiments on three popular benchmark MSA datasets and attain results that are superior to the existing state-of-the-art MSA baselines on all metrics, demonstrating the effectiveness of the proposed methods in ConKI.
+
+# 2 Related Work
+
+In this section, we discuss related research in multimodal sentiment analysis, knowledge injection, and contrastive learning.
+
+# 2.1 Multimodal Sentiment Analysis
+
+Research on MSA mainly focuses on multimodal fusion and representation learning. For multimodal fusion, existing methods are typically divided into early fusion and late fusion techniques. Early fusion refers to joining multimodal inputs into a single feature before single-model encoding. For example, Williams et al. (2018) concatenate initial input features and then use LSTM to capture the temporal dependencies in the sequence. On the contrary, late fusion learns unimodal representations via separate models and fuses them in a later stage for inference. Zadeh et al. (2017) introduce a tensor fusion network that first encodes each modality with corresponding sub-networks and then models the unimodal, bimodal, and trimodal interactions by a three-fold Cartesian product. For representation learning methods, Hazarika et al. (2020)
+
+propose to project each modality into a modality-invariant and modality-specific representation. Different from the above work, we propose to decompose each modality into two representations based on knowledge types. Both representations can complement each other, leading to a richer unimodal representation.
+
+# 2.2 Knowledge Injection
+
+Injecting knowledge into pretrained language models (PLMs) has been proven to outperform vanilla pretrained models on various NLP tasks (Wei et al., 2021; Wang et al., 2021a; Tian et al., 2020; Ke et al., 2020; Lin et al., 2019; Wang et al., 2021b). Adapters are commonly used as a knowledge injection model plugged outside or inside of PLMs. For instance, Wang et al. (2021a) infuse factual knowledge from Wikidata (Vrandecic and Krötzsch, 2014) and linguistic knowledge from web text to RoBERTa (Liu et al., 2019) via two kinds of adapters. In this work, we build different adapters for different modalities, not limited to text, to learn specific multimodal knowledge from an external dataset for the downstream task. To the best of our knowledge, we are the first to explore knowledge injection in the multimodal domain.
+
+# 2.3 Contrastive Learning
+
+Contrastive learning (CL) aims to learn effective representations such that positive pairs of samples are close while negative pairs of samples are far apart (Liu et al., 2021; Li et al., 2020; Chen et al., 2020a; Khosla et al., 2020; He et al., 2020). Existing works can be divided into two categories: self-supervised CL (Akbari et al., 2021; Chen et al., 2020a,b; He et al., 2020; You et al., 2020; Tao et al., 2020) and supervised CL (Khosla et al., 2020; Mai et al., 2021). The difference between them is whether the label information is used to form positive/negative pairs. For example, Khosla et al. (2020) propose supervised CL to pull samples of the same class together and push samples from different classes away. In our work, we design contrastive pairs in finer granularity. That is, we consider contrasts between knowledge types, between modalities, and across samples.
+
+# 3 Method
+
+In this section, we explain the Contrastive Knowledge Injection framework (ConKI) in detail. The goal of ConKI is to generate pan- and specific
+
+![](images/c4a12855f8534fde42d34e03d8bcf7d087186325c73d9b38bc83248e43895e8e.jpg)  
+Figure 1: The overall architecture of ConKI. The solid and dashed arrows represent the procedure of the main MSA task and the hierarchical contrastive learning subtask, respectively. Inside the contrastive learning procedure, cyan and pink boxes illustrate samples that fall in different sentiment score intervals.
+
+knowledge modality representations via knowledge injection and hierarchical contrastive learning. Knowledge injection intends to obtain knowledge-specific representations that could complement the pan-knowledge representations offered by pretrained models. Hierarchical contrastive learning further optimizes these knowledge-specific and pan-knowledge representations by considering contrasts between knowledge types, modalities, and samples.
+
+# 3.1 Problem Definition
+
+The task of multimodal sentiment analysis (MSA) is to detect sentiments in videos based on multimodal signals, including text  $(t)$ , vision  $(v)$ , and audio  $(a)$  modalities. These signals are represented as sequences of low-level features, i.e.,  $I_{t} \in \mathbb{R}^{l_{t} \times d_{t}}$ ,  $I_{v} \in \mathbb{R}^{l_{v} \times d_{v}}$ , and  $I_{a} \in \mathbb{R}^{l_{a} \times d_{a}}$ , respectively. Here  $l_{m \in \{t, v, a\}}$  denotes the length of the sequence for each modality, while  $d_{m \in \{t, v, a\}}$  denotes the corresponding feature vector dimension. The detail for acquiring these features is described in Appendix B. Given these sequences  $I_{m \in \{t, v, a\}}$ , the primary task is to make accurate predictions on the sentiment intensity by extracting and fusing higher-level multimodal information.
+
+# 3.2 Overall Architecture
+
+Figure 1 shows the overall architecture of ConKI. We first process raw multimodal input to low-level features  $I_{m \in \{t, v, a\}}$  with their corresponding feature extractors and tokenizers. Then we encode  $I_{m}$  into knowledge-specific representations (i.e.,  $A_{m}$ ) generated by some adapters and pan-knowledge representations (i.e.,  $O_{m}$ ) generated by pretrained encoders. The text encoder is from publicly-available pretrained backbones like BERT (Devlin et al., 2019), and the vision/audio encoder is a designed model with random initialization since there is no suitable backbone that is pretrained by the above low-level features. After generating the knowledge-specific and pan-knowledge representations, ConKI is trained simultaneously with two different tasks on the downstream target dataset – the primary MSA regression task and the contrastive learning subtask.
+
+For the MSA task, we concatenate the knowledge-specific representation and pan-knowledge representation of each modality before feeding them into a fully-connected (FC) layer for inner-modality fusion. We then design a fusion network that consists of a concatenation layer and a fusion module for multi-modality fusion, as shown in Figure 2. The fused representations are
+
+![](images/56727cdbf1b02d0110454a39a6e2645096fe0e3485bd4c93e8c49cc6a18637d2.jpg)  
+Figure 2: The Fusion Network. The fusion module marked in the dashed box is used to get the weighted fused embedding.  $\otimes$  means element-wise multiplication.
+
+passed into a multilayer perceptron (MLP) network to produce the sentiment predictions,  $\hat{y}$ .
+
+For the subtask of hierarchical contrastive learning, we carefully construct the negative and positive sample pairs at the knowledge level, modality level, and sample level. The intuition of our pairing policy is as follows. We expect  $A_{m}$  and  $O_{m}$  to capture different knowledge, so we disentangle them and make them complement each other to get richer modality representations by knowledge-level contrasts. Since a video's sentiment is determined by all modalities, we learn the commonalities among the six representations by modality-level contrasts. Besides, videos that express close sentiments should share some correlations. We capture the correlations by sample-level contrasts to help further learn the commonalities among samples under close sentiments. By integrating these hierarchical contrasts, ConKI is able to catch full dynamics among representations which can significantly benefit the main MSA task.
+
+# 3.3 Encoding with Knowledge Injection
+
+We encode each modality into a pan-knowledge representation via the pretrained encoders and a knowledge-specific representation via the adapters.
+
+Pan-knowledge representations. We use the pretrained BERT (Devlin et al., 2019) to encode the input sentence for the text modality. The pooled output vector in the last layer is extracted as the whole sentence representation  $O_{t}$ :
+
+$$
+O _ {t}, H _ {t} = \operatorname {B e r t} \left(I _ {t}; \theta_ {t} ^ {\text {B e r t}}\right), \tag {1}
+$$
+
+where  $H_{t}$  denotes the hidden states of all layers. For audio and vision modalities, we employ encoders of stacked transformer layers (Vaswani et al., 2017) to capture the temporal features  $O_{m}$ :
+
+$$
+O _ {m}, H _ {m} = \operatorname {E n c o d e r} \left(I _ {m}; \theta_ {m} ^ {\text {e n c o d e r}}\right), m \in \{v, a \}. \tag {2}
+$$
+
+![](images/db03cc2e9c0a741caad06051988d68e3f8b151f0a8bf4e80c05e705d9783054c.jpg)  
+Figure 3: Adapter and its connection with the backbone encoder.
+
+Here,  $O_{t}$ ,  $O_{a}$ , and  $O_{v}$  are regarded as three pan-knowledge representations since they mainly contain general knowledge such as the generic facts encoded by BERT (Devlin et al., 2019) pretrained on big text data.
+
+Knowledge-specific representations. We infuse specific domain knowledge from external multimodal sources through knowledge injection models (adapters). The adapter is commonly used in natural language processing (NLP) to enhance existing pretrained language models' knowledge awareness (Wei et al., 2021). The outputs of adapters are taken as knowledge-specific representations. Specifically, the adapter for each modality is plugged outside of the respective pretrained encoder, as shown in Figure 3. It consists of multiple modules with the same sandwich structure: two FC layers with two transformer layers in between. Each module can be inserted before any transformer layers of the backbone models (encoders), e.g., the second and fourth transformer layers in Figure 3. Therefore, each module takes the intermediate layers' hidden states of the pretrained encoder and the output of the previous adapter module as input. The output of the adapter is denoted as  $A_{m}$ , where
+
+$$
+A _ {m} = \operatorname {A d a p t e r} \left(H _ {m}; \theta_ {m} ^ {\text {a d a p t e r}}\right), m \in \{t, v, a \}. \tag {3}
+$$
+
+With the objective of learning specific multimodal sentiment knowledge, we pretrain one adapter for each modality, i.e.,  $\mathrm{Adapter}_t$ ,  $\mathrm{Adapter}_a$  and  $\mathrm{Adapter}_v$ , concurrently using an external dataset while keeping the pretrained encoders frozen. Since the external dataset we select is also from the multimodal sentiment domain, the pretraining task remains the MSA task. That is, we
+
+Algorithm 1: Learning Procedure of ConKI  
+Stage 1: Adapter Pretraining  
+Input: External dataset  $\mathcal{E}$ , its corresponding features  $I_{m}$  and labels  $y$ ,  
+Output: Pretrained adapters  $\{\theta_{m}^{\text{adapter}} \mid m \in \{t, v, a\}\}$   
+for each training epoch do  
+    for batch  $\{(I_t^i, I_v^i, I_a^i)\}_{i=1}^{|B|}$  from  $\mathcal{E}$  do  
+        Encode  $I_{m}^{i}$  to  $O_{m}^{i}$  and  $A_{m}^{i}$  via Eq. (1-3)  
+        Inner-modality fusion:  $F_{m}^{i} = F C([O_{m}^{i}; A_{m}^{i}])$ , where  $[\cdot; \cdot]$  denotes the concatenation of two vectors  
+        Multi-modality fusion:  $F^{i} = F N(F_{t}^{i}, F_{v}^{i}, F_{a}^{i})$ , where  $FN$  is the fusion network  
+        Compute the predictions using  $\hat{y}^{i} = M L P(F^{i})$   
+        Compute  $\mathcal{L}_{\text{task}}$  via Eq. (4)  
+        Update parameters except  $\{\theta_{t}^{\text{Bert}}, \theta_{m}^{\text{encoder}} \mid m \in \{v, a\}\}$   
+    end  
+    Save  $\{\theta_{m}^{\text{adapter}} \mid m \in \{t, v, a\}\}$  when reaching the best validation result  
+end  
+Stage 2: Downstream Fine-tuning  
+Input: Target dataset  $\mathcal{D}$ , its corresponding features  $I_{m}$ , its labels  $y$ , and the pretrained adapters  
+Output: Predictions  $\hat{y}$   
+for each training epoch do  
+    for batch  $\{(I_t^i, I_v^i, I_a^i)\}_{i=1}^{|B|}$  from  $\mathcal{D}$  do  
+        Perform Steps 1 - 4 in Stage 1  
+        Compute  $\mathcal{L}$  via Eq. (5)  
+        Update parameters except  $\{\theta_{m}^{\text{adapter}} \mid m \in \{t, v, a\}\}$   
+    end  
+end
+
+pretrain adapter parts in Figure 1 with only the MSA task on the external dataset, then utilize the pretrained adapters to produce knowledge-specific representations  $A_{m}$  for the downstream target task that includes both the MSA task and the hierarchical contrastive learning subtask. Algorithm 1 summarizes this pretraining procedure of adapters.
+
+# 3.4 Hierarchical Contrastive Learning
+
+In our framework, we propose a hierarchical contrastive learning method to enhance the learned representations by considering the following four aspects in a batch  $B$ :
+
+- For a single video sample  $i$ , all the modalities share common motives of the speaker that determine the overall sentiment. The pan-knowledge representations of different modalities are expected to represent similar meanings and thus need to be pulled closer to each other. And the same applies for knowledge-specific representations. This intuition leads to the construction of intra-sample positive pairs:
+
+$$
+\mathcal {P} _ {1} ^ {i} = \left\{\left(O _ {m} ^ {i}, O _ {n} ^ {i}\right), \left(A _ {m} ^ {i}, A _ {n} ^ {i}\right) \mid \right.
+$$
+
+$$
+m, n \in \{t, v, a \} \& m \neq n \& i \in B \};
+$$
+
+- The pan-knowledge representations and the knowledge-specific representations should be disentangled from each other since they belong to different knowledge domains and are designed to complement each other. This exists inside each sample ( $i$  and  $j$  represent the same sample) as well as across samples in the batch ( $i$  and  $j$  represent two different samples). Therefore, we can build the inter-knowledge negative pairs within a batch:
+
+$$
+\mathcal {N} _ {1} ^ {i} = \left\{\left(O _ {m} ^ {i}, A _ {n} ^ {j}\right) \mid \right.
+$$
+
+$$
+m, n \in \{t, v, a \} \& i, j \in B \};
+$$
+
+- For two arbitrary samples  $i$  and  $j$  having close sentiments, i.e., their sentiment scores can be rounded to the same integer, six representations of sample  $i$  (i.e.,  $O_{m}^{i}$  and  $A_{m}^{i}$ ) should be close to the corresponding representations of sample  $j$  (i.e.,  $O_{n}^{j}$  and  $A_{n}^{j}$ ). Note that the subscripts  $m$  and  $n$  represent the modality for sample  $i$  and  $j$ , respectively. We then form the inter-sample positive pairs as
+
+$$
+\begin{array}{l} \mathcal {P} _ {2} ^ {i} = \{(O _ {m} ^ {i}, O _ {n} ^ {j}), (A _ {m} ^ {i}, A _ {n} ^ {j}) | m, n \in \{t, v, a \} \\ \& r \left(y ^ {i}\right) = r \left(y ^ {j}\right) \& i, j \in B \& i \neq j \}, \\ \end{array}
+$$
+
+where  $y^{i}$  denotes the ground-truth of sample  $i$ , and  $r(\cdot)$  stands for the round function;
+
+Except for the pairs derived from the above three aspects, the remaining pairs with sample  $i$  in the same batch are set as negative pairs  $\mathcal{N}_2^i$ . Please refer to Appendix A.1 for a more detailed pairing policy.
