Add Batch 56780c47-647b-47c5-9c21-04bf2084bb02 data

.gitattributes

@@ -7193,3 +7193,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2023/Length[[:space:]]is[[:space:]]a[[:space:]]Curse[[:space:]]and[[:space:]]a[[:space:]]Blessing[[:space:]]for[[:space:]]Document-level[[:space:]]Semantics/72fb1c19-e28c-4d00-91c0-c703f771355e_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Let[[:space:]]GPT[[:space:]]be[[:space:]]a[[:space:]]Math[[:space:]]Tutor_[[:space:]]Teaching[[:space:]]Math[[:space:]]Word[[:space:]]Problem[[:space:]]Solvers[[:space:]]with[[:space:]]Customized[[:space:]]Exercise[[:space:]]Generation/cce65de3-d100-49f9-abfb-d3d917648e30_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2023/Let’s[[:space:]]Sample[[:space:]]Step[[:space:]]by[[:space:]]Step_[[:space:]]Adaptive-Consistency[[:space:]]for[[:space:]]Efficient[[:space:]]Reasoning[[:space:]]and[[:space:]]Coding[[:space:]]with[[:space:]]LLMs/87f84b4a-fa49-4c49-98e9-354a13ef41c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Let’s[[:space:]]Think[[:space:]]Frame[[:space:]]by[[:space:]]Frame[[:space:]]with[[:space:]]VIP_[[:space:]]A[[:space:]]Video[[:space:]]Infilling[[:space:]]and[[:space:]]Prediction[[:space:]]Dataset[[:space:]]for[[:space:]]Evaluating[[:space:]]Video[[:space:]]Chain-of-Thought/4f07856a-09e6-4f3e-ad73-e9ecc85309e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lifelong[[:space:]]Sequence[[:space:]]Generation[[:space:]]with[[:space:]]Dynamic[[:space:]]Module[[:space:]]Expansion[[:space:]]and[[:space:]]Adaptation/0474ffe8-7b94-4eda-aa63-2522d4a8cfa0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Linear-Time[[:space:]]Modeling[[:space:]]of[[:space:]]Linguistic[[:space:]]Structure_[[:space:]]An[[:space:]]Order-Theoretic[[:space:]]Perspective/88551d68-b602-471f-9566-62b6a0746f09_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Ling-CL_[[:space:]]Understanding[[:space:]]NLP[[:space:]]Models[[:space:]]through[[:space:]]Linguistic[[:space:]]Curricula/555b9eb9-8ec1-48b5-b32f-200ee26764bf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Linking[[:space:]]Surface[[:space:]]Facts[[:space:]]to[[:space:]]Large-Scale[[:space:]]Knowledge[[:space:]]Graphs/9a8770e9-9bcc-4ad5-8318-6ea414c765f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lion_[[:space:]]Adversarial[[:space:]]Distillation[[:space:]]of[[:space:]]Proprietary[[:space:]]Large[[:space:]]Language[[:space:]]Models/317da3bb-97cc-44eb-b7cf-4be112125836_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Localizing[[:space:]]Active[[:space:]]Objects[[:space:]]from[[:space:]]Egocentric[[:space:]]Vision[[:space:]]with[[:space:]]Symbolic[[:space:]]World[[:space:]]Knowledge/2bad6413-23a3-4118-b2c2-882a5a9baf49_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Location-Aware[[:space:]]Visual[[:space:]]Question[[:space:]]Generation[[:space:]]with[[:space:]]Lightweight[[:space:]]Models/0e4f617e-b625-4a64-9c41-5b1c70b50c7f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Log-FGAER_[[:space:]]Logic-Guided[[:space:]]Fine-Grained[[:space:]]Address[[:space:]]Entity[[:space:]]Recognition[[:space:]]from[[:space:]]Multi-Turn[[:space:]]Spoken[[:space:]]Dialogue/4166b175-fae8-4a02-8116-9769fd80da5b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Longtriever_[[:space:]]a[[:space:]]Pre-trained[[:space:]]Long[[:space:]]Text[[:space:]]Encoder[[:space:]]for[[:space:]]Dense[[:space:]]Document[[:space:]]Retrieval/ec8e6764-9d98-4bdb-8fe9-a4eb590ae00f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Look-back[[:space:]]Decoding[[:space:]]for[[:space:]]Open-Ended[[:space:]]Text[[:space:]]Generation/d72bedbc-93c1-4f69-9dab-6e80120ec201_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Lost[[:space:]]in[[:space:]]Translation,[[:space:]]Found[[:space:]]in[[:space:]]Spans_[[:space:]]Identifying[[:space:]]Claims[[:space:]]in[[:space:]]Multilingual[[:space:]]Social[[:space:]]Media/1f4e1d6c-0296-4ade-9d3d-96707fd503e9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/M2DF_[[:space:]]Multi-grained[[:space:]]Multi-curriculum[[:space:]]Denoising[[:space:]]Framework[[:space:]]for[[:space:]]Multimodal[[:space:]]Aspect-based[[:space:]]Sentiment[[:space:]]Analysis/9c9dc993-9770-4a86-882e-097c303dfef0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/M3Seg_[[:space:]]A[[:space:]]Maximum-Minimum[[:space:]]Mutual[[:space:]]Information[[:space:]]Paradigm[[:space:]]for[[:space:]]Unsupervised[[:space:]]Topic[[:space:]]Segmentation[[:space:]]in[[:space:]]ASR[[:space:]]Transcripts/fe468f50-969b-4694-b6d7-7ed30c9af0db_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MADNet_[[:space:]]Maximizing[[:space:]]Addressee[[:space:]]Deduction[[:space:]]Expectation[[:space:]]for[[:space:]]Multi-Party[[:space:]]Conversation[[:space:]]Generation/8463e9b5-74e6-4599-89c0-ed34069e8010_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MAF_[[:space:]]Multi-Aspect[[:space:]]Feedback[[:space:]]for[[:space:]]Improving[[:space:]]Reasoning[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/9e520e38-75e1-4f3b-92f1-e53f1dd3cb44_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MAGNIFICo_[[:space:]]Evaluating[[:space:]]the[[:space:]]In-Context[[:space:]]Learning[[:space:]]Ability[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]to[[:space:]]Generalize[[:space:]]to[[:space:]]Novel[[:space:]]Interpretations/9eb8af8c-c614-4b7a-8c59-0687f598cc8a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MAUD_[[:space:]]An[[:space:]]Expert-Annotated[[:space:]]Legal[[:space:]]NLP[[:space:]]Dataset[[:space:]]for[[:space:]]Merger[[:space:]]Agreement[[:space:]]Understanding/007cdf79-c4ab-480b-a8e2-6e57d205a8ee_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MEGA_[[:space:]]Multilingual[[:space:]]Evaluation[[:space:]]of[[:space:]]Generative[[:space:]]AI/aff4a3ab-bf29-405f-881c-99ee9f1a2833_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MILDSum_[[:space:]]A[[:space:]]Novel[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]for[[:space:]]Multilingual[[:space:]]Summarization[[:space:]]of[[:space:]]Indian[[:space:]]Legal[[:space:]]Case[[:space:]]Judgments/d31d32cb-7026-446a-b615-3976794a6972_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MMNMT_[[:space:]]Modularizing[[:space:]]Multilingual[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation[[:space:]]with[[:space:]]Flexibly[[:space:]]Assembled[[:space:]]MoE[[:space:]]and[[:space:]]Dense[[:space:]]Blocks/5f10c047-6d3e-44a4-b36b-8cc567d4f0f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MProto_[[:space:]]Multi-Prototype[[:space:]]Network[[:space:]]with[[:space:]]Denoised[[:space:]]Optimal[[:space:]]Transport[[:space:]]for[[:space:]]Distantly[[:space:]]Supervised[[:space:]]Named[[:space:]]Entity[[:space:]]Recognition/ccd60ec6-5be1-402b-8b03-ea42740b4112_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MQuAKE_[[:space:]]Assessing[[:space:]]Knowledge[[:space:]]Editing[[:space:]]in[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Multi-Hop[[:space:]]Questions/1715db53-750e-41ae-bb12-acc4b2aae009_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MT2_[[:space:]]Towards[[:space:]]a[[:space:]]Multi-Task[[:space:]]Machine[[:space:]]Translation[[:space:]]Model[[:space:]]with[[:space:]]Translation-Specific[[:space:]]In-Context[[:space:]]Learning/e7a23c2b-0a07-42dc-a946-81a71d14c549_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MULTITuDE_[[:space:]]Large-Scale[[:space:]]Multilingual[[:space:]]Machine-Generated[[:space:]]Text[[:space:]]Detection[[:space:]]Benchmark/ee2f47ee-d4fe-48af-a30c-309db5da6476_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MaNtLE_[[:space:]]Model-agnostic[[:space:]]Natural[[:space:]]Language[[:space:]]Explainer/2e496c96-92d8-4ea6-96de-df3a6f2531c6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MailEx_[[:space:]]Email[[:space:]]Event[[:space:]]and[[:space:]]Argument[[:space:]]Extraction/b61c58ac-1484-4cfe-8f60-2e7f26dfc87a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Make[[:space:]]Every[[:space:]]Example[[:space:]]Count_[[:space:]]On[[:space:]]the[[:space:]]Stability[[:space:]]and[[:space:]]Utility[[:space:]]of[[:space:]]Self-Influence[[:space:]]for[[:space:]]Learning[[:space:]]from[[:space:]]Noisy[[:space:]]NLP[[:space:]]Datasets/10a9fc90-064c-415d-bd49-30f3f105723c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Making[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Better[[:space:]]Data[[:space:]]Creators/065c3c0d-dd01-4b5a-986a-d256f1c38e76_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MarkQA_[[:space:]]A[[:space:]]large[[:space:]]scale[[:space:]]KBQA[[:space:]]dataset[[:space:]]with[[:space:]]numerical[[:space:]]reasoning/8130c6c2-5d4b-46fc-bb27-689c8d6f26a2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MeaeQ_[[:space:]]Mount[[:space:]]Model[[:space:]]Extraction[[:space:]]Attacks[[:space:]]with[[:space:]]Efficient[[:space:]]Queries/0a899d4b-02c7-4726-9790-e10e6159685f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MedEval_[[:space:]]A[[:space:]]Multi-Level,[[:space:]]Multi-Task,[[:space:]]and[[:space:]]Multi-Domain[[:space:]]Medical[[:space:]]Benchmark[[:space:]]for[[:space:]]Language[[:space:]]Model[[:space:]]Evaluation/d4dbe76e-e627-4ae4-baad-c451803c2208_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MediaHG_[[:space:]]Rethinking[[:space:]]Eye-catchy[[:space:]]Features[[:space:]]in[[:space:]]Social[[:space:]]Media[[:space:]]Headline[[:space:]]Generation/b0ad8624-11d5-47cf-9d52-1f6ca4d52e78_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MemeCap_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Captioning[[:space:]]and[[:space:]]Interpreting[[:space:]]Memes/d5dba9d1-734b-4db7-bc9a-d23be682cc90_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Memorisation[[:space:]]Cartography_[[:space:]]Mapping[[:space:]]out[[:space:]]the[[:space:]]Memorisation-Generalisation[[:space:]]Continuum[[:space:]]in[[:space:]]Neural[[:space:]]Machine[[:space:]]Translation/08b9fc11-1171-41ef-9508-44e34806c788_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Memory-Based[[:space:]]Invariance[[:space:]]Learning[[:space:]]for[[:space:]]Out-of-Domain[[:space:]]Text[[:space:]]Classification/7e3fcbed-a432-477f-9e0f-3f710fc80cc0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Merging[[:space:]]Experts[[:space:]]into[[:space:]]One_[[:space:]]Improving[[:space:]]Computational[[:space:]]Efficiency[[:space:]]of[[:space:]]Mixture[[:space:]]of[[:space:]]Experts/61bbb5db-7cbd-441a-9ca6-c5c5ae84501a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Merging[[:space:]]Generated[[:space:]]and[[:space:]]Retrieved[[:space:]]Knowledge[[:space:]]for[[:space:]]Open-Domain[[:space:]]QA/1c46efc0-f189-4a20-9ce6-ec0282011869_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Meta-Learning[[:space:]]Online[[:space:]]Adaptation[[:space:]]of[[:space:]]Language[[:space:]]Models/b6be4825-85b3-4929-b1ec-a9ea4dfd7631_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MingOfficial_[[:space:]]A[[:space:]]Ming[[:space:]]Official[[:space:]]Career[[:space:]]Dataset[[:space:]]and[[:space:]]a[[:space:]]Historical[[:space:]]Context-Aware[[:space:]]Representation[[:space:]]Learning[[:space:]]Framework/e7f25693-6175-426a-ab34-32683e7e5b12_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mirages.[[:space:]]On[[:space:]]Anthropomorphism[[:space:]]in[[:space:]]Dialogue[[:space:]]Systems/7b27b482-f385-4711-9a61-e3f6ce3c7292_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mirror_[[:space:]]A[[:space:]]Universal[[:space:]]Framework[[:space:]]for[[:space:]]Various[[:space:]]Information[[:space:]]Extraction[[:space:]]Tasks/38e6604e-3823-4ee4-aef4-e7310f264e75_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mitigating[[:space:]]Backdoor[[:space:]]Poisoning[[:space:]]Attacks[[:space:]]through[[:space:]]the[[:space:]]Lens[[:space:]]of[[:space:]]Spurious[[:space:]]Correlation/a4767899-feeb-4a61-a27e-1023e79a902a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mitigating[[:space:]]Over-Generation[[:space:]]for[[:space:]]Unsupervised[[:space:]]Keyphrase[[:space:]]Extraction[[:space:]]with[[:space:]]Heterogeneous[[:space:]]Centrality[[:space:]]Detection/f956e7a5-13a7-4a49-90cb-cd291f170269_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Mitigating[[:space:]]Temporal[[:space:]]Misalignment[[:space:]]by[[:space:]]Discarding[[:space:]]Outdated[[:space:]]Facts/62633764-09c9-492c-810d-54e26ba434f5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MoPe_[[:space:]]Model[[:space:]]Perturbation[[:space:]]based[[:space:]]Privacy[[:space:]]Attacks[[:space:]]on[[:space:]]Language[[:space:]]Models/4367b34e-a971-42d3-b4f1-85ca4b4c543b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MoT_[[:space:]]Memory-of-Thought[[:space:]]Enables[[:space:]]ChatGPT[[:space:]]to[[:space:]]Self-Improve/aa47f336-4e55-4e67-b3fe-628cc2e4e9e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Model-tuning[[:space:]]Via[[:space:]]Prompts[[:space:]]Makes[[:space:]]NLP[[:space:]]Models[[:space:]]Adversarially[[:space:]]Robust/af751e26-66cc-4f62-b977-76bb96609f26_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Modeling[[:space:]]Conceptual[[:space:]]Attribute[[:space:]]Likeness[[:space:]]and[[:space:]]Domain[[:space:]]Inconsistency[[:space:]]for[[:space:]]Metaphor[[:space:]]Detection/df86c499-66b6-4e51-9add-aa7b946b4f8b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Modeling[[:space:]]Empathic[[:space:]]Similarity[[:space:]]in[[:space:]]Personal[[:space:]]Narratives/18c7013c-4a98-4f43-a26c-844ef748fedb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Modeling[[:space:]]Legal[[:space:]]Reasoning_[[:space:]]LM[[:space:]]Annotation[[:space:]]at[[:space:]]the[[:space:]]Edge[[:space:]]of[[:space:]]Human[[:space:]]Agreement/14a02cb9-4c37-4731-88ad-4215bf7033b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Models[[:space:]]See[[:space:]]Hallucinations_[[:space:]]Evaluating[[:space:]]the[[:space:]]Factuality[[:space:]]in[[:space:]]Video[[:space:]]Captioning/1a8b7244-5b39-4253-8c2c-d062d3456fea_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MolCA_[[:space:]]Molecular[[:space:]]Graph-Language[[:space:]]Modeling[[:space:]]with[[:space:]]Cross-Modal[[:space:]]Projector[[:space:]]and[[:space:]]Uni-Modal[[:space:]]Adapter/3acb0b16-0421-40e4-ae30-e4f7a9f4fa9f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/More[[:space:]]Than[[:space:]]Spoken[[:space:]]Words_[[:space:]]Nonverbal[[:space:]]Message[[:space:]]Extraction[[:space:]]and[[:space:]]Generation/b38727a0-6cff-45cf-ac6a-4b5e01335e85_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-Source[[:space:]]Multi-Type[[:space:]]Knowledge[[:space:]]Exploration[[:space:]]and[[:space:]]Exploitation[[:space:]]for[[:space:]]Dialogue[[:space:]]Generation/ec09dcf9-3d72-4c25-988e-fa82eabc5949_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-Source[[:space:]]Probing[[:space:]]for[[:space:]]Open-Domain[[:space:]]Conversational[[:space:]]Understanding/844b8d54-0fc3-4f4f-bb2a-6bac880f6029_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-Task[[:space:]]Knowledge[[:space:]]Distillation[[:space:]]with[[:space:]]Embedding[[:space:]]Constraints[[:space:]]for[[:space:]]Scholarly[[:space:]]Keyphrase[[:space:]]Boundary[[:space:]]Classification/fa34acd0-77b2-43a9-8384-815473d0290e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-level[[:space:]]Adaptive[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Knowledge[[:space:]]Internalization[[:space:]]in[[:space:]]Dialogue[[:space:]]Generation/203255b7-7b35-4ddb-a668-369e58aa7fcf_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-level[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Script-based[[:space:]]Character[[:space:]]Understanding/2f394ab4-d6da-450d-bc33-69a322119ad2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multi-view[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Entity[[:space:]]Typing[[:space:]]over[[:space:]]Knowledge[[:space:]]Graphs/7c331706-974a-407e-98ce-3cd184c9de3b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/MultiTurnCleanup_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Multi-Turn[[:space:]]Spoken[[:space:]]Conversational[[:space:]]Transcript[[:space:]]Cleanup/85b186d4-64ac-40b2-a013-f27d644389ad_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multilingual[[:space:]]Holistic[[:space:]]Bias_[[:space:]]Extending[[:space:]]Descriptors[[:space:]]and[[:space:]]Patterns[[:space:]]to[[:space:]]Unveil[[:space:]]Demographic[[:space:]]Biases[[:space:]]in[[:space:]]Languages[[:space:]]at[[:space:]]Scale/19eb0759-3301-4b8b-b16a-18078864402b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multilingual[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Are[[:space:]]Not[[:space:]](Yet)[[:space:]]Code-Switchers/aaddb6b8-305c-4641-933d-66442103f062_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2023/Multilingual[[:space:]]Pixel[[:space:]]Representations[[:space:]]for[[:space:]]Translation[[:space:]]and[[:space:]]Effective[[:space:]]Cross-lingual[[:space:]]Transfer/41b5f147-ea83-4898-b04e-3dc4a13da7af_origin.pdf filter=lfs diff=lfs merge=lfs -text

2023/Let’s Think Frame by Frame with VIP_ A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought/4f07856a-09e6-4f3e-ad73-e9ecc85309e4_content_list.json

@@ -0,0 +1,1936 @@
+[
+    {
+        "type": "text",
+        "text": "Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought",
+        "text_level": 1,
+        "bbox": [
+            129,
+            87,
+            868,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Vaishnavi Himakunthala*, Andy Ouyang*, Daniel Rose*, Ryan He*, Alex Mei, Yujie Lu, Chinmay Sonar, Michael Saxon, William Yang Wang",
+        "bbox": [
+            159,
+            142,
+            840,
+            175
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "University of California, Santa Barbara",
+        "bbox": [
+            341,
+            177,
+            660,
+            192
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{vaishnavi, andyouyang, danielrose, ryanhe, yujielu, saxon}@ucsb.edu {alexmei, csonar, william}@cs.ucsb.edu",
+        "bbox": [
+            156,
+            193,
+            845,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite exciting recent results showing vision-language systems' capacity to reason about images using natural language, their capacity for video reasoning remains underexplored. We motivate framing video reasoning as the sequential understanding of a small number of keyframes, thereby leveraging the power and robustness of vision-language while alleviating the computational complexities of processing videos. To evaluate this novel application, we introduce VIP<sup>1</sup>, an inference-time challenge dataset designed to explore models' reasoning capabilities through video chain-of-thought. Inspired by visually descriptive scene plays, we propose two formats for keyframe description: unstructured dense captions and structured scene descriptions that identify the focus, action, mood, objects, and setting (FAMOuS) of the keyframe. To evaluate video reasoning, we propose two tasks: Video Infilling and Video Prediction, which test abilities to generate multiple intermediate keyframes and predict future keyframes, respectively. We benchmark GPT-4, GPT-3, and VICUNA on VIP, demonstrate the performance gap in these complex video reasoning tasks, and encourage future work to prioritize language models for efficient and generalized video reasoning.",
+        "bbox": [
+            142,
+            282,
+            460,
+            680
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            695,
+            260,
+            709
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Constituting  $65\\%$  of all internet traffic in 2023, videos are an area of huge potential for the next chapter of leveraging artificial intelligence (Fu et al., 2021; Zellers et al., 2021; Fu et al., 2023a). For example, Video Question Answering (Lei et al., 2018) and Video Summarization (Xu et al., 2016) are two existing datasets that empirically evaluate video models. Yet, they do not assess more challenging tasks, such as reasoning through specific relationships between multiple frames. Just",
+        "bbox": [
+            112,
+            721,
+            490,
+            883
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "like how humans understand videos by processing frames across time steps, AI's ability to accomplish multi-frame reasoning is a core task for video understanding.",
+        "bbox": [
+            507,
+            253,
+            884,
+            316
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Multi-frame video reasoning is bottlenecked by the sheer computing resources needed to process videos, which typically contain 24 frames per second and can vary widely. However, intelligible videos tend to have little variation from frame to frame. Akin to selecting principle components on the axes containing the highest orthogonal variance, picking a small sample of keyframes can capture much of the video's meaning. Understanding a video via extracted keyframes poses the challenge of multi-hop reasoning, which requires stronger language model capabilities.",
+        "bbox": [
+            507,
+            319,
+            884,
+            512
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To evaluate this challenging multi-hop multiframe video reasoning, we elicit video chain-of-thought (VIDEOCOT). We propose two tasks to test such capabilities in existing models - Video Infilling and Video Prediction. In the Video Infilling setting, the goal is to predict the masked keyframes' descriptions when given the previous and next keyframes' descriptions in sequence as context, following the spirit of masked language modeling. In the Video Prediction, the goal is to predict the descriptions of the next keyframes in the sequence when given a set of previous keyframes' descriptions, similar to the next token prediction task. These tasks can help analyze video generation and whether video models truly understand the dynamic relations between subsequent video frames, given the variable context gap between keyframes.",
+        "bbox": [
+            507,
+            514,
+            884,
+            788
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To benchmark our proposed Video Infilling and Prediction tasks, we construct a dataset by proposing an automated method to extract keyframes. In addition to the frame itself, we propose two textual representations – unstructured dense captions and FAMouS structured scene descriptions (Figure 1). The unstructured captions are intended to extract more significant and visually descriptive informa",
+        "bbox": [
+            507,
+            790,
+            885,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "\\*Denotes equal contribution. 1https://github.com/vaishnaviHimakunthala/VIP",
+        "bbox": [
+            137,
+            891,
+            431,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "204",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 204-219 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            166,
+            945,
+            828,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/519217f7101b82444bf1b32abe8a5d8a8dd4ac083767d022567e1c117d760e80.jpg",
+        "image_caption": [
+            "Figure 1: The Video Infilling and Prediction Dataset consists of two ways to describe keyframes: an unstructured dense caption and a structured scene description with five components: Focus, Action, Mood, Objects, and Setting (FAMOUS). The unstructured dense captions are highly detailed dense captions that can promote visually descriptive reasoning tasks, while structured scene description provide a concise, visual description of the keyframe that can aid in more focused reasoning tasks."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            127,
+            92,
+            857,
+            334
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "tion compared to existing captioning systems, a necessary component to provide enough context for more challenging tasks. In addition, we systematically create structured scene descriptions from these unstructured dense captions by extracting the frame's focus, action, mood, objects, and setting using weak human supervision for quality assurance. The FAMOuS categories are inspired by play scripts, which maintain much of the visual detail of the unstructured captions while providing a clear, structured way to reason through visual tasks with high degrees of freedom.",
+        "bbox": [
+            112,
+            453,
+            487,
+            645
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We propose the following contributions:",
+        "bbox": [
+            132,
+            649,
+            433,
+            665
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We systematically collect an inference-time challenge dataset of keyframes for video reasoning augmented with two textual representations: unstructured dense captions for visually-descriptive information and FAMOuS scene descriptions for structured reasoning.",
+            "- We propose the Video Infilling and Video Prediction tasks to benchmark the video chain-of-thought capabilities in existing models.",
+            "- We empirically demonstrate that existing models have the potential for multi-hop multi-frame video reasoning tasks but have a significant area for improvement as future work."
+        ],
+        "bbox": [
+            114,
+            687,
+            490,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            509,
+            451,
+            663,
+            467
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "AI Reasoning. Large language models (LLMs) demonstrate considerable gains on existing reasoning benchmarks with strategies such as chain-of-thought (CoT) (Wei et al., 2022) and few-shot demonstrations (Brown et al., 2020). Vision-language models (VLM) (Alayrac et al., 2022; Chowdhery et al., 2022; Driess et al., 2023) have furthered LLMs' capabilities by adding the visual modality to perform tasks such as visually-guided text generation (Rose et al., 2023; Zhu et al., 2023), vision question-answering (Wang et al., 2022a; Kim et al., 2021), and image captioning (Li et al., 2022; Liu et al., 2023). We believe the logical next step is to extend these existing models to the video domain. VIP's annotated collection of extracted keyframes from real-world videos offers a resource to evaluate reasoning abilities within the video domain.",
+        "bbox": [
+            507,
+            476,
+            884,
+            765
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Datasets for Video Understanding. Existing video datasets are often limited by domain specificity or require a supplementary representation (e.g., audio, text) (Lei et al., 2018, 2020a; Tapaswi et al., 2016; Miech et al., 2019; Kim et al., 2017; Mun et al., 2017). These datasets also provide simplifications as textual summaries (Xu et al., 2016; Guadarrama et al., 2013) or a single video frame (Yu et al., 2019; Zeng et al., 2017; Maharaj et al.,",
+        "bbox": [
+            507,
+            774,
+            884,
+            919
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "205",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2017), which by themselves can be sufficient to complete the task. While some datasets consider the multi-frame component (Jang et al., 2017; Yi et al., 2020; Mun et al., 2017; Fu et al., 2023b) for higher order complexity, VIP differs in trying to reduce the computational intensity of video reasoning without reducing the task difficulty and generality. VIP is a dataset of real-life videos that spans a breadth of domains and assesses multi-hop, multiframe video reasoning without requiring significant computation to train on videos.",
+        "bbox": [
+            112,
+            84,
+            492,
+            261
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Textual Representations of Videos. Early video models are trained with visual keyframes and textual questions as input and return textual answers as output (Sukhbaatar et al., 2015; Kim et al., 2017; Jang et al., 2017). Then, researchers started to unify the video, keyframe, and text embedding spaces (Miech et al., 2019; Kim et al., 2018; Zellers et al., 2021; Kim et al., 2017; Guadarrama et al., 2013; Bhattacharya et al., 2023). VidIL leverages the contemporary in-context inference paradigm with few-shot demonstrations, including frames, captions, and visual tokens to prompt language models to solve VidL tasks (Lei et al., 2020b; Wang et al., 2022b). In contrast to these existing works, VIP introduces textual representations at the keyframe level and then leverages them to reason about specific video segments using VideoCOT.",
+        "bbox": [
+            115,
+            272,
+            490,
+            546
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 VIP Dataset Construction",
+        "text_level": 1,
+        "bbox": [
+            112,
+            560,
+            371,
+            574
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To construct the dataset, we first outsource the video corpus to stem from the YouTube-8m dataset (Abu-El-Haija et al., 2016), whose diverse, realistic videos with human-labeled topics align well with our desiderata. To effectively enable multi-frame video reasoning, we downsample visually static categories such as weather, which may not contain much change throughout the video. The weights of each category $^2$  is described in Figure 2.",
+        "bbox": [
+            112,
+            587,
+            487,
+            731
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Then, to reduce the computational complexity of video processing, we reduce a video into a set of keyframes that seek to capture the video's meaning (§3.1). To accommodate the limitations of existing models, we also generate two forms of visually-descriptive, textual representations—unstructured dense captions and FAMOuS scene descriptions (§3.2). Figure 3 summarizes this automated pipeline, which reduces the cost ordinarily spent to manually construct such a dataset.",
+        "bbox": [
+            112,
+            733,
+            489,
+            892
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/526900e08b3f683b32dae2628e1cc1bf32c9787f3f73e375da87eda7853477ff.jpg",
+        "image_caption": [
+            "Figure 2: Distribution of VIP's real-world video domains, weighted to emphasize videos containing significant visual change."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            529,
+            95,
+            875,
+            449
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Representative Keyframe Selection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            536,
+            830,
+            552
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Selecting Video Frames. The bottleneck to models in the video modality is the computational intensity. We select video frames that best capture the overall video content to mitigate this issue. Instead of training a model to choose a dynamic number of keyframes – which would be computationally expensive – we design an algorithm to prune semantically similar keyframes (Algorithm 1). However, selecting keyframes in this manner comes with the tradeoff that too many frames can introduce redundancy while too few can remove critical context. We choose a large set of candidate keyframes to balance these considerations, which we then dynamically prune. We employ an off-the-shelf keyframe extractor instead of learning a model ourselves. We choose to use  $\\mathrm{KATNA}^3$  as the baseline keyframe extraction tool as it is open-sourced and easy to onboard. KATNA selects keyframes by leveraging the differences in LUV colorspace, brightness, contrast, blur, and k-means clustering of images.",
+        "bbox": [
+            505,
+            563,
+            885,
+            885
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "2Videos may be in multiple categories.",
+        "bbox": [
+            134,
+            903,
+            373,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://github.com/keplerlab/katna",
+        "bbox": [
+            529,
+            903,
+            749,
+            919
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "206",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/5c2ea6303b1e49648f8eaf107ea6ff22d6249a16256d22bd10a5b29b7852b9e9.jpg",
+        "image_caption": [
+            "Figure 3: Overview of the pipeline to generate the scene descriptions provided in the VIP Dataset. We first process a video and extract the important frames (§3.1), then generate scene descriptions by extracting visual information from each keyframe, along with grounding information from the video to offset model hallucinations. We then feed in the extracted information into GPT-4 to generate the dense captions and structured scene descriptions. (§3.2)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            122,
+            84,
+            873,
+            275
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "code",
+        "sub_type": "algorithm",
+        "code_caption": [
+            "Algorithm 1: frame.extract(v, c, f)"
+        ],
+        "code_body": "Data: video  $v$  ,intsc,  $f$  of the candidate and finalized keyframe counts,respectively. Result: List of  $f$  finalized keyframes from  $v$  Extract initial frames and embeddings: 1  $k_{1},\\ldots ,k_{n}\\gets Katna(v,c)$  2  $t_1,\\dots ,t_n\\gets CLIP(Detic(k_1,\\dots ,k_n))$  3  $i_1,\\dots ,i_n\\gets CLIP(k_1,\\dots ,k_n)$  4 while len  $(k) > f$  do Remove frame with highest adjacent similarity: 5 for  $j$  in range(len  $(k)$  )do 6 costext  $\\leftarrow$  cos  $(t_j,t_{j - 1},t_{j + 1})$  7 cosimage  $\\leftarrow$  cos  $(i_j,i_{j - 1},i_{j + 1})$  8 scores[j]  $\\leftarrow$  mean(cosText,cos image) 9 remove k[s] where s = argmaxs(scores[s])   \n10 return  $k_{1},\\ldots ,k_{f}$",
+        "bbox": [
+            115,
+            391,
+            485,
+            583
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 4: frame.extract returns a list of  $f$  selected keyframes from a video  $v$ . First, we extract  $c$  candidates using Katna. These keyframes are embedded using CLIP in the image space; Detic extracts objects from the keyframes into a textual representation, which are also embedded with CLIP. Then, we iteratively prune the keyframe with the highest cosine similarity with adjacent frames until  $f$  keyframes remain.",
+        "bbox": [
+            112,
+            598,
+            485,
+            713
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Pruning Redundant Frames. Once baseline candidate keyframes are selected, we prune them by removing low-quality, semantically similar frames. First, we remove blurry keyframes with low Laplacian scores, which indicate the absence of intensity changes. Then, we use object detection models DETIC (Zhou et al., 2022) and GRiT (Wu et al., 2022) to filter keyframes that contain minimal objects, which indicates blurriness as these models are quite sensitive to all background objects.",
+        "bbox": [
+            112,
+            740,
+            485,
+            901
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "After removing low-quality frames, we use CLIP",
+        "bbox": [
+            132,
+            903,
+            485,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "(Radford et al., 2021) to create embeddings for the keyframe image and its list of detected objects and positions from the previous step in the pipeline. This combination helps us compare frames using pixel similarity and object invariance. We take the average cosine similarity score for the keyframe's image and object embeddings compared to the surrounding keyframes and prune the frames with the highest similarity. As people tend to be the primary subject of these videos, we add an additional check only to prune keyframes containing people if either of the surrounding frames also includes people.",
+        "bbox": [
+            507,
+            370,
+            884,
+            565
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "3.2 Textual Representations of Keyframes",
+        "text_level": 1,
+        "bbox": [
+            507,
+            579,
+            853,
+            594
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Next, to complement the keyframe images, we construct two textual representations of scenes: an unstructured, dense caption that provides visually descriptive insight into the scene; and second, a FAMOuS scene description that offers a structured approach to the reasoning process. We first generate the dense caption, which we then use to extract specific information for the structured scene description. These frame descriptions allow for leveraging existing LLM/VLM capabilities for video reasoning and generation.",
+        "bbox": [
+            505,
+            601,
+            884,
+            778
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Unstructured, Dense Captions. To create visually descriptive frame descriptions, we first extract three things from each keyframe: a caption, an object list, and a dense caption list. Together, these outputs paint a visual description of the keyframe – the object list and the dense captions describe the focus, objects, and setting, while the caption details the focus, action, and mood. We specifically use",
+        "bbox": [
+            507,
+            790,
+            884,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "207",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/07487577e4575311c44296a95a2a6ce48e8b3940afdc326c5d2d1a009c73ed70.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Frame</td><td>Structured</td><td>Domain</td><td>Vid. Len.</td><td>Cap. Len.</td><td>Test Samples</td></tr><tr><td>MSR-VTT (Xu et al., 2016)</td><td>X</td><td>X</td><td>Open</td><td>20.7s</td><td>9.6</td><td>3K</td></tr><tr><td>YouCook2 (Zhou et al., 2018)</td><td>X</td><td>X</td><td>Cooking</td><td>5.26m</td><td>8.8</td><td>2K</td></tr><tr><td>ActyNet-Cap (Krishna et al., 2017)</td><td>X</td><td>X</td><td>Open</td><td>2m</td><td>13.5</td><td>5K</td></tr><tr><td>HowTo100M (Miech et al., 2019)</td><td>X</td><td>X</td><td>Instructional</td><td>18s</td><td>4</td><td>24K</td></tr><tr><td>VATEX (Wang et al., 2019)</td><td>X</td><td>X</td><td>Open</td><td>10s</td><td>15.2</td><td>6K</td></tr><tr><td>VideoStory (Li et al., 2020)</td><td>✓</td><td>X</td><td>Events</td><td>12.6m</td><td>12.1</td><td>16</td></tr><tr><td>WebVid-2M (Bain et al., 2021)</td><td>X</td><td>X</td><td>Open</td><td>18s</td><td>12</td><td>5K</td></tr><tr><td>VIP</td><td>✓</td><td>✓</td><td>Open</td><td>3.6m</td><td>114.2</td><td>1.5K</td></tr></table>",
+        "bbox": [
+            121,
+            80,
+            877,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Statistics of video datasets: whether they include frame-level (Frame) and structured descriptions (Structured) as well as the average video token length (Vid. Len.), average caption token length (Cap. Len.), and number of inference samples (Test Samples). VIP is the only video reasoning dataset that has an open video domain and frame-level descriptions and provides novel structured frame descriptions.",
+        "bbox": [
+            112,
+            222,
+            884,
+            280
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "DETIC (Zhou et al., 2022), a tool that accurately detects objects without much detail, to simply list the frame's objects. To extract more descriptive, object-level detail for high-quality scene descriptions, we use GRiT (Wu et al., 2022), which returns dense captions describing each object. Finally, we obtain the keyframe's overall caption using LLAVA (Liu et al., 2023).",
+        "bbox": [
+            112,
+            305,
+            487,
+            432
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Because these individual models are prone to hallucination (Dai et al., 2023), we ensure the accuracy of our unstructured descriptions by engineering DETIC and GRIT to return confidence scores for each of their outputs. Additionally, we utilize the Wiki descriptions of each video topic in the YouTube-8M dataset to extract a grounding list of baseline objects using GPT-4. Finally, we feed in all of the extracted outputs into GPT-4 to generate the final dense caption.",
+        "bbox": [
+            112,
+            435,
+            489,
+            595
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "FAMOUS Structured Scene Descriptions. Structure can improve the reasoning ability of a model by providing concrete targets. To provide structure, we take inspiration from scene plays which clearly label and describe the scene. Specifically, we identify and extract the focus, action, mood, objects, and setting from the dense caption using GPT-4, categories which should capture the most important visual information in a concise, structured manner.",
+        "bbox": [
+            112,
+            608,
+            489,
+            769
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3.3 Dataset Contributions",
+        "text_level": 1,
+        "bbox": [
+            112,
+            784,
+            334,
+            797
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The VIP dataset is the first to evaluate multi-hop video reasoning via a video-chain of thought. This novel paradigm promotes efficient and robust video reasoning through automated keyframe extraction (Algorithm 1) over a breadth of domains (Figure 2). Our two textual representations of keyframes (Figure 1) add significantly granularity to videos (with",
+        "bbox": [
+            112,
+            806,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "an average caption length of 114 tokens) compared to traditional video caption datasets (Table 1). This enables reasoning on more specific visual and semantic changes which occur between frames, more closely mimicking how humans process videos by thinking frame by frame.",
+        "bbox": [
+            507,
+            305,
+            884,
+            400
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To ensure the quality of our collected dataset, we verify correctness via crowdsourcing on Amazon Mechanical Turk (Appendix A). Workers are paid to evaluate the quality of structured scene descriptions and edit those of low quality. Unstructured dense captions are corrected using the validated structured scene descriptions with GPT-4 and verified with another round of human evaluation.",
+        "bbox": [
+            507,
+            401,
+            885,
+            529
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Video Reasoning Tasks",
+        "text_level": 1,
+        "bbox": [
+            507,
+            541,
+            742,
+            558
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Taking inspiration from existing natural language tasks, we propose two tasks for videos that explore a model's multi-frame reasoning capabilities. The Video Infilling task requires models to predict a set of keyframes given the preceding and following frames, akin to masked language modeling for keyframes. Video Prediction tasks models to predict the most likely sequence of frames to follow a given series of frames - parallel to the text completion task. Video infilling and prediction of keyframes are two general tasks with several downstream contexts that can benefit from video understanding and completion.",
+        "bbox": [
+            507,
+            565,
+            884,
+            774
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To concretely define the tasks below, we represent the sequence of chronological keyframes as  $k_{1}, \\ldots, k_{n}$ , their respective unstructured dense captions as  $u_{1}, \\ldots, u_{n}$ , and FAMoS structured scene descriptions as  $s_{1}, \\ldots, s_{n}$ .",
+        "bbox": [
+            507,
+            776,
+            884,
+            857
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Video Infilling Task",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            710,
+            881
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Suppose a subsequence of frames  $k_{i},\\ldots ,k_{j}$  is masked. In the video infilling task, the target is for",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "208",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/3e12ea018b11af31140901a571e2b2d4e2ad3aea58fd72607ba6d3adcc507534.jpg",
+        "image_caption": [
+            "Figure 5: Given a number of context frames, the frame prediction task requires models to predict the following  $n$  frames. In this example, we provide two FAMOUS scene descriptions and use VICUNA and GPT-4 to predict the next three frames. Results emphasized in red differ from the ground truth."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            117,
+            92,
+            880,
+            381
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "a model to learn to reconstruct these masked frames using preceding context frames  $k_{i - n},\\ldots ,k_{i - 1}$  and following context frames  $k_{j + 1},\\ldots ,k_{j + n}$  where  $n$  is the number of frames provided as context. Without loss of generality, this task follows for both textual representations using  $u$  and  $s$  as inputs and outputs instead of  $k$ . In the multimodal setting, we can use pairs  $(k,u)$  or  $(k,s)$  as inputs and outputs.",
+        "bbox": [
+            112,
+            466,
+            489,
+            595
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "This task requires models to capture a scene's temporal variations and transitions, including changes in visual elements, object positions, and contextual factors. Furthermore, the task's difficulty scales two-fold. First, decreasing the context window  $n$  will reduce the ability to leverage hints from surrounding keyframes to infill informatively; combined with the necessity to perform multi-hop reasoning between each pair of frames in sequence, insufficient context could result in training divergence. Second, increasing the number of frames to predict in sequence between  $i,j$  also raises similar challenges as too large a gap could add several degrees of freedom, resulting in significant infilling variability. Successfully predicting intermediate keyframes may illuminate models' abilities to reason through the dynamic evolution of scenes and identify critical deltas in videos.",
+        "bbox": [
+            112,
+            629,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/81708e56551518d95cffe39bb5fd57a8e5aac42d03a7907ba93bc72f33d5f458.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metric</td><td>Mean ± STD</td></tr><tr><td>ROUGE_L</td><td>17.75 ± 0.43</td></tr><tr><td>BERTSCORE</td><td>18.97 ± 0.47</td></tr><tr><td>SENTENCEBERT</td><td>53.50 ± 0.50</td></tr></table>",
+        "bbox": [
+            574,
+            462,
+            815,
+            527
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Average GPT-4 performance across three prompts (Figure 6) on Infilling-1 task reported as mean  $\\pm$  one standard deviation. The low standard deviation indicates prompt stability.",
+        "bbox": [
+            507,
+            537,
+            884,
+            595
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.2 Video Prediction Task",
+        "text_level": 1,
+        "bbox": [
+            507,
+            620,
+            727,
+            634
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Suppose we are given a sequence of context frames  $k_{i - n},\\ldots ,k_{i}$  . In the video prediction task, we aim to predict the  $f$  following frames  $k_{i + 1},\\dots ,k_{f}$  Without loss of generality, this task follows for the unimodal text and multimodal representations. Much like the infilling task, the difficulty increases by decreasing the context window or increasing the prediction span. Since the prediction task only provides past context, predicting a longer sequence following may be harder as the possibilities increase exponentially.",
+        "bbox": [
+            507,
+            640,
+            884,
+            818
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            829,
+            653,
+            845
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Setup. Although it would be ideal to benchmark multi-modal language models on our proposed tasks, the current pre-trained models (e.g., Open Flamingo and Otter (Awadalla et al., 2023; Li et al.,",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "209",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/c59be81bdb56ec727263a3c043a012a4e52148f61cdc57f82f9ffd2b23d3fe38.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">Metric</td><td rowspan=\"2\">Model</td><td colspan=\"2\">Infilling-1</td><td colspan=\"2\">Infilling-2</td><td colspan=\"2\">Prediction-1</td><td colspan=\"2\">Prediction-2</td></tr><tr><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td></tr><tr><td rowspan=\"3\">ROUGE\\( _L \\)</td><td>GPT-4</td><td>17.34</td><td>24.92</td><td>18.44</td><td>25.25</td><td>15.52</td><td>23.31</td><td>16.66</td><td>25.22</td></tr><tr><td>GPT-3</td><td>17.83</td><td>26.26</td><td>19.34</td><td>25.50</td><td>16.39</td><td>28.43</td><td>17.20</td><td>25.96</td></tr><tr><td>VICUNA</td><td>17.37</td><td>25.34</td><td>18.85</td><td>26.69</td><td>15.86</td><td>23.75</td><td>16.59</td><td>25.88</td></tr><tr><td rowspan=\"3\">BERTSCORE</td><td>GPT-4</td><td>18.61</td><td>24.81</td><td>19.66</td><td>25.67</td><td>16.24</td><td>20.57</td><td>17.24</td><td>24.79</td></tr><tr><td>GPT-3</td><td>18.20</td><td>26.24</td><td>19.56</td><td>23.10</td><td>16.60</td><td>29.96</td><td>17.24</td><td>23.47</td></tr><tr><td>VICUNA</td><td>17.67</td><td>24.07</td><td>18.98</td><td>28.14</td><td>15.80</td><td>19.76</td><td>16.68</td><td>24.34</td></tr><tr><td rowspan=\"3\">SENTENCEBERT</td><td>GPT-4</td><td>53.05</td><td>58.53</td><td>53.87</td><td>58.22</td><td>50.57</td><td>53.55</td><td>51.54</td><td>57.06</td></tr><tr><td>GPT-3</td><td>52.95</td><td>58.54</td><td>54.57</td><td>55.69</td><td>51.04</td><td>59.83</td><td>51.81</td><td>53.99</td></tr><tr><td>VICUNA</td><td>52.19</td><td>54.66</td><td>53.33</td><td>58.80</td><td>50.40</td><td>51.86</td><td>50.96</td><td>54.86</td></tr></table>",
+        "bbox": [
+            122,
+            80,
+            877,
+            250
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 3: Model performance on infilling and prediction tasks when outputting three frames, with the best results underlined.. We vary the number of context frames as indicated by the dash in task name (e.g., Infilling-2 uses two previous and two future context frames to predict three intermediate masked keyframes; Prediction-2, uses two preceding keyframes to predict the subsequent three keyframes). We evaluate models using both FAMOUS structured scene descriptions and unstructured dense captions. Models show weak performance overall.",
+        "bbox": [
+            112,
+            258,
+            884,
+            332
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "2023)) are not designed to accommodate multiple image inputs off-the-shelf. As a result, we chose to benchmark the video infilling and prediction tasks as a language task, generating keyframes as represented by dense captions or FAMOUS descriptions. We use GPT-3, GPT-4, and VICUNA $^{4}$  as leading models, with in-context inference using one demonstration in both our infilling and prediction tasks. To mitigate hallucination, we leverage greedy decoding. In each task, the goal is to infill or predict three intermediate or subsequent keyframes, respectively. Evaluation metrics are computed as the mean of these three generated keyframes compared to the ground truth. Results are reported using one prompt, but a follow-up analysis shows prompt stability through low-variation among other prompts (Table 2).",
+        "bbox": [
+            112,
+            356,
+            489,
+            630
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Metrics. We use three standard text comparison metrics: ROUGE $_L$ , BERTSCORE (Zhang* et al., 2020), and SENTENCEBERT (Reimers and Gurevych, 2019). ROUGE $_L$  is best suited for tasks aimed to generate text that exactly matches the ground truth. BERTSCORE leverages BERT embeddings, which utilize the surrounding context. SENTENCEBERT is similar to BERTSCORE but computes the similarity of texts using sentence-level embeddings instead of word embeddings. These metrics combined provide initial scope into keyframe generation performance from both the semantic and contextual perspective.",
+        "bbox": [
+            112,
+            661,
+            489,
+            870
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.1 Primary Results",
+        "text_level": 1,
+        "bbox": [
+            507,
+            356,
+            685,
+            372
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We break down the primary results (Table 3) into four key points.",
+        "bbox": [
+            507,
+            376,
+            882,
+            409
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Number of Context Frames. Although the output size is fixed, we investigate how varying the input size affects the complexity of the VIP tasks. Consistent with intuition, we observe higher performance given additional context. However, the performance boost with each additional keyframe is marginal. With scores for all three metrics significantly lower than other tasks of similar spirit, it appears that our multi-hop, multi-frame prediction task is quite challenging using only textual representations for existing state-of-the-art language models. This low baseline performance may overshadow the change in difficulty as a result of varying context frames.",
+        "bbox": [
+            507,
+            418,
+            884,
+            643
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Dense Captions vs FAMOUS Descriptions. Dense captions consistently show stronger performance using our selected metrics than FAMOUS descriptions. As our evaluation metrics emphasize word similarity, they may favor dense captions which contain filler words used to form complete sentences. In the FAMOUS structure, descriptions are broken down by category, which reduces the vocabulary, thereby increasing the difficulty for word comparison metrics.",
+        "bbox": [
+            507,
+            652,
+            884,
+            813
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Infilling vs Prediction Tasks. We consistently observe that models have stronger performance on the infilling task compared to the prediction task. To most fairly compare the two tasks, compare the Infilling-1 task, which aims to predict three intermediate frames given one predecessor and one",
+        "bbox": [
+            507,
+            822,
+            885,
+            919
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "4We use the pre-trained VICUNA-13B checkpoint.",
+        "bbox": [
+            134,
+            903,
+            440,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "210",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/5dddc383975e5f435353692b89672aa13a2c02e74244fe9222797843c63b6ec3.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metric</td><td>Model</td><td>FAMOuS</td><td>Dense</td><td>Focus</td><td>Action</td><td>Mood</td><td>Objects</td><td>Setting</td></tr><tr><td rowspan=\"3\">ROUGE_L</td><td>GPT-4</td><td>17.14</td><td>26.05</td><td>13.12</td><td>9.64</td><td>19.56</td><td>28.82</td><td>14.52</td></tr><tr><td>GPT-3</td><td>17.44</td><td>29.35</td><td>14.96</td><td>10.47</td><td>19.09</td><td>27.85</td><td>14.82</td></tr><tr><td>VICUNA</td><td>17.37</td><td>26.99</td><td>15.31</td><td>10.59</td><td>19.53</td><td>28.14</td><td>13.30</td></tr><tr><td rowspan=\"3\">BERTSCORE</td><td>GPT-4</td><td>17.63</td><td>26.48</td><td>20.70</td><td>13.26</td><td>18.49</td><td>19.86</td><td>15.81</td></tr><tr><td>GPT-3</td><td>17.43</td><td>30.95</td><td>22.25</td><td>14.08</td><td>16.54</td><td>19.00</td><td>15.26</td></tr><tr><td>VICUNA</td><td>17.21</td><td>26.91</td><td>21.77</td><td>13.57</td><td>18.80</td><td>18.12</td><td>13.79</td></tr><tr><td rowspan=\"3\">SENTENCEBERT</td><td>GPT-4</td><td>51.87</td><td>57.89</td><td>46.83</td><td>44.48</td><td>51.34</td><td>58.84</td><td>57.88</td></tr><tr><td>GPT-3</td><td>52.15</td><td>56.40</td><td>47.72</td><td>45.71</td><td>51.43</td><td>58.35</td><td>57.51</td></tr><tr><td>VICUNA</td><td>51.57</td><td>55.44</td><td>47.65</td><td>45.36</td><td>51.44</td><td>57.33</td><td>56.12</td></tr></table>",
+        "bbox": [
+            168,
+            80,
+            828,
+            230
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Model performance on the Prediction-3 task; the two leftmost numerical columns report the aggregate results, while the five rightmost columns report the individual FAMOUS component results. The best results on the component level is underlined for each metric. Language models perform worse predicting Action and Setting with textual representation inputs, highlighting where video representations could be most beneficial.",
+        "bbox": [
+            112,
+            240,
+            882,
+            299
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "successor keyframe, with the Prediction-2 task, which aims to predict the three keyframes following the two context frames. Aside from a few nonsignificant outliers, these models perform better across all metrics and both textual representations. This is inline with intuition as bidirectional context reduces the complexity of the problem.",
+        "bbox": [
+            112,
+            323,
+            489,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Individual Model Performance. Across models, the performance does not follow any obvious trends, which is surprising considering the size difference in the open-source VICUNA compared to the human-reinforced GPT-3 and GPT-4 models. By metrics, we observe GPT-3 and VICUNA performs slightly better performances on  $\\mathrm{ROGUE}_L$  compared to GPT-4, suggesting exact word consistency may be better in these earlier models. GPT-4's and GPT-3's edge in SENTENCE-BERT suggest their generations may align more semantically than an off-the-shelf VICUNA.",
+        "bbox": [
+            112,
+            448,
+            489,
+            642
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.2 FAMOUS Component Analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            655,
+            410,
+            671
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We decompose model performance on FAMOUS structured scene descriptions on a component level (Table 4) for the Prediction-3 task, which aims to predict the three keyframes following the three input keyframes. Through ROUGE $_L$ , we observe models perform significantly better identifying objects, compared to the other four components. Comparing BERTSCORE, models appear to semantically align with the ground truth on the focus component better and more poorly in understanding of the keyframe's action. Finally, the SENTENCEBERT results suggest that models better maintain overall sentence similarity when considering components of the image's environment, such as mood, objects, and setting. These trends",
+        "bbox": [
+            112,
+            677,
+            489,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "highlight that reasoning through textual representations for basic components such as keyframe focus and objects is a strength of language models, while reasoning about more dynamic components such as the action necessitates a more intricate understanding of the keyframes and could benefit from a video representation.",
+        "bbox": [
+            507,
+            323,
+            884,
+            437
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.3 Causal Aspect Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            447,
+            741,
+            463
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We examine the difference in performance between physical and social causal reasoning (Table 5). A task necessitates physical causal reasoning when the video changes stem from external, tangible forces, like a wave crashing on a beach. Conversely, a task involves social causal reasoning when video changes result from social cues, such as a character becoming joyful during a surprise party. Observation of the results show that social causal reasoning tasks scored higher on BERTSCORE while physical causal reasoning tasks scored higher on SENTENCEBERT. These results may be an outgrowth of the FAMOuS Component Analysis §5.2, where a consistent character focus and objects present in many social scenarios yield higher token-level similarity with BERTSCORE. By contrast, the consistent environmental qualities like action or mood- present in many physical scenarios- result in a greater SENTENCEBERT score.",
+        "bbox": [
+            507,
+            468,
+            884,
+            774
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Domain Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            690,
+            800
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We also outline the overall results from all experiments corresponding to the different visual domains of our videos in Table 6. Although we found several categories to be noisy due to low sample sizes, certain categories like Games perform well, while others like Jobs & Education fall behind. We hypothesize that the availability of domain",
+        "bbox": [
+            507,
+            806,
+            884,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "211",
+        "bbox": [
+            485,
+            928,
+            512,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/8fb0d6d1fc922bca768c825884c002ba7489e5e838fee5c946bbc2e8096e1778.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Metric</td><td>Social</td><td>Physical</td></tr><tr><td>ROUGE\\( _L \\)</td><td>18.68</td><td>17.70</td></tr><tr><td>BERTSCORE</td><td>20.81</td><td>16.61</td></tr><tr><td>SENTENCEBERT</td><td>52.70</td><td>55.44</td></tr></table>",
+        "bbox": [
+            166,
+            80,
+            431,
+            146
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Table 5: GPT-4 performance for videos partitioned by the physical (changes due to outside, real forces) and social (changes related to social cues) on the Infilling-1 task with structured, FAMOUS descriptions. Significant differences between partitions are underlined.",
+        "bbox": [
+            112,
+            155,
+            487,
+            242
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "specific training data as well as intrinsic dimensionality needed to model interactions within these topics jointly contribute to such observations.",
+        "bbox": [
+            112,
+            268,
+            489,
+            317
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "5.5 Qualitative Observations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            326,
+            356,
+            343
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Figure 5 provides a visual depiction of the outputs generated by GPT-4 and VICUNA for the prediction task. Inspecting the depicted outputs from both models, it's evident that they lack some semantic congruence with the ground truth, underscoring the limitations that language model-based approaches face in in video reasoning. Figure 7 and Figure 5 further demonstrate impressive early performance using video-chain of thought, though the examples' strikingly similar output suggests some overfitting to training data. Despite the observable limitations, it's clear that the language models have a clear baseline video understanding. Still, both the quantitative and qualitative axes highlight that only using unimodal language doesn't generalize their strong language task performance to VIP's video reasoning tasks, which naturally opens several areas for subsequent research threads.",
+        "bbox": [
+            115,
+            349,
+            489,
+            638
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "6 Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            651,
+            263,
+            667
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we aim to lay the groundwork for exploring the challenging topic of multi-frame, multihop reasoning within the existing capabilities of deep learning models. Naturally, this opens several directions for exciting future work.",
+        "bbox": [
+            112,
+            677,
+            487,
+            757
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In the language model space, we benchmark the performance of several leading models using textual representations of keyframes for video-related reasoning following a standard in-context inference procedure with few-shot demonstrations. This invites the opportunity to discover more targeted inference-time techniques using language models or vision-language models to improve the performance of video reasoning tasks beyond a general paradigm. Similarly, additional training-time ef",
+        "bbox": [
+            112,
+            758,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "fort could be worthwhile through fine-tuning or a more traditional train-validate-test paradigm to learn skills beyond the general pre-trained learnset. In this vein, collecting additional data samples could improve the feasibility of these research threads.",
+        "bbox": [
+            507,
+            84,
+            884,
+            179
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Beyond the language modality, bridging the video reasoning task end-to-end with video is a longer-term research direction with immediate benefits in animation. Our paper reduces the video reasoning task into a language task with a textual output. Image synthesis would be an immediate step to reconstruct the keyframe image. Then, video synthesis from a set of images would naturally follow. Finally, unifying these disjoint tasks could benefit from error reduction and improved usability.",
+        "bbox": [
+            507,
+            184,
+            884,
+            344
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "As video reasoning is a new space, developing robust evaluation metrics would be a valuable contribution. Some desirable but difficult properties to consider in this area include the ability to capture both the spatial and temporal invariance that could occur through videos, as multiple interchangeable actions are plausible within different areas of the frame and sequences.",
+        "bbox": [
+            507,
+            348,
+            882,
+            475
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Finally, our general video reasoning tasks pose the prospect of efficient transfer learning where improving on such a task could benefit several new applications, similar to the contemporary boom of language technologies.",
+        "bbox": [
+            507,
+            479,
+            882,
+            558
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            509,
+            580,
+            640,
+            596
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "We present the inference-time Video Infilling and Prediction dataset to evaluate models' video reasoning abilities by performing a video chain-of-thought on video keyframes. To collect this dataset, we introduce a novel pipeline to systematically extract keyframes and generate corresponding textual representations – unstructured dense captions and structured FAMOUS scene descriptions. We benchmark state-of-the-art language models on VIP tasks and evaluate their ability to generate these textual representations of keyframes through a video chain-of-thought. These models display potential to perform video-related reasoning yet have significant potential to improve. By testing multi-hop, multi-frame reasoning abilities on sparse keyframes, we hope to promote research into developing models that support real-world video understanding and video generation while being resource efficient.",
+        "bbox": [
+            507,
+            613,
+            882,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "212",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Model Selection for Benchmarking. The primary limitation of our work is that current multimodal models do not support multiple image inputs per input query as a trivial use-case. As a result, despite our paper proposing a novel dataset intended for video-related reasoning, we currently only benchmark large language models. We encourage future research to explore the reasoning capabilities for video scene description generation.",
+        "bbox": [
+            112,
+            109,
+            492,
+            256
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Automation Process. While our work aims to systematically generate samples to evaluate the video-related reasoning capabilities of existing AI systems, we acknowledge the potential for error when using other AI systems to generate such examples. As a result, we add a layer of human supervision where crowd workers are used to first to classify whether generated scene descriptions are sufficiently correct. Then, we must make use of expert annotators to manually correct the generated scene descriptions that were flagged as poor quality for quality assurances purposes of this dataset.",
+        "bbox": [
+            112,
+            262,
+            490,
+            457
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            466,
+            267,
+            482
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Potential for Bias. We acknowledge the potential for bias in the data collection process and inference task design. We have taken a number of steps to mitigate bias, including ensuring diversity in video content selection, and regularly reviewing and refining the annotation process to minimize any unconscious bias. In addition, we are committed to addressing and correcting any biases that may arise during the evaluation and analysis of VIP model performance. Our dataset is restricted to English captions, thereby containing a bias toward culturally Western scenes. In a multilingual setting differential behaviors between language classes would probably be observed (Saxon and Wang, 2023).",
+        "bbox": [
+            112,
+            492,
+            489,
+            718
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Crowdsourcing. Crowdsourcing via the Amazon Mechanical Turk platform was utilized for conducting human evaluation. To ensure reliable results, we restricted participation to workers based in Australia, Canada, New Zealand, the United Kingdom, or the United States, with a minimum HIT approval rating of  $98\\%$ . Compensation for scene description checking and correction was set at a rate of $15 per hour. Our data collection is classified as an approved exempt protocol from the Institutional Review Board. Details of the interface can be found in Appendix A.",
+        "bbox": [
+            112,
+            726,
+            492,
+            919
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Acknowledgements",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            684,
+            99
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "We thank our reviewers for their supportive feedback. This material is based upon work supported in part by the National Science Foundation under Grants #1821415 and #2048122 REU. The authors are solely responsible for the contents of the paper, and the opinions expressed in this publication do not necessarily reflect the official policy or position of associated funding agencies or past or present employers of the authors. The contents of this paper are not intended to provide, and should not be relied upon for, investment advice.",
+        "bbox": [
+            507,
+            108,
+            885,
+            287
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            312,
+            610,
+            326
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. 2016. Youtube8m: A large-scale video classification benchmark. 3",
+            "Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736. 2",
+            "Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. 2023. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390. 6",
+            "Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. 2021. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision. 5",
+            "Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. 2023. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. 3",
+            "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. 2"
+        ],
+        "bbox": [
+            509,
+            334,
+            885,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "213",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. 2",
+            "Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. 2023. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2136-2148, Dubrovnik, Croatia. Association for Computational Linguistics. 5",
+            "Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. 2023. Palm-e: An embodied multimodal language model. 2",
+            "Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. 2021. VIOLET: End-to-End Video-Language Transformers with Masked Visual-token Modeling. In arXiv:2111.1268.1",
+            "Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. 2023a. An Empirical Study of End-to-End Video-Language Transformers with Masked Visual Modeling. In Conference on Computer Vision and Pattern Recognition (CVPR). 1",
+            "Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. 2023b. Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR). 3",
+            "Sergio Guadarrama, Niveda Krishnamoorthy, Girish Malkarnenkar, Subhashini Venugopalan, Raymond Mooney, Trevor Darrell, and Kate Saenko. 2013. Youtube2text: Recognizing and describing arbitrary activities using semantic hierarchies and zero-shot recognition. In Proceedings of the 2013 IEEE International Conference on Computer Vision, ICCV '13, page 2712-2719, USA. IEEE Computer Society. 2, 3",
+            "Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. 2017. Tgif-qa: Toward spatiotemporal reasoning in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 3",
+            "Kyung-Min Kim, Seong-Ho Choi, Jin-Hwa Kim, and Byoung-Tak Zhang. 2018. Multimodal dual attention memory for video story question answering. In Proceedings of the European Conference on Computer Vision (ECCV). 3"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. 2017. Deepstory: Video story qa by deep embedded memory networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, IJCAI'17, page 2016-2022. AAAI Press. 2, 3",
+            "Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. 2",
+            "Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. 2017. Dense-captioning events in videos. In International Conference on Computer Vision (ICCV). 5",
+            "Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. 2018. TVQA: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1369-1379, Brussels, Belgium. Association for Computational Linguistics. 1, 2",
+            "Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. 2020a. TVQA+: Spatio-temporal grounding for video question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8211-8225, Online. Association for Computational Linguistics. 2",
+            "Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. 2020b. What is more likely to happen next? video-and-language future event prediction. arXiv preprint arXiv:2010.07999. 3",
+            "Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. 2023. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726. 6",
+            "Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In ICML. 2",
+            "Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S. Kankanhalli. 2020. Video storytelling: Textual summaries for events. IEEE Transactions on Multimedia, 22(2):554-565. 5",
+            "Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. 2, 5",
+            "T. Maharaj, N. Ballas, A. Rohrbach, A. Courville, and C. Pal. 2017. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368, Los Alamitos, CA, USA. IEEE Computer Society. 2",
+            "Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. 2019. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV. 2, 3, 5"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "214",
+        "bbox": [
+            485,
+            928,
+            515,
+            939
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. 2017. Marioqa: Answering questions by watching gameplay videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 2867-2875. 2, 3",
+            "Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. 4",
+            "Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. 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 3982-3992, Hong Kong, China. Association for Computational Linguistics. 7",
+            "Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. 2023. Visual chain of thought: Bridging logical gaps with multimodal infillings. 2",
+            "Michael Saxon and William Yang Wang. 2023. Multilingual conceptual coverage in text-to-image models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4831-4848, Toronto, Canada. Association for Computational Linguistics. 10",
+            "Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. 2015. Weakly supervised memory networks. ArXiv, abs/1503.08895. 3",
+            "Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. 2016. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2",
+            "Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022a. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052. 2",
+            "Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. 2019. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 5",
+            "Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, Shih-Fu Chang, Mohit Bansal, and Heng Ji. 2022b. Language models with image descriptors are strong few-shot"
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "video-language learners. In Advances in Neural Information Processing Systems. 3",
+            "Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems. 2",
+            "Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. 2022. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280. 4,5",
+            "Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. Msr-vtt: A large video description dataset for bridging video and language. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). 1, 2, 5",
+            "Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. 2020. CLEVRER: collision events for video representation and reasoning. In ICLR. 3",
+            "Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. 2019. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127-9134. 2",
+            "Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. 2021. Merlot: Multimodal neural script knowledge models. In Advances in Neural Information Processing Systems, volume 34, pages 23634-23651. Curran Associates, Inc. 1, 3",
+            "Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. 2017. Leveraging video descriptions to learn video question answering. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1). 2",
+            "Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. *Bertscore: Evaluating text generation with bert.* In International Conference on Learning Representations. 7",
+            "Luowei Zhou, Chenliang Xu, and Jason J. Corso. 2018. Towards automatic learning of procedures from web instructional videos. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press. 5",
+            "Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. 2022. Detecting twenty-thousand classes using image-level supervision. In ECCV. 4, 5"
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "215",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Wang, Miguel Eckstein, and William Yang Wang. 2023. Visualize before you write: Imagination-guided open-ended text generation. In Findings of the Association for Computational Linguistics: EACL 2023, pages 78-92, Dubrovnik, Croatia. Association for Computational Linguistics. 2",
+        "bbox": [
+            115,
+            85,
+            489,
+            178
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A Appendix",
+        "text_level": 1,
+        "bbox": [
+            114,
+            190,
+            238,
+            206
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "A.1 Human Evaluation Interface",
+        "text_level": 1,
+        "bbox": [
+            114,
+            216,
+            389,
+            231
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "We demonstrate the interface of our human evaluation in scene description checking in Figure 9. We employ manual procedures to guarantee the exclusion of personal information and the absence of offensive content during human evaluations.",
+        "bbox": [
+            112,
+            237,
+            489,
+            317
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "216",
+        "bbox": [
+            485,
+            928,
+            515,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Prompt 1 (Used for experimentation): You are given the scene descriptions of 2 video keyframes (frames 1 and 5). Let's think frame by frame: identify the scene descriptions of the 3 intermediate keyframes that occur between the 2 given scene descriptions (frames 2, 3, and 4). Copy the structure and length exactly.",
+        "bbox": [
+            149,
+            148,
+            853,
+            219
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Prompt 2: Your job is to interpolate the scene descriptions for the three intermediate keyframes (frames 2, 3, and 4) using the scene descriptions of the first and fifth keyframes. Ensure that your descriptions match the provided structure and length.",
+        "bbox": [
+            149,
+            233,
+            826,
+            285
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Prompt 3: Examine the scene depictions of two video keyframes labeled Scene Description 1 (Frame 1) and Scene Description 5 (Frame 5). Let's break down the progression chronologically: deduce the scene depictions for the three intermediate keyframes (Frames 2, 3, and 4) situated between the two provided scenes. Maintain the identical structure and length precisely.",
+        "bbox": [
+            149,
+            300,
+            818,
+            370
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Figure 6: Input prompt templates used to test for prompt variation (Table 2). To reduce complexity, all subsequent results are reported using Prompt 1, which was designed to align with the widely recognized chain-of-thought prompting approach, adapted for video: \"Let's think frame by frame.\"",
+        "bbox": [
+            112,
+            413,
+            882,
+            456
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "You are given the scene descriptions of 2 video keyframes (frames 1 and 5). Let's think frame by frame: identify the scene descriptions of the 3 intermediate keyframes that occur between the 2 given scene descriptions (frames 2, 3, and 4). Copy the structure and length exactly.",
+        "bbox": [
+            139,
+            561,
+            803,
+            612
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 1: Focus: Man on fishing boat; Action: Man standing and looking at river; Mood: Focused, calm, enjoyable; Objects: Man, boat, fish food, river, trees, sky; Setting: River in the middle of forest",
+        "bbox": [
+            141,
+            613,
+            831,
+            661
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 5: Focus: Man releasing fish into river; Action: Man pushing fish back into water; Mood: Kind, happy, excited; Objects: Man, fish, river, trees, boat, rod; Setting: River surrounded by trees\"",
+        "bbox": [
+            141,
+            662,
+            825,
+            712
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 2: Focus: Man on boat with fishing rod; Action: Man fishing with rod; Mood: Focused, calm, enjoyable; Objects: Man, fishing rod, boat, river, trees; Setting: River in the forest Scene Description 3: Focus: Fish in the river; Action: Fish swimming; Mood: Peaceful, calm, still; Objects: Fish, river; Setting: River",
+        "bbox": [
+            139,
+            728,
+            806,
+            795
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 4: Focus: Man holding fish on boat; Action: Man taking fish hook out of fish; Mood: Proud, focused; Objects: Man, fish, boat, rod, river; Setting: River",
+        "bbox": [
+            141,
+            796,
+            796,
+            831
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "text",
+        "text": "Figure 7: Input prompt with a single in-context demonstration for the Infilling-1 task.",
+        "bbox": [
+            196,
+            858,
+            798,
+            873
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "217",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "table",
+        "img_path": "images/581a8bff779c73b00438b791ea25aefaedca16271b5f7f073045c87d18dae154.jpg",
+        "table_caption": [],
+        "table_footnote": [
+            "Table 6: Average model performance across the Infilling-1, Infilling-2, Prediction-1, Prediction-2, and Prediction-3 tasks, partitioned by domain. Categories such as Games, Sports, People & Society, Health, and Science perform better on average, while others such as Jobs & Education, Law & Government, and Food & Drink perform less well. The Best results for each metric and textual representation are underlined."
+        ],
+        "table_body": "<table><tr><td rowspan=\"2\">Visual Domain</td><td rowspan=\"2\">Model</td><td colspan=\"2\">ROUGEL</td><td colspan=\"2\">BERTScore</td><td colspan=\"2\">SentenceBERT</td></tr><tr><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td></tr><tr><td rowspan=\"2\">Jobs &amp; Education</td><td>GPT-4</td><td>0.1570</td><td>0.2444</td><td>0.1640</td><td>0.2337</td><td>0.4774</td><td>0.5101</td></tr><tr><td>VICUNA</td><td>0.1662</td><td>0.2570</td><td>0.1694</td><td>0.2599</td><td>0.4829</td><td>0.5391</td></tr><tr><td rowspan=\"2\">Games</td><td>GPT-4</td><td>0.1793</td><td>0.2519</td><td>0.2476</td><td>0.2619</td><td>0.5776</td><td>0.6794</td></tr><tr><td>VICUNA</td><td>0.1821</td><td>0.2625</td><td>0.2251</td><td>0.2892</td><td>0.5929</td><td>0.6344</td></tr><tr><td rowspan=\"2\">Sports</td><td>GPT-4</td><td>0.1686</td><td>0.2410</td><td>0.1940</td><td>0.2337</td><td>0.5440</td><td>0.5808</td></tr><tr><td>VICUNA</td><td>0.1768</td><td>0.2512</td><td>0.2000</td><td>0.2415</td><td>0.5507</td><td>0.5546</td></tr><tr><td rowspan=\"2\">Pets &amp; Animals</td><td>GPT-4</td><td>0.1644</td><td>0.2426</td><td>0.1828</td><td>0.2366</td><td>0.5173</td><td>0.5535</td></tr><tr><td>VICUNA</td><td>0.1642</td><td>0.2492</td><td>0.1735</td><td>0.2309</td><td>0.5127</td><td>0.5152</td></tr><tr><td rowspan=\"2\">Law &amp; Government</td><td>GPT-4</td><td>0.1484</td><td>0.2434</td><td>0.1685</td><td>0.2223</td><td>0.5091</td><td>0.5521</td></tr><tr><td>VICUNA</td><td>0.1501</td><td>0.2522</td><td>0.1628</td><td>0.2285</td><td>0.5016</td><td>0.5304</td></tr><tr><td rowspan=\"2\">Hobbies &amp; Leisure</td><td>GPT-4</td><td>0.1835</td><td>0.2590</td><td>0.1948</td><td>0.2660</td><td>0.5511</td><td>0.6209</td></tr><tr><td>VICUNA</td><td>0.1835</td><td>0.2640</td><td>0.1848</td><td>0.2617</td><td>0.5434</td><td>0.5967</td></tr><tr><td rowspan=\"2\">Home &amp; Garden</td><td>GPT-4</td><td>0.1656</td><td>0.2526</td><td>0.1754</td><td>0.2430</td><td>0.5060</td><td>0.5628</td></tr><tr><td>VICUNA</td><td>0.1601</td><td>0.2544</td><td>0.1634</td><td>0.2282</td><td>0.4955</td><td>0.5350</td></tr><tr><td rowspan=\"2\">Travel</td><td>GPT-4</td><td>0.1709</td><td>0.2484</td><td>0.1855</td><td>0.2429</td><td>0.5248</td><td>0.5539</td></tr><tr><td>VICUNA</td><td>0.1758</td><td>0.2481</td><td>0.1816</td><td>0.2272</td><td>0.5165</td><td>0.4978</td></tr><tr><td rowspan=\"2\">Food &amp; Drink</td><td>GPT-4</td><td>0.1647</td><td>0.2482</td><td>0.1476</td><td>0.2381</td><td>0.5171</td><td>0.5587</td></tr><tr><td>VICUNA</td><td>0.1651</td><td>0.2509</td><td>0.1397</td><td>0.2258</td><td>0.5048</td><td>0.5206</td></tr><tr><td rowspan=\"2\">Business &amp; Industrial</td><td>GPT-4</td><td>0.1656</td><td>0.2535</td><td>0.1716</td><td>0.2448</td><td>0.5183</td><td>0.5820</td></tr><tr><td>VICUNA</td><td>0.1648</td><td>0.2570</td><td>0.1654</td><td>0.2366</td><td>0.5134</td><td>0.5501</td></tr><tr><td rowspan=\"2\">Autos &amp; Vehicles</td><td>GPT-4</td><td>0.1667</td><td>0.2476</td><td>0.1795</td><td>0.2393</td><td>0.5396</td><td>0.5845</td></tr><tr><td>VICUNA</td><td>0.1668</td><td>0.2509</td><td>0.1795</td><td>0.2351</td><td>0.5311</td><td>0.5575</td></tr><tr><td rowspan=\"2\">People &amp; Society</td><td>GPT-4</td><td>0.1975</td><td>0.2517</td><td>0.2242</td><td>0.2668</td><td>0.5568</td><td>0.6402</td></tr><tr><td>VICUNA</td><td>0.2149</td><td>0.2638</td><td>0.2263</td><td>0.2803</td><td>0.5672</td><td>0.6304</td></tr><tr><td rowspan=\"2\">Reference</td><td>GPT-4</td><td>0.1704</td><td>0.2590</td><td>0.1944</td><td>0.2434</td><td>0.5147</td><td>0.5704</td></tr><tr><td>VICUNA</td><td>0.1691</td><td>0.2550</td><td>0.1800</td><td>0.2449</td><td>0.5072</td><td>0.5480</td></tr><tr><td rowspan=\"2\">Real Estate</td><td>GPT-4</td><td>0.1687</td><td>0.2590</td><td>0.1795</td><td>0.2440</td><td>0.5270</td><td>0.6144</td></tr><tr><td>VICUNA</td><td>0.1698</td><td>0.2647</td><td>0.1740</td><td>0.2459</td><td>0.5172</td><td>0.6144</td></tr><tr><td rowspan=\"2\">Health</td><td>GPT-4</td><td>0.2407</td><td>0.2442</td><td>0.2219</td><td>0.2262</td><td>0.5141</td><td>0.5528</td></tr><tr><td>VICUNA</td><td>0.2175</td><td>0.2551</td><td>0.1992</td><td>0.2260</td><td>0.4894</td><td>0.5090</td></tr><tr><td rowspan=\"2\">Science</td><td>GPT-4</td><td>0.1934</td><td>0.2436</td><td>0.2026</td><td>0.2362</td><td>0.5492</td><td>0.5923</td></tr><tr><td>VICUNA</td><td>0.1906</td><td>0.2448</td><td>0.1922</td><td>0.2210</td><td>0.5434</td><td>0.5482</td></tr></table>",
+        "bbox": [
+            114,
+            127,
+            880,
+            800
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "page_number",
+        "text": "218",
+        "bbox": [
+            485,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 14
+    },
+    {
+        "type": "text",
+        "text": "You are given the scene descriptions of 1 video keyframes (frames 1). Let's think frame by frame: predict the scene descriptions of the next 3 keyframes (frames 2, 3, and 4). Copy the structure and length exactly.",
+        "bbox": [
+            142,
+            143,
+            826,
+            193
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 1: Focus: Man on fishing boat; Action: Man standing and looking at river; Mood: Focused, calm, enjoyable; Objects: Man, boat, fish food, river, trees, sky; Setting: River in the middle of forest",
+        "bbox": [
+            142,
+            195,
+            840,
+            244
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 2: Focus: Man on boat with fishing rod; Action: Man fishing with rod; Mood: Focused, calm, enjoyable; Objects: Man, fishing rod, boat, river, trees; Setting: River in the forest Scene Description 3: Focus: Fish in the river; Action: Fish swimming; Mood: Peaceful, calm, still; Objects: Fish, river; Setting: River",
+        "bbox": [
+            142,
+            279,
+            815,
+            346
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Scene Description 4: Focus: Man holding fish on boat; Action: Man taking fish hook out of fish; Mood: Proud, focused; Objects: Man, fish, boat, rod, river; Setting: River",
+        "bbox": [
+            142,
+            347,
+            805,
+            382
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Instructions",
+        "text_level": 1,
+        "bbox": [
+            119,
+            521,
+            206,
+            533
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "You will be given a description of an image that is part of a video that describes the image's focus, action, mood, objects, and setting. Your job is to rate how well the scene description matches the given image.",
+        "bbox": [
+            119,
+            539,
+            868,
+            558
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "Focus: Reflects the center of activity, attraction, or attention in the image Action: Reflects the primary action occurring in the image",
+            "Mood: Reflects the feeling or atmosphere set in the image",
+            "Objects: Reflect all characters and entities that are present in the image",
+            "Setting: Reflects the place and circumstances at which the image takes place"
+        ],
+        "bbox": [
+            119,
+            564,
+            457,
+            606
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Task 1",
+        "text_level": 1,
+        "bbox": [
+            122,
+            615,
+            169,
+            627
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "image",
+        "img_path": "images/a9320a43bd0fc9013c67383c9ba422737ed95601d868d2d2ed4b69f8b52d243b.jpg",
+        "image_caption": [
+            "Figure 8: Input prompt with a single in-context demonstration for the Prediction-1 task.",
+            "Figure 9: Amazon Mechanical Turk Platform Interface, where crowdworkers are asked to qualitatively judge the accuracy of a FAMOuS scene description."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            218,
+            636,
+            423,
+            746
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Focus: Man playing golf",
+        "bbox": [
+            119,
+            753,
+            233,
+            764
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Action: Hitting a golf ball",
+        "bbox": [
+            122,
+            765,
+            233,
+            772
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Mood: Pleasant and relaxed",
+        "bbox": [
+            122,
+            772,
+            248,
+            781
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Objects: Golf ball, golf club, 14 people, tall grass",
+        "bbox": [
+            122,
+            782,
+            337,
+            791
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "Setting: Green field, golf course",
+        "bbox": [
+            122,
+            791,
+            265,
+            799
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "The scene description provided may have incorrect or missing information. Rate on a scale of 1-5 on how well the scene description matches the image.",
+        "bbox": [
+            119,
+            808,
+            815,
+            819
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "1 (very poor)",
+        "bbox": [
+            121,
+            824,
+            191,
+            835
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "$\\bigcirc 2$  (poor)",
+        "bbox": [
+            243,
+            824,
+            290,
+            835
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "$\\bigcirc 3$  (neutral)",
+        "bbox": [
+            364,
+            824,
+            421,
+            835
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "$\\bigcirc 4$  (good)",
+        "bbox": [
+            485,
+            824,
+            534,
+            835
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "text",
+        "text": "$\\bigcirc 5$  (very good)",
+        "bbox": [
+            606,
+            824,
+            675,
+            835
+        ],
+        "page_idx": 15
+    },
+    {
+        "type": "page_number",
+        "text": "219",
+        "bbox": [
+            485,
+            927,
+            515,
+            940
+        ],
+        "page_idx": 15
+    }
+]