+
+Specifically, our hierarchical contrastive loss  $\mathcal{L}_{con}$  is computed by
+
+$$
+\begin{array}{l} \mathcal {L} _ {c o n} = \sum_ {i \in B} \frac {- 1}{| \mathcal {P} _ {1} ^ {i} \cup \mathcal {P} _ {2} ^ {i} |} \times \\ \sum_ {(p, q) \in \mathcal {P} _ {1} ^ {i} \cup \mathcal {P} _ {2} ^ {i}} \log \frac {f ((p , q))}{\sum_ {(p ^ {\prime} , q ^ {\prime}) \in \mathrm {A l l P a i r s}} f ((p ^ {\prime} , q ^ {\prime}))}, \\ \end{array}
+$$
+
+where,
+
+$$
+f ((p, q)) = \exp (\frac {p}{\| p \| _ {2}} \cdot \frac {q}{\| q \| _ {2}} \cdot \frac {1}{\tau}),
+$$
+
+$$
+\text {A l l P a i r s} = \mathcal {P} _ {1} ^ {i} \cup \mathcal {P} _ {2} ^ {i} \cup \mathcal {N} _ {1} ^ {i} \cup \mathcal {N} _ {2} ^ {i}.
+$$
+
+In the above equation,  $|\mathcal{P}_1^i \cup \mathcal{P}_2^i|$  means the number of positive pairs with sample  $i$  in a batch  $B$ ,  $(\cdot, \cdot)$  denotes a pair in the corresponding set, e.g.,  $(O_t^i, O_v^i)$ , and  $\tau$  is a scalar temperature parameter.
+
+The rationale behind this hierarchical contrastive learning subtask is as follows. First, we capture the commonalities across the three modalities within each knowledge type of each sample to reduce the modality gaps under a shared motive. Second, we model the commonalities across samples of close sentiments within each knowledge type to reduce the sample gaps. Third, we capture the differences between the pan-knowledge representations and the knowledge-specific representations in each sample which results in a complementary effect of the two knowledge types of representations. Last but not least, we capture the differences across samples of different sentiments within each knowledge type in order to learn the dynamics of different sentiment intervals.
+
+# 3.5 Training Procedure
+
+Given the ground truth  $y$  and the predictions  $\hat{y}$ , we can calculate the main MSA task loss by the mean squared error:
+
+$$
+\mathcal {L} _ {\text {t a s k}} = \frac {1}{| B |} \sum_ {i} ^ {| B |} \left(\hat {y} ^ {i} - y ^ {i}\right) ^ {2}, \tag {4}
+$$
+
+where  $|B|$  is the number of samples in a batch.
+
+ConKI adopts the learning regime of pretraining followed by fine-tuning. We first pretrain the adapters in ConKI with  $\mathcal{L}_{task}$  using an external dataset while fixing the model parameters of the pretrained backbones, considering ConKI only encodes specific knowledge in adapters which have
+
+<table><tr><td>Dataset</td><td>#Train</td><td>#Valid</td><td>#Test</td><td>#Total</td></tr><tr><td>CMU-MOSI</td><td>1284</td><td>229</td><td>686</td><td>2199</td></tr><tr><td>CMU-MOSEI</td><td>16326</td><td>1871</td><td>4659</td><td>22856</td></tr><tr><td>SIMS</td><td>1368</td><td>456</td><td>457</td><td>2281</td></tr></table>
+
+Table 1: The statistics of CMU-MOSI, CMU-MOSEI and SIMS.
+
+much fewer trainable parameters compared to backbones. Then we fine-tune ConKI with the downstream target dataset by optimizing the overall loss  $\mathcal{L}$ :
+
+$$
+\mathcal {L} = \mathcal {L} _ {\text {t a s k}} + \lambda \mathcal {L} _ {\text {c o n}}, \tag {5}
+$$
+
+where  $\lambda$  is a hyperparameter that balances the MSA task loss and the hierarchical contrastive loss. Algorithm 1 shows the full training procedure of ConKI.
+
+# 4 Experiments
+
+In this section, we present some experimental details, including datasets, evaluation metrics, baseline models, and experimental results. The implementation details are shown in Appendix B.
+
+# 4.1 Datasets and Metrics
+
+We conduct experiments on three publicly available benchmark datasets in MSA: CMU-MOSI (Zadeh et al., 2016), CMU-MOSEI (Zadeh et al., 2018) and SIMS (Yu et al., 2020). Table 1 shows the statistics of the datasets. Appendix C describes the details of these datasets.
+
+Following the previous works (Sun et al., 2020; Rahman et al., 2020; Hazarika et al., 2020; Yu et al., 2021; Mai et al., 2021; Han et al., 2021; Yu et al., 2020), we report our experimental results in two forms: regression and classification. For regression, we report mean absolute error (MAE) and Pearson correlation (Corr). For classification, we report binary classification accuracy (Acc-2) and F1 score. Specifically, for CMU-MOSI and CMU-MOSEI datasets, we calculate Acc-2 and F1 scores in negative/positive (zero excluded) and non-negative/positive (zero included) settings as well as seven-class classification accuracy (Acc-7) which shows the percentage of predictions that correctly classified into the same interval of seven intervals between  $-3$  and  $+3$ . Higher values indicate better performance for all metrics except for MAE.
+
+# 4.2 Baselines
+
+We compare ConKI with the following state-of-the-art baseline models in MSA: TFN (Zadeh et al.,
+
+<table><tr><td rowspan="2">Models*</td><td colspan="5">CMU-MOSI</td><td colspan="5">CMU-MOSEI</td></tr><tr><td>MAE</td><td>Corr</td><td>Acc-7</td><td>Acc-2</td><td>F1</td><td>MAE</td><td>Corr</td><td>Acc-7</td><td>Acc-2</td><td>F1</td></tr><tr><td>TFN†</td><td>0.901</td><td>0.698</td><td>34.9</td><td>-/80.8</td><td>-/80.7</td><td>0.593</td><td>0.700</td><td>50.2</td><td>-/82.5</td><td>-/82.1</td></tr><tr><td>LMF†</td><td>0.917</td><td>0.695</td><td>33.2</td><td>-/82.5</td><td>-/82.4</td><td>0.623</td><td>0.677</td><td>48.0</td><td>-/82.0</td><td>-/82.1</td></tr><tr><td>MulT†</td><td>0.861</td><td>0.711</td><td>-</td><td>81.5/84.1</td><td>80.6/83.9</td><td>0.580</td><td>0.703</td><td>-</td><td>-/82.5</td><td>-/82.3</td></tr><tr><td>ICCN†</td><td>0.862</td><td>0.714</td><td>39.0</td><td>-/83.0</td><td>-/83.0</td><td>0.565</td><td>0.713</td><td>51.6</td><td>-/84.2</td><td>-/84.2</td></tr><tr><td>MISA†</td><td>0.804</td><td>0.764</td><td>-</td><td>80.79/82.10</td><td>80.77/82.03</td><td>0.568</td><td>0.724</td><td>-</td><td>82.59/84.23</td><td>82.67/83.97</td></tr><tr><td>MAG-BERT†</td><td>0.727</td><td>0.781</td><td>43.62</td><td>82.37/84.43</td><td>82.50/84.61</td><td>0.543</td><td>0.755</td><td>52.67</td><td>82.51/84.82</td><td>82.77/84.71</td></tr><tr><td>Self-MM†</td><td>0.712</td><td>0.795</td><td>45.79</td><td>82.54/84.77</td><td>82.68/84.91</td><td>0.529</td><td>0.767</td><td>53.46</td><td>82.68/84.96</td><td>82.95/84.93</td></tr><tr><td>HyCon‡</td><td>0.713</td><td>0.790</td><td>46.6</td><td>-/85.2</td><td>-/85.1</td><td>0.601</td><td>0.776</td><td>52.8</td><td>-/85.4</td><td>-/85.6</td></tr><tr><td>MMIM†</td><td>0.700</td><td>0.800</td><td>46.65</td><td>84.14/86.06</td><td>84.00/85.98</td><td>0.526</td><td>0.772</td><td>54.24</td><td>82.24/85.97</td><td>82.66/85.94</td></tr><tr><td>ConKI</td><td>0.681</td><td>0.816</td><td>48.43</td><td>84.37/86.13</td><td>84.33/86.13</td><td>0.529</td><td>0.782</td><td>54.25</td><td>82.73/86.25</td><td>83.08/86.15</td></tr></table>
+
+2017), LMF (Liu and Shen, 2018), MulT (Tsai et al., 2019), ICCN (Sun et al., 2020), MISA (Hazarika et al., 2020), MAG-BERT (Rahman et al., 2020), Self-MM (Yu et al., 2021), HyCon (Mai et al., 2021), and MMIM (Han et al., 2021). The details of these baseline models are shown in Appendix D.
+
+# 4.3 Results
+
+In accordance with previous work, we run our model five times under the same hyper-parameter settings and report the average performance of all metrics in Table 2 and Table 3. We can observe from these tables that ConKI yields better or competitive results to a range of baseline models on CMU-MOSI, CMU-MOSEI, and SIMS. Specifically, ConKI outperforms all state-of-the-art baseline models in all metrics on CMU-MOSI and SIMS as well as in Corr, Acc-7, Acc-2, F1 scores on CMU-MOSEI. It also achieves closed performance to the best baseline model in MAE on CMU-MOSEI.
+
+Table 2: Results on CMU-MOSI and CMU-MOSEI. In Acc-2 and F1, the left of the “/” corresponds to “negative/non-negative” and the right corresponds to “negative/positive”. *: all models use BERT as the text encoder; †: from (Han et al., 2021); ‡: from (Mai et al., 2021). Best results are marked in bold.  
+
+<table><tr><td>Models</td><td>MAE</td><td>Corr</td><td>Acc-2</td><td>F1</td></tr><tr><td>TFN</td><td>0.488</td><td>0.496</td><td>75.27</td><td>75.56</td></tr><tr><td>LMF</td><td>0.487</td><td>0.502</td><td>75.36</td><td>75.78</td></tr><tr><td>MulT</td><td>0.485</td><td>0.504</td><td>75.62</td><td>75.84</td></tr><tr><td>MISA</td><td>0.472</td><td>0.542</td><td>75.49</td><td>75.85</td></tr><tr><td>MAG-BERT</td><td>0.553</td><td>0.242</td><td>71.43</td><td>63.68</td></tr><tr><td>Self-MM</td><td>0.458</td><td>0.535</td><td>77.37</td><td>77.54</td></tr><tr><td>MMIM</td><td>0.607</td><td>-</td><td>69.37</td><td>58.00</td></tr><tr><td>ConKI</td><td>0.454</td><td>0.542</td><td>77.94</td><td>78.17</td></tr></table>
+
+Table 3: Results on SIMS. All baseline model codes are from https://github.com/thuiar/MMSA.
+
+It is notable that the MAE of ConKI on CMU-MOSI outperforms the best baseline model MMIM by around 0.02, which shows ConKI is able to learn effective representations for the MSA task since MAE is the most commonly used evaluation metric in regression tasks. ConKI also presents an excellent performance in the Corr scores on both CMU-MOSI and CMU-MOSEI datasets. The possible reasoning behind this excellent performance is that ConKI uses contrastive learning for recognizing the samples under different sentiments, which could lead to effective ranking results among samples and thus produce a higher Corr score (Swinscow et al., 2002).
+
+Furthermore, Acc-7 of ConKI on CMU-MOSI surpasses the best baseline by 1.78. Though performing classification, especially seven-class classification, is difficult in a regression task, ConKI successfully leverages the contrasts across samples that are classified into seven intervals (by the round function described in Section 3.4) to model the sample dynamics, which brings a great improvement to Acc-7 and Acc-2, demonstrating the efficacy of ConKI in representation learning for MSA. In addition, ConKI shows excellent F1 scores on all datasets, which endorse its potential in real-world applications since F1 is valuable for evaluating imbalanced datasets.
+
+# 4.4 Ablation Study
+
+We first conduct an ablation study about modalities, as shown in Table 4. We can observe that the inclusion of all three modalities significantly improves the performance of ConKI.
+
+To show the benefits of the proposed knowledge
+
+<table><tr><td>Models</td><td>MAE</td><td>Corr</td><td>Acc-7</td><td>Acc-2</td><td>F1</td></tr><tr><td>V+A</td><td>1.408</td><td>0.248</td><td>18.72</td><td>55.71/54.33</td><td>54.37/53.24</td></tr><tr><td>T+A</td><td>0.700</td><td>0.799</td><td>48.22</td><td>82.45/84.18</td><td>82.38/84.16</td></tr><tr><td>T+V</td><td>0.718</td><td>0.798</td><td>45.45</td><td>82.97/84.88</td><td>82.89/84.86</td></tr><tr><td>ConKI</td><td>0.681</td><td>0.816</td><td>48.43</td><td>84.37/86.13</td><td>84.33/86.13</td></tr></table>
+
+Table 4: Ablation results when using different modalities.  
+
+<table><tr><td>Models</td><td>MAE</td><td>Corr</td><td>Acc-7</td><td>Acc-2</td><td>F1</td></tr><tr><td>ConKI</td><td>0.681</td><td>0.816</td><td>48.43</td><td>84.37/86.13</td><td>84.33/86.13</td></tr><tr><td>w/o C1</td><td>0.734</td><td>0.794</td><td>43.56</td><td>82.39/84.21</td><td>82.35/84.22</td></tr><tr><td>w/o C2</td><td>0.753</td><td>0.789</td><td>43.50</td><td>82.16/83.84</td><td>82.15/83.89</td></tr><tr><td>w/o C3</td><td>0.710</td><td>0.811</td><td>44.75</td><td>83.29/84.82</td><td>83.23/84.80</td></tr><tr><td>w/o C4</td><td>0.683</td><td>0.815</td><td>48.05</td><td>84.23/86.01</td><td>84.16/85.99</td></tr><tr><td>w/o N1</td><td>0.689</td><td>0.812</td><td>47.90</td><td>83.88/85.76</td><td>83.8/85.74</td></tr></table>
+
+injection and hierarchical contrastive learning in ConKI, we conduct a series of ablation experiments on CMU-MOSI, as shown in Table 5 and Table 6. ConKI mainly includes four components: the use of the external dataset (C1), adapters for knowledge injection (C2), pretrained encoders for panknowledge (C3), and hierarchical contrastive learning (C4). Table 5 shows that C1 provides advantages by comparing w/o C1 and ConKI. Similarly, C2 provides benefits by comparing w/o C2 and ConKI. C3 is beneficial by comparing w/o C3 and ConKI. C4 is beneficial by comparing w/o C4 and ConKI.
+
+Since the spotlight in our hierarchical contrastive learning is the contrasts between knowledge types, we also compare our model with the model w/o  $\mathcal{N}_1$  trained with  $\mathcal{L}_{con}$  but without negative pairs  $\mathcal{N}_1$ , i.e., without disentangling the pan knowledge and specific knowledge. We can conclude that learning differentiated pan-knowledge and knowledge-specific representations is essential in our hierarchical contrastive learning. To better understand the learned pan- and specific-knowledge representations by our hierarchical contrastive learning, we visualize and analyze these representations in Appendix A.2.
+
+To further examine if our performance gain is from the external dataset instead of the proposed knowledge injection and contrastive learning technique, we compare our model with the state-of-the-art baseline models which are fine-tuned by the external dataset. The results from Table 6 show that ConKI still outperforms those baseline mod
+
+Table 5: Ablation results of ConKI's components on CMU-MOSI.  
+
+<table><tr><td>Models</td><td>MAE</td><td>Corr</td><td>Acc-7</td><td>Acc-2</td><td>F1</td></tr><tr><td>MISA</td><td>0.711</td><td>0.804</td><td>44.96</td><td>82.04/83.99</td><td>81.98/84.0</td></tr><tr><td>Self-MM</td><td>0.712</td><td>0.798</td><td>45.51</td><td>83.59/85.18</td><td>83.47/85.12</td></tr><tr><td>MMIM</td><td>0.716</td><td>0.791</td><td>45.42</td><td>81.81/83.54</td><td>81.67/83.46</td></tr><tr><td>ConKI</td><td>0.681†</td><td>0.816†</td><td>48.43†</td><td>84.37†/86.13†</td><td>84.33†/86.13†</td></tr></table>
+
+Table 6: Ablation results when introducing CMU-MOSEI as an external dataset on CMU-MOSI.  $\dagger$  means the corresponding result is significantly better than SOTA with  $p$ -value  $< 0.05$  based on paired  $t$ -test.
+
+els even though they are trained with the external dataset.
+
+Therefore, our gain from ConKI is not solely from adding more data, but from knowledge injection with multi-step transfer learning. Considering the size of CMU-MOSEI is much larger than CMUMOSI, injecting CMU-MOSEI's knowledge into CMU-MOSI thus has more effects on the downstream task than injecting CMU-MOSI into CMUMOSEI, as shown in Table 2.
+
+# 5 Conclusion
+
+In this paper, we present ConKI, a Contrastive Knowledge Injection framework for multimodal sentiment analysis, which learns knowledge-specific representations along with pan-knowledge representations via knowledge injection and hierarchical contrastive learning. ConKI utilizes the pretrained encoders to obtain pan-knowledge representations while generating knowledge-specific representations based on injected adapters that are trained on an external knowledge source. With the specific knowledge, ConKI is able to produce more accurate sentiment predictions than solely using the pan-knowledge representations. To further improve the learning of these representations, we specifically design a hierarchical contrastive learning procedure taking into account the contrasts between knowledge types within each modality, across modalities within one sample, and across samples. Experimental results on three benchmark datasets show that ConKI outperforms all state-of-the-art methods on a range of performance metrics.
+
+# Limitations
+
+Our research presents an initial step toward a knowledge injection framework for MSA and still has some limitations to be tackled in the future. Firstly, we can learn more disentangled representations by carefully selecting contrastive pairs for further improvement. Secondly, it will be interest-
+
+ing if we extend our method with multiple external sources that come from different knowledge domains.
+
+# References
+
+Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, and Boqing Gong. 2021. Vatt: Transformers for multimodal self-supervised learning from raw video, audio and text. Advances in Neural Information Processing Systems, 34.  
+Minghai Chen, Sen Wang, Paul Pu Liang, Tadas Baltrusaitis, Amir Zadeh, and Louis-Philippe Morency. 2017. Multimodal sentiment analysis with word-level fusion and reinforcement learning. In Proceedings of the 19th ACM international conference on multimodal interaction, pages 163-171.  
+Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020a. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR.  
+Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. 2020b. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems, 33:22243-22255.  
+Gilles Degottex, John Kane, Thomas Drugman, Tuomo Raitio, and Stefan Scherer. 2014. Covarep — a collaborative voice analysis repository for speech technologies. In 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 960–964.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In *NAACL*.  
+Paul Ekman and Erika L. Rosenberg. 2005. What the face reveals: basic and applied studies of spontaneous expression using the facial action coding system (facs).  
+Wei Han, Hui Chen, and Soujanya Poria. 2021. Improving multimodal fusion with hierarchical mutual information maximization for multimodal sentiment analysis. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9180-9192.  
+Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. 2020. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122-1131.
+
+Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. 2020. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738.  
+Pei Ke, Haozhe Ji, Siyang Liu, Xiaoyan Zhu, and Minlie Huang. 2020. Sentilare: Sentiment-aware language representation learning with linguistic knowledge. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6975-6988.  
+Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning. Advances in Neural Information Processing Systems, 33:18661-18673.  
+Anne Lauscher, Olga Majewska, Leonardo FR Ribeiro, Iryna Gurevych, Nikolai Rozanov, and Goran Glavaš. 2020. Common sense or world knowledge? investigating adapter-based knowledge injection into pretrained transformers. Association for Computational Linguistics.  
+Junnan Li, Pan Zhou, Caiming Xiong, and Steven Hoi. 2020. Prototypical contrastive learning of unsupervised representations. In International Conference on Learning Representations.  
+Bill Yuchen Lin, Xinyue Chen, Jamin Chen, and Xiang Ren. 2019. Kagnet: Knowledge-aware graph networks for commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2829-2839.  
+Chen Liu, Yanwei Fu, C. Xu, Siqian Yang, Jilin Li, Chengjie Wang, and Li Zhang. 2021. Learning a few-shot embedding model with contrastive learning. In AAAI.  
+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.  
+Zhun Liu and Ying Shen. 2018. Efficient low-rank multimodal fusion with modality-specific factors. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Long Papers).  
+Sijie Mai, Ying Zeng, Shuangjia Zheng, and Haifeng Hu. 2021. Hybrid contrastive learning of tri-modal representation for multimodal sentiment analysis. arXiv preprint arXiv:2109.01797.  
+Prem Melville, Wojciech Gryc, and Richard D. Lawrence. 2009. Sentiment analysis of blogs by combining lexical knowledge with text classification. In Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '09, page 1275-1284, New York, NY, USA. Association for Computing Machinery.
+
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32.  
+Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, Amir Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. 2020. Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2020, page 2359. NIH Public Access.  
+Zhongkai Sun, Prathusha Sarma, William Sethares, and Yingyu Liang. 2020. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8992-8999.  
+Thomas Douglas Victor Swinscow, Michael J Campbell, et al. 2002. Statistics at square one. Bmj London.  
+Li Tao, Xueting Wang, and Toshihiko Yamasaki. 2020. Self-supervised video representation learning using inter-intra contrastive framework. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2193-2201.  
+Hao Tian, Can Gao, Xinyan Xiao, Hao Liu, Bolei He, Hua Wu, Haifeng Wang, and Feng Wu. 2020. Skep: Sentiment knowledge enhanced pre-training for sentiment analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4067-4076.  
+Yao-Hung Hubert Tsai, Shaojie Bai, Paul Pu Liang, J Zico Kolter, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2019. Multimodal transformer for unaligned multimodal language sequences. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2019, page 6558. NIH Public Access.  
+Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of machine learning research, 9(11).  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Denny Vrandecic and Markus Krötzsch. 2014. Wikidata: a free collaborative knowledgebase. Communications of the ACM, 57(10):78-85.  
+Ruize Wang, Duyu Tang, Nan Duan, Zhongyu Wei, Xuan-Jing Huang, Jianshu Ji, Guihong Cao, Daxin Jiang, and Ming Zhou. 2021a. K-adapter: Infusing knowledge into pre-trained models with adapters. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1405–1418.
+
+Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. 2021b. Kepler: A unified model for knowledge embedding and pre-trained language representation. Transactions of the Association for Computational Linguistics, 9:176-194.  
+Xiaokai Wei, Shen Wang, Dejiao Zhang, Parminder Bhatia, and Andrew Arnold. 2021. Knowledge enhanced pretrained language models: A comprehensive survey. arXiv preprint arXiv:2110.08455.  
+Jennifer Williams, Steven Kleinegesse, Ramona Comanescu, and Oana Radu. 2018. Recognizing emotions in video using multimodal DNN feature fusion. In Proceedings of Grand Challenge and Workshop on Human Multimodal Language (Challenge-HML), pages 11-19, Melbourne, Australia. Association for Computational Linguistics.  
+Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. Advances in Neural Information Processing Systems, 33:5812-5823.  
+Wenmeng Yu, Hua Xu, Fanyang Meng, Yilin Zhu, Yixiao Ma, Jiele Wu, Jiyun Zou, and Kaicheng Yang. 2020. CH-SIMS: A Chinese multimodal sentiment analysis dataset with fine-grained annotation of modality. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3718-3727, Online. Association for Computational Linguistics.  
+Wenmeng Yu, Hua Xu, Ziqi Yuan, and Jiele Wu. 2021. Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 10790-10797.  
+Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. 2017. Tensor fusion network for multimodal sentiment analysis. In EMNLP.  
+Amir Zadeh, Paul Pu Liang, Soujanya Poria, E. Cambria, and Louis-Philippe Morency. 2018. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In ACL.  
+Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis Philippe Morency. 2016. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259.
+
+# A Hierarchical Contrastive Learning
+
+# A.1 Pairing Policy
+
+To further elaborate on the pairing policy of our hierarchical contrastive learning, we show
+
+![](images/d7d4dc9546d2b5b719df4cdeca907bbab901138450d850c134f95c8c0020efec.jpg)  
+Figure 4: Pairing example of three samples where sample 1 and sample 2 are in the same sentiment interval while sample 3 is in a different sentiment interval. Grey cells with "1" stand for the positive pairs, and white cells with "0" represent the negative pairs.
+
+an example batch that consists of three samples  $\left(\{O_{t}^{i}, O_{v}^{i}, O_{a}^{i}, A_{t}^{i}, A_{v}^{i}, A_{a}^{i}\}_{i=1,2,3} \in B\right)$  where sample 1 and sample 2 belong to the same sentiment interval while sample 3 falls in a different sentiment interval in Figure 4. In this figure, the "1"s in the heatmap represent the positive pairs of the row vectors and column vectors. The "0"s represent the negative pairs of each two vectors.
+
+From this figure, we can get the intra-sample positive pairs as the "1"s with red borders:
+
+$$
+\begin{array}{l} \mathcal {P} _ {1} ^ {B} = \left\{\left(O _ {t} ^ {1}, O _ {v} ^ {1}\right), \left(O _ {t} ^ {1}, O _ {a} ^ {1}\right), \left(O _ {a} ^ {1}, O _ {v} ^ {1}\right), \right. \\ \left(A _ {t} ^ {1}, A _ {v} ^ {1}\right), \left(A _ {t} ^ {1}, A _ {a} ^ {1}\right), \left(A _ {a} ^ {1}, A _ {v} ^ {1}\right), \\ \left(O _ {t} ^ {2}, O _ {v} ^ {2}\right), \left(O _ {t} ^ {2}, O _ {a} ^ {2}\right), \left(O _ {a} ^ {2}, O _ {v} ^ {2}\right), \\ \left(A _ {t} ^ {2}, A _ {v} ^ {2}\right), \left(A _ {t} ^ {2}, A _ {a} ^ {2}\right), \left(A _ {a} ^ {2}, A _ {v} ^ {2}\right), \\ (O _ {t} ^ {3}, O _ {v} ^ {3}), (O _ {t} ^ {3}, O _ {a} ^ {3}), (O _ {a} ^ {3}, O _ {v} ^ {3}), \\ \left. \left(A _ {t} ^ {3}, A _ {v} ^ {3}\right), \left(A _ {t} ^ {3}, A _ {a} ^ {3}\right), \left(A _ {a} ^ {3}, A _ {v} ^ {3}\right) \right\}; \\ \end{array}
+$$
+
+- We represent the inter-knowledge negative pairs  $\mathcal{N}_1^B$  as the "0"s in the blue zone;  
+- Since sample 1 and sample 2 have close sentiment scores, we form the inter-sample positive pairs as the "1"s without red borders:
+
+$$
+\begin{array}{l} \mathcal {P} _ {2} ^ {B} = \left\{\left(O _ {t} ^ {1}, O _ {t} ^ {2}\right), \left(O _ {t} ^ {1}, O _ {v} ^ {2}\right), \left(O _ {t} ^ {1}, O _ {a} ^ {2}\right), \right. \\ (O _ {v} ^ {1}, O _ {t} ^ {2}), (O _ {v} ^ {1}, O _ {v} ^ {2}), (O _ {v} ^ {1}, O _ {a} ^ {2}), \\ \left(O _ {a} ^ {1}, O _ {t} ^ {2}\right), \left(O _ {a} ^ {1}, O _ {v} ^ {2}\right), \left(O _ {a} ^ {1}, O _ {a} ^ {2}\right), \\ \left(A _ {t} ^ {1}, A _ {t} ^ {2}\right), \left(A _ {t} ^ {1}, A _ {v} ^ {2}\right), \left(A _ {t} ^ {1}, A _ {a} ^ {2}\right), \\ \left(A _ {v} ^ {1}, A _ {t} ^ {2}\right), \left(A _ {v} ^ {1}, A _ {v} ^ {2}\right), \left(A _ {v} ^ {1}, A _ {a} ^ {2}\right), \\ \left. \left(A _ {a} ^ {1}, A _ {t} ^ {2}\right), \left(A _ {a} ^ {1}, A _ {v} ^ {2}\right), \left(A _ {a} ^ {1}, A _ {a} ^ {2}\right) \right\}; \\ \end{array}
+$$
+
+- The remaining white cells with "0" show the negative pairs in  $\mathcal{N}_2^B$  which aim to push sample 3 away from sample 1 and sample 2 because they have different sentiments.
+
+# A.2 Visualization of Modality Representations
+
+The motivation for us to propose hierarchical contrastive learning into ConKI is that we think modalities will be closer to each other within one sample and across samples in the same sentiment interval and will be far away across samples in different intervals while two knowledge contained in one modality will also be different. We use t-SNE (Van der Maaten and Hinton, 2008) to visualize the distributions of the six representations learned by ConKI with and without hierarchical contrastive learning, as shown in Figure 5.
+
+Though we divide all samples into seven intervals to perform contrastive learning, we take samples of two intervals from the testing set to show the learned representations before and after contrastive learning due to the simplicity of the visualization. From Figure 5 (a), we can easily observe that some of the representations such as the pan-knowledge representations in light blue and the knowledge-specific representations in dark green of samples in two different intervals overlap extremely with each other.
+
+In contrast, these overlapping representations are pushed further in Figure 5 (b) due to sample-level contrasts. It is also obvious that the three knowledge-specific representations of samples in the same interval, e.g.,  $A_{t}$ ,  $A_{v}$ ,  $A_{a}$  of Interval 2 in dark colors and star shape become closer because of both modality-level and sample-level contrasts. Moreover, the distance between the knowledge-specific representations and the pan-knowledge representations, e.g.,  $A_{v}$  in dark green and  $O_{v}$  in light green of Interval 2, becomes larger in Figure 5 (b) by knowledge-level contrasts. All of these indicate ConKI is able to perform desired contrastive learning for learning better representations that help improve the performance, even in the generalized scenario, i.e., in the testing set.
+
+# B Implementation Details
+
+We use unaligned raw data in all experiments as the previous works (Yu et al., 2021; Han et al., 2021) for fair comparisons. For audio and video modalities, two commonly-used toolkits (COVAREP (Degottex et al., 2014) and Facial Action Coding System (FACS) (Ekman and Rosenberg, 2005)) act as
+
+![](images/cf5223a05a2421066797dc45888e748d23f8c3a47c0f06d805dd31db46450126.jpg)  
+(a)
+
+![](images/3b7cf5adfcdda4bfadb1d23fde05ad80d625e76909be5be6007ce8f9efaf0cd0.jpg)  
+(b)  
+Figure 5: The visualization of the six decomposed representations of samples in the same sentiment interval and different intervals in (a) w/o h-CL; (b) ConKI. In each subfigure, light yellow, light blue, and light green represent pan-knowledge representations of text, audio, and video modalities, respectively, while dark yellow, dark blue, and dark green represent knowledge-specific representations accordingly. Each point or star stands for a sample in Interval 1 or Interval 2.