2023/Let’s Think Frame by Frame with VIP_ A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought/4f07856a-09e6-4f3e-ad73-e9ecc85309e4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7b719027f588dcf72139a6d59ce2f9872a7384fa912c744656eb8d2069bf570f
+size 1959079

2023/Let’s Think Frame by Frame with VIP_ A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought/full.md

@@ -0,0 +1,341 @@
+# Let's Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought
+
+Vaishnavi Himakunthala*, Andy Ouyang*, Daniel Rose*, Ryan He*, Alex Mei, Yujie Lu, Chinmay Sonar, Michael Saxon, William Yang Wang
+
+University of California, Santa Barbara
+
+{vaishnavi, andyouyang, danielrose, ryanhe, yujielu, saxon}@ucsb.edu {alexmei, csonar, william}@cs.ucsb.edu
+
+# Abstract
+
+Despite exciting recent results showing vision-language systems' capacity to reason about images using natural language, their capacity for video reasoning remains underexplored. We motivate framing video reasoning as the sequential understanding of a small number of keyframes, thereby leveraging the power and robustness of vision-language while alleviating the computational complexities of processing videos. To evaluate this novel application, we introduce VIP<sup>1</sup>, an inference-time challenge dataset designed to explore models' reasoning capabilities through video chain-of-thought. Inspired by visually descriptive scene plays, we propose two formats for keyframe description: unstructured dense captions and structured scene descriptions that identify the focus, action, mood, objects, and setting (FAMOuS) of the keyframe. To evaluate video reasoning, we propose two tasks: Video Infilling and Video Prediction, which test abilities to generate multiple intermediate keyframes and predict future keyframes, respectively. We benchmark GPT-4, GPT-3, and VICUNA on VIP, demonstrate the performance gap in these complex video reasoning tasks, and encourage future work to prioritize language models for efficient and generalized video reasoning.
+
+# 1 Introduction
+
+Constituting  $65\%$  of all internet traffic in 2023, videos are an area of huge potential for the next chapter of leveraging artificial intelligence (Fu et al., 2021; Zellers et al., 2021; Fu et al., 2023a). For example, Video Question Answering (Lei et al., 2018) and Video Summarization (Xu et al., 2016) are two existing datasets that empirically evaluate video models. Yet, they do not assess more challenging tasks, such as reasoning through specific relationships between multiple frames. Just
+
+like how humans understand videos by processing frames across time steps, AI's ability to accomplish multi-frame reasoning is a core task for video understanding.
+
+Multi-frame video reasoning is bottlenecked by the sheer computing resources needed to process videos, which typically contain 24 frames per second and can vary widely. However, intelligible videos tend to have little variation from frame to frame. Akin to selecting principle components on the axes containing the highest orthogonal variance, picking a small sample of keyframes can capture much of the video's meaning. Understanding a video via extracted keyframes poses the challenge of multi-hop reasoning, which requires stronger language model capabilities.
+
+To evaluate this challenging multi-hop multiframe video reasoning, we elicit video chain-of-thought (VIDEOCOT). We propose two tasks to test such capabilities in existing models - Video Infilling and Video Prediction. In the Video Infilling setting, the goal is to predict the masked keyframes' descriptions when given the previous and next keyframes' descriptions in sequence as context, following the spirit of masked language modeling. In the Video Prediction, the goal is to predict the descriptions of the next keyframes in the sequence when given a set of previous keyframes' descriptions, similar to the next token prediction task. These tasks can help analyze video generation and whether video models truly understand the dynamic relations between subsequent video frames, given the variable context gap between keyframes.
+
+To benchmark our proposed Video Infilling and Prediction tasks, we construct a dataset by proposing an automated method to extract keyframes. In addition to the frame itself, we propose two textual representations – unstructured dense captions and FAMouS structured scene descriptions (Figure 1). The unstructured captions are intended to extract more significant and visually descriptive informa
+
+![](images/519217f7101b82444bf1b32abe8a5d8a8dd4ac083767d022567e1c117d760e80.jpg)  
+Figure 1: The Video Infilling and Prediction Dataset consists of two ways to describe keyframes: an unstructured dense caption and a structured scene description with five components: Focus, Action, Mood, Objects, and Setting (FAMOUS). The unstructured dense captions are highly detailed dense captions that can promote visually descriptive reasoning tasks, while structured scene description provide a concise, visual description of the keyframe that can aid in more focused reasoning tasks.
+
+tion compared to existing captioning systems, a necessary component to provide enough context for more challenging tasks. In addition, we systematically create structured scene descriptions from these unstructured dense captions by extracting the frame's focus, action, mood, objects, and setting using weak human supervision for quality assurance. The FAMOuS categories are inspired by play scripts, which maintain much of the visual detail of the unstructured captions while providing a clear, structured way to reason through visual tasks with high degrees of freedom.
+
+We propose the following contributions:
+
+- We systematically collect an inference-time challenge dataset of keyframes for video reasoning augmented with two textual representations: unstructured dense captions for visually-descriptive information and FAMOuS scene descriptions for structured reasoning.  
+- We propose the Video Infilling and Video Prediction tasks to benchmark the video chain-of-thought capabilities in existing models.  
+- We empirically demonstrate that existing models have the potential for multi-hop multi-frame video reasoning tasks but have a significant area for improvement as future work.
+
+# 2 Related Work
+
+AI Reasoning. Large language models (LLMs) demonstrate considerable gains on existing reasoning benchmarks with strategies such as chain-of-thought (CoT) (Wei et al., 2022) and few-shot demonstrations (Brown et al., 2020). Vision-language models (VLM) (Alayrac et al., 2022; Chowdhery et al., 2022; Driess et al., 2023) have furthered LLMs' capabilities by adding the visual modality to perform tasks such as visually-guided text generation (Rose et al., 2023; Zhu et al., 2023), vision question-answering (Wang et al., 2022a; Kim et al., 2021), and image captioning (Li et al., 2022; Liu et al., 2023). We believe the logical next step is to extend these existing models to the video domain. VIP's annotated collection of extracted keyframes from real-world videos offers a resource to evaluate reasoning abilities within the video domain.
+
+Datasets for Video Understanding. Existing video datasets are often limited by domain specificity or require a supplementary representation (e.g., audio, text) (Lei et al., 2018, 2020a; Tapaswi et al., 2016; Miech et al., 2019; Kim et al., 2017; Mun et al., 2017). These datasets also provide simplifications as textual summaries (Xu et al., 2016; Guadarrama et al., 2013) or a single video frame (Yu et al., 2019; Zeng et al., 2017; Maharaj et al.,
+
+2017), which by themselves can be sufficient to complete the task. While some datasets consider the multi-frame component (Jang et al., 2017; Yi et al., 2020; Mun et al., 2017; Fu et al., 2023b) for higher order complexity, VIP differs in trying to reduce the computational intensity of video reasoning without reducing the task difficulty and generality. VIP is a dataset of real-life videos that spans a breadth of domains and assesses multi-hop, multiframe video reasoning without requiring significant computation to train on videos.
+
+Textual Representations of Videos. Early video models are trained with visual keyframes and textual questions as input and return textual answers as output (Sukhbaatar et al., 2015; Kim et al., 2017; Jang et al., 2017). Then, researchers started to unify the video, keyframe, and text embedding spaces (Miech et al., 2019; Kim et al., 2018; Zellers et al., 2021; Kim et al., 2017; Guadarrama et al., 2013; Bhattacharya et al., 2023). VidIL leverages the contemporary in-context inference paradigm with few-shot demonstrations, including frames, captions, and visual tokens to prompt language models to solve VidL tasks (Lei et al., 2020b; Wang et al., 2022b). In contrast to these existing works, VIP introduces textual representations at the keyframe level and then leverages them to reason about specific video segments using VideoCOT.
+
+# 3 VIP Dataset Construction
+
+To construct the dataset, we first outsource the video corpus to stem from the YouTube-8m dataset (Abu-El-Haija et al., 2016), whose diverse, realistic videos with human-labeled topics align well with our desiderata. To effectively enable multi-frame video reasoning, we downsample visually static categories such as weather, which may not contain much change throughout the video. The weights of each category $^2$  is described in Figure 2.
+
+Then, to reduce the computational complexity of video processing, we reduce a video into a set of keyframes that seek to capture the video's meaning (§3.1). To accommodate the limitations of existing models, we also generate two forms of visually-descriptive, textual representations—unstructured dense captions and FAMOuS scene descriptions (§3.2). Figure 3 summarizes this automated pipeline, which reduces the cost ordinarily spent to manually construct such a dataset.
+
+![](images/526900e08b3f683b32dae2628e1cc1bf32c9787f3f73e375da87eda7853477ff.jpg)  
+Figure 2: Distribution of VIP's real-world video domains, weighted to emphasize videos containing significant visual change.
+
+# 3.1 Representative Keyframe Selection
+
+Selecting Video Frames. The bottleneck to models in the video modality is the computational intensity. We select video frames that best capture the overall video content to mitigate this issue. Instead of training a model to choose a dynamic number of keyframes – which would be computationally expensive – we design an algorithm to prune semantically similar keyframes (Algorithm 1). However, selecting keyframes in this manner comes with the tradeoff that too many frames can introduce redundancy while too few can remove critical context. We choose a large set of candidate keyframes to balance these considerations, which we then dynamically prune. We employ an off-the-shelf keyframe extractor instead of learning a model ourselves. We choose to use  $\mathrm{KATNA}^3$  as the baseline keyframe extraction tool as it is open-sourced and easy to onboard. KATNA selects keyframes by leveraging the differences in LUV colorspace, brightness, contrast, blur, and k-means clustering of images.
+
+![](images/5c2ea6303b1e49648f8eaf107ea6ff22d6249a16256d22bd10a5b29b7852b9e9.jpg)  
+Figure 3: Overview of the pipeline to generate the scene descriptions provided in the VIP Dataset. We first process a video and extract the important frames (§3.1), then generate scene descriptions by extracting visual information from each keyframe, along with grounding information from the video to offset model hallucinations. We then feed in the extracted information into GPT-4 to generate the dense captions and structured scene descriptions. (§3.2).
+
+Algorithm 1: frame.extract(v, c, f)  
+Data: video  $v$  ,intsc,  $f$  of the candidate and finalized keyframe counts,respectively. Result: List of  $f$  finalized keyframes from  $v$  Extract initial frames and embeddings: 1  $k_{1},\ldots ,k_{n}\gets Katna(v,c)$  2  $t_1,\dots ,t_n\gets CLIP(Detic(k_1,\dots ,k_n))$  3  $i_1,\dots ,i_n\gets CLIP(k_1,\dots ,k_n)$  4 while len  $(k) > f$  do Remove frame with highest adjacent similarity: 5 for  $j$  in range(len  $(k)$  )do 6 costext  $\leftarrow$  cos  $(t_j,t_{j - 1},t_{j + 1})$  7 cosimage  $\leftarrow$  cos  $(i_j,i_{j - 1},i_{j + 1})$  8 scores[j]  $\leftarrow$  mean(cosText,cos image) 9 remove k[s] where s = argmaxs(scores[s])   
+10 return  $k_{1},\ldots ,k_{f}$
+
+Figure 4: frame.extract returns a list of  $f$  selected keyframes from a video  $v$ . First, we extract  $c$  candidates using Katna. These keyframes are embedded using CLIP in the image space; Detic extracts objects from the keyframes into a textual representation, which are also embedded with CLIP. Then, we iteratively prune the keyframe with the highest cosine similarity with adjacent frames until  $f$  keyframes remain.
+
+Pruning Redundant Frames. Once baseline candidate keyframes are selected, we prune them by removing low-quality, semantically similar frames. First, we remove blurry keyframes with low Laplacian scores, which indicate the absence of intensity changes. Then, we use object detection models DETIC (Zhou et al., 2022) and GRiT (Wu et al., 2022) to filter keyframes that contain minimal objects, which indicates blurriness as these models are quite sensitive to all background objects.
+
+After removing low-quality frames, we use CLIP
+
+(Radford et al., 2021) to create embeddings for the keyframe image and its list of detected objects and positions from the previous step in the pipeline. This combination helps us compare frames using pixel similarity and object invariance. We take the average cosine similarity score for the keyframe's image and object embeddings compared to the surrounding keyframes and prune the frames with the highest similarity. As people tend to be the primary subject of these videos, we add an additional check only to prune keyframes containing people if either of the surrounding frames also includes people.
+
+# 3.2 Textual Representations of Keyframes
+
+Next, to complement the keyframe images, we construct two textual representations of scenes: an unstructured, dense caption that provides visually descriptive insight into the scene; and second, a FAMOuS scene description that offers a structured approach to the reasoning process. We first generate the dense caption, which we then use to extract specific information for the structured scene description. These frame descriptions allow for leveraging existing LLM/VLM capabilities for video reasoning and generation.
+
+Unstructured, Dense Captions. To create visually descriptive frame descriptions, we first extract three things from each keyframe: a caption, an object list, and a dense caption list. Together, these outputs paint a visual description of the keyframe – the object list and the dense captions describe the focus, objects, and setting, while the caption details the focus, action, and mood. We specifically use
+
+<table><tr><td>Dataset</td><td>Frame</td><td>Structured</td><td>Domain</td><td>Vid. Len.</td><td>Cap. Len.</td><td>Test Samples</td></tr><tr><td>MSR-VTT (Xu et al., 2016)</td><td>X</td><td>X</td><td>Open</td><td>20.7s</td><td>9.6</td><td>3K</td></tr><tr><td>YouCook2 (Zhou et al., 2018)</td><td>X</td><td>X</td><td>Cooking</td><td>5.26m</td><td>8.8</td><td>2K</td></tr><tr><td>ActyNet-Cap (Krishna et al., 2017)</td><td>X</td><td>X</td><td>Open</td><td>2m</td><td>13.5</td><td>5K</td></tr><tr><td>HowTo100M (Miech et al., 2019)</td><td>X</td><td>X</td><td>Instructional</td><td>18s</td><td>4</td><td>24K</td></tr><tr><td>VATEX (Wang et al., 2019)</td><td>X</td><td>X</td><td>Open</td><td>10s</td><td>15.2</td><td>6K</td></tr><tr><td>VideoStory (Li et al., 2020)</td><td>✓</td><td>X</td><td>Events</td><td>12.6m</td><td>12.1</td><td>16</td></tr><tr><td>WebVid-2M (Bain et al., 2021)</td><td>X</td><td>X</td><td>Open</td><td>18s</td><td>12</td><td>5K</td></tr><tr><td>VIP</td><td>✓</td><td>✓</td><td>Open</td><td>3.6m</td><td>114.2</td><td>1.5K</td></tr></table>
+
+Table 1: Statistics of video datasets: whether they include frame-level (Frame) and structured descriptions (Structured) as well as the average video token length (Vid. Len.), average caption token length (Cap. Len.), and number of inference samples (Test Samples). VIP is the only video reasoning dataset that has an open video domain and frame-level descriptions and provides novel structured frame descriptions.
+
+DETIC (Zhou et al., 2022), a tool that accurately detects objects without much detail, to simply list the frame's objects. To extract more descriptive, object-level detail for high-quality scene descriptions, we use GRiT (Wu et al., 2022), which returns dense captions describing each object. Finally, we obtain the keyframe's overall caption using LLAVA (Liu et al., 2023).
+
+Because these individual models are prone to hallucination (Dai et al., 2023), we ensure the accuracy of our unstructured descriptions by engineering DETIC and GRIT to return confidence scores for each of their outputs. Additionally, we utilize the Wiki descriptions of each video topic in the YouTube-8M dataset to extract a grounding list of baseline objects using GPT-4. Finally, we feed in all of the extracted outputs into GPT-4 to generate the final dense caption.
+
+FAMOUS Structured Scene Descriptions. Structure can improve the reasoning ability of a model by providing concrete targets. To provide structure, we take inspiration from scene plays which clearly label and describe the scene. Specifically, we identify and extract the focus, action, mood, objects, and setting from the dense caption using GPT-4, categories which should capture the most important visual information in a concise, structured manner.
+
+# 3.3 Dataset Contributions
+
+The VIP dataset is the first to evaluate multi-hop video reasoning via a video-chain of thought. This novel paradigm promotes efficient and robust video reasoning through automated keyframe extraction (Algorithm 1) over a breadth of domains (Figure 2). Our two textual representations of keyframes (Figure 1) add significantly granularity to videos (with
+
+an average caption length of 114 tokens) compared to traditional video caption datasets (Table 1). This enables reasoning on more specific visual and semantic changes which occur between frames, more closely mimicking how humans process videos by thinking frame by frame.
+
+To ensure the quality of our collected dataset, we verify correctness via crowdsourcing on Amazon Mechanical Turk (Appendix A). Workers are paid to evaluate the quality of structured scene descriptions and edit those of low quality. Unstructured dense captions are corrected using the validated structured scene descriptions with GPT-4 and verified with another round of human evaluation.
+
+# 4 Video Reasoning Tasks
+
+Taking inspiration from existing natural language tasks, we propose two tasks for videos that explore a model's multi-frame reasoning capabilities. The Video Infilling task requires models to predict a set of keyframes given the preceding and following frames, akin to masked language modeling for keyframes. Video Prediction tasks models to predict the most likely sequence of frames to follow a given series of frames - parallel to the text completion task. Video infilling and prediction of keyframes are two general tasks with several downstream contexts that can benefit from video understanding and completion.
+
+To concretely define the tasks below, we represent the sequence of chronological keyframes as  $k_{1}, \ldots, k_{n}$ , their respective unstructured dense captions as  $u_{1}, \ldots, u_{n}$ , and FAMoS structured scene descriptions as  $s_{1}, \ldots, s_{n}$ .
+
+# 4.1 Video Infilling Task
+
+Suppose a subsequence of frames  $k_{i},\ldots ,k_{j}$  is masked. In the video infilling task, the target is for
+
+![](images/3e12ea018b11af31140901a571e2b2d4e2ad3aea58fd72607ba6d3adcc507534.jpg)  
+Figure 5: Given a number of context frames, the frame prediction task requires models to predict the following  $n$  frames. In this example, we provide two FAMOUS scene descriptions and use VICUNA and GPT-4 to predict the next three frames. Results emphasized in red differ from the ground truth.
+
+a model to learn to reconstruct these masked frames using preceding context frames  $k_{i - n},\ldots ,k_{i - 1}$  and following context frames  $k_{j + 1},\ldots ,k_{j + n}$  where  $n$  is the number of frames provided as context. Without loss of generality, this task follows for both textual representations using  $u$  and  $s$  as inputs and outputs instead of  $k$ . In the multimodal setting, we can use pairs  $(k,u)$  or  $(k,s)$  as inputs and outputs.
+
+This task requires models to capture a scene's temporal variations and transitions, including changes in visual elements, object positions, and contextual factors. Furthermore, the task's difficulty scales two-fold. First, decreasing the context window  $n$  will reduce the ability to leverage hints from surrounding keyframes to infill informatively; combined with the necessity to perform multi-hop reasoning between each pair of frames in sequence, insufficient context could result in training divergence. Second, increasing the number of frames to predict in sequence between  $i,j$  also raises similar challenges as too large a gap could add several degrees of freedom, resulting in significant infilling variability. Successfully predicting intermediate keyframes may illuminate models' abilities to reason through the dynamic evolution of scenes and identify critical deltas in videos.
+
+<table><tr><td>Metric</td><td>Mean ± STD</td></tr><tr><td>ROUGE_L</td><td>17.75 ± 0.43</td></tr><tr><td>BERTSCORE</td><td>18.97 ± 0.47</td></tr><tr><td>SENTENCEBERT</td><td>53.50 ± 0.50</td></tr></table>
+
+Table 2: Average GPT-4 performance across three prompts (Figure 6) on Infilling-1 task reported as mean  $\pm$  one standard deviation. The low standard deviation indicates prompt stability.
+
+# 4.2 Video Prediction Task
+
+Suppose we are given a sequence of context frames  $k_{i - n},\ldots ,k_{i}$  . In the video prediction task, we aim to predict the  $f$  following frames  $k_{i + 1},\dots ,k_{f}$  Without loss of generality, this task follows for the unimodal text and multimodal representations. Much like the infilling task, the difficulty increases by decreasing the context window or increasing the prediction span. Since the prediction task only provides past context, predicting a longer sequence following may be harder as the possibilities increase exponentially.
+
+# 5 Experiments
+
+Setup. Although it would be ideal to benchmark multi-modal language models on our proposed tasks, the current pre-trained models (e.g., Open Flamingo and Otter (Awadalla et al., 2023; Li et al.,
+
+<table><tr><td rowspan="2">Metric</td><td rowspan="2">Model</td><td colspan="2">Infilling-1</td><td colspan="2">Infilling-2</td><td colspan="2">Prediction-1</td><td colspan="2">Prediction-2</td></tr><tr><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td></tr><tr><td rowspan="3">ROUGE\( _L \)</td><td>GPT-4</td><td>17.34</td><td>24.92</td><td>18.44</td><td>25.25</td><td>15.52</td><td>23.31</td><td>16.66</td><td>25.22</td></tr><tr><td>GPT-3</td><td>17.83</td><td>26.26</td><td>19.34</td><td>25.50</td><td>16.39</td><td>28.43</td><td>17.20</td><td>25.96</td></tr><tr><td>VICUNA</td><td>17.37</td><td>25.34</td><td>18.85</td><td>26.69</td><td>15.86</td><td>23.75</td><td>16.59</td><td>25.88</td></tr><tr><td rowspan="3">BERTSCORE</td><td>GPT-4</td><td>18.61</td><td>24.81</td><td>19.66</td><td>25.67</td><td>16.24</td><td>20.57</td><td>17.24</td><td>24.79</td></tr><tr><td>GPT-3</td><td>18.20</td><td>26.24</td><td>19.56</td><td>23.10</td><td>16.60</td><td>29.96</td><td>17.24</td><td>23.47</td></tr><tr><td>VICUNA</td><td>17.67</td><td>24.07</td><td>18.98</td><td>28.14</td><td>15.80</td><td>19.76</td><td>16.68</td><td>24.34</td></tr><tr><td rowspan="3">SENTENCEBERT</td><td>GPT-4</td><td>53.05</td><td>58.53</td><td>53.87</td><td>58.22</td><td>50.57</td><td>53.55</td><td>51.54</td><td>57.06</td></tr><tr><td>GPT-3</td><td>52.95</td><td>58.54</td><td>54.57</td><td>55.69</td><td>51.04</td><td>59.83</td><td>51.81</td><td>53.99</td></tr><tr><td>VICUNA</td><td>52.19</td><td>54.66</td><td>53.33</td><td>58.80</td><td>50.40</td><td>51.86</td><td>50.96</td><td>54.86</td></tr></table>
+
+Table 3: Model performance on infilling and prediction tasks when outputting three frames, with the best results underlined.. We vary the number of context frames as indicated by the dash in task name (e.g., Infilling-2 uses two previous and two future context frames to predict three intermediate masked keyframes; Prediction-2, uses two preceding keyframes to predict the subsequent three keyframes). We evaluate models using both FAMOUS structured scene descriptions and unstructured dense captions. Models show weak performance overall.
+
+2023)) are not designed to accommodate multiple image inputs off-the-shelf. As a result, we chose to benchmark the video infilling and prediction tasks as a language task, generating keyframes as represented by dense captions or FAMOUS descriptions. We use GPT-3, GPT-4, and VICUNA $^{4}$  as leading models, with in-context inference using one demonstration in both our infilling and prediction tasks. To mitigate hallucination, we leverage greedy decoding. In each task, the goal is to infill or predict three intermediate or subsequent keyframes, respectively. Evaluation metrics are computed as the mean of these three generated keyframes compared to the ground truth. Results are reported using one prompt, but a follow-up analysis shows prompt stability through low-variation among other prompts (Table 2).
+
+Metrics. We use three standard text comparison metrics: ROUGE $_L$ , BERTSCORE (Zhang* et al., 2020), and SENTENCEBERT (Reimers and Gurevych, 2019). ROUGE $_L$  is best suited for tasks aimed to generate text that exactly matches the ground truth. BERTSCORE leverages BERT embeddings, which utilize the surrounding context. SENTENCEBERT is similar to BERTSCORE but computes the similarity of texts using sentence-level embeddings instead of word embeddings. These metrics combined provide initial scope into keyframe generation performance from both the semantic and contextual perspective.
+
+# 5.1 Primary Results
+
+We break down the primary results (Table 3) into four key points.
+
+Number of Context Frames. Although the output size is fixed, we investigate how varying the input size affects the complexity of the VIP tasks. Consistent with intuition, we observe higher performance given additional context. However, the performance boost with each additional keyframe is marginal. With scores for all three metrics significantly lower than other tasks of similar spirit, it appears that our multi-hop, multi-frame prediction task is quite challenging using only textual representations for existing state-of-the-art language models. This low baseline performance may overshadow the change in difficulty as a result of varying context frames.
+
+Dense Captions vs FAMOUS Descriptions. Dense captions consistently show stronger performance using our selected metrics than FAMOUS descriptions. As our evaluation metrics emphasize word similarity, they may favor dense captions which contain filler words used to form complete sentences. In the FAMOUS structure, descriptions are broken down by category, which reduces the vocabulary, thereby increasing the difficulty for word comparison metrics.
+
+Infilling vs Prediction Tasks. We consistently observe that models have stronger performance on the infilling task compared to the prediction task. To most fairly compare the two tasks, compare the Infilling-1 task, which aims to predict three intermediate frames given one predecessor and one
+
+<table><tr><td>Metric</td><td>Model</td><td>FAMOuS</td><td>Dense</td><td>Focus</td><td>Action</td><td>Mood</td><td>Objects</td><td>Setting</td></tr><tr><td rowspan="3">ROUGE_L</td><td>GPT-4</td><td>17.14</td><td>26.05</td><td>13.12</td><td>9.64</td><td>19.56</td><td>28.82</td><td>14.52</td></tr><tr><td>GPT-3</td><td>17.44</td><td>29.35</td><td>14.96</td><td>10.47</td><td>19.09</td><td>27.85</td><td>14.82</td></tr><tr><td>VICUNA</td><td>17.37</td><td>26.99</td><td>15.31</td><td>10.59</td><td>19.53</td><td>28.14</td><td>13.30</td></tr><tr><td rowspan="3">BERTSCORE</td><td>GPT-4</td><td>17.63</td><td>26.48</td><td>20.70</td><td>13.26</td><td>18.49</td><td>19.86</td><td>15.81</td></tr><tr><td>GPT-3</td><td>17.43</td><td>30.95</td><td>22.25</td><td>14.08</td><td>16.54</td><td>19.00</td><td>15.26</td></tr><tr><td>VICUNA</td><td>17.21</td><td>26.91</td><td>21.77</td><td>13.57</td><td>18.80</td><td>18.12</td><td>13.79</td></tr><tr><td rowspan="3">SENTENCEBERT</td><td>GPT-4</td><td>51.87</td><td>57.89</td><td>46.83</td><td>44.48</td><td>51.34</td><td>58.84</td><td>57.88</td></tr><tr><td>GPT-3</td><td>52.15</td><td>56.40</td><td>47.72</td><td>45.71</td><td>51.43</td><td>58.35</td><td>57.51</td></tr><tr><td>VICUNA</td><td>51.57</td><td>55.44</td><td>47.65</td><td>45.36</td><td>51.44</td><td>57.33</td><td>56.12</td></tr></table>
+
+Table 4: Model performance on the Prediction-3 task; the two leftmost numerical columns report the aggregate results, while the five rightmost columns report the individual FAMOUS component results. The best results on the component level is underlined for each metric. Language models perform worse predicting Action and Setting with textual representation inputs, highlighting where video representations could be most beneficial.
+
+successor keyframe, with the Prediction-2 task, which aims to predict the three keyframes following the two context frames. Aside from a few nonsignificant outliers, these models perform better across all metrics and both textual representations. This is inline with intuition as bidirectional context reduces the complexity of the problem.
+
+Individual Model Performance. Across models, the performance does not follow any obvious trends, which is surprising considering the size difference in the open-source VICUNA compared to the human-reinforced GPT-3 and GPT-4 models. By metrics, we observe GPT-3 and VICUNA performs slightly better performances on  $\mathrm{ROGUE}_L$  compared to GPT-4, suggesting exact word consistency may be better in these earlier models. GPT-4's and GPT-3's edge in SENTENCE-BERT suggest their generations may align more semantically than an off-the-shelf VICUNA.
+
+# 5.2 FAMOUS Component Analysis
+
+We decompose model performance on FAMOUS structured scene descriptions on a component level (Table 4) for the Prediction-3 task, which aims to predict the three keyframes following the three input keyframes. Through ROUGE $_L$ , we observe models perform significantly better identifying objects, compared to the other four components. Comparing BERTSCORE, models appear to semantically align with the ground truth on the focus component better and more poorly in understanding of the keyframe's action. Finally, the SENTENCEBERT results suggest that models better maintain overall sentence similarity when considering components of the image's environment, such as mood, objects, and setting. These trends
+
+highlight that reasoning through textual representations for basic components such as keyframe focus and objects is a strength of language models, while reasoning about more dynamic components such as the action necessitates a more intricate understanding of the keyframes and could benefit from a video representation.
+
+# 5.3 Causal Aspect Analysis
+
+We examine the difference in performance between physical and social causal reasoning (Table 5). A task necessitates physical causal reasoning when the video changes stem from external, tangible forces, like a wave crashing on a beach. Conversely, a task involves social causal reasoning when video changes result from social cues, such as a character becoming joyful during a surprise party. Observation of the results show that social causal reasoning tasks scored higher on BERTSCORE while physical causal reasoning tasks scored higher on SENTENCEBERT. These results may be an outgrowth of the FAMOuS Component Analysis §5.2, where a consistent character focus and objects present in many social scenarios yield higher token-level similarity with BERTSCORE. By contrast, the consistent environmental qualities like action or mood- present in many physical scenarios- result in a greater SENTENCEBERT score.
+
+# 5.4 Domain Analysis
+
+We also outline the overall results from all experiments corresponding to the different visual domains of our videos in Table 6. Although we found several categories to be noisy due to low sample sizes, certain categories like Games perform well, while others like Jobs & Education fall behind. We hypothesize that the availability of domain
+
+<table><tr><td>Metric</td><td>Social</td><td>Physical</td></tr><tr><td>ROUGE\( _L \)</td><td>18.68</td><td>17.70</td></tr><tr><td>BERTSCORE</td><td>20.81</td><td>16.61</td></tr><tr><td>SENTENCEBERT</td><td>52.70</td><td>55.44</td></tr></table>
+
+Table 5: GPT-4 performance for videos partitioned by the physical (changes due to outside, real forces) and social (changes related to social cues) on the Infilling-1 task with structured, FAMOUS descriptions. Significant differences between partitions are underlined.
+
+specific training data as well as intrinsic dimensionality needed to model interactions within these topics jointly contribute to such observations.
+
+# 5.5 Qualitative Observations
+
+Figure 5 provides a visual depiction of the outputs generated by GPT-4 and VICUNA for the prediction task. Inspecting the depicted outputs from both models, it's evident that they lack some semantic congruence with the ground truth, underscoring the limitations that language model-based approaches face in in video reasoning. Figure 7 and Figure 5 further demonstrate impressive early performance using video-chain of thought, though the examples' strikingly similar output suggests some overfitting to training data. Despite the observable limitations, it's clear that the language models have a clear baseline video understanding. Still, both the quantitative and qualitative axes highlight that only using unimodal language doesn't generalize their strong language task performance to VIP's video reasoning tasks, which naturally opens several areas for subsequent research threads.
+
+# 6 Future Work
+
+In this paper, we aim to lay the groundwork for exploring the challenging topic of multi-frame, multihop reasoning within the existing capabilities of deep learning models. Naturally, this opens several directions for exciting future work.
+
+In the language model space, we benchmark the performance of several leading models using textual representations of keyframes for video-related reasoning following a standard in-context inference procedure with few-shot demonstrations. This invites the opportunity to discover more targeted inference-time techniques using language models or vision-language models to improve the performance of video reasoning tasks beyond a general paradigm. Similarly, additional training-time ef
+
+fort could be worthwhile through fine-tuning or a more traditional train-validate-test paradigm to learn skills beyond the general pre-trained learnset. In this vein, collecting additional data samples could improve the feasibility of these research threads.
+
+Beyond the language modality, bridging the video reasoning task end-to-end with video is a longer-term research direction with immediate benefits in animation. Our paper reduces the video reasoning task into a language task with a textual output. Image synthesis would be an immediate step to reconstruct the keyframe image. Then, video synthesis from a set of images would naturally follow. Finally, unifying these disjoint tasks could benefit from error reduction and improved usability.
+
+As video reasoning is a new space, developing robust evaluation metrics would be a valuable contribution. Some desirable but difficult properties to consider in this area include the ability to capture both the spatial and temporal invariance that could occur through videos, as multiple interchangeable actions are plausible within different areas of the frame and sequences.
+
+Finally, our general video reasoning tasks pose the prospect of efficient transfer learning where improving on such a task could benefit several new applications, similar to the contemporary boom of language technologies.
+
+# 7 Conclusion
+
+We present the inference-time Video Infilling and Prediction dataset to evaluate models' video reasoning abilities by performing a video chain-of-thought on video keyframes. To collect this dataset, we introduce a novel pipeline to systematically extract keyframes and generate corresponding textual representations – unstructured dense captions and structured FAMOUS scene descriptions. We benchmark state-of-the-art language models on VIP tasks and evaluate their ability to generate these textual representations of keyframes through a video chain-of-thought. These models display potential to perform video-related reasoning yet have significant potential to improve. By testing multi-hop, multi-frame reasoning abilities on sparse keyframes, we hope to promote research into developing models that support real-world video understanding and video generation while being resource efficient.
+
+# Limitations
+
+Model Selection for Benchmarking. The primary limitation of our work is that current multimodal models do not support multiple image inputs per input query as a trivial use-case. As a result, despite our paper proposing a novel dataset intended for video-related reasoning, we currently only benchmark large language models. We encourage future research to explore the reasoning capabilities for video scene description generation.
+
+Automation Process. While our work aims to systematically generate samples to evaluate the video-related reasoning capabilities of existing AI systems, we acknowledge the potential for error when using other AI systems to generate such examples. As a result, we add a layer of human supervision where crowd workers are used to first to classify whether generated scene descriptions are sufficiently correct. Then, we must make use of expert annotators to manually correct the generated scene descriptions that were flagged as poor quality for quality assurances purposes of this dataset.
+
+# Ethics Statement
+
+Potential for Bias. We acknowledge the potential for bias in the data collection process and inference task design. We have taken a number of steps to mitigate bias, including ensuring diversity in video content selection, and regularly reviewing and refining the annotation process to minimize any unconscious bias. In addition, we are committed to addressing and correcting any biases that may arise during the evaluation and analysis of VIP model performance. Our dataset is restricted to English captions, thereby containing a bias toward culturally Western scenes. In a multilingual setting differential behaviors between language classes would probably be observed (Saxon and Wang, 2023).
+
+Crowdsourcing. Crowdsourcing via the Amazon Mechanical Turk platform was utilized for conducting human evaluation. To ensure reliable results, we restricted participation to workers based in Australia, Canada, New Zealand, the United Kingdom, or the United States, with a minimum HIT approval rating of  $98\%$ . Compensation for scene description checking and correction was set at a rate of $15 per hour. Our data collection is classified as an approved exempt protocol from the Institutional Review Board. Details of the interface can be found in Appendix A.
+
+# Acknowledgements
+
+We thank our reviewers for their supportive feedback. This material is based upon work supported in part by the National Science Foundation under Grants #1821415 and #2048122 REU. The authors are solely responsible for the contents of the paper, and the opinions expressed in this publication do not necessarily reflect the official policy or position of associated funding agencies or past or present employers of the authors. The contents of this paper are not intended to provide, and should not be relied upon for, investment advice.
+
+# References
+
+Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. 2016. Youtube8m: A large-scale video classification benchmark. 3  
+Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. 2022. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736. 2  
+Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. 2023. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390. 6  
+Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. 2021. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision. 5  
+Aanisha Bhattacharya, Yaman K Singla, Balaji Krishnamurthy, Rajiv Ratn Shah, and Changyou Chen. 2023. A video is worth 4096 tokens: Verbalize story videos to understand them in zero shot. 3  
+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. 2
+
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2022. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311. 2  
+Wenliang Dai, Zihan Liu, Ziwei Ji, Dan Su, and Pascale Fung. 2023. Plausible may not be faithful: Probing object hallucination in vision-language pre-training. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2136-2148, Dubrovnik, Croatia. Association for Computational Linguistics. 5  
+Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. 2023. Palm-e: An embodied multimodal language model. 2  
+Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. 2021. VIOLET: End-to-End Video-Language Transformers with Masked Visual-token Modeling. In arXiv:2111.1268.1  
+Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. 2023a. An Empirical Study of End-to-End Video-Language Transformers with Masked Visual Modeling. In Conference on Computer Vision and Pattern Recognition (CVPR). 1  
+Tsu-Jui Fu, Licheng Yu, Ning Zhang, Cheng-Yang Fu, Jong-Chyi Su, William Yang Wang, and Sean Bell. 2023b. Tell Me What Happened: Unifying Text-guided Video Completion via Multimodal Masked Video Generation. In Conference on Computer Vision and Pattern Recognition (CVPR). 3  
+Sergio Guadarrama, Niveda Krishnamoorthy, Girish Malkarnenkar, Subhashini Venugopalan, Raymond Mooney, Trevor Darrell, and Kate Saenko. 2013. Youtube2text: Recognizing and describing arbitrary activities using semantic hierarchies and zero-shot recognition. In Proceedings of the 2013 IEEE International Conference on Computer Vision, ICCV '13, page 2712-2719, USA. IEEE Computer Society. 2, 3  
+Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. 2017. Tgif-qa: Toward spatiotemporal reasoning in visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 3  
+Kyung-Min Kim, Seong-Ho Choi, Jin-Hwa Kim, and Byoung-Tak Zhang. 2018. Multimodal dual attention memory for video story question answering. In Proceedings of the European Conference on Computer Vision (ECCV). 3
+
+Kyung-Min Kim, Min-Oh Heo, Seong-Ho Choi, and Byoung-Tak Zhang. 2017. Deepstory: Video story qa by deep embedded memory networks. In Proceedings of the 26th International Joint Conference on Artificial Intelligence, IJCAI'17, page 2016-2022. AAAI Press. 2, 3  
+Wonjae Kim, Bokyung Son, and Ildoo Kim. 2021. Vilt: Vision-and-language transformer without convolution or region supervision. 2  
+Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. 2017. Dense-captioning events in videos. In International Conference on Computer Vision (ICCV). 5  
+Jie Lei, Licheng Yu, Mohit Bansal, and Tamara Berg. 2018. TVQA: Localized, compositional video question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1369-1379, Brussels, Belgium. Association for Computational Linguistics. 1, 2  
+Jie Lei, Licheng Yu, Tamara Berg, and Mohit Bansal. 2020a. TVQA+: Spatio-temporal grounding for video question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8211-8225, Online. Association for Computational Linguistics. 2  
+Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. 2020b. What is more likely to happen next? video-and-language future event prediction. arXiv preprint arXiv:2010.07999. 3  
+Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. 2023. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726. 6  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In ICML. 2  
+Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S. Kankanhalli. 2020. Video storytelling: Textual summaries for events. IEEE Transactions on Multimedia, 22(2):554-565. 5  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. 2, 5  
+T. Maharaj, N. Ballas, A. Rohrbach, A. Courville, and C. Pal. 2017. A dataset and exploration of models for understanding video data through fill-in-the-blank question-answering. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 7359-7368, Los Alamitos, CA, USA. IEEE Computer Society. 2  
+Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. 2019. HowTo100M: Learning a Text-Video Embedding by Watching Hundred Million Narrated Video Clips. In ICCV. 2, 3, 5
+
+Jonghwan Mun, Paul Hongsuck Seo, Ilchae Jung, and Bohyung Han. 2017. Marioqa: Answering questions by watching gameplay videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 2867-2875. 2, 3  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning transferable visual models from natural language supervision. 4  
+Nils Reimers and Iryna Gurevych. 2019. SentenceBERT: Sentence embeddings using Siamese BERTnetworks. 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 3982-3992, Hong Kong, China. Association for Computational Linguistics. 7  
+Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. 2023. Visual chain of thought: Bridging logical gaps with multimodal infillings. 2  
+Michael Saxon and William Yang Wang. 2023. Multilingual conceptual coverage in text-to-image models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4831-4848, Toronto, Canada. Association for Computational Linguistics. 10  
+Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. 2015. Weakly supervised memory networks. ArXiv, abs/1503.08895. 3  
+Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. 2016. Movieqa: Understanding stories in movies through question-answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2  
+Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022a. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. CoRR, abs/2202.03052. 2  
+Xin Wang, Jiawei Wu, Junkun Chen, Lei Li, Yuan-Fang Wang, and William Yang Wang. 2019. Vatex: A large-scale, high-quality multilingual dataset for video-and-language research. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 5  
+Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, Shih-Fu Chang, Mohit Bansal, and Heng Ji. 2022b. Language models with image descriptors are strong few-shot
+
+video-language learners. In Advances in Neural Information Processing Systems. 3  
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian richter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. Chain of thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems. 2  
+Jialian Wu, Jianfeng Wang, Zhengyuan Yang, Zhe Gan, Zicheng Liu, Junsong Yuan, and Lijuan Wang. 2022. Grit: A generative region-to-text transformer for object understanding. arXiv preprint arXiv:2212.00280. 4,5  
+Jun Xu, Tao Mei, Ting Yao, and Yong Rui. 2016. Msr-vtt: A large video description dataset for bridging video and language. In IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE International Conference on Computer Vision and Pattern Recognition (CVPR). 1, 2, 5  
+Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. 2020. CLEVRER: collision events for video representation and reasoning. In ICLR. 3  
+Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. 2019. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127-9134. 2  
+Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. 2021. Merlot: Multimodal neural script knowledge models. In Advances in Neural Information Processing Systems, volume 34, pages 23634-23651. Curran Associates, Inc. 1, 3  
+Kuo-Hao Zeng, Tseng-Hung Chen, Ching-Yao Chuang, Yuan-Hong Liao, Juan Carlos Niebles, and Min Sun. 2017. Leveraging video descriptions to learn video question answering. Proceedings of the AAAI Conference on Artificial Intelligence, 31(1). 2  
+Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. *Bertscore: Evaluating text generation with bert.* In International Conference on Learning Representations. 7  
+Luowei Zhou, Chenliang Xu, and Jason J. Corso. 2018. Towards automatic learning of procedures from web instructional videos. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence, AAAI'18/IAAI'18/EAAI'18. AAAI Press. 5  
+Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. 2022. Detecting twenty-thousand classes using image-level supervision. In ECCV. 4, 5
+
+Wanrong Zhu, An Yan, Yujie Lu, Wenda Xu, Xin Wang, Miguel Eckstein, and William Yang Wang. 2023. Visualize before you write: Imagination-guided open-ended text generation. In Findings of the Association for Computational Linguistics: EACL 2023, pages 78-92, Dubrovnik, Croatia. Association for Computational Linguistics. 2
+
+# A Appendix
+
+# A.1 Human Evaluation Interface
+
+We demonstrate the interface of our human evaluation in scene description checking in Figure 9. We employ manual procedures to guarantee the exclusion of personal information and the absence of offensive content during human evaluations.
+
+Prompt 1 (Used for experimentation): You are given the scene descriptions of 2 video keyframes (frames 1 and 5). Let's think frame by frame: identify the scene descriptions of the 3 intermediate keyframes that occur between the 2 given scene descriptions (frames 2, 3, and 4). Copy the structure and length exactly.
+
+Prompt 2: Your job is to interpolate the scene descriptions for the three intermediate keyframes (frames 2, 3, and 4) using the scene descriptions of the first and fifth keyframes. Ensure that your descriptions match the provided structure and length.
+
+Prompt 3: Examine the scene depictions of two video keyframes labeled Scene Description 1 (Frame 1) and Scene Description 5 (Frame 5). Let's break down the progression chronologically: deduce the scene depictions for the three intermediate keyframes (Frames 2, 3, and 4) situated between the two provided scenes. Maintain the identical structure and length precisely.
+
+Figure 6: Input prompt templates used to test for prompt variation (Table 2). To reduce complexity, all subsequent results are reported using Prompt 1, which was designed to align with the widely recognized chain-of-thought prompting approach, adapted for video: "Let's think frame by frame."
+
+You are given the scene descriptions of 2 video keyframes (frames 1 and 5). Let's think frame by frame: identify the scene descriptions of the 3 intermediate keyframes that occur between the 2 given scene descriptions (frames 2, 3, and 4). Copy the structure and length exactly.
+
+Scene Description 1: Focus: Man on fishing boat; Action: Man standing and looking at river; Mood: Focused, calm, enjoyable; Objects: Man, boat, fish food, river, trees, sky; Setting: River in the middle of forest
+
+Scene Description 5: Focus: Man releasing fish into river; Action: Man pushing fish back into water; Mood: Kind, happy, excited; Objects: Man, fish, river, trees, boat, rod; Setting: River surrounded by trees"
+
+Scene Description 2: Focus: Man on boat with fishing rod; Action: Man fishing with rod; Mood: Focused, calm, enjoyable; Objects: Man, fishing rod, boat, river, trees; Setting: River in the forest Scene Description 3: Focus: Fish in the river; Action: Fish swimming; Mood: Peaceful, calm, still; Objects: Fish, river; Setting: River
+
+Scene Description 4: Focus: Man holding fish on boat; Action: Man taking fish hook out of fish; Mood: Proud, focused; Objects: Man, fish, boat, rod, river; Setting: River
+
+Figure 7: Input prompt with a single in-context demonstration for the Infilling-1 task.
+
+<table><tr><td rowspan="2">Visual Domain</td><td rowspan="2">Model</td><td colspan="2">ROUGEL</td><td colspan="2">BERTScore</td><td colspan="2">SentenceBERT</td></tr><tr><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td><td>FAMOuS</td><td>Dense</td></tr><tr><td rowspan="2">Jobs &amp; Education</td><td>GPT-4</td><td>0.1570</td><td>0.2444</td><td>0.1640</td><td>0.2337</td><td>0.4774</td><td>0.5101</td></tr><tr><td>VICUNA</td><td>0.1662</td><td>0.2570</td><td>0.1694</td><td>0.2599</td><td>0.4829</td><td>0.5391</td></tr><tr><td rowspan="2">Games</td><td>GPT-4</td><td>0.1793</td><td>0.2519</td><td>0.2476</td><td>0.2619</td><td>0.5776</td><td>0.6794</td></tr><tr><td>VICUNA</td><td>0.1821</td><td>0.2625</td><td>0.2251</td><td>0.2892</td><td>0.5929</td><td>0.6344</td></tr><tr><td rowspan="2">Sports</td><td>GPT-4</td><td>0.1686</td><td>0.2410</td><td>0.1940</td><td>0.2337</td><td>0.5440</td><td>0.5808</td></tr><tr><td>VICUNA</td><td>0.1768</td><td>0.2512</td><td>0.2000</td><td>0.2415</td><td>0.5507</td><td>0.5546</td></tr><tr><td rowspan="2">Pets &amp; Animals</td><td>GPT-4</td><td>0.1644</td><td>0.2426</td><td>0.1828</td><td>0.2366</td><td>0.5173</td><td>0.5535</td></tr><tr><td>VICUNA</td><td>0.1642</td><td>0.2492</td><td>0.1735</td><td>0.2309</td><td>0.5127</td><td>0.5152</td></tr><tr><td rowspan="2">Law &amp; Government</td><td>GPT-4</td><td>0.1484</td><td>0.2434</td><td>0.1685</td><td>0.2223</td><td>0.5091</td><td>0.5521</td></tr><tr><td>VICUNA</td><td>0.1501</td><td>0.2522</td><td>0.1628</td><td>0.2285</td><td>0.5016</td><td>0.5304</td></tr><tr><td rowspan="2">Hobbies &amp; Leisure</td><td>GPT-4</td><td>0.1835</td><td>0.2590</td><td>0.1948</td><td>0.2660</td><td>0.5511</td><td>0.6209</td></tr><tr><td>VICUNA</td><td>0.1835</td><td>0.2640</td><td>0.1848</td><td>0.2617</td><td>0.5434</td><td>0.5967</td></tr><tr><td rowspan="2">Home &amp; Garden</td><td>GPT-4</td><td>0.1656</td><td>0.2526</td><td>0.1754</td><td>0.2430</td><td>0.5060</td><td>0.5628</td></tr><tr><td>VICUNA</td><td>0.1601</td><td>0.2544</td><td>0.1634</td><td>0.2282</td><td>0.4955</td><td>0.5350</td></tr><tr><td rowspan="2">Travel</td><td>GPT-4</td><td>0.1709</td><td>0.2484</td><td>0.1855</td><td>0.2429</td><td>0.5248</td><td>0.5539</td></tr><tr><td>VICUNA</td><td>0.1758</td><td>0.2481</td><td>0.1816</td><td>0.2272</td><td>0.5165</td><td>0.4978</td></tr><tr><td rowspan="2">Food &amp; Drink</td><td>GPT-4</td><td>0.1647</td><td>0.2482</td><td>0.1476</td><td>0.2381</td><td>0.5171</td><td>0.5587</td></tr><tr><td>VICUNA</td><td>0.1651</td><td>0.2509</td><td>0.1397</td><td>0.2258</td><td>0.5048</td><td>0.5206</td></tr><tr><td rowspan="2">Business &amp; Industrial</td><td>GPT-4</td><td>0.1656</td><td>0.2535</td><td>0.1716</td><td>0.2448</td><td>0.5183</td><td>0.5820</td></tr><tr><td>VICUNA</td><td>0.1648</td><td>0.2570</td><td>0.1654</td><td>0.2366</td><td>0.5134</td><td>0.5501</td></tr><tr><td rowspan="2">Autos &amp; Vehicles</td><td>GPT-4</td><td>0.1667</td><td>0.2476</td><td>0.1795</td><td>0.2393</td><td>0.5396</td><td>0.5845</td></tr><tr><td>VICUNA</td><td>0.1668</td><td>0.2509</td><td>0.1795</td><td>0.2351</td><td>0.5311</td><td>0.5575</td></tr><tr><td rowspan="2">People &amp; Society</td><td>GPT-4</td><td>0.1975</td><td>0.2517</td><td>0.2242</td><td>0.2668</td><td>0.5568</td><td>0.6402</td></tr><tr><td>VICUNA</td><td>0.2149</td><td>0.2638</td><td>0.2263</td><td>0.2803</td><td>0.5672</td><td>0.6304</td></tr><tr><td rowspan="2">Reference</td><td>GPT-4</td><td>0.1704</td><td>0.2590</td><td>0.1944</td><td>0.2434</td><td>0.5147</td><td>0.5704</td></tr><tr><td>VICUNA</td><td>0.1691</td><td>0.2550</td><td>0.1800</td><td>0.2449</td><td>0.5072</td><td>0.5480</td></tr><tr><td rowspan="2">Real Estate</td><td>GPT-4</td><td>0.1687</td><td>0.2590</td><td>0.1795</td><td>0.2440</td><td>0.5270</td><td>0.6144</td></tr><tr><td>VICUNA</td><td>0.1698</td><td>0.2647</td><td>0.1740</td><td>0.2459</td><td>0.5172</td><td>0.6144</td></tr><tr><td rowspan="2">Health</td><td>GPT-4</td><td>0.2407</td><td>0.2442</td><td>0.2219</td><td>0.2262</td><td>0.5141</td><td>0.5528</td></tr><tr><td>VICUNA</td><td>0.2175</td><td>0.2551</td><td>0.1992</td><td>0.2260</td><td>0.4894</td><td>0.5090</td></tr><tr><td rowspan="2">Science</td><td>GPT-4</td><td>0.1934</td><td>0.2436</td><td>0.2026</td><td>0.2362</td><td>0.5492</td><td>0.5923</td></tr><tr><td>VICUNA</td><td>0.1906</td><td>0.2448</td><td>0.1922</td><td>0.2210</td><td>0.5434</td><td>0.5482</td></tr></table>
+
+Table 6: Average model performance across the Infilling-1, Infilling-2, Prediction-1, Prediction-2, and Prediction-3 tasks, partitioned by domain. Categories such as Games, Sports, People & Society, Health, and Science perform better on average, while others such as Jobs & Education, Law & Government, and Food & Drink perform less well. The Best results for each metric and textual representation are underlined.
+
+You are given the scene descriptions of 1 video keyframes (frames 1). Let's think frame by frame: predict the scene descriptions of the next 3 keyframes (frames 2, 3, and 4). Copy the structure and length exactly.
+
+Scene Description 1: Focus: Man on fishing boat; Action: Man standing and looking at river; Mood: Focused, calm, enjoyable; Objects: Man, boat, fish food, river, trees, sky; Setting: River in the middle of forest
+
+Scene Description 2: Focus: Man on boat with fishing rod; Action: Man fishing with rod; Mood: Focused, calm, enjoyable; Objects: Man, fishing rod, boat, river, trees; Setting: River in the forest Scene Description 3: Focus: Fish in the river; Action: Fish swimming; Mood: Peaceful, calm, still; Objects: Fish, river; Setting: River
+
+Scene Description 4: Focus: Man holding fish on boat; Action: Man taking fish hook out of fish; Mood: Proud, focused; Objects: Man, fish, boat, rod, river; Setting: River
+
+# Instructions
+
+You will be given a description of an image that is part of a video that describes the image's focus, action, mood, objects, and setting. Your job is to rate how well the scene description matches the given image.
+
+Focus: Reflects the center of activity, attraction, or attention in the image Action: Reflects the primary action occurring in the image  
+Mood: Reflects the feeling or atmosphere set in the image  
+Objects: Reflect all characters and entities that are present in the image  
+Setting: Reflects the place and circumstances at which the image takes place
+
+# Task 1
+
+![](images/a9320a43bd0fc9013c67383c9ba422737ed95601d868d2d2ed4b69f8b52d243b.jpg)  
+Figure 8: Input prompt with a single in-context demonstration for the Prediction-1 task.  
+Figure 9: Amazon Mechanical Turk Platform Interface, where crowdworkers are asked to qualitatively judge the accuracy of a FAMOuS scene description.
+
+Focus: Man playing golf
+
+Action: Hitting a golf ball
+
+Mood: Pleasant and relaxed
+
+Objects: Golf ball, golf club, 14 people, tall grass
+
+Setting: Green field, golf course
+
+The scene description provided may have incorrect or missing information. Rate on a scale of 1-5 on how well the scene description matches the image.
+
+1 (very poor)
+
+$\bigcirc 2$  (poor)
+
+$\bigcirc 3$  (neutral)
+
+$\bigcirc 4$  (good)
+
+$\bigcirc 5$  (very good)