+
+the feature extractors, respectively. We use uncased 12-layers BERT pretrained model<sup>1</sup> as the text encoder, and two 2-layer transformers as the video and audio encoders, respectively.  $\text{Adapter}_t$  has three modules inserted before the first, sixth, and eleventh layers of BERT sequentially.  $\text{Adapter}_v$  and  $\text{Adapter}_a$  have one module inserted before the second layer of the corresponding encoder. We use CMU-MOSEI as the external dataset for CMU-MOSI and SIMS while using CMU-MOSI for CMU-MOSEI. During the pretraining, the learning rate is set to  $5e - 5$  and we train for 10 epochs with one epoch for a linear warm-up scheduler. During the fine-tuning of CMU-MOSI and SIMS, the learning rates for encoders and other components are set to  $5e - 6$  and  $1e - 6$  respectively with a weight decay 0.001. The temperature parameter  $\tau$  is set to 0.07 and  $\lambda$  is set to 0.01 after grid-search. We fine tune for 200 epochs with batch size 32. For the fine-tuning of CMU-MOSEI, the learning rates for the text encoder and others are  $5e - 6$  and  $5e - 5$ , respectively.  $\lambda$  for CMU-MOSEI is set to 0.001. The best performance on the validation dataset is used for testing. We implement our experiments using PyTorch (Paszke et al., 2019) on an Nvidia RTX 2080Ti GPU.
+
+# C Datasets
+
+CMU-MOSI is a popular benchmark dataset collected from YouTube. It contains 2,199 video clips sliced from 93 videos where a speaker shares opinions on topics such as movies. Each video is an-
+
+notated with sentiment scores ranging from  $-3$  (strongly negative) to  $+3$  (strongly positive). CMUMOSEI is the largest MSA dataset that has greater diversity in speakers, topics, and annotations. It contains 22,856 annotated video segments from 1,000 distinct speakers and 250 topics. Each clip also has sentiment scores between  $[-3, +3]$ . SIMS is a Chinese MSA dataset that contains 2,281 refined video segments. Each sample has one multimodal label and three unimodal labels, with sentiment scores from  $-1$  to  $+1$ . We translate the Chinese text into English so that we can inject knowledge from English MSA datasets into SIMS. For fair comparisons, all baseline models use the English version to evaluate the performance.
+
+# D Baseline Models
+
+TFN. The Tensor Fusion Network (TFN) (Zadeh et al., 2017) encodes three modalities with corresponding embedding subnetworks and uses outer-product to model the unimodal, bimodal, and trimodal interactions as the fusion results.
+
+LMF. The Low-rank Multimodal Fusion (LMF) (Liu and Shen, 2018) utilizes low-rank tensors to improve efficiency of multimodal fusion.
+
+MulT. The Multimodal Transformer (MulT) (Tsai et al., 2019) proposes directional pairwise cross-modal attention that adapts one modality into another for multimodal fusion.
+
+ICCN. The Interaction Canonical Correlation Network (ICCN) (Sun et al., 2020) learns text-based audio and text-based video features by op
+
+timizing canonical loss. These features are concatenated with the text features for downstream classifiers such as logistic regression.
+
+MISA. The Modality-Invariant and -Specific Representations (MISA) (Hazarika et al., 2020) designs a multitask loss including task prediction loss, reconstruction loss, similarity loss, and difference loss to learn modality-invariant and modality-specific representations.
+
+MAG-BERT. The Multimodal Adaptation Gate for Bert (MAG-BERT) (Rahman et al., 2020) builds an alignment gate that allows audio and video information to leak into the BERT model for multimodal fusion.
+
+Self-MM. The Self-Supervised Multitask Learning (Self-MM) (Yu et al., 2021) proposes a label generation module based on self-supervised learning to obtain unimodal supervision. Then they joint train the multimodal and unimodal tasks for better fusion results.
+
+HyCon. The Hybrid Contrastive Learning (HyCon) (Mai et al., 2021) performs intra- and inter-modal contrastive learning as well as semicontroactive learning within a modality to explore cross-modal interactions.
+
+MMIM. MultiModal InfoMax (MMIM) (Han et al., 2021) maximizes the mutual information in unimodal input pairs as well as between multimodal fusion result and unimodal input to aid the main MSA task.
+
+A For every submission:
+
+A1. Did you describe the limitations of your work?
+
+See Limitations section.
+
+A2. Did you discuss any potential risks of your work?
+
+Not applicable. Left blank.
+
+A3. Do the abstract and introduction summarize the paper's main claims?
+
+Left blank.
+
+A4. Have you used AI writing assistants when working on this paper?
+
+Left blank.
+
+B Did you use or create scientific artifacts?
+
+Not applicable. Left blank.
+
+B1. Did you cite the creators of artifacts you used?
+
+Not applicable. Left blank.
+
+B2. Did you discuss the license or terms for use and / or distribution of any artifacts?
+
+Not applicable. Left blank.
+
+B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)?
+
+Not applicable. Left blank.
+
+B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it?
+
+Not applicable. Left blank.
+
+B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.?
+
+Not applicable. Left blank.
+
+B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be.
+
+See Table 1 and Appendix C.
+
+C Did you run computational experiments?
+
+See Section 4.
+
+C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used?
+
+See Appendix B.
+
+C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? See Appendix B.  
+C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? See Section 4.3.  
+C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? See Appendix B.
+
+D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.
+
+D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? Not applicable. Left blank.  
+D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? Not applicable. Left blank.  
+D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? Not applicable. Left blank.  
+D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? Not applicable. Left blank.  
+D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? Not applicable. Left blank.

2023/ConKI_ Contrastive Knowledge Injection for Multimodal Sentiment Analysis/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcdd422c25e1a498003e453e09c2ff217397e10beca3aa94f1c9a2ead5964c82
+size 544039

2023/Concept2Box_ Joint Geometric Embeddings for Learning Two-View Knowledge Graphs/a7ac92f6-3015-4bc3-abcd-68379068a363_content_list.json

@@ -0,0 +1,2118 @@
+[
+    {
+        "type": "text",
+        "text": "Concept2Box: Joint Geometric Embeddings for Learning Two-View Knowledge Graphs",
+        "text_level": 1,
+        "bbox": [
+            147,
+            79,
+            850,
+            118
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Zijie Huang $^{1}$ , Daheng Wang $^{2}$ , Binxuan Huang $^{2}$ , Chenwei Zhang $^{2}$ , Jingbo Shang $^{4}$ ,",
+        "bbox": [
+            147,
+            124,
+            853,
+            143
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yan Liang $^{2}$ , Zhengyang Wang $^{2}$ , Xian Li $^{2}$ , Christos Faloutsos $^{3}$ , Yizhou Sun $^{1}$ , Wei Wang $^{1}$",
+        "bbox": [
+            121,
+            153,
+            873,
+            170
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "<sup>1</sup>University of California, Los Angeles, CA, USA <sup>2</sup>Amazon.com Inc, CA, USA",
+        "bbox": [
+            174,
+            171,
+            820,
+            186
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{3}$ Carnegie Mellon University, PA, USA  $^{4}$ University of California, San Diego, CA, USA",
+        "bbox": [
+            144,
+            187,
+            850,
+            203
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{zijiehuang,yzsun,weiwang}@cs.ucla.edu,",
+        "bbox": [
+            267,
+            205,
+            732,
+            219
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{dahenwan,binxuan,cwzhang,ynliang,zhengywa,xianlee}@amazon.com, jshang@ucsd.edu,christos@cs.cmu.edu",
+        "bbox": [
+            122,
+            221,
+            875,
+            253
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            253,
+            339,
+            266
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Knowledge graph embeddings (KGE) have been extensively studied to embed large-scale relational data for many real-world applications. Existing methods have long ignored the fact many KGs contain two fundamentally different views: high-level ontology-view concepts and fine-grained instance-view entities. They usually embed all nodes as vectors in one latent space. However, a single geometric representation fails to capture the structural differences between two views and lacks probabilistic semantics towards concepts' granularity. We propose Concept2Box, a novel approach that jointly embeds the two views of a KG using dual geometric representations. We model concepts with box embeddings, which learn the hierarchy structure and complex relations such as overlap and disjoint among them. Box volumes can be interpreted as concepts' granularity. Different from concepts, we model entities as vectors. To bridge the gap between concept box embeddings and entity vector embeddings, we propose a novel vector-to-box distance metric and learn both embeddings jointly. Experiments on both the public DBpedia KG and a newly-created industrial KG showed the effectiveness of Concept2Box.",
+        "bbox": [
+            141,
+            281,
+            460,
+            664
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            692,
+            258,
+            707
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Knowledge graphs (KGs) like Freebase (Bollacker et al., 2008) and DBpedia (Lehmann et al., 2015) have motivated various knowledge-driven applications (Yasunaga et al., 2021; Lin et al., 2021). They contain structured and semantic information among nodes and relations where knowledge is instantiated as factual triples. One critical fact being ignored is that many KGs consist of two fundamentally different views as depicted in Figure 1: (1) an ontology-view component with high-level concepts and meta-relations, e.g., (\"Singer\", \"is_a\",",
+        "bbox": [
+            112,
+            719,
+            489,
+            896
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/9a060859468b097cf36514a62e6211b865b0b2974010dbfcce1b270a7a3623fa.jpg",
+        "image_caption": [
+            "Figure 1: A two-view KG consisting of (1) an ontology-view KG with high-level concepts and meta-relations, (2) an instance-view KG with fine-grained instances and relations, and, (3) a set of cross-view links between two views of KG. Concept2Box learns dual geometric embeddings where each concept is modeled as a box in the latent space and entities are learned as point vectors."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            263,
+            878,
+            395
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "\"Artist\"), and (2) an instance-view component with specific entities and their relations, e.g., (\"Taylor Swift\", \"has_award\", \"Grammy Award\"). Also, there are valuable cross-view links that connect ontological concepts with entities instantiated from them, which bridge the semantics of the two views.",
+        "bbox": [
+            505,
+            544,
+            884,
+            639
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Modern KGs are usually far from complete as new knowledge is rapidly emerging (Huang et al., 2022). How to automatically harvest unknown knowledge from data has been an important research area for years. Knowledge graph embedding (KGE) methods achieve promising results for modeling KGs by learning low-dimensional vector representations of nodes and relations. However, most existing methods (Bordes et al., 2013; Wang et al., 2014; Sun et al., 2019; Wang et al., 2021b) have a major limitation of only considering KGs of a single view. The valuable semantics brought by jointly modeling the two views are omitted. In contrast, modeling two-view KGs in a holistic way has a key advantage: entity embeddings provide rich information for understanding corresponding ontological concepts; and, in turn, a concept embedding",
+        "bbox": [
+            505,
+            645,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "*Part of work was done during internship at Amazon.",
+        "bbox": [
+            141,
+            904,
+            470,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "10105",
+        "bbox": [
+            475,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: ACL 2023, pages 10105-10118 July 9-14, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            220,
+            945,
+            778,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "provides a high-level summary of linked entities, guiding the learning process of entities with only few relational facts.",
+        "bbox": [
+            112,
+            84,
+            489,
+            133
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Modeling two-view KG is non-trivial where the challenges are three-fold: First, there exists a significant structural difference between the two views, where the ontology-view KG usually forms a hierarchical structure with concept nodes of different granularity, and the instance-view KG usually follows a flat or cyclical structure (Iyer et al., 2022) with each entity associated with a specific meaning. Second, concepts often exhibit complex relations such as intersect, contain and disjoint. Traditional vector-based embeddings (Bordes et al., 2013; Sun et al., 2019) fail to adequately express them and capture interpretable probabilistic semantics to explain the granularity of concepts. Third, to bridge the knowledge from two views, effective semantic mappings from fine-grained entities to their corresponding concepts need to be carefully considered. For work that models two-view KGs, they either model concepts and entities all as points in the Euclidean space (Hao et al., 2019) and in the product manifold space (Wang et al., 2021c), which can fail to capture the structural differences among the two views and result in suboptimal performance, or they embed concepts and entities separately as points in the hyperbolic and spherical space respectively (Iyer et al., 2022), which cannot provide interpretation towards concepts' granularity.",
+        "bbox": [
+            112,
+            140,
+            492,
+            575
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "To this end, we propose Concept2Box, a novel two-view KG embedding model for jointly learning the representations of concepts and instances using dual geometric objects. Specifically, we propose to leverage the geometric objects of boxes (i.e. axis-aligned hyperrectangle) to embed concept nodes in the ontology-view KG. This is motivated by the desirable geometric properties of boxes to represent more complex relationships among concepts such as intersect, contain, and disjoint, and also capture the hierarchy structure. For example, in Figure 1 we can see City is contained within Location, denoting it's a sub-concept of Location; Artist and Scientist intersects with each other, denoting they are correlated. In addition, box embeddings provide probabilistic measures of concepts' granularity based on the volumes, such as Person is a larger box than Scientist. Entities in the instance-view KG are modeled with classic vector representations to capture their specific meanings. As we model concepts and entities using boxes and vectors respectively,",
+        "bbox": [
+            112,
+            581,
+            490,
+            921
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "we further design a new metric function to measure the distance from a vector to a box in order to bridge the semantics of the two views. Based on this vector-to-box distance function, the proposed model is trained to jointly model the semantics of the instance-view KG, the ontology-view KG, and the cross-view links. Empirically, Concept2Box outperforms popular baselines on both public and our newly-created industrial recipe datasets.",
+        "bbox": [
+            507,
+            84,
+            885,
+            231
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Our contributions are as follows: (1) We propose a novel model for learning two-view KG representations by jointly embedding concepts and instances with different geometric objects. (2) We design a new metric function to measure the distance between concept boxes and entity vectors to bridge the two views. (3) We construct a new industrial-level recipe-related KG dataset. (4) Extensive experiments verify the effectiveness of Concept2Box.",
+        "bbox": [
+            507,
+            235,
+            885,
+            382
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            412,
+            665,
+            426
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Knowledge Graph Embedding (KGE)",
+        "text_level": 1,
+        "bbox": [
+            507,
+            451,
+            853,
+            468
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The majority of work on knowledge graph embeddings only considers a single view, where models learn the latent vector representations for nodes and relations and measure triple plausibility via varying score functions. Representatives include translation-based TransE (Bordes et al., 2013), TransH (Wang et al., 2014); rotation-based RotatE (Sun et al., 2019), and neural network-based ConvE (Dettmers et al., 2018). Most recently, some work employs graph neural networks (GNNs) to aggregate neighborhood information in KGs (Huang et al., 2022; Zhang et al., 2020), which have shown superior performance on graph-structured data by passing local message (Huang et al., 2020, 2021; Javari et al., 2020; Wang et al., 2021a). For work that pushes two-view or multiple KGs together (Hao et al., 2019; Huang et al., 2022), they either model all nodes as points in the Euclidean space (Hao et al., 2019; Huang et al., 2022) and in the product manifold space (Wang et al., 2021c), or separately embed entities and concepts as points in two spaces respectively (Iyer et al., 2022). Our work uses boxes and vectors for embedding concepts and entities respectively, which not only gives better performance but also provides a probabilistic explanation of concepts' granularity based on box volumes.",