2023/Let’s Think Frame by Frame with VIP_ A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e729dfd9f59bdd8c9a6459b2e587d109d5be649254902fa13e73220395b8e72
+size 899675

2023/Lifelong Sequence Generation with Dynamic Module Expansion and Adaptation/0474ffe8-7b94-4eda-aa63-2522d4a8cfa0_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5716c5cf7bdca85ef6cf6fea14d30d1363b8c7465a75198d6253074af8ca139c
+size 991241

2023/Lifelong Sequence Generation with Dynamic Module Expansion and Adaptation/full.md

@@ -0,0 +1,441 @@
+# Lifelong Sequence Generation with Dynamic Module Expansion and Adaptation
+
+Chengwei Qin\*, Chen Chen\*, Shafiq Joty\*
+
+\* Nanyang Technological University
+
+Salesforce AI
+
+{chengwei003@e.ntu, chen1436@e.ntu, srjoty@ntu}.edu.sg
+
+# Abstract
+
+Lifelong sequence generation (LSG), a problem in continual learning, aims to continually train a model on a sequence of generation tasks to learn constantly emerging new generation patterns while avoiding the forgetting of previous knowledge. Existing LSG methods mainly focus on maintaining old knowledge while paying little attention to knowledge transfer across tasks. In contrast, humans can better learn new tasks by leveraging previously acquired knowledge from similar tasks. Inspired by the learning paradigm of humans, we propose Dynamic Module Expansion and Adaptation (DMEA), which enables the model to dynamically determine the architecture for acquiring new knowledge based on task correlation and select the most similar previous tasks to facilitate adaptation to new tasks. In addition, as the learning process can easily be biased towards the current task which might cause more severe forgetting of previously learned knowledge, we propose dynamic gradient scaling to balance the learning of the current task and replayed tasks. With extensive experiments, we demonstrate that DMEA can consistently outperform existing methods in different LSG settings.
+
+# 1 Introduction
+
+With the recent advancements in pre-trained language models (LMs), current sequence generation methods have achieved impressive performance on a variety of generation tasks (Radford et al., 2019; Raffel et al., 2020). Typically, these models are trained on a fixed corpus, assuming the underlying data distribution to be static (Ham et al., 2020; El-Kassas et al., 2021). However, real cognitive tasks are generally more complex involving changing contexts and dynamic environments. The ever-changing data distribution causes the models to face challenges in acquiring new knowledge, while retaining the prior knowledge. Speaking about what is next for NLP, Kathleen McKeown
+
+in a recent interview said: "Most models are static. But the world changes every minute, every second. Dealing with a dynamic world is a new area that's up and coming." (Source)
+
+A potential solution is to formalize sequence generation as lifelong sequence generation or LSG (Sun et al., 2020), where the model is expected to learn sequentially from a stream of generation tasks with potentially different data distributions. In such cases of distribution shift, the model might forget previously acquired knowledge upon learning new tasks, a phenomenon known as catastrophic forgetting (McCloskey and Cohen, 1989). Previous LSG methods (Mi et al., 2020; Sun et al., 2020; Madotto et al., 2021) mainly explore different ways to alleviate forgetting. Recently, Zhang et al. (2022) propose Adaptive Compositional Modules (ACM) which dynamically adds modules for new tasks depending on whether there are reusable previous modules, achieving SOTA performance on LSG.
+
+Despite its effectiveness, ACM has several key limitations. First, it mainly focuses on mitigating forgetting of previously acquired knowledge while paying little attention to transferring learned knowledge to new tasks which is as important for continual learning as preventing forgetting (Ke et al., 2020). In fact, a hallmark of human intelligence is that humans can better learn new tasks by leveraging previously acquired knowledge from similar tasks (Lake et al., 2017). They can not only determine whether previously acquired skills are sufficient to solve a new task, but also exploit the most similar learned skills to facilitate the learning of the task; see Appendix A.1 for an illustration. Second, ACM does not consider the correlation between learned tasks and the new task when adding modules, which might hinder finding the optimal architecture (case study in Appendix A.9). Finally, the learning process in ACM can be biased towards the new task as the gradient norm of the new task on reused modules is typically much larger than
+
+that of replayed tasks, which may affect previously acquired knowledge; see Appendix A.2 for an explanation.
+
+Inspired by the learning paradigm of humans and to address the above limitations of ACM, in this work we propose Dynamic Module<sup>1</sup> Expansion and Adaptation (DMEA). We divide the learning process of a new task into three stages: expansion, selection and adaptation. In the expansion stage, DMEA determines whether to reuse modules of previous tasks or insert new modules for learning novel knowledge. Inspired by Zhang et al. (2022), it utilizes differentiable architecture search (Liu et al., 2019) to enable the model to dynamically determine the architecture for solving the new task. The learnable coefficients in architecture search are initialized based on the cosine similarity of word frequency distributions between learned tasks and the new task, aiming to discover the optimal model architecture. After searching, the module with the largest coefficient in every layer is chosen for the new task. In the selection stage, DMEA selects the top- $K$  most similar previous tasks through input subspace (Lin et al., 2022b). Finally, in the adaptation stage, it utilizes the selected similar tasks to facilitate adaptation to the new task. The output of selected similar tasks is fused with that of the new task using learnable coefficients in every transformer layer to enable forward knowledge transfer. This is indeed an instance of mixture-of-experts (Masoudnia and Ebrahimpour, 2014).
+
+In addition, when the model learns a new task, DMEA also incorporates pseudo-sample replay (Sun et al., 2020) to further mitigate catastrophic forgetting. To address the "bias to the new task" in the gradient update, we introduce dynamic gradient scaling to balance the learning of the new task and replayed tasks. To verify the effectiveness of DMEA, we conduct extensive experiments on various generation tasks in different LSG settings. The empirical results show that DMEA can consistently outperform previous state-of-the-art baselines.
+
+In summary, our main contributions are:
+
+- To the best of our knowledge, we are the first to explore solving LSG from the perspective of human learning. We propose DMEA, a novel method based on dynamic module expansion and adaptation, to alleviate catastrophic forgetting and facilitate knowledge transfer in LSG.
+
+- With extensive experiments and analysis, we demonstrate the effectiveness of our method compared to existing ones in different LSG settings.
+
+# 2 Related Work
+
+Lifelong Learning (LL) aims to continually learn knowledge from a sequence of tasks with different distributions. The goal is twofold: alleviate catastrophic forgetting (McCloskey and Cohen, 1989) of learned tasks, and facilitate knowledge transfer (Lopez-Paz and Ranzato, 2017) across tasks.
+
+Catastrophic forgetting typically means that the model forgets previously acquired knowledge after learning new tasks. Prior LL methods mainly focus on mitigating this problem and can be divided into three categories. First, regularization-based methods constrain the update of parameters that are important to learned tasks to retain previous knowledge (Kirkpatrick et al., 2017; Li and Hoiem, 2017; Zenke et al., 2017; Ritter et al., 2018). Second, architecture-based methods dynamically adjust the model architecture to acquire new information while preventing the forgetting of previously learned tasks (Rusu et al., 2016; Chen et al., 2016; Fernando et al., 2017; Madotto et al., 2021; Zhang et al., 2022). Finally, memory-based methods keep a number of key samples from previous tasks in memory to alleviate forgetting (Rebuffi et al., 2017; Shin et al., 2017; Chaudhry et al., 2019; Qin and Joty, 2022a). The memory data can be either real examples (Han et al., 2020) or generated by language models (Sun et al., 2020; Qin and Joty, 2022b).
+
+More recently, researchers have considered exploring knowledge transfer in LL, i.e., learning on a task can benefit from learning on another task by transferring related knowledge. This includes CTR (Ke et al., 2021) and CUBER (Lin et al., 2022a). Despite their effectiveness, these methods mainly focus on classification tasks, while generation tasks typically have more complex label space. Note that this line of research is different from transfer learning (Ruder et al., 2019), which mainly focuses on exploring better ways to reuse learned knowledge which is usually static, e.g., a frozen language model. In contrast, the acquired knowledge is continually accumulated in lifelong learning.
+
+Lifelong Sequence Generation (LSG) enables the model to learn sequentially from a stream of generation tasks. Sun et al. (2020) propose LAMOL which formalizes different types of tasks as ques
+
+tion answering and utilizes pseudo-sample replay to alleviate forgetting. Chuang et al. (2020) further improve LAMOL by knowledge distillation (Hinton et al., 2015). AdapterCL (Madotto et al., 2021) inserts task-specific modules into every transformer layer to learn new tasks while keeping the pre-trained LM and previous modules frozen. On the basis of AdapterCL, Zhang et al. (2022) introduce ACM which dynamically adds modules for learning new tasks depending on whether there are reusable previously inserted modules. Though ACM can enable knowledge transfer to some extent via module sharing, there is no explicit mechanism to encourage knowledge transfer across tasks, a common phenomenon of human learning.
+
+Summary. Existing work in LSG mainly focuses on mitigating the catastrophic forgetting of previously learned knowledge while paying little attention to knowledge transfer across tasks. In contrast to these lines of work, we aim to explicitly encourage forward knowledge transfer in LSG inspired by the way humans learn (Lake et al., 2017).
+
+# 3 Problem Formulation
+
+LSG involves learning from a stream of sequence generation tasks  $\mathbb{T} = (\mathcal{T}^1,\dots,\mathcal{T}^n)$ , where every task  $\mathcal{T}^i$  has its own training set  $D_{\mathrm{train}}^i$ , validation set  $D_{\mathrm{valid}}^i$ , and test set  $D_{\mathrm{test}}^i$ . Every dataset  $D$  contains a set of examples  $\{(X_j,Y_j)\}_{j=1}^{|D|}$ , where  $X_j$  and  $Y_j$  denote the input and output texts, respectively. At time step  $k$ , the model is trained on the training set  $D_{\mathrm{train}}^k$  of task  $\mathcal{T}^k$  and has no access to real samples of previously learned tasks.
+
+After the training on  $D_{\mathrm{train}}^k$ , the model is expected to perform well on all the tasks learned so far, i.e.,  $\mathcal{T}^1, \dots, \mathcal{T}^k$ , and will be evaluated on the test set  $D_{\mathrm{test}}^i$  of each task  $\mathcal{T}^i (1 \leq i \leq k)$  with corresponding evaluation metrics separately. Therefore, to achieve the goal of LSG, the model is required to alleviate the forgetting of acquired knowledge and better learn new patterns through possible forward knowledge transfer.
+
+# 3.1 Data Format
+
+Given an input-output text pair  $(X,Y)$  for a task, the model learns to decode the output text  $Y$  after reading the input  $X$ . Following Zhang et al. (2022), a natural language question  $Q$  describing the purpose of each task (task instruction) is inserted after the input to form a triple  $(X,Q,Y)$ ; see Appendix A.3 for an example. To learn a new
+
+![](images/450eca5aab1915246535faeb0fce7e60873cfbd4532ae4cc27d698383c52697d.jpg)  
+Figure 1: In the expansion stage, after inserting a new module (purple dashed hexagon) into each layer, DMEA dynamically determines the architecture by differentiable architecture search. Specifically, the outputs of all modules in the same layer are fused through dynamically initialized learnable coefficients. The weighted average is then passed to the next layer of the model for learning. Note that only newly added modules (purple dashed hexagons) are learnable modules in this stage. After several epochs of training, the module with the largest coefficient in every layer (green polygon) is selected for the new task. The selected module can be either a previous module (rectangle) or the newly added one (hexagon). Finally, newly added modules that are not selected (red dashed hexagons) will be discarded.
+
+task, the model is optimized to decode  $Y$  given  $X$  and  $Q$ . Denoting the concatenation of  $X, Q$  and  $Y$  as  $A$ , the autoregressive training objective is:
+
+$$
+\mathcal {L} _ {\text {t a s k}} = - \sum_ {j = m + 1} ^ {n} \log p _ {\theta} \left(A _ {j} \mid A _ {<   j}\right) \tag {1}
+$$
+
+where  $n$  is the total number of tokens in  $A$  and  $(A_{1},\dots,A_{m})$  is the concatenation of  $X$  and  $Q$ , and  $\theta$  denotes the model parameters.
+
+# 4 Methodology
+
+Inspired by how humans learn a new task (Fig. 5), DMEA divides the learning process into three stages. The expansion stage (\$4.1) first determines the model architecture dynamically. The selection stage (\$4.2) then selects the top-K most similar previous tasks which are utilized in the final adaptation stage (\$4.3) to facilitate adaptation to the new task. We also employ pseudo-sample replay along with a dynamic gradient scaling method to balance the learning of the new and replayed tasks.
+
+# 4.1 Expansion Stage
+
+Humans are able to determine whether previously acquired skills are sufficient to solve a new task. Our method DMEA aims to mimic this learning process in the expansion stage. It can dynamically decide whether to reuse modules of previous tasks or insert a new module in every transformer layer to learn novel knowledge. Inspired by Zhang et al. (2022), we utilize differentiable architecture search (Liu et al., 2019) to achieve this goal.
+
+Specifically, assuming that there are  $k$  modules (i.e., Adapter (Houlsby et al., 2019))  $\{m_1^l,\dots,m_k^l\}$  in layer  $l$  of the transformer model before learning a new task  $\mathcal{T}^j$ , we temporarily insert a new module  $m_{k + 1}^{l}$  into this layer at the beginning of the expansion stage. For each forward pass, after calculating the output  $h_t^l$  of every module  $m_t^l$  in the layer separately, we fuse all outputs  $\{h_1^l,\dots,h_{k + 1}^l\}$  through learnable coefficients  $\{\lambda_1^l,\dots,\lambda_{k + 1}^l\}$  as follows.
+
+$$
+\hat {h} ^ {l} = \sum_ {t = 1} ^ {k + 1} \frac {e ^ {\lambda_ {t} ^ {l}}}{\sum_ {s = 1} ^ {k + 1} e ^ {\lambda_ {s} ^ {l}}} h _ {t} ^ {l} \tag {2}
+$$
+
+The weighted average  $\hat{h}^l$  is then passed to the next part of the model for learning. After training the model on  $D_{\mathrm{train}}^j$  for several epochs using  $\mathcal{L}_{\mathrm{train}}$  (defined in §4.3), we select the module with the largest coefficient in every layer for the new task  $\mathcal{T}^j$ .
+
+Different from Zhang et al. (2022) which initialize  $\{\lambda_1^l,\dots,\lambda_{k + 1}^l\}$  with predefined hyperparameters, we propose to dynamically initialize learnable coefficients based on the correlation between the learned tasks  $\mathcal{T}^1,\ldots ,\mathcal{T}^{j - 1}$  and new task  $\mathcal{T}^j$  .Denoting the word frequency distribution of  $\mathcal{T}^i$  as  $f^i$  and all previous tasks sharing the module  $m_t^l$  as  $\mathcal{Z}_t^l$  the learnable coefficient  $\lambda_t^l$  is initialized as:
+
+$$
+\lambda_ {t} ^ {l} = \left\{ \begin{array}{l l} \max  _ {\mathcal {T} ^ {i} \in \mathcal {Z} _ {t} ^ {l}} \cos \left(f ^ {i}, f ^ {k + 1}\right), & 1 \leq t \leq k \\ \min  _ {1 \leq i \leq k} \lambda_ {i} ^ {l}, & t = k + 1 \end{array} \right. \tag {3}
+$$
+
+where  $\cos$  is the cosine similarity function and  $f^i$  is calculated based on the training set  $D_{\mathrm{train}}^i$ . In this way, a previous module shared by tasks with higher word frequency distribution similarity to the new task has a larger initial coefficient, increasing the tendency to reuse it. In addition, the coefficient  $\lambda_{k+1}^l$  of the newly added module  $m_{k+1}^l$  is initialized to the minimum value of the initial coefficients  $\{\lambda_1^l, \dots, \lambda_k^l\}$  of previously added modules  $\{m_1^l, \dots, m_k^l\}$  to encourage module reuse.
+
+![](images/89a8d5535fc74d549949e06feb59c9dc44a7f12bcdbdc1c3dd1fd105d9afa79c.jpg)  
+Figure 2: In the selection stage, DMEA selects the top-  $K$  most similar previous tasks through input subspace to facilitate adaptation to a new task. During adaptation, the output of the selected similar tasks is fused with that of the new task in every layer to enable forward knowledge transfer. Note that only modules selected for the new task (green polygons) are learnable modules in the adaptation stage. In addition, DMEA introduces dynamic gradient scaling to balance the learning of the new task and replayed tasks.
+
+The selected module in layer  $l$  can be either from previous modules  $\{m_1^l,\dots,m_k^l\}$  or the newly added one  $m_{k + 1}^{l}$  and will be tuned in the adaptation stage to accommodate new knowledge. We then discard newly added modules that are not selected. Note that only newly added modules and coefficients are learnable in the expansion stage; the pre-trained LM and previous modules are kept frozen.
+
+# 4.2 Selection Stage
+
+As humans, we can better acquire new knowledge by recognizing and utilizing knowledge from previously learned tasks that are similar (Lake et al., 2017). Based on the observation that the norm of one task's gradient projection onto the subspace of another task can characterize the correlation between them when the model architecture is static (Lin et al., 2022b), we further extend it to dynamic modules. Specifically, we obtain the input subspace of each task using modules of it and select the top- $K$  most similar previous tasks by input subspace similarity to facilitate adaptation to the new task  $\mathcal{T}^j$ . The model architecture induced from the expansion stage is used for selection and adaptation.
+
+Similar to Lin et al. (2022b), we adopt Singular Value Decomposition (SVD) to obtain the input subspace of each task. After training the model on  $D_{\mathrm{train}}^{j}$  for several epochs in the expansion stage, we randomly select  $n$  samples
+
+$\{X_{1},\ldots ,X_{n}\}$  from  $D_{\mathrm{train}}^j$  and obtain their representations  $\{\pmb {X}_1,\dots,\pmb {X}_n\} \in \mathbb{R}^m$  by forwardpropagating them through the network. We use the final-layer representation of the last non-padding token in the input as the sample representation.
+
+After obtaining the representation matrix  $\pmb{R}^{j} = [X_{1},\dots,X_{n}] \in \mathbb{R}^{m\times n}$  for task  $\mathcal{T}^j$ , we apply SVD to  $\pmb{R}^{j}$ , i.e.,  $\pmb{R}^{j} = U^{j}\pmb{\Sigma}^{j}(\pmb{V}^{j})'$ , where  $\pmb{U}^{j} = [\pmb{u}_{1}^{j},\dots,\pmb{u}_{m}^{j}] \in \mathbb{R}^{m\times m}$  is composed of left-singular vectors  $\pmb{u}_i^j$ ,  $\pmb{\Sigma}^{j} \in \mathbb{R}^{m\times n}$  is a rectangular diagonal matrix with singular values on the diagonal, and  $V^{j} = [\pmb{v}_{1}^{j},\dots,\pmb{v}_{n}^{j}] \in \mathbb{R}^{n\times n}$  is composed of right-singular vectors  $\pmb{v}_i^j$ . To obtain the input subspace  $S^j$  of  $\mathcal{T}^j$ , we select the first  $k$  left-singular vectors in  $U^{j}$  to form the bases  $B^{j} = [u_{1}^{j},\dots,u_{k}^{j}]$  for  $S^j$ , where  $k$  is determined by the requirement:  $||\pmb{R}_k^j||_F^2 \geq \epsilon^j ||\pmb{R}^j||_F^2$  with  $R_k^j$  being the  $k$ -rank approximation of  $R^j$ ,  $F$  being the Frobenius norm, and  $\epsilon^j$  being a predefined threshold.
+
+For the new task  $\mathcal{T}^j$ , the norm of its subspace projection onto the subspace of a previously learned task  $\mathcal{T}^i$  could characterize the similarity  $Q_{j,i}$  between these two tasks. More formally,
+
+$$
+\boldsymbol {Q} _ {j, i} = \frac {\left| \left| \operatorname {P r o j} _ {\boldsymbol {S} ^ {i}} \left(\boldsymbol {S} ^ {j}\right) \right| \right| _ {2}}{\left| \left| \boldsymbol {B} ^ {j} \right| \right| _ {2}} \tag {4}
+$$
+
+where  $\operatorname{Proj}_{\mathbf{S}^i}(\mathbf{S}^j) = \mathbf{B}^j\mathbf{B}^i (\mathbf{B}^i)'$  denotes the subspace projection. After getting the similarity scores  $Q_{j,i}, 1 \leq i < j$  of all previous tasks, we pick  $K$  tasks  $\mathbb{T}^{\mathrm{sim}} = (\mathcal{T}^{1},\dots,\mathcal{T}^{K})$  with the top- $K$  highest scores to facilitate adaptation to the new task  $\mathcal{T}^j$ .
+
+# 4.3 Adaptation Stage
+
+For adaptation to  $\mathcal{T}^j$ , assume that  $\mathbb{T}^{\mathrm{all}} = (\mathcal{T}^1,\dots,\mathcal{T}^K,\mathcal{T}^j)$  contains a total of  $r$  modules  $\{m_1^l,\ldots,m_r^l\}$  in layer  $l$ . During the training on  $D_{\mathrm{train}}^j$  using  $\mathcal{L}_{\mathrm{train}}$  (see Eq. (7)), for each sample in  $D_{\mathrm{train}}^j$ , we fuse the output  $h_s^l$  of each module  $m_s^l\in \{m_1^l,\dots,m_r^l\}$  by learnable coefficients  $\{\alpha_1^l,\dots,\alpha_r^l\}$  to enable forward knowledge transfer:
+
+$$
+\tilde {h} ^ {l} = \sum_ {s = 1} ^ {r} \frac {e ^ {\alpha_ {s} ^ {l}}}{\sum_ {u = 1} ^ {r} e ^ {\alpha_ {u} ^ {l}}} h _ {s} ^ {l} \tag {5}
+$$
+
+The learnable coefficients  $\{\alpha_1^l,\dots,\alpha_r^l\}$  are equally initialized to 1.0. Similar to the expansion stage, the fused output  $\tilde{h}^l$  is passed to the next part of the model for learning. After training, the learnable coefficients will be saved for inference. Note that we only tune modules selected in the expansion stage (can be modules of previous tasks or newly
+
+added modules) and learnable coefficients while keeping the pre-trained language model and other modules frozen.
+
+As there is no saved real sample of previously learned tasks when the model adapts to a new task, we also incorporate pseudo-sample replay (Sun et al., 2020) to alleviate the forgetting of acquired knowledge. We achieve this by simultaneously training the model as a task solver  $(\mathcal{L}_{\mathrm{task}}$  in §3.1) and as a data generator. When training as a data generator, the model learns to generate the triple  $(X,Q,Y)$  given a task-specific generation token  $G$  as input. Then before learning a new task, the model can generate pseudo samples of previous tasks, which are combined with new data for training to mitigate forgetting. Denoting the concatenation of  $G,X,Q$  and  $Y$  as  $A^{\prime}$ , the data generation loss is expressed as:
+
+$$
+\mathcal {L} _ {\mathrm {d a t a}} = - \sum_ {i = 2} ^ {m} \log p _ {\theta} \left(A _ {i} ^ {\prime} \mid A _ {<   i} ^ {\prime}\right) \tag {6}
+$$
+
+where  $m$  is the total number of tokens in  $A^{\prime}$ . The overall loss that DMEA optimizes for adapting to a new task is:
+
+$$
+\mathcal {L} _ {\text {t r a i n}} = \mathcal {L} _ {\text {t a s k}} + \mu \mathcal {L} _ {\text {d a t a}} \tag {7}
+$$
+
+where  $\mu$  is the weight of data generation loss.
+
+After the expansion stage, if the new task reuses some modules of previously learned tasks, the model will generate some pseudo samples of these tasks and train the model using  $\mathcal{L}_{\mathrm{train}}$  on the combination of new data and pseudo data. As the model has not seen new data before, the gradient norm of the new task on reused modules is much larger than that of replayed tasks. The learning process can easily be biased towards the new task which may affect previously acquired knowledge.
+
+Therefore, to balance the learning of the new task and replayed tasks, we introduce dynamic gradient scaling. Specifically, assuming that the new task  $\mathcal{T}^j$  reuses  $s$  modules  $\{m_1,\dots,m_s\}$  of a previous task  $\mathcal{T}^i$  in all layers, we randomly select  $q$  examples from  $D_{\mathrm{train}}^j$  and pseudo samples of  $\mathcal{T}^i$  separately and forwards them through the model to obtain the gradient of  $\mathcal{T}^j$  and  $\mathcal{T}^i$  using  $\mathcal{L}_{\mathrm{train}}$  with regard to reused modules  $\{m_1,\dots,m_s\}$ , denoted as  $g^{j}$  and  $g^{i}$ , respectively. The dynamic scale factor  $\eta_t^i$  is then calculated as:
+
+$$
+\eta_ {t} ^ {i} = \left(\frac {\left| \left| g ^ {j} \right| \right| _ {2}}{\left| \left| g ^ {i} \right| \right| _ {2}} - 1\right) e ^ {- t} + 1 \tag {8}
+$$
+
+<table><tr><td colspan="2">Methods</td><td>Finetune</td><td>EWC</td><td>LAMOL</td><td>Metac</td><td>Adapter +LAMOL</td><td>AdapterCL</td><td>ACM</td><td>DMEA</td><td>MTL</td></tr><tr><td colspan="2">Tune Whole Model?</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✘</td><td>✘</td><td>✘</td><td>✓</td></tr><tr><td colspan="2">Pseudo-sample Replay?</td><td>✘</td><td>✘</td><td>✓</td><td>✓</td><td>✓</td><td>✘</td><td>✓</td><td>✓</td><td>-</td></tr><tr><td rowspan="4">Similar Tasks</td><td># 1</td><td>42.9±0.2</td><td>56.6±0.1</td><td>65.5±0.3</td><td>65.2±0.4</td><td>64.5±0.2</td><td>63.4±0.3</td><td>64.8±0.3</td><td>65.8±0.2</td><td>67.1±0.2</td></tr><tr><td># 2</td><td>51.8±0.1</td><td>61.0±0.2</td><td>65.2±0.4</td><td>65.0±0.2</td><td>64.6±0.4</td><td>63.4±0.3</td><td>64.9±0.2</td><td>65.6±0.2</td><td>-</td></tr><tr><td># 3</td><td>45.2±0.2</td><td>57.8±0.1</td><td>65.7±0.2</td><td>65.4±0.4</td><td>64.2±0.3</td><td>63.4±0.3</td><td>64.5±0.2</td><td>65.5±0.1</td><td>-</td></tr><tr><td># 4</td><td>31.4±0.4</td><td>46.6±0.3</td><td>65.6±0.3</td><td>65.5±0.1</td><td>65.1±0.1</td><td>63.4±0.3</td><td>65.4±0.3</td><td>66.2±0.3</td><td>-</td></tr><tr><td colspan="2">Average</td><td>42.9±8.5</td><td>55.5±6.2</td><td>65.5±0.2</td><td>65.3±0.2</td><td>64.6±0.4</td><td>63.4±0.0</td><td>64.9±0.4</td><td>65.8±0.3</td><td>67.1±0.0</td></tr><tr><td rowspan="4">Random Tasks</td><td># 1</td><td>33.8±0.4</td><td>37.6±0.5</td><td>55.7±0.4</td><td>55.9±0.3</td><td>53.2±0.2</td><td>56.1±0.5</td><td>56.7±0.3</td><td>57.5±0.3</td><td>59.7±0.2</td></tr><tr><td># 2</td><td>33.1±0.3</td><td>38.4±0.4</td><td>62.6±0.2</td><td>62.4±0.4</td><td>61.8±0.3</td><td>64.2±0.2</td><td>64.9±0.4</td><td>65.6±0.3</td><td>67.5±0.1</td></tr><tr><td># 3</td><td>25.7±0.1</td><td>43.2±0.3</td><td>54.8±0.1</td><td>55.4±0.2</td><td>53.6±0.3</td><td>55.6±0.4</td><td>56.3±0.1</td><td>57.3±0.2</td><td>60.4±0.3</td></tr><tr><td># 4</td><td>34.2±0.3</td><td>48.9±0.1</td><td>64.7±0.4</td><td>65.3±0.3</td><td>62.5±0.1</td><td>65.4±0.3</td><td>66.2±0.2</td><td>67.4±0.1</td><td>69.8±0.1</td></tr><tr><td colspan="2">Average</td><td>31.7±4.0</td><td>42.0±5.2</td><td>59.5±4.9</td><td>59.8±4.9</td><td>57.8±5.1</td><td>60.3±5.2</td><td>61.0±5.3</td><td>62.0±5.3</td><td>64.4±5.1</td></tr></table>
+
+Table 1: The average performance score for each task sequence after learning all tasks. Bold indicates the best score. 'MTL' stands for 'multi-task learning', serving as the upper bound for LSG. In each scenario, DMEA is significantly better than ACM with  $p$ -value  $< 0.05$  (paired t-test). Note that while LAMOL and Metac are not directly comparable to other adapter-based methods as their learnable parameters are orders of magnitude larger, DMEA still outperforms them in most cases. The comparison of learnable parameters and computational resources between ACM and DMEA is reported in Appendix A.8.
+
+where  $t$  is the number of completed training epochs. After dynamic gradient scaling, the total loss for jointly learning  $\mathcal{T}^j$  and  $\mathcal{T}^i$  is:
+
+$$
+\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {t r a i n}} ^ {j} + \eta_ {t} ^ {i} \mathcal {L} _ {\text {t r a i n}} ^ {i} \tag {9}
+$$
+
+Note that in the early stage of training, the value of  $t$  is small.  $\eta_t$  is greater than 1 to balance the gradient of the new task  $\mathcal{T}^j$  and the replayed task  $\mathcal{T}^i$ . When the model has seen enough new data in the late stage of training (no need to balance),  $\eta_t$  is approximately equal to 1 as the value of  $t$  is large.
+
+# 5 Experimental Setup
+
+In this section, we first describe investigated tasks and then introduce methods compared in our work.
+
+# 5.1 Tasks
+
+Four representative sequence generation tasks are investigated in our work: natural language generation, summarization, task-oriented dialogue and SQL query generation. Following Zhang et al. (2022), we consider two different scenarios: (i) LSG on similar tasks where the model learns a sequence of tasks of the same type but different domains, and (ii) LSG on random tasks where the model learns knowledge from different types of tasks. For LSG on similar tasks, we use five different domains from two natural language generation datasets (RNNLG (Wen et al., 2015) and E2ENLG (Novikova et al., 2017)) to form the task sequences. We further incorporate summarization (CNNDM (See et al., 2017)), task-oriented dialogue (MultiWOZ (Budzianowski et al., 2018)) and SQL query generation (WikisSQL (Zhong et al., 2017)) to form the task se
+
+quences for LSG on random tasks. For each scenario, we randomly select four different orders $^2$  (Appendix A.4) and run experiments for every order five times with different random seeds (20 runs per scenario). For each order, we report the average of all learned tasks' performance scores following Zhang et al. (2022); see Appendix A.5 for details of task-specific evaluation metrics.
+
+# 5.2 Methods Compared
+
+Following Zhang et al. (2022), we use GPT-2 (Radford et al., 2019) as the backbone model and Adapter (Houlsby et al., 2019) as the insertable module, and compare with the following methods:
+
+- Finetune tunes the whole GPT-2 model only on the training data of the new task during the LSG process.  
+- EWC (Kirkpatrick et al., 2017) constrains the update of parameters that are important to previously learned tasks to alleviate forgetting.  
+- LAMOL (Sun et al., 2020) tunes the whole GPT-2 model with pseudo-sample replay.  
+- Metac-Adapt (Metac) (Wang et al., 2023) adapts LAMOL towards better semantic space for generating pseudo samples.  
+- Adapter+LAMOL only inserts adapter modules for the first task and tunes these modules with pseudo-sample replay while keeping the backbone model frozen.  
+- AdapterCL (Madotto et al., 2021) inserts task
+
+<table><tr><td>Method</td><td>Similar Tasks</td><td>Random Tasks</td></tr><tr><td>DMEA</td><td>65.8</td><td>57.3</td></tr><tr><td>w.o. transfer</td><td>64.9</td><td>56.5</td></tr><tr><td>w.o. scaling</td><td>65.5</td><td>56.8</td></tr><tr><td>w.o. initialization</td><td>65.4</td><td>57.0</td></tr></table>
+
+specific adapter modules for every new task while keeping the backbone model and previous modules frozen.
+
+- ACM (Zhang et al., 2022) dynamically adds adapter modules for new tasks depending on whether there are reusable previous modules to improve the performance and parameter efficiency of AdapterCL. It is the state-of-the-art on LSG.
+
+# 6 Results and Analysis
+
+# 6.1 Main Results
+
+Table 1 shows the average performance score for each task sequence after learning all tasks (see Appendix A.7 for the performance of each task). From the results, we can see that DMEA outperforms previous baselines in all LSG settings, which demonstrates the superiority of our method. Note that while the learnable parameters of LAMOL are orders of magnitude larger, DMEA still achieves better performance than LAMOL in 7 out of 8 runs, showing its effectiveness in LSG.
+
+Simply fine-tuning the model with new samples leads to poor performance due to catastrophic forgetting. Although EWC adopts Fisher information matrix to alleviate forgetting, its performance is still much worse than other memory-based baselines, indicating the importance of pseudo-sample replay. When learning from a sequence of similar tasks, Adapter+LAMOL performs better than AdapterCL as AdapterCL applies parameter isolation to different tasks which might prevent positive knowledge transfer across tasks. However, this is not the case when learning from random tasks: AdapterCL achieves much better results than Adapter+LAMOL as AdapterCL can avoid catastrophic forgetting by assigning different learnable parameters to each task. The performance of ACM is superior to Adapter+LAMOL and AdapterCL in both scenarios, showing the effectiveness of
+
+Table 2: The average performance score for different ablations: (i) without forward knowledge transfer, (ii) without dynamic gradient scaling, and (iii) without dynamic initialization. All components improve the performance of our method.  
+
+<table><tr><td colspan="2">Time Step</td><td>AdapterCL</td><td>ACM</td><td>DMEA</td></tr><tr><td rowspan="4">Similar</td><td>2</td><td>55.8(+0.0)</td><td>56.0(+0.1)</td><td>56.3(+0.3)</td></tr><tr><td>3</td><td>58.6(+0.0)</td><td>59.1(+0.4)</td><td>59.5(+0.6)</td></tr><tr><td>4</td><td>61.2(+0.0)</td><td>62.5(+0.6)</td><td>63.2(+0.9)</td></tr><tr><td>5</td><td>63.4(+0.0)</td><td>64.8(+0.3)</td><td>65.8(+1.0)</td></tr><tr><td rowspan="4">Random</td><td>2</td><td>55.4(+0.0)</td><td>56.3(+0.9)</td><td>57.1(+2.1)</td></tr><tr><td>3</td><td>58.4(+0.0)</td><td>58.9(+0.3)</td><td>59.7(+1.3)</td></tr><tr><td>4</td><td>64.0(+0.0)</td><td>64.3(+0.7)</td><td>65.4(+1.4)</td></tr><tr><td>5</td><td>64.2(+0.0)</td><td>64.9(+0.6)</td><td>65.6(+1.1)</td></tr></table>
+
+Table 3: The average performance score and forward knowledge transfer (FKT) of different methods at every time step. FKT is reported in parentheses.
+
+its adaptive compositional architecture. However, ACM has no explicit mechanism to encourage forward knowledge transfer in LSG, which is actually the human learning paradigm. Our proposed DMEA consistently outperforms ACM by dynamically leveraging previously acquired knowledge to facilitate adaptation to new tasks.
+
+# 6.2 Ablation Study
+
+We conduct several ablations to analyze the contribution of different components of DMEA. In particular, we investigate three variants of DMEA  $(a)$  without selecting similar previous tasks for forward knowledge transfer (w.o. transfer),  $(b)$  removing dynamic gradient scaling (w.o. scaling), and  $(c)$  without dynamically initializing learnable coefficients (w.o. initialization). For each scenario, i.e., similar tasks or random tasks, we randomly pick one sequence for experiments. Table 2 reports the average performance score after learning all tasks for different ablations.
+
+From the results, we can observe that all components contribute to the average performance. Removing forward knowledge transfer leads to a significant performance drop in both scenarios, indicating that selecting top-  $K$  most similar previous tasks can indeed discover and transfer useful learned knowledge to facilitate adaptation to the new task. The adoption of dynamic gradient scaling yields a moderate performance boost as it can balance the learning of the new task and replayed tasks to mitigate catastrophic forgetting. Dynamic initialization of learnable coefficients also facilitates performance improvement, demonstrating the effectiveness of leveraging the similarity of word frequency distributions between tasks.
+
+# 6.3 Further Analysis
+
+Quantify Forward Knowledge Transfer. Following Ke et al. (2020), we define metrics quanti
+
+<table><tr><td>Metrics</td><td>Similar Tasks</td><td>Random Tasks</td></tr><tr><td>Input Subspace</td><td>65.8</td><td>57.3</td></tr><tr><td>Frequency</td><td>65.3</td><td>56.9</td></tr><tr><td>Representation</td><td>65.2</td><td>56.9</td></tr><tr><td>w.o. transfer</td><td>64.9</td><td>56.5</td></tr></table>
+
+fying forward knowledge transfer (FKT) at every time step  $t$  as:
+
+$$
+\mathrm {F W T} = \frac {1}{t - 1} \sum_ {i = 2} ^ {t} R _ {i, i} - \bar {d} _ {i}. \tag {10}
+$$
+
+where  $R_{i,j}$  is the performance score on  $\mathcal{T}^j$  after learning  $\mathcal{T}^i$  and  $\bar{d}_i$  refers to the performance of training  $\mathcal{T}^i$  individually, which is actually the result of AdapterCL. For each scenario, we randomly select one sequence for analysis and report the average performance score along with FKT at each step in Table 3. From the results, we can see that DMEA consistently outperforms ACM in terms of the average performance score and FKT at all steps, demonstrating that DMEA can better facilitate positive knowledge transfer.
+
+Input Subspace vs. Other Similarity Metrics. The ablation (w.o. transfer) in §6.2 demonstrates the importance of selecting similar learned tasks. To further investigate whether different similarity metrics influence the performance of DMEA, we conduct controlled experiments with two new metrics: (a) cosine similarity of word frequency distributions between different tasks (frequency), and (b) cosine similarity of the representations of selected samples from different tasks $^3$  (representation). For each scenario, we use the same sequence as §6.2. From the results in Table 4, we can observe that selecting similar previous tasks by input subspace consistently outperforms using other similarity metrics, demonstrating its superiority.
+
+Robustness to Module Type To verify whether the performance gain of DMEA is consistent across different types of modules, we extend the experiments to prefix-tuning (Li and Liang, 2021) and LoRA (Hu et al., 2022). We randomly pick four sequences for experiments and report the average result in Table 5. we can see that DMEA still outperforms ACM when using other architecture as
+
+Table 4: The average performance score using different similarity metrics.  
+
+<table><tr><td>Module Type</td><td>ACM</td><td>DMEA</td></tr><tr><td>Prefix-tuning</td><td>62.6</td><td>63.4</td></tr><tr><td>LoRA</td><td>63.1</td><td>64.2</td></tr></table>
+
+Table 5: The average performance score of ACM and DMEA with different module types.
+
+the insertable module, showing its robustness to module type.
+
+Longer Sequence. As mentioned in §5.1, we mainly conduct experiments on sequences consisting of 5 tasks following Zhang et al. (2022). To verify whether DMEA can still outperform the baselines when learning from a larger number of tasks, we further combine all tasks investigated in this work to form a longer sequence of 8 tasks. We evaluate ACM and DMEA on this longer sequence with 3 different orders and report the average performance score for each order after learning all tasks in Fig. 3. We can observe that DMEA is still superior to ACM when learning from longer sequences.
+
+Quality of Pseudo Data Fig. 4 shows several pseudo samples generated by DMEA. We can see that DMEA can indeed generate high-quality pseudo samples to mitigate the forgetting of previously learned knowledge. However, the generated pseudo data could also be noisy as shown at the bottom of the figure, which might hinder further performance improvement.
+
+Other Types of Tasks To explore whether the performance gain of DMEA is consistent on other types of tasks, we further include three new tasks: sentiment analysis (SST (Socher et al., 2013)), semantic role labeling (SRL (He et al., 2015)) and question answering (SQuAD (Rajpurkar et al., 2016)). We randomly select two tasks from the original task set three times and combine them with new tasks to form three task sequences. From the results shown in Table 6, we can observe that DMEA performs better than ACM on all sequences, showing its robustness to task types.
+
+Different Pseudo-data Sampling Ratios Following Zhang et al. (2022), we set the pseudo-data sampling ratio to 0.2. To validate whether different pseudo-data sampling rates influence the performance gain of DMEA, we conduct controlled experiments with sampling rates  $\{0.05, 0.1, 0.4\}$ . We randomly pick three sequences for experiments and report the performance comparison between
+
+![](images/14dfefe1a01b04bb167cec3a94f88c5617885f02298dccfb5404242347cf3e6e.jpg)  
+Figure 3: The average performance score for every order after learning all 8 tasks of the longer sequence.  
+Figure 4: Some examples of generated pseudo data. We color the task instruction in blue and output text in gray. Missing/wrong information is colored in red.
+
+# High-quality Data
+
+name[The Cricketers], eatType[restaurant], food[French], priceRange[moderate], near[Rainbow Vegetarian Cafe] what is the natural language form? The Cricketers is a French restaurant next to the Rainbow Vegetarian Cafe with moderate prices and a French taste.
+
+# Noisy Data
+
+the table has columns week number, date, opponent, result, record and key words max, min, count, sum, avg,  $= , > , < ,$  op, select, where, and, col, table, caption - - who the opponent was on the weekend where the record was 0-0? what is the translation from english to sql? select opponent from table where record  $= 0$
+
+ACM and DMEA in Table 7. We can see that DMEA consistently outperforms ACM in all cases, demonstrating its effectiveness.
+
+In addition, we show case studies of learned model architecture, model output, dynamic gradient scaling and task selection, generalization of dynamic initialization, and potential real-world applications in Appendix A.9  $\sim$  A.14, respectively.
+
+# 7 Conclusion
+
+In this work, we have introduced DMEA for lifelong sequence generation (LSG). DMEA leverages task correlations to dynamically determine the suitable architecture required to acquire novel knowledge of a new task and selects the most similar previous tasks through input subspace to facilitate knowledge transfer. It uses pseudo-sample replay along with dynamic gradient scaling to balance the learning of the new task and replayed tasks to further alleviate forgetting. With extensive experiments and analysis we have shown that DMEA consistently outperforms previous methods in different LSG settings. In the future, we would like to investigate ways to improve the quality of pseudo data and explore more metrics for task similarity.
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">Sequence</td><td rowspan="2">Average</td></tr><tr><td>(i)</td><td>(ii)</td><td>(iii)</td></tr><tr><td>ACM</td><td>68.4</td><td>63.6</td><td>71.8</td><td>67.9</td></tr><tr><td>DMEA</td><td>69.5</td><td>64.4</td><td>73.0</td><td>69.0</td></tr></table>
+
+Table 6: The average performance score for every sequence after learning all new types of tasks.  
+
+<table><tr><td>Sampling Ratio</td><td>0.05</td><td>0.1</td><td>0.4</td></tr><tr><td>ACM</td><td>61.7</td><td>62.0</td><td>62.1</td></tr><tr><td>DMEA</td><td>62.5</td><td>63.1</td><td>62.8</td></tr></table>
+
+Table 7: The average performance score of ACM and DMEA with different pseudo-data sampling ratios.
+
+# Limitations
+
+Although effective, DMEA has couple of limitations:
+
+- DMEA mainly focuses on the setting where every task has plenty of training samples. In contrast, humans can easily learn to perform new tasks with only few data, which is a hallmark of human intelligence. We leave how to explore lifelong sequence generation in few-shot settings as future work.  
+- DMEA does not consider machine translation, a sequence generation task that might involve vocabulary changes. One potential solution is to use multilingual pre-trained language models.
+
+# References
+
+Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Inigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gašić. 2018. MultiWOZ - a large-scale multi-domain Wizard-of-Oz dataset for task-oriented dialogue modelling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 5016-5026, Brussels, Belgium. Association for Computational Linguistics.
+
+Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. 2019. Efficient lifelong learning with A-Geom. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net.
+
+Tianqi Chen, Ian J. Goodfellow, and Jonathon Shlens. 2016. Net2net: Accelerating learning via knowledge transfer. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings.
+
+Yung-Sung Chuang, Shang-Yu Su, and Yun-Nung Chen. 2020. Lifelong language knowledge distillation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2914-2924, Online. Association for Computational Linguistics.  
+Wafaa S El-Kassas, Cherif R Salama, Ahmed A Rafea, and Hoda K Mohamed. 2021. Automatic text summarization: A comprehensive survey. Expert Systems with Applications, 165:113679.  
+Chrisantha Fernando, Dylan Banarse, Charles Blundell, Yori Zwols, David Ha, Andrei A Rusu, Alexander Pritzel, and Daan Wierstra. 2017. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint arXiv:1701.08734.  
+Donghoon Ham, Jeong-Gwan Lee, Youngsoo Jang, and Kee-Eung Kim. 2020. End-to-end neural pipeline for goal-oriented dialogue systems using GPT-2. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 583-592, Online. Association for Computational Linguistics.  
+Xu Han, Yi Dai, Tianyu Gao, Yankai Lin, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. 2020. Continual relation learning via episodic memory activation and reconsolidation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6429-6440, Online. Association for Computational Linguistics.  
+Luheng He, Mike Lewis, and Luke Zettlemoyer. 2015. Question-answer driven semantic role labeling: Using natural language to annotate natural language. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 643-653, Lisbon, Portugal. Association for Computational Linguistics.  
+Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2(7).  
+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 International Conference on Machine Learning, pages 2790-2799. PMLR.  
+Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.  
+Zixuan Ke, Bing Liu, and Xingchang Huang. 2020. Continual learning of a mixed sequence of similar and dissimilar tasks. In Advances in Neural Information Processing Systems, volume 33, pages 18493-18504. Curran Associates, Inc.
+
+Zixuan Ke, Bing Liu, Nianzu Ma, Hu Xu, and Lei Shu. 2021. Achieving forgetting prevention and knowledge transfer in continual learning. Advances in Neural Information Processing Systems, 34:22443-22456.  
+James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526.  
+Brenden M. Lake, Tomer D. Ullman, Joshua B. Tenenbaum, and Samuel J. Gershman. 2017. Building machines that learn and think like people. Behavioral and Brain Sciences, 40:e253.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics.  
+Zhizhong Li and Derek Hoiem. 2017. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947.  
+Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. 2022a. Beyond not-forgetting: Continual learning with backward knowledge transfer. In Advances in Neural Information Processing Systems.  
+Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. 2022b. TRGP: Trust region gradient projection for continual learning. In International Conference on Learning Representations.  
+Hanxiao Liu, Karen Simonyan, and Yiming Yang. 2019. DARTS: Differentiable architecture search. In International Conference on Learning Representations.  
+David Lopez-Paz and Marc'Aurelio Ranzato. 2017. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30.  
+Andrea Madotto, Zhaojiang Lin, Zhenpeng Zhou, Seungwhan Moon, Paul Crook, Bing Liu, Zhou Yu, Eunjoon Cho, Pascale Fung, and Zhiguang Wang. 2021. Continual learning in task-oriented dialogue systems. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7452-7467, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Saeed Masoudnia and Reza Ebrahimpour. 2014. Mixture of experts: a literature survey. Artificial Intelligence Review, 42(2):275-293.  
+Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in connectionist networks: The
+
+sequential learning problem. In Psychology of learning and motivation, volume 24, pages 109-165. Elsevier.  
+Fei Mi, Liangwei Chen, Mengjie Zhao, Minlie Huang, and Boi Faltings. 2020. Continual learning for natural language generation in task-oriented dialog systems. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 3461-3474, Online. Association for Computational Linguistics.  
+Jekaterina Novikova, Ondrej Dušek, and Verena Rieser. 2017. The E2E dataset: New challenges for end-to-end generation. In Proceedings of the 18th Annual SIGdial Meeting on Discourse and Dialogue, pages 201-206, Saarbrücken, Germany. 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.  
+Chengwei Qin and Shafiq Joty. 2022a. Continual few-shot relation learning via embedding space regularization and data augmentation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2776-2789, Dublin, Ireland. Association for Computational Linguistics.  
+Chengwei Qin and Shafiq Joty. 2022b. LFPT5: A unified framework for lifelong few-shot language learning based on prompt tuning of t5. In International Conference on Learning Representations.  
+Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1-67.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383-2392, Austin, Texas. Association for Computational Linguistics.  
+Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. 2017. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 5533-5542. IEEE Computer Society.
+
+Hippolyt Ritter, Aleksandar Botev, and David Barber. 2018. Online structured laplace approximations for overcoming catastrophic forgetting. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, December 3-8, 2018, Montreal, Canada, pages 3742-3752.  
+Sebastian Ruder, Matthew E. Peters, Swabha Swayamdipta, and Thomas Wolf. 2019. Transfer learning in natural language processing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Tutorials, pages 15-18, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. 2016. Progressive neural networks. arXiv preprint arXiv:1606.04671.  
+Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get to the point: Summarization with pointer-generator networks. arXiv preprint arXiv:1704.04368.  
+Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. 2017. Continual learning with deep generative replay. 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 2990-2999.  
+Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1631-1642, Seattle, Washington, USA. Association for Computational Linguistics.  
+Fan-Keng Sun, Cheng-Hao Ho, and Hung-Yi Lee. 2020. Lamol: Language modeling for lifelong language learning. In International Conference on Learning Representations.  
+Han Wang, Ruiliu Fu, Xuejun Zhang, Jun Zhou, and Qingwei Zhao. 2023. Metacognitive adaptation to enhance lifelong language learning. IEICE TRANSACTIONS on Information and Systems, 106(1):86-90.  
+Tsung-Hsien Wen, Milica Gašić, Nikola Mrkšić, PeiHao Su, David Vandyke, and Steve Young. 2015. Semantically conditioned LSTM-based natural language generation for spoken dialogue systems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1711-1721, Lisbon, Portugal. Association for Computational Linguistics.
+
+![](images/e865bc29368232fe97d71edc7bc78327bb0f61bcb86bf8267b178c97d3fac0a2.jpg)  
+Figure 5: Given three learned skills, i.e., swimming, running and boating, humans can determine that these skills are not sufficient for diving. And after realizing that swimming is the most similar learned skill, they only need to learn the new aspect, i.e., how to safely jump off the diving platform to master the new diving skill.  
+Figure 7: An example of the task instruction for E2ENLG. We color the task instruction in blue.
+
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi 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 Rush. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.
+
+Friedemann Zenke, Ben Poole, and Surya Ganguli. 2017. Continual learning through synaptic intelligence. In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pages 3987-3995. PMLR.
+
+Yanzhe Zhang, Xuezhi Wang, and Diyi Yang. 2022. Continual sequence generation with adaptive compositional modules. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3653-3667, Dublin, Ireland. Association for Computational Linguistics.
+
+Victor Zhong, Caiming Xiong, and Richard Socher. 2017. Seq2sql: Generating structured queries from natural language using reinforcement learning. arXiv preprint arXiv:1709.00103.
+
+# A Appendix
+
+# A.1 Illustration of Human Learning
+
+We show the illustration of human learning in Fig. 5.
+
+![](images/f01ec581e163dc5a7f91933658c5ea134271224886b2068f233acf134c650805.jpg)  
+Figure 6: The effect of large gradient norm.  $g_{old}$ ,  $g_{new}$ , and  $g_{sum}$  represent the gradient of replayed tasks, the gradient of the new task, and the aggregated gradient, respectively. The blue arrows show the projection of  $g_{sum}$  onto  $g_{old}$ . If the gradient norm of  $g_{new}$  is large (right part of the figure), this projection might deviate too much from  $g_{old}$ .
+
+name[Thevaults], eatType[pub], priceRange[moderate], customer rating[1 out of 5], near[Caf Adriatic] what is the natural language form? A moderately priced pub, named The vaults, is located near Caf Adriatic. It has a customer rating of 1 out of 5.
+
+# A.2 Effect of Large Gradient Norm
+
+As shown in the Fig. 6, if the gradient norm of the new task  $g_{new}$  is large, the projection of the aggregated gradient  $g_{sum}$  onto the gradient of replayed tasks  $g_{old}$  might deviate too much from  $g_{old}$ , leading to more severe forgetting.
+
+# A.3 Task Instruction Example
+
+Following Zhang et al. (2022), we insert a natural language question describing the purpose of every task (task instruction) after the input of each sample. Fig. 7 shows an example of the task instruction for E2ENLG (Novikova et al., 2017).
+
+# A.4 Task Orders
+
+We present different task orders for two LSG scenarios in Table 8.
+
+# A.5 Task-specific Evaluation Metrics
+
+We report details of task-specific evaluation metrics in Table 9.
+
+# A.6 Implementation Details
+
+All methods are implemented with PyTorch/Transformers library (Wolf et al., 2020). We adopt AdapterHub (Pfeiffer et al., 2020) to implement adapter modules. For hyperparameters, we mainly follow the settings in Zhang et al. (2022) to have a fair comparison. In the expansion stage, we train the model for 6 epochs before selecting modules. In the adaptation stage, we set the number  $(n)$  of samples selected to obtain the
+
+<table><tr><td colspan="2">Order</td><td>Task Sequence</td></tr><tr><td rowspan="4">Similar Tasks</td><td># 1</td><td>e2e → res → hotel → tv → laptop</td></tr><tr><td># 2</td><td>e2e → tv → res → laptop → hotel</td></tr><tr><td># 3</td><td>res → hotel → e2e → laptop → tv</td></tr><tr><td># 4</td><td>laptop → hotel → res → tv → e2e</td></tr><tr><td rowspan="4">Random Tasks</td><td># 1</td><td>mwoz → cnn → e2e → res → hotel</td></tr><tr><td># 2</td><td>e2e → sql → hotel → mwoz → res</td></tr><tr><td># 3</td><td>cnn → hotel → sql → e2e → mwoz</td></tr><tr><td># 4</td><td>e2e → mwoz → laptop → sql → tv</td></tr></table>
+
+Table 8: Different task orders for each scenario. 'e2e' stands for E2ENLG. 'res', 'hotel', 'laptop' and 'tv' are four domains in RNNLG (restaurant, hotel, laptop and television). 'sql', 'cnn' and 'mwoz' respectively stand for 'WikiSQL' (SQL query generation), 'CN-NDM' (summarization) and 'MultiWOZ' (task-oriented dialogue).  
+
+<table><tr><td>Dataset</td><td>Metric</td></tr><tr><td>RNNLG</td><td></td></tr><tr><td>E2ENLG</td><td>A-RG</td></tr><tr><td>CNNDM</td><td></td></tr><tr><td>WikiSQL</td><td>lfEM</td></tr><tr><td>MultiWOZ</td><td>dsEM</td></tr></table>
+
+input subspace as 100. The threshold  $\epsilon$  is set as 0.95 for selecting left-singular vectors. We adopt 1 for the number of similar tasks  $K$ . For dynamic gradient scaling, we set 100 for the number  $(q)$  of examples selected to calculate the gradient.
+
+# A.7 Performance of Each Task
+
+Table 10 shows the performance of each task for every task sequence after learning all tasks.
+
+# A.8 Number of Learnable Parameters and Computational Resources
+
+We present the average number of learnable parameters and average running time for ACM and DMEA in Table 11. From the comparison, we can observe that DMEA can outperform ACM with a negligible increase in learnable parameters and computational resources.
+
+# A.9 Learned Model Architecture
+
+To further demonstrate that dynamically initializing learnable coefficients can facilitate finding the
+
+Table 9: Details of task-specific evaluation metrics. 'ARG', 'lfEM' and 'dsEM' respectively stand for 'average of ROUGE-1, ROUGE-2 and ROUGE-L scores', 'exact match of logical forms' and 'exact match of dialogue state'.  
+
+<table><tr><td>Similar #1</td><td>e2e</td><td>res</td><td>hotel</td><td>tv</td><td>laptop</td><td>Avg</td></tr><tr><td>ACM</td><td>49.6</td><td>65.7</td><td>65.8</td><td>71.6</td><td>71.1</td><td>64.8</td></tr><tr><td>DMEA</td><td>49.2</td><td>67.1</td><td>68.1</td><td>72.5</td><td>72.0</td><td>65.8</td></tr><tr><td>Similar #2</td><td>e2e</td><td>tv</td><td>res</td><td>laptop</td><td>hotel</td><td>Avg</td></tr><tr><td>ACM</td><td>48.7</td><td>74.0</td><td>64.4</td><td>72.6</td><td>65.0</td><td>64.9</td></tr><tr><td>DMEA</td><td>47.9</td><td>74.9</td><td>64.9</td><td>74.2</td><td>65.9</td><td>65.6</td></tr><tr><td>Similar #3</td><td>res</td><td>hotel</td><td>e2e</td><td>laptop</td><td>tv</td><td>Avg</td></tr><tr><td>ACM</td><td>65.6</td><td>67.3</td><td>48.5</td><td>72.0</td><td>69.3</td><td>64.5</td></tr><tr><td>DMEA</td><td>66.9</td><td>66.6</td><td>49.5</td><td>73.7</td><td>70.9</td><td>65.5</td></tr><tr><td>Similar #4</td><td>laptop</td><td>hotel</td><td>res</td><td>tv</td><td>e2e</td><td>Avg</td></tr><tr><td>ACM</td><td>73.1</td><td>66.8</td><td>66.9</td><td>72.4</td><td>47.6</td><td>65.4</td></tr><tr><td>DMEA</td><td>74.6</td><td>67.6</td><td>67.4</td><td>72.9</td><td>48.3</td><td>66.2</td></tr><tr><td>Random #1</td><td>mwoz</td><td>cnn</td><td>e2e</td><td>res</td><td>hotel</td><td>Avg</td></tr><tr><td>ACM</td><td>81.6</td><td>26.0</td><td>47.5</td><td>64.5</td><td>64.1</td><td>56.7</td></tr><tr><td>DMEA</td><td>81.6</td><td>26.5</td><td>48.1</td><td>65.7</td><td>65.4</td><td>57.5</td></tr><tr><td>Random #2</td><td>e2e</td><td>sql</td><td>hotel</td><td>mwoz</td><td>res</td><td>Avg</td></tr><tr><td>ACM</td><td>48.4</td><td>62.7</td><td>64.6</td><td>84.8</td><td>64.0</td><td>64.9</td></tr><tr><td>DMEA</td><td>48.7</td><td>64.9</td><td>64.9</td><td>84.9</td><td>64.6</td><td>65.6</td></tr><tr><td>Random #3</td><td>cnn</td><td>hotel</td><td>sql</td><td>e2e</td><td>mwoz</td><td>Avg</td></tr><tr><td>ACM</td><td>26.3</td><td>63.2</td><td>62.1</td><td>47.7</td><td>82.4</td><td>56.3</td></tr><tr><td>DMEA</td><td>26.6</td><td>65.0</td><td>63.1</td><td>48.4</td><td>83.5</td><td>57.3</td></tr><tr><td>Random #4</td><td>e2e</td><td>mwoz</td><td>laptop</td><td>sql</td><td>tv</td><td>Avg</td></tr><tr><td>ACM</td><td>48.6</td><td>80.5</td><td>70.3</td><td>63.5</td><td>68.2</td><td>66.2</td></tr><tr><td>DMEA</td><td>49.0</td><td>82.8</td><td>71.7</td><td>64.4</td><td>68.9</td><td>67.4</td></tr></table>
+
+Table 10: The performance of each task for every sequence after learning all tasks.  
+
+<table><tr><td>Method</td><td>Avg Para Num</td><td>Avg Time (min)</td></tr><tr><td>ACM</td><td>4.6M</td><td>218.1</td></tr><tr><td>DMEA</td><td>4.8M</td><td>223.5</td></tr></table>
+
+Table 11: The comparison of the average number of learnable parameters (Avg Para Num) and average running time (Avg Time) between ACM and DMEA.
+
+optimal model architecture, we analyze the model expansion stage of ACM and DMEA using sequence #4 in random scenario. For the final task tv, ACM decides to reuse modules from the first  $(e2e)$  and the third task (laptop) while DMEA reuses all modules from laptop which is consistent with the observation that the similarity between tv and laptop is much higher than that between tv and  $e2e$ .
+
+# A.10 Case Study of Model Output
+
+We select RNNLGhotel (sequence #1 in similar scenario) and WikiSQL (sequence #4 in random scenario) as two representative tasks and show several examples of output in Table 12. Compared with ACM, DMEA possesses the capability to convey more precise and relevant information from the input without introducing superfluous details.
+
+<table><tr><td colspan="2">RNNLGhotel: inform(name=&#x27;mandarin oriental san francisco&#x27;;dogsallowed=&#x27;yes&#x27;;pricerange=&#x27;ultra high end&#x27;)</td></tr><tr><td>Reference</td><td>the mandarin oriental san francisco is in the ultra high end price range and allows dogs.</td></tr><tr><td>ACM</td><td>the mandarin oriental san francisco is a hotel in the ultra high end range (missing: and allows dogs).</td></tr><tr><td>DMEA</td><td>the mandarin oriental san francisco offers ultra high end accommodations and allows dogs.</td></tr><tr><td colspan="2">WikiSQL: on which date was the winning driver alain prost and had damon hill in the pole position?</td></tr><tr><td>Reference</td><td>select date from table where winning driver = alain prost and pole position = damon hill</td></tr><tr><td>ACM</td><td>select date from table where winning driver = alain prost and pole position = damon (missing: hill)</td></tr><tr><td>DMEA</td><td>select date from table where pole position = damon hill and winning driver = alain prost</td></tr></table>
+
+Table 12: Output examples of different methods after learning the whole sequence. We color missing/wrong information in red and redundant information in blue.  
+
+<table><tr><td>Method</td><td>Similar Tasks</td><td>Random Tasks</td></tr><tr><td>ACM</td><td>64.5</td><td>66.2</td></tr><tr><td>ACM w DI</td><td>64.7</td><td>66.5</td></tr></table>
+
+Table 13: The performance comparison between ACM and ACM with dynamic initialization (ACM w DI).
+
+# A.11 Case Study of Dynamic Gradient Scaling
+
+The ablation study in §6.2 demonstrates the importance of dynamic gradient scaling. We further conduct a case study using sequence #1 in random scenario. During the learning of this sequence, the fourth task  $res$  reuses several modules from the third task  $e2e$ . After applying dynamic gradient scaling, the performance of  $e2e$  is improved by 0.3 without compromising  $res$ , indicating that it does mitigate the bias towards the new task.
+
+# A.12 Case Study of Task Selection
+
+To verify that the previous task chosen in the selection stage is indeed the most similar to the new task, we analyze several cases using sequence #2 in random scenario. For the third task hotel, the selected first task  $e2e$  has the highest similarity score as they share the same task type. In addition, the third task hotel shares a similar semantic space with the final task res. Therefore, it is selected for forward knowledge transfer when learning res.
+
+# A.13 Generalization of Dynamic Initialization
+
+To demonstrate the generalization ability of dynamic initialization, we apply it to the expansion stage of ACM. For each scenario, we randomly pick one sequence for experiments. As reported in Table 13, dynamic initialization does benefit ACM, verifying its generalization capability.
+
+# A.14 Real World Application
+
+Apart from the aforementioned sequence generation tasks, DMEA demonstrates the potential to be applied to various real-world lifelong learning scenarios. For example, it can continually train a model to perform summarization and question-answering based on news articles from different domains during the onset of an emerging event like Covid-19.
+
+# A.15 Hyperparameter Search
+
+We select the number of training epochs before modules selection from  $\{6,9,12\}$ , the number  $(n)$  of samples picked to obtain the input subspace from  $\{50,100,200,500\}$  and the threshold  $\epsilon$  for selecting left-singular vectors from  $\{0.90,0.95,0.99\}$ . The number of similar previous tasks  $K$  is selected from  $\{1,2,3\}$ . The number  $(q)$  of examples for calculating the gradient in dynamic gradient scaling is selected from  $\{20,50,100,200\}$ .

2023/Lifelong Sequence Generation with Dynamic Module Expansion and Adaptation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48de539d71527037e31b068256ee6fad0d02f00ad14a7d95b37d395a4dfaf81a
+size 577842

2023/Linear-Time Modeling of Linguistic Structure_ An Order-Theoretic Perspective/88551d68-b602-471f-9566-62b6a0746f09_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4220280cb71c24a21d8b77ba924601ee4c7fd186b910d6555725f053aaaae082
+size 890629

2023/Linear-Time Modeling of Linguistic Structure_ An Order-Theoretic Perspective/full.md

@@ -0,0 +1,717 @@
+# Linear-Time Modeling of Linguistic Structure: An Order-Theoretic Perspective
+
+Tianyu Liu
+
+Afra Amini
+
+Mrinmaya Sachan
+
+Ryan Cotterell
+
+{tianyu.liu, afra.amini, mrinmaya.sachan, ryan.cotterell}@inf.ethz.ch
+
+ETH zürich
+
+# Abstract
+
+Tasks that model the relation between pairs of tokens in a string are a vital part of understanding natural language. Such tasks, in general, require exhaustive pair-wise comparisons of tokens, thus having a quadratic runtime complexity in the length of the string. We show that these exhaustive comparisons can be avoided, and, moreover, the complexity of such tasks can be reduced to linear by casting the relation between tokens as a partial order over the string. Our method predicts real numbers for each token in a string in parallel and sorts the tokens accordingly, resulting in total orders of the tokens in the string. Each total order implies a set of arcs oriented from smaller to greater tokens, sorted by their predicted numbers. The intersection of total orders results in a partial order over the set of tokens in the string, which is then decoded into a directed graph representing the desired linguistic structure. Our experiments on dependency parsing and coreference resolution show that our method achieves state-of-the-art or comparable performance. Moreover, the linear complexity and parallelism of our method double the speed of graph-based coreference resolution models, and bring a 10-times speed-up over graph-based dependency parsers.
+
+![](images/27daac235ca7c34d82906cf327ffb42d6e51d47b928757a3eb80a52e3c940f51.jpg)
+
+https://github.com/lyutyuh/partial
+
+# 1 Introduction
+
+Strings of tokens in natural language are not constructed arbitrarily. Indeed, which tokens co-occur within the same string is highly structured according to the rules of the language. Understanding such structures is critical to the comprehension of natural language. In natural language processing (NLP), many structured prediction tasks aim to automatically extract the underlying structure that dictates the relationship between the tokens in a string of text. Examples of such tasks include
+
+dependency parsing, semantic parsing, and coreference resolution. These tasks involve predicting complex and hierarchical output structures, making them inherently more challenging than their classification or regression counterparts. This paper contributes a novel and generic framework for structured prediction with empirical evidence from dependency parsing and coreference resolution.
+
+Many machine learning models for structured prediction score and predict graphs (McDonald et al., 2005; McDonald and Pereira, 2006), in which the vertices represent the tokens in the string and the edges represent the relations between them. One common strategy to model a graph is to decompose it into smaller subgraphs that are tractable (Taskar et al., 2004; Smith, 2011, §2.2). For example, arc-factored models (Eisner, 1996) score a graph only using the score of each constituent edge. However, even with such simplification, the computational costs of arc-factored models are superlinear. The reason is that one needs to exhaustively compute scores for all possible edges in the graph, which, in general, requires at least quadratic number of computations with respect to the length of the string. Another common strategy employs weighted transition-based systems (Knuth, 1965; Yamada and Matsumoto, 2003; Nivre, 2003). They decompose structures into transitions between intermediate model states and  $do$  offer linear-time algorithms. However, in general, predicting the transitions between states cannot be parallelized, which is another worrying limitation. The authors of this paper contend the limitations of both graph-based and transition-based models are frustrating in an era when researchers are processing longer and longer texts (Tay et al., 2021).
+
+From a more abstract perspective, the mathematical and algorithmic foundation on which structured prediction models rest can be regarded as a design choice. Graph-based and transition-based modeling are both specific design choices. These design
+
+![](images/fc803244064119bd2efc7c1412a6c2b2a50418aa06408fcf1419e7e257d76226.jpg)  
+(a) An example dependency structure. The root arc and arc labels are omitted.
+
+![](images/459a79d5900f58c8f5d603d83eadddbdcb9132ed65a9cc6d135465b205fb23c1.jpg)  
+(b) The token-split structure of Fig. 1a, which is a partially ordered set.  
+(c) A realizer of Fig. 1b with 2 total orders such that  $E = E_{1} \cap E_{2}$ .  $E_{1}$  and  $E_{2}$  contain the arcs oriented from  $V^{r}$  (red nodes) to  $V^{b}$  (blue nodes) and from left to right.  
+Figure 1: An overview of our method. To model a linguistic structure, represented as a directed graph in Fig. 1a, we first convert it into a token-split structure (see §3.4) in Fig. 1b, which is a partial order, to remove undesired transitivity. Then, 2 real numbers are predicted for each vertex in Fig. 1b. The positions of vertices in Fig. 1c in the inequalities indicate the real numbers the vertices are mapped to. The vertices are sorted twice accordingly, resulting in a realizer (see Def. 3.8) of 2 total orderings, each possessing a set of edges  $E_1$  and  $E_2$ . The exact set of desired edges in the original structure can be restored from the intersection of  $E_1$  and  $E_2$  (see §3.3). Some qualitative examples are included in App. J.
+
+$$
+E _ {1} = \left\{\left(b, a ^ {\prime}\right), \left(b, d ^ {\prime}\right), \left(b, c ^ {\prime}\right), \left(d, c ^ {\prime}\right) \right\}
+$$
+
+$$
+\mathcal {T} _ {1}: \begin{array}{c} \underline {{b}} ^ {\prime} \\ \text {r e a d} ^ {\prime} \end{array} <   \begin{array}{c} \underline {{b}} \\ \text {r e a d} \end{array} <   \begin{array}{c} \underline {{a}} ^ {\prime} \\ \text {t h e y} ^ {\prime} \end{array} <   \begin{array}{c} \underline {{d}} ^ {\prime} \\ \text {p a p e r s} ^ {\prime} \end{array} <   \begin{array}{c} \underline {{d}} \\ \text {p a p e r s} \end{array} <   \begin{array}{c} \underline {{c}} ^ {\prime} \\ \text {c o o l} ^ {\prime} \end{array} <   \begin{array}{c} \underline {{a}} \\ \text {t h e y} \end{array} <   \begin{array}{c} \underline {{c}} \\ \text {c o o l} \end{array}
+$$
+
+$$
+\mathcal {T} _ {2}: \begin{array}{c} \underline {{b}} ^ {\prime} <   \underline {{d}} <   \underline {{c}} ^ {\prime} <   \underline {{b}} <   \underline {{a}} ^ {\prime} <   \underline {{d}} ^ {\prime} <   \underline {{c}} <   \underline {{a}} \\ \text {r e a d} ^ {\prime} \text {p a p e r s} \quad \text {c o o l} ^ {\prime} \quad \text {r e a d} \quad \text {t h e y} ^ {\prime} \quad \text {p a p e r s} ^ {\prime} \quad \text {c o o l} \quad \text {t h e y} \end{array}
+$$
+
+$$
+E _ {2} = \left\{\left(d, c ^ {\prime}\right), \left(d, a ^ {\prime}\right), \left(d, d ^ {\prime}\right), \left(b, a ^ {\prime}\right), \left(b, d ^ {\prime}\right) \right\}
+$$
+
+choices impose substantial inductive biases by confining the class of models available to be utilized to solve the task and set limits on the efficiency of the models. In this paper, we propose a fresh design choice for structured prediction. Specifically, we propose an order-theoretic perspective to understand and model structures in NLP. Our approach can predict many structures in natural language in  $\mathcal{O}(N)$  time where  $N$  is the length of the string and is easily parallelizable. The linear-time complexity means our method avoids comparing all  $\mathcal{O}\big(N^2\big)$  token pairs. The key innovation that enables this speed-up is the following: Rather than considering structures as graphs, we view them as partial orderings of the tokens in the strings.
+
+Concretely, we treat structured prediction as a regression task. Because the set of real numbers  $\mathbb{R}$  is naturally ordered by  $<$ , we use real numbers as the proxy for determining the partial order. We predict  $K$  numbers for each token and sort the tokens  $K$  times accordingly. Two tokens are partially ordered by  $\prec$  if and only if they are ordered by  $<$  in all of the  $K$  orders above. We further provide an efficiency guarantee based on the well-established result in order theory that partial orders satisfying particular conditions can be represented as the intersection of as few as  $K = 2$  total orders. We show that most structures in natural language, including trees, alignments, and set partitions, satisfy these conditions. This result enables us to develop a linear-time algorithm for predicting such structures. Fig. 1 gives an illustrative example of our framework applied to dependency parsing, in which the structure being modeled is a tree.
+
+On dependency parsing, our experimental results
+
+show that our method achieves 96.1 labeled attachment score (LAS) and 97.1 unlabeled attachment score (UAS) by using an intersection of only 2 total orders, 96.4 LAS and 97.4 UAS using an intersection of 4 total orders on the English Penn Treebank (Marcus et al., 1993). Furthermore, our method sets the new state of the art on Universal Dependencies 2.2 (Nivre et al., 2018), while being 10 times faster and more memory efficient than graph-based models. Our method also achieves 79.2 F1 score with only 4 total orders on the English OntoNotes coreference resolution benchmark (Pradhan et al., 2012), which is on par with the state of the art, while being twice as fast and using less memory.
+
+# 2 Motivation
+
+We now provide high-level motivation for order-theoretic structured prediction.
+
+# 2.1 Linearization of Structure
+
+The NLP literature abounds with linear-time structured prediction models. Many are derived from the classical shift--reduce parsers (Knuth, 1965) from the compiler literature. One recent line of research has derived linear-time parsers by reducing parsing to tagging (Gómez-Rodríguez and Vilares, 2018; Strzyz et al., 2020; Kitaev and Klein, 2020; Amini et al., 2023, inter alia). In these methods, a finite set of tags  $\mathcal{C}$  is chosen such that all structures for parsing a string can be embedded in  $\mathcal{C}^N$  for a string of length  $N$ . Tagging-based parsers often yield strong empirical performance in both constituency parsing and projective dependency parsing. A natural question is, then, why do we need another method?
+
+We give two motivations. The first linguistic
+
+and the second mathematical. Linguistically, the underlying structures of natural language, e.g., syntax, semantics, and discourse, are often not aligned with the surface form of a sequence due to the existence of displacement (Chomsky, 2015, Chapter 1, p. 44). The strong performance of parsing-as-tagging schemes relies, in part, on there being a tight correspondence between the surface string and structure (Amini and Cotterell, 2022, Proposition 1). Mathematically, the maximum number of structures that a discrete tag sequence can represent is at most  $\mathcal{O}\big(|\mathcal{C}|^N\big)$ . This set is simply not large enough to capture many structures of interest in NLP. For instance, the space of non-projective dependency trees of  $N$  tokens has a cardinality of  $\mathcal{O}\big(N^{N - 2}\big)$  (Cayley, 1889). Therefore, to parse non-projective dependency trees with tagging, the size of the tag set has to grow with  $N$ . However, this implies performing a classification task with an infinite number of classes.
+
+# 2.2 An Illuminating Example
+
+Order-theoretic approaches appear across computer science. For instance, it is well-known that a binary tree can be uniquely restored from its inorder traversal and either the pre- or postorder traversal. Consider the following binary tree.
+
+Example 2.1 (Binary Tree).
+
+![](images/f97df7ce464019c2f375a948ca4e80f9f5bb51349a4e77ffb7b149f1d1f3c03f.jpg)  
+Inorder:abcdefg
+
+![](images/33afaefc14d76428a532b8c70759cf57092417a978f48e74c7d39beb2b604b27.jpg)  
+Postorder: acbegfd  
+Figure 2: An example binary tree and a partial order over the vertices induced by two total orders.
+
+In a binary tree, a vertex  $x$  is a left descendant of vertex  $y$  if and only if  $x$  is visited before  $y$  in both of the in- and postorder traversal. E.g., in Ex. 2.1,  $a$  is the left descendant of  $d$  and is visited before  $d$  in both the in- and postorder traversal.
+
+Another way of stating the above fact is that a binary tree can be recovered from the combination of two total orders, the one induced by the inorder traversal and the one induced by the postorder traversal. Combining these two total orders yields a partial order, i.e., left descendant, from which the left child of each vertex can be identified. This partial order is shown on the right of Ex. 2.1. See App. B and (Knuth, 1997, §2.3.1, Ex. 7) for further discussion. In light of these observations, we conceive an order-theoretic treatment that constructs
+
+a tree by predicting multiple total orders and intersecting them. In terms of computation, predicting total orders only requires labeling each node with real numbers and then sorting, the complexity of which is linear under radix sort. On the other hand, an arc-factored model necessarily computes all  $\mathcal{O}\big(N^2\big)$  pair-wise scores for every pair of vertices to decide the existence of each edge.
+
+Next, we generalize the intuitions gained from this example. In §3, we explore the class of graphs that can be efficiently represented with partial orders. In §4, we show how to learn the ordering efficiently with neural networks.
+
+# 3 Order and Structure
+
+In this section, we describe an order-theoretic treatment for linguistic structure prediction. Specifically, we treat the structure to be predicted as a partially ordered set, i.e., a set equipped with a transitive relation  $\prec$ . We begin by revisiting how linguistic structures are represented as graphs.
+
+# 3.1 Linguistic Structures as Directed Graphs
+
+Let  $\Sigma$  be an alphabet, i.e., a finite set of natural language tokens, and let  $w = w_{1}w_{2}\dots w_{N}\in \Sigma^{*}$  be a string. Linguistic structure prediction is the task of assigning a structure, e.g., a dependency tree, to a given string  $w$  in natural language.
+
+A wide range of linguistic structures are built upon the relations between pairs of tokens. Many structured prediction models are thus arc-factored, i.e., they predict the arcs between a pair of tokens and then combine them back into structures, which are our focus in this work. Formally, their major goal is to model the homogeneous relation<sup>1</sup> on the spanning node set  $V = \{w_{1}, w_{2}, \dots, w_{N}\}$  of a sentence  $w = w_{1} \cdots w_{N}$  (Kübler et al., 2009). The output space is defined by the input itself, in contrast to the external label spaces in other tasks such as classification or language generation.
+
+Definition 3.1 (Structure). A structure over a string  $\boldsymbol{w} = w_{1}w_{2}\dots w_{N}$  is a directed graph  $\mathcal{G} = (V,E)$ , where  $V = \{w_{1},w_{2},\dots ,w_{N}\}$ ,  $E\subseteq V\times V$  is the set of arcs. A typed structure  $\mathcal{G} = (V,E,R)$  is a structure with  $E\subseteq V\times V\times R$ , where  $R$  is a finite set of relation labels.
+
+Most linguistic structures are naturally subsumed under this definition. We give two examples of linguistic structure prediction tasks.
+
+Example 3.2 (Dependency Parsing; Kübler et al., 2009, Def. 2.3). A dependency structure is a structure  $\mathcal{G} = (V,E,R)$ , where  $E\subseteq V\times V\times R$  and  $R$  is the set of dependency relation types. If  $(x,y,r)\in E$ , then  $\forall r^{\prime}\neq r,(x,y,r^{\prime})\notin E$
+
+Example 3.3 (Coreference Resolution). A coref- erence structure is a structure  $\mathcal{G} = (V,E,R)$  where  $E\subseteq V\times V\times R$  , and  $R = \{r,r^{\prime}\}$  . The rela-tions  $r,r^{\prime}$  represent the entity mention and coreference, respectively. We have  $(x,y,r)\in E$  if and only if the textual span  $x:y$  in  $\pmb{w}$  is a mention of an entity.  $(x_{1},x_{2},r^{\prime})\in E\land (y_{1},y_{2},r^{\prime})\in E$  if and only if the textual spans  $x_{1}:y_{1}$  and  $x_{2}:y_{2}$  corefer.
+
+# 3.2 From Directed Graphs to Partial Orders
+
+Our treatment constructs linguistic structures with techniques from order theory. The key is to cast the relation between tokens as an order, which is defined as follows.
+
+Definition 3.4 (Order; Hausdorff, 1914). An order over a set  $V$  is a relation  $\prec$  such that the following hold for all  $x, y, z \in V$ :
+
+(a)irreflexivity:  $x\neq x$  
+(b) asymmetry:  $x \prec y \implies y \not\prec x$ ;  
+(c) transitivity:  $x \prec y \land y \prec z \implies x \prec z$ .
+
+Natural language exhibits structural sparsity in that each token in a string usually only interacts with very few other tokens with a particular relation. For instance, in a dependency graph, there are no direct paths between most of the word pairs. Such sparsity, from an order-theoretic point of view, can be characterized by incomparability in a partially ordered set (Birkhoff, 1967, Chapter 1, p. 2).
+
+By analogy, we define the following partially ordered structure, which is a partially ordered set mathematically. Its elements are the tokens of a string, and its order encodes a linguistic structure.
+
+Definition 3.5 (Partially Ordered Structure). Let  $\mathcal{G} = (V,E)$  be a structure. Define the following relation  $\prec$ : For  $x,y\in V$ ,  $x\prec y \iff (x,y)\in E$ . We call  $\mathcal{P} = (V,E,\prec)$  a partially ordered structure if  $\prec$  satisfies Def. 3.4.
+
+The essential theoretical foundation of our linguistic structure prediction framework is the classic result that partial orders can be represented by an intersection of total orders (Dushnik and Miller, 1941). It is this result that enables us to use
+
+real numbers as a proxy to determine the partial ordering of tokens.
+
+Definition 3.6 (Totally Ordered Structure). A partially ordered structure  $\mathcal{P} = (V, E, \prec)$  is totally ordered if  $\forall x, y \in V: x \prec y \lor y \prec x$ .
+
+Due to the transitivity of the ordering relation  $\prec$ , a totally ordered structure of  $|V|$  elements always contains  $|E| = \binom{|V|}{2}$  relations. Given a collection of structures  $\{(V, E_k)\}_{k \in [K]}$  defined over the same set of vertices  $V$ , their intersection is also a structure—namely  $(V, \cap_{k \in [K]} E_k)$ , where  $K \in \mathbb{N}, [K] \stackrel{\text{def}}{=} \{1, \dots, K\}$ . The intersection of partially ordered structures remains partially ordered.
+
+We now cite a famous theorem from order theory.
+
+Theorem 3.7 (Szpilrajn (1930)). Every partially ordered structure is contained in a totally ordered structure, i.e., for every partially ordered structure  $\mathcal{P} = (V,E,\prec)$ , there exists a totally ordered structure  $\mathcal{T} = (V,\widehat{E},\prec)$  such that  $E\subseteq \widehat{E}$ .
+
+Thm. 3.7 ensures that every partially ordered structure can be embedded in some totally ordered structure in the sense that the totally ordered structure contains all the relations in the partially ordered structure. More importantly, a stronger result can be shown: Partially ordered structures can always be represented as intersections of a collection of totally ordered structures.
+
+Definition 3.8 (Realizer). Let  $\mathcal{P} = (V,E,\prec)$  be a partially ordered structure. A realizer  $R_{\mathcal{P}}$  of  $\mathcal{P}$  is a set of totally ordered structures  $\left\{\mathcal{T}_1,\mathcal{T}_2,\dots ,\mathcal{T}_K\right\}$  over  $V$ , i.e., each  $\mathcal{T}_k = (V,E_k,\prec_k)$ , such that  $E = \bigcap_{k\in [K]}E_k$ . In other words,  $\forall x,y\in V,x\prec y\iff \bigwedge_{k\in [K]}x\prec_ky$ .
+
+Theorem 3.9 (Dushnik and Miller, 1941, Thm. 2.32). There exists a realizer  $R_{\mathcal{P}}$  for every partially ordered structure  $\mathcal{P} = (V,E,\prec)$
+
+A corollary of the above theorem is that the complexity of a partially ordered structure can be characterized by its order dimension, which is defined as follows.
+
+Definition 3.10 (Order Dimension; Dushnik and Miller, 1941). Let  $\mathcal{P} = (V,E,\prec)$  be a partially ordered structure. The order dimension  $D_{\mathcal{P}}$  of  $\mathcal{P}$  is the cardinality of the smallest realizer of  $\mathcal{P}$ .
+
+# 3.3 Efficiency Guarantees
+
+In this section, we give an efficiency guarantee of order-theoretic structured prediction. These efficiency guarantees come from a series of results in
+
+order theory and lattice theory (Dushnik and Miller, 1941; Hiraguchi, 1955; Birkhoff, 1967, inter alia).
+
+First, it is important to note that not all partially ordered structures can be represented as an intersection of a constant number of totally ordered structures (Dushnik and Miller, 1941, Thm. 4.1).
+
+In fact, testing whether the order dimension of a partial order  $\mathcal{P}$  is at most  $K$ ,  $\forall K \geq 3$  is NP-complete (Yannakakis, 1982). However, we contend that most of the linguistic structures found in natural language processing (Smith, 2011)—including trees, equivalence classes (i.e., set partitioning), and alignment (i.e., bipartite matching)—can be represented as the intersection of 2 totally ordered structures. We postulate that this is possible due to their innate sparsity, i.e., a token tends to only interact with a few other tokens. These assumptions are formalized as follows.
+
+Assumption 3.11 (Sparsity). A class of linguistic structures  $\mathcal{G} = (V,E)$  over natural language strings  $\boldsymbol {w}\in \Sigma^{*}$  with  $N = |\pmb {w}|$  is called sparse if  $\mathcal{O}(|E|) = \mathcal{O}(N)$
+
+Assumption 3.12 (Linguistic Structures are 2-dimensional). Structures in natural language can be represented as intersections of 2 totally ordered structures.
+
+We justify Assumptions 3.11-3.12 in App. D. Empirical evidence is also provided in §5, where 2-dimensional order-theoretic models are trained to tackle two linguistic structure prediction tasks with high performance.
+
+# 3.4 Token-Split Structures
+
+An obvious limitation of our formulation of linguistic structures as partial orders is that by Def. 3.4, partial order is transitive. In other words,  $x \prec y \land y \prec z$  implies  $x \prec z$ , which, however, does not hold in the structures characterized by the directed graph formalization in Def. 3.1. In addition, we note that our notation of structures generalizes to cyclic graphs. However, partially ordered structures are inherently acyclic due to the transitivity of  $\prec$ . We now introduce the token-split structure, which enables cycles and removes redundant edges introduced by transitivity in partially ordered structures.
+
+Definition 3.13 (Token-Split Structure). A token-split structure induced by a structure  $\mathcal{G} = (V,E)$  is a structure  $\mathcal{P} = (\widehat{V},\widehat{E},\prec)$  such that
+
+(a)  $\widehat{V} \stackrel{\mathrm{def}}{=} V^r \cup V^b$ , where  $V^r = \{x^r \mid x \in V\}$ ,  $V^b = \{x^b \mid x \in V\}$ ;
+
+(b)  $V^r\cap V^b = \varnothing$  
+(c)  $\widehat{E} = \left\{(x^r, y^b) \mid (x, y) \in E\right\}$ .
+
+In other words, a token-split structure maps the edges from the original structure, including self-loops, into a bipartite graph in which the edges are oriented from  $V^r$  to  $V^b$ . An example is displayed in Fig. 1b.
+
+Given a token-split structure  $\mathcal{P} = (\widehat{V},\widehat{E},\prec)$ , we can recover the original structure  $\mathcal{G} = (V,E)$  from which  $\mathcal{P}$  is induced using the following equation
+
+$$
+E = \left\{\left(x, y\right) \mid x ^ {r} \in V ^ {r} \wedge y ^ {b} \in V ^ {b} \wedge x ^ {r} \prec y ^ {b} \right\} \quad (1)
+$$
+
+Theorem 3.14. Token-split structures are partially ordered.
+
+Proof. See App. C.1.
+
+Remark 3.15 (Conversion between Structures and Partially Ordered Structures). Thm. 3.14 and Eq. (1) ensure that we can convert back and forth between any structure under Def. 3.1 and a partially ordered structure. Specifically, they enable us to first convert a structure to a partially ordered structure, predict it order-theoretically, and then finally convert it back to a structure.
+
+# 4 A Neural Parameterization
+
+In this section, we describe the core technical contribution of our work. We show how to model partially ordered structures with a neural model. Specifically, we define a parameterized realizer of Def. 3.8 and an objective function for training the realizer to model the token-split structures. We also give algorithms for efficient training and decoding.
+
+# 4.1 Neuralized Total Order
+
+We now discuss a parameterized neural network that induces partial orders as the intersection of several total orders.
+
+Definition 4.1 (Functional Realizer). A functional realizer of a partially ordered structure  $\mathcal{P} = (V,E,\prec)$  is a set of mappings  $\mathcal{F}_{\theta} = \{f_{\theta}^{(1)},\dots ,f_{\theta}^{(K)}\}$ , where  $\pmb{\theta}$  is the set of learnable parameters shared among  $f_{\theta}^{(k)}$ , and the order dimension  $K\in \mathbb{N}$  is a hyperparameter of the realizer. The realize element  $f_{\theta}^{(k)}:V\to \mathbb{R},\forall k\in [K]$  maps each vertex in the input structure to a real number. We overload  $\mathcal{F}_{\theta}$  as a mapping  $\mathcal{F}_{\theta}:V\to \mathbb{R}^{K}$ , defined as  $\mathcal{F}_{\theta}(x)\stackrel {\mathrm{def}}{=}\left[f_{\theta}^{(1)}(x),\dots ,f_{\theta}^{(K)}(x)\right]^{\top}$ .
+
+The set of real numbers  $\mathbb{R}$  is totally ordered, in which the order is given by the  $<$  (less than) relation. Each individual  $f_{\theta}^{(k)} \in \mathcal{F}_{\theta}$  induces a total order  $\mathcal{T}_k = \left(V, \{(x,y) \mid x,y \in V, f_{\theta}^{(k)}(x) < f_{\theta}^{(k)}(y)\}, \prec_k\right)$ .
+
+The functional realizer assigns  $K$  total orders  $\{\mathcal{T}_1,\mathcal{T}_2,\dots ,\mathcal{T}_K\}$  to the input string. During decoding, an edge from  $x$  to  $y$  exists in  $\mathcal{P}$  if and only if  $x\prec_{k}y$  holds in  $\mathcal{T}_k,\forall k\in [K]$
+
+Implementing Def. 4.1 with neural networks is straightforward. To obtain  $f_{\theta}^{(k)}(x^r)$  and  $f_{\theta}^{(k)}(x^b)$ , where  $x^r, x^b$  are two vertices introduced by the token-split formulation (Def. 3.13) corresponding to the same token  $w_x$  in the input, we apply two linear projections on the contextualized representation of  $x$  given by a pretrained model parameterized by  $\theta$ . In total,  $2K$  real numbers are predicted for each input token.
+
+# 4.2 Learning a Functional Realizer
+
+To learn the functional realizers with a gradient-based procedure, we need a differentiable objective. In a partially ordered structure  $\mathcal{P}$  with functional realizer  $\mathcal{F}_{\theta} = \{f_{\theta}^{(1)}, f_{\theta}^{(2)}, \dots, f_{\theta}^{(K)}\}$ , we have  $x \prec y$  if and only if  $\bigwedge_{k \in [K]} \left(f_{\theta}^{(k)}(x) < f_{\theta}^{(k)}(y)\right)$ . We re-express this condition as follows:
+
+$$
+F _ {\boldsymbol {\theta}} (x, y) \stackrel {\text {d e f}} {=} \max  _ {k \in [ K ]} \left(f _ {\boldsymbol {\theta}} ^ {(k)} (x) - f _ {\boldsymbol {\theta}} ^ {(k)} (y)\right) <   0 \tag {2}
+$$
+
+We call  $F_{\theta}$  a pair-wise function. On the other hand, we have  $x \not\prec y$  if and only if  $\bigvee_{k \in [K]} \left(f_{\theta}^{(k)}(x) \geq f_{\theta}^{(k)}(y)\right)$ . This condition can be re-expressed as  $F_{\theta}(x,y) \geq 0$ . Thus, empirically, the smaller  $F_{\theta}(x,y)$  is, the more likely the relation  $x \not\prec y$  exists.
+
+We now define a training objective, which encourages the model to make decisions that comply with the order constraints enforced by the structures, described by Eq. (2). Given the token-split version  $\mathcal{P} = (V,E,\prec)$  induced by the structure being modeled, we consider the following objective
+
+$$
+\begin{array}{c} \mathcal {L} (\boldsymbol {\theta}) = \log \sum_ {(x, y) \in V ^ {2} \backslash E} \exp - F _ {\boldsymbol {\theta}} (x, y) + \\ \log \sum_ {(x, y) \in E} \exp F _ {\boldsymbol {\theta}} (x, y) \end{array} \tag {3}
+$$
+
+The first term maximizes  $F_{\theta}(x,y)$  for  $x \neq y$ , while the second minimizes  $F_{\theta}(x,y)$  for  $x \prec y$ . Note that in the second term, we assume  $\mathcal{O}(|E|) = \mathcal{O}(N)$  in a linguistic structure following Assumption 3.11.
+
+# 4.3 An Efficient Algorithm
+
+We remark that both training and decoding of the proposed model can be regarded as performing an aggregation for every token  $x \in V$ .
+
+Definition 4.2 (Aggregation). An  $\oplus$ -aggregation given a token  $x$  for a pair-wise function  $F_{\theta}$  over the set  $V$  is an operation  $\bigoplus_{y \in V} F_{\theta}(x, y)$ , where  $\oplus$  is a commutative and associative operation over which real number addition  $+$  is distributive.
+
+Aggregation is a common abstraction for computing the relation between a token  $x$  and every other token. The aggregation operator is associative and commutative, thus can be computed in parallel. The number of required computations is  $\mathcal{O}(|V|)$  for naively computing an aggregation of  $x$ .
+
+During training, we  $\oplus$ -aggregate using negative log-sum-exp, i.e., we compute  $-\log \sum_{y} \exp(-F_{\theta}(x, y))$  for all  $x$ , to compute the first term of Eq. (3). In greedy decoding, we  $\oplus$ -aggregate by computing  $\min_{y} F_{\theta}(x, y)$  to find the optimal relation arc from each  $x$ . Naively,  $\oplus$ -aggregating for every token  $x \in V$  takes  $\mathcal{O}(N^2)$  in total, as each aggregate has a complexity of  $\mathcal{O}(N)$ . However, the partial order we assigned over  $V$  allows us to efficiently compute the aggregands.
+
+For  $K = 2$ , we can inspect Eq. (2) to see that  $F_{\pmb{\theta}}(x,y)$  is equal to either  $f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)$  or  $f_{\pmb{\theta}}^{(2)}(x) - f_{\pmb{\theta}}^{(2)}(y)$ . We now define the following two subsets of  $V$  for  $k \in \{1,2\}$ :
+
+$$
+\mathcal {S} _ {k} (x) = \left\{y \mid F _ {\boldsymbol {\theta}} (x, y) = f _ {\boldsymbol {\theta}} ^ {(k)} (x) - f _ {\boldsymbol {\theta}} ^ {(k)} (y) \right\}
+$$
+
+Using this notation, we can write the following
+
+$$
+\begin{array}{l} \bigoplus_ {(x, y) \in V ^ {2}} F _ {\boldsymbol {\theta}} (x, y) = \bigoplus_ {x \in V} \bigoplus_ {y \in V} F _ {\boldsymbol {\theta}} (x, y) (5a) \\ = \underbrace {\bigoplus_ {x \in V} \bigoplus_ {y \in \mathcal {S} _ {1} (x)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right)} _ {\stackrel {\text {d e f}} {=} G _ {1}} (5b) \\ \oplus \underbrace{\bigoplus_{x\in V}\bigoplus_{y\in\mathcal{S}_{2}(x)}\left(f_{\boldsymbol{\theta}}^{(2)}(x) - f_{\boldsymbol{\theta}}^{(2)}(y)\right)}_{\stackrel{\mathrm{def}}{= G_{2}}} \\ \end{array}
+$$
+
+We now give an efficient algorithm to compute  $G_{1}$  and, by symmetry,  $G_{2}$ . Our first observation is that,
+
+by distributivity, we can write
+
+$$
+\begin{array}{l} G _ {1} = \bigoplus_ {x \in V} \bigoplus_ {y \in \mathcal {S} _ {1} (x)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) (6a) \\ = \bigoplus_ {x \in V} \underbrace {\left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) + \bigoplus_ {y \in \mathcal {S} _ {1} (x)} - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right)} _ {\stackrel {\text {d e f}} {=} G _ {1} (x)} (6b) \\ \end{array}
+$$
+
+Alone, this application of dynamic programming does not reduce the complexity from  $\mathcal{O}\big(N^2\big)$  to  $\mathcal{O}(N)$  as desired because the inner aggregand,  $\bigoplus_{y\in S_1(x)} - f_{\pmb{\theta}}^{(1)}(y)$ , itself still takes  $\mathcal{O}(N)$  time. However, we are able to compute  $\bigoplus_{y\in S_1(x)} - f_{\pmb{\theta}}^{(1)}(y)$  in amortized  $\mathcal{O}(1)$  time due to Fredman's (1976, Eq. 1) algebraic trick.
+
+The strategy is to sort<sup>4</sup> the vertices of the partially ordered structure according to  $f_{\pmb{\theta}}^{(1)}(y) - f_{\pmb{\theta}}^{(2)}(y)$ . Thus, if we have  $f_{\pmb{\theta}}^{(1)}(y) - f_{\pmb{\theta}}^{(2)}(y) < f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(2)}(x)$ , simple algebra reveals that  $f_{\pmb{\theta}}^{(2)}(x) - f_{\pmb{\theta}}^{(2)}(y) < f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)$ . Thus, for a given  $x$ , every vertex  $y$  that comes before  $x$  in the sorted order satisfies  $F_{\pmb{\theta}}(x,y) = f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)$ . Aggregating in this order enables intermediate results to be reused.
+
+Algorithm 1 Computing  $G_{1}$  when  $K = 2$ .  
+1: procedure COMPUTE-G1(f(1),f(2),V)  
+2: U← sort (V,key=f(1)-f(2))  
+3: G1,s1←0,0▷ 0 is the zero element of ⊕  
+4: for n=1 up to N:  
+5: q1=f(1)(Un)+s1▷ q1=G1(Un)  
+6: G1 ⊕= q1  
+7: s1 ⊕=-f(1)(Un)  
+8: return G1
+
+Likewise, if we sorted in reverse, i.e., according to  $f_{\pmb{\theta}}^{(2)}(y) - f_{\pmb{\theta}}^{(1)}(y)$ , the same manipulation yields that for a given  $x$ , every vertex  $y$  that comes before  $x$  in the reverse sorted order satisfies  $F_{\pmb{\theta}}(x,y) = f_{\pmb{\theta}}^{(2)}(x) - f_{\pmb{\theta}}^{(2)}(y)$ .
+
+The algorithm for computing  $G_{1}$  is given in Algorithm 1, which has  $\mathcal{O}(N)$  computations in total. Moreover, if parallelized, it can be run in  $\mathcal{O}(\log N)$  time. For  $K > 2$ , we speculate that the aggregation algorithm can be done in  $\mathcal{O}\bigl (KN\log^{K - 2}N\bigr)$ . We leave this to future work. See App. E.2 for further discussion.
+
+As before, we take the complexity of sorting to be  $\mathcal{O}(N)$  where we can apply radix sort as implemented by Pytorch.
+
+# 5 Experiments
+
+We report the experimental results on two representative linguistic structure prediction problems in NLP, namely dependency parsing and coreference resolution. The graph-theoretic definitions of these tasks are given in Examples 3.2 and 3.3. We first convert the linguistic structures to partially ordered (token-split) structures as described in §3.4, and then apply the neural method described in §4 to model the partially ordered structures.
+
+# 5.1 Dependency Parsing
+
+Modeling. Orders  $\prec$  are not typed in Def. 3.5. In other words, under Def. 3.5, all relations in a partially ordered structure are of the same type. To model dependency type labels, we apply a token-level classifier on the contextualized representation. During decoding, similar to arc-factored models for dependency parsing, we keep the head word that minimizes  $F_{\theta}(x,y)$  for a given  $x$ , i.e.,  $\mathrm{argmin}_{y\in V}F_{\theta}(x,y)$ .
+
+For pretrained language models, we use XLNet-large-cased $^5$  (Yang et al., 2019) for PTB, bert-base-chinese $^6$  for CTB, and bert-base-multilingual-cased $^7$  for UD.
+
+Datasets. We conduct experiments on the English Penn Treebank (PTB; Marcus et al., 1993), the Chinese Penn Treebank (CTB; Xue et al., 2005), and the Universal Dependencies 2.2 (UD; Nivre et al., 2018). Hyperparameter settings and dataset statistics are given in Apps. F.1 and G.1.
+
+Accuracy. We report the experimental results in Tab. 1. The full results on UD are included in App. I.1. On PTB and UD, our method achieves state-of-the-art performance compared with  $\mathcal{O}(N^3)$  (Yang and Tu, 2022),  $\mathcal{O}(N^2)$  (Mrini et al., 2020), and  $\mathcal{O}(N)$  (Amini et al., 2023) methods. Although Amini et al.'s (2023) method has the same complexity as ours, it is worth noting that our method is more general since it can handle non-projective dependencies without using pseudoprojectivization (Nivre and Nilsson, 2005).
+
+Efficiency. We evaluate the efficiency of our method with two representative baseline models. As depicted in Tab. 2, we observe that our method with  $K = 2$  is almost 10 times as fast as Biaff
+
+<sup>5</sup>https://huggingface.co/xlnet-large-cased  
+<sup>6</sup>https://huggingface.co/bert-base-chinese  
+<sup>7</sup>https://huggingface.co/
+
+bert-base-multilingual-cased
+
+<table><tr><td rowspan="2">Model</td><td colspan="2">PTB</td><td colspan="2">CTB</td><td>UD</td></tr><tr><td>UAS</td><td>LAS</td><td>UAS</td><td>LAS</td><td>LAS</td></tr><tr><td>Zhou and Zhao*</td><td>97.0</td><td>95.4</td><td>91.2</td><td>89.2</td><td>-</td></tr><tr><td>Mrini et al.*</td><td>97.4</td><td>96.3</td><td>94.6</td><td>89.3</td><td>-</td></tr><tr><td>Chen and Manning</td><td>91.8</td><td>89.6</td><td>83.9</td><td>82.4</td><td>-</td></tr><tr><td>Dozat and Manning</td><td>95.7</td><td>94.1</td><td>89.3</td><td>88.2</td><td>91.8</td></tr><tr><td>Yang and Tu#</td><td>97.4</td><td>95.8</td><td>93.3</td><td>92.3</td><td>91.9</td></tr><tr><td>Amini et al.</td><td>97.4</td><td>96.4</td><td>93.2</td><td>91.9</td><td>91.8</td></tr><tr><td>Ours (K=2)</td><td>97.1</td><td>96.1</td><td>90.7</td><td>89.5</td><td>91.2</td></tr><tr><td>Ours (K=4)</td><td>97.4</td><td>96.4</td><td>92.4</td><td>91.4</td><td>92.1</td></tr></table>
+
+Table 1: Experimental results on PTB, CTB, and UD. * indicates usage of extra constituency annotation. # is our re-implementation using the same pretrained encoder as ours.  $K$  is the dimension of the realizer used.  
+
+<table><tr><td rowspan="2">#token</td><td colspan="3">Speed (sent/s) ↑</td><td colspan="3">Memory (GB) ↓</td></tr><tr><td>Ours</td><td>Hexa</td><td>Biaff</td><td>Ours</td><td>Hexa</td><td>Biaff</td></tr><tr><td>32</td><td>3232</td><td>2916</td><td>493</td><td>1.7</td><td>2.9</td><td>4.5</td></tr><tr><td>64</td><td>3332</td><td>3011</td><td>328</td><td>1.7</td><td>3.0</td><td>10.1</td></tr><tr><td>128</td><td>3182</td><td>2649</td><td>202</td><td>1.9</td><td>3.7</td><td>30.6</td></tr><tr><td>256</td><td>3314</td><td>3270</td><td>98</td><td>3.1</td><td>4.5</td><td>56.2</td></tr><tr><td>overall</td><td>3347</td><td>3176</td><td>338</td><td>1.7</td><td>3.0</td><td>10.6</td></tr></table>
+
+(Dozat and Manning, 2017), and consumes less memory than Hexa (Amini et al., 2023), which is  $\mathcal{O}(N)$  in space complexity. We further include some qualitative examples using  $K = 2$  in App. J for a more intuitive picture of our method.
+
+# 5.2 Coreference Resolution
+
+Modeling. Our method operates in a two-stage manner to accommodate the two relations in Ex. 3.3. First, it extracts a list of entity mentions using the partial order induced by  $r$  (mention relation). In other words,  $x \prec y \iff \text{span } x : y$  is an entity mention. Then, it models the partial order induced by  $r'$  (coreference relation) over the extracted mentions. In other words,  $m_1 \prec m_2 \iff \text{mention } m_1 \text{ corefers to } m_2$ . To find the optimal coreferent antecedent for each mention  $m$ , we keep  $m'$  that minimizes  $F_{\theta}(m, m')$ .
+
+The overall complexity of the coreference resolution model is  $\mathcal{O}(N)$ , since the complexity of the encoder used (Beltagy et al., 2020) and the number of valid mentions are both  $\mathcal{O}(N)$ ,
+
+assuming entity mentions are constituents (Liu et al., 2022). We experiment on the CoNLL-2012 English shared task dataset (OntoNotes; Pradhan et al., 2012). Hyperparameter settings and dataset statistics are given in Apps. F.2 and G.2.
+
+Accuracy. The experimental results are displayed in Tab. 3. Similar to the results for dependency parsing, an intersection of 2 total orders can already achieve reasonable performance on coreference resolution. This provides empirical evidence for our assertion in §3.3 that most structures in NLP can be represented as the intersection of at most 2 total orders. When  $K = 4$ , the performance of our method is comparable to Kirstain et al. (2021), which uses the same pretrained encoder as ours and requires an  $\mathcal{O}(N^2)$  biaffine product computation for token-pair scores.
+
+Efficiency. We compare the efficiency of our method with Kirstain et al.'s (2021) method. It is worth noting that Kirstain et al. (2021) has already performed aggressive optimization in both the speed and memory footprint of coreference modeling. Our method is still 2 times as fast, achieving a speed of 82.8 documents per second vs. 41.9, while using less memory, especially on long documents. The full efficiency statistics are given in App. H.
+
+Table 2: Speed and memory consumption comparison on PTB test set. We report results averaged over 3 random runs of our method with  $K = 2$ . The other settings and the results for Hexa and Biaff are taken from Amini et al. (2023, Tab. 3).  
+
+<table><tr><td></td><td>Avg. P</td><td>Avg. R</td><td>Avg. F1</td></tr><tr><td>Lee et al. (2017)</td><td>69.9</td><td>64.7</td><td>67.2</td></tr><tr><td>Kantor and Globerson</td><td>76.1</td><td>77.1</td><td>76.6</td></tr><tr><td>Joshi et al. (2020)</td><td>80.1</td><td>78.9</td><td>79.6</td></tr><tr><td>Xu and Choi (2020)</td><td>80.3</td><td>79.5</td><td>79.9</td></tr><tr><td>Kirstain et al. (2021)</td><td>81.2</td><td>79.4</td><td>80.3</td></tr><tr><td>0urs (K=2)</td><td>75.2</td><td>74.8</td><td>75.0</td></tr><tr><td>0urs (K=4)</td><td>79.3</td><td>79.0</td><td>79.2</td></tr></table>
+
+Table 3: Experimental results on the OntoNotes benchmark.  $K$  is the dimension of the realizer.
+
+# 6 Related Work<sup>8</sup>
+
+# 6.1 Structured Prediction
+
+Structured prediction constitutes an important part of natural language processing. It involves the modeling of interrelated variables or outputs with structural constraints. Some representative structured prediction problems are sequence tagging (Church, 1988), dependency parsing (Kübler et al., 2009), and coreference resolution (Stede, 2012).
+
+Structured prediction can often be formulated as learning and inference of probabilistic graphical models (Smith, 2011, §2.2). The key idea is to represent the probability distribution over the output space using a graph, in which each vertex corresponds to a random variable, and each edge corresponds to a dependence relation between two random variables.
+
+# 6.2 Graph-Based Parsing
+
+Graph-based parsers, or arc-factored parsers, construct graphs by scoring all possible arcs (Eisner, 1996; McDonald and Pereira, 2006) between each pair of words. At inference time, they use either maximum spanning tree (MST) finding algorithms (Chu and Liu, 1965; Edmonds, 1967; Tarjan, 1977), or the projective MST algorithm (Eisner, 1996) to build the valid dependency trees with maximum score. Kiperwasser and Goldberg (2016) present a neural graph-based parser that uses the same kind of attention mechanism as Bahdanau et al. (2015) for computing arc scores. Greedy decoding that independently assigns a head word to each word (Dozat and Manning, 2017) is also widely used as an approximation to exact inference algorithms.
+
+# 6.3 Tagging-Based Parsing
+
+Inspired by transition-based parsers (Knuth, 1965) and Bangalore and Joshi's (1999) seminal work on supertagging, one line of work uses pretrained models to parse dependency trees by inferring tags for each word in the input sequence. Li et al. (2018) and Kiperwasser and Ballesteros (2018) predict the relative position of the dependent with respect to its head in a sequence-to-sequence manner. Strzyz et al. (2019) give a framework for analyzing similar tagging schemes. Gómez-Rodríguez et al. (2020) infer a chunk of actions in a transition-based system for each word in the sequence.
+
+For non-projective dependency parsing, Gómez-Rodríguez and Nivre (2010, 2013) show that efficient parsers exist for 2-planar trees (Yli-Jyvä, 2003), a sub-class of non-projective trees whose arcs can be partitioned into 2 sets while arcs in the same set do not cross each other. Strzyz et al. (2020) propose an encoding scheme for 2-planar trees, enabling a tagging-based parser for such trees. As mentioned in §2.1, to handle the entire set of non-projective trees, the size of the tag set has to be unrestricted, which limits the efficiency and applicability of this series of approaches of approaches.
+
+# 6.4 Parsing with Syntactic Distance
+
+Shen et al. (2018a,b) introduce a constituent parsing scheme which is also based on the comparison of real numbers. In this scheme, a neural model is trained to assign one real number, termed the syntactic distance, to the gap between every pair of neighboring tokens. To parse a span into two sub-constituents, the gap with the largest syntactic distance within that span is chosen as the split point. Parsing can be done by recursively performing the above splitting procedure starting from a given string. The algorithm has a runtime complexity of  $\mathcal{O}(N\log N)$ , which is significantly more efficient than chart-based parsers with  $\mathcal{O}\left(N^{2}\right)$  complexity. However, this method does not generalize easily to perform non-context-free parsing, since it cannot handle the possible discontinuity of constituents. Moreover, the recursive splitting procedure restricts the output space of parse trees to be a subset of phrase-structure trees (Dyer et al., 2019).
+
+# 7 Conclusion
+
+In this paper, we propose an order-theoretic treatment of linguistic structured prediction. Theoretical and empirical results show that most linguistic structure prediction problems can be solved in linear time and memory by framing them as partial orderings of the tokens in the input string. We demonstrate the effectiveness of our method on dependency parsing and coreference resolution, setting the new state-of-the-art accuracy in some cases and achieving significant efficiency improvements.
+
+# 8 Limitations
+
+# 8.1 Decoding Algorithms
+
+This work does not provide algorithms for particular structures or algorithms that ensure the well-formedness of structures, such as maximum spanning trees and projective trees. It remains to be investigated whether existing constrained decoding algorithms for arc-factored models (Chu and Liu, 1965; Edmonds, 1967; Eisner, 1997, inter alia) have their counterparts in the order-theoretic method. We would like to explore decoding algorithms for structured prediction under order-theoretic formulation in future work.
+
+# 8.2 Interpretability
+
+In our method, the interactions between tokens are not directly modeled as in graph-based structured
+