+        "bbox": [
+            507,
+            483,
+            885,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "10106",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Geometric Representation Learning",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            443,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Representing elements with more complex geometric objects than vectors is an active area of study (Chen et al., 2021). Some representative models include Poincaré embeddings (Nickel and Kiela, 2017), which embeds data into the hyperbolic space with the inductive basis of negative curvature. Hyperbolic entailment cones (Ganea et al., 2018) combine order embeddings with hyperbolic geometry, where relations are viewed as partial orders defined over nested geodesically convex cones. Elliptical distributions (Muzellec and Cuturi, 2018) model nodes as distributions and define distances based on the 2-Wasserstein metric. While they are promising for embedding hierarchical structures, they do not provide interpretable probabilistic semantics. Box embeddings not only demonstrate improved performance on modeling hierarchies but also model probabilities via box volumes to explain the granularity of concepts.",
+        "bbox": [
+            112,
+            108,
+            489,
+            413
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Preliminaries",
+        "text_level": 1,
+        "bbox": [
+            112,
+            430,
+            265,
+            444
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Problem Formulation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            458,
+            329,
+            472
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We consider a two-view knowledge graph  $\\mathcal{G}$  consisting of a set of entities  $\\mathcal{E}$ , a set of concepts  $\\mathcal{C}$ , and a set of relations  $\\mathcal{R} = \\mathcal{R}^{\\mathcal{I}} \\cup \\mathcal{R}^{\\mathcal{O}} \\cup \\mathcal{S}$ , where  $\\mathcal{R}^{\\mathcal{I}}, \\mathcal{R}^{\\mathcal{O}}$  are the entity-entity and concept-concept relation sets respectively.  $\\mathcal{S} = \\{(e, c)\\}$  is the cross-view entity-concept relation set with  $e \\in \\mathcal{E}, c \\in \\mathcal{C}$ . Entities and concepts are disjoint sets, i.e.,  $\\mathcal{E} \\cap \\mathcal{C} = \\emptyset$ .",
+        "bbox": [
+            112,
+            481,
+            487,
+            594
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We denote the instance-view KG as  $\\mathcal{G}^{\\mathcal{I}} := \\{\\mathcal{E}, \\mathcal{R}^{\\mathcal{I}}, \\mathcal{T}^{\\mathcal{I}}\\}$  where  $\\mathcal{T}^{\\mathcal{I}} = \\{(h^{\\mathcal{I}}, r^{\\mathcal{I}}, t^{\\mathcal{I}})\\}$  are factual triples connecting head and tail entities  $h^{\\mathcal{I}}, t^{\\mathcal{I}} \\in \\mathcal{E}$  with a relation  $r^{\\mathcal{I}} \\in \\mathcal{R}^{\\mathcal{I}}$ . The instance-view triples can be any real-world instance associations and are often of a flat structure such as (\"Taylor Swift\", \"has_award\", \"Grammy Award\"). Similarly, the ontology-view KG is denoted as  $\\mathcal{G}^{\\mathcal{O}} := \\{\\mathcal{C}, \\mathcal{R}^{\\mathcal{O}}, \\mathcal{T}^{\\mathcal{O}}\\}$  where  $\\mathcal{T}^{\\mathcal{O}} = \\{(h^{\\mathcal{O}}, r^{\\mathcal{O}}, t^{\\mathcal{O}})\\}$  are triples connecting  $h^{\\mathcal{O}}, t^{\\mathcal{O}} \\in \\mathcal{C}$  and  $r^{\\mathcal{O}} \\in \\mathcal{R}^{\\mathcal{O}}$ . The ontology-view triples describe the hierarchical structure among abstract concepts and their meta-relations such as (\"Singer\", \"is_a\", \"Artist\").",
+        "bbox": [
+            112,
+            596,
+            489,
+            804
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Given: a two-view KG  $\\mathcal{G}$ , we want to learn a model which effectively embeds entities, concepts, and relations to latent representations with different geometric objects to better capture the structural differences of two views. We evaluate the quality of the learned embeddings on the KG completion task and concept linking task, described in Section 5.",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Probabilistic Box Embedding",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            786,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Box embeddings represent an element as a hyperrectangle characterized with two parameters (Vilnis et al., 2018): the minimum and the maximum corners  $(\\pmb{x}_m,\\pmb{x}_M)\\in \\mathbb{R}^d$  where  $x_{m}^{i} < x_{M}^{i}$  for each coordinates  $i\\in 1,2,\\dots d$ . Box volume is computed as the product of side-lengths  $\\mathrm{Vol}(x) = \\prod_{i = 1}^{d}\\left(x_{M}^{i} - x_{m}^{i}\\right)$ . Given a box space  $\\Omega_{\\mathrm{Box}}\\subseteq \\mathbb{R}^d$ , the set of boxes  $\\mathcal{B}(\\Omega_{\\mathrm{Box}})$  is closed under intersection (Chen et al., 2021) and box volumes can be viewed as the unnormalized probabilities to interpret concepts' granularity. The intersection volume of two boxes is defined as  $\\mathrm{Vol}(x\\cap y) = \\prod_{i}\\max \\left(\\min \\left(x_{M}^{i},y_{M}^{i}\\right) - \\max \\left(x_{m}^{i},y_{m}^{i}\\right),0\\right)$  and the conditional probability is computed as  $P(x|y) = \\frac{\\mathrm{Vol}(x\\cap y)}{\\mathrm{Vol}(y)}$ , with value between 0 and 1 (Vilnis et al., 2018).",
+        "bbox": [
+            507,
+            107,
+            882,
+            380
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Directly computing the condition probabilities would pose difficulties during training. This can be caused by a variety of settings like when two boxes are disjoint but should overlap (Chen et al., 2021), there would be no gradient for some parameters. To alleviate this issue, we rely on Gumbel boxes (Das-gupta et al., 2020), where the min and max corners of boxes are modeled via Gumbel distributions:",
+        "bbox": [
+            507,
+            382,
+            882,
+            510
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\operatorname {V o l} (x) = \\prod_ {i = 1} ^ {d} \\left[ x _ {m} ^ {i}, x _ {M} ^ {i} \\right] \\quad \\text {w h e r e} \\tag {1}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            566,
+            526,
+            882,
+            575
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {m} ^ {i} \\sim \\operatorname {G u m b e l M a x} \\left(\\mu_ {m} ^ {i}, \\beta\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            594,
+            571,
+            818,
+            590
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nx _ {M} ^ {i} \\sim \\operatorname {G u m b e l M i n} \\left(\\mu_ {M} ^ {i}, \\beta\\right),\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            593,
+            592,
+            816,
+            611
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "$\\mu_{m}^{i}, \\mu_{M}^{i}$  are the minimum and maximum coordinates parameters;  $\\beta$  is the global variance. Gumbel distribution provides min/max stability (Chen et al., 2021) and the expected volume is approximated as",
+        "bbox": [
+            507,
+            621,
+            882,
+            686
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} [ \\operatorname {V o l} (x) ] \\approx \\prod_ {i = 1} ^ {d} \\beta \\log \\left(1 + \\exp \\left(\\frac {\\mu_ {M} ^ {i} - \\mu_ {m} ^ {i}}{\\beta} - 2 \\gamma\\right)\\right), \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            700,
+            882,
+            745
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma$  is the Euler-Mascheroni constant (Onoe et al., 2021). The conditional probability becomes",
+        "bbox": [
+            507,
+            747,
+            880,
+            778
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathbb {E} [ P (x | y) ] \\approx \\frac {\\mathbb {E} [ \\operatorname {V o l} (x \\cap y) ]}{\\mathbb {E} [ \\operatorname {V o l} (y) ]}. \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            583,
+            791,
+            882,
+            826
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "This also leads to improved learning as the noise ensembles over a large collection of boxes which allows the model to escape plateaus in the loss function (Chen et al., 2020; Onoe et al., 2021). We use this method when training box embeddings.",
+        "bbox": [
+            507,
+            839,
+            880,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "10107",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/2379f0be122f4b862455e0d691c9d97d8952aba9d0bf47be383e8075d5eaef72.jpg",
+        "image_caption": [
+            "Figure 2: The overall framework of our proposed Concept2Box includes 3 modules. First, we employ probabilistic box embeddings to model the ontology-view KG, capturing the hierarchical structure and preserving the granularity of concepts (upper). Second, we model the instance-view KG by applying the vector-based KG embedding method (bottom). Third, to bridge these two views, we design a novel distance metric defined from a box to a vector (right). The model is learned by jointly optimizing three loss terms of each corresponding module."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            114,
+            83,
+            884,
+            373
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            478,
+            218,
+            494
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce Concept2Box, a novel KGE model to learn two-view knowledge graph embeddings with dual geometric representations. As shown in Figure 2, Concept2Box consists of three main components: (1). The ontology-view box-based KG embedding module that captures the hierarchical structure and complex relations of concepts trained via the KG completion loss over  $\\mathcal{G}^{\\mathcal{O}}$ ; (2). The instance-view vector-based KG embedding module trained via the KG completion loss over  $\\mathcal{G}^{\\mathcal{I}}$ . It has the flexibility to incorporate any off-the-shelf vector-based KGE methods (Bordes et al., 2013; Sun et al., 2019); (3). The cross-view module trained via the concept linking loss over  $S$ . This module relies on a new distance metric for bridging the semantics between a vector and a box. We now present each component in detail.",
+        "bbox": [
+            112,
+            508,
+            489,
+            782
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.1 Ontology-View Box Embedding",
+        "text_level": 1,
+        "bbox": [
+            112,
+            797,
+            410,
+            814
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "For the ontology-view KG, we model each concept as a Gumbel box to capture the hierarchical structure and interpret concept granularity based on the box volume. Specifically, each Gumbel box  $x$  is parameterized using a center  $\\operatorname{cen}(x) \\in \\mathbb{R}^d$  and offset  $\\operatorname{off}(x) \\in \\mathbb{R}_{+}^{d}$ , and the minimum and",
+        "bbox": [
+            112,
+            822,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "maximum location parameters are given by  $\\mu_{\\mathbf{m}} =$ $\\operatorname {cen}(x) - \\operatorname {off}(x),\\mu_{\\mathbf{M}} = \\operatorname {cen}(x) + \\operatorname {off}(x).$  To compute the triple score from box embeddings, we use the approximate conditional probability of observing a triple given its relation and tail, as shown",
+        "bbox": [
+            507,
+            480,
+            884,
+            558
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\phi \\left(h ^ {\\mathcal {O}}, r ^ {\\mathcal {O}}, t ^ {\\mathcal {O}}\\right) = \\frac {\\mathbb {E} \\left[ \\operatorname {V o l} \\left(f _ {r \\mathcal {O}} \\left(\\operatorname {B o x} \\left(h ^ {\\mathcal {O}}\\right)\\right) \\cap f _ {r \\mathcal {O}} \\left(\\operatorname {B o x} \\left(t ^ {\\mathcal {O}}\\right)\\right)\\right) \\right]}{\\mathbb {E} \\left[ \\operatorname {V o l} \\left(f _ {r \\mathcal {O}} \\left(\\operatorname {B o x} \\left(t ^ {\\mathcal {O}}\\right)\\right)\\right) \\right]}, \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            514,
+            573,
+            882,
+            615
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where  $f_{r^{\\mathcal{O}}}$  denotes the box transformation function to account for the influence of the relation  $r^{\\mathcal{O}}$ . As shown in Figure 1,  $f_{\\text{lives\\_in}}(\\text{Artist})$ ,  $f_{\\text{is\\_a}}(\\text{Artist})$  are the transformed box embeddings of the concept Artist in the context of the living locations and supertypes. The bounded score  $\\phi$  should be close to 1 for positive triples and 0 for negative ones.",
+        "bbox": [
+            505,
+            615,
+            882,
+            728
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Previous work Chen et al. (2021) defined the transformation function  $f_{r}$  as shifting the original center and offset of the input box  $x$ :  $\\operatorname{cen}(f_r(x)) = \\operatorname{cen}(x) + \\pmb{\\tau}^r$ ,  $\\operatorname{off}(f_r(x)) = \\operatorname{off}(x) \\circ \\Delta^r$ , where  $\\circ$  is the Hadamard product and  $(\\pmb{\\tau}^r, \\pmb{\\Delta}^r)$  are relation-specific parameters. To further incorporate semantic information from the text of a concept  $c$  and relation  $r$ , we propose a new transformation function as:",
+        "bbox": [
+            507,
+            730,
+            882,
+            872
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathbf {h} ^ {[ \\mathrm {C L S} ]} = \\operatorname {B E R T} \\left(c _ {\\text {t e x t}} + ^ {\\prime} \\mathrm {S E P} ^ {\\prime} + r _ {\\text {t e x t}}\\right) \\\\ \\operatorname {c e n} \\left(f _ {r} (\\operatorname {B o x} (c))\\right), \\operatorname {o f f} \\left(f _ {r} (\\operatorname {B o x} (c))\\right) = f _ {\\operatorname {p r o j}} \\left(\\mathbf {h} ^ {[ \\mathrm {C L S} ]}\\right). \\end{array} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            519,
+            885,
+            882,
+            923
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "10108",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we first concatenate the textual content of the concept and relation for feeding into a pretrained language model BERT (Devlin et al., 2019) that captures the rich semantics of the text. We obtain the embedding by taking the hidden vector at the [CLS] token. Then, we employ a projection function  $f_{\\mathrm{proj}}$  to map the hidden vector  $\\mathbf{h}^{[\\mathrm{CLS}]}\\in \\mathbb{R}^{\\ell}$  to the  $\\mathbb{R}^{2d}$  space, where  $\\ell$  is the hidden dimension of the BERT encoder, and  $d$  is the dimension of the box space. In practice, we implement  $f_{\\mathrm{proj}}$  as a simple two-layer Multi-Layer Perception (MLP) with a nonlinear activation function. We further split the hidden vector into two equal-length vectors as the center and offset of the transformed box.",
+        "bbox": [
+            112,
+            84,
+            489,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Remark 1. Instead of encoding relation text first to get the shifting and scaling parameters and then conducting the box transformation as introduced in (Chen et al., 2020), we feed the concept and relation text jointly to the BERT encoder to allow interaction between them through the attention mechanism. The goal is to preserve both concept and relation semantics by understanding their textual information as a whole.",
+        "bbox": [
+            112,
+            313,
+            489,
+            456
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We train the ontology-view KG embedding module with the binary cross entropy loss as below:",
+        "bbox": [
+            112,
+            461,
+            489,
+            494
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} \\mathcal {I} ^ {\\mathcal {G} ^ {\\mathcal {C}}} = \\sum_ {\\substack {(h ^ {\\mathcal {O}}, r ^ {\\mathcal {O}}, t ^ {\\mathcal {O}}) \\in \\mathcal {T} ^ {\\mathcal {O}}, \\\\ (\\tilde {h} ^ {\\mathcal {O}}, r ^ {\\mathcal {O}}, \\tilde {t} ^ {\\mathcal {O}}) \\notin \\mathcal {T} ^ {\\mathcal {O}}}} | \\phi (h ^ {\\mathcal {O}}, r ^ {\\mathcal {O}}, t ^ {\\mathcal {O}}) - 1 | ^ {2} \\tag{6} \\\\ { + } { | \\phi ( \\widetilde { h } ^ { \\mathcal { O } } , r ^ { \\mathcal { O } } , \\widetilde { t } ^ { \\mathcal { O } } ) | ^ { 2 } , } \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            126,
+            513,
+            489,
+            588
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $(\\widetilde{h}^{\\mathcal{O}},r^{\\mathcal{O}},\\widetilde{t}^{\\mathcal{O}})$  is a negative sampled triple obtained by replacing head or tail of the true triple  $(h^{O},r^{O},t^{O})$  . We do not use the hinge loss defined in vector-based KGE models (Bordes et al., 2013), as the triple scores are computed from the conditional probability and strictly fall within O and 1.",
+        "bbox": [
+            112,
+            596,
+            489,
+            695
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Instance-View and Cross-View Modeling",
+        "text_level": 1,
+        "bbox": [
+            112,
+            705,
+            480,
+            721
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Next, we introduce learning the instance-view KG and the cross-view module for bridging two views.",
+        "bbox": [
+            112,
+            726,
+            487,
+            757
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Instance-View KG Modeling. As entities in the instance-view KG are fine-grained, we choose to model entities and relations as classic point vectors. We employ the standard hinge loss defined as follows to train the instance-view KGE module.",
+        "bbox": [
+            112,
+            758,
+            489,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {I} ^ {\\mathcal {G} ^ {\\mathcal {I}}} = \\sum_ {\\substack {(h ^ {\\mathcal {I}}, r ^ {\\mathcal {I}}, t ^ {\\mathcal {I}}) \\in \\mathcal {T} ^ {\\mathcal {I}} \\\\ (\\tilde {h} ^ {\\mathcal {I}}, r ^ {\\mathcal {I}}, \\tilde {t} ^ {\\mathcal {I}}) \\notin \\mathcal {T} ^ {\\mathcal {I}}}} \\left[ \\begin{array}{c} f _ {\\text {vec}} \\left(h ^ {\\mathcal {I}}, r ^ {\\mathcal {I}}, t ^ {\\mathcal {I}}\\right) \\\\ - f _ {\\text {vec}} \\left(\\tilde {h} ^ {\\mathcal {I}}, r ^ {\\mathcal {I}}, \\tilde {t} ^ {\\mathcal {I}}\\right) \\\\ + \\gamma^ {\\mathrm {KG}} \\end{array} \\right] _ {+}, \\tag{7}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            122,
+            857,
+            489,
+            921
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma^{\\mathrm{KG}} > 0$  is a positive margin,  $f_{\\mathrm{vec}}$  is an arbitrary vector-based KGE model such as TransE (Bordes et al., 2013) and RotatE (Sun et al., 2019). And  $(\\widetilde{h}^{\\mathcal{I}}, r^{\\mathcal{I}}, \\widetilde{t}^{\\mathcal{I}})$  is a sampled negative triple.",
+        "bbox": [
+            507,
+            82,
+            884,
+            149
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Cross-View KG Modeling. The goal of the cross-view module is to bridge the semantics between the entity embeddings and the concept embeddings by using the cross-view links. It forces any entity  $e \\in \\mathcal{E}$  to be close to its corresponding concept  $c \\in \\mathcal{C}$ . As we model concepts and entities with boxes and vectors respectively, existing distance measures between boxes (or between vectors) would be inappropriate for measuring the distance between them. To tackle this challenge, we now define a new metric function to measure the distance from a vector to a box. Our model is trained to minimize the distances for the concept-entity pairs in the cross-view links based on this metric function. Specifically, given a concept  $c$  and an entity  $e$ , if we denote the minimum and maximum location parameters of the concept box as  $\\mu_{m}, \\mu_{M}$ , and the vector for the entity as  $e$ , we define the distance function  $f_{d}$  as:",
+        "bbox": [
+            507,
+            149,
+            884,
+            456
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} f _ {d} (e, c) = \\operatorname {d i s t} _ {\\text {o u t}} (e, c) + \\alpha \\cdot \\operatorname {d i s t} _ {\\text {i n}} (e, c) \\\\ \\operatorname {d i s t} _ {\\text {o u t}} (e, c) = \\left\\| \\operatorname {M a x} \\left(\\boldsymbol {e} - \\boldsymbol {\\mu} _ {M}, \\mathbf {0}\\right) + \\operatorname {M a x} \\left(\\boldsymbol {\\mu} _ {m} - \\boldsymbol {e}, \\mathbf {0}\\right) \\right\\| _ {1} \\\\ \\operatorname {d i s t} _ {\\text {i n}} (e, c) = \\left\\| \\operatorname {C e n} (\\operatorname {B o x} (c)) - \\operatorname {M i n} \\left(\\boldsymbol {\\mu} _ {M}, \\operatorname {M a x} \\left(\\boldsymbol {\\mu} _ {m}, e\\right)\\right) \\right\\| _ {1} \\tag {8} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            469,
+            880,
+            530
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\mathrm{dist}_{\\mathrm{out}}$  corresponds to the distance between the entity and closest corner of the box;  $\\mathrm{dist}_{\\mathrm{in}}$  corresponds to the distance between the center of the box and its side or the entity itself if the entity is inside the box (Ren et al., 2020).  $0 < \\alpha < 1$  is a fixed scalar to balance the outside and inside distances, with the goal to punish more if the entity falls outside of the concept, and less if it is within the box but not close to the box center. An example is shown in Figure 2. The proposed vector-to-box distance metric is nonnegative and commutative.",
+        "bbox": [
+            507,
+            532,
+            882,
+            708
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "However, there is one drawback to the above formulation: different boxes should have customized volumes reflecting their granularity. A larger box may be a more general concept (e.g., Person), so it is suboptimal for all paired entities to be as close as to its center, in contrast to more concrete concepts like Singer, as illustrated in Figure 2. Moreover, for instances that belong to multiple concepts, roughly pushing them to be equally close to the centers of relevant concepts can result in poor embedding results. Consider we have two recipe concepts KungPao chicken and Curry chicken, suppose KungPao Chicken is more complicated and requires",
+        "bbox": [
+            507,
+            709,
+            884,
+            919
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "10109",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "more ingredients. If we roughly push cooking wine, salt, and chicken breast these shared ingredient instances to be as close as to both of their centers, it will possibly result in forcing the centers and volume of the two recipes to be very similar to each other. Therefore, we propose to associate the balancing scalar  $\\alpha$  to be related to each concept box volume. Specifically, we define  $\\alpha(x)$  with regard to a box  $x$  as follows:",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\alpha (x) = \\alpha \\cdot \\operatorname {S i g m o i d} \\left(\\frac {1}{\\operatorname {V o l} (x)}\\right). \\tag {9}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            186,
+            239,
+            487,
+            272
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We multiply the original constant coefficient  $\\alpha$  with a box-specific value negatively proportional to its volume. The idea is to force entities to be closer to the box center when the box has a smaller volume.",
+        "bbox": [
+            112,
+            275,
+            487,
+            338
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The cross-view module is trained by minimizing the following loss with negative sampling:",
+        "bbox": [
+            112,
+            340,
+            485,
+            372
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {J} ^ {\\text {C r o s s}} = \\sum_ {\\substack {(e, c) \\in \\mathcal {S} \\\\ (e ^ {\\prime}, c ^ {\\prime}) \\notin \\mathcal {S}}} \\left[ \\sigma \\left(f _ {d} (e, c)\\right) - \\sigma \\left(f _ {d} \\left(e ^ {\\prime}, c ^ {\\prime}\\right)\\right) + \\gamma^ {\\text {C r o s s}} \\right], \\tag{10}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            112,
+            380,
+            497,
+            428
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\gamma^{\\mathrm{Cross}}$  represents a fixed scalar margin.",
+        "bbox": [
+            112,
+            429,
+            450,
+            445
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 Overall Training Objective",
+        "text_level": 1,
+        "bbox": [
+            112,
+            456,
+            371,
+            470
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Now that we have presented all model components, the overall loss function is a linear combination of the instance-view and ontology-view KG completion loss, and the cross-view loss, as shown below:",
+        "bbox": [
+            112,
+            476,
+            489,
+            539
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\mathcal {J} = \\mathcal {J} ^ {\\mathcal {G} ^ {\\mathcal {O}}} + \\lambda_ {1} \\mathcal {J} ^ {\\mathcal {G} ^ {\\mathcal {I}}} + \\lambda_ {2} \\mathcal {J} ^ {\\text {C r o s s}}, \\tag {11}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            548,
+            487,
+            567
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda_1, \\lambda_2 > 0$  are two positive hyperparameters to balance the scale of the three loss terms.",
+        "bbox": [
+            112,
+            579,
+            485,
+            609
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.4 Time Complexity Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            621,
+            366,
+            636
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The time complexity of Concept2Box is  $\\mathcal{O}(3d|\\mathcal{T}^{\\mathcal{O}}| + d|\\mathcal{T}^{\\mathcal{I}}| + 5d|S|)$ , where  $\\mathcal{T}^{\\mathcal{O}},\\mathcal{T}^{\\mathcal{I}},\\mathcal{S}$  are the ontology-view KG triples, instance-view KG triples, and cross-view set defined in Sec 3.1. Here  $d$  is the embedding dimension of vectors. The time complexity of other vector-based methods is  $\\mathcal{O}(d|\\mathcal{T}^{\\mathcal{O}}| + d|\\mathcal{T}^{\\mathcal{I}}| + d|S|)$  which is asymptotically the same as Concept2Box.",
+        "bbox": [
+            112,
+            640,
+            487,
+            770
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            781,
+            258,
+            797
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In this section, we introduce our empirical results against existing baselines to evaluate our model performance. We start with the datasets and baselines introduction and then conduct the performance evaluation. Finally, we conduct a model generalization experiment and a case study to better illustrate our model design.",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Dataset and Baselines",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            726,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We conducted experiments over two datasets: one public dataset from DBpedia (Hao et al., 2019), which describes general concepts and fine-grained entities crawled from DBpedia. We also created a new recipe-related dataset, where concepts are recipes, general ingredient and utensil names, etc, and the entities are specific products for cooking each recipe searched via Amazon.com, along with some selected attributes such as brand. The detailed data structure can be found in Appendix A. The statistics of the two datasets are shown in Table 1.",
+        "bbox": [
+            505,
+            105,
+            882,
+            282
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We compare against both single-view KGE models and two-view KGE models as follows:",
+        "bbox": [
+            507,
+            284,
+            882,
+            313
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Vector-based KGE models:: Single-view vector-based KGE models that represent triples as vectors, including TransE (Bordes et al., 2013), RotatE (Sun et al., 2019), DistMult (Yang et al., 2015), and ComplexE (Trouillon et al., 2016).",
+            "- KGBert (Yao et al., 2020): Single-view KG embedding method which employs BERT to incorporate text information of triples.",
+            "- Box4ET (Onoe et al., 2021): Single-view KGE model that represents triples using boxes.",
+            "- Hyperbolic GCN (Chami et al., 2019): Single-view KGE method that represents triples in the hyperbolic space.",
+            "JOIE (Hao et al., 2019): Two-view KGE method that embeds entities and concepts all as vectors."
+        ],
+        "bbox": [
+            507,
+            317,
+            882,
+            560
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 KG Completion Performance",
+        "text_level": 1,
+        "bbox": [
+            507,
+            574,
+            786,
+            589
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The KG completion (KGC) task is to predict missing triples based on the learned embeddings, which helps to facilitate the knowledge learning process in KG construction. Without loss of generality, we discuss the case of predicting missing tails, which we also refer to as a query  $q = (h,r,\\text {?}t)$  (Huang et al., 2022; Chen et al., 2021). We conduct the KGC task for both views against baselines.",
+        "bbox": [
+            505,
+            596,
+            882,
+            724
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Protocol. We evaluate the KGC performance using mean reciprocal ranks (MRR), accuracy (Hits@1) and the proportion of correct answers ranked within the top 10 (Hits@10). All three metrics are preferred to be higher, so as to indicate better KGC performance. For both  $\\mathcal{G}^{\\mathcal{I}}$  and  $\\mathcal{G}^{\\mathcal{O}}$ , we randomly split the triples into three parts:  $80\\%$  for training,  $10\\%$  for validation and  $10\\%$  for testing. During model training and testing, we utilize all cross-view links  $S$ .",
+        "bbox": [
+            507,
+            726,
+            882,
+            884
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Results. The evaluation results are shown in Table 2. Firstly, by comparing the average per",
+        "bbox": [
+            507,
+            887,
+            882,
+            919
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "10110",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/0887dab03b2661e7aa87989724fb8cf9d6ec0af204636fecd9323335f3fec6fb.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td colspan=\"3\">Instance-View KG</td><td colspan=\"3\">Ontology-View KG</td><td>Cross-View Links</td></tr><tr><td></td><td># Entities</td><td># Relations</td><td># Triples</td><td># Concepts</td><td># Relations</td><td># Triples</td><td># Links</td></tr><tr><td>DBpedia</td><td>111,762</td><td>305</td><td>863,643</td><td>174</td><td>20</td><td>763</td><td>99,748</td></tr><tr><td>Recipe</td><td>21,457</td><td>9</td><td>200,288</td><td>32,922</td><td>5</td><td>445,632</td><td>11,474</td></tr></table>",
+        "bbox": [
+            161,
+            82,
+            833,
+            145
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d3feff077c0a4fefe145212f401e18353ca8691ce88b587e98a9533d96befeb8.jpg",
+        "table_caption": [
+            "Table 1: Statistics of the DBpedia and Recipe datasets."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"6\">Recipe</td><td colspan=\"6\">DBpedia</td></tr><tr><td colspan=\"3\">G KG Completion</td><td colspan=\"3\">G KG Completion</td><td colspan=\"3\">G KG Completion</td><td colspan=\"3\">G KG Completion</td></tr><tr><td></td><td>MRR</td><td>H@1</td><td>H@10</td><td>MRR</td><td>H@1</td><td>H@10</td><td>MRR</td><td>H@1</td><td>H@10</td><td>MRR</td><td>H@1</td><td>H@10</td></tr><tr><td>TransE</td><td>0.28</td><td>26.31</td><td>38.59</td><td>0.18</td><td>11.24</td><td>30.67</td><td>0.32</td><td>22.70</td><td>48.12</td><td>0.54</td><td>47.90</td><td>61.84</td></tr><tr><td>RotatE</td><td>0.30</td><td>28.31</td><td>40.01</td><td>0.22</td><td>15.68</td><td>35.47</td><td>0.36</td><td>29.31</td><td>54.60</td><td>0.56</td><td>49.16</td><td>68.19</td></tr><tr><td>DistMult</td><td>0.29</td><td>27.33</td><td>39.04</td><td>0.18</td><td>12.37</td><td>32.04</td><td>0.28</td><td>27.24</td><td>29.70</td><td>0.50</td><td>45.52</td><td>64.73</td></tr><tr><td>ComplexE</td><td>0.27</td><td>26.58</td><td>37.99</td><td>0.19</td><td>14.76</td><td>34.02</td><td>0.32</td><td>27.39</td><td>46.63</td><td>0.55</td><td>47.80</td><td>62.23</td></tr><tr><td>JOIE</td><td>0.41</td><td>29.73</td><td>60.14</td><td>0.36</td><td>22.58</td><td>53.62</td><td>0.48</td><td>35.21</td><td>72.38</td><td>0.60</td><td>52.48</td><td>79.71</td></tr><tr><td>Hyperbolic GCN</td><td>0.36</td><td>28.58</td><td>52.48</td><td>0.25</td><td>16.04</td><td>40.35</td><td>0.30</td><td>33.68</td><td>46.72</td><td>0.54</td><td>47.59</td><td>62.11</td></tr><tr><td>KG-Bert</td><td>0.30</td><td>29.92</td><td>51.32</td><td>0.24</td><td>19.04</td><td>38.78</td><td>0.39</td><td>31.10</td><td>60.46</td><td>0.63</td><td>55.69</td><td>81.07</td></tr><tr><td>Box4ET</td><td>0.42</td><td>28.81</td><td>59.88</td><td>0.35</td><td>23.01</td><td>53.79</td><td>0.42</td><td>33.69</td><td>68.12</td><td>0.64</td><td>55.89</td><td>81.45</td></tr><tr><td>Concept2Box</td><td>0.44</td><td>30.33</td><td>62.01</td><td>0.37</td><td>23.16</td><td>54.92</td><td>0.51</td><td>36.52</td><td>73.11</td><td>0.65</td><td>56.82</td><td>83.01</td></tr></table>",
+        "bbox": [
+            115,
+            184,
+            882,
+            356
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "formance between single-view and two-view KG models, we can observe that two-view methods are able to achieve better results for the KGC task. This indicates that the intuition behind modeling two-view KGs to conduct KG completion is indeed beneficial. Secondly, by comparing vector-based methods with other geometric representation-based methods such as Hyperbolic GCN, there is a performance improvement on most metrics, indicating that simple vector representations have its limitation, especially in capturing the hierarchical structure among concepts (shown by the larger performance gap in  $\\mathcal{G}^{\\mathcal{O}}$ ). Our Concept2Box is able to achieve higher performance in most cases, verifying the effectiveness of our design.",
+        "bbox": [
+            112,
+            406,
+            489,
+            646
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Concept Linking Performance",
+        "text_level": 1,
+        "bbox": [
+            112,
+            657,
+            400,
+            671
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The concept linking task predicts the associating concepts of given entities, where each entity can be potentially mapped to multiple concepts. It tests the quality of the learned embeddings.",
+        "bbox": [
+            112,
+            677,
+            487,
+            741
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Evaluation Protocol. We split the cross-view links  $S$  into  $80\\%$ ,  $10\\%$ ,  $10\\%$  for training, validation, testing respectively. We utilize all the triples in both views for training and testing. For each concept-entity pair, we rank all concept candidates for the given entity based on the distance metric from a box to a vector introduced in Sec 4.2, and report MRR, Hits@1 and Hits@3 for evaluation.",