+prediction models, which makes it more difficult to interpret the output of our model. In addition, we leave to future work the investigation of the total ordering metrics (see App. J) learned by the realizers in our method.
+
+# 8.3 Hardness of Learning
+
+Intuitively, it is harder to learn partial orders over strings than directly modeling the arcs in a graph, since our order-theoretic treatment has much fewer parameters when scoring token pairs. We also observed in our experiments that order-theoretic models take more training iterations to converge than arc-factored models.
+
+For instance, consider the modeling of a tree structure with  $N$  nodes with  $N - 1$  arcs using partial order, which implies  $N - 1$  constraints of the form  $x \prec y$  and  $N^2 - 2N + 1$  constraints of  $x \nprec y$ . From a theoretical perspective,  $K = 2$  is sufficient to represent such a structure as shown in §3. In other words, there always exist 2 total orders whose intersection satisfies the aforementioned  $N(N - 1)$  constraints. However, it might not be easy to find such orders in practice.
+
+A realizer with  $K$  beyond 2 can more easily satisfy the constraints, especially of the form  $x \neq y$  since there are more constraints of this form. It allows more possibilities for  $\bigvee_{k \in [K]} f_{\pmb{\theta}}^{(k)}(x) \geq f_{\pmb{\theta}}^{(k)}(y)$  (i.e., more choices of  $k$  to satisfy the expression). On the other hand, using a small  $K$  might make it harder to satisfy the constraints.
+
+We plan to further investigate the hardness of learning a string partial order in future work.
+
+# 8.4 Precision of floating-point numbers and numerical stability
+
+Our method might be affected by the finite precision of floating-point numbers and numerical instability when applied to very long strings. Although we did not encounter such issues in our experiments ( $N \leq 4096 = 2^{12}$ ), issues might arise when  $N > 65536 = 2^{16}$  if bfloat16 or half precision is used. In such extreme cases, our assumption that  $\forall k \in [K]$ ,  $f_{\theta}^{(k)}$  is injective cannot be fulfilled. Thus, not all totally ordered structures of  $N$  elements can be represented, and our method might not exhibit the desired behavior.
+
+# Ethics Statement
+
+We do not believe the work presented here further amplifies biases already present in the datasets and
+
+pretrained models. Therefore, we foresee no ethical concerns in this work.
+
+# Acknowledgments
+
+We would like to thank Zhaofeng Wu, Clément Guerner, and Tim Vieira for their invaluable feedback. We are grateful to the anonymous reviewers for their insightful comments and suggestions. Afra Amini is supported by ETH AI Center doctoral fellowship. MS acknowledges support from the Swiss National Science Foundation (Project No. 197155), a Responsible AI grant by the Haslerstiftung; and an ETH Grant (ETH-19 21-1).
+
+# References
+
+Kazimierz Adjukiewicz. 1935. Die syntaktische Konnexität. Studia Philosophica, 1:1-27.  
+Afra Amini and Ryan Cotterell. 2022. On parsing as tagging. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8884-8900, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Afra Amini, Tianyu Liu, and Ryan Cotterell. 2023. Hexatagging: Projective dependency parsing as tagging. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1453-1464, Toronto, Canada. Association for Computational Linguistics.  
+Ben Athiwaratkun and Andrew Gordon Wilson. 2018. On modeling hierarchical data via probabilistic order embeddings. In International Conference on Learning Representations.  
+Dzmitry Bahdanau, Kyung Hyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015.  
+K. A. Baker, P. C. Fishburn, and F. S. Roberts. 1972. Partial orders of dimension 2. Networks, 2(1):11-28.  
+Srinivas Bangalore and Aravind K. Joshi. 1999. Supertagging: An approach to almost parsing. Computational Linguistics, 25(2):237-265.  
+Yehoshua Bar-Hillel. 1953. A quasi-arithmetical notation for syntactic description. Language, 29(1):47-58.  
+Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv:2004.05150.  
+Jon Louis Bentley. 1979. Decomposable searching problems. Information Processing Letters, 8(5):244-251.
+
+Jon Louis Bentley. 1980. Multidimensional divide-and-conquer. Commun. ACM, 23(4):214-229.  
+Mark de Berg, Ottfried Cheong, Marc van Kreveld, and Mark Overmars. 2008. Computational Geometry: Algorithms and Applications, 3rd edition. Springer-Verlag TELOS, Santa Clara, CA, USA.  
+G. Birkhoff. 1967. Lattice Theory. American Mathematical Society colloquium publications. American Mathematical Society.  
+Arthur Cayley. 1889. A theorem on trees. Quarterly Journal of Mathematics, 23:376-378.  
+Bernard Chazelle. 1988. A functional approach to data structures and its use in multidimensional searching. SIAM Journal on Computing, 17(3):427-462.  
+Bernard Chazelle. 1990a. Lower bounds for orthogonal range searching: I. The reporting case. Journal of the ACM, 37(2):200-212.  
+Bernard Chazelle. 1990b. Lower bounds for orthogonal range searching: Part II. The arithmetic model. Journal of the ACM, 37(3):439-463.  
+Danqi Chen and Christopher Manning. 2014. A fast and accurate dependency parser using neural networks. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 740-750, Doha, Qatar. Association for Computational Linguistics.  
+Noam Chomsky. 2015. The Minimalist Program, 20 edition. The MIT Press.  
+Yoeng-Jin Chu and Tseng-Hong Liu. 1965. On the shortest arborescence of a directed graph. Scientia Sinica, 14:1396-1400.  
+Kenneth Ward Church. 1988. A stochastic parts program and noun phrase parser for unrestricted text. In Second Conference on Applied Natural Language Processing, pages 136-143, Austin, Texas, USA. Association for Computational Linguistics.  
+Marie-Catherine de Marneffe and Christopher D. Manning. 2008. Stanford typed dependencies manual. Technical report, Technical report, Stanford University.  
+Timothy Dozat and Christopher D. Manning. 2017. Deep bioaffine attention for neural dependency parsing. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.  
+Ben Dushnik and E. W. Miller. 1941. Partially ordered sets. American Journal of Mathematics, 63(3):600-610.  
+Chris Dyer, Gábor Melis, and Phil Blunsom. 2019. A critical analysis of biased parsers in unsupervised parsing. CoRR, abs/1909.09428.
+
+Jack Edmonds. 1967. Optimum branchings. Journal of Research of the national Bureau of Standards B, 71(4):233-240.  
+Jason Eisner. 1997. Bilexical grammars and a cubic-time probabilistic parser. In Proceedings of the Fifth International Workshop on Parsing Technologies, pages 54-65, Boston/Cambridge, Massachusetts, USA. Association for Computational Linguistics.  
+Jason M. Eisner. 1996. Three new probabilistic models for dependency parsing: An exploration. In COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics.  
+Michael L. Fredman. 1976. New bounds on the complexity of the shortest path problem. SIAM Journal on Computing, 5(1):83-89.  
+Carlos Gómez-Rodríguez, Michalina Strzyz, and David Vilares. 2020. A unifying theory of transition-based and sequence labeling parsing. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3776-3793, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Carlos Gómez-Rodríguez and David Vilares. 2018. Constituent parsing as sequence labeling. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 1314-1324, Brussels, Belgium. Association for Computational Linguistics.  
+Carlos Gómez-Rodríguez and Joakim Nivre. 2010. A transition-based parser for 2-planar dependency structures. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1492-1501, Uppsala, Sweden. Association for Computational Linguistics.  
+Carlos Gomez-Rodriguez and Joakim Nivre. 2013. Divisible Transition Systems and Multiplanar Dependency Parsing. Computational Linguistics, 39(4):799-845.  
+F. Hausdorff. 1914. Grundzüge der Mengenlehre. Göschems Lehrbucherei/Gruppe I: Reine und Angewandte Mathematik Series. Veit & Company.  
+Toshio Hiraguchi. 1955. On the dimension of orders. The Science Reports of the Kanazawa University, 4:1-20.  
+Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. Neural Computation, 9(8):1735-1780.  
+Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. Span-BERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64-77.
+
+Ben Kantor and Amir Globerson. 2019. Coreference resolution with entity equalization. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 673-677, Florence, Italy. Association for Computational Linguistics.  
+Eliyahu Kiperwasser and Miguel Ballesteros. 2018. Scheduled multi-task learning: From syntax to translation. Transactions of the Association for Computational Linguistics, 6:225-240.  
+Eliyahu Kiperwasser and Yoav Goldberg. 2016. Simple and accurate dependency parsing using bidirectional LSTM feature representations. Transactions of the Association for Computational Linguistics, 4:313-327.  
+Yuval Kirstain, Ori Ram, and Omer Levy. 2021. Coreference resolution without span representations. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers), pages 14-19, Online. Association for Computational Linguistics.  
+Nikita Kitaev and Dan Klein. 2020. Tetra-tagging: Word-synchronous parsing with linear-time inference. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6255-6261, Online. Association for Computational Linguistics.  
+Donald E. Knuth. 1965. On the translation of languages from left to right. Information and Control, 8(6):607-639.  
+Donald E. Knuth. 1997. The Art of Computer Programming: Fundamental Algorithms, 3 edition, volume 1. Addison Wesley Longman Publishing Co., Inc., USA.  
+Sandra Kübler, Ryan McDonald, and Joakim Nivre. 2009. Dependency Parsing. Springer Cham.  
+E. L. Lawler. 1978. Sequencing jobs to minimize total weighted completion time subject to precedence constraints. In B. Alspach, P. Hell, and D.J. Miller, editors, Algorithmic Aspects of Combinatorics, volume 2 of Annals of Discrete Mathematics, pages 75-90. Elsevier.  
+Kenton Lee, Luheng He, Mike Lewis, and Luke Zettlemoyer. 2017. End-to-end neural coreference resolution. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 188-197, Copenhagen, Denmark. Association for Computational Linguistics.  
+Zuchao Li, Jiaxun Cai, Shexia He, and Hai Zhao. 2018. Seq2seq dependency parsing. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3203-3214, Santa Fe, New Mexico, USA. Association for Computational Linguistics.
+
+Tianyu Liu, Yuchen Jiang, Ryan Cotterell, and Mrimaya Sachan. 2022. A structured span selector. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2629-2641, Seattle, United States. Association for Computational Linguistics.  
+Ilya Loshchilov and Frank Hutter. 2019. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Holbrook Mann MacNeill. 1937. Partially ordered sets. Transactions of the American Mathematical Society, 42(3):416-460.  
+Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313-330.  
+Peter McCullagh. 1980. Regression models for ordinal data. Journal of the Royal Statistical Society. Series B (Methodological), 42(2):109-142.  
+Ryan McDonald and Fernando Pereira. 2006. Online learning of approximate dependency parsing algorithms. In 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 81-88, Trento, Italy. Association for Computational Linguistics.  
+Ryan McDonald, Fernando Pereira, Kiril Ribarov, and Jan Hajic. 2005. Non-projective dependency parsing using spanning tree algorithms. In Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, pages 523-530, Vancouver, British Columbia, Canada. Association for Computational Linguistics.  
+George A. Miller. 1994. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994.  
+Khalil Mrini, Franck Dernoncourt, Quan Hung Tran, Trung Bui, Walter Chang, and Ndapa Nakashole. 2020. Rethinking self-attention: Towards interpretability in neural parsing. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 731-742, Online. Association for Computational Linguistics.  
+Joakim Nivre. 2003. An efficient algorithm for projective dependency parsing. In Proceedings of the Eighth International Conference on Parsing Technologies, pages 149-160, Nancy, France.  
+Joakim Nivre, Mitchell Abrams, Željko Agić, Lars Ahrenberg, Lene Antonsen, Maria Jesus Aranzabe, Gashaw Arutie, Masayuki Asahara, Luma Ateyah, Mohammed Attia, Aitziber Atutxa, Liesbeth Augustinus, Elena Badmaeva, Miguel Ballesteros, Esha Banerjee, Sebastian Bank, Verginica
+
+Barbu Miticelu, John Bauer, Sandra Bellato, Kepa Bengoetxea, Riyaz Ahmad Bhat, Erica Biagetti, Eckhard Bick, Rogier Blokland, Victoria Bobicev, Carl Borstell, Cristina Bosco, Gosse Bouma, Sam Bowman, Adriane Boyd, Aljoscha Burchardt, Marie Candido, Bernard Caron, Gauthier Caron, Gülşen Cebroğlu Eryigit, Giuseppe G. A. Celano, Savas Cetin, Fabricio Chalub, Jinho Choi, Yongseok Cho, Jayeol Chun, Silvie Cinková, Aurélie Collomb, Cagrölekin, Miriam Connor, Marine Courtin, Elizabeth Davidson, Marie-Catherine de Marneffé, Valeria de Paiva, Arantza Diaz de Ilarraza, Carly Dickerson, Peter Dirix, Kaja Dobrovoljc, Timothy Dozat, Kira Droganova, Puneet Dwivedi, Marhaba Eli, Ali Elkahky, Binyam Ephrem, Tomaz Erjavec, Aline Etienne, Richard Farkas, Hector Fernandez Alcalde, Jennifer Foster, Cláudia Freitas, Katarína Gajdsova, Daniel Galbraith, Marcos Garcia, Moa Gartenfors, Kim Gerdes, Filip Ginter, lakes Goenaga, Koldo Gojenola, Memduh Gökirmak, Yoav Goldberg, Xavier Gómez Guinovart, Berta Gonzáles Saavedra, Matias Grioni, Normunds Grūzītis, Bruno Guillaume, Céline Guillot-Barbance, Nizar Habash, Jan Hajic, Jan Hajic jr., Linh Hà My, Na-Rae Han, Kim Harris, Dag Haug, Barbora Hladka, Jaroslava Hlaváčová, Florinel Hociung, Petter Hohle, Jena Hwang, Radu Ion, Elena Irimia, Tomás Jelinek, Anders Johannsen, Fredrik Jørgensen, Hüner Kaskikara, Sylvain Kahane, Hiroshi Kanayama, Jenna Kanerva, Tolga Kayadelen, Václava Kettnerová, Jesse Kirchner, Natalia Kotsyba, Simon Krek, Sookyoung Kwak, Veronika Laippala, Lorenzo Lambertino, Tatiana Lando, Septina Dian Larasati, Alexei Lavrentiev, John Lee, Phuong Lê Hong, Alessandro Lenci, Saran Lertpradit, Herman Leung, Cheuk Ying Li, Josie Li, Keying Li, KyungTae Lim, Nikola Ljubesić, Olga Loginova, Olga Lyashevskaya, Teresa Lynn, Vivien Macketanz, Aibek Makazhanov, Michael Mandl, Christopher Manning, Ruli Manurung, Catalina Máranduc, David Mareček, Katrin Marheinecke, Héctor Martínez Alonso, André Martins, Jan Mašek, Yuji Matsumoto, Ryan McDonald, Gustavo Mendonça, Niko Miekka, Anna Missilä, Cathalin Mitteltu Yusuke Miyao, Simonetta Montemagni, Amir More, Laura Moreno Romero, Shinsuke Mori Bjartur Mortensen Bohdan Moskalevskyi Kadri Muischnek Yugo Murawaki Kaili Müürisep Pinkey Nainwani Juan Ignacio Navarro Horniacek Anna Nedoluzhko Gunta Nespore-Berzkalne Luong Nguyen Thi Huyen Nguyen Thi Minh Vitaly Nikolaev Rattima Nitisaroj Hanna Nurmi Stina Ojala Adedayo Olúokun Mai Omura Petya Osenova Robert Ostling Lilja Øvrelid Niko Partanen Elena Pascual Marco Passarotti Agnieszka Patejuk Siyao Peng Cenel-Augusto Perez Guy Perrier Slav Petrov Jussi Piṭulainen Emily Pitler Barbara Plank Thierry Poibeau Martin Popel Lauma Pretkalnina Sophie Prévost Prokopis Prokopidis Adam Przepiórkowski Tiina Puolakainen Sampa Pyysalo Andriela Raabis Alexandre Rademaker Loganathan Ramasamy Taraka Rama Carlos Ramisch Vinit Ravishankar Livy Real Siva Reddy Georg Rehm Michael Rießler Larissa Rinaldi Laura Rituma Luisa Rocha Mykhailo Ro
+
+manenko, Rudolf Rosa, Davide Rovati, Valentin Rosca, Olga Rudina, Shoval Sadde, Shadi Saleh, Tanja Samardžić, Stephanie Samson, Manuela Sanguinetti, Baiba Saulite, Yanin Sawanakunanon, Nathan Schneider, Sebastian Schuster, Djame Seddah, Wolfgang Seeker, Mojgan Seraji, Mo Shen, Atsuko Shimada, Muh Shohibussirri, Dmitry Sichinava, Natalia Silveira, Maria Simi, Radu Simionescu, Katalin Simko, María Šimková, Kiril Simov, Aaron Smith, Isabela Soares-Bastos, Antonio Stella, Milan Straka, Jana Strnadová, Alane Suhr, Umut Sulubacak, Zsolt Szántó, Dima Taji, Yuta Takahashi, Takaaki Tanaka, Isabelle Tellier, Trond Trosterud, Anna Trukhina, Reut Tsarfaty, Francis Tyers, Sumire Uematsu, Zdenka Urešova, Larraitz Uria, Hans Uszkoreit, Sowmya Vajjala, Daniel van Niekerk, Gertjan van Noord, Viktor Varga, Veronika Vincze, Lars Wallin, Jonathan North Washington, Seyi Williams, Mats Wiren, Tsegay Wdomemariam, Tak-sum Wong, Chunxiao Yan, Marat M. Yavrumyan, Zhuoran Yu, Zdeněk Žabokrtský, Amir Zeldes, Daniel Zeman, Manying Zhang, and Hanzhi Zhu. 2018. Universal dependencies 2.2. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (UFAL), Faculty of Mathematics and Physics, Charles University.
+
+Joakim Nivre and Jens Nilsson. 2005. Pseudoprojective dependency parsing. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL'05), pages 99-106, Ann Arbor, Michigan. Association for Computational Linguistics.
+
+Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Olga Uryupina, and Yuchen Zhang. 2012. CoNLL-2012 shared task: Modeling multilingual unrestricted coreference in OntoNotes. In Joint Conference on EMNLP and CoNLL - Shared Task, pages 1-40, Jeju Island, Korea. Association for Computational Linguistics.
+
+John Shawe-Taylor and Nello Cristianini. 2004. Kernel Methods for Pattern Analysis. Cambridge University Press.
+
+Yikang Shen, Zhouhan Lin, Chin-Wei Huang, and Aaron Courville. 2018a. Neural language modeling by jointly learning syntax and lexicon. In International Conference on Learning Representations.
+
+Yikang Shen, Zhouhan Lin, Athul Paul Jacob, Alessandro Sordoni, Aaron Courville, and Yoshua Bengio. 2018b. Straight to the tree: Constituency parsing with neural syntactic distance. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1171-1180, Melbourne, Australia. Association for Computational Linguistics.
+
+N.A. Smith. 2011. Linguistic Structure Prediction. Synthesis digital library of engineering and computer science. Morgan & Claypool.
+
+M. Stede. 2012. *Discourse Processing*. Synthesis lectures on human language technologies. Morgan & Claypool.  
+Mark Steedman. 1987. Combinatory grammars and parasitic gaps. Natural Language & Linguistic Theory, 5(3):403-439.  
+Mark Steedman. 2000. The Syntactic Process. MIT Press, Cambridge, MA, USA.  
+Michalina Strzyz, David Vilaras, and Carlos Gomez-Rodriguez. 2019. Viable dependency parsing as sequence labeling. 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 717-723, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Michalina Strzyz, David Vilares, and Carlos Gomez-Rodriguez. 2020. Bracketing encodings for 2-planar dependency parsing. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2472-2484, Barcelona, Spain (Online). International Committee on Computational Linguistics.  
+Edward Szpilrajn. 1930. Sur l'extension de l'ordre partiel. Fundamenta Mathematicae, 16(1):386-389.  
+Robert Endre Tarjan. 1977. Finding optimum branchings. Networks, 7(1):25-35.  
+Ben Taskar, Dan Klein, Mike Collins, Daphne Koller, and Christopher Manning. 2004. Max-margin parsing. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 1-8, Barcelona, Spain. Association for Computational Linguistics.  
+Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. 2021. Long range arena: A benchmark for efficient transformers. In International Conference on Learning Representations.  
+L. Tesnière. 1959. Éléments de Syntaxe Structurale. C. Klincksieck.  
+Jacobo Valdes, Robert E. Tarjan, and Eugene L. Lawler. 1979. The recognition of series parallel digraphs. In Proceedings of the Eleventh Annual ACM Symposium on Theory of Computing, STOC '79, page 1-12, New York, NY, USA. Association for Computing Machinery.  
+Ivan Vendrov, Ryan Kiros, Sanja Fidler, and Raquel Urtasun. 2015. Order-embeddings of images and language. In International Conference on Learning Representations.  
+Xinyu Wang and Kewei Tu. 2020. Second-order neural dependency parsing with message passing and end-to-end training. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association
+
+for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 93-99, Suzhou, China. Association for Computational Linguistics.  
+Douglas B. West. 2018. Introduction to Graph Theory. Pearson Modern Classics for Advanced Mathematics Series. Pearson.  
+Liyan Xu and Jinho D. Choi. 2020. Revealing the myth of higher-order inference in coreference resolution. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8527-8533, Online. Association for Computational Linguistics.  
+Naiwen Xue, Fei Xia, Fu-dong Chiou, and Marta Palmer. 2005. The penn chinese treebank: Phrase structure annotation of a large corpus. Natural Language Engineering, 11(2):207-238.  
+Hiroyasu Yamada and Yuji Matsumoto. 2003. Statistical dependency analysis with support vector machines. In Proceedings of the Eighth International Conference on Parsing Technologies, pages 195-206, Nancy, France.  
+Songlin Yang and Kewei Tu. 2022. Headed-span-based projective dependency parsing. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2188-2200, Dublin, Ireland. Association for Computational Linguistics.  
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. XLNet: Generalized autoregressive pretraining for language understanding. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc.  
+Mihalis Yannakakis. 1982. The complexity of the partial order dimension problem. SIAM Journal on Algebraic Discrete Methods, 3(3):351-358.  
+Anssi Mikael Yli-Jyrä. 2003. Multiplanarity-a model for dependency structures in treebanks. In TLT 2003, Proceedings of the Second Workshop on Treebanks and Linguistic Theories. Växjö University Press.  
+Yu Zhang, Zhenghua Li, and Min Zhang. 2020. Efficient second-order TreeCRF for neural dependency parsing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3295-3305, Online. Association for Computational Linguistics.  
+Yue Zhang and Stephen Clark. 2008. A tale of two parsers: Investigating and combining graph-based and transition-based dependency parsing. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 562-571, Honolulu, Hawaii. Association for Computational Linguistics.
+
+Junru Zhou and Hai Zhao. 2019. Head-Driven Phrase Structure Grammar parsing on Penn Treebank. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2396-2408, Florence, Italy. Association for Computational Linguistics.
+
+# A Related Work
+
+# A.1 Ordinal Regression
+
+Ordinal regression is a family of problems that involve ranking a set of objects. Unlike classification, the label spaces in ordinal regression exhibit some natural ordering in its elements (McCullagh, 1980). For instance, in information retrieval, a ranking model sorts a set of documents typically according to the document's relevance to the query. Practically, ordinal regression can either be tackled as either regression or classification by treating the ranks as real-values or the assignment to a particular rank value as a classification (Shawe-Taylor and Cristianini, 2004).
+
+# A.2 Order Embeddings of Lexicons
+
+The notion of partial order has also been explored for learning word embeddings. The lexicons of natural languages exhibit hierarchical structures according to the concepts that the words represent (Miller, 1994). For instance, 'cat' and 'dog' are 'animal', 'animal' and 'plant' are 'living thing'. Order embeddings (Vendrov et al., 2015; Athiwaratkun and Wilson, 2018) propose to learn such property by learning embeddings that encode such partial order on the lexicon, resulting in improved performance on downstream tasks such as image caption retrieval.
+
+# B An Order-Theoretic Re-evaluation of §2.2
+
+![](images/00d1cf4f1b259ddfa43fd5c89048077535e1310153f34b7a459d5a84f44fd706.jpg)  
+(a) The example binary tree in §2.2 and its traversal sequences.
+
+![](images/6c492d770978b383367264bfd6cd4a27407cc01c27d1c9b89da39ea1ddcfd444.jpg)  
+Inorder:abcdefg  
+Postorder: acbegfd  
+(b) Partial order of §2.2 defined by the intersection of in- and postorder.  $A\to B$  represents the relation  $A\prec_{1}B$
+
+Figure 3: An order-theoretic re-evaluation of Thm. B.1.  
+![](images/c2eb0058cade6211c8e3e42711ce6d62caa0e7fcca5e76cc684bc33c487e3cf6.jpg)  
+(c) Partial order of §2.2 defined by the intersection of reversed in- and postorder.  $A \to B$  represents the relation  $A \prec_{2} B$ .
+
+Theorem B.1 (A binary tree and its traversal; Knuth, 1997, §2.3.1, Ex. 7). Given the inorder and either the pre- or postorder traversal of the vertices in a binary tree, the binary tree structure can be reconstructed.
+
+Proof Sketch (order-theoretic). Without loss of generality, we explain the case of the combination of in- and postorder.  $V$  denotes the set of vertices in the binary tree. First, the intersection of in- and postorder defines a partial order relation  $\mathcal{P}_1 = (V,E_1,\prec_1)$ . For any 2 vertices  $x,y$  in the binary tree,  $x\prec_{1}y$  if and only if  $x$  is a left descendant of  $y$ . I.e.,  $x$  is either the left child or a descendant of the left child of  $y$  (see Fig. 3b). Since  $x$  is visited before visiting  $y$  in both inorder traversal and postorder traversal, if and only if  $x$  is the left descendant of  $y$ . The left child of each vertex in  $V$  can be decoded from  $\mathcal{P}_1$  by finding
+
+the child with the deepest subtree. Second, the intersection of reversed inorder and postorder defines a partial order relation  $\mathcal{P}_2 = (V, E_2, \prec_2)$ . For any 2 vertices  $x, y$  in the binary tree,  $x \prec_2 y$  if and only if  $x$  is a right descendant of  $y$  (see Fig. 3c). Since  $x$  is visited before visiting  $y$  in both the reversed inorder traversal and postorder traversal, if and only if  $x$  is the right descendant of  $y$ . The right child of each vertex in  $V$  can be decoded from  $\mathcal{P}_2$  also by finding the child with the deepest subtree. Thus, the original binary tree can be reconstructed.
+
+# C Proofs on the Partially Ordered Properties of Structures
+
+# C.1 Proof of Thm. 3.14
+
+Theorem 3.14. Token-split structures are partially ordered.
+
+Proof. We show that a token-split structure  $\mathcal{P} = \left(\widehat{V},\widehat{E},\prec\right)$  satisfies all the properties of partially ordered structure defined in Def. 3.4.
+
+1. irreflexivity: By Def. 3.13(c), for all  $x \in \widehat{V}$ ,  $x \neq x$ .  
+2. asymmetry: Suppose that  $\exists x, y, x \neq y$ , s.t.  $x \prec y \land y \prec x$ . By Definitions 3.13(b) and 3.13(c),  $x, y \in V^r \cap V^b = \emptyset$ . Thus,  $x \prec y \Rightarrow y \nprec x$ .  
+3. transitivity:  $x \prec y \land y \prec z$  cannot hold by Def. 3.13 (c). Since  $x \prec y$  implies  $x \in V^r \land y \in V^b$ , while  $y \prec z$  implies  $y \in V^r \land x \in V^b$ , a contradiction occurs due to  $y \in V^r \cap V^b = \emptyset$  by Def. 3.13 (b)  $x \prec y \land y \prec z \Rightarrow x \prec z$  holds since the antecedent of the proposition is always false.
+
+Thus, token-split structures are partially ordered.
+
+# D Guarantees of Order Dimension of Linguistic Structures
+
+We justify the guarantees of order dimension of linguistic structures. One conventional way to characterize the dimension of partial orders is from a lattice-theoretic point of view. A basic result tells us that a partial order is 2-dimensional if and only if its complete lattice embedding has a planar Hasse diagram (Baker et al., 1972). In other words, its complete lattice embedding can be drawn on a plane without any crossing edges.
+
+Theorem D.1 (Baker et al., 1972, Thm. 4.1). Suppose  $\mathcal{P} = (V,E,\prec)$  is a partially ordered structure. Then the following are equivalent:
+
+(a)  $D(\mathcal{P})\leq 2$  
+(b) The complete lattice embedding of  $\mathcal{P}$  has a planar Hasse diagram.
+
+Remark D.2. MacNeill (1937) and Birkhoff (1967, Chapter 5) introduced the construction of complete lattice embeddings for any partial order. Although it is difficult in practice to compute the complete lattice embedding for a partially ordered structure (MacNeill, 1937), Thm. D.1 can still provide an empirical characterization of the class of structures that can be efficiently represented. According to Euler's formula, the average degree of a vertex in a planar graph cannot exceed 6 (West, 2018, §6.1.23), which intuitively forces the partially ordered structures that can be represented as an intersection of 2 totally ordered structures to be sparse enough—thus to have planar complete lattice embeddings.
+
+Fortunately, this is often the case in natural language. Such phenomenon is closely related to what is termed valency by Tesnière (1959, Part 1, Book D). The number of actants (i.e., arguments) needed to implement the function of a word is a property of the word itself—a constant that does not change with the context (cf. categories<sup>9</sup> in categorial grammars (Adjukiewicz, 1935; Bar-Hillel, 1953; Steedman, 1987)). In natural language, the valency of a word is often a small constant. For instance, Steedman (2000, Chapter 3, fn. 10 and Chapter 8, p. 212) observes that the highest valency in the Dutch and English lexicon can be regarded as bounded by 4.
+
+We refer interested readers to MacNeill (1937) and Birkhoff (1967, Chapter 5) for the construction of complete lattice embeddings. Here, we give a weaker but more practical efficiency guarantee, based on a method to construct large partially ordered structures from smaller partially ordered structures.
+
+Definition D.3 (Series-Parallel Partial Orders; Valdes et al., 1979). A partially ordered structure is series-parallel if it satisfies the following inductive definition:
+
+(a) A single-vertex structure with no edges is series-parallel;  
+(b) If partially ordered structures  $\mathcal{P}_1 = (V_1, E_1, \prec)$  and  $\mathcal{P}_2 = (V_2, E_2, \prec)$  are series-parallel, so is the partially ordered structures constructed by either of the following operations:
+
+i. Parallel composition:
+
+$$
+\mathcal {P} _ {\mathrm {p}} = \left(V _ {1} \cup V _ {2}, E _ {1} \cup E _ {2}, \prec\right).
+$$
+
+ii. Series composition:
+
+$\mathcal{P}_{\mathrm{s}} = (V_{1}\cup V_{2},E_{1}\cup E_{2}\cup (\mathcal{M}_{1}\times \mathcal{N}_{2}),\prec)$ , where  $\mathcal{M}_1$  is the set of sinks of  $\mathcal{P}_1$  and  $\mathcal{N}_2$  the set of sources of  $\mathcal{P}_2$ .<sup>10</sup>
+
+Theorem D.4 (Series-parallel partially ordered structures are 2-dimensional; Valdes et al., 1979). The dimension of series-parallel partially ordered structures is at most 2.
+
+Thm. D.4 provides the guarantee that many structures in natural language processing can be represented as the intersection of 2 totally ordered structures. Since most structures of interest in NLP, such as trees and forests (thereby alignments and set partitioning), can be subsumed under series-parallel partially ordered structures, therefore have an order dimension of at most 2.
+
+Proposition D.5 (Trees are 2-dimensional; Lawler, 1978). Directed tree partially ordered structures are series-parallel. The order dimension of tree structures is at most 2.
+
+Proposition D.6 (Forests are 2-dimensional). Forests are series-parallel. The order dimension of forest structures is at most 2.
+
+Proof. Forests are parallel compositions of trees. Thus, the proposition holds.
+
+# E Efficient Algorithm for  $\oplus$ -Aggregation
+
+# E.1 Correctness of Algorithm 1
+
+Algorithm 1 Computing  $G_{1}$  when  $K = 2$ .  
+1: procedure  $\mathrm{COMPU}E - G_{1}(f_{\pmb{\theta}}^{(1)},f_{\pmb{\theta}}^{(2)},V)$   
+2:  $U\gets \mathsf{sort}\left(V,\mathsf{key} = f_{\pmb{\theta}}^{(1)} - f_{\pmb{\theta}}^{(2)}\right)$   
+3:  $G_{1},s_{1}\gets 0,0\succ 0$  is the zero element of  $\oplus$   
+4: for  $n = 1$  up to  $N$  :  
+5:  $q_{1} = f_{\pmb{\theta}}^{(1)}(U_{n}) + s_{1}\triangleright q_{1} = G_{1}(U_{n})$   
+6:  $G_{1}\oplus = q_{1}$   
+7:  $s_1\oplus = -f_\pmb{\theta}^{(1)}(U_n)$   
+8: return  $G_{1}$
+
+Proposition E.1. In Algorithm 1,  $G_{1} = \bigoplus_{x\in V}\bigoplus_{y\in S_{1}(x)}\Big(f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)\Big).$
+
+Proof. By induction, we show that upon finishing step  $n$ ,  $s_1 = \bigoplus_{y \in S_1(U_{n+1})} -f_{\pmb{\theta}}^{(1)}(y)$ ,  $G_1 = \bigoplus_{x \in \{U_1, \dots, U_n\}} \bigoplus_{y \in S_1(x)} \left(f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)\right)$ . First,  $S_1(U_n) = \{U_1, \dots, U_{n-1}\}$  holds as discussed in §4.3. When  $n = 1$ , we have  $s_1 = -f_{\pmb{\theta}}^{(1)}(U_1)$ ,  $G_1 = \mathbf{0} = \bigoplus_{x \in \{U_1\}} \bigoplus_{y \in S_1(x)} \left(f_{\pmb{\theta}}^{(1)}(x) - f_{\pmb{\theta}}^{(1)}(y)\right)$ .
+
+since  $S_{1}(U_{1}) = \emptyset$ . Assume that our statements hold for  $n = j$ , when  $n = j + 1$ , it is straightforward that  $s_{1} = \bigoplus_{y\in S_{1}(U_{j + 2})} - f_{\pmb{\theta}}^{(1)}(y)$ . For  $G_{1}$ , we have
+
+$$
+\begin{array}{l} G _ {1} = \bigoplus_ {x \in \left\{U _ {1}, \dots , U _ {j} \right\}} \bigoplus_ {y \in \mathcal {S} _ {1} (x)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) \oplus \left(f _ {\boldsymbol {\theta}} ^ {(1)} \left(U _ {j + 1}\right) + \bigoplus_ {y \in \mathcal {S} _ {1} \left(U _ {j + 1}\right)} - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) (7a) \\ = \bigoplus_ {x \in \left\{U _ {1}, \dots , U _ {j} \right\}} \bigoplus_ {y \in \mathcal {S} _ {1} (x)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) \oplus \bigoplus_ {y \in \mathcal {S} _ {1} \left(U _ {j + 1}\right)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} \left(U _ {j + 1}\right) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) (7b) \\ = \bigoplus_ {x \in \left\{U _ {1}, \dots , U _ {j + 1} \right\}} \bigoplus_ {y \in \mathcal {S} _ {1} (x)} \left(f _ {\boldsymbol {\theta}} ^ {(1)} (x) - f _ {\boldsymbol {\theta}} ^ {(1)} (y)\right) (7c) \\ \end{array}
+$$
+
+Thus, the claims hold for  $n = j + 1$ , establishing the induction step.
+
+Proposition E.2. Algorithm 1 runs in  $\mathcal{O}(N)$  time and space. With parallel computing, Algorithm 1 runs in  $\mathcal{O}(\log N)$  span.
+
+Proof. The sorting step in line 2 can be executed in  $\mathcal{O}(N)$  time and space. The for loop in lines 4 to 7 runs in  $\mathcal{O}(N)$  time and space. In total, Algorithm 1 runs in  $\mathcal{O}(N)$  time and space. Computing  $s_1$  in each step is a prefix-sum of  $-f_{\pmb{\theta}}^{(1)}(U_n)$ , which can be done in  $\mathcal{O}(\log N)$  span with parallel computing.  $q_1, G_1$  in each step can be computed in  $\mathcal{O}(1)$  in parallel following the computation of all  $s_1$ . Thus, the total span of Algorithm 1 is  $\mathcal{O}(\log N)$ .
+
+# E.2 Order Dimension  $K > 2$
+
+Finding all  $y \in V$  such that  $x \prec y$  in a partial order for a given  $x \in V$  requires efficiently finding all  $y$  that satisfy  $\bigwedge_{k \in [K]} (f_{\theta}^{(k)}(x) < f_{\theta}^{(k)}(y))$ . We remark that this problem bears a resemblance to orthogonal range searching in a  $K$ -dimensional space (Berg et al., 2008, Chapter 5), i.e., for a given  $x$ , we aim to find all  $y$  such that  $(f_{\theta}^{(1)}(y), f_{\theta}^{(2)}(y), \dots, f_{\theta}^{(K)}(y))$  is within the range  $(f_{\theta}^{(1)}(x), \infty) \times (f_{\theta}^{(2)}(x), \infty) \times \dots \times (f_{\theta}^{K}(x), \infty)$ . This problem can be naively solved in  $\mathcal{O}\left(\log^{K-1} N + \ell\right)$  using a range tree (Bentley, 1979, 1980; Chazelle, 1988, 1990a,b), where  $\ell$  is the cardinality of query results, as opposed to arc-factored models in which solving the same problem takes  $\mathcal{O}(N)$  computations.
+
+For  $\oplus$ -aggregation, a more efficient algorithm which makes use of  $(K - 1)$ -dimensional range trees can be designed. In future work, we show that computing the complexity of  $\oplus$ -aggregation for all  $x \in V$  can be further reduced to  $\mathcal{O}\big(KN\log^{K - 2}N\big)$  by applying Fredman's (1976) trick which we used in Algorithm 1. Extending the notation in §4.3, the set of all vertices  $V$  can be partitioned into  $K$  subsets  $S_{1}(x), \dots, S_{K}(x)$  for each  $x \in V$ , where  $S_{k}(x) = \{y \mid y \in V \land F_{\pmb{\theta}}(x,y) = f_{\pmb{\theta}}^{(k)}(x) - f_{\pmb{\theta}}^{(k)}(y)\} \oplus_{y \in V} F_{\pmb{\theta}}(x,y)$  can be decomposed into a  $\oplus$ -aggregation of  $K$  terms.
+
+$$
+G (x) \stackrel {\text {d e f}} {=} \bigoplus_ {y \in V} F _ {\theta} (x, y) \tag {8a}
+$$
+
+$$
+G (x) = \bigoplus_ {k \in [ K ]} \underbrace {\left(\bigoplus_ {y \in \mathcal {S} _ {k}} F _ {\boldsymbol {\theta}} (x , y)\right)} _ {\stackrel {\text {d e f}} {=} G _ {k} (x)} \tag {8b}
+$$
+
+We leave to future work showing that computing each  $G_{k}(x)$  takes  $\mathcal{O}\big(\log^{K - 2}N\big)$ .
+
+# F Hyperparameter Settings
+
+# F.1 Dependency Parsing
+
+For pretrained language models, we use XLNet-large-cased $^{11}$  (Yang et al., 2019) for PTB, bert-base-chinese $^{12}$  for CTB, and bert-base-multilingual-cased $^{13}$  for UD. We set the dimension of POS tag embedding to 256 for all experiments. On top of concatenated pretrained representations
+
+and POS embedding, we use a 3-layer BiLSTM (Hochreiter and Schmidhuber, 1997) with a hidden size of 768 for base-sized models (bert-base-chinese on CTB and bert-multilingual-cased on UD) and 1024 for large-sized models (x1net-large-cased on PTB). We apply dropout with a rate of 0.33 to the concatenated embedding layer, between LSTM layers, and before the linear projection layer of the realizer. We employ AdamW (Loshchilov and Hutter, 2019) with a learning rate of  $2\mathrm{e}-5$  for pretrained LMs and  $1\mathrm{e}-4$  for POS embedding, BiLSTM, and linear projection during training. The gradient clipping threshold is set to 1.0. The batch size for training is 32. The number of training epochs is 50.
+
+# F.2 Coreference Resolution
+
+We use longformer-large-cased $^{14}$  (Beltagy et al., 2020) as the pretrained encoder. We use the same hyperparameter settings as Kirstain et al. (2021). We use AdamW with a learning rate of  $1\mathrm{e} - 5$  for pretrained LM and  $3\mathrm{e} - 4$  for the linear projection during training, with 5600 linear warmup steps. Training documents are batched into batches with maximum 5000 tokens in total. The number of training epochs is 129.
+
+# G Datasets
+
+# G.1 Dependency Parsing
+
+Preprocessing. We follow previous work (Kiperwasser and Goldberg, 2016; Dozat and Manning, 2017) to derive the dependency annotations from the treebank annotations using the Stanford Dependency converter v3.3.0 (de Marneffe and Manning, 2008). During evaluation, punctuations are omitted. Following Amini et al. (2023), we provide gold part-of-speech tags to the model during training and decoding.
+
+Splits. The dataset splits are consistent with previous work. For PTB, we follow the standard split of Marcus et al. (1993), resulting in 39,832 sentences for training, 1,700 for development, and 2,416 for testing. For CTB, we follow the split of Zhang and Clark (2008), resulting in 16,091 sentences for training, 803 for development, and 1,910 for testing. For UD, we follow previous work (Zhang et al., 2020; Yang and Tu, 2022) and use the standard splits of the following corpora for experiments: BG-btb, CA-ancora, CS-pdt, DE-gsd, EN-ewt, ES-ancora, FR-gsd, IT-isdt, NL-alpino, NO-rrt, RO-rrt, RU-syntagrus.
+
+Licenses. The PTB and CTB datasets are licensed under LDC User Agreement. The UD dataset is licensed under the Universal Dependencies License Agreement.
+
+# G.2 Coreference Resolution
+
+Preprocessing. We experiment on the CoNLL-2012 English shared task dataset (OntoNotes; Pradhan et al., 2012). We follow the preprocessing procedure of (Kirstain et al., 2021). During training and decoding, the speaker information is provided to the model.
+
+Splits. The OntoNotes dataset contains 2,802 documents for training, 343 for validation, and 348 for testing. We use this official split following previous work (Lee et al., 2017; Kirstain et al., 2021).
+
+Licenses. The OntoNotes dataset is licensed under LDC User Agreement.
+
+# H Efficiency Evaluation
+
+# H.1 Dependency Parsing
+
+For efficiency evaluation, BERT-large-cased $^{15}$  is used as the pretrained encoder for our method with  $K = 2$ , hexatagger (Hexa; Amini et al., 2023), and biaffine model (Biaff). We use the English PTB test set and truncate or pad the input sentences to the control length. The results are averaged over 3 random runs on the same server with one Nvidia A100-80GB GPU. The other experimental settings are kept the same (i.e., the version of PyTorch and transformers, FP32 precision, batching).
+
+# H.2 Coreference Resolution
+
+<table><tr><td rowspan="2">Doc length</td><td colspan="2">Speed (doc/s) ↑</td><td colspan="2">Memory (GB) ↓</td></tr><tr><td>Ours (K = 4)</td><td>Kirstain et al.</td><td>Ours (K = 4)</td><td>Kirstain et al.</td></tr><tr><td>512</td><td>72.5</td><td>35.7</td><td>7.3</td><td>7.4</td></tr><tr><td>1024</td><td>54.3</td><td>26.7</td><td>7.3</td><td>7.4</td></tr><tr><td>2048</td><td>33.8</td><td>15.9</td><td>9.4</td><td>9.5</td></tr><tr><td>4096</td><td>19.3</td><td>8.6</td><td>17.8</td><td>21.0</td></tr><tr><td>overall</td><td>82.8</td><td>41.9</td><td>7.3</td><td>7.4</td></tr></table>
+
+We compare the efficiency of our order-theoretic method with baseline coreference resolution model. The full results are given in Tab. 4. On the OntoNotes coreference resolution benchmark, our method is twice as fast as Kirstain et al.'s (2021) model while using less memory, especially on long documents. It is worth noting that Kirstain et al. (2021) has already performed aggressive optimization in both the speed and memory footprint of coreference modeling. I.e., they abandon the computation for textual span representations and entity-pair representations, and use biaffine scorers to compute coreference scores.
+
+# I Additional Experimental Results
+
+# I.1 Dependency Parsing
+
+We report additional experimental results on the UD dependency parsing dataset in Tab. 5. On average, our model has state-of-the-art performance and outperforms all other baseline models on 5 languages.
+
+Table 4: Comparison of speed and memory consumption on OntoNotes test set using Longformer-base $^{16}$  as pretrained encoder. Results are averaged over 3 random runs on the same server with one Nvidia A100-80GB GPU using BERT-large as encoder. We use a batch size of 32 documents.  
+
+<table><tr><td></td><td>bg</td><td>ca</td><td>cs</td><td>de</td><td>en</td><td>es</td><td>fr</td><td>it</td><td>nl</td><td>no</td><td>ro</td><td>ru</td><td>Avg.</td></tr><tr><td>Zhang et al. (2020)</td><td>90.77</td><td>91.29</td><td>91.54</td><td>80.46</td><td>87.32</td><td>90.86</td><td>87.96</td><td>91.91</td><td>88.62</td><td>91.02</td><td>86.90</td><td>93.33</td><td>89.33</td></tr><tr><td>Wang and Tu (2020)</td><td>90.53</td><td>92.83</td><td>92.12</td><td>81.73</td><td>89.72</td><td>92.07</td><td>88.53</td><td>92.78</td><td>90.19</td><td>91.88</td><td>85.88</td><td>92.67</td><td>90.07</td></tr><tr><td colspan="14">+BERTmultilingual</td></tr><tr><td>Wang and Tu (2020)</td><td>91.30</td><td>93.60</td><td>92.09</td><td>82.00</td><td>90.75</td><td>92.62</td><td>89.32</td><td>93.66</td><td>91.21</td><td>91.74</td><td>86.40</td><td>92.61</td><td>90.61</td></tr><tr><td>Dozat and Manning (2017)</td><td>90.30</td><td>94.49</td><td>92.65</td><td>85.98</td><td>91.13</td><td>93.78</td><td>91.77</td><td>94.72</td><td>91.04</td><td>94.21</td><td>87.24</td><td>94.53</td><td>91.82</td></tr><tr><td>Yang and Tu (2022)</td><td>91.10</td><td>94.46</td><td>92.57</td><td>85.87</td><td>91.32</td><td>93.84</td><td>91.69</td><td>94.78</td><td>91.65</td><td>94.28</td><td>87.48</td><td>94.45</td><td>91.96</td></tr><tr><td>Amini et al. (2023)</td><td>92.87</td><td>93.79</td><td>92.82</td><td>85.18</td><td>90.85</td><td>93.17</td><td>91.50</td><td>94.72</td><td>91.89</td><td>93.95</td><td>87.54</td><td>94.03</td><td>91.86</td></tr><tr><td>ours (K=2)</td><td>92.81</td><td>93.26</td><td>92.52</td><td>83.33</td><td>90.38</td><td>92.55</td><td>89.83</td><td>93.82</td><td>91.29</td><td>93.61</td><td>87.40</td><td>94.10</td><td>91.24</td></tr><tr><td>ours (K=4)</td><td>93.82</td><td>94.23</td><td>93.03</td><td>84.68</td><td>91.40</td><td>93.62</td><td>90.95</td><td>94.59</td><td>92.58</td><td>94.22</td><td>88.45</td><td>94.40</td><td>92.16</td></tr></table>
+
+Table 5: LAS scores on the test sets of 12 languages in UD. Our method with an order dimension of  $K = 4$  achieves competitive performance in all languages, being state-of-the-art on 5 languages and on average.
+
+# J Qualitative Examples
+
+We present some qualitative examples from the PTB development set and one non-projective example using our method with a 2-dimensional realizer, with their ground truth annotations on the right in Figures 4-9. For a more intuitive and compact exhibition, we plot the 2 total orders output by our model in a 2-dimensional plane. Each axis corresponds to one of the 2 orders. The relation  $x \prec y$  encoded by  $\bigwedge_{k \in \{1,2\}} f_{\pmb{\theta}}^{(k)}(x) < f_{\pmb{\theta}}^{(k)}(y)$  is equivalent to  $x$  being located below and to the left of  $y$ .
+
+Tokens in  $V^r$  and  $V^b$  are represented by  $\times$  and  $\bullet$ , respectively. The line segments between  $\times$  and  $\bullet$  are the extracted dependency relations. In each of the plots, every  $\times$  (token in  $V^r$ ) except for the root is connected to a  $\bullet$  (token in  $V^b$ ), which indicates  $\times$  is the modifier of  $\bullet$ . The roots (about, moving, ready, had, adds, bought represented by  $\times$ ) are not connected to any other word.
+
+![](images/2bea9d5faf9b4e7ad8fafd5e0c1960701e2f8159a8f729397aa3b6e9e99fea9a.jpg)  
+Figure 4: We 're about to see if advertising works
+
+![](images/5194bfe676904a59b6cb4ee0754ad07dd1b9290d96822f279ad33e10bf5535a9.jpg)
+
+![](images/6af36febf64e700431f0713cb7294f62a44c1ddd1c6de5ee1a40b0f88b41739c.jpg)  
+Figure 5: This time around, they 're moving even faster
+
+![](images/11b1920fe582db1578760c5b036f5e284f86a40db14a8475b3902ab2dbf5735f.jpg)
+
+![](images/4a1f1767466e0f02c19e6ed19587bafd90594c550ecc306e2855d1f28637262d.jpg)  
+Figure 6: This time, the firms were ready
+
+![](images/f57a0a17bd40bf2d42c02f424dcf0336b99bd03b17cfe38ce73cc3b96830918d.jpg)
+
+![](images/9402c0b532750ea369497392588ea2de0919f0e165abeb011e044abd50d2e275.jpg)  
+Figure 7: We had to think about it ahead of time
+
+![](images/f7f3fe60d2f00be52f3781f0b9c2f6c5b95fd806b8e722bf8ed27df0481920c8.jpg)
+
+![](images/402d7407b8605310189ec89362441593ac2cbae6f8318461b88e6c3e09efdb7b.jpg)  
+Figure 8: He adds, "This isn't 1987 revisited"
+
+![](images/690b9679a7930c472d80790b69d32b1822230eeda4571e4745a7bc829ed4e926.jpg)
+
+![](images/8c1832c902b56ab8e25deb29e9fc5bdc20bae5661ff262e388baad759bb9468a.jpg)  
+Figure 9: A sentence with a non-projective dependency structure: John bought a car yesterday which is red
+
+![](images/1ae19cf5079238321f85f3a7a1d4d65e926827f11be42d218395bc23649e277c.jpg)

2023/Linear-Time Modeling of Linguistic Structure_ An Order-Theoretic Perspective/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32e6777b6c2f884d83b5964e536aa1bc18479ae030da473659ef45229f7bde0b
+size 576339

2023/Ling-CL_ Understanding NLP Models through Linguistic Curricula/555b9eb9-8ec1-48b5-b32f-200ee26764bf_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37b8436a560ec6297cfd6f795183a73c640f72752ce8d896834f726dbac033b0
+size 4253687

2023/Ling-CL_ Understanding NLP Models through Linguistic Curricula/full.md

@@ -0,0 +1,447 @@
+# Ling-CL: Understanding NLP Models through Linguistic Curricula
+
+Mohamed Elgaar and Hadi Amiri
+
+University of Massachusetts Lowell
+
+{melgaar,hadi}@cs.uml.edu
+
+# Abstract
+
+We employ a characterization of linguistic complexity from psycholinguistic and language acquisition research to develop data-driven curricula to understand the underlying linguistic knowledge that models learn to address NLP tasks. The novelty of our approach is in the development of linguistic curricula derived from data, existing knowledge about linguistic complexity, and model behavior during training. By analyzing several benchmark NLP datasets, our curriculum learning approaches identify sets of linguistic metrics (indices) that inform the challenges and reasoning required to address each task. Our work will inform future research in all NLP areas, allowing linguistic complexity to be considered early in the research and development process. In addition, our work prompts an examination of gold standards and fair evaluation in NLP.
+
+# 1 Introduction
+
+Linguists devised effective approaches to determine the linguistic complexity of text data (Wolfe-Quintero et al., 1998; Bulté and Housen, 2012; Housen et al., 2019). There is a spectrum of linguistic complexity indices for English, ranging from lexical diversity (Malvern et al., 2004; Yu, 2010) to word sophistication (O'Dell et al., 2000; Harley and King, 1989) to higher-level metrics such as readability, coherence, and information entropy (van der Sluis and van den Broek, 2010). These indices have not been fully leveraged in NLP.
+
+We investigate the explicit incorporation of linguistic complexity of text data into the training process of NLP models, aiming to uncover the linguistic knowledge that models learn to address NLP tasks. Figure 1 shows data distribution and accuracy trend of Roberta-large (Liu et al., 2019) against the linguistic complexity index "verb variation" (ratio of distinct verbs). This analysis is conducted on ANLI (Nie et al., 2020) validation data, where balanced accuracy scores are computed
+
+![](images/79bac9adca20ad0907e8c5e4e0ba07811549397aabb7dad2650b61700bc60f65.jpg)  
+Figure 1: Data distribution and trend of model accuracy against the linguistic index verb variation computed on ANLI (Nie et al., 2020) validation data. Samples with greater verb variation are more complex and also harder for the model to classify. Such linguistic indices can inform difficulty estimation and linguistic curriculum development for NLP tasks.
+
+for individual bins separately. The accuracy trend indicates that verb variation can describe the difficulty of ANLI samples to the model. In addition, the data distribution illustrates potential linguistic disparity in ANLI; see §3.4
+
+To reveal the linguistic knowledge NLP models learn during their training, we will employ known linguistic complexity indices to build multiview linguistic curricula for NLP tasks. A curriculum is a training paradigm that schedules data samples in a meaningful order for iterative training, e.g., by starting with easier samples and gradually introducing more difficult ones (Bengio et al., 2009). Effective curricula improve learning in humans (Tabibian et al., 2019; Nishimura, 2018) and machines (Bengio et al., 2009; Kumar et al., 2010; Zhou et al., 2020; Castells et al., 2020). Curriculum learning has been found effective in many NLP tasks (Settles and Meeder, 2016; Amiri et al., 2017; Platanios et al., 2019; Zhang et al., 2019; Amiri, 2019; Xu et al., 2020; Lalor and Yu, 2020; Jafarpour et al., 2021; Kreutzer et al., 2021; Agrawal and Carpuat, 2022; Maharana and Bansal, 2022). A multiview curriculum is a curriculum able to integrate multiple difficulty scores simultaneously and leverage their collective value (Vakil and Amiri, 2023).
+
+We assume there exists a subset of linguistic complexity indices that are most influential to learning an NLP task by a particular model. To identify these indices for each model and NLP task, we derive a weight factor  $\rho_{i} \in [-1, 1]$  for each linguistic index that quantifies how well the index estimates the true difficulty of data samples to the model, determined by model instantaneous loss against validation data. By learning these weight factors, we obtain precise estimations that shed light on the core linguistic complexity indices that each model needs at different stages of its training to learn an NLP task. In addition, these estimates can be readily used for linguistic curriculum development, e.g., by training models with linguistically easy samples (with respect to the model) and gradually introducing linguistically challenging samples.
+
+To achieve the above goals, we should address two gaps in the existing literature: First, existing curricula are often limited to a single criterion of difficulty and are not applicable to multiview settings. This is while difficulty is a condition that can be realized from multiple perspectives, can vary across a continuum for different models, and can dynamically change as the model improves. Second, existing approaches quantify the difficulty of data based on instantaneous training loss. However, training loss provides noisy estimates of sample difficulty due to data memorization (Zhang et al., 2017; Arpit et al., 2017) in neural models. We will address both issues as part of this research.
+
+The contributions of this paper are:
+
+- incorporating human-verified linguistic complexity information to establish an effective scoring function for assessing the difficulty of text data with respect to NLP models,  
+- deriving linguistic curricula for NLP models based on linguistic complexity of data and model behavior during training, and  
+- identifying the core sets of linguistic complexity indices that most contribute to learning NLP tasks by models.
+
+We evaluate our approach on several NLP tasks that require significant linguistic knowledge and reasoning to be addressed. Experimental results show that our approach can uncover latent linguistic knowledge that is most important for addressing NLP tasks. In addition, our approach obtains consistent performance gain over competing models. Source code and data is available at https://github.com/CLU-UML/Ling-CL.
+
+Algorithm 1 Correlation Method  
+Require:  $D_{train}$ ,  $D_{val}$ , Model  $\Theta$ , Optimizer  $g$ , Loss function  $f$ , Curriculum  $C$   
+1: step  $\leftarrow 0$   
+2:  $\rho \leftarrow$  random initialization  
+3: while step  $<$  total_steps do  
+4: training_batch  $\leftarrow$  SampleBatch(step,  $D_{train}$ )  
+5: loss  $\leftarrow$  ComputeLoss (training_batch,  $\Theta$ )  
+6: ling  $\leftarrow$  GetLinguisticFeatures(training_batch)  
+7: difficulty  $\leftarrow$  CalculateDifficulty(ling,  $\rho$ )  
+8: confidence  $\leftarrow$  DetermineConfidence(step, difficulty)  
+9: weighted_loss  $\leftarrow$  loss  $\otimes$  confidence  
+10:  $\Theta \leftarrow$  UpdateModel (weighted_loss,  $\Theta$ )  
+11: if step \% validation_step = 0 then  
+12: l  $\leftarrow$  ComputeLoss ( $D_{val}$ ,  $\Theta$ )  
+13: ling  $\leftarrow$  GetLinguisticFeatures( $D_{val}$ )  
+14: for  $\rho_i \in \rho$  do  
+15:  $\rho_i \leftarrow$  pearsonr(l, ling[:i])  
+16: end for  
+17: end if  
+18: step  $\leftarrow$  step + 1  
+19: end while
+
+# 2 Multiview Linguistic Curricula
+
+We present a framework for multiview curriculum learning using linguistic complexity indices. Our framework estimates the importance of various linguistic complexity indices, aggregates the resulting importance scores to determine the difficulty of samples for learning NLP tasks, and develops novel curricula for training models using complexity indices. The list of all indices used is in Appendix A.
+
+# 2.1 Linguistic Index Importance Estimation
+
+# 2.1.1 Correlation Approach
+
+Given linguistic indices  $\{\mathbf{X}_j\}_{j=1}^k$  of  $n$  data samples, where  $k$  is the number of linguistic indices and  $\mathbf{X}_j \in \mathbb{R}^n$ , we start by standardizing the indices  $\{\mathbf{Z}_j = \frac{\mathbf{X}_j - \mu_j}{\sigma_j}\}_{j=1}^k$ . We consider importance weight factors for indices  $\{\rho_j\}_{j=1}^k$ , which are randomly initialized at the start of training. At every validation step, the weights are estimated using the validation dataset by computing the Pearson's correlation coefficient between loss and linguistic indices of the validation samples  $\rho_j = r(\mathbf{l}, \mathbf{Z}_j)$  where  $r$  is the correlation function and  $\mathbf{l} \in \mathbb{R}^n$  is the loss of validation samples. The correlation factors are updated periodically. It is important to use validation loss as opposed to training loss because the instantaneous loss of seen data might be affected by memorization in neural networks (Zhang et al., 2017; Arpit et al., 2017; Wang et al., 2020). This is while unseen data points more accurately represent the difficulty of samples for a model. Algorithm 1 presents the correlation approach.
+
+# 2.1.2 Optimization Approach
+
+Let  $\mathbf{Z} \in \mathbb{R}^{n \times k}$  be the matrix of  $k$  linguistic indices computed for  $n$  validation samples and  $\mathbf{l} \in \mathbb{R}^n$  indicate the corresponding loss vector of validation samples. We find the optimal weights for linguistic indices to best approximate validation loss:
+
+$$
+\boldsymbol {\rho} ^ {*} = \arg \min  _ {\boldsymbol {\rho}} \| \mathbf {l} - \mathbf {Z} \boldsymbol {\rho} \| _ {2} ^ {2} + \lambda_ {\rho} \| \boldsymbol {\rho} \| _ {1}, \tag {1}
+$$
+
+where  $\lambda_{\rho}\in \mathbb{R}$  and  $\pmb {\rho}^{*}\in \mathbb{R}^{k}$  is jointly optimized over all indices. The index that best correlates with loss can be obtained as follows:
+
+$$
+i ^ {*} = \arg \min  _ {i} \| \mathbf {l} - \mathbf {Z} _ {* i} \boldsymbol {\rho} _ {i} \| _ {2} ^ {2}, \tag {2}
+$$
+
+where  $\mathbf{Z}_{*\mathbf{i}}$  denotes the  $i^{th}$  column of  $\mathbf{Z}$ . Algorithm 2 presents this approach.
+
+We note that the main distinction between the two correlation and optimization approaches lies in their scope: the correlation approach operates at the index level, whereas the optimization approach uses the entire set of indices.
+
+# 2.1.3 Scoring Linguistic Complexity
+
+We propose two methods for aggregating linguistic indices  $\{X_{j}\}^{k}$  and their corresponding importance factors  $\{\rho_j\}^k$  into a linguistic complexity score. The first method selects the linguistic index with the maximum importance score at each timestep:
+
+$$
+S _ {i} = Z _ {i \hat {j}}, \quad \hat {j} = \underset {j} {\arg \max } \rho_ {j}, \tag {3}
+$$
+
+which provides insights into the specific indices that determine the complexity to the model.
+
+The second method computes a weighted average of linguistic indices, which serves as a difficulty score. This is achieved as follows:
+
+$$
+S _ {i} = \frac {\sum_ {j} \rho_ {j} \mathbf {Z} _ {i j}}{\sqrt {\sum_ {j} \rho_ {j} ^ {2}}}, \tag {4}
+$$
+
+where  $S_{i}\in \mathbb{R},(\mu_{S_{i}},\sigma_{S_{i}}) = (0,1)$ , is an aggregate of linguistic complexity indices for the input text. If an index  $\mathbf{Z}_j$  is negatively correlated with loss,  $\rho_{j}$  will be negative, and  $\rho_{j}\mathbf{Z}_{j}$  will be positively correlated with loss. Therefore,  $S_{i}$  is an aggregate complexity that is positively correlated with loss. And using weighted average results in the indices that are most highly correlated with loss to have the highest contribution to  $S_{i}$ .
+
+Algorithm 2 Optimization Method  
+Require:  $D_{train}, D_{val}$ , Model  $\Theta$ , Optimizer  $g$ , Optimizer  $h$ , Loss function  $f$ , [Optional] Curriculum  $C$   
+1: step  $\leftarrow 0$   
+2:  $\rho \leftarrow$  random initialization  
+3: while step  $<$  total_steps do  
+4: training_batch  $\leftarrow$  SampleBatch(step,  $D_{train}$ )  
+5: loss  $\leftarrow$  ComputeLoss (training_batch,  $\Theta$ )  
+6: ling  $\leftarrow$  GetLinguisticFeatures(training_batch)  
+7: difficulty  $\leftarrow$  CalculateDifficulty(ling,  $\rho$ )  
+8: confidence  $\leftarrow$  DetermineConfidence(step, difficulty)  
+9: weighted_loss  $\leftarrow$  loss  $\otimes$  confidence  
+10:  $\Theta \leftarrow$  UpdateModel (weighted_loss,  $\Theta$ )  
+11: if step \% validation_step = 0 then  
+12: l  $\leftarrow$  ComputeLoss ( $D_{val}, \Theta$ )  
+13: ling  $\leftarrow$  GetLinguisticFeatures( $D_{val}$ )  
+14:  $\rho \leftarrow$  arg  $\min_{\rho} \| \text{ling} \cdot \rho - l \|_2^2 + \lambda_{\rho} \| \rho \|_1$   
+15: end if  
+16: step  $\leftarrow$  step + 1  
+17: end while
+
+# 2.2 Linguistic Curriculum
+
+We evaluate the quality of weighted linguistic indices as a difficulty score and introduce three new curricula based on a moving logistic (Richards, 1959) and Gaussian functions, see Figure 2.
+
+# 2.2.1 Time-varying Sigmoid
+
+We develop a time-varying sigmoid function to produce weights (Eq. 3). The sigmoid function assigns a low weight to samples with small difficulty scores and a high weight to larger difficulty scores. Weights are used to emphasize or de-emphasize the loss of different samples. For this purpose, we use the training progress  $t \in [0,1]$  as a shift parameter, to move the sigmoid function to the left throughout training, so that samples with a small difficulty score are assigned a higher weight in the later stages of training. By the end of the training, all samples are assigned a weight close to 1. Additionally, we add a scale parameter  $\beta \in [1,\infty)$  that controls the growth rate of weight (upper bounded by 1) for all samples.
+
+$$
+w \left(S _ {i}, t; \beta\right) = \frac {1}{1 + \exp \left(- S _ {i} - t \cdot \beta\right)}. \tag {5}
+$$
+
+The sigmoid function saturates as the absolute value of its input increases. To account for this issue, our input aggregated linguistic index follows the standard scale, mean of zero, and variance of one, in (4) and (3).
+
+# 2.2.2 Moving Negative-sigmoid
+
+The positive-sigmoid function assigns greater weights to samples with a large value for  $S$  that
+
+![](images/bdd6a946ae458f7c2d617771d6dadd3d0743319402494f374513e45a2ca61fd5.jpg)  
+(a) Sigmoid
+
+![](images/ca9eed6139a54038ebb0e20daaac777e09a1d0958f2209968bcbedd9539d9f9e.jpg)  
+(b) Negative Sigmoid  
+Figure 2: At the beginning of training, the sigmoid function with the lowest opacity is used. It is the right-most curve in (a), the left-most curve in (b), and the middle curve in (c). Then, as training progresses, the function is shifted using the parameter  $t$  in (5) and (7), causing samples with a higher complexity to be assigned higher confidence if (a) is used, samples with a lower complexity to be assigned higher confidence if (b) is used, and samples with medium complexity to be assigned higher confidence if (c) is used.
+
+![](images/b31fec10a94ec7e899fedd84d84583aca761cf9aff015d702ecd336938f2e932.jpg)  
+(c) Gaussian
+
+are linguistically more complex. In order to establish a curriculum that starts with easy samples and gradually proceeds to harder ones, we use a negative sigmoid function:
+
+$$
+w \left(S _ {i}, t; \beta\right) = \frac {1}{1 + \exp \left(S _ {i} - t \cdot \beta\right)}. \tag {6}
+$$
+
+Figure 2 illustrates the process of time-varying positive and negative sigmoid functions. Over the course of training, larger intervals of linguistic complexity are assigned full confidence, until the end of training when almost all samples have a confidence of one and are fully used in training.
+
+# 2.2.3 Time-varying Gaussian Function
+
+We hypothesize that training samples that are not too hard and not too easy are the most useful in training, and should receive the most focus. In fact, samples that are too easy or hard may contain artifacts that are harmful to training, may contain noise, and may not be generalizable to the target task. Therefore, we use the Gaussian function to prioritize learning from medium-level samples. The function starts with a variance of 1, and scales up during the course of training so that the easier and harder samples, having lower and higher linguistic complexity values, respectively, are assigned increasing weights and are learned by the end of training. We propose the following function:
+
+$$
+w \left(S _ {i}, t; \gamma\right) = \exp \left(\frac {- S _ {i} ^ {2}}{2 (1 + t \cdot \gamma)}\right), \tag {7}
+$$
+
+where  $\gamma$  is the rate of growth of variance and  $t$  is the training progress, see Figure 2.
+
+# 2.2.4 Weighting-based Curriculum
+
+We define a curriculum by weighting sample losses according to their confidence. Samples that are most useful for training receive higher weights, and those that are redundant or noisy receive smaller weights. Weighting the losses effectively causes the gradient update direction to be dominated by the samples that the curriculum thinks are most useful. Weights  $w$  are computed using either Equation 5, 6 or 7:
+
+$$
+\mathcal {L} = \frac {1}{\sum_ {i} w \left(S _ {i} , t ; \beta\right)} \sum_ {i} w \left(S _ {i}, t; \beta\right) \cdot \ell_ {i}, \tag {8}
+$$
+
+where  $\ell_{i}$  is the loss of sample  $i$ ,  $t$  the current training progress, and  $\mathcal{L}$  is the average weighted loss.
+
+# 2.3 Reducing Redundancy in Indices
+
+We have curated a list of 241 linguistic complexity indices. In the case of a text pair input (e.g. NLI), we concatenate the indices of the two text inputs, for a total of 482. Our initial data analysis reveals significant correlation among these indices in their estimation of linguistic complexity. To optimize computation, avoid redundancy, and ensure no single correlated index skews the complexity aggregation approach 2.1.3, we propose two methods to select a diverse and distinct set of indices for our study. We consider the choice of using full indices or filtering them as a hyper-parameter.
+
+In the first approach, for each linguistic index, we split the dataset into  $m$  partitions based on the index values (similar to Figure 1). Next, using a trained No-CL (§3.3) model, we compute the accuracy for each partition. Then, we find the first-order
+
+![](images/471a04e7b9fa541925fba2aef3535905396a38725bd48b4df8983cb4b4f54c84.jpg)  
+(a) Pair-wise correlation between indices
+
+![](images/fb2abc1de9d12ec2fdb609e1291bb44be60019d55804794722f797bc82ebffae.jpg)  
+(b) Clustered correlation matrix.  
+Figure 3: Eliminating redundancy in linguistic indices. (a) shows the Pearson's correlation coefficient between each pair of linguistic indices. (b) is created by reordering the rows and columns of (a), such that mutually correlated indices are clustered into blocks using hierarchical clustering (Kumar et al., 2000). Best seen in color; lighter areas indicate greater correlations among index pairs or groups.
+
+accuracy trend across these partitions. Linguistic indices with a pronounced slope describe great variance in the data and are considered for our study; we select the top  $30\%$  of indices, reducing their count from 482 to 144 for text pair inputs.
+
+In the second approach, we compute pair-wise correlations between all indices. Then, we group highly correlated indices, as shown in Figure 3. From each cluster, we select a representative index, aiming to prevent correlated indices from dominating the aggregation approach and to eliminate redundancy. This method narrows our focus to the following 16 key indices: 1) type-token ratio (TTR), 2) semantic richness, 3) ratio of verbs to tokens, 4) mean TTR of all  $k$  word segments, 5) Total number of verbs, 6) number of unique words, 7) adverbs per sentence, 8) number of unique words in the first  $k$  tokens, 9) ratio of nouns to verbs, 10) semantic noise, 11) lexical sophistication, 12) verb sophistication, 13) clauses per sentence, 14) average SubtexUS CDlow value per token, 15) adjective variation, 16) ratio of unique verbs. Please refer to Appendix A for definitions and references to indices.
+
+# 3 Experiments
+
+# 3.1 Datasets
+
+We evaluate NLP models in learning the tasks of the following datasets:
+
+- SNLI: Stanford Natural Language Inference (Bowman et al., 2015). The task is to classify a pair of sentences by the relation between them as one of entailment, neutral, or contradiction.  
+- CoLA: Corpus of Linguistic Acceptability (Warstadt et al., 2019). It is a task of classifying sentences as grammatical vs. ungrammatical.  
+- ANLI: Adversarial Natural Language Inference (Nie et al., 2020). This NLI dataset was created with a model in the loop, by only adding samples to the dataset that fool the model. We train only on the ANLI training set of 162k samples.  
+- SST-2: Stanford Sentiment Treebank (Socher et al., 2013). The task is to predict the sentiment of a given sentence as positive or negative.  
+- RTE: Recognizing Textual Entailment (Wang et al., 2018). The task is to determine if a given sentence is entailed by another given sentence.  
+- AN-Pairs: Adjective-noun pairs from the Cambridge ESOL First Certificate in English (FCE) exams (Kochmar and Briscoe, 2014). The task is to detect if an adjective-noun pair, including pairs that are typically confusing to language learners, is used correctly in the context of a sentence.  
+- GED: Grammatical Error Detection (Yannakoudakis et al., 2011). The task is to identify grammar errors at word level in given sentences.
+
+# 3.2 Difficulty Scoring Functions
+
+The curriculum learning approaches in §2.2 use difficulty scores or compute confidence to quantify sample difficulty in order to rank sentences. We use as difficulty scores: aggregate linguistic complexity Ling, see Section 2.1.3, and Loss (Xu et al., 2020; Wu et al., 2021; Zhou et al., 2020). We take the loss from a proxy model (No-CL in §3.3) by recording all samples losses two times per epoch during training and computing the sample-wise average.
+
+# 3.3 Baselines
+
+We consider a no-curriculum baseline as well as several recent curriculum learning approaches.
+
+- No-CL: no-curriculum uses standard random mini-batch sampling from the whole dataset without sample weighting.  
+- Sampling (Bengio et al., 2009) uses the easiest subset of the dataset at each stage of training. Instead of randomly sampling a mini-batch from the whole dataset, a custom data sampler is created that provides the subset consisting of the easiest  $\alpha\%$  of data when training progress is at  $\alpha\%$ .  
+- SL-CL & WR-CL (Platanios et al., 2019) is a curriculum learning approach that defines a time-varying function of the model's competence (defined as the fraction of training data that the model uses at every step), and a difficulty score of the data. At each iteration, a minibatch is sampled from the subset of data with difficulty smaller than the model's competence—a pre-defined non-linear function. The model employs sentence length (SL-CL) and word rarity (WR-CL) as difficulty measures. Sampling is the same as Competence-based curriculum learning with a linear competence function.  
+- SuperLoss (Castells et al., 2020) uses instantaneous loss to compute task-agnostic sample confidence. It emphasizes easy samples and de-emphasizes hard samples based on the global average loss as the difficulty threshold.  
+- Concat (Lee et al., 2021) concatenates linguistic indices to language embeddings before classification. Lee et al. (2021) and Meng et al. (2020) reported low performance as a result of appending features to embeddings. However, our approach succeeds in utilizing concatenated features.  
+- Data Selection (Mohiuddin et al., 2022) is an online curriculum learning approach. It evaluates the training data at every epoch and uses loss as the difficulty score. It selects the middle  $40\%$  of samples according to difficulty.
+
+We compare the above models against our approaches, Ling-CL, which aggregates linguistic indices using weighted average or max-index aggregation, and applies different curriculum strategies:
+
+sigmoid, negative-sigmoid, and Gaussian weighting, as well as sampling an competence-based approaches, see §3.3. We test variants of our approach with the correlation method, optimization method, and indices filtering. We report results of the max aggregation (§2.1.3) approach as it performs better than the weighted average and is computationally cheaper. Loss-CL computes loss as a difficulty score by recording the losses of samples during training of No-CL. The loss during the early stages of training generated by an under-trained model is a good measure of the relative difficulty of both training and validation samples.
+
+# 3.4 Evaluation Metrics
+
+Linguistic disparity can be quantified by the extent of asymmetry in the probability distribution of the linguistic complexity of samples in a dataset, e.g., see Figure 1 in §1. A natural solution to evaluate models is to group samples based on their linguistic complexity. Such grouping is crucial because if easy samples are overrepresented in a dataset, then models can result in unrealistically high performance on that dataset. Therefore, we propose to partition datasets based on a difficulty metric (linguistic index or loss) and compute balanced accuracy of different models on the resulting groups. This evaluation approach reveals great weaknesses in models, and benchmark datasets or tasks that seemed almost "solved" such as the complex tasks of NLI.
+
+# 3.5 Experimental Settings
+
+We use the transformer model roberta-base (Liu et al., 2019) from (Wolf et al., 2020), and run each experiment with at least two random seeds and report the average performance. We use AdamW (Loshchilov and Hutter, 2018) optimizer with a learning rate of  $1 \times 10^{-5}$ , batch size of 16, and weight decay of  $1 \times 10^{-2}$  for all models. The model checkpoint with the best validation accuracy is used for final evaluation. In NLI tasks with a pair of text inputs, the indices of both texts are used. For Ling-CL, we optimize the choice of index importance estimation method and aggregation method. For the baselines, we optimize the parameters of SuperLoss ( $\lambda$  and moving average method), and the two parameters of SL-CL and WR-CL models for each dataset. For the data selection, we use a warm-up period of  $20\%$  of the total training iterations.
+
+<table><tr><td></td><td>ANLI</td><td>COLA</td><td>RTE</td><td>SNLI</td><td>SST2</td><td>AN-Pairs</td><td>GED</td><td>Average</td></tr><tr><td>Ling-CL [NegSig]</td><td>59.3 ± 2.55</td><td>72.4 ± 0.40</td><td>79.1 ± 8.47</td><td>82.8 ± 8.35</td><td>92.2 ± 0.22</td><td>79.1 ± 1.55</td><td>75.3 ± 0.67</td><td>77.2 ± 3.17</td></tr><tr><td>Ling-CL [Gauss]</td><td>60.9 ± 1.41</td><td>73.0 ± 0.02</td><td>77.2 ± 8.08</td><td>83.5 ± 8.39</td><td>92.4 ± 0.27</td><td>82.9 ± 1.24</td><td>75.5 ± 0.41</td><td>77.9 ± 2.83</td></tr><tr><td>Ling-CL [Sig]</td><td>58.1 ± 0.17</td><td>64.6 ± 8.91</td><td>78.7 ± 8.87</td><td>83.0 ± 8.48</td><td>92.3 ± 0.01</td><td>82.3 ± 0.93</td><td>75.9 ± 0.10</td><td>76.4 ± 3.92</td></tr><tr><td>Loss-CL [NegSig]</td><td>59.0 ± 0.31</td><td>55.6 ± 0.64</td><td>68.1 ± 1.59</td><td>75.1 ± 0.05</td><td>91.6 ± 0.26</td><td>76.4 ± 5.70</td><td>75.1 ± 1.44</td><td>71.6 ± 1.43</td></tr><tr><td>Loss-CL [Sig]</td><td>49.7 ± 9.58</td><td>56.6 ± 0.37</td><td>66.8 ± 0.29</td><td>83.6 ± 8.37</td><td>90.9 ± 0.42</td><td>81.4 ± 0.61</td><td>73.3 ± 0.29</td><td>71.8 ± 2.85</td></tr><tr><td>Loss-CL [Gauss]</td><td>49.4 ± 11.07</td><td>57.0 ± 1.29</td><td>67.2 ± 1.41</td><td>75.1 ± 0.52</td><td>91.8 ± 0.12</td><td>80.5 ± 2.08</td><td>74.5 ± 0.07</td><td>70.8 ± 2.37</td></tr><tr><td>Sampling</td><td>49.9 ± 10.00</td><td>64.6 ± 8.89</td><td>67.9 ± 0.03</td><td>83.2 ± 8.72</td><td>91.5 ± 0.07</td><td>82.6 ± 3.93</td><td>73.8 ± 1.23</td><td>73.4 ± 4.7</td></tr><tr><td>Competence</td><td>50.1 ± 11.27</td><td>63.4 ± 9.08</td><td>68.8 ± 0.64</td><td>74.7 ± 0.06</td><td>91.6 ± 0.03</td><td>84.0 ± 1.14</td><td>74.1 ± 0.39</td><td>72.4 ± 3.23</td></tr><tr><td>SL-CL</td><td>50.3 ± 10.05</td><td>55.8 ± 0.06</td><td>67.7 ± 1.27</td><td>82.6 ± 8.35</td><td>93.1 ± 0.00</td><td>81.6 ± 0.72</td><td>75.2 ± 0.26</td><td>72.3 ± 2.96</td></tr><tr><td>WR-CL</td><td>50.9 ± 9.80</td><td>56.1 ± 0.53</td><td>68.4 ± 0.73</td><td>74.5 ± 0.16</td><td>91.5 ± 0.16</td><td>80.1 ± 0.81</td><td>75.2 ± 0.17</td><td>71.0 ± 1.77</td></tr><tr><td>SuperLoss</td><td>39.5 ± 0.14</td><td>56.9 ± 0.69</td><td>69.6 ± 0.50</td><td>75.2 ± 0.14</td><td>91.7 ± 0.26</td><td>77.8 ± 1.89</td><td>74.2 ± 0.15</td><td>69.3 ± 0.54</td></tr><tr><td>Concat</td><td>51.3 ± 9.83</td><td>64.3 ± 8.03</td><td>71.4 ± 0.51</td><td>75.2 ± 0.24</td><td>91.9 ± 0.14</td><td>81.8 ± 1.66</td><td>73.8 ± 0.91</td><td>72.8 ± 3.05</td></tr><tr><td>Data Selection</td><td>46.8 ± 6.12</td><td>55.1 ± 1.71</td><td>66.6 ± 1.49</td><td>74.4 ± 0.49</td><td>91.5 ± 0.30</td><td>79.6 ± 1.03</td><td>75.5 ± 0.52</td><td>69.9 ± 1.67</td></tr><tr><td>No-CL</td><td>51.7 ± 8.21</td><td>57.0 ± 0.22</td><td>70.0 ± 0.45</td><td>83.3 ± 8.42</td><td>83.7 ± 8.22</td><td>82.1 ± 0.51</td><td>74.0 ± 0.14</td><td>71.7 ± 3.74</td></tr></table>
+
+Table 1: Balanced accuracy by linguistic index (Word rarity). Accuracy is the metric for all datasets except CoLA and GED, CoLA uses Matthew's correlation and GED uses  $F_{\beta = 0.5}$  score. Ling-CL uses aggregate linguistic complexity as a difficulty score we create, and Loss-CL uses the average loss of a sample throughout a full training.
+
+# 3.6 Enhanced Linguistic Performance
+
+Tables 1 show the performance of different models when test samples are grouped based on word rarity. The results show that the performance of the baseline models severely drops compared to standard training (No-CL). This is while our Ling-CL approach results in 4.5 absolute points improvement in accuracy over the best-performing baseline averaged across tasks, owing to its effective use of linguistic indices. Appendix D shows the overall results on the entire test sets, and results when test samples are grouped based on their loss; we use loss because it is a widely-used measure of difficulty in curriculum learning. These groupings allow for a detailed examination of the model's performance across samples with varying difficulty, providing insights into the strengths and weaknesses of models. For example, the performance on SNLI varies from 89.8 to 90.6. However, when word rarity is used to group data based on difficulty, the performance range significantly drops from 74.4 to 83.6, indicating the importance of the proposed measure of evaluation. We observe that such grouping does not considerably change the performance on ANLI, which indicates the high quality of the dataset. In addition, it increases model performance on AN-Pair and GED, which indicates a greater prevalence of harder examples in these datasets.
+
+On average, the optimization approach outperforms correlation by  $1.6\% \pm 1.9\%$  accuracy in our experiments. Also notably, on average, the argmax index aggregation outperforms the weighted average by  $1.9\% \pm 1.9\%$ , and the filtered indices outperform the full list of indices by  $1.4\% \pm 1.1\%$ .
+
+# 3.7 Learning Dynamics for NLP Tasks
+
+Identification of Key Linguistic Indices We analyze the linguistic indices that most contribute to learning NLP tasks. For this purpose, we use the evaluation approach described in §3.4 for computing balanced accuracy according to linguistic indices. Table 2 shows the top three important linguistic indices for each dataset as identified by our optimization algorithm using the Gaussian curriculum. Importance is measured by the average  $\rho$  value. Early, middle, and late divide the training progress into three equal thirds. The top index in the early stage is the index with the highest average  $\rho$  during the first  $33.3\%$  of training. The top indices are those that most accurately estimate the true difficulty of samples, as they should highly correlate with validation loss.
+
+Table 2 shows that different indices are important for different tasks. This means that it is not possible to use a single set of linguistic indices as a general text difficulty score, important indices can be identified for each task, which can be achieved by our index importance estimation approach (§2.1) and evaluation metric (§3.4).
+
+Analysis of Linguistic Indices for Grammar Tasks We consider the grammar tasks for analysis. For AN-Pairs (adjective-noun pair), during the early stage, the top indices are the number of tokens per sentence, age of acquisition (AoA) of words, and mean length of sentences. This is meaningful because longer sentences might introduce modifiers or sub-clauses that can create ambiguity or make it more challenging to discern the intended adjective-noun relationship accurately. Regarding AoA, words that are acquired later in life or belong
+
+<table><tr><td></td><td>Early</td><td>Middle</td><td>Late</td></tr><tr><td rowspan="3">AN-Pairs</td><td># Tokens per sentence</td><td>Lemmas age of acquisition</td><td># Adverbs per sentence</td></tr><tr><td>Lemmas age of acquisition</td><td># Tokens per sentence</td><td>Corrected TTR</td></tr><tr><td>Mean sentence length</td><td># Adverbs per sentence</td><td>Nouns to adjective ratio</td></tr><tr><td rowspan="3">GED</td><td># Tokens per sentence</td><td colspan="2">Corrected noun variation</td></tr><tr><td rowspan="2"># Nouns per sentence</td><td># Nouns per sentence</td><td># Tokens per sentence</td></tr><tr><td># Tokens per sentence</td><td># Nouns per sentence</td></tr><tr><td rowspan="3">RTE</td><td colspan="3">Ratio of Adverbs to Verbs (P)</td></tr><tr><td colspan="2">Ratio of Subordinating Conjunctions to Verbs (P)</td><td>Adverb Variation (P)</td></tr><tr><td colspan="2">Verb sophistication (P)</td><td>Adverbs per sentence (P)</td></tr><tr><td rowspan="3">ANLI</td><td>Lexical verb variation (P)</td><td colspan="2">Function words per sentence (H)</td></tr><tr><td colspan="2">Unique Entities (P)</td><td>Log Tokens per log sentences</td></tr><tr><td colspan="2">Unique Entities per token (P)</td><td></td></tr><tr><td rowspan="3">SST-2</td><td colspan="3"># Complex nominals</td></tr><tr><td colspan="3">Noun variation</td></tr><tr><td>Ratio of nouns to verbs</td><td colspan="2">Verb variation</td></tr><tr><td rowspan="3">CoLA</td><td rowspan="3"># Function words</td><td colspan="2"># Coordinating Conjunctions</td></tr><tr><td colspan="2">Number of T-units</td></tr><tr><td colspan="2">T-units per sentence</td></tr><tr><td rowspan="3">SNLI</td><td colspan="3">Lemmas age of acquisition (P)</td></tr><tr><td colspan="3">Linsear Write Formula Score (P)</td></tr><tr><td colspan="3">Gunning Fog Count Score (P)</td></tr></table>
+
+Table 2: Top three important linguistic indices at each stage of learning. For datasets with a premise (P) and hypothesis (H), they are indicated in parentheses.
+
+![](images/0639a3f47452340d0cd54888c9215a5dfa2880821e895fdb1a2d3c27002cbbf3.jpg)  
+(a) # Unique words
+
+![](images/e0980fd2dc240e67e1fa38800ca29b3b995888484c7a033c697f2cb9d2a59548.jpg)  
+(b) # Unique sophisticated words  
+Figure 4: The progression of the estimated importance factors  $\rho$ , and balanced accuracy for groups of linguistic indices.
+
+![](images/4c5a7114f84813ec9a2aca836e358b3b09f215b26062fdd70bc71de590cc3ed9.jpg)  
+(c) # Verb sophistication
+
+to more specialized domains might pose challenges in accurately judging the correct usage of adjectivenoun pairs because of their varying degrees of familiarity and potential difficulty associated with specific vocabulary choices.
+
+During the middle stage the AoA increases in importance and remains challenging to the model, the number of adverbs per sentence increases in rank and joins the top three indices. In the context of adjective-noun pairs, the presence of multiple adverbs in a sentence can potentially affect the interpretation and intensity of the adjective's meaning. This is because adverbs often modify verbs, adjectives, or other adverbs in sentences. In addition, depending on the specific adverbs used, they may enhance, weaken, or alter the intended relationship between the adjective and the noun. Moreover, the presence of several adverbs can simply introduce potential challenges in identifying and correctly interpreting the relationship between adjectives and
+
+nouns due to increasing syntactic complexity.
+
+In the third stage, the number of adverbs per sentence becomes the top important index, while AoA and the number of tokens per sentence drop out of the top three. In the early stage, AoA and the number of tokens has  $\rho$  values of 0.168 and 0.164, respectively. In the late stage, they drop to 0.11 and 0.13, while the number of adverbs per sentence is 0.138 early, and increases to 0.181 in the late stage. We see that indices may become dominant not only by increasing their  $\rho$  value but also by waiting for other indices to drop down when they have been learned by the model. Therefore, Ling-CL can determine the order to learn linguistic indices, and then learn them sequentially.
+
+Regarding GED, noun variation is the dominant index throughout the training process. Such variation is important because it affects syntactic agreement, subject-verb agreement, modifier placement, and determiner selection. These factors affect gram
+
+matical consistency and coherence within the sentence structure, leading to the importance of noun variation throughout the training process.
+
+Dominant Indices for CoLA Task Regarding CoLA, the number of function words and coordinating conjunctions indices are the dominant indices at the early stage, and middle and late stages of training respectively. These words are crucial in establishing the syntactic structure of a sentence. They directly contribute to agreement and references, coherence, and adherence to grammar rules. We note that T-units (independent/main clauses clauses with their associated subordinate clauses) are higher-order linguistic constituents that provide information about the dependency relations between sub-constituents, and the overall coherence of sentences. Indices related to T-units are among the top three crucial indices.
+
+Trends and Relationships between  $\rho$  and Balanced Accuracy We use the GED dataset (§3.1) to analyze the trends of  $\rho$  throughout training, and the relation between  $\rho$  and balanced accuracy. Figure 4 shows the progression of  $\rho$  with the progression of balanced accuracy for selected linguistic indices. This figure is produced using No-CL. We observe across several indices that  $\rho$  is high when balanced accuracy is low, indicating that the index is challenging to the model and therefore used for learning with a high  $\rho$ , and decreases as the index is learned. However, Figure 4a shows that it is not necessary that when balanced accuracy increases  $\rho$  would decrease. In this case, it means that the model is performing relatively well on the index, but the index remains predictive of loss. So, although the average performance increased, the variation in performance among different values of the index remains high. We find that numerous indices follow the same trend of  $\rho$ . In Appendix B, we propose a method for clustering  $\rho$  to effectively uncover patterns and similarities in the learning of different indices. However, further analysis of the dynamics of  $\rho$  is the subject of our future work.
+
+In addition, we find that the rank of top indices is almost constant throughout the training. This quality may be useful in creating an approach that gathers the indices rankings early on and utilizes them for training. Appendix E lists influential indices by their change in  $\rho$  across stages of training. We note that the "number of input sentences" index is the least important metric because the index is al
+
+most constant across samples— $75\%$  of the samples consist of a single sentence in the datasets.
+
+# 4 Conclusion and Future Work
+
+We propose a new approach to linguistic curriculum learning. Our approach estimates the importance of multiple linguistic indices and aggregates them, provides effective difficulty estimates through correlation and optimization methods, and introduces novel curricula for using difficulty estimates, to uncover the underlying linguistic knowledge that NLP models learn during training. Furthermore, we present a method for a more accurate and fair evaluation of computational models for NLP tasks according to linguistic indices. Furthermore, the estimated importance factors present insights about each dataset and NLP task, the linguistic challenges contained within each task, and the factors that most contribute to model performance on the task. Further analysis of such learning dynamics for each NLP task will shed light on the linguistic capabilities of computational models at different stages of their training.
+
+Our framework and the corresponding tools serve as a guide for assessing linguistic complexity for various NLP tasks and uncover the learning dynamics of the corresponding NLP models during training. While we conducted our analysis on seven tasks and extracted insights on the key indices for each task, NLP researchers have the flexibility to either build on our results or apply our approach to other NLP tasks to extract relevant insights. Promising areas for future work include investigations on deriving optimal linguistic curriculum tailored for each NLP task; examining and enhancing linguistic capabilities of different computational models, particularly with respect to linguistically complex inputs; and developing challenge datasets that carry a fair distribution of linguistically complex examples for various NLP tasks. In addition, future work could study why specific indices are important, how they connect to the linguistic challenges of each task, and how different linguistic indices jointly contribute to learning a target task. We expect other aggregation functions, such as log-average, exponential-average, and probabilistic selection of maximum, to be effective approaches for difficulty estimation based on validation loss. Finally, other variations of the proposed Gaussian curriculum could be investigated for model improvement.
+
+# 5 Limitations
+
+Our work requires the availability of linguistic indices, which in turn requires expert knowledge. Such availability requirements may not be fulfilled in many languages. Nevertheless, some linguistic complexity indices are language independent, such as the commonly-used "word rarity" measure, which facilitates extending our approach to other languages. Moreover, our approach relies on the effectiveness of specific linguistic complexity indices for target tasks and datasets employed for evaluation; different linguistic complexity indices may not capture all aspects of linguistic complexity and may yield different results for the same task or dataset. In addition, the incorporation of linguistic complexity indices and the generation of data-driven curricula can introduce additional computational overhead during the training process. Finally, our approach does not provide insights into the interactions between linguistic indices during training.
+
+# References
+
+Sweta Agrawal and Marine Carpuat. 2022. An imitation learning curriculum for text editing with non-autoregressive models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7550-7563, Dublin, Ireland. Association for Computational Linguistics.  
+Hadi Amiri. 2019. Neural self-training through spaced repetition. 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 21-31, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Hadi Amiri, Timothy Miller, and Guergana Savova. 2017. Repeat before forgetting: Spaced repetition for efficient and effective training of neural networks. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2401-2410.  
+Devansh Arpit, Stanisław Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, et al. 2017. A closer look at memorization in deep networks. In International conference on machine learning, pages 233-242. PMLR.  
+Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. 2009. Curriculum learning. In ACM International Conference Proceeding Series, volume 382, pages 1-8, New York, New York, USA. ACM Press.
+
+Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326.  
+Bram Bulté and Alex Housen. 2012. Defining and operationalising 12 complexity. *Dimensions of L2 performance and proficiency: Complexity, accuracy and fluency in SLA*, pages 23-46.  
+Thibault Castells, Philippe Weinzaepfel, and Jerome Revaud. 2020. Superloss: A generic loss for robust curriculum learning. Advances in Neural Information Processing Systems, 33:4308-4319.  
+Birgit Harley and Mary Lou King. 1989. Verb lexis in the written compositions of young 12 learners. Studies in Second Language Acquisition, 11(4):415-439.  
+Alex Housen, Bastien De Clercq, Folkert Kuiken, and Ineke Vedder. 2019. Multiple approaches to complexity in second language research. Second language research, 35(1):3-21.  
+Kenneth Hyltenstam. 1988. Lexical characteristics of near-native second-language learners of swedish. Journal of Multilingual & Multicultural Development, 9(1-2):67-84.  
+Borna Jafarpour, Dawn Sepehr, and Nick Pogrebnyakov. 2021. Active curriculum learning. In Proceedings of the First Workshop on Interactive Learning for Natural Language Processing, pages 40-45.  
+Ekaterina Kochmar and Ted Briscoe. 2014. Detecting learner errors in the choice of content words using compositional distributional semantics. In Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1740-1751.  
+Julia Kreutzer, David Vilar, and Artem Sokolov. 2021. Bandits don't follow rules: Balancing multi-facet machine translation with multi-armed bandits. In *Findings of the Association for Computational Linguistics: EMNLP* 2021, pages 3190–3204, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+M Kumar, Benjamin Packer, and Daphne Koller. 2010. Self-paced learning for latent variable models. Advances in neural information processing systems, 23:1189-1197.  
+Michael Steinbach George Karypis Vipin Kumar, M Steinbach, and G Karypis. 2000. A comparison of document clustering techniques. Department of Computer Science and Engineering, University of Minnesota.  
+John P Lalor and Hong Yu. 2020. Dynamic data selection for curriculum learning via ability estimation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, volume 2020, page 545. NIH Public Access.
+
+Bruce W Lee, Yoo Sung Jang, and Jason Lee. 2021. Pushing on text readability assessment: A transformer meets handcrafted linguistic features. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 10669-10686.  
+Moira Linnarud. 1986. Lexis in composition: A performance analysis of Swedish learners' written English. 74. CWK Gleerup.  
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692.  
+Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Conference on Learning Representations.  
+Xiaofei Lu. 2010. Automatic analysis of syntactic complexity in second language writing. International Journal of Corpus Linguistics, 15:474-496.  
+Xiaofei Lu. 2012. The Relationship of Lexical Richness to the Quality of ESL Learners' Oral Narratives. Source: The Modern Language Journal, 96(2):190-208.  
+Adyasha Maharana and Mohit Bansal. 2022. On curriculum learning for commonsense reasoning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 983-992, Seattle, United States. Association for Computational Linguistics.  
+David Malvern, Brian Richards, Ngoni Chipere, and Pilar Durán. 2004. Lexical diversity and language development. Springer.  
+Erica McClure. 1991. A comparison of lexical strategies in l1 and l2 written english narratives. *Pragmatics and language learning*, 2:141-154.  
+Gerard McKee, David Malvern, and Brian Richards. 2000. Measuring vocabulary diversity using dedicated software. *Literary and linguistic computing*, 15(3):323-338.  
+Changping Meng, Muhao Chen, Jie Mao, and Jennifer Neville. 2020. Readnet: A hierarchical transformer framework for web article readability analysis. In *European Conference on Information Retrieval*, pages 33-49. Springer.  
+Tasnim Mohiuddin, Philipp Koehn, Vishrav Chaudhary, James Cross, Shruti Bhosale, and Shafiq Joty. 2022. Data selection curriculum for neural machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, Online. Association for Computational Linguistics.
+
+Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and Douwe Kiela. 2020. Adversarial nli: A new benchmark for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4885-4901.  
+Joel Nishimura. 2018. Critically slow learning in flashcard learning models. *Chaos*, 28(8):083115.  
+Felicity O'Dell, John Read, Michael McCarthy, et al. 2000. Assessing vocabulary. Cambridge university press.  
+Lourdes Ortega. 2003. Syntactic complexity measures and their relationship to l2 proficiency: A research synthesis of college-level l2 writing. Applied linguistics, 24(4):492-518.  
+Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabas Poczos, and Tom M Mitchell. 2019. Competence-based curriculum learning for neural machine translation. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT, pages 1162-1172.  
+FJ Richards. 1959. A flexible growth function for empirical use. Journal of experimental Botany, 10(2):290-301.  
+Burr Settles and Brendan Meeder. 2016. A trainable spaced repetition model for language learning. In Proceedings of the 54th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1848-1858.  
+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 Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642.  
+Behzad Tabibian, Utkarsh Upadhyay, Abir De, Ali Zarezade, Bernhard Scholkopf, and Manuel Gomez-Rodriguez. 2019. Enhancing human learning via spaced repetition optimization. Proc. Natl. Acad. Sci. U. S. A., 116(10):3988-3993.  
+Mildred C Templin. 1957. Certain language skills in children: Their development and interrelationships, volume 10. JSTOR.  
+Nidhi Vakil and Hadi Amiri. 2023. Multiview competence-based curriculum learning for graph neural networks. In Proceedings of the 61th Annual Meeting of the Association for Computational Linguistics (ACL 2023). Association for Computational Linguistics.  
+Frans van der Sluis and Egon L van den Broek. 2010. Using complexity measures in information retrieval. In Proceedings of the third symposium on information interaction in context, pages 383-388.
+
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. 2018. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461.  
+Xinyi Wang, Hieu Pham, Paul Michel, Antonios Anastasopoulos, Jaime Carbonell, and Graham Neubig. 2020. Optimizing data usage via differentiable rewards. In International Conference on Machine Learning, pages 9983-9995. PMLR.  
+Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45.  
+Kate Wolfe-Quintero, Shunji Inagaki, and Hae-Young Kim. 1998. Second language development in writing: Measures of fluency, accuracy, and complexity. Honolulu, HI: University of Hawai'i, Second Language Teaching & Curriculum Center.  
+Xiaoxia Wu, Ethan Dyer, and Behnam Neyshabur. 2021. When do curricula work? In International Conference on Learning Representations (ICLR).  
+Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, and Yongdong Zhang. 2020. Curriculum learning for natural language understanding. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6095-6104.  
+Helen Yannakoudakis, Ted Briscoe, and Ben Medlock. 2011. A new dataset and method for automatically grading esol texts. In Proceedings of the 49th annual meeting of the association for computational linguistics: human language technologies, pages 180-189.  
+Guoxing Yu. 2010. Lexical diversity in writing and speaking task performances. Applied linguistics, 31(2):236-259.  
+Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. 2017. Understanding deep learning requires rethinking generalization. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.  
+Xuan Zhang, Pamela Shapiro, Gaurav Kumar, Paul McNamee, Marine Carpuat, and Kevin Duh. 2019. Curriculum learning for domain adaptation in neural machine translation. In Proceedings of NAACL-HLT, pages 1903-1915.
+
+Tianyi Zhou, Shengjie Wang, and Jeffrey Bilmes. 2020. Curriculum learning by dynamic instance hardness. Advances in Neural Information Processing Systems, 33:8602-8613.
+
+# A List of indices
+
+<table><tr><td># Unique words</td><td># Unique sophisticated words</td></tr><tr><td># Unique lexical words</td><td># Unique sophisticated lexical words</td></tr><tr><td># Total words</td><td># Total sophisticated words</td></tr><tr><td># Total lexical words</td><td># Total sophisticated lexical words</td></tr><tr><td>Lexical density</td><td>Lexical sophistication (total)</td></tr><tr><td>Lexical sophistication (unique)</td><td>Verb sophistication</td></tr><tr><td>Verb sophistication (squared numerator)</td><td>Verb sophistication (sqrt denominator)</td></tr><tr><td>Type-token ratio (TTR)</td><td>Mean TTR of all k word segments</td></tr><tr><td>Corrected TTR (sqrt(2N) denominator)</td><td>Root TTR (sqrt(N) denominator)</td></tr><tr><td>Log TTR</td><td>Uber index</td></tr><tr><td>Noun variation</td><td>Adjective variation</td></tr><tr><td>Adverb variation</td><td>(Ajd + Adv) variation</td></tr><tr><td>D Measure</td><td>Ratio of unique verbs</td></tr><tr><td>Verb variation with squared numerator</td><td>Verb variation with (sqrt(2N)) denominator</td></tr><tr><td>Verb variation over all lexical words</td><td>Unique words in first k tokens</td></tr><tr><td>Unique words in random k tokens</td><td>Unique words in random sequence of k tokens</td></tr></table>
+
+Table 3: Lexical indices  
+
+<table><tr><td># Words</td><td># Sentences</td></tr><tr><td># Verb phrases</td><td># Clauses</td></tr><tr><td># T-units</td><td># Dependent clauses</td></tr><tr><td># Complex T-units</td><td># Coordinate phrases</td></tr><tr><td># Complex nominals</td><td>Mean length of sentence</td></tr><tr><td>Mean length of T-unit</td><td>Mean unit of clause</td></tr><tr><td>Clauses per sentence</td><td>Verb phrases per T-unit</td></tr><tr><td>Clauses per T-unit</td><td>Dependent clause ratio</td></tr><tr><td>Dependent clause per T-unit</td><td>T-units per sentence</td></tr><tr><td>Complex T-unit ratio</td><td>Coordinate phrases per T-unit</td></tr><tr><td>Coordinate phrases per clause</td><td>Complex nominals per T-unit</td></tr><tr><td>Complex nominals per clause</td><td></td></tr></table>
+
+Table 4: Syntactic indices
+
+In our work we make use indices from Lu (2010), Lu (2012), and Lee et al. (2021). Table 3 lists the lexical indices (33 indices) and table 4 lists the syntactic indices (23 indices) that we use. For their full descriptions please refer to Lu (2010) and Lu (2012). In this section, we provide descriptions of a few relevant indices. Please refer to Lee et al. (2021) for the comprehensive list of lingfeat (185 indices) indices.
+
+TTR is the ratio of unique words in the text. D-measure is a modification to TTR that is not biased by sample size. Lexical words are nouns, verbs, adjectives, and adverbs. Sophisticated words are the unconventional words. We consider words beyond the 2000 most frequent words in the American National Corpus as sophisticated. Uber index is a transformation of TTR. SubtlexUS CDlow is a word frequency measure, specifically, "document frequency" of words starting with a lower case letter.
+
+# B Clustering  $\rho$
+
+![](images/3de7bc61ceae3b43a3f29f71eaa600e65b0c86fd6d4dd6bde76293a4f3e1b135.jpg)  
+(a) Pair-wise differences between  $\rho$
+
+![](images/29ad733e4f2a517128b93960a5b604578da29248681d89e75e07d0d315f73d53.jpg)  
+(b) Clustered differences map.
+
+![](images/3fd7be4d6c8e5291373c455f1fd322c131b76222e0f853ba16cba8d5085e2826.jpg)  
+Figure 5: Figure (a) shows mean absolute difference between  $\rho$  of each pair of linguistic indices, averaged over the whole training. Figure (b) is created by re-ordering the rows and columns of (a), such that groups of indices have minimal difference between them.
+
+![](images/37394aa9ce26f1f9a4b9f92d77bfa86364c7ad32961465b926c9ab3b17a54c6e.jpg)
+
+![](images/11e9bf57a92bd424feadf5f61077642de9f5cd9e8ba2eb092f355679efcf3238.jpg)
+
+![](images/3a897e579642f17ab2c6a170a30a04602520da129523a8bd9e5cba37c8c234c5.jpg)  
+(a) CoLA  $\rho$  
+(d) CoLA Balanced Accuracy  
+Figure 6: The progression of the estimated importance factors  $\rho$ , and balanced accuracy for groups of linguistic indices. Each pair of figures (a) and (d), (b) and (e), (c) and (f) share the legend. The solid line is the mean value of the group of lines, and the shaded area is the  $95\%$  confidence interval.
+
+![](images/52b830ba3a61d4101c2400e7beedbfeefb72974d65f9e36e4ae2f6eea3e6c29f.jpg)  
+(b) SNLI  $\rho$  
+(e) SNLI Balanced Accuracy
+
+![](images/3e4b86055ffa21a6bbc5864e566efb374314a121df8b567094a12d70159be89a.jpg)  
+(c) SST-2  $\rho$  
+(f) SST-2 Balanced Accuracy
+
+We observe that several indices follow the same patterns. Therefore, we devise a method to group indices that follow the same pattern of  $\rho$ . We compute the mean absolute difference between  $\rho$  of each pair of linguistic indices. Then, we cluster the groups of indices that all have a minimum distance between them. Appendix B displays the effect of clustering. Note that the common trends among lines in Figures 6a, 6b, and 6c are because they are all governed by the trend of the validation loss (using both optimization and correlation approaches). Figures 6d, 6e, and 6f show the trend of balanced accuracy for the same groups. The grouped balanced accuracy has a very high variance. It shows that indices with similar  $\rho$  do not have similar values of balanced accuracy. Moreover, it shows that indices with the highest  $\rho$  do not necessarily have the highest mean balanced accuracy. Furthermore, indices that have  $\rho = 0$  perform comparably to other indices, indicating that the model performs well according to such linguistic
+
+indices, despite them not being correlated with loss.
+
+Figure 5 illustrates the process of clustering together linguistic indices based on their matching  $\rho$  curves. We cluster the indices using hierarchical clustering with complete linkage using the flat clustering method $^{23}$ .
+
+# C Linguistic Complexity Indices
+
+We consider linguistic complexity in terms of variability and sophistication in productive vocabulary and grammatical structures in textual content. We employ a characterization of such complexity based on existing findings in language acquisition research (Wolfe-Quintero et al., 1998; Lu, 2010, 2012). Specifically, we obtain 56 complexity measures from Lu (2010) and Lu (2012), including lexical and syntactic measures. Additionally, we use 185 linguistic features from the lingfeat library (Lee et al., 2021), including semantic, lexical, syntactic, discourse, and traditional features. In total, we use 241 indices. For inputs that consist of a pair of sentences, we concatenate the indices for a total of 482 indices.
+
+# C.1 Lexical Complexity
+
+In terms of lexical complexity, we consider three dimensions: lexical density, sophistication, and variation described below:
+
+Lexical density: is quantified by the ratio of the number of open-class words to the total number of words in a given text. Texts with higher lexical density are expected to be more complex as they contain larger amounts of information-carrying words.
+
+Lexical sophistication: measures the proportion of sophisticated—relatively unusual or advanced—words in the input text (O'Dell et al., 2000), e.g., words not in the top K ( $\mathrm{K} = 5000$ ) frequent words in the target dataset or language. Example indices include the ratio of the number of sophisticated lexical words (Linnarud, 1986; Hyltenstam, 1988), sophisticated word types (Wolfe-Quintero et al., 1998) and sophisticated verb types (Harley and King, 1989) in texts, which include several variations as reported in Appendix A, Table 3. We use the top K most frequent words of each dataset and consider different inflections of the same lemma as one type for computing lexical sophistication.
+
+Lexical Variation: refers to the diversity of vocabulary in a given text. Examples of such variations include type-token ratio (Templin, 1957) which is the ratio of the number of word types to the number of words in the text and several different variations of this metric (Malvern et al., 2004; McKee et al., 2000; McClure, 1991) including  $D$ -measure (Malvern et al., 2004), which determines lexical variation of an input text by finding the curve that best matches the actual curve of type-token ratio against tokens of the input.
+
+# C.2 Syntactic Complexity
+
+Syntactic complexity determines variability and sophistication with respect to grammatical structures. Simple sentences such as "the mouse ate the cheese" can be converted to their linguistically-complex counterparts, e.g., "the mouse the cat the dog bit chased ate the cheese," which are still well-formed sentences but force readers to suspend their partial understanding of the entire sentence by encountering subordinate clauses that substantially increase the cognitive load of the sentences. We employ syntactic complexity measures that quantify the length of production units at the clausal, sentential, or T-unit levels; indices that reflect the amount of subordination, e.g., T-unit complexity ratio (clauses per T-unit) or dependent clause ratio (dependent clauses per clause); indices that quantify the amount of coordination, e.g., number of coordinate phrases per clause, T-unit or complex T-unit; as well as those that quantify the range of surface and particular syntactic and morphological structures (e.g., frequency and variety of tensed forms or extent of affixation) (Wolfe-Quintero et al., 1998; Ortega, 2003). See Appendix A, Table 3.
+
+# D Full results
+
+Tables 5 and 6 show the overall performance and performance balanced by loss. Our Ling-CL approach results in 1.3 absolute points improvement in accuracy over the best-performing baseline averaged across tasks, balanced by loss.
+
+<table><tr><td></td><td>ANLI</td><td>COLA</td><td>RTE</td><td>SNLI</td><td>SST2</td><td>AN-Pairs</td><td>GED</td><td>Average</td></tr><tr><td>Ling-CL [NegSig]</td><td>49.6 ± 0.44</td><td>62.8 ± 0.19</td><td>81.0 ± 0.10</td><td>90.0 ± 0.10</td><td>95.0 ± 0.06</td><td>74.0 ± 1.04</td><td>71.6 ± 0.37</td><td>74.9 ± 0.33</td></tr><tr><td>Ling-CL [Gauss]</td><td>51.5 ± 0.39</td><td>64.5 ± 0.35</td><td>79.8 ± 0.21</td><td>90.4 ± 0.01</td><td>95.2 ± 0.00</td><td>75.0 ± 2.08</td><td>71.9 ± 0.20</td><td>75.5 ± 0.46</td></tr><tr><td>Ling-CL [Sig]</td><td>51.6 ± 0.17</td><td>65.1 ± 0.71</td><td>81.9 ± 0.21</td><td>90.2 ± 0.07</td><td>95.1 ± 0.11</td><td>74.5 ± 2.60</td><td>71.8 ± 0.05</td><td>75.7 ± 0.56</td></tr><tr><td>Loss-CL [NegSig]</td><td>52.7 ± 0.03</td><td>63.5 ± 0.66</td><td>80.5 ± 0.36</td><td>90.5 ± 0.12</td><td>95.0 ± 0.00</td><td>69.3 ± 1.54</td><td>72.2 ± 0.03</td><td>74.8 ± 0.39</td></tr><tr><td>Loss-CL [Gauss]</td><td>51.9 ± 0.06</td><td>65.7 ± 1.66</td><td>82.3 ± 1.08</td><td>90.3 ± 0.38</td><td>94.7 ± 0.11</td><td>72.9 ± 1.04</td><td>72.1 ± 0.08</td><td>75.7 ± 0.63</td></tr><tr><td>Loss-CL [Sig]</td><td>51.1 ± 1.20</td><td>65.0 ± 0.57</td><td>78.9 ± 1.26</td><td>90.4 ± 0.26</td><td>94.9 ± 0.06</td><td>74.3 ± 1.93</td><td>71.7 ± 0.15</td><td>75.2 ± 0.78</td></tr><tr><td>Sampling</td><td>49.3 ± 0.33</td><td>64.0 ± 1.02</td><td>78.4 ± 0.70</td><td>90.3 ± 0.14</td><td>94.6 ± 0.10</td><td>72.9 ± 4.17</td><td>71.1 ± 0.07</td><td>74.4 ± 0.93</td></tr><tr><td>Competence</td><td>50.7 ± 0.03</td><td>63.2 ± 0.63</td><td>77.8 ± 0.73</td><td>90.3 ± 0.17</td><td>95.0 ± 0.11</td><td>74.5 ± 0.52</td><td>71.2 ± 0.10</td><td>74.7 ± 0.33</td></tr><tr><td>SL-CL</td><td>51.1 ± 0.95</td><td>63.6 ± 0.42</td><td>80.3 ± 2.35</td><td>90.1 ± 0.11</td><td>94.9 ± 0.06</td><td>69.3 ± 1.56</td><td>71.9 ± 0.12</td><td>74.5 ± 0.80</td></tr><tr><td>WR-CL</td><td>51.9 ± 0.16</td><td>64.5 ± 1.00</td><td>80.0 ± 0.54</td><td>90.2 ± 0.18</td><td>94.4 ± 0.23</td><td>64.6 ± 4.17</td><td>71.7 ± 0.32</td><td>73.9 ± 0.94</td></tr><tr><td>Concat</td><td>52.2 ± 0.30</td><td>65.2 ± 0.33</td><td>81.8 ± 0.90</td><td>90.6 ± 0.00</td><td>94.6 ± 0.11</td><td>72.4 ± 2.60</td><td>71.7 ± 0.21</td><td>75.5 ± 0.64</td></tr><tr><td>SuperLoss</td><td>51.7 ± 0.21</td><td>64.3 ± 0.99</td><td>80.5 ± 1.04</td><td>90.5 ± 0.15</td><td>94.8 ± 0.13</td><td>67.7 ± 2.08</td><td>71.7 ± 0.10</td><td>74.5 ± 0.67</td></tr><tr><td>Data Selection</td><td>48.5 ± 0.94</td><td>59.4 ± 0.03</td><td>79.2 ± 0.90</td><td>89.8 ± 0.12</td><td>94.2 ± 0.17</td><td>72.4 ± 1.56</td><td>71.1 ± 0.13</td><td>73.5 ± 0.55</td></tr><tr><td>No-CL</td><td>51.4 ± 0.06</td><td>64.1 ± 0.42</td><td>81.4 ± 0.77</td><td>90.4 ± 0.17</td><td>94.9 ± 0.10</td><td>71.9 ± 2.08</td><td>72.0 ± 0.40</td><td>75.2 ± 0.57</td></tr></table>
+
+Table 5: Overall performance of each approach. Unlike Tables 1 and 6, this table presents the standard un-balanced accuracy.  
+
+<table><tr><td></td><td>ANLI</td><td>COLA</td><td>RTE</td><td>SNLI</td><td>SST2</td><td>AN-Pairs</td><td>GED</td><td>Average</td></tr><tr><td>Ling-CL [NegSig]</td><td>23.2 ± 4.61</td><td>27.1 ± 0.66</td><td>45.4 ± 6.23</td><td>26.3 ± 0.44</td><td>25.7 ± 1.31</td><td>36.7 ± 4.98</td><td>73.9 ± 0.83</td><td>36.9 ± 2.72</td></tr><tr><td>Ling-CL [Gauss]</td><td>21.0 ± 1.89</td><td>26.7 ± 1.62</td><td>37.1 ± 0.52</td><td>42.9 ± 0.12</td><td>22.7 ± 2.91</td><td>34.7 ± 6.92</td><td>73.9 ± 0.36</td><td>37.0 ± 2.05</td></tr><tr><td>Ling-CL [Sig]</td><td>22.2 ± 0.51</td><td>26.4 ± 0.14</td><td>38.5 ± 0.27</td><td>35.2 ± 8.99</td><td>25.0 ± 0.58</td><td>38.4 ± 3.26</td><td>74.6 ± 0.11</td><td>37.2 ± 1.98</td></tr><tr><td>Loss-CL [NegSig]</td><td>20.9 ± 1.59</td><td>26.7 ± 0.42</td><td>37.7 ± 1.03</td><td>33.8 ± 8.37</td><td>23.8 ± 3.13</td><td>29.8 ± 3.76</td><td>73.4 ± 0.84</td><td>35.2 ± 2.73</td></tr><tr><td>Loss-CL [Sig]</td><td>19.5 ± 0.21</td><td>27.6 ± 0.51</td><td>36.3 ± 2.43</td><td>34.6 ± 7.58</td><td>24.2 ± 3.59</td><td>31.6 ± 1.95</td><td>72.6 ± 0.01</td><td>35.2 ± 2.33</td></tr><tr><td>Loss-CL [Gauss]</td><td>20.6 ± 0.99</td><td>26.3 ± 1.29</td><td>40.8 ± 3.89</td><td>26.2 ± 0.18</td><td>23.1 ± 1.63</td><td>31.7 ± 3.46</td><td>73.0 ± 0.38</td><td>34.5 ± 1.69</td></tr><tr><td>Sampling</td><td>19.0 ± 0.12</td><td>26.2 ± 0.36</td><td>32.9 ± 0.75</td><td>42.8 ± 0.50</td><td>24.1 ± 3.57</td><td>33.6 ± 1.09</td><td>72.8 ± 1.44</td><td>35.9 ± 1.12</td></tr><tr><td>Competence</td><td>21.5 ± 0.55</td><td>26.8 ± 1.84</td><td>36.9 ± 2.57</td><td>27.0 ± 0.34</td><td>27.1 ± 0.63</td><td>35.5 ± 2.01</td><td>73.4 ± 1.06</td><td>35.5 ± 1.29</td></tr><tr><td>SL-CL</td><td>21.3 ± 0.28</td><td>26.2 ± 0.00</td><td>31.6 ± 0.58</td><td>34.6 ± 8.35</td><td>26.7 ± 0.00</td><td>32.3 ± 2.59</td><td>74.8 ± 0.77</td><td>35.4 ± 1.8</td></tr><tr><td>WR-CL</td><td>19.3 ± 0.52</td><td>26.4 ± 0.69</td><td>35.7 ± 0.51</td><td>26.1 ± 0.03</td><td>24.2 ± 1.26</td><td>29.0 ± 0.73</td><td>73.5 ± 0.61</td><td>33.5 ± 0.62</td></tr><tr><td>SuperLoss</td><td>20.7 ± 1.77</td><td>29.0 ± 2.51</td><td>35.6 ± 0.99</td><td>26.0 ± 0.38</td><td>23.9 ± 0.83</td><td>25.9 ± 2.28</td><td>72.8 ± 0.03</td><td>33.4 ± 1.26</td></tr><tr><td>Concat</td><td>21.7 ± 0.17</td><td>28.1 ± 0.85</td><td>37.8 ± 3.09</td><td>26.6 ± 0.33</td><td>27.4 ± 0.29</td><td>33.4 ± 1.20</td><td>72.5 ± 0.30</td><td>35.4 ± 0.89</td></tr><tr><td>Data Selection</td><td>19.5 ± 0.76</td><td>25.7 ± 0.28</td><td>33.8 ± 2.23</td><td>25.9 ± 0.65</td><td>22.6 ± 1.24</td><td>30.6 ± 2.01</td><td>74.3 ± 0.28</td><td>33.2 ± 1.06</td></tr><tr><td>No-CL</td><td>20.1 ± 1.41</td><td>28.1 ± 0.85</td><td>37.8 ± 3.09</td><td>26.6 ± 0.33</td><td>27.4 ± 0.29</td><td>31.8 ± 0.28</td><td>72.2 ± 0.46</td><td>34.9 ± 0.96</td></tr></table>
+
+Table 6: Balanced accuracy by loss.
+
+# E Index Importance Changes
+
+<table><tr><td rowspan="4">AN-Pairs</td><td>Index</td><td>Change Stages</td><td>Magnitude</td></tr><tr><td>Corrected TTR</td><td>Medium to late</td><td>8.70%</td></tr><tr><td>Ratio of nx entity grid transitions</td><td>Medium to late</td><td>7.89%</td></tr><tr><td>Semantic Richness</td><td>Medium to late</td><td>7.86%</td></tr><tr><td rowspan="3">GED</td><td>Ratio of nx entity grid transitions</td><td>Medium to late</td><td>3.18%</td></tr><tr><td>Lexical sophistication</td><td>Medium to late</td><td>2.26%</td></tr><tr><td>Ratio of Coordinating Conjunction to Adjectives</td><td>Medium to late</td><td>1.84%</td></tr><tr><td rowspan="3">SNLI</td><td>Ratio of Subordinating Conjunction to Adverbs</td><td>Early to late</td><td>1.00%</td></tr><tr><td>Noun-subject transitions</td><td>Early to late</td><td>0.79%</td></tr><tr><td>Number of topics (Weebit-based)</td><td>Early to late</td><td>0.77%</td></tr><tr><td rowspan="3">ANLI</td><td>Verb sophistication</td><td>Medium to late</td><td>0.98%</td></tr><tr><td>T-unit length</td><td>Medium to late</td><td>0.96%</td></tr><tr><td>Log tokens over log sentences</td><td>Early to medium</td><td>0.94%</td></tr><tr><td rowspan="3">CoLA</td><td>Object-noun transitions</td><td>Medium to late</td><td>0.78%</td></tr><tr><td>TTR</td><td>Early to late</td><td>0.75%</td></tr><tr><td># Clauses</td><td>Medium to late</td><td>0.75%</td></tr><tr><td rowspan="3">RTE</td><td>Adverb to adjective ratio</td><td>Early to late</td><td>3.30%</td></tr><tr><td># T-units</td><td>Early to late</td><td>3.27%</td></tr><tr><td>Mean sentence length</td><td>Early to late</td><td>3.15%</td></tr><tr><td rowspan="3">SST-2</td><td>Noun-subject transitions</td><td>Early to late</td><td>1.76%</td></tr><tr><td>Coordinating conjunction per sentence</td><td>Early to late</td><td>1.75%</td></tr><tr><td>Verbs per token</td><td>Medium to late</td><td>1.72%</td></tr></table>
+
+Table 7: Top three moving linguistic indices.
+
+Table 7 shows indices with maximum change in their  $\rho$ s between any two stages. Only the relative differences and ranking of  $\rho$  values are important. Therefore, the table displays relative changes in the magnitude of the importance factors. The indices with a large change magnitude indicate that they are either influential at an early stage of training and drop in importance at the later stage, or vise versa. See our analysis on these indices in §3.7.

2023/Ling-CL_ Understanding NLP Models through Linguistic Curricula/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ac03b2d2c4df50d67206d996532ebedb0be7a5dd1cfd47965779ca7a518b666
+size 1203379

2023/Linking Surface Facts to Large-Scale Knowledge Graphs/9a8770e9-9bcc-4ad5-8318-6ea414c765f4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4dd7403792ca2002721978a661a6e0c00214bb244c08766a323d44200fd2bad
+size 645006