+        "bbox": [
+            112,
+            741,
+            487,
+            869
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Results. The evaluation results are shown in Table 3. We can see that Concept2Box is able to achieve the highest performance across most",
+        "bbox": [
+            112,
+            871,
+            487,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "metrics, indicating its effectiveness. By comparing with Box4ET where both entities and concepts are modeled as boxes, our Concept2Box performs better, which indicates our intuition that entities and concepts are indeed two fundamentally different types of nodes and should be modeled with different geometric objects. Finally, we observe that Box4ET, and Concept2Box are able to surpass methods that are not using box embeddings, which shows the advantage of box embeddings in learning the hierarchical structure and complex behaviors of concepts with varying granularity.",
+        "bbox": [
+            507,
+            406,
+            884,
+            599
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/659ecc02ba807238ee63c03ba84e3e08b32b91ff6ab1d35805812fe53f68350f.jpg",
+        "table_caption": [
+            "Table 2: Results of KG completion task. Best results are bold-faced"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Datasets Metrics</td><td colspan=\"3\">Recipe</td><td colspan=\"3\">DBPedia</td></tr><tr><td>MRR</td><td>Acc.</td><td>Hit@3</td><td>MRR</td><td>Acc.</td><td>Hit@3</td></tr><tr><td>TransE</td><td>0.23</td><td>21.36</td><td>30.68</td><td>0.53</td><td>43.67</td><td>60.78</td></tr><tr><td>RotatE</td><td>0.25</td><td>23.01</td><td>33.34</td><td>0.72</td><td>61.48</td><td>75.67</td></tr><tr><td>DistMult</td><td>0.24</td><td>22.34</td><td>31.05</td><td>0.55</td><td>49.83</td><td>68.01</td></tr><tr><td>ComplexE</td><td>0.22</td><td>22.50</td><td>31.45</td><td>0.58</td><td>55.07</td><td>70.17</td></tr><tr><td>JOIE</td><td>0.57</td><td>54.58</td><td>66.39</td><td>0.85</td><td>75.58</td><td>96.02</td></tr><tr><td>Hyper. GCN</td><td>0.60</td><td>53.49</td><td>67.03</td><td>0.86</td><td>76.11</td><td>96.50</td></tr><tr><td>Box4ET</td><td>0.59</td><td>54.36</td><td>68.88</td><td>0.88</td><td>77.04</td><td>97.38</td></tr><tr><td>Concept2Box</td><td>0.61</td><td>55.33</td><td>69.01</td><td>0.87</td><td>78.09</td><td>97.44</td></tr></table>",
+        "bbox": [
+            510,
+            612,
+            895,
+            715
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Results of concept linking task.",
+        "bbox": [
+            557,
+            726,
+            831,
+            740
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Abaltion Studies. To evaluate the effectiveness of our model design, we conduct an ablation study on the DBpedia dataset as shown in Table 4. First, to show the effect of joint modeling the text semantics of relations and concepts for generating the relation-transformed box embeddings as in Eqn 5, we compare with separately modeling relations as box shifting and scaling introduced in (Chen et al., 2021). The latter one gives poorer results, indicating the effectiveness of our proposed box encoder.",
+        "bbox": [
+            507,
+            758,
+            885,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "10111",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Next, we study the effect of the volume-based balancing scalar defined in Eqn 9. Compared with a fixed balancing scalar for all boxes, Concept2Box is able to generate better results.",
+        "bbox": [
+            112,
+            84,
+            489,
+            149
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/3e0572db47266ad4ddf46fec07f86d5df77c0aa3e441e1b040ec6dac8f4fc6a0.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\"></td><td colspan=\"3\">G^T KG Completion</td><td colspan=\"3\">G^O KG Completion</td><td colspan=\"3\">Concept Linking</td></tr><tr><td>MRR</td><td>H@1</td><td>H@10</td><td>MRR</td><td>H@1</td><td>H@10</td><td>MRR</td><td>H@1</td><td>H@3</td></tr><tr><td>Concept2Box</td><td>0.506</td><td>36.52</td><td>73.11</td><td>0.644</td><td>56.82</td><td>83.01</td><td>0.874</td><td>78.09</td><td>97.44</td></tr><tr><td>- relation shift&amp;scale</td><td>0.483</td><td>34.99</td><td>72.18</td><td>0.631</td><td>55.34</td><td>81.97</td><td>0.865</td><td>77.64</td><td>96.34</td></tr><tr><td>- fixed distance coefficient</td><td>0.479</td><td>35.13</td><td>71.88</td><td>0.635</td><td>54.98</td><td>82.33</td><td>0.854</td><td>76.91</td><td>96.13</td></tr></table>",
+        "bbox": [
+            115,
+            160,
+            505,
+            231
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Model Generalization",
+        "text_level": 1,
+        "bbox": [
+            112,
+            281,
+            332,
+            294
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We test our model's generalization ability by training our model on the Recipe dataset and evaluating it on the concept linking task with recipe-product pairs collected from other online resources. Specifically, in the evaluation data, all products are known entities in the Recipe dataset during training, where we can directly obtain their embeddings. For recipes, only  $3\\%$  of them are observed, and for unseen recipes, we generate their box embeddings by sending their title as input to the box encoder. We divide these recipe-product pairs into two groups based on whether the recipes are known. We use the first group with known recipes to test the generalization ability in the transductive setting: When test data follows a different distribution as opposed to the training data, how does the model perform? It differs from Sec 5.3 in that the test data share the same distribution with the training data as we randomly split them from the same dataset. We use the second group with unknown recipes to test in the inductive setting: When unknown concepts are sent as input, how does the model perform?",
+        "bbox": [
+            112,
+            302,
+            489,
+            655
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Diversity-Aware Evaluation. To verify our learned embeddings can capture complex relations and hierarchy structure among concepts, we use diversity-aware evaluation for the generalization experiment similar to that in (Hao et al., 2020). The idea is that for each recipe, a more preferred ranking list is one consisting of products from multiple necessary ingredients, instead of the one with products all from one ingredient. Since we model",
+        "bbox": [
+            112,
+            657,
+            489,
+            802
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/7488292c3da6a56b51a6179f97238827eb311c3f20054a7c295719d3962f1be6.jpg",
+        "table_caption": [
+            "Table 4: Results of ablation study on DBpedia."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"3\">Known Recipes</td><td colspan=\"3\">Unknown Recipes</td></tr><tr><td># types *</td><td>10*12</td><td>12*10</td><td>24*5</td><td>10*12</td><td>12*10</td><td>24*5</td></tr><tr><td># items</td><td></td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>JOIE</td><td>68.33</td><td>69.17</td><td>68.45</td><td>57.34</td><td>57.56</td><td>56.38</td></tr><tr><td>Concept2Box</td><td>71.97</td><td>73.33</td><td>69.43</td><td>61.32</td><td>66.77</td><td>59.68</td></tr></table>",
+        "bbox": [
+            117,
+            822,
+            492,
+            890
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 5: H@120 with different number of item types.",
+        "bbox": [
+            117,
+            898,
+            480,
+            914
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "recipes and ingredients as concepts, and products as entities, for each recipe, we first extract its top  $K$  ingredients (types) from  $\\mathcal{G}^{\\mathcal{O}}$ , and then for each of the  $K$  ingredient we extract the top  $M$  products (items). We set  $M \\times K = 120$  and let  $K = 10, 12, 24$ .",
+        "bbox": [
+            507,
+            84,
+            884,
+            162
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Results. Table 5 shows the results of H@120 among two-view KGE methods. We can see that Concept2Box is able to achieve the best across different settings, showing its great generalization ability. Note that when changing the number of types, Concept2Box is able to bring more performance gain than JOIE. This can be understood as the hierarchy structure of concepts is well captured by the box embeddings, when properly selecting the number of types, we first narrow down the concept (recipe) to the set of related concepts (ingredients) for a better understanding, which will generate better results.",
+        "bbox": [
+            507,
+            165,
+            884,
+            375
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Case Study",
+        "text_level": 1,
+        "bbox": [
+            507,
+            388,
+            645,
+            404
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We conduct a case study to investigate the semantic meaning of learned box embeddings of concepts. We select recipe concepts and show the top 3 concepts that have the largest intersection with them after the relation transformation of \"using ingredients\". As shown in Figure 3, the dominant ingredients for each recipe are at the top of the list, verifying the effectiveness of Concept2Box. Also, we found that Egyptian lentil Soup has a larger intersection with Black Rice Pudding than with Thai Iced Tea, since the first two are dishes with more similarities and the third one is beverage. We also examine the box volume of concepts which reflects their granularity. Among ingredient concepts, we observe that Water, Sugar, and Rice are of the largest volumes, indicating they are common ingredients shared across recipes. This observation confirms that the box volume effectively represents probabilistic semantics, which we believe to be a reason for the performance improvement.",
+        "bbox": [
+            507,
+            410,
+            884,
+            733
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            747,
+            643,
+            762
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose Concept2Box, a novel two-view knowledge graph embedding method with dual geometric representations. We model high-level concepts as boxes to capture the hierarchical structure and complex relations among them and reflect concept granularity based on box volumes. For the instance-view KG, we model fine-grained entities as vectors and propose a novel metric function to define the distance between a",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "10112",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/1a14d75361ebfdeedc8962417da938c8343cecd24d2d2426ed6062913eecd24b.jpg",
+        "image_caption": [
+            "Egyptian Lentil Soup:"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            90,
+            295,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Water (0.891)",
+        "bbox": [
+            131,
+            244,
+            226,
+            256
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Red Lentil (0.864)",
+        "bbox": [
+            132,
+            256,
+            250,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Rom Tomato (0.776)",
+        "bbox": [
+            132,
+            269,
+            267,
+            281
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/c082caf89532b9eb8bc5d98b31a311ef0216f27ae598302532dabb29c103ef86.jpg",
+        "image_caption": [
+            "Black Rice Pudding:"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            315,
+            91,
+            473,
+            204
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Black Rice (0.765)",
+        "bbox": [
+            319,
+            243,
+            443,
+            256
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "White Rice (0.632)",
+        "bbox": [
+            321,
+            256,
+            443,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Palm Sugar (0.618)",
+        "bbox": [
+            321,
+            269,
+            448,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/96de4f75c98742fed16c6bca9f4d8d9fbc111081922fcc1acd2504c1aea358eb.jpg",
+        "image_caption": [
+            "Thai Iced Tea:"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            494,
+            91,
+            653,
+            203
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Water (0.904)",
+        "bbox": [
+            504,
+            244,
+            596,
+            256
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Black Tea Leaf (0.811)",
+        "bbox": [
+            504,
+            256,
+            650,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "White Sugar (0.674)",
+        "bbox": [
+            504,
+            269,
+            638,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "image",
+        "img_path": "images/53c1225ed2d887cb7ecaff06cd74cfcb63db1b5df19bea3384fa86c8be04eedf.jpg",
+        "image_caption": [
+            "Marrakesh Vegetable Curry:",
+            "Figure 3: Case Study on the Recipe Dataset."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            682,
+            91,
+            843,
+            203
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Potato (0.881)",
+        "bbox": [
+            678,
+            244,
+            773,
+            256
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Eggplant (0.764)",
+        "bbox": [
+            678,
+            256,
+            788,
+            269
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Green Bell Pepper (0.653)",
+        "bbox": [
+            678,
+            269,
+            850,
+            282
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "box to a vector to bridge the semantics of the two views. Concept2Box is jointly trained to model the instance-view KG, the ontology-view KG, and the cross-view links. Empirical experiment results on two real-world datasets including a newly-created recipe dataset verify the effectiveness of Concept2Box.",
+        "bbox": [
+            112,
+            325,
+            489,
+            437
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            460,
+            250,
+            474
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our model is developed to tackle the structural difference between the ontology and instance views of a KG. However, many modern KGs are multilingual, where different portions of a KG may not only differ in structure but also differ in the text semantics. How to jointly capture these differences remain unsolved. Also since box embeddings naturally provide interpretability to the granularity of the learned concepts, how to use the current model to discover unknown concepts from these embeddings is also challenging.",
+        "bbox": [
+            112,
+            493,
+            489,
+            671
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "8 Ethical Impact",
+        "text_level": 1,
+        "bbox": [
+            112,
+            692,
+            278,
+            709
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Our paper proposed Concept2Box, a novel two-view knowledge graph embedding model with dual geometric representations. Concept2Box neither introduces any social/ethical bias to the model nor amplifies any bias in the data. For the created recipe dataset, we did not include any customers'/sellers' personal information and only preserved information related to products' attributes and their relations. Our model implementation is built upon public libraries in Pytorch. Therefore, was e do not foresee any direct social consequences or ethical issues.",
+        "bbox": [
+            112,
+            726,
+            490,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "9 Acknowledgement",
+        "text_level": 1,
+        "bbox": [
+            507,
+            324,
+            702,
+            341
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work was partially supported by NSF 1829071, 2106859, 2200274, 2211557, 1937599, 2119643, 2303037, DARPA #HR00112290103/HR0011260656, NASA,",
+        "bbox": [
+            507,
+            350,
+            884,
+            414
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Amazon Research Awards, and an NEC research award.",
+        "bbox": [
+            507,
+            414,
+            882,
+            445
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            473,
+            608,
+            488
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. Freebase: A collaboratively created graph database for structuring human knowledge. In Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data, SIGMOD '08, page 1247-1250.",
+            "Antoine Bordes, Nicolas Usunier, Alberto Garcia-Durán, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. In Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2, page 2787-2795.",
+            "Ines Chami, Zhitao Ying, Christopher Ré, and Jure Leskovec. 2019. Hyperbolic graph convolutional neural networks. In Advances in Neural Information Processing Systems, pages 4869-4880.",
+            "Xuelu Chen, Michael Boratko, Muhao Chen, Shib Sankar Dasgupta, Xiang Lorraine Li, and Andrew McCallum. 2021. Probabilistic box embeddings for uncertain knowledge graph reasoning. In Proceedings of the 19th Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL).",
+            "Xuelu Chen, Muhao Chen, Changjun Fan, Ankith Uppunda, Yizhou Sun, and Carlo Zaniolo. 2020. Multilingual knowledge graph completion via ensemble knowledge transfer. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: Findings, pages 3227-3238."