2023/Linking Surface Facts to Large-Scale Knowledge Graphs/full.md

@@ -0,0 +1,396 @@
+# Linking Surface Facts to Large-Scale Knowledge Graphs
+
+Gorjan Radevski $^{1,2}$ , Kiril Gashteovski $^{1,3}$ , Chia-Chien Hung $^{1}$ , Carolin Lawrence $^{1}$ , Goran Glavaš $^{4}$
+
+<sup>1</sup>NEC Laboratories Europe, Heidelberg, Germany; <sup>2</sup>KU Leuven, Leuven, Belgium;
+
+$^{3}$ CAIR, Ss. Cyril and Methodius University, Skopje, North Macedonia;
+
+$^{4}$ CAIDAS, University of Würzburg, Würzburg, Germany
+
+$^{1}$ firstname_lastname@neclab.eu,  $^{2}$ gorjan.radevski@kuleuven.be,
+
+4goran.glavas@uni-wuerzburg.de
+
+# Abstract
+
+Open Information Extraction (OIE) methods extract facts from natural language text in the form of ("subject"; "relation"; "object") triples. These facts are, however, merely surface forms, the ambiguity of which impedes their downstream usage; e.g., the surface phrase "Michael Jordan" may refer to either the former basketball player or the university professor. Knowledge Graphs (KGs), on the other hand, contain facts in a canonical (i.e., unambiguous) form, but their coverage is limited by a static schema (i.e., a fixed set of entities and predicates). To bridge this gap, we need the best of both worlds: (i) high coverage of free-text OIEs, and (ii) semantic precision (i.e., monosemy) of KGs. In order to achieve this goal, we propose a new benchmark with novel evaluation protocols that can, for example, measure fact linking performance on a granular triple slot level, while also measuring if a system has the ability to recognize that a surface form has no match in the existing KG. Our extensive evaluation of several baselines shows that detection of out-of-KG entities and predicates is more difficult than accurate linking to existing ones, thus calling for more research efforts on this difficult task. We publicly release all resources (data, benchmark and code)<sup>1</sup>.
+
+# 1 Introduction
+
+Open Information Extraction (OIE) methods extract surface ("subject"; "relation"; "object")- triples from natural language text in a schema-free manner (Banko et al., 2007). For example, given the sentence "Michael Jordan, who grew up in Wilmington, played for Chicago Bulls", an OIE system should extract the triples (i.e., surface facts):  $t_1 =$  ("Michael Jordan"; "played for"; "Chicago Bulls") and  $t_2 =$  ("Michael Jordan"; "grew up in"; "Wilmington"). The output of such systems is used in a diverse range of downstream tasks, including
+
+summmarization (Ribeiro et al., 2022), question answering (QA) (Wu et al., 2022), event extraction (Dukic et al., 2023), text clustering (Viswanathan et al., 2023) and video grounding (Nan et al., 2021). OIE triples, however, consist of surface-form entities and relations, which are frequently ambiguous (e.g., in  $t_1$  and  $t_2$ , the entity mention "Michael Jordan" may refer to several entities, e.g. the basketball player or the computer scientist). Resolving such ambiguities, by linking the OIE triple slots to inventories of unambiguous concepts, improves their downstream utility, e.g., in QA (Saxena et al., 2020), automatic medical diagnosing (Li et al., 2022) or dialogue (Joko et al., 2021).
+
+Knowledge Graphs (KGs), on the other hand, are inventories of canonical facts in the form of (subject; predicate; object)-triples, where each slot is a unique (i.e., unambiguous) concept (Vrandecic, 2012). KGs are, however, limited by their own static and often hand-crafted schema (i.e., fixed set of entities and predicates). As a consequence, methods that directly extract canonical KG triples from text (Trisedya et al., 2019; Josifoski et al., 2022), preemptively discard any information outside of the reference KG schema. Acknowledging this, Ye et al. (2023) recently proposed schema-adaptable KG construction, with the goal of extracting information for a KG with an evolving schema. Ye et al. (2023) conclude that OIE methods indeed extract meaningful new knowledge for such KGs, however, they point precisely to the ambiguity of the surface forms as the major obstacle.
+
+To combine the best of both worlds, we need to bridge the gap between the schema-free (but ambiguous) surface facts extracted from text and the schema-fixed (but unambiguous) KG knowledge. However, existing benchmarks and models only partially address the problem. One line of work (Zhang et al., 2019; Jiang et al., 2021; Wood et al., 2021) assumes a setting, which is arguably unrealistic, because the OIE entity slots are a priori
+
+![](images/8acf4e26ac2f95c194832928e0343b3ca7f6cc0528509398b2c7f61b4494266c.jpg)  
+Figure 1: Open Information Extractions (OIEs) represent free-text surface-form triples, which may be ambiguous. Knowledge Graphs (KGs) represent canonical, unambiguous concepts, yet are limited by a hand-crafted schema. By linking OIEs to KG facts we bridge the gap between the schema-free (but ambiguous) surface facts extracted from text and the schema-fixed (but unambiguous) KG knowledge. Our benchmark addresses all aspects of the OIE-to-KG linking. In transductive evaluation, the KG facts consist of entities seen during training, while in inductive evaluation the KG facts contain entities strictly outside of the training data. Polysemous OIEs match to multiple KG concepts (e.g., "Michael Jordan"). Finally, some OIEs might refer to Out-of-KG entities or predicates.
+
+linked to KG entities and thus only addresses relation linking. A second line of research (Trisedya et al., 2019; Cabot and Navigli, 2021; Josifoski et al., 2022; Sakor et al., 2020; Elsahar et al., 2018) entails canonical KG triples directly from text paragraphs, bypassing OIE, which makes such models tied to the fixed KG schema. Critically, all these works conjecture that each OIE slot has a corresponding KG entry. Gashteovski et al. (2019) showed that this is hardly ever the case in practice.
+
+Contributions and Findings. We move away from the unrealistic and incomplete assumptions of prior work and propose ① a novel large-scale benchmark for OIE-to-KG linking; ② multifaceted evaluation protocols that cover all aspects of linking OIE facts to KGs (for an overview, see Fig. 1); ③ several strong baselines, inspired by state-of-the-art entity linking (Wu et al., 2019) and cross-modal retrieval (Miech et al., 2021; Geigle et al., 2022) approaches.
+
+Through our experimental study, we found that the methods (i) perform well transductively but (ii) their performance deteriorates in an inductive evaluation. Further, we find that (iii) a dedicated OIE-to-KG fact re-ranking model improves the linking performance of both inductive and polysemous OIEs, and that (iv) we obtain high performance by training models solely on a synthetic variant of our dataset (i.e., with the KG as the only human-annotated data). Lastly, we investigate the largely underexplored issue of detecting Out-of-Knowledge-Graph extractions. We show that (v) it is possible to detect Out-of-KG entities to an extent, however, the same does not hold for predicates: a task that our experiments identify as a difficult open
+
+problem, which requires more research attention.
+
+# 2 Fact Linking: Problem Statement
+
+For a given surface-form OIE triple  $t_1 =$  ("s"; "r"; "o"), the goal is to link each slot to a canonical concept in a KG (if the corresponding concept exists in the KG): "s" → e1 ∈ E; "r" → p ∈ P; "o" → e2 ∈ E, with E and P as the (fixed) sets of KG entities and predicates. Importantly, our problem definition (and consequently evaluation) focuses on linking at the fact level, where each OIE slot is contextualized with the other two OIE slots.
+
+Additionally and crucially, we want linking models that can assign an empty set to OIE surface forms (e.g., “ $s$ ”  $\rightarrow \emptyset$ ) when they refer to concepts not present in the KG (i.e., out-of-KG concepts). To enable a realistic setup for linking free-text OIE triples to a KG, we build a benchmark which considers four different facets (see Fig. 1).
+
+Transductive Fact Linking. In transductive linking, we measure how well the models link OIEs to KG facts consisting of entities and predicates seen during training (as components of training KG facts). Note that the testing KG facts (as whole triples) are not in the training data. Consider, for example, the extraction  $t_1$  in Fig. 1. In the transductive linking task, the mentions "played for" and "Chicago Bulls" refer to the KG predicate (P54) and entity (Q128109) respectively; both seen by the model as part of other training KG facts. However, the whole triple (Q41421; P54; Q128109) to which
+
+$t_1$  is linked was not part of the training data.
+
+Inductive Fact Linking. The inductive setup evaluates the linking to KG facts that consist of entities that are not seen during the training of the models. In other words, this setup tests the generalization of fact linking models over entities. In Fig. 1, given  $t_2$  as a test instance, the OIE entity "Wilmington" is inductive as it refers to a KG entity (Q659400) that is not part of any training KG fact.
+
+Polysemous Fact Linking. We focus on OIEs for which the "s" and "o" slots are ambiguous w.r.t. the KG, i.e., in isolation they refer to a set of KG entities rather than a single entity. The mention "Michael Jordan" from either  $t_1$  and  $t_2$  (Fig. 1), in isolation, refers to both the basketball player (Q41421) and the computer scientist (Q3308205). Here, the other two OIE slots offer the disambiguation signal that is necessary for successful linking.
+
+Out-of-KG Detection. We introduce a novel out-of-KG detection task, in which the models are to recognize that an OIE triple component (i.e., "s", "r" or "o") cannot be linked because they do not have a corresponding KG concept (e.g., the relation "grew up in" from the triple  $t_2$  in Fig. 1).
+
+# 3 FaLB: Fact Linking Benchmark
+
+We set up an automatic data processing pipeline to derive an OIE-to-KG fact linking benchmark, which supports all four evaluation facets from §2. We refer to both the process (i.e., pipeline) and the resulting benchmark as FaLB. FaLB's input is a dataset with (gold) alignments between natural language sentences and KG facts entailed by the sentence; i.e., each data instance is (sentence, KG fact) pair. Consequently, the creation of FaLB entails five design decisions: selection of (i) sentence-to-KG fact dataset(s), and (ii) a reference KG; (iii) generating OIE triples, (iv) high-precision OIE-KG fact alignments, and (v) a data augmentation strategy to increase the diversity of the data. Below is an example instance of the FaLB dataset:
+
+# Example Data Instance from FaLB
+
+Sentence: "M. J., who was born in Brooklyn, played for the Bulls."  
+OIE surface facts:  $t_1 =$  ("M. J."; "played for"; "the Bulls") and  $t_2 =$  ("M. J."; "was born in"; "Brooklyn")  
+KG canonical fact identifiers: (Q41421; P54; Q128109); (Q41421; P19; Q18419)  
+KG text facts: (Michael Jordan; member of sports team; Chicago Bulls); (Michael Jordan; place of birth; Brooklyn)  
+KG entity aliases: (Q41421: Air Jordan, Michael Jeffrey Jordan, His Airness); (Q128109: Bulls, The Bulls), ...
+
+Sentence-to-KG Fact Datasets. We build FaLB benchmarks for two such existing datasets: REBEL (Cabot and Navigli, 2021) and SynthIE (Josifoski et al., 2023). REBEL is built from Wikipedia abstracts, where the manually hyperlinked entities in each sentence are linked to Wikidata entities, thus making them golden. To match the sentence with a KG fact, for each pair of KG entities  $e_i$  and  $e_j$  within the sentence, REBEL obtains all KG predicates  $p_k$  such that  $(e_i, p_k, e_j)$  or  $(e_j, p_k, e_j)$  exist. However, two Wikidata entity nodes connected with a predicate (thus constituting a KG fact) do not guarantee the predicate validity in the sentence. For that reason, Cabot and Navigli (2021) use a Natural Language Inference pre-trained RoBERTa (Liu et al., 2019) to filter the predicates not entailed by the Wikipedia sentence; see (Cabot and Navigli, 2021) for details. We also use SynthIE, which is synthetically generated using a Large Language Model (LLM) (Brown et al., 2020). Given a set of KG facts, the LLM is prompted to generate a sentence that covers the KG facts. $^3$
+
+Reference KG. We use Wikidata (Vrandečić, 2012) as our reference KG, because REBEL and SynthIE align sentences to Wikidata facts. We select the subgraph of Wikidata that contains all entities which have a corresponding Wikipedia page, as per Wu et al. (2019). Following Lerer et al. (2019), we additionally filter out the most infrequent entities and predicates, appearing less than 5 times in the whole Wikidata dump. This results in a large reference KG with 5,794,782 unique entities and 4,153 unique predicates. We further create two smaller, dataset-specific reference KGs; one for each of the two datasets for which we apply FaLB: REBEL and SynthIE. These Benchmark-Restricted KGs (BRKGs) contain only the Wikidata entities and predicates referenced by at least one OIE triple extracted from their respective datasets. The BRKG for REBEL contains 625,125 entities and 565 predicates, while the one for SynthIE contains 702,334 entities and 758 predicates.
+
+Generating OIE Triples. We use four state-of-the-art OIE methods to obtain a set of OIE triples for each of the sentences in the dataset. To increase diversity, we use two state-of-the-art rule-based OIE methods: MinIE (Gashteovski et al., 2017) and StanfordOIE (Angeli et al., 2015); and two
+
+state-of-the-art neural OIE models: Millie (Kotnis et al., 2022c) and Multi $^2$ OIE (Ro et al., 2020).
+
+# High-precision OIE-KG Fact Alignments.
+
+Next, we need to match the extracted OIE triples against the set of KG facts associated with the sentences. Crucially, this automatic step needs to have a high-precision in order to create a high quality benchmark. As per the distant supervision assumption of Mintz et al. (2009), we create an alignment  $t \leftrightarrow f$  between any OIE triple  $t$  and any KG fact  $f$  that exactly match (i.e., have identical text form) in subject and object, assuming that the relation of  $t$  is aligned with the predicate of  $f$ . When multiple OIE triples (excluding exact duplicates)  $t_1, t_2, \ldots, t_k$  (e.g., extracted with different OIE systems) match with the same KG fact  $f$ , we obtain all  $k$  alignments:  $t_1 \leftrightarrow f, \ldots, t_k \leftrightarrow f$ . Finally, we remove all training alignments  $t \leftrightarrow f$  that exist in the test portion. To verify that this strategy indeed has high precision, we randomly sample 100 test instances, and conduct a human-study with two expert annotators. The annotators found  $97\%$  correct pairings (inter-annotator agreement of  $99\%$ ; Kohen's kappa of 0.80); see Appendix F for details. We therefore confirm the reliability of this design choice to produce high quality data.
+
+Data Augmentation to Increase Diversity. Lack of example diversity—most linked OIE entity mentions exactly match the text of the corresponding KG entities—is a common problem in existing linking benchmarks (Cabot and Navigli, 2021). This stems from strict data curation procedures that aim for high alignment precision, representing a mismatch with practice, where often the entity mentions do not exactly match the canonical KG entity denotation. To train and evaluate fact linking methods on more complex linking examples (e.g., linking from initials and abbreviations), we augment the data using the additional information available in the reference KG. For each alignment  $t \leftrightarrow f$ , we fetch the Wikidata aliases (i.e., non-canonical denotations) for the subject and object entities of  $f$ . We then create additional pairs  $t' \leftrightarrow f$  with augmented OIE triples  $t'$  for all possible alias combinations. For example, for the original OIE triple ("Michael Jordan"; "played for"; "Chicago Bulls"), this process results in augmented triples such as ("Air Jordan"; "played for"; "Chicago Bulls"), ("M.J."; "played for"; "The Bulls"), etc.
+
+# 4 OIE-to-KG Fact Linking Models
+
+Our goal is to link surface-form ("s"; "r"; "o") OIE triples, to canonical Knowledge Graph facts  $(e_1; p; e_2)$ , where  $e_1, e_2 \in \mathcal{E}$ , and  $p \in \mathcal{P}$ . Each KG entity or predicate is represented as its surface-form KG label (e.g., "Michael Jordan"), and its KG-provided description (e.g., "American basketball player and businessman"). We henceforth refer to the entities and predicates as KG entries. Intuitively, since both data streams—the OIE triples and the KG facts—are in natural language, we opt to obtain their representations (i.e., embeddings) with a pre-trained language model. We decouple the OIE-to-KG linking in two steps: pre-ranking and re-ranking (see §4.1 and §4.2, respectively), as is done commonly in entity retrieval (Wu et al., 2019) and image-text matching (Geigle et al., 2022; Li et al., 2021). See Fig. 2 for an overview and Appendix C for implementation details.
+
+# 4.1 Pre-ranking OIE Slots to KG Entries
+
+We denote this model as  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ . It aims to generate OIE slot embeddings and KG entry embeddings, such that an OIE slot embedding yields higher similarity with its aligned KG entry's embedding compared to the other KG entries. Therefore, during training, we contrast the positive pairs against a set of negatives, thereby training the model to generate embeddings for a matching OIE slot and KG entry that lie close in the latent space. The motivation for such formulation is two-fold: (i) the number of entities is large, and could practically grow further, therefore computing the softmax over all KG entities during training is prohibitive; (ii) there may be unseen KG entries that we encounter during inference, therefore posing the problem as a standard classification inevitably leads to the model ignoring them. We use RoBERTa (Liu et al., 2019) to encode the OIE slots and the KG entries.
+
+OIE Embeddings. We first add special tokens to indicate the start of each OIE slot:  $<\text{SUBJ}>$  for "subject",  $<\text{REL}>$  for "relation",  $<\text{OBJ}>$  for "object". Hence,  $t_1 = (\text{"M. Jordan"}; \text{"grew up in"}; \text{"Wilmington")}$  is represented as  $<\text{SUBJ}>$ $M.$  Jordan  $<\text{REL}>$  grew up in  $<\text{OBJ}>$  Wilmington". We then tokenize the OIE representation, denoted as  $\hat{t}$ ,
+
+![](images/67872ea64270e3ee1ec85c6af8b3fedd6f44a30d189a652a696fe59d3764f58f.jpg)  
+Figure 2: Left:  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  which performs pre-ranking of the OIE slots to KG entries independently (i.e., no context between them is considered). Trained with negatives sampled from the whole KG; Right: Factre ranker which attends between the whole OIE triple and KG fact to output their similarity; Trained with hard-negatives.
+
+![](images/4018c66f7e26bbd6762706e50f5e540ddb92fd288c96035948adb9f71e7f0bc5.jpg)
+
+and obtain RoBERTa token embeddings. Finally, we pool only the special tokens' embeddings subsequently linearly projected in the desired latent space:  $\hat{o}_i = \mathrm{Linear}(\mathrm{RoBERTa}(\hat{t}))$ , where  $\hat{o}$  is the slot embedding, and  $i \in \mathbb{R}^3$  is the OIE slot index.
+
+KG Embeddings. We represent both the entities and predicates as their label followed by their description (if available in the KG); e.g., the entity  $(e)$  Michael Jordan is represented as "Michael Jordan <DESC> American basketball player and businessman...", and the predicate  $(p)$  place of birth is represented as "place of birth <DESC> most specific known birth location of a person...". The <DESC> special token indicates the start of the description. We then tokenize the representation ( $\hat{e}$  for entity,  $\hat{p}$  for predicate), and obtain embeddings using the same RoBERTa model. We pool the <CLS> token embedding and linearly project it in the OIE representation space as:  $\hat{k}_j = \text{Linear}(\text{RoBERTa}(b_j))$ , where  $j \in \{\hat{e}, \hat{p}\}$ ,  $\hat{k}$  is the KG entry embedding, and  $b$  is the tokenized KG entry representation.
+
+Linking OIEs to KG Facts. Given an OIE slot embedding  $o_i$  and a KG entry embedding  $k_j$ , we obtain their dot-product as:  $\hat{s}_{pre} = o_i^T k_j$ , where  $o_i$  and  $k_j$  are norm-scaled (i.e.,  $\hat{s}_{pre}$  represents the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  cosine similarity). During inference, we link an OIE to a KG fact by selecting the most similar KG entry for each of the OIE slots.
+
+$\mathbf{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  Training. We train the model using standard contrastive loss: we sample  $N - 1$  in-batch negative KG entries for each positive OIE slot  $\leftrightarrow$  KG entry pair, where  $N$  is the batch size. As per standard practice (van den Oord et al., 2018), we train the model using temperature-scaled InfoNCE contrastive loss:  $\mathcal{L} = -\log \frac{e^{\hat{o}^{\mathrm{T}\hat{k}} / \tau}}{e^{\hat{o}^{\mathrm{T}\hat{k}} / \tau} + \sum_{n = 1}^{\mathrm{N - 1}}e^{\hat{o}^{\mathrm{T}\hat{k}_n} / \tau}}$
+
+where  $\tau$  is the temperature. Note that during training, we only sample negative KG entry embeddings for each OIE slot, but not the other way around.
+
+Sampling only in-batch negatives presents an issue as the training data represents only a limited subset of the whole KG (i.e., only the KG entries with paired OIE). During inference, however, we contrast each OIE slot against the whole KG to find the KG entry with which it exhibits the highest similarity. Therefore, for each OIE slot, we additionally sample  $e$  negative entities and  $p$  negative predicates at random from the whole KG (e.g., we would sample  $\sum_{n=1}^{N-1+p} \hat{k}_n^-$  negative predicates).
+
+# 4.2 Re-ranking OIEs to KG Facts
+
+In certain scenarios (e.g., in the case of polysemous OIEs), matching whole OIEs with whole KG facts (i.e., not decoupled per OIE slot) could resolve the ambiguity and thus improve performance. To that end, for each OIE slot, we re-rank the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  top-k most probable KG links. We denote this model as  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$ . We perform self-attention between the OIE and the KG fact (both provided as input, separated by a <FACT> special token) with a single RoBERTa transformer, and return their similarity as:  $\hat{s}_{re} = \sigma(\mathrm{Linear}(\mathrm{RoBERTa}(\hat{c}))),$  where  $\hat{c}$  is the concatenated OIE and KG fact representation, and  $\hat{s}_{re}$  is the sigmoid ( $\sigma$ ) normalized similarity.
+
+Fact $_{\text{ranker}}^{\text{pre}}$  Training. We train by sampling matching OIE  $\leftrightarrow$  KG fact pairs as positives, and negatives, where we replace some KG fact slots (subject, predicate, object) with incorrect ones, generated as follows: We first obtain embeddings for each KG entry using the  $OIE_{\text{ranker}}^{\text{pre}}$ , and find its top-k most similar candidates w.r.t. all other KG entries. We then corrupt the ground truth KG fact by randomly sampling only from the top-k (hard)
+
+negative candidates. Lastly, with  $50\%$  probability, we randomly mask (i.e., replace with a <mask> token) the description of the KG fact entries.
+
+# 5 Experiments and Discussions
+
+We measure accuracy to evaluate both OIE slot linking to KGs (§5.1), and Out-of-KG detection of OIE slots (§5.2). To measure OIE fact linking, we score a hit if all OIE slots are linked correctly. The error bars represent the standard error of the mean.
+
+# 5.1 Linking OIEs to Knowledge Graphs
+
+Setup. We explore the extent to which we can link OIE slots to a large-scale KG (Wikidata). We address two main research questions concerning the OIE-to-KG fact linking task: (i) To what extent do methods generalize to different KG entity facets? We consider transductive, inductive, or polysemous entities (see §2 for detailed definition); (ii) What is the performance impact of the KG size? We test two reference KG sizes: Benchmark Restricted KG ( $\sim$ 650k entities,  $\sim$ 0.6k predicates) and Large KG ( $\sim$ 5.9M entities,  $\sim$ 4k predicates).
+
+Methods. We use the following methods to obtain results for the OIE slot linking task: (i) RANDOM: for each OIE slot, we sample a random KG entry; (ii) FREQUENCY: based on the training data statistics, we link each OIE slot to the most frequent KG entry (entity or predicate) in the training set; (iii) SIMCSE: we use a pretrained SimCSE model (Gao et al., 2021), where we represent each OIE slot and each KG entry (optionally, its description as well – if available in the KG) in a natural language format, and obtain their embeddings. We finally obtain the cosine similarity between each OIE slot and KG entry to perform the linking; (iv) \(\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}\) : We train a pre-ranking model as per the setup described in §4.1; (v) + Context: We append the context sentence to the \(\mathrm{OIE}^5\); (vi) + Fact_{ranker}^{\mathrm{re}}\) : We additionally re-rank the top-k (\(k = 3\)) pre-ranked OIE slot links as per the setup in §4.2.
+
+Results. In Table 1 we report the OIE-to-KG linking performance for each OIE slot, as well as linking on fact level. Overall, across the data splits, we observe that all unsupervised baselines perform poorly compared to models proposed in this work; indicating that OIE linking is not trivial, hence
+
+off-the-shelf zero-shot models fail.6 Additionally, there is a significant performance drop on the inductive and polysemous split compared to the transductive split, suggesting that the models are neither robust w.r.t. entities unseen during training, nor can cope with polysemous entities. Expectedly, leveraging extra context (via the sentence from where the OIE is obtained) aids the linking process in the inductive and polysemous split, as it helps generalization and disambiguation. Similarly, the  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$  brings a significant performance gain especially prominent for linking complete facts. Finally, we observe a significant performance impact of the KG size: across all splits, OIE-to-KG linking is more challenging on large-scale KG compared to the smaller Benchmark-Restricted KG.
+
+Training on Synthetic Data Improves Performance. We explore the extent to which we can learn fact linking models using synthetic data. SynthIE (Josifoski et al., 2023) is a dataset that features natural language sentences paired with KG facts, where the sentences are obtained using a LLM. Namely, given a set of KG facts, Josifoski et al. (2023) prompt the LLM to generate a sentence which mentions (i.e., entails) all of the KG facts. Notably, if we can link OIE slots to KG facts by training on such synthetic dataset, then the KG remains the only human-annotated component for learning the OIE-to-KG fact linking task.
+
+To measure to what extent we can link OIEs to KG facts using only synthetic data, we train  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  models on both REBEL and SynthIE, and report results (in Table 2) on inductive testing splits w.r.t. each dataset. We observe that models trained on SynthIE are overall better OIE-to-KG fact linkers than models trained on REBEL (i.e., higher macro accuracy across datasets). This indicates that learning to link OIEs to KGs is possible using only synthetic data, thus the only human-annotated requirement remains to be a reference KG.
+
+Ablation Study: Importance of Entity Alias Augmentation. We observe that in current datasets most surface form entities (in the natural language sentences) appear only with their "canonical" la
+
+<table><tr><td>Method</td><td>Split Type</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td></tr><tr><td colspan="6">Benchmark-Restricted Knowledge Graph</td><td colspan="4">Large Knowledge Graph</td></tr><tr><td>Random</td><td>Transductive</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td></tr><tr><td>Frequency</td><td>Transductive</td><td>0.0 ± 0.0</td><td>53.3 ± 0.1</td><td>8.1 ± 0.1</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>53.3 ± 0.1</td><td>8.1 ± 0.1</td><td>0.0 ± 0.0</td></tr><tr><td>SimCSE</td><td>Transductive</td><td>5.4 ± 0.1</td><td>0.0 ± 0.0</td><td>0.9 ± 0.1</td><td>0.0 ± 0.0</td><td>2.6 ± 0.1</td><td>0.0 ± 0.0</td><td>0.5 ± 0.0</td><td>0.0 ± 0.0</td></tr><tr><td>OIEpreranker</td><td>Transductive</td><td>86.8 ± 0.1</td><td>93.5 ± 0.1</td><td>95.7 ± 0.0</td><td>79.1 ± 0.1</td><td>78.7 ± 0.2</td><td>93.5 ± 0.1</td><td>93.1 ± 0.1</td><td>70.7 ± 0.2</td></tr><tr><td>+ Context</td><td>Transductive</td><td>84.9 ± 0.1</td><td>92.2 ± 0.1</td><td>94.8 ± 0.1</td><td>77.7 ± 0.1</td><td>77.8 ± 0.1</td><td>92.2 ± 0.1</td><td>92.2 ± 0.1</td><td>70.2 ± 0.1</td></tr><tr><td>Frequency</td><td>Inductive</td><td>0.0 ± 0.0</td><td>0.1 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.1 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td></tr><tr><td>SimCSE</td><td>Inductive</td><td>12.6 ± 0.3</td><td>0.4 ± 0.1</td><td>6.8 ± 0.3</td><td>0.0 ± 0.0</td><td>8.1 ± 0.3</td><td>0.0 ± 0.0</td><td>3.2 ± 0.2</td><td>0.0 ± 0.0</td></tr><tr><td>OIEpreranker</td><td>Inductive</td><td>71.9 ± 0.4</td><td>69.8 ± 0.5</td><td>59.5 ± 0.5</td><td>34.9 ± 0.5</td><td>62.0 ± 0.5</td><td>69.8 ± 0.5</td><td>48.2 ± 0.5</td><td>25.4 ± 0.4</td></tr><tr><td>+ Context</td><td>Inductive</td><td>74.5 ± 0.4</td><td>73.8 ± 0.4</td><td>62.3 ± 0.5</td><td>38.9 ± 0.5</td><td>64.5 ± 0.5</td><td>73.8 ± 0.4</td><td>50.8 ± 0.5</td><td>29.2 ± 0.5</td></tr><tr><td>+ Factpreranker</td><td>Inductive</td><td>76.2 ± 0.4</td><td>67.6 ± 0.5</td><td>60.8 ± 0.5</td><td>40.6 ± 0.5</td><td>64.8 ± 0.5</td><td>66.5 ± 0.5</td><td>54.8 ± 0.5</td><td>32.9 ± 0.5</td></tr><tr><td>Frequency</td><td>Polysemous</td><td>0.0 ± 0.0</td><td>68.1 ± 0.4</td><td>11.7 ± 0.3</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>68.1 ± 0.4</td><td>11.7 ± 0.3</td><td>0.0 ± 0.0</td></tr><tr><td>SimCSE</td><td>Polysemous</td><td>1.3 ± 0.1</td><td>0.0 ± 0.0</td><td>0.3 ± 0.1</td><td>0.0 ± 0.0</td><td>0.4 ± 0.1</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td><td>0.0 ± 0.0</td></tr><tr><td>OIEpreranker</td><td>Polysemous</td><td>68.8 ± 0.4</td><td>93.0 ± 0.2</td><td>94.6 ± 0.2</td><td>62.5 ± 0.4</td><td>58.2 ± 0.4</td><td>93.0 ± 0.2</td><td>92.4 ± 0.2</td><td>51.7 ± 0.4</td></tr><tr><td>+ Context</td><td>Polysemous</td><td>77.9 ± 0.3</td><td>93.3 ± 0.2</td><td>95.8 ± 0.2</td><td>71.6 ± 0.4</td><td>69.5 ± 0.4</td><td>93.3 ± 0.2</td><td>94.0 ± 0.2</td><td>63.5 ± 0.4</td></tr><tr><td>+ Factpreranker</td><td>Polysemous</td><td>75.3 ± 0.4</td><td>90.6 ± 0.2</td><td>95.3 ± 0.2</td><td>69.5 ± 0.4</td><td>74.1 ± 0.4</td><td>90.6 ± 0.2</td><td>95.6 ± 0.2</td><td>69.3 ± 0.4</td></tr></table>
+
+Table 1: OIE slot and fact linking accuracy with models trained and evaluated on REBEL. Evaluation on a smaller Benchmark-Restricted KG, and a Large KG. The error bars indicate standard error of the mean.  
+
+<table><tr><td>Train Data</td><td>Test Data</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td></tr><tr><td>REBEL</td><td>REBEL</td><td>62.0 ± 0.5</td><td>69.8 ± 0.5</td><td>48.2 ± 0.5</td><td>25.4 ± 0.4</td></tr><tr><td>REBEL</td><td>SynthIE</td><td>53.6 ± 0.4</td><td>57.2 ± 0.4</td><td>44.9 ± 0.4</td><td>17.4 ± 0.3</td></tr><tr><td colspan="2">Macro Score</td><td>57.8</td><td>63.5</td><td>46.6</td><td>21.4</td></tr><tr><td>SynthIE</td><td>REBEL</td><td>69.7 ± 0.3</td><td>63.2 ± 0.3</td><td>41.6 ± 0.3</td><td>22.4 ± 0.2</td></tr><tr><td>SynthIE</td><td>SynthIE</td><td>64.1 ± 0.6</td><td>81.2 ± 0.5</td><td>57.6 ± 0.6</td><td>34.5 ± 0.6</td></tr><tr><td colspan="2">Macro Score</td><td>66.9</td><td>72.2</td><td>49.6</td><td>28.5</td></tr></table>
+
+bel. $^{8}$  Since the OIEs represent surface-form facts, such lack of diversity prevents the models from learning more complex linking patterns. To overcome this, we perform entity alias augmentation in FaLB by adding the surface form aliases of the entities—available in Wikidata and manually curated—and ablate its impact on the OIE linking task. Besides  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  models trained on REBEL and SynthIE with entity alias augmentation, we train additional models without the alias augmented samples. We report results in Table 3 on inductive REBEL and SynthIE testing data, which does and does not feature entity aliases.
+
+Unexpectedly, we observe that training with entity aliases allows us to link such OIE entity mentions more successfully than training without them. This was reflected on models trained on both REBEL
+
+Table 2: Experiments on human-created (REBEL) and synthetically generated (SynthIE) datasets. OIE linking to a Large KG variant on inductive splits w.r.t. each dataset (see Appendix D.1 for details).  
+
+<table><tr><td>Training</td><td>Testing</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td></tr><tr><td colspan="6">Augmentation</td></tr><tr><td colspan="6">Models trained and evaluated on REBEL</td></tr><tr><td>X</td><td>X</td><td>90.7 ± 0.1</td><td>91.9 ± 0.1</td><td>87.1 ± 0.2</td><td>74.0 ± 0.2</td></tr><tr><td>X</td><td>✓</td><td>61.6 ± 0.2</td><td>64.8 ± 0.2</td><td>41.5 ± 0.2</td><td>25.2 ± 0.2</td></tr><tr><td colspan="2">Macro Score</td><td>76.2</td><td>78.4</td><td>64.3</td><td>49.6</td></tr><tr><td>✓</td><td>X</td><td>89.0 ± 0.1</td><td>92.8 ± 0.1</td><td>85.1 ± 0.2</td><td>72.4 ± 0.2</td></tr><tr><td>✓</td><td>✓</td><td>79.0 ± 0.2</td><td>92.2 ± 0.1</td><td>89.0 ± 0.1</td><td>67.9 ± 0.2</td></tr><tr><td colspan="2">Macro Score</td><td>84.0</td><td>92.5</td><td>87.1</td><td>70.15</td></tr><tr><td colspan="6">Models trained and evaluated on SynthIE</td></tr><tr><td>X</td><td>X</td><td>91.8 ± 0.2</td><td>86.3 ± 0.2</td><td>90.7 ± 0.2</td><td>72.8 ± 0.3</td></tr><tr><td>X</td><td>✓</td><td>61.2 ± 0.2</td><td>70.1 ± 0.2</td><td>40.7 ± 0.2</td><td>22.6 ± 0.2</td></tr><tr><td colspan="2">Macro Score</td><td>76.5</td><td>78.2</td><td>65.7</td><td>47.7</td></tr><tr><td>✓</td><td>X</td><td>91.1 ± 0.2</td><td>88.9 ± 0.2</td><td>90.5 ± 0.2</td><td>74.3 ± 0.3</td></tr><tr><td>✓</td><td>✓</td><td>80.1 ± 0.2</td><td>89.9 ± 0.1</td><td>86.3 ± 0.2</td><td>64.8 ± 0.2</td></tr><tr><td colspan="2">Macro Score</td><td>85.6</td><td>89.4</td><td>88.4</td><td>69.6</td></tr></table>
+
+Table 3: Impact of the FaLB entity alias augmentation step on REBEL and SynthIE. Inference is done on the full test set. The model is  $OIE_{\text{ranker}}^{\text{pre}}$ .
+
+and SynthIE. We further observe that this step hurts the linking of specific OIE entity mentions only moderately, suggesting that OIE linking methods could be trained to be robust w.r.t. entity synsets. Finally, across all OIE slots and fact linking, the macro scores are significantly in favor of the model trained with entity aliases, on both datasets.
+
+Ablation Study: Importance of Fact-eranking. We ablate the number of KG facts we rerank  $(k)$  with  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$  and report results in Fig. 3. We observe the highest fact linking accuracy when performing reranking using the top-2 highest scoring KG entries for each OIE slot, although performance is within 1 standard deviation for  $k = 2, 3, 4$ . If reranking  $k = 2$  facts, effectively, the  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$
+
+![](images/88ec2732062cdd2daf0be889467c9024abdec0c33bb0f171704326caccf8aa45.jpg)  
+Figure 3: Ablating the impact of K in the Fact<sup>re</sup> ranker.
+
+constructs  $2^{3}$  OIE-KG fact pairs, reranks the list, and returns the highest scoring KG fact.
+
+# 5.2 Detecting Out-of-Knowledge Graph OIEs
+
+Setup. Prior works which study the OIE to KG linking problem (Zhang et al., 2019; Wood et al., 2021; Jiang et al., 2021) make the assumption that all OIE slots that need to be linked are present in the KG, which is rarely the case in practice. Here, we study (i) whether OIE slot linking methods can be converted to Out-of-KG detectors; (ii) the performance impact of the KG size on the Out-of-KG detection task; and (iii) to what extent is it more difficult to recognize the presence or absence of aliased OIE entity mentions.
+
+Datasets. We create the Out-of-KG test split by imposing constraints such that we mimic an out-of-distribution setting. The constraints are: (i) both the KG entities and predicates referred by OIEs are unseen during training (i.e., not in the KG); (ii) the testing OIE-to-KG pairs do not come from the training data distribution: we use models trained on REBEL, but evaluate on SynthIE (which we use to create the Out-of-KG split). We measure Out-of-KG detection performance on the original OIEs, and OIEs with alias augmented entity slots.
+
+Evaluation Protocol. We evaluate to what extent we can detect if an OIE slot refers to a concept outside of the KG. For each OIE slot, we add its corresponding concept (KG entity or predicate) to the KG, and score a hit if the model outputs a positive score for that slot. Conversely, for each OIE slot, we remove its corresponding entity or predicate from the KG, and score a hit if the model outputs a negative score. We report the averaged
+
+accuracy over the two scenarios for each OIE slot.
+
+Methods. We create three methods on top of  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  trained on REBEL: (i) CONFIDENCE@1-BASED HEURISTIC: We get the cosine similarities of the top-5 predictions from  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ , compute the softmax, and threshold the top-1 confidence; (ii) ENTROPY-BASED HEURISTIC: We also obtain the top-5 cosine similarities, however, after softmax normalization compute their entropy. Finally, we threshold the entropy to obtain the prediction; (iii) QUERY-KEY-VALUE CROSS-ATTENTION: We train a lightweight query-key-value cross-attention module on top of the frozen  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  embeddings. Given a query OIE slot embedding, the model attends over the KG entry embeddings representing the keys and values, and outputs a probability indicating the presence of the OIE slot in the KG (See Appendix E.1 for details).
+
+Results. We report the Out-of-KG detection performance in Table 4. First, we observe that none of the models we evaluate are able to recognize whether an OIE relation has a corresponding KG predicate, indicating that models do not cope with zero-shot relations. Intuitively, the number of entities is orders of magnitude more than the relations, thus the models learn features which generalize to unseen data. On the other hand, the number of relations is limited ( $\sim$ 600 during training) and therefore, the models overfit on this limited set. Second, similar to our observations on the OIE linking task, detection of Out-of-KG slots is significantly more difficult on data which features entity aliases, and even more difficult on larger KGs. Lastly, the best performing model is an entropy-based heuristic on top of the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  output scores. Overall, our results indicate that Out-of-KG detection remains an open research problem.
+
+# 6 Related Work
+
+Prior work (Zhang et al., 2019; Wood et al., 2021; Jiang et al., 2021)—based on the ReVerb45k dataset (Vashishth et al., 2018)—considers fact linking as inductive and polysemous for the entities, but perform inductive inference for the relations. They also assume the OIE entities are linked to the KG entities a priori. In contrast, FaLB requires linking of all OIE slots to KG entries, with multiple evaluation facets (including the out-of-KG
+
+<table><tr><td></td><td>Method</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td><td>Subject</td><td>Relation</td><td>Object</td><td>Fact</td></tr><tr><td colspan="6">Benchmark-Restricted Knowledge Graph</td><td colspan="4">Large Knowledge Graph</td></tr><tr><td colspan="10">Testing data with entity alias augmentation</td></tr><tr><td>Random</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>12.5 ± 0.8</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>12.5 ± 0.8</td><td></td></tr><tr><td>Confidence@1-based Heuristic</td><td>63.8 ± 1.2</td><td>49.7 ± 1.2</td><td>65.0 ± 1.2</td><td>21.5 ± 0.4</td><td>62.0 ± 1.2</td><td>49.7 ± 1.2</td><td>62.7 ± 1.2</td><td>18.5 ± 0.4</td><td></td></tr><tr><td>Entropy-based Heuristic</td><td>67.7 ± 1.2</td><td>49.0 ± 1.2</td><td>68.6 ± 1.1</td><td>22.8 ± 0.4</td><td>63.1 ± 1.1</td><td>49.0 ± 1.2</td><td>64.5 ± 1.1</td><td>23.0 ± 0.4</td><td></td></tr><tr><td>Query-Key-Value Cross-Attention</td><td>62.4 ± 1.2</td><td>48.8 ± 1.0</td><td>63.8 ± 1.2</td><td>17.2 ± 0.2</td><td>55.3 ± 1.2</td><td>49.2 ± 1.2</td><td>56.8 ± 1.2</td><td>16.6 ± 0.2</td><td></td></tr><tr><td colspan="10">Testing data without entity alias augmentation</td></tr><tr><td>Random</td><td>50.0 ± 2.8</td><td>50.0 ± 2.8</td><td>50.0 ± 2.8</td><td>12.5 ± 1.8</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>50.0 ± 1.3</td><td>12.5 ± 1.8</td><td></td></tr><tr><td>Confidence@1-based Heuristic</td><td>71.2 ± 2.3</td><td>49.6 ± 2.6</td><td>72.1 ± 2.3</td><td>30.2 ± 1.5</td><td>71.2 ± 2.5</td><td>49.6 ± 2.7</td><td>69.1 ± 2.5</td><td>24.5 ± 1.5</td><td></td></tr><tr><td>Entropy-based Heuristic</td><td>77.5 ± 2.3</td><td>49.0 ± 2.7</td><td>78.4 ± 2.2</td><td>29.9 ± 1.7</td><td>73.3 ± 2.3</td><td>49.1 ± 2.7</td><td>72.8 ± 2.4</td><td>27.0 ± 1.7</td><td></td></tr><tr><td>Query-Key-Value Cross-Attention</td><td>70.7 ± 2.5</td><td>49.6 ± 2.7</td><td>70.9 ± 2.5</td><td>25.9 ± 1.0</td><td>59.8 ± 2.7</td><td>49.4 ± 2.7</td><td>60.8 ± 2.5</td><td>19.7 ± 0.9</td><td></td></tr></table>
+
+Table 4: Detection accuracy of Out-of-Knowledge Graph entities and predicates on the Out-of-KG split – built on top of SynthIE. All models trained on REBEL with entity alias augmentation.  
+
+<table><tr><td>Dataset</td><td>Golden entities</td><td>Manually validated refs.</td><td>Multifaceted</td><td>Out-of-KG</td><td>Inductive</td><td>Transductive</td><td>Polysemous</td><td># OIE Entities</td><td># OIE Relations</td></tr><tr><td>FaLB (REBEL)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>936,655</td><td>159,597</td></tr><tr><td>FaLB (SynthIE)</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>✓</td><td>1,049,922</td><td>147,056</td></tr><tr><td>ReVerb45k</td><td>X</td><td>X</td><td>X</td><td>X</td><td>—</td><td>—</td><td>✓</td><td>28,798</td><td>21,925</td></tr></table>
+
+Table 5: Comparison of OIE-to-KG datasets: FaLB v.s. ReVerb45k. As transductivity and inductivity are defined only w.r.t. what the model has observed during training, we leave these entries blank, because ReVerb45k has only validation and testing dataset.
+
+setup). In ReVerb45k (Vashishth et al., 2018) the links from the OIE entities to the KG entities are not golden (i.e., human labelled), but rather automatically obtained with an outdated entity linker (Lin et al., 2012); thus, the poor performance of the entity linker caps the performance of the fact linking models. In addition, the number of benchmark KG predicates is unrealistically small (only 250) compared to modern KGs (e.g., Wikidata). In this work, we mitigate these issues and build a benchmark (i) with golden links to KG entities that also (ii) reflects the size of modern KGs.
+
+Furthermore, all prior work to date has relied on the strict assumption that all OIE surface form slots have a corresponding reference KG entity or predicate (Wu et al., 2019; Jiang et al., 2021; Zhang et al., 2019; Josifoski et al., 2022). This is obviously a false assumption, as any text corpus "in the wild" contains entities and predicates not present in even the largest of KGs (Gashteovski et al., 2020a). See App. A for detailed related work discussion.
+
+Finally, most existing publicly available datasets do not address the problem of OIE-to-KG linking. Popular datasets like T-REx (Elsahar et al., 2018) and REBEL (Cabot and Navigli, 2021)—if considered without modification—address only Text-to-KG alignment, thus lack the OIE component. Therefore, these datasets are not directly comparable to FaLB. To the best of our knowledge, the only publicly-available OIE-to-KG dataset is Re
+
+Verb45k (Vashishth et al., 2018), which, as indicated above, has several drawbacks. For detailed comparisons with and FaLB, see Table5.
+
+# 7 Conclusion
+
+We shed light on the OIE to KG linking problem, allowing us to fuse the surface-form and opened-ended knowledge found in OIEs, with the canonical real-world KG facts. We introduced a novel multifaceted benchmark which fixes prior work deficiencies, and proposed a set of task-specific baselines. Our experiments uncover that (i) linking inductive or polysemous OIEs to large KGs is challenging; (ii) we can learn OIE linking methods using only synthetic data; and (iii) detecting whether OIEs are Out-of-KG is an open research problem.
+
+# Acknowledgements
+
+We thank Dina Trajkovska for the help with the figures, and Mike Zhang for the fruitful discussions and feedback at the initial stages of the project.
+
+# Limitations
+
+Notably, the set of models we explore ignore the KG structure to obtain KG entry embeddings. Leveraging the underlying graph structure should, in theory, yield representations which generalize better to zero-shot samples (e.g., as is the case with detecting out-of-KG relations). Such KG entry embeddings could be even trained offline (i.e., as a separate step) with standard KG embedding methods (Bordes et al., 2013).
+
+Last but not least, all data, resources, and models used in this work are specific to the English language. Notice however, our approach can be readily extended to languages other than English, while Wikipedia and Wikidata have versions in other languages – which we leave for future work.
+
+# Ethical Impact
+
+We are not aware of any direct ethical impact generated by our work. However, in general, care should be taken when applying our technology to sensitive use cases in high risk domains, such as healthcare.
+
+# References
+
+Gabor Angeli, Melvin Johnson, and Christopher D. Manning. 2015. Leveraging Linguistic Structure For Open Domain Information Extraction. In Annual Meeting of the Association for Computational Linguistics (ACL), page 344-354.  
+Michele Banko, Michael J. Cafarella, Stephen Soderland, Matthew Broadhead, and Oren Etzioni. 2007. Open Information Extraction from the Web. In Communications of the ACM.  
+Farimah Bayat, Nikita Bhutani, and H. V. Jagadish. 2022. CompactIE: Compact Facts in Open Information Extraction. In North American Chapter of the Association for Computational Linguistics (NAACL), pages 900-910.  
+Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and Oksana Yakhnenko. 2013. Translating Embeddings for Modeling Multi-relational Data. Advances in Neural Information Processing Systems (NeurIPS), 26.  
+Samuel Broscheit, Kiril Gashteovski, and Martin Achenbach. 2017. OpenIE for Slot Filling at TAC KBP 2017 - System Description. Text Analysis Conference (TAC).  
+Samuel Broscheit, Kiril Gashteovski, Yanjie Wang, and Rainer Gemulla. 2020. Can We Predict New Facts with Open Knowledge Graph Embeddings? A Benchmark for Open Link Prediction. In Annual Meeting of
+
+the Association for Computational Linguistics (ACL), page 2296-2308.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, T. J. Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901.  
+Pere-Lluis Huguet Cabot and Roberto Navigli. 2021. REBEL: Relation Extraction By End-to-end Language generation. In Findings of the Association for Computational Linguistics: Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2370-2381.  
+Luciano Del Corro and Rainer Gemulla. 2013. ClausIE: Clause-based Open Information Extraction. In Proceedings of the International Conference on World Wide Web (WWW), page 355-366.  
+Daniel Daza, Michael Cochez, and Paul T. Groth. 2021. Inductive Entity Representations from Text via Link Prediction. Proceedings of the Web Conference (WWW), page 798-808.  
+David Dukić, Kiril Gashteovski, Goran Glavaš, and Jan Šnajder. 2023. Leveraging Open Information Extraction for Improving Few-Shot Trigger Detection Domain Transfer. arXiv preprint arXiv:2305.14163.  
+Hady Elsahar, Pavlos Vougiouklis, Arslen Remaci, Christophe Gravier, Jonathon Hare, Frederique Laforest, and Elena Simperl. 2018. T-REx: A Large Scale Alignment of Natural Language with Knowledge Base Triples. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018).  
+David Friede and Kiril Gashteovski. 2022. Context-Aware Entity Linking for Knowledge Graphs. US Patent App. 17/392,319.  
+Niklas Friedrich, Kiril Gashteovski, Mingying Yu, Bhushan Kotnis, Caroline V. Lawrence, Mathias Niepert, and Goran Glavavs. 2022. AnnIE: An Annotation Platform for Constructing Complete Open Information Extraction Benchmark. In Annual Meeting of the Association for Computational Linguistics (ACL). System Demonstrations., page 44-60.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 6894-6910.
+
+Kiril Gashteovski, Rainer Gemulla, and Luciano Del Corro. 2017. MinIE: Minimizing Facts in Open Information Extraction. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 2630-2640.  
+Kiril Gashteovski, Rainer Gemulla, Bhushan Kotnis, Sven Hertling, and Christian Meilicke. 2020a. On Aligning OpenIE Extractions with Knowledge Bases: A Case Study. Proceedings of the First Workshop on Evaluation and Comparison of NLP Systems (Eval4NLP@EMNLP).  
+Kiril Gashteovski, Rainer Gemulla, Bhushan Kotnis, Sven Hertling, and Christian Meilicke. 2020b. On Aligning OpenIE Extractions with Knowledge Bases: A Case Study. In Workshop on Evaluation and Comparison of NLP Systems (Eval4NLP@EMNLP), page 143-154.  
+Kiril Gashteovski, Sebastian Wanner, Sven Hertling, Samuel Broscheit, and Rainer Gemulla. 2019. OPIEC: An Open Information Extraction Corpus. In Conference on Automatic Knowledge Base Construction (AKBC).  
+Kiril Gashteovski, Mingying Yu, Bhushan Kotnis, Carolin Lawrence, Mathias Niepert, and Goran Glavaš. 2022a. BenchIE: A Framework for Multi-Faceted Fact-Based Open Information Extraction Evaluation. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 4472-4490.  
+Kiril Gashteovski, Mingying Yu, Bhushan Kotnis, Carolin Lawrence, Mathias Niepert, and Goran Glavaš. 2022b. BenchIE: A framework for multi-faceted fact-based open information extraction evaluation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4472-4490, Dublin, Ireland. Association for Computational Linguistics.  
+Gregor Geigle, Jonas Pfeiffer, Nils Reimers, Ivan Vulic, and Iryna Gurevych. 2022. Retrieve Fast, Rerank Smart: Cooperative and Joint Approaches for Improved Cross-Modal Retrieval. Transactions of the Association for Computational Linguistics (TACL), 10:503-521.  
+Patrick Hohenecker, Frank Martin Mtumbuka, Vid Kocijan, and Thomas Lukasiewicz. 2020. Systematic Comparison of Neural Architectures and Training Approaches for Open Information Extraction. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 8554-8565.  
+Sameer Jain, Vaishakh Keshava, Swarnashree Mysore Sathyendra, Patrick Fernandes, Pengfei Liu, Graham Neubig, and Chunting Zhou. 2023. Multi-Dimensional Evaluation of Text Summarization with In-Context Learning. In Annual Meeting of the Association for Computational Linguistics (ACL). Findings., page 8487-8495.
+
+Zhengbao Jiang, Jialong Han, Bunyamin Sisman, and Xin Dong. 2021. CoRI: Collective Relation Integration with Data Augmentation for Open Information Extraction. In Annual Meeting of the Association for Computational Linguistics (ACL), page 4706-4716.  
+Jeff Johnson, Matthijs Douze, and Hervé Jégou. 2019. Billion-Scale Similarity Search with GPUs. IEEE Transactions on Big Data, 7(3):535-547.  
+Hideaki Joko, Faegheh Hasibi, Krisztian Balog, and Arjen P de Vries. 2021. Conversational Entity Linking: Problem Definition and Datasets. In Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR), pages 2390-2397.  
+Martin Josifoski, Nicola De Cao, Maxime Peyrard, and Robert West. 2022. GenIE: Generative Information Extraction. In North American Chapter of the Association for Computational Linguistics (NAACL), pages 4626-4643.  
+Martin Josifoski, Marija Sakota, Maxime Peyrard, and Robert West. 2023. Exploiting Asymmetry for Synthetic Training Data Generation: SynthIE and the Case of Information Extraction. ArXiv, abs/2303.04132.  
+Bhushan Kotnis, Kiril Gashteovski, J. Gastinger, Giuseppe Serra, Francesco Alesiani, Timo Sztyler, Ammar Shaker, Na Gong, Carolin (Haas) Lawrence, and Zhao Xu. 2022a. Human-Centric Research for NLP: Towards a Definition and Guiding Questions. ArXiv, abs/2207.04447.  
+Bhushan Kotnis, Kiril Gashteovski, and Carolin Lawrence. 2022b. Open Information Extraction from Low Resource Languages. US Patent App. 17/342,575.  
+Bhushan Kotnis, Kiril Gashteovski, and Carolin Lawrence. 2023. Keyword Based Open Information Extraction for Fact-Relevant Knowledge Graph Creation and Link Prediction. US Patent App. 17/734,129.  
+Bhushan Kotnis, Kiril Gashteovski, Daniel Rubio, Ammar Shaker, Vanessa Rodriguez-Tembras, Makoto Takamoto, Mathias Niepert, and Carolin Lawrence. 2022c. MILIE: Modular & Iterative Multilingual Open Information Extraction. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 6939-6950.  
+Anne Lauscher, Yide Song, and Kiril Gashteovski. 2019. MinSciIE: Citation-Centered Open Information Extraction. ACM/IEEE Joint Conference on Digital Libraries (JCDL), pages 386-387.  
+Adam Lerer, Ledell Wu, Jiajun Shen, Timothee Lacroix, Luca Wehrstedt, Abhijit Bose, and Alex Peysakhovich. 2019. PyTorch-BigGraph: A Large-scale Graph Embedding System. Proceedings of Machine Learning and Systems, 1:120-131.
+
+Irene Li, Jessica Pan, Jeremy Goldwasser, Neha Verma, Wai Pan Wong, Muhammed Yavuz Nuzumlali, Benjamin Rosand, Yixin Li, Matthew Zhang, David Chang, Richard Andrew Taylor, Harlan M. Krumholz, and Dragomir R. Radev. 2022. Neural Natural Language Processing for Unstructured Data in Electronic Health Records: a Review. Computer Science Review, 46:100511.  
+Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 2021. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. Advances in Neural Information Processing Systems (NeurIPS), 34:9694-9705.  
+Thomas Lin, Mausam, and Oren Etzioni. 2012. Entity Linking at Web Scale. In Proceedings of the Joint Workshop on Automatic Knowledge Base Construction and Web-scale Knowledge Extraction (AKBC-WEKEX@NAACL-HLT), pages 84-88.  
+Xueling Lin, Haoyang Li, Hao Xin, Zijian Li, and Lei Chen. 2020. KBPearl: A Knowledge Base Population System Supported by Joint Entity and Relation Linking. Proceedings of the VLDB Endowment, 13(7):1035-1049.  
+Pengfei Liu, Jinlan Fu, Yanghua Xiao, Weizhe Yuan, Shuaichen Chang, Junqi Dai, Yixin Liu, Zihuiwen Ye, Zi-Yi Dou, and Graham Neubig. 2021. *ExplainaBoard: An Explanable Leaderboard for NLP*. In Annual Meeting of the Association for Computational Linguistics (ACL). System Demonstrations, page 280-289.  
+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.  
+Mausam, Michael Schmitz, Stephen Soderland, Robert Bart, and Oren Etzioni. 2012. Open Language Learning for Information Extraction. In *Conference on Empirical Methods in Natural Language Processing (EMNLP)*, page 523-534.  
+Mary L. McHugh. 2012. Interrater Reliability: the Kappa Statistic. Biochemia Medica, 22(3):276-282.  
+Christian Meilicke, Manuel Fink, Yanjie Wang, Daniel Ruffinelli, Rainer Gemulla, and Heiner Stuckenschmidt. 2018. Fine-Grained Evaluation of Rule- and Embedding-Based Systems for Knowledge Graph Completion. In International Semantic Web Conference (ISWC), page 3-20.  
+Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Andrew Zisserman. 2021. Thinking Fast and Slow: Efficient Text-to-Visual Retrieval with Transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9826-9836.
+
+Mike Mintz, Steven Bills, Rion Snow, and Dan Jurafsky. 2009. Distant Supervision for Relation Extraction without Labeled Data. In Proceedings of the Joint Conference of the Annual Meeting of the ACL and the International Joint Conference on Natural Language Processing of the AFNLP (ACL/IJCNLP), pages 1003-1011.  
+Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. 2021. Interventional Video Grounding with Dual Contrastive Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2765-2775.  
+Federico Nanni, Jingyi Zhang, Ferdinand Betz, and Kiril Gashteovski. 2019. EAL: A toolkit and dataset for entity-aspect linking. In ACM/IEEE Joint Conference on Digital Libraries (JCDL), pages 430-431. IEEE.  
+Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zach DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic Differentiation in PyTorch.  
+Miao Peng, Ben Liu, Qianqian Xie, Wenjie Xu, Hua Wang, and Min Peng. 2022. SMiLE: Schemaaugmented Multi-level Contrastive Learning for Knowledge Graph Link Prediction. In Findings of the Association for Computational Linguistics: Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4194-4206.  
+Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vassilis Plachouras, Tim Rocktaschel, and Sebastian Riedel. 2020. KILT: a Benchmark for Knowledge Intensive Language Tasks. In North American Chapter of the Association for Computational Linguistics (NAACL), page 2523-2544.  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR.  
+Leonardo F. R. Ribeiro, Mengwen Liu, Iryna Gurevych, Markus Dreyer, and Mohit Bansal. 2022. FactGraph: Evaluating Factuality in Summarization with Semantic Graph Representations. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 3238-3253.  
+Marco Tulio Ribeiro, Tongshuang Sherry Wu, Carlos Guestrin, and Sameer Singh. 2020. Beyond Accuracy: Behavioral Testing of NLP Models with CheckList. In Annual Meeting of the Association for Computational Linguistics (ACL), page 4902-4912.
+
+Wiem Ben Rim, Carolin (Haas) Lawrence, Kiril Gashteovski, Mathias Niepert, and Naoaki Okazaki. 2021. Behavioral Testing of Knowledge Graph Embedding Models for Link Prediction. In Conference on Automated Knowledge Base Construction (AKBC).  
+Youngbin Ro, Yukyung Lee, and Pilsung Kang. 2020. Multi $^2$ OIE: Multilingual Open Information Extraction Based on Multi-Head Attention with BERT. In Findings of the Association for Computational Linguistics: Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1107–1117.  
+Ahmad Sakor, Kuldeep Singh, Anery Patel, and Maria-Esther Vidal. 2020. Falcon 2.0: An Entity and Relation Linking Tool over Wikidata. In Proceedings of the ACM International Conference on Information & Knowledge Management (CIKM), pages 3141-3148.  
+Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.  
+Sascha Saralajew, Ammar Shaker, Zhao Xu, Kiril Gashteovski, Bhushan Kotnis, Wiem Ben-Rim, Jürgen Quittek, and Carolin Lawrence. 2022. A Human-Centric Assessment Framework for AI. ArXiv, abs/2205.12749.  
+Apoory Saxena, Aditay Tripathi, and Partha Talukdar. 2020. Improving Multi-hop Question Answering over Knowledge Graphs using Knowledge Base Embeddings. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), pages 4498-4507.  
+Gabriel Stanovsky, Julian Michael, Luke Zettlemoyer, and Ido Dagan. 2018. Supervised Open Information Extraction. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pages 885-895.  
+Dianbo Sui, Chenhao Wang, Yubo Chen, Kang Liu, Jun Zhao, and Wei Bi. 2021. Set Generation Networks for End-to-End Knowledge Base Population. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9650–9660.  
+Bayu Distiawan Trisedya, Gerhard Weikum, Jianzhong Qi, and Rui Zhang. 2019. Neural Relation Extraction for Knowledge Base Enrichment. In Annual Meeting of the Association for Computational Linguistics (ACL), pages 229-240.  
+Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. ArXiv, abs/1807.03748.  
+Shikhar Vashisth, Prince Jain, and Partha Pratim Talukdar. 2018. CESI: Canonicalizing Open Knowledge Bases using Embeddings and Side Information. Proceedings of the World Wide Web Conference (WWW), page 1317-1327.
+
+Vijay Viswanathan, Kiril Gashteovski, Carolin Lawrence, Tongshuang Wu, and Graham Neubig. 2023. Large Language Models Enable Few-Shot Clustering. arXiv preprint arXiv:2307.00524.  
+Denny Vrandecic. 2012. Wikidata: A new platform for collaborative data collection. In Proceedings of the International Conference on World Wide Web (WWW), pages 1063-1064.  
+Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple Contrastive Knowledge Graph Completion with Pre-trained Language Models. In Association for Computational Linguistics (ACL), pages 4281-4294.  
+Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu, Zhengyan Zhang, Zhiyuan Liu, Juanzi Li, and Jian Tang. 2021. KEPLER: A Unified Model for Knowledge Embedding and Pre-trained Language Representation. Transactions of the Association for Computational Linguistics (TACL), 9:176-194.  
+Haris Widjaja, Kiril Gashteovski, Wiem Ben Rim, Pengfei Liu, Christopher Malon, Daniel Ruffinelli, Carolin (Haas) Lawrence, and Graham Neubig. 2022. KGxBoard: Explainable and Interactive Leaderboard for Evaluation of Knowledge Graph Completion Models. In Conference on Empirical Methods in Natural Language Processing (EMNLP). System Demonstrations., page 338-350.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. 2020. Transformers: State-of-the-Art Natural Language Processing. In *Conference on Empirical Methods in Natural Language Processing* (EMNLP). System Demonstrations, page 38-45.  
+Ian D. Wood, Mark Johnson, and Stephen Wan. 2021. Integrating Lexical Information into Entity Neighbourhood Representations for Relation Prediction. In North American Chapter of the Association for Computational Linguistics (NAACL), page 3429-3436.  
+Chengmin Wu, Enrui Hu, Ke Zhan, Lan Luo, Xinyu Zhang, Hao Jiang, Qirui Wang, Zhao Cao, Fan Yu, and Lei Chen. 2022. Triple-Fact Retriever: An Explanable Reasoning Retrieval Model for Multi-hop QA Problem. In International Conference on Data Engineering (ICDE), pages 1206-1218. IEEE.  
+Ledell Yu Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer. 2019. Scalable Zero-shot Entity Linking with Dense Entity Retrieval. In Conference on Empirical Methods in Natural Language Processing (EMNLP), page 6397-6407.  
+Hongbin Ye, Honghao Gui, Xin Xu, Huajun Chen, and Ningyu Zhang. 2023. Schema-adaptable Knowledge Graph Construction. arXiv preprint arXiv:2305.08703.
+
+Zihuiwen Ye, Pengfei Liu, Jinlan Fu, and Graham Neubig. 2021. Towards More Fine-grained and Reliable NLP Performance Prediction. In Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics (EACL), page 3703-3714.  
+Dongxu Zhang, Subhabrata Mukherjee, Colin Lockard, Xin Dong, and Andrew McCallum. 2019. OpenKI: Integrating Open Information Extraction and Knowledge Bases with Relation Inference. In North American Chapter of the Association for Computational Linguistics (NAACL).  
+Markus Zopf and Kiril Gashteovski. 2023. FineGrained Concept Identification for Open Information Knowledge Graph Population. US Patent App. 17/572,685.
+
+# A Detailed Discussion on Related Work
+
+In §2, we go over the problem statement, and in §6 we discussed how the different facets of our benchmark relate to prior and closely related work. Here, we provide a broader discussion where we group the related work based on the problem they address, and provide a more detailed discussion w.r.t. the differences with our work.
+
+Open Information Extraction (OIE). OIE methods extract structured surface-form factual information from natural language text data, in the form of ("subject"; "relation"; "object")-triples (Banko et al., 2007). Such systems are typically either rule-based (Mausam et al., 2012; Corro and Gemulla, 2013; Angeli et al., 2015; Gashteovski et al., 2017; Lauscher et al., 2019; Zopf and Gashteovski, 2023) or neural-based (Stanovsky et al., 2018; Hohenecker et al., 2020; Kotnis et al., 2022c,b; Bayat et al., 2022). Recent research showed that neural OIE systems still fall behind rule-based systems (Gashteovski et al., 2022a; Friedrich et al., 2022). To use the best of both worlds, in this work we use two rule-based systems: MinIE (Gashteovski et al., 2017) and Stanford OIE (Angeli et al., 2015); and two neural systems: MI-LIE (Kotnis et al., 2022c) and Multi $^{2}$ OIE (Ro et al., 2020).
+
+Closed Information Extraction. Given an input text, ClosedIE methods extract a set of (subject; relation; object)-triples where each triple can be expressed within the predefined schema—fixed sets of entities and predicates—of the reference KG (Josifoski et al., 2022; Trisedya et al., 2019; Sui et al., 2021; Josifoski et al., 2023). To ensure that each triple belongs to the KG, these methods prune candidates outside of the KG schema. Because of the pruning, these methods are unable to generate triples which contain entities or predicates that are not already in the KG. In turn, the task we address—OIE linking to KGs—allows us to both link OIE to existing KG candidates, as well as detect novel candidates which are outside of the KG schema.
+
+Open Information Extraction and Knowledge Graphs. The context of OIE facts is used for many tasks for knowledge graphs, such as knowledge graph population (Broscheit et al., 2017; Lin et al., 2020; Zopf and Gashteovski, 2023), (open) link prediction (Broscheit et al., 2020; Kotnis et al., 2023), entity linking (Nanni et al., 2019) and entity
+
+alignment (Friede and Gashteovski, 2022). Therefore, the task of OIE fact linking to KGs is of great importance, as OIE facts can provide knowledge that is outside of the KG schema, as well as knowledge that aligns with the KG schema (Gashteovski et al., 2020b). Several methods (Zhang et al., 2019; Jiang et al., 2021; Wood et al., 2021) address the problem of linking OIEs to KGs. These methods, however, make the assumption that the subject and object OIE slots (which link to KG entities) are linked a priori. These methods can link inductive OIE relations (outside of the training data), however, the assumption that the OIE entity slots are linked beforehand renders these methods not applicable for our task of linking (free-text) surface facts—also known as OIEs—to canonical large-scale KGs.
+
+Knowledge Graph Link Prediction. These methods (Daza et al., 2021; Wang et al., 2021, 2022; Peng et al., 2022) address the problem of Knowledge Graph population by predicting the missing facts in the KG given only the current set of facts. They assign higher scores to valid KG facts (composed of entities and predicates currently in the graph), which need to be added in the graph, and lower scores to erroneous facts (i.e., wrong fact proposals). Even though, during inference, these methods deal with entities and predicates which were unseen during training, an implausible assumption is made that information is already extracted and canonicalized as a KG facts, which is never the case in practice.
+
+Multifaceted Evaluation. NLP and KG tasks are typically evaluated on a held-out test set, by using evaluation frameworks that assign performance scores on a single value; e.g., accuracy (Petroni et al., 2020). In recent years, researchers have observed that such evaluation protocols are somewhat limited (Jain et al., 2023; Ye et al., 2021; Liu et al., 2021), because they do not expose the particular types of problems that the models might have. Hence, with such evaluation protocols, the tested models are more difficult to diagnose when they make errors (Ribeiro et al., 2020). Following prior work on multifaceted OIE (Gashteovski et al., 2022b) and multifaceted KG evaluation (Meilicke et al., 2018; Rim et al., 2021; Widjaja et al., 2022), we propose FaLB: a multifaceted evaluation framework for fact linking. FaLB allows fine-grained evaluation that helps users to pinpoint the source
+
+of error (e.g., on which slot an examined model makes an error), which makes the benchmark more interpretable and useful for subsequent diagnostics of the models. In addition, our benchmark evaluates the performance of the models in different scenarios: inductive, transductive, polysemous, and out-of-KG. With this, the evaluation framework is more human-centric, in a sense that it can help users identify the model that they want for their needs instead of relying on single-score metrics (Kotnis et al., 2022a; Saralajew et al., 2022).
+
+# B Benchmark Datasets Statistics
+
+In Table 6 we report statistics of the benchmark datasets we create on top of REBEL and SynthIE, using the Wikidata Knowledge Graph.
+
+# C Implementation Details
+
+We train all models for 10 epochs using AdamW with a learning rate of 5e-5 and weight decay of 1e-3. We use a RoBERTa (Liu et al., 2019) model, distilled following the procedure of Sanh et al. (2019). The model consists of 6 layers, a hidden size of 768, and has 12 self-attention heads. To reduce the computational complexity and memory demands, we further linearly project the embeddings obtained from the RoBERTa model to a 200-dimensional latent space. When training the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  models, we initialize the temperature  $\tau$  to 0.07 as per Radford et al. (2021). We train  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  models with negatives that are sampled from the whole KG, where we sample 128 negative KG entities and 64 negative KG predicates. When training the  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$ , we sample negatives such that we corrupt the slots of the KG fact by replacing them with incorrect ones. Instead of choosing the negatives at random, for each KG entry, we find its top-10 most similar KG entries, and sample negatives from this subset. This ensures that the  $\mathrm{Fact}_{\mathrm{ranker}}^{\mathrm{re}}$  model learns how to refine the predictions of the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ . We implement everything using PyTorch (Paszke et al., 2017), while we use HuggingFace transformers (Wolf et al., 2020) for the RoBERTa implementation. Lastly, we use Faiss (Johnson et al., 2019) to enable fast linking to Knowledge Graphs.
+
+# D Discussion on the Quality of REBEL and SynthIE
+
+While REBEL's entities are golden (i.e., obtained as the hyperlinks from Wikipedia abstracts which link to Wikipedia pages), the predicates between
+
+them are obtained using a set of heuristics. This leads to imbalanced data, where most predicates occur only few times, and others occur orders of magnitude more. Consequently, such issue is reflected in the OIE-to-KG fact linking data that we obtain. To address this problem of REBEL, Josifoski et al. (2023) proposed SynthIE: a synthetically generated dataset which deals with the imbalance.
+
+Josifoski et al. (2023) perform human evaluation to verify whether their synthetically generated dataset (of natural language sentences pair with KG facts) is of higher quality than REBEL. They obtain 44 data samples from REBEL, and generate a sentence using a Large Language Model (LLM) given only the KG triplets from the sample. Finally, they verify the triplet-set-to-text compatibility for both REBEL and SynthIE. Importantly, human evaluators find that the LLM generated sentences (i.e., the synthetic ones) are more compatible with the set of the KG facts (that is, the validity of the KG facts is higher in SynthIE compared to REBEL).
+
+On the other hand, the issue we observe with SynthIE, is that due to the way the data is provided to LLM, the surface-form entities remain in their canonical text-form (i.e., their canonical denotation in the KG), and thus contain little variation. This diverges from the data that we find "in the wild". Critically, when people refer to entities in free-form natural language, they commonly use synonyms, aliases, abbreviations, nicknames, etc. For example, the former basketball player "Michael Jordan" could be referred to as "Air Jordan" and "M.J."
+
+To cope with this issue, we leverage the entity alias augmentation step in FaLB. By increasing the diversity of the OIE entity surface form, we are able to obtain a high quality OIE-to-KG fact linking synthetic dataset, thus the only human-annotated component remains to be the KG.
+
+# D.1 Inductive Splits between Datasets
+
+Importantly, the inductive evaluation is testing OIE linking to KG facts that consist of entities which were not part of the models' training data. In §5.1<sup>10</sup>, we perform experiments by training models on REBEL and SynthIE, and then evaluate how well they perform on inductive data splits w.r.t. each of the datasets. Namely, to obtain an inductive REBEL testing split w.r.t. SynthIE, we find all testing samples from REBEL, which contain KG enti
+
+<table><tr><td>Backbone Dataset</td><td>Split Type</td><td># Total Samples</td><td># Unique Entities</td><td># Unique Predicates</td><td># Unique Facts</td></tr><tr><td>REBEL</td><td>Training</td><td>5,638,244</td><td>572,020</td><td>555</td><td>613,047</td></tr><tr><td>SynthIE</td><td>Training</td><td>7,749,603</td><td>685,959</td><td>757</td><td>766,032</td></tr><tr><td>REBEL</td><td>Full Validation</td><td>421,547</td><td>48,103</td><td>303</td><td>41,092</td></tr><tr><td>SynthIE</td><td>Full Validation</td><td>48,347</td><td>6,827</td><td>466</td><td>4,766</td></tr><tr><td>REBEL</td><td>Full Testing split</td><td>427,961</td><td>48,807</td><td>314</td><td>41,767</td></tr><tr><td>SynthIE</td><td>Full Testing split</td><td>240,300</td><td>28,660</td><td>635</td><td>22,661</td></tr><tr><td>REBEL</td><td>Testing Transductivity</td><td>241,995</td><td>14,730</td><td>192</td><td>13,496</td></tr><tr><td>REBEL</td><td>Testing Inductivity</td><td>10,300</td><td>4,423</td><td>185</td><td>2,297</td></tr><tr><td>SynthIE</td><td>Testing Inductivity</td><td>6,592</td><td>3,372</td><td>196</td><td>1,775</td></tr><tr><td>REBEL</td><td>Testing Polysemy</td><td>15,339</td><td>3,504</td><td>117</td><td>3,290</td></tr><tr><td>SynthIE</td><td>Testing Out-of-KG</td><td>1,604</td><td>443</td><td>106</td><td>264</td></tr></table>
+
+Table 6: REBEL and SynthIE: overview of the number of samples in each dataset, number of unique KG entities, number of unique KG predicates, number of unique KG facts. Reported for all data splits of REBEL and SynthIE.  
+
+<table><tr><td>Dataset Type</td><td># Samples</td><td># Entities</td><td># Predicates</td><td># Facts</td></tr><tr><td>REBEL w.r.t. Synthie</td><td>26,937</td><td>8,781</td><td>181</td><td>5,304</td></tr><tr><td>SynthIE w.r.t. REBEL</td><td>17,931</td><td>6,318</td><td>488</td><td>3,499</td></tr></table>
+
+Table 7: Overview of the total number of samples, number of unique KG entities, number of unique KG predicates, number of unique KG facts, and the. Reported for inductive splits w.r.t. each dataset.
+
+ties that are not part of any SynthIE training samples. In turn, to obtain an inductive SynthIE testing split w.r.t. REBEL, we find all SynthIE testing samples which contain KG entities that are not part of any training samples from REBEL. We report statistics of the inductive splits w.r.t. each dataset in Table 7.
+
+# E Details on the Out-of-Knowledge Graph Detection Task
+
+In §5.2 we evaluate the ability of the models to detect whether an OIE slot (surface-form entity, or surface-form relation) is present in the Knowledge Graph. Intuitively, this task is more difficult than the OIE linking task, as the models need to generalize beyond the training data distribution to perform well on this task. Namely, when linking OIEs to a KG, all prior work (Zhang et al., 2019; Jiang et al., 2021; Wood et al., 2021) makes the conjecture that the testing data (in the open-world) is independent and identically distributed (i.i.d.) w.r.t. the training data, which is an invalid assumption in certain scenarios. Notably, the Out-of-KG OIEs lie outside of the training data distribution. Therefore, a model that performs well on the OIE linking task does not warrant high performance on the Out-of-KG detection task. We observed that this is the case
+
+(i.e., a model performs well on the linking task, but performs poorly on the Out-of-KG task) when detecting out-of-KG relations.
+
+Therefore, to ensure that the Out-of-KG data is indeed out-of-distribution (as it would be the case in practice), we impose the following constraints: (i) We select OIE-to-KG pairs from SynthIE, while the "backbone" model we build on top of is trained on REBEL; (ii) All entities and predicates—which are part of the KG facts from the testing data—are not in the KG at the time of training the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ .
+
+To perform the evaluation, for each OIE slot we either leave its corresponding KG entry outside of the KG, and score a hit if the models predict a negative score for that slot; or, we perform imputation of the Out-of-KG entries (thus, they are now part of the KG), and score a hit if the models predict a positive score for that slot. Finally, we report the averaged accuracy over the two scenarios for each OIE slot. Note that, in this case both micro- and macro-average yield the same number, because we have the same number of samples for each scenario.
+
+# E.1 Out-of-Knowledge Graph Detection Models
+
+All models that we use for the Out-of-KG detection task in §5.2 are built on top of a OIE<sup>pre</sup> ranker, which is trained on REBEL. For all models, to obtain an out-of-KG indicator—True (1), or False (0)—we threshold the output of the models. For each model, we determine the optimal threshold (the confidence, or the entropy) on a hold-out validation dataset which we build on top of REBEL.
+
+CONFIDENCE@1-BASED HEURISTIC: We obtain the KG links for each of the OIE slots using
+
+the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ . The linking is characterized by the cosine similarity between the embeddings of each OIE slot and the KG entries. We then compute the softmax of the top-5 highest cosine similarities, and finally threshold the confidence@1 to obtain a prediction; such that a confidence  $< T_{c}$  indicates an out-of-KG instance, and a confidence  $>T_{c}$  indicates an instance inside the KG, where  $T_{c}$  is the confidence threshold. We use  $T_{c} = [0.235; 0.260; 0.235]$  for detecting out-of-KG subjects, relations and objects respectively.
+
+ENTROPY-BASED HEURISTIC: Similar to the confidence@1-based method, we obtain the cosine similarities with the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$ . However, instead of using the top-1 probability, we obtain the entropy of the top-5 predictions. We finally threshold the entropy to obtain a prediction, such that an entropy  $>T_{e}$  indicates out-of-KG instance, and an entropy  $< T_{e}$  indicates an instance inside the KG, where  $T_{e}$  is the entropy threshold. We use  $T_{e} = [1.60; 1.58; 1.60]$  for detecting out-of-KG subjects, relations and objects respectively.
+
+# QUERY-KEY-VALUE CROSS-ATTENTION:
+
+Using the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  embeddings, we train a lightweight query-key-value cross-attention module on top with weights that are initialized with the identity matrix—at the start of training the  $\mathrm{OIE}_{\mathrm{ranker}}^{\mathrm{pre}}$  embeddings are used as is. Given a query OIE slot embedding, the model attends over the KG entry embeddings representing the keys and values, and outputs a sigmoid normalized score (i.e., a probability) indicating the presence confidence of the OIE slot in the KG. During training, to obtain in-batch negatives, for each OIE slot we drop the KG entry counterpart with  $50\%$  probability. We obtain additional negatives by sampling KG entries from the whole KG, which do not match any of the OIE slots in the batch. We train the model using the binary cross-entropy loss, such that, if the corresponding KG entry for an OIE slot is in the sampled KG subset, the model predicts a positive score averaged over the KG graph subset, and negative otherwise. We use a uniform threshold of  $T_{a} = 0.3$  for all three slots.
+
+# F Data Quality
+
+To assess FaLB's data quality, we performed manual human evaluation. In particular, we did the following steps:
+
+1. We randomly selected 100 data points, where
+
+each data point contained information about the provenance sentence, a KG fact that is contained in the sentence, and a corresponding OIE surface fact that was extracted from the sentence.
+
+2. Two expert annotators annotated each data point independently of whether the KG fact matches the OIE extraction semantically (see two labelled examples in Table 8).  
+3. We considered a data point as "correct" only if the two annotators agreed that the KG fact semantically matches the information in the OIE triple.  
+4. We computed accuracy, inter-annotator agreement and Kohen's kappa score.
+
+We found that  $97\%$  of the data points are considered "correct" by both annotators. We also observed that the inter-annotator agreement was high: the annotators agreed in  $99\%$  of the cases, with high Kohen's kappa score (McHugh, 2012) of 0.80. Please refer to the supplementary material for the subset of samples which was provided to the expert annotators.
+
+<table><tr><td>KG fact (IDs)</td><td>KG fact (names)</td><td>OIE Triple</td></tr><tr><td>Sentence: “Hekimoglu Ali Pasha Mosque was built between 1734–1735 in the Fatih district of Istanbul by Hekimoglu Ali Pasha, who was born in Istanbul in 1689.”(Q1584693; P19; Q406) Label: correct</td><td>(Hekimoglu Ali Pasha; place of birth; Istanbul)</td><td>(“H. Ali Pasha”, “was born in”, “Istanbul”)</td></tr><tr><td>Sentence: “Pierre Hétu (April 22, 1936 in Montreal – December 3, 1998 in Montreal) was a conductor and pianist.”(Q3385492, P19, Q340) Label: incorrect</td><td>(Pierre Hétu; place of birth; Montreal)</td><td>(“Pierre Hétu”, “April in”, “Montreal”)</td></tr></table>
+
+Table 8: Example annotations for FaLB data. The annotator sees: (1) the input sentence; (2) the KG fact with the original IDs (thus, the user can further check the meaning of the entity and predicate); (3) the surface text of the KG fact; (4) the OIE triple, extracted from the input sentence. The first matching is labelled as "correct", because the KG fact semantically matches the OIE triple. The second matching is labelled as "incorrect", due to the incorrect OIE relation "April in".