+        ],
+        "bbox": [
+            509,
+            495,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "10113",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Shib Dasgupta, Michael Boratko, Dongxu Zhang, Luke Vilnis, Xiang Li, and Andrew McCallum. 2020. Improving local identifiability in probabilistic box embeddings. Advances in Neural Information Processing Systems, 33:182-192.",
+            "Tim Dettmers, Minervini Pasquale, Stenetorp Pontus, and Sebastian Riedel. 2018. Convolutional 2d knowledge graph embeddings. In Proceedings of the 32th AAAI Conference on Artificial Intelligence, pages 1811-1818.",
+            "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.",
+            "Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. 2018. Hyperbolic entailment cones for learning hierarchical embeddings. In International Conference on Machine Learning, pages 1646-1655. PMLR.",
+            "Junheng Hao, Muhao Chen, Wenchao Yu, Yizhou Sun, and Wei Wang. 2019. Universal representation learning of knowledge bases by jointly embedding instances and ontological concepts. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '19, page 1709-1719.",
+            "Junheng Hao, Tong Zhao, Jin Li, Xin Luna Dong, Christos Faloutsos, Yizhou Sun, and Wei Wang. 2020. P-companion: A principled framework for diversified complementary product recommendation. In Proceedings of the 29th ACM International Conference on Information Knowledge Management, CIKM '20, page 2517-2524.",
+            "Zijie Huang, Zheng Li, Haoming Jiang, Tianyu Cao, Hanqing Lu, Bing Yin, Karthik Subbian, Yizhou Sun, and Wei Wang. 2022. Multilingual knowledge graph completion with self-supervised adaptive graph alignment. In Annual Meeting of the Association for Computational Linguistics (ACL).",
+            "Zijie Huang, Yizhou Sun, and Wei Wang. 2020. Learning continuous system dynamics from irregularly-sampled partial observations. In Advances in Neural Information Processing Systems.",
+            "Zijie Huang, Yizhou Sun, and Wei Wang. 2021. Coupled graph ode for learning interacting system dynamics. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.",
+            "Roshni G. Iyer, Yunsheng Bai, Wei Wang, and Yizhou Sun. 2022. Dual-geometric space embedding model for two-view knowledge graphs. In Proceedings"
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, page 676-686.",
+            "Amin Javari, Zhankui He, Zijie Huang, Raj Jeetu, and Kevin Chen-Chuan Chang. 2020. Weakly supervised attention for hashtag recommendation using graph data. In Proceedings of The Web Conference 2020, WWW '20.",
+            "Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. In International Conference on Learning Representations.",
+            "Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, and Soren Auer. 2015. Dbpedia-a large-scale, multilingual knowledge base extracted from wikipedia. In Semantic Web, pages 167-195.",
+            "Bill Yuchen Lin, Haitian Sun, Bhuwan Dhingra, Manzil Zaheer, Xiang Ren, and William Cohen. 2021. Differentiable open-ended commonsense reasoning. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4611-4625.",
+            "Boris Muzellec and Marco Cuturi. 2018. Generalizing point embeddings using the wasserstein space of elliptical distributions. NIPS'18, page 10258-10269.",
+            "Maximillian Nickel and Douwe Kiela. 2017. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30.",
+            "Yasumasa Onoe, Michael Boratko, Andrew McCallum, and Greg Durrett. 2021. Modeling fine-grained entity types with box embeddings. In ACL.",
+            "Hongyu Ren, Weihua Hu, and Jure Leskovec. 2020. Query2box: Reasoning over knowledge graphs in vector space using box embeddings. In International Conference on Learning Representations.",
+            "Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations.",
+            "Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48, ICML'16, page 2071-2080. JMLR.org.",
+            "Luke Vilnis, Xiang Lorraine Li, Shikhar Murty, and Andrew McCallum. 2018. Probabilistic embedding of knowledge graphs with box lattice measures. In ACL."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "10114",
+        "bbox": [
+            477,
+            928,
+            524,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ruijie Wang, Zijie Huang, Shengzhong Liu, Hua-jie Shao, Dongxin Liu, Jinyang Li, Tianshi Wang, Dachun Sun, Shuochao Yao, and Tarek Abdelzaher. 2021a. Dydiff-vae: A dynamic variational framework for information diffusion prediction. In SIGIR'21.",
+            "Shen Wang, Xiaokai Wei, Cicero Nogueira Nogueira dos Santos, Zhiguo Wang, Ramesh Nallapati, Andrew Arnold, Bing Xiang, Philip S. Yu, and Isabel F. Cruz. 2021b. Mixed-curvature multi-relational graph neural network for knowledge graph completion. In Proceedings of the Web Conference 2021, WWW '21, page 1761-1771.",
+            "Shen Wang, Xiaokai Wei, Cicero Nogueira Nogueira dos Santos, Zhiguo Wang, Ramesh Nallapati, Andrew Arnold, Bing Xiang, Philip S Yu, and Isabel F Cruz. 2021c. Mixed-curvature multi-relational graph neural network for knowledge graph completion. In Proceedings of the Web Conference 2021, pages 1761-1771.",
+            "Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. Proceedings of the 28th AAAI Conference on Artificial Intelligence, 28.",
+            "Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In International Conference on Learning Representations (ICLR).",
+            "Liang Yao, Chengsheng Mao, and Yuan Luo. 2020. Kgbert: Bert for knowledge graph completion. Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence.",
+            "Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. 2021. Qa-gnn: Reasoning with language models and knowledge graphs for question answering. In North American Chapter of the Association for Computational Linguistics (NAACL).",
+            "Zhao Zhang, Fuzhen Zhuang, Hengshu Zhu, Zhiping Shi, Hui Xiong, and Qing He. 2020. Relational graph neural network with hierarchical attention for knowledge graph completion. Proceedings of the AAAI Conference on Artificial Intelligence, pages 9612-9619."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            746
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "A Data Construction",
+        "text_level": 1,
+        "bbox": [
+            509,
+            83,
+            709,
+            98
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We introduce the data generation process of the Recipe dataset. The overall data structure is depicted in Fig 4. For the ontology-view KG, concepts consist of recipes, ingredients, and utensils for cooking them and their associated types. Dietary styles and regions where each recipe originated from are also considered concepts. We obtain these high-level recipe-related concepts from recipeDB with some data cleaning procedures. Triples in the ontology-view KG describe the connectivity among these concepts such as (Kungpao Chicken, useingredient, Chicken Breast), (Salt, is_of_ingredient_type, Additive).",
+        "bbox": [
+            507,
+            109,
+            884,
+            318
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "For the instance-view KG, entities are real ingredient and utensil products sold on Amazon.com by searching their corresponding concept names (ingredient and utensil names) as keywords. For each ingredient and utensil name, we select the top 10 matched products as entities in the instance-view KG. Afterward, we manually select 8 representative attributes such as flavor and brand, that are commonly shared by these products as additional entities in the instance-view KG, to provide more knowledge for the embedding learning process. The triples in the instance-view KG describe facts connecting a product to its attributes.",
+        "bbox": [
+            507,
+            319,
+            884,
+            527
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "B Implementation Details.",
+        "text_level": 1,
+        "bbox": [
+            507,
+            539,
+            754,
+            556
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "We use Adam (Kingma and Ba, 2014) as the optimizer to train our model and use TransE (Bordes et al., 2013) as the KGE model for learning the instance-view KG, whose margin  $\\gamma^{\\mathrm{KG}}$  is set to be 0.3. The margin for cross-view module  $\\gamma^{\\mathrm{Cross}}$  is set to be 0.15. We set the latent dimensions for instance-view KG as 256, and the box dimensions for ontology-view KG as 512. We use a batch size of 128 and a learning rate  $lr = 0.005$  during training.",
+        "bbox": [
+            507,
+            565,
+            882,
+            725
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "text",
+        "text": "Instead of directly optimizing  $\\mathcal{I}$  in Eqn 11, we alternately update  $\\mathcal{J}^{\\mathcal{G}^{\\mathcal{O}}}$ ,  $\\mathcal{J}^{\\mathcal{G}^{\\mathcal{I}}}$  and  $\\mathcal{J}^{\\mathrm{Cross}}$  with different learning rate. Specifically, in our implementation, we optimize with  $\\theta_{new} \\gets \\theta_{old} - \\eta \\nabla \\mathcal{J}^{\\mathcal{G}^{\\mathcal{O}}}$ ,  $\\theta_{new} \\gets \\theta_{old} - (\\lambda_1\\eta)\\nabla \\mathcal{J}^{\\mathcal{G}^{\\mathcal{I}}}$ ,  $\\theta_{new} \\gets \\theta_{old} - (\\lambda_2\\eta)\\nabla \\mathcal{J}^{\\mathrm{cross}}$  in consecutive steps within one epoch, where  $\\theta_{new}$  denotes our model parameters. The procedure is depicted in Algorithm 1.",
+        "bbox": [
+            507,
+            726,
+            882,
+            854
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://cosylab.iitd.edu.in/recipientdb/>",
+        "bbox": [
+            529,
+            903,
+            759,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "10115",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "image",
+        "img_path": "images/6cbe4cda457fa3bd0ca7286db4b640b3c5fc9fe94ef86863d1482cb19e20c501.jpg",
+        "image_caption": [
+            "Figure 4: Recipe Dataset Structure"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            144,
+            185,
+            845,
+            340
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "table",
+        "img_path": "images/c5c8db172f2fdf3b59bf9d9e9b9b6f23b1c468b7755f15f1b9a208dcb4fd9b4f.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Algorithm 1: Concept2Box training pro-cure.</td></tr><tr><td colspan=\"2\">Input: A two-view KG G. Output: Model parameters θ.</td></tr><tr><td>1</td><td>while model not converged do</td></tr><tr><td>2</td><td>//For the KGC loss of G&#x27;O:</td></tr><tr><td>3</td><td>Optimize the box-based KGC loss in Eqn 6:</td></tr><tr><td>4</td><td>θnew← θold - η∇G&#x27;O</td></tr><tr><td>5</td><td>//For the KGC loss of G&#x27;I:</td></tr><tr><td>6</td><td>Optimize the vector-based KGC loss in Eqn 7:</td></tr><tr><td>7</td><td>θnew← θold - (λ1η)∇G&#x27;I</td></tr><tr><td>8</td><td>//For the cross-view loss:</td></tr><tr><td>9</td><td>Optimize with the cross-view loss in Eqn 10:</td></tr><tr><td>10</td><td>θnew← θold - (λ2η)∇Jcross</td></tr><tr><td>11</td><td>end</td></tr></table>",
+        "bbox": [
+            115,
+            527,
+            485,
+            822
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "10116",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "A For every submission:",
+        "bbox": [
+            115,
+            107,
+            322,
+            122
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "A1. Did you describe the limitations of your work? Section 7.",
+            "A2. Did you discuss any potential risks of your work? No risk concerns for this work.",
+            "A3. Do the abstract and introduction summarize the paper's main claims? Sec.1 for the introduction.",
+            "A4. Have you used AI writing assistants when working on this paper? Left blank."
+        ],
+        "bbox": [
+            129,
+            126,
+            695,
+            287
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "B Did you use or create scientific artifacts?",
+        "bbox": [
+            114,
+            300,
+            489,
+            316
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            321,
+            215,
+            336
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "B1. Did you cite the creators of artifacts you used? No response.",
+            "B2. Did you discuss the license or terms for use and / or distribution of any artifacts? No response.",
+            "B3. Did you discuss if your use of existing artifact(s) was consistent with their intended use, provided that it was specified? For the artifacts you create, do you specify intended use and whether that is compatible with the original access conditions (in particular, derivatives of data accessed for research purposes should not be used outside of research contexts)? No response.",
+            "B4. Did you discuss the steps taken to check whether the data that was collected / used contains any information that names or uniquely identifies individual people or offensive content, and the steps taken to protect / anonymize it? No response.",
+            "B5. Did you provide documentation of the artifacts, e.g., coverage of domains, languages, and linguistic phenomena, demographic groups represented, etc.? No response.",
+            "B6. Did you report relevant statistics like the number of examples, details of train / test / dev splits, etc. for the data that you used / created? Even for commonly-used benchmark datasets, include the number of examples in train / validation / test splits, as these provide necessary context for a reader to understand experimental results. For example, small differences in accuracy on large test sets may be significant, while on small test sets they may not be. No response."
+        ],
+        "bbox": [
+            127,
+            347,
+            880,
+            753
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C Did you run computational experiments?",
+        "bbox": [
+            114,
+            764,
+            492,
+            781
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Left blank.",
+        "bbox": [
+            132,
+            787,
+            213,
+            801
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C1. Did you report the number of parameters in the models used, the total computational budget (e.g., GPU hours), and computing infrastructure used? Left blank.",
+        "bbox": [
+            129,
+            812,
+            880,
+            860
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "The Responsible NLP Checklist used at ACL 2023 is adopted from NAACL 2022, with the addition of a question on AI writing assistance.",
+        "bbox": [
+            112,
+            868,
+            877,
+            892
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "header",
+        "text": "ACL 2023 Responsible NLP Checklist",
+        "bbox": [
+            132,
+            84,
+            433,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "10117",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "C2. Did you discuss the experimental setup, including hyperparameter search and best-found hyperparameter values? Left blank.",
+            "C3. Did you report descriptive statistics about your results (e.g., error bars around results, summary statistics from sets of experiments), and is it transparent whether you are reporting the max, mean, etc. or just a single run? Left blank.",
+            "C4. If you used existing packages (e.g., for preprocessing, for normalization, or for evaluation), did you report the implementation, model, and parameter settings used (e.g., NLTK, Spacy, ROUGE, etc.)? Left blank."
+        ],
+        "bbox": [
+            129,
+            83,
+            878,
+            282
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "D Did you use human annotators (e.g., crowdworkers) or research with human participants? Left blank.",
+        "bbox": [
+            112,
+            293,
+            877,
+            330
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "D1. Did you report the full text of instructions given to participants, including e.g., screenshots, disclaimers of any risks to participants or annotators, etc.? No response.",
+            "D2. Did you report information about how you recruited (e.g., crowdsourcing platform, students) and paid participants, and discuss if such payment is adequate given the participants' demographic (e.g., country of residence)? No response.",
+            "D3. Did you discuss whether and how consent was obtained from people whose data you're using/curating? For example, if you collected data via crowdsourcing, did your instructions to crowdworkers explain how the data would be used? No response.",
+            "D4. Was the data collection protocol approved (or determined exempt) by an ethics review board? No response.",
+            "D5. Did you report the basic demographic and geographic characteristics of the annotator population that is the source of the data? No response."
+        ],
+        "bbox": [
+            127,
+            340,
+            878,
+            640
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "10118",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 13
+    }
+]