2023/Linking Surface Facts to Large-Scale Knowledge Graphs/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9d9cabfaa8b85bfe45ae416684ed523d55aae32bb919b1a3076d25a9ae420a9
+size 562966

2023/Lion_ Adversarial Distillation of Proprietary Large Language Models/317da3bb-97cc-44eb-b7cf-4be112125836_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05110530f85f2ff06fbee259037d48d38cd1219ff2f418009a196b97ce495364
+size 6204518

2023/Lion_ Adversarial Distillation of Proprietary Large Language Models/full.md

@@ -0,0 +1,483 @@
+# Lion: Adversarial Distillation of Proprietary Large Language Models
+
+Yuxin Jiang $^{1,2}$  Chunkit Chan $^{2*}$  Mingyang Chen $^{1,2*}$  Wei Wang $^{1,2}$
+
+<sup>1</sup>The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China
+
+$^{2}$ The Hong Kong University of Science and Technology, Hong Kong SAR, China {yjiangcm, ckchancc, mchenbt}@connect.ust.hk, weiwcs@ust.hk
+
+# Abstract
+
+The practice of transferring knowledge from a sophisticated, proprietary large language model (LLM) to a compact, open-source LLM has garnered considerable attention. Previous works have focused on a unidirectional knowledge distillation way by aligning the responses of the student model with those of the teacher models to a set of instructions. Nevertheless, they overlooked the possibility of incorporating any "feedback"—identifying challenging instructions where the student model's performance falls short—to boost the student model's proficiency iteratively. To this end, we propose a novel adversarial distillation framework for a more efficient knowledge transfer. Leveraging the versatile role adaptability of LLMs, we prompt the teacher model to identify "hard" instructions and generate new "hard" instructions for the student model, creating a three-stage adversarial loop of imitation, discrimination, and generation. By applying this adversarial framework, we successfully transfer knowledge from ChatGPT to a student model (named Lion), using a mere 70k training data. Our results show that Lion-13B not only achieves comparable open-ended generation capabilities to ChatGPT but surpasses conventional state-of-the-art (SOTA) instruction-tuned models like Vicuna-13B by  $55.4\%$  in challenging zero-shot reasoning benchmarks such as BIG-Bench Hard (BBH) and  $16.7\%$  on AGIEval. $^{1}$
+
+# 1 Introduction
+
+Large language models (LLMs) capable of following natural language instructions have exhibited tremendous success in generalizing zero-shot to new tasks (Mishra et al., 2022; Wei et al., 2022a). Due to various concerns, the most advanced LLMs, such as ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023) that boasting billions of parameters, are
+
+![](images/a5326d200f826d9bfe7fee71a7dbdf362d79ecd59fe6b986ffbde5b20710b09d.jpg)  
+Figure 1: An illustration of the distinction between our approach and earlier ones. Previous methods facilitate a one-way knowledge transfer from the teacher to the student (solid arrow). Our approach, however, incorporates an innovative step (dashed arrow) that completes a loop: it enables the feedback"—identifying the student model's weaknesses—to be relayed back to the teacher, in order to foster tailored learning.
+
+![](images/f7ced28f5acd6057cdf767ddac393a6cf6473d3b03bd9b8ff118c7296b5287b8.jpg)
+
+<table><tr><td>Instruction: What are the three primary colors?
+Teacher&#x27;s Response: The three primary colors are red, blue, and yellow.
+Student&#x27;s Response: The three primary colors are: red, blue, and yellow.</td><td>easy</td></tr><tr><td>Instruction: Calculate the following mathematical expression: 4 + 5(6-3).
+Teacher&#x27;s Response: The result of the expression is 29. The calculation proceeds as follows: 4 + 5(6-3) = 4 + 5(3) = 4 + 15 = 19.
+Student&#x27;s Response: After calculating, the result is 4 + 5(6-3) = 15. ×</td><td>hard</td></tr><tr><td>…</td><td>…</td></tr></table>
+
+typically proprietary, comprising both the model parameter and the training data. To foster increased transparency regarding their intricate operational mechanics, a surge in research efforts focusing on knowledge distillation from a proprietary "teacher" LLM to an open-source "student" LLM. This is typically accomplished by aligning the responses of the student model with those of the teacher model to a set of instructions, which can be manually or automatically generated (Wang et al., 2022; Taori et al., 2023; Chiang et al., 2023; Xu et al., 2023).
+
+However, previous works employ a unidirectional approach to knowledge transfer (solid arrow in Figure 1), where the teacher imparts knowledge to the student without considering any "feedback".
+
+To better illustrate this using a tangible classroom scenario, the "feedback" refers to identifying the "hard" examples or problems where the student's performance falls short. This feedback guarantees that the teacher can provide bespoke training that centers on "hard" examples, thereby paving the way for more effective and tailored learning experiences for the student.
+
+Inspired by adversarial knowledge distillation (AKD), which aims to iteratively improve the student model's performance by learning from generated hard samples (Fang et al., 2019; Micaelli and Storkey, 2019a; Heo et al., 2019), we propose an adversarial framework for distilling a proprietary LLM into a compact student model. Nevertheless, these AKD methodologies necessitate accessibility to the weights or gradients of the teacher model, which cannot be directly adapted to our setting. To circumvent this problem, we leverage the unparalleled role adaptability of LLMs, which can be effectively employed through a diverse range of prompts (Sanh et al., 2022). In particular, we prompt the proprietary teacher LLM to serve as a "referee" to discriminate hard instructions where there exists a significant performance discrepancy between the teacher's and student's responses, and serve as a "generator" to produce new instructions that emulate the data distributions corresponding to the discriminated hard instructions. Our framework, as depicted in Figure 2, consists of three stages in an iteration: 1) an imitation stage to align the student's response with the teacher's response; 2) a discrimination stage to identify hard instructions; 3) A generation stage to produce new hard instructions for escalating the challenges presented to the student model. In essence, our adversarial framework forms a positive feedback loop that efficiently bootstraps the student model's proficiency.
+
+To verify the efficiency and efficacy of our method, we apply our AKD framework to transfer the knowledge of ChatGPT 2 onto an open-source foundation LLM, known as LLaMA (Touvron et al., 2023). We select Alpaca's training data (generated from only 175 manually selected seed instructions) as the initial training instructions and execute three iterations of AKD, resulting in a total of 70K data that our model is trained on. We've christened our model as Lion, drawing inspiration from the art of "distillation". By conducting extensive exper
+
+iments on open-ended generation and reasoning datasets, which include a total of 40 sub-tasks, our Lion-13B showcases superior performance surpassing instruction-tuned baseline models such as Vicuna (Chiang et al., 2023). Our main contributions are as follows:
+
+- Our work is the first attempt to adopt the idea of adversarial knowledge distillation to large language models.  
+- Our proposed framework demonstrates impressive efficiency and efficacy. With instruction tuning performed on 70k data without any human annotation, our Lion-13B approximates ChatGPT's capabilities on open-ended generation dataset and largely outperforms the current SOTA model Vicuna-13B on reasoning tasks.  
+- The versatility of our framework allows for broad application: it is not exclusive to ChatGPT but can be conveniently adapted to suit a variety of other proprietary LLMs.
+
+# 2 Related Work
+
+# 2.1 Instruction-Following Language Models
+
+With the impressive ability of instruction-following large language models such as ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023), the techniques of instruction tuning (Wei et al., 2022b) have attracted a lot of attention (Wei et al., 2022c; Bubeck et al., 2023; Bang et al., 2023; Chan et al., 2023a). The early research of instruction tuning aims to enhance the generalization ability of language models, allowing these models to perform new tasks by comprehending task descriptions without relying on a few exemplars. By fine-tuning these instruction-following language models (e.g., T5 (Raffel et al., 2020), FLAN (Aribandi et al., 2022), T0 (Sanh et al., 2022), and ExT5 (Aribandi et al., 2022)) on multi-task datasets in the form of natural language phrased as instructions, these models have been shown to perform well on unseen tasks with the instructions.
+
+However, these models are only fine-tuned on simple task-specific instructions, and it is challenging to comprehend the sophisticated and diverse intent of users in real-world scenarios. Therefore, InstructGPT (Wei et al., 2022b), ChatGPT (OpenAI, 2022), and GPT-4 (OpenAI, 2023) trained on the diverse forms and abundant task types of
+
+human-crafted instructions annotated by a considerable number of annotators. Since these instructions were not open-sourced, recent works such as Alpaca (Taori et al., 2023), Vicuna (Chiang et al., 2023), and WizardLM (Xu et al., 2023) investigate how to generate high-quality instructions and fine-tune the open-source large language model LLaMA (Touvron et al., 2023) with them to approach the performance of ChatGPT.
+
+# 2.2 Knowledge Distillation
+
+Knowledge Distillation (KD) (Hinton et al., 2015; Radosavovic et al., 2018; Chen et al., 2019) represents a crucial strategy within the sphere of model compression and acceleration, wherein a compact student model is instructed to emulate the performance traits of a more cumbersome teacher model. In practical contexts, the availability of training data is often constrained due to concerns regarding privacy, legality, security, or confidentiality. To address the absence of training data, data-free KD methods were proposed to align the student model to the teacher model, capitalizing on either related proxy data (Orekondy et al., 2019; Papernot et al., 2017) or synthetic data generated by learnable generators (e.g., Generative Adversarial Network (GAN)) (Addepalli et al., 2020; Fang et al., 2019; Micaelli and Storkey, 2019b) or teacher model inversions (Yin et al., 2020; Chawla et al., 2021; Fang et al., 2022). Nevertheless, these KD methodologies necessitate the accessibility to the weights or gradients of the teacher model. Consequently, an alternative line of research, commonly denoted as data-free model extraction (or stealing), endeavors to bridge this gap by employing zero-order estimation methodologies to approximate the authentic gradients of the teacher model to guide the update of the optimized generators (Kariyappa et al., 2021; Truong et al., 2021). However, adapting these methods to our distillation task presents two main hurdles. First, these techniques are primarily designed for image-based classification tasks, assuming access to a continuous softmax vector from the teacher model. Estimating zero-order gradients becomes problematic in our case, as responses are typically sequence-oriented. Second, developing an effective instruction generator capable of producing diverse, high-quality instructions that mirror the teacher model's training data distribution proves more challenging than in the image domain.
+
+# 3 Methodology
+
+Harnessing the learned knowledge of a sophisticated teacher model  $\mathcal{T}(x;\theta^{\mathcal{T}})$  where the parameter  $\theta^{\mathcal{T}}$  is inaccessible, our goal is to craft a more lightweight student model  $S(x;\theta^{S})$ . Ideally, a student model is optimal if the expectation of model discrepancy (which indicates the prediction differences between teacher  $\mathcal{T}$  and student  $S$ ) on the uniform data distribution is minimized. Inspired by the success of adversarial knowledge distillation (AKD) (Fang et al., 2019; Micaelli and Storkey, 2019a; Heo et al., 2019), we turn to optimize an upper bound of the expectation—the expectation of the model discrepancy on "hard samples", where the teacher  $\mathcal{T}$  and the student  $S$  have a relatively large performance gap. These "hard samples" are inclined to dominate the expectation of the model discrepancy. Thus, the overall expected model discrepancy can be effectively and efficiently reduced by optimizing the student model  $S$  on these "hard samples". The underlying rationale is rather straightforward and can be analogized to a real-world educational scenario: continuously concentrating on the "hard" knowledge that the student finds challenging to grasp is the most effective manner of enhancing a student's proficiency.
+
+However, in the process of training the student model  $S$ , hard samples will be mastered by the student and converted into easy samples. Hence we need a mechanism to continuously generate hard samples, which can be achieved by an adversarial framework.
+
+The whole framework of our Adversarial Knowledge Distillation is depicted in Figure 2, which contains three stages in an iteration: 1) an imitation stage to align the student's response with the teacher's response; 2) a discrimination stage to identify hard samples; 3) A generation stage to produce new hard samples for escalating the challenges presented to the student model.
+
+# 3.1 Initilization
+
+As shown in Figure 2, four roles and two data pools are established in our framework, and we will comprehensively illustrate their functions later. We initialize our student model  $S$  using a foundation LLM such as LLaMA (Touvron et al., 2023). We initialize our teacher model  $\mathcal{T}$ , referee  $\mathcal{R}$ , and generator  $\mathcal{G}$  by using the same proprietary LLM such as ChatGPT (OpenAI, 2022). The multiple roles that this proprietary LLM serves are accomplished
+
+![](images/acf4ef917718692a90fc501ae83aa929f9b4bfe1e72716c5ab892f0b611a850b.jpg)  
+Figure 2: The overview of our adversarial distillation framework, where we craft a compact Student LLM  $S$  based on a superior proprietary LLM that serves three roles: the Teacher  $\mathcal{T}$ , the Referee  $\mathcal{R}$ , and the Generator  $\mathcal{G}$ . From left to right, there are three stages in an iteration: 1) Imitation; 2) Discrimination; 3) Generation.
+
+through the use of varied prompt templates. We start the iteration from a given initial Train Pool  $X^{A} = \{x_{i}^{A}\}_{i\in [1,N^{A}]}$ , where  $x_{i}^{A}$  is the  $i$ -th instruction in  $X^{A}$ , and  $N^{A}$  is the number of samples in  $X^{A}$ . The Cache Pool  $X^{B}$  is initialized as identical to  $X^{A}$ , consisting of instructions to evaluate the performance of  $S$  and  $\mathcal{T}$ .
+
+# 3.2 Imitation Stage
+
+To impart the knowledge of the teacher to the student, we construct the instruction-response data  $\{x_i^A,\mathcal{T}(x_i^A)\}_{i\in [1,N^A ]}$  by forward propagating instructions in the Train Pool  $X^{A}$  through the teacher  $\mathcal{T}$ . The prompt template used for model inference is shown in Table 10. Like the imitation training of previous work (Taori et al., 2023; Chiang et al., 2023), we fine-tune our student model  $S$  to align the response of the teacher model, by optimizing the autoregressive language modeling objective.
+
+# 3.3 Discrimination Stage
+
+Figure 2 demonstrates that the discrimination stage starts from the Cache Pool, denoted as  $X^B$ . Even though this pool begins with the same initialization as the Train Pool, their uses diverge. The Train Pool is rejuvenated by replacing its existing instructions with freshly generated instructions, whereas the Cache Pool is enriched by incorporating these generated instructions. As a result, the growing storage capacity of the Cache Pool provides a more extensive space for evaluating the performance gap between teacher  $\mathcal{T}$  and student  $S$ . This allows for
+
+more thorough detection of hard instructions.
+
+In the discrimination stage, we ask the proprietary LLM to serve as a "referee", which quantifies the performance gap between  $\mathcal{T}$  and  $S$ . Specifically, we feed each instruction  $x_{i}^{B}$  in the Cache Pool  $X^{B}$  through both the teacher  $\mathcal{T}$  and student  $S$  to generate the outputs  $\mathcal{T}(x_{i}^{B})$  and  $\mathcal{S}(x_{i}^{B})$ , respectively. Then we ask the referee  $\mathcal{R}$  to quantitatively measure the quality difference between teacher's response  $\mathcal{T}(x_{i}^{B})$  and student's response  $\mathcal{S}(x_{i}^{B})$ , conditioned on  $x_{i}^{B}$ :
+
+$$
+d _ {i} = \mathcal {R} (\mathcal {T} (x _ {i} ^ {B}), \mathcal {S} (x _ {i} ^ {B}) \mid x _ {i} ^ {B}) \tag {1}
+$$
+
+The above process is conducted by using the prompt template (as shown in Table 11) inspired by (Chiang et al., 2023), which requires the LLM to consider the helpfulness, relevance, accuracy, and level of detail of two responses and output two scores. To mitigate the positional bias (Wang et al., 2023) of the LLM referee, we conduct two runs by exchanging the positions of the teacher's response and the student's response and compute the final score as the average of the two runs. Then  $d_{i}$  is calculated as the difference between the teacher's score and the student's score. By setting a threshold  $\tau$  (1.0 used in our experiments), we discriminate hard instructions as those instructions with  $d_{i} \geq \tau$  and the others are identified as easy ones. Figure 3b provides a clear and intuitive demonstration of which kinds of instructions are discriminated as hard in the first iteration. Compared with the instructions in the Cache Pool (Figure 3a), the dis
+
+![](images/1ec7f110da1afda9f7fa4398611b2545608e73cb13d0de2b7fb5d3c425cf6ea7.jpg)  
+(a) Instructions of the Cache Pool in the first iteration.
+
+![](images/7f4259dcb8998bea617c8c8a2e386d623886f6803c72f1bc9702ceade73f6d6d.jpg)  
+(b) Identified hard instructions in the first iteration.  
+Figure 3: The top 20 most common root verbs (inner circle) and their top 4 direct noun objects (outer circle) in the instructions.
+
+![](images/281d7c923190a3b579e3fa0bf1d7122274dd6f076dfba19d1dcc0377d82abac4.jpg)  
+(c) Generated hard instructions in the first iteration.
+
+tribution of the identified hard instructions is quite different, focusing more on complex tasks such as math, coding, etc.
+
+# 3.4 Generation Stage
+
+After carefully discerning the hard instructions, the generation stage aims to produce samples that mirror the data distributions corresponding to these challenging directives. This process is achieved by employing the proprietary LLM as a generator, denoted as  $\mathcal{G}$ , leveraging its exceptional prowess in content creation. Inspired by (Xu et al., 2023), we randomly sample an instruction from the hard instructions and prompt the generator  $\mathcal{G}$  to generate a new instruction. The newly generated instruction is required to pertain to the same domain and match the task type of the sampled instruction. The template utilized for this prompt is exhibited in Table 12. As shown in Figure 3c, the distribution of the newly generated hard instructions appears to be comparable to that of the previously identified hard instructions. To mitigate the issue of catastrophic forgetting and to augment the diversity of the generated instructions, we also randomly sample an instruction from the easy instructions and prompt the generator  $\mathcal{G}$  to generate a new instruction that belongs to the same domain as the sampled one, but exhibit a more long-tailed distribution. The template we use to prompt this process is displayed in Table 13.
+
+In each iteration, we define  $N$  as the total count of newly generated instructions and maintain a 1:1 ratio  $r$  between the generated hard instructions and the generated easy instructions. To promote diversity, a new instruction will be deemed valid only if its ROUGE-L overlap with any existing instructions in the Cache Pool is below 0.7. Finally, as
+
+aforementioned in Section 3.3, we proceed to rejuvenate the Train Pool, replacing its existing instructions with freshly generated ones. Concurrently, we enrich the Cache Pool by incorporating these newly generated instructions.
+
+# 3.5 Min-Max Game Interpretation
+
+Our adversarial knowledge distillation framework can be interpreted as a dynamic min-max game: in the imitation stage, we fine-tune our student to minimize the model discrepancy between itself and the teacher on hard samples; in the discrimination and generation stage, we craft new hard samples to maximize the model discrepancy, based on the learning progress of the student model. This dialectic framework propels the student model towards uncovering otherwise hidden knowledge, paving the way to complete understanding. As the training progresses through several iterations, the system should ideally achieve equilibrium. This is the point where the student model has mastered all the hard samples and the referee  $\mathcal{R}$  can no longer distinguish between the student  $S$  and teacher  $\mathcal{T}$  models. At this juncture,  $S$  becomes functionally indistinguishable from  $\mathcal{T}$ .
+
+# 4 Experiments Setting
+
+# 4.1 Datasets
+
+In our experiments, we implemented a comprehensive LLM evaluation protocol that considers a diverse range of abilities, such as writing, coding, commonsense, math, and logical reasoning. The datasets we utilized can be classified into two main categories: open-ended generation and reasoning.
+
+# 4.1.1 Open-ended Generation Datasets
+
+Vicuna-Instructions (Chiang et al., 2023) is a set of 80 questions spanning 9 distinct task categories. This dataset has gained extensive usage in evaluating the capabilities of LLMs. Within our work, we examine LLMs' performance on this dataset in two different settings:
+
+- Setting1: Following Vicuna (Chiang et al., 2023), we leverage GPT-4 to automatically assess the quality of responses (rated on a scale of 1 to 10) between a reference model (ChatGPT) and a candidate model. Subsequently, we calculate the candidate model's performance as the percentage of the total score it achieves compared to the reference model.  
+- Setting2: A recent work (Wang et al., 2023) pointed out that a systematic bias may exist in the above-mentioned GPT-4 automatic evaluation. To mitigate this, they propose two strategies, namely Multiple Evidence Calibration and Balanced Position Calibration, to obtain closer alignment with human judgments.
+
+# 4.1.2 Reasoning Datasets
+
+AGIEval (Zhong et al., 2023) is a well-known benchmark that quantifies the reasoning capability of foundation models in the context of human-centric standardized exams, including college entrance exams, math competitions, lawyer qualification tests, etc. We choose all English multiple-choice questions (8 tasks, 2,546 samples) among AGIEval for our experiments. The data statistics are shown in Table 6.
+
+BIG-Bench Hard (BBH) (Suzgun et al., 2022) consists of a suite of challenging tasks from BIG-Bench (Srivastava et al., 2022), designed to assess the capabilities and limitations of large language models. These are the tasks on which prior language models underperform the average human rater. We choose all tasks that can be formatted into multiple-choice questions (23 tasks, 5,511 samples) among BBH for our experiments. The data statistics are shown in Table 7.
+
+Setting We evaluate reasoning capabilities under a zero-shot setting without any exemplars and without Chain-of-Thought (CoT). For both AGIEval and BBH, we use the prompt format and parsing following (Zhong et al., 2023; Mukherjee et al.,
+
+2023). Given the free-form response from the generative models, only the first capital character in the response is considered to compare with the gold answer (exact match). The result we report is accuracy  $(\%)$ .
+
+# 4.2 Baselines
+
+We select five superior LLMs as baselines, including LLaMA (Touvron et al., 2023), Alpaca (Taori et al., 2023), WizardLM (Xu et al., 2023), Vicuna (Chiang et al., 2023), and ChatGPT (OpenAI, 2022). It is worth noting that Vicuna has consistently ranked as the top open-source language model on multiple leaderboards, such as Chatbot Arena<sup>3</sup>. Therefore, we will conduct a comprehensive comparison with Vicuna. See detailed descriptions of these baselines in Appendix B.
+
+# 4.3 Implementation Details
+
+Training Details Our student model is initialized using the pre-trained LLaMA. The Train Pool and Cache Pool are initialized with the 52K automatically generated instructions from Alpaca (Taori et al., 2023). The total number of iterations is set to 3, with 6K newly generated instructions added at each iteration. This results in a total of 70K data that our model is trained on in order to make a fair comparison with current SOTA baselines, including WizardLM and Vicuna. The training hyperparameters are listed in Appendix C.
+
+Inference Details To draw inferences from Lion and ChatGPT, we calibrated the temperature to 0.7 and set the maximum generation length at 1024. All other parameters adhere to their default settings. For LLaMA, Alpaca, WizardLM, and Vicuna, we configured their inference parameters in line with the specifications given in their respective original papers. When engaging with the gpt-3.5-turbo API for various roles, we employ an array of hyperparameters, the specifics of which can be located in Appendix C.
+
+# 5 Experimental Results
+
+# 5.1 Results for Open-ended Generation
+
+Table 1 shows the performance comparison of various models against ChatGPT as the reference model, where GPT-4 is used as a referee/rater. Our Lion-7B and Lion-13B remarkably outperform their counterparts under two evaluation settings.
+
+<table><tr><td>Model</td><td>Setting1</td><td>Setting2</td><td>Avg.</td></tr><tr><td>LLaMA-7B</td><td>58.46</td><td>59.12</td><td>58.79</td></tr><tr><td>Alpaca-7B</td><td>69.29</td><td>67.20</td><td>68.25</td></tr><tr><td>WizardLM-7B</td><td>89.29</td><td>86.67</td><td>87.98</td></tr><tr><td>Vicuna-7B</td><td>87.79</td><td>89.96</td><td>88.88</td></tr><tr><td>Lion-7B</td><td>94.74</td><td>92.88</td><td>93.81</td></tr><tr><td>LLaMA-13B</td><td>69.23</td><td>68.21</td><td>68.72</td></tr><tr><td>Alpaca-13B</td><td>76.87</td><td>74.69</td><td>75.78</td></tr><tr><td>Vicuna-13B</td><td>92.25</td><td>92.97</td><td>92.61</td></tr><tr><td>Lion-13B</td><td>96.57</td><td>100.18</td><td>98.38</td></tr></table>
+
+Table 1: Relative response quality (%) against ChatGPT (assessed by GPT-4) on Vicuna-Instructions.
+
+![](images/3d23657f69371698af7292fece730806c0471c727e3ab0413e5158ee1772df9e.jpg)  
+Figure 4: Relative response quality against ChatGPT on diverse task categories of Vicuna-Instructions.
+
+Noticeably, Lion-13B shows an 8-point improvement over Vicuna-13B on aggregate, achieving  $98.38\%$  capabilities of ChatGPT.
+
+To comprehensively compare with other baseline models on the capability to generate high-quality responses on various types of instruction, the relative response quality (Setting2) among different task categories is depicted in Figure 4. Our model impressively and slightly surpasses ChatGPT in the generic, knowledge, common-sense, and counterfactual task categories. Furthermore, for the two difficulty task categories described in the previous study (Chiang et al., 2023; Xu et al., 2023), our model significantly outperforms other baseline models with at least  $32.32\%$  relative score in the math task category while exceeding most of the baseline in the coding generation task category.
+
+# 5.2 Results for Reasoning
+
+AGIEval Results Table 2 presents the standard zero-shot performance comparison between Lion and baseline models on the AGIEval benchmark for multiple-choice English questions. Lion demonstrates significantly stronger performance compared to Vicuna, surpassing it in most task cate
+
+gories and achieving an average relative improvement of over  $16\%$ . However, Lion-13B still significantly lags behind ChatGPT, only retaining  $72.5\%$  of its reasoning capability.
+
+Big-Bench Hard Results Table 3 displays the zero-shot performance comparison between Lion and baseline models on BIG-Bench Hard with standard zero-shot prompting. Similar to AGIEval, Vicuna exhibits poor performance on sophisticated reasoning tasks within this benchmark, while Lion substantially surpasses Vicuna by around  $50\%$  on average. Particularly, Lion demonstrates significant performance enhancements of over  $100\%$  on tasks involving data understanding, semantic understanding (Disambiguation QA and Snarks), logical and geometric reasoning (Logical Deduction and Geometric Shapes), and position reasoning (Tracking Shuffled Objects). Despite achieving an average ability of nearly  $74\%$  compared to ChatGPT on BBH, Lion-13B surpasses ChatGPT in several tasks, including Movie Recommendation, Snarks (identifying sarcastic sentences from two nearly-identical ones), and Tracking Shuffled Objects. This demonstrates the effectiveness of our method.
+
+# 6 Analyses
+
+# 6.1 Ablation Studies
+
+The threshold  $\tau$  for distinguishing between hard and easy instructions We systematically explored  $\tau$  ranging from 0.0 to 2.0 and documented its influence on average performance across three datasets. Table 4 reveals an optimal range of  $\tau$  between 1.0 and 1.5 for all datasets. Notably, elevating  $\tau$  from 0.0 to 1.0 consistently enhances performance across all datasets, indicating effective differentiation between hard and easy instructions. However, a continuous increase from 1.0 to 2.0 gradually degrades performance due to decreased diversity in hard instructions. The ablation results demonstrate that our method is not quite sensitive to a large value of  $\tau$ .
+
+The ratio  $r$  of generated hard and easy instructions We change the ratio of generated hard instructions to generated easy instructions from 1:0 (all hard) to 0:1 (all easy) and investigate its impact on average performance across three datasets. It can be seen from Table 5 that higher ratios of hard to easy instructions generally lead to improved performance, with a balanced ratio of 1:1 yielding the
+
+<table><tr><td rowspan="2">Task</td><td colspan="2">Human</td><td rowspan="2">ChatGPT</td><td rowspan="2">Vicuna-7B</td><td rowspan="2">Lion-7B</td><td rowspan="2">Vicuna-13B</td><td rowspan="2">Lion-13B</td></tr><tr><td>Avg</td><td>Top</td></tr><tr><td>AQuA-RAT</td><td>85.0</td><td>100.0</td><td>31.9</td><td>23.2</td><td>18.5 (-20.3%)</td><td>20.1</td><td>26.0 (29.4%)</td></tr><tr><td>LogiQA</td><td>86.0</td><td>95.0</td><td>35.0</td><td>21.4</td><td>31.8 (48.6%)</td><td>29.8</td><td>31.3 (5.0%)</td></tr><tr><td>LSAT-AR</td><td>56.0</td><td>91.0</td><td>24.4</td><td>22.2</td><td>17.4 (-21.6%)</td><td>20.4</td><td>23.0 (12.7%)</td></tr><tr><td>LSAT-LR</td><td>56.0</td><td>91.0</td><td>52.6</td><td>18.6</td><td>28.2 (51.6%)</td><td>32.6</td><td>32.6 (0.0%)</td></tr><tr><td>LSAT-RC</td><td>56.0</td><td>91.0</td><td>65.4</td><td>21.9</td><td>29.4 (34.2%)</td><td>32.7</td><td>40.9 (25.1%)</td></tr><tr><td>SAT-Math</td><td>66.0</td><td>94.0</td><td>42.7</td><td>21.4</td><td>20.9 (-2.3%)</td><td>28.6</td><td>29.4 (2.8%)</td></tr><tr><td>SAT-English</td><td>66.0</td><td>94.0</td><td>81.1</td><td>25.7</td><td>36.4 (41.6%)</td><td>44.2</td><td>53.9 (21.9%)</td></tr><tr><td>SAT-English (w/o Psg.)</td><td>66.0</td><td>94.0</td><td>44.2</td><td>26.2</td><td>27.7 (5.7%)</td><td>26.2</td><td>36.2 (38.2%)</td></tr><tr><td>Average</td><td>67.1</td><td>93.8</td><td>47.2</td><td>22.6</td><td>26.3 (16.4%)</td><td>29.3</td><td>34.2 (16.7%)</td></tr></table>
+
+Table 2: Zero-shot performance comparison of ChatGPT, Vicuna, and Lion on AGIEval (multiple-choice English questions). We report the performance of Human, ChatGPT, and Vicuna from (Mukherjee et al., 2023). Performance improvements obtained by Lion over Vicuna are shown in parenthesis.  
+
+<table><tr><td>Task</td><td>ChatGPT</td><td>Vicuna-7B</td><td colspan="2">Lion-7B</td><td>Vicuna-13B</td><td colspan="2">Lion-13B</td></tr><tr><td>Boolean Expressions</td><td>82.8</td><td>39.2</td><td>55.2</td><td>(40.8%)</td><td>40.8</td><td>65.6</td><td>(60.8%)</td></tr><tr><td>Causal Judgement</td><td>57.2</td><td>39.7</td><td>50.3</td><td>(26.7%)</td><td>42.2</td><td>43.9</td><td>(4.0%)</td></tr><tr><td>Date Understanding</td><td>42.8</td><td>8.6</td><td>34.0</td><td>(295.3%)</td><td>10.0</td><td>40.4</td><td>(304.0%)</td></tr><tr><td>Disambiguation QA</td><td>57.2</td><td>15.2</td><td>35.6</td><td>(134.2%)</td><td>18.4</td><td>44.8</td><td>(143.5%)</td></tr><tr><td>Formal Fallacies</td><td>53.6</td><td>40.0</td><td>46.0</td><td>(15.0%)</td><td>47.2</td><td>52.4</td><td>(11.0%)</td></tr><tr><td>Geometric Shapes</td><td>25.6</td><td>3.6</td><td>8.8</td><td>(144.4%)</td><td>3.6</td><td>8.8</td><td>(144.4%)</td></tr><tr><td>Hyperbaton</td><td>69.2</td><td>42.8</td><td>51.6</td><td>(20.6%)</td><td>44.0</td><td>56.8</td><td>(29.1%)</td></tr><tr><td>Logical Deduction (5 objects)</td><td>38.8</td><td>4.8</td><td>19.6</td><td>(308.3%)</td><td>4.8</td><td>20.8</td><td>(333.3%)</td></tr><tr><td>Logical Deduction (7 objects)</td><td>39.6</td><td>1.2</td><td>14.4</td><td>(1100.0%)</td><td>1.2</td><td>21.2</td><td>(1666.7%)</td></tr><tr><td>Logical Deduction (3 objects)</td><td>60.4</td><td>19.6</td><td>40.4</td><td>(106.1%)</td><td>16.8</td><td>38.0</td><td>(126.2%)</td></tr><tr><td>Movie Recommendation</td><td>55.4</td><td>24.4</td><td>26.8</td><td>(9.8%)</td><td>43.4</td><td>57.6</td><td>(32.7%)</td></tr><tr><td>Navigate</td><td>55.6</td><td>43.6</td><td>49.2</td><td>(12.8%)</td><td>46.4</td><td>45.2</td><td>(-2.6%)</td></tr><tr><td>Penguins in a Table</td><td>45.9</td><td>17.5</td><td>24.7</td><td>(41.1%)</td><td>15.1</td><td>26.7</td><td>(76.8%)</td></tr><tr><td>Reasoning about Colored Objects</td><td>47.6</td><td>14.0</td><td>15.2</td><td>(8.6%)</td><td>12.0</td><td>17.6</td><td>(46.7%)</td></tr><tr><td>Ruin Names</td><td>56.0</td><td>12.2</td><td>14.4</td><td>(18.0%)</td><td>15.7</td><td>29.2</td><td>(86.0%)</td></tr><tr><td>Salient Translation Error Detection</td><td>40.8</td><td>2.0</td><td>12.0</td><td>(500.0%)</td><td>2.0</td><td>12.4</td><td>(520.0%)</td></tr><tr><td>Snarks</td><td>59.0</td><td>28.0</td><td>56.2</td><td>(100.7%)</td><td>28.1</td><td>61.2</td><td>(117.8%)</td></tr><tr><td>Sports Understanding</td><td>79.6</td><td>40.4</td><td>48.4</td><td>(19.8%)</td><td>48.4</td><td>51.6</td><td>(6.6%)</td></tr><tr><td>Temporal Sequences</td><td>35.6</td><td>21.2</td><td>24.4</td><td>(15.1%)</td><td>16.0</td><td>10.4</td><td>(-35.0%)</td></tr><tr><td>Tracking Shuffled Objects (5 objects)</td><td>18.4</td><td>6.4</td><td>14.4</td><td>(125.0%)</td><td>9.2</td><td>24.8</td><td>(169.6%)</td></tr><tr><td>Tracking Shuffled Objects (7 objects)</td><td>15.2</td><td>4.0</td><td>13.6</td><td>(240.0%)</td><td>5.6</td><td>13.2</td><td>(135.7%)</td></tr><tr><td>Tracking Shuffled Objects (3 objects)</td><td>31.6</td><td>26.8</td><td>34.0</td><td>(26.9%)</td><td>23.2</td><td>34.4</td><td>(48.3%)</td></tr><tr><td>Web of Lies</td><td>56.0</td><td>49.4</td><td>47.2</td><td>(-4.5%)</td><td>41.2</td><td>54.8</td><td>(33.0%)</td></tr><tr><td>Average</td><td>48.9</td><td>21.9</td><td>32.0</td><td>(45.9%)</td><td>23.3</td><td>36.2</td><td>(55.4%)</td></tr></table>
+
+Table 3: Zero-shot performance comparison of ChatGPT, Vicuna, and Lion on BIGBench Hard (multiple-choice questions) without CoT. We report the performance of ChatGPT and Vicuna from (Mukherjee et al., 2023). Performance improvements obtained by Lion over Vicuna are shown in parenthesis.
+
+highest average scores.
+
+# 6.2 The Learning Dynamics of Lion
+
+In Figure 5, we delve into the learning dynamics of Lion by visualizing its performance on AGIEval and BBH throughout the training iterations. The results clearly demonstrate that our adversarial knowledge distillation framework consistently enhances the performance of the student model as the iterations progress. Notably, the most significant improvement in capability occurs in the first iteration, suggesting the usefulness of the identification of challenging example patterns (refer Figure 3b).
+
+# 6.3 Case Studies
+
+To clearly compare the generated response quality between our model and other baselines, we provide nine case studies sampled from Vicuna-instruction, AGIEval, and BBH in Appendix E. Table 14 showcases the responses of various models to a math instruction. It can be seen that only Lion and ChatGPT provide the correct answer and follow the correct problem-solving steps. A counterfactual case is shown in Table 15, where ChatGPT provides a relevant answer that considers the potential impacts of Newton focusing on biology instead of physics, but it lacked details and depth. Lion, on
+
+<table><tr><td>Threshold τ</td><td>Vicuna-Instructions (Avg.)</td><td>AGIEval (Avg.)</td><td>BBH (Avg.)</td></tr><tr><td>0.0</td><td>89.58</td><td>22.4</td><td>26.5</td></tr><tr><td>0.5</td><td>92.16</td><td>23.5</td><td>29.8</td></tr><tr><td>1.0</td><td>93.81</td><td>26.3</td><td>32.0</td></tr><tr><td>1.5</td><td>94.09</td><td>25.7</td><td>31.6</td></tr><tr><td>2.0</td><td>92.23</td><td>24.6</td><td>31.3</td></tr></table>
+
+Table 4: Ablation study of the threshold  $\tau$  for Lion-7B.  
+
+<table><tr><td>Ratio r</td><td>Vicuna-Instructions (Avg.)</td><td>AGIEval (Avg.)</td><td>BBH (Avg.)</td></tr><tr><td>1:0</td><td>89.60</td><td>24.3</td><td>30.8</td></tr><tr><td>2:1</td><td>92.95</td><td>25.7</td><td>33.1</td></tr><tr><td>1:1</td><td>93.81</td><td>26.3</td><td>32.0</td></tr><tr><td>1:2</td><td>91.77</td><td>23.9</td><td>29.6</td></tr><tr><td>0:1</td><td>90.02</td><td>22.1</td><td>24.3</td></tr></table>
+
+Table 5: Ablation study of the ratio  $r$  for Lion-7B.
+
+![](images/756e8a9d539b12c6c83f6750f6e00fece3739e74d6f4b65918654359ab2642e8.jpg)  
+Figure 5: Performance of Lion-7B and Lion-13B on AGIEval and BBH through the training iterations.
+
+the other hand, offered a more detailed and engaging response that explored different possibilities such as the development of biophysics or discovering new principles that could be applied to both fields. Lion's response also considered the potential implications of Newton's work on motion, force, gravity, and thermodynamics in biology, providing a more comprehensive answer.
+
+# 7 Conclusion
+
+This paper presents an innovative adversarial knowledge distillation framework for distilling a proprietary LLM into a compact, open-source student model. While previous methodologies have concentrated on unidirectional knowledge transfer, our approach seeks to integrate "feedback" into the learning process. Leveraging the versatile role adaptability of LLMs, we prompt the proprietary model to identify "hard" instructions and generate new "hard" instructions for the student model, creating a three-stage adversarial loop of imitation, discrimination, and generation. This approach al
+
+ows us to refine the student model's performance iteratively, efficiently bootstrapping its proficiency. We aspire that our model, named Lion, may serve as a baseline to reflect the performance of ChatGPT, especially the open-source instruction-following language model baseline for our community.
+
+# Limitations and Discussions
+
+The Model Capability We have identified that Lion is subject to certain constraints: 1) A recent study (Gudibandé et al., 2023) asserts that "model imitation is a false promise" since imitation models are adept at mimicking ChatGPT's style but fall short in improving LMs across more challenging tasks. While Lion still lags behind its teacher model ChatGPT in handling intricate reasoning tasks (as shown in our experiments), it demonstrates promising improvements compared to previous imitation models. Therefore, our adversarial knowledge distillation framework may provide a more effective way for knowledge transfer. 2) Since our training data doesn't encompass dialogues, Lion struggles to manage multi-turn conversations. 3) Due to computational resource constraints, Lion's maximum sequence length is limited to 1024. Consequently, it faces challenges when dealing with long documents. Despite these limitations, we envision Lion serving as an accessible springboard for future research endeavors aimed at addressing these limitations.
+
+The Training Process To train a single student model, we request the gpt-3.5-turbo API around 450k times, a number that is roughly  $70\%$  of the WizardLM's usage of 624k (Xu et al., 2023).
+
+Nonetheless, this utilization incurs a considerable expense, nearing $900. In contrast to methods like Alpaca (Taori et al., 2023) and WizardLM (Xu et al., 2023), which only fine-tune the student model once, our adversarial knowledge distillation method employs iterative parametric updates to the student model. While this iterative approach inevitably leads to slower iteration speed, it offers additional benefits. Finally, different from traditional adversarial knowledge distillation where the weights of the generator are iteratively updated, we use a black-box and parameter-frozen LLM (ChatGPT in our paper) to serve the role. Therefore, the quality of the LLM is quite essential in the generation of new instructions.
+
+The Evaluation Metrics Though automated evaluations leveraging GPT-4 have showcased promising prospects in appraising chatbot performance, the technique is yet to reach a level of maturity and accuracy, especially considering the propensity of large language models to generate non-existent or "hallucinated" information. Evaluating the efficacy of LLM across various tasks presents a considerable challenge since different tasks require quite different expertise (Wang et al., 2022). Therefore, the creation of a comprehensive, standardized evaluation system for chatbots is a prevailing research challenge that demands additional exploration and study.
+
+# Ethics Statement
+
+Inherited Biases It is important to consider that the behavior of our distilled student models may exhibit potential toxicity, biases, or privacy issues (Li et al., 2023a,b) inherited from the larger teacher LLM. We anticipate that the advancements made in reducing anti-social behaviors in LLMs can also be utilized to enhance student language models.
+
+License and Legality Based on Stanford Alpaca's guidelines (Taori et al., 2023), we have determined that the weights of Lion will be exclusively licensed for research purposes in the future. Utilizing Lion's weights alongside LLaMA's original weights must adhere to Meta's LLaMA License Agreement. Users are responsible for acquiring and utilizing LLaMA in accordance with the license agreement.
+
+Safety Unlike ChatGPT (OpenAI, 2022), Lion does not rely on human feedback to mitigate undesired behaviors. Instead, Lion learns to avoid such
+
+behaviors by imitating ChatGPT. However, it is important to acknowledge the potential risks associated with using Lion for malicious purposes, especially upon releasing its weights in the future. For future work, we aim to incorporate the technique of Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022) to enhance access control. Additionally, Meta has implemented an access application process that can help regulate the distribution of LLaMA models and minimize the potential risks associated with their usage, providing an alternative option.
+
+# Acknowledgements
+
+W. Wang was also affiliated with Guangzhou Municipal Key Laboratory of Materials Informatics, The Hong Kong University of Science and Technology (Guangzhou), China. He was supported by HKUST(GZ) Grant G0101000028, GZU-HKUST Joint Research Collaboration Grant GZU22EG04, CCF-HuaweiDBC202302, and Guangzhou Municipal Science and Technology Project (No. 2023A03J0003).
+
+# References
+
+Sravanti Addepalli, Gaurav Kumar Nayak, Anirban Chakraborty, and Venkatesh Babu Radhakrishnan. 2020. Degan: Data-enriching gan for retrieving representative samples from a trained classifier. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 3130-3137.  
+Vamsi Aribandi, Yi Tay, Tal Schuster, Jinfeng Rao, Huaxiu Steven Zheng, Sanket Vaibhav Mehta, Honglei Zhuang, Vinh Q. Tran, Dara Bahri, Jianmo Ni, Jai Prakash Gupta, Kai Hui, Sebastian Ruder, and Donald Metzler. 2022. Ext5: Towards extreme multitask scaling for transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. Open-Review.net.  
+Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. CoRR, abs/2302.04023.  
+Sebastien Buebeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott M. Lundberg, Harsha Nori, Hamid Palangi, Marco Túlio Ribeiro, and Yi Zhang. 2023. Sparks of artificial general intelligence: Early experiments with GPT-4. CoRR, abs/2303.12712.  
+Chunkit Chan and Tsz Ho Chan. 2023. Discourse-aware prompt for argument impact classification. In Proceedings of the 15th International Conference on Machine Learning and Computing, ICMLC 2023, Zhuhai, China, February 17-20, 2023, pages 165-171. ACM.  
+Chunkit Chan, Jiayang Cheng, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2023a. Chatgpt evaluation on sentence level relations: A focus on temporal, causal, and discourse relations. CoRR, abs/2304.14827.  
+Chunkit Chan, Xin Liu, Tsz Ho Chan, Jiayang Cheng, Yangqiu Song, Ginny Y. Wong, and Simon See. 2023b. Self-consistent narrative prompts on abductive natural language inference. CoRR, abs/2309.08303.  
+Chunkit Chan, Xin Liu, Jiayang Cheng, Zihan Li, Yangqiu Song, Ginny Y. Wong, and Simon See. 2023c. Discoprompt: Path prediction prompt tuning for implicit discourse relation recognition. In Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023, pages 35-57. Association for Computational Linguistics.  
+Akshay Chawla, Hongxu Yin, Pavlo Molchanov, and Jose Alvarez. 2021. Data-free knowledge distillation for object detection. In Proceedings of the IEEE/CVF
+
+Winter Conference on Applications of Computer Vision, pages 3289-3298.  
+Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2019. Distilling knowledge learned in bert for text generation. arXiv preprint arXiv:1911.03829.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An opensource chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Gongfan Fang, Kanya Mo, Xinchao Wang, Jie Song, Shitao Bei, Haofei Zhang, and Mingli Song. 2022. Up to 100x faster data-free knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 6597-6604.  
+Gongfan Fang, Jie Song, Chengchao Shen, Xinchao Wang, Da Chen, and Mingli Song. 2019. Data-free adversarial distillation. CoRR, abs/1912.11006.  
+Google. 2023. Bard.  
+Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. 2023. The false promise of imitating proprietary llms. CoRR, abs/2305.15717.  
+Byeongho Heo, Minsik Lee, Sangdoo Yun, and Jin Young Choi. 2019. Knowledge distillation with adversarial samples supporting decision boundary. 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 3771-3778. AAAI Press.  
+Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531.  
+Yuxin Jiang, Linhan Zhang, and Wei Wang. 2022. Improved universal sentence embeddings with prompt-based contrastive learning and energy-based learning. In *Findings of the Association for Computational Linguistics: EMNLP* 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 3021-3035. Association for Computational Linguistics.  
+Cheng Jiayang, Lin Qiu, Tsz Ho Chan, Tianqing Fang, Weiqi Wang, Chunkit Chan, Dongyu Ru, Qipeng Guo, Hongming Zhang, Yangqiu Song, Yue Zhang, and Zheng Zhang. 2023. Storyanology: Deriving story-level analogies from large language models to unlock analogical understanding. CoRR, abs/2310.12874.
+
+Sanjay Kariyappa, Atul Prakash, and Moinuddin K Qureshi. 2021. Maze: Data-free model stealing attack using zeroth-order gradient estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13814-13823.  
+Haoran Li, Yulin Chen, Jinglong Luo, Yan Kang, Xiaojin Zhang, Qi Hu, Chunkit Chan, and Yangqiu Song. 2023a. Privacy in large language models: Attacks, defenses and future directions.  
+Haoran Li, Dadi Guo, Wei Fan, Mingshi Xu, and Yangqiu Song. 2023b. Multi-step jailbreaking privacy attacks on chatgpt. CoRR, abs/2304.05197.  
+Paul Micaelli and Amos J. Storkey. 2019a. Zero-shot knowledge transfer via adversarial belief matching. In Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 9547-9557.  
+Paul Micaelli and Amos J Storkey. 2019b. Zero-shot knowledge transfer via adversarial belief matching. Advances in Neural Information Processing Systems, 32.  
+Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. 2022. Cross-task generalization via natural language crowdsourcing instructions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3470-3487. Association for Computational Linguistics.  
+Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Hassan Awadallah. 2023. Orca: Progressive learning from complex explanation traces of GPT-4. CoRR, abs/2306.02707.  
+OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.  
+TB OpenAI. 2022. Chatgpt: Optimizing language models for dialogue. OpenAI.  
+Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. 2019. Knockoff nets: Stealing functionality of black-box models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4954-4963.  
+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 Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. CoRR, abs/2203.02155.
+
+Nicolas Papernot, Patrick McDaniel, Ian Goodfellow, Somesh Jha, Z Berkay Celik, and Ananthram Swami. 2017. Practical black-box attacks against machine learning. In Proceedings of the 2017 ACM on Asia conference on computer and communications security, pages 506-519.  
+Ilija Radosavovic, Piotr Dólar, Ross Girshick, Georgia Gkioxari, and Kaiming He. 2018. Data distillation: Towards omni-supervised learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4119-4128.  
+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.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H. Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal V. Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng Shen, Zheng Xin Yong, Harshit Pandey, Rachel Bawden, Thomas Wang, Trishala Neeraj, Jos Rozen, Abheesht Sharma, Andrea Santilli, Thibault Fevry, Jason Alan Fries, Ryan Teehan, Teven Le Scao, Stella Biderman, Leo Gao, Thomas Wolf, and Alexander M. Rush. 2022. Multitask prompted training enables zero-shot task generalization. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.  
+Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. 2022. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615.  
+Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V. Le, Ed H. Chi, Denny Zhou, and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. CoRR, abs/2210.09261.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
+
+Jean-Baptiste Truong, Pratyush Maini, Robert J Walls, and Nicolas Papernot. 2021. Data-free model extraction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4771-4780.
+
+Peiyi Wang, Lei Li, Liang Chen, Dawei Zhu, Binghuai Lin, Yunbo Cao, Qi Liu, Tianyu Liu, and Zhifang Sui. 2023. Large language models are not fair evaluators. CoRR, abs/2305.17926.
+
+Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-instruct: Aligning language model with self generated instructions. CoRR, abs/2212.10560.
+
+Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022a. Finetuned language models are zero-shot learners. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.
+
+Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. 2022b. Finetuned language models are zero-shot learners. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net.
+
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022c. Emergent abilities of large language models. CoRR, abs/2206.07682.
+
+Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022d. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS.
+
+Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large language models to follow complex instructions.
+
+Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. 2020. Dreaming to distill: Data-free knowledge transfer via deepinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8715-8724.
+
+Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan. 2023. Agieval: A human-centric benchmark for evaluating foundation models. CoRR, abs/2304.06364.
+
+# A Data Statistics
+
+Table 6 and Table 7 show the data statistics of AGIEval and BIG-Bench Hard, respectively.
+
+<table><tr><td>Task</td><td># Examples</td><td># Choices</td></tr><tr><td>AQuA-RAT</td><td>254</td><td>5</td></tr><tr><td>LogiQA</td><td>651</td><td>4</td></tr><tr><td>LSAT-AR</td><td>230</td><td>5</td></tr><tr><td>LSAT-LR</td><td>510</td><td>5</td></tr><tr><td>LSAT-RC</td><td>269</td><td>5</td></tr><tr><td>SAT-Math</td><td>220</td><td>4</td></tr><tr><td>SAT-English</td><td>206</td><td>4</td></tr><tr><td>SAT-English (w/o Psg.)</td><td>206</td><td>4</td></tr></table>
+
+Table 6: Statistics of AGIEval dataset.  
+
+<table><tr><td>Task</td><td># Examples</td><td># Choices</td></tr><tr><td>Boolean Expressions</td><td>250</td><td>2</td></tr><tr><td>Causal Judgement</td><td>187</td><td>2</td></tr><tr><td>Date Understanding</td><td>250</td><td>6</td></tr><tr><td>Disambiguation QA</td><td>250</td><td>4</td></tr><tr><td>Formal Fallacies</td><td>250</td><td>2</td></tr><tr><td>Geometric Shapes</td><td>250</td><td>11</td></tr><tr><td>Hyperbaton</td><td>250</td><td>2</td></tr><tr><td>Logical Deduction (5 objects)</td><td>250</td><td>5</td></tr><tr><td>Logical Deduction (7 objects)</td><td>250</td><td>7</td></tr><tr><td>Logical Deduction (3 objects)</td><td>250</td><td>3</td></tr><tr><td>Movie Recommendation</td><td>250</td><td>5</td></tr><tr><td>Navigate</td><td>250</td><td>2</td></tr><tr><td>Penguins in a Table</td><td>146</td><td>5</td></tr><tr><td>Reasoning about Colored Objects</td><td>250</td><td>18</td></tr><tr><td>Ruin Names</td><td>250</td><td>11</td></tr><tr><td>Salient Translation Error Detection</td><td>250</td><td>6</td></tr><tr><td>Snarks</td><td>178</td><td>2</td></tr><tr><td>Sports Understanding</td><td>250</td><td>2</td></tr><tr><td>Temporal Sequences</td><td>250</td><td>4</td></tr><tr><td>Tracking Shuffled Objects (5 objects)</td><td>250</td><td>5</td></tr><tr><td>Tracking Shuffled Objects (7 objects)</td><td>250</td><td>7</td></tr><tr><td>Tracking Shuffled Objects (3 objects)</td><td>250</td><td>3</td></tr><tr><td>Web of Lies</td><td>250</td><td>2</td></tr></table>
+
+Table 7: Statistics of BIG-Bench Hard dataset.
+
+# B Baselines
+
+- LLaMA (Touvron et al., 2023) is a collection of foundation language models ranging from 7B to 65B parameters. It is trained on trillions of tokens from publicly available datasets and is demonstrated to outperform larger-size LLMs such as GPT-3 (175B) across a multitude of benchmarks. We use the official code from LLaMA  $^{4}$ .  
+- Alpaca (Taori et al., 2023) is a project initiated by Stanford University with the objective of developing and disseminating an open-source model that adeptly follows instructions. It is based on LLaMA and fine-tuned on 52K instruction-following examples generated by
+
+querying OpenAI's text-davinci-003 model. On the self-instruct evaluation set, Alpaca mirrors text-davinci-003, but is notably more compact and cost-effective to reproduce. We use the official code from Alpaca  $^{5}$ .
+
+- WizardLM (Xu et al., 2023) employs LLMs instead of humans to automatically mass-produce open-domain instructions of various difficulty levels, to improve the performance of LLMs. It uses an Evol-Instruct method to bootstrap the 52k instruction-following examples of Alapca into a larger set of 250k more intricate instructions. Out of this larger set, 70k examples were selected to fine-tune LLaMA. We use WizardLM-7B-V1.0 from the official code  $^{6}$ .  
+- Vicuna (Chiang et al., 2023), a superior open-source chatbot, excels in generating fluid and captivating responses to user queries. It is based on LLaMA and fine-tuned on 70K user-shared conversations collected from ShareGPT, a platform designed for sharing interactions with ChatGPT. Its impressive capabilities make it one of the leading open instruction-following models today. Vicuna achieves competitive performance against proprietary models such as ChatGPT and Bard (Google, 2023). We use Vicuna-7B-V1.1 and Vicuna-13B-V1.1 from FastChat<sup>7</sup>.  
+- ChatGPT (OpenAI, 2022), a product of OpenAI, is an advanced AI chatbot renowned for its ability to interact with users in an authentically human and engaging manner. The chatbot is built on powerful LLMs such as GPT-3.5 and GPT-4, which are trained on a vast corpus of internet text data. ChatGPT undergoes fine-tuning via both supervised and reinforcement learning techniques, with the human trainers providing necessary feedback and direction.
+
+# C Implementation Details
+
+Training Hyperparameters The training process is conducted on 8 A100 GPUs. During each iteration of adversarial knowledge distillation, the hyperparameters for training are shown in Table 8.
+
+<table><tr><td>Hyperparameter</td><td>Lion-7B</td><td>Lion-13B</td></tr><tr><td>Batch size</td><td>128</td><td>128</td></tr><tr><td>Learning rate</td><td>2e-5</td><td>2e-5</td></tr><tr><td>Epoches</td><td>3</td><td>3</td></tr><tr><td>Max length</td><td>1024</td><td>1024</td></tr><tr><td>Optimizer</td><td>AdamW</td><td>AdamW</td></tr><tr><td>Scheduler</td><td>cosine</td><td>cosine</td></tr><tr><td>Weight decay</td><td>0</td><td>0</td></tr><tr><td>Warmup ratio</td><td>0.03</td><td>0.03</td></tr></table>
+
+Querying the gpt-3.5-turbo API We use different sets of hyperparameters when querying the gpt-3.5-turbo API for different roles (Teacher, Referee, Generator). These hyperparameters are found to work well and we listed them in Table 9.
+
+Table 8: Training hyperparameters.  
+
+<table><tr><td>Role</td><td>temperature</td><td>top_p</td><td>beam_size (n)</td><td>max_tokens</td></tr><tr><td>Teacher</td><td>0.7</td><td>1.0</td><td>1</td><td>1024</td></tr><tr><td>Referee</td><td>0.2</td><td>1.0</td><td>1</td><td>512</td></tr><tr><td>Generator</td><td>1.0</td><td>1.0</td><td>1</td><td>512</td></tr></table>
+
+Table 9: Hyperparameters for querying OpenAI gpt-3.5-turbo API under different roles.
+
+# D Prompt Templates for Our Adversarial Distillation Framework
+
+Fine-tuning an LLM (i.e. ChatGPT) is costly and intricate, human-tailored prompt templates are utilized to solve various tasks (Wei et al., 2022d; Chan et al., 2023b,c; Jiang et al., 2022; Jiayang et al., 2023; Chan and Chan, 2023). The prompt template of the Teacher for generating responses is shown in Table 10. The prompt template of the Referee for comparing the quality of two responses generated by two AI assistants is shown in Table 11. The prompt templates of the Generator for generating new hard instructions and new easy instructions are shown in Table 12 and Table 13, respectively.
+
+# E Case Studies
+
+Here we show 3 cases in Table 14, 15, and 16 to clearly compare the open-ended generation performance among various models including our Lion-13B, LLaMA-13B, Alpaca-13B, Vicuna-13B, and ChatGPT.
+
+Besides, we show 6 cases in Table 17, 18, 19, 20, 21, and 22 to clearly compare the reasoning capability among various models including our Lion-13B, Vicuna-13B, and ChatGPT. We utilize  $\checkmark$  and  $\times$  to denote whether the response is correct or incorrect, respectively.
+
+<table><tr><td>system content</td><td>You are a helpful assistant that generates a response to a given task instruction.</td></tr><tr><td rowspan="2">user content</td><td>Instruction: {instruction}</td></tr><tr><td>Response:</td></tr></table>
+
+Table 10: Prompt template of gpt-3.5-turbo for generating responses. Note that the original instruction in Alpaca is composed of an instruction prompt and an instance input. For example, the instruction prompt is "write an abstract about the following method", and the instance input is "knowledge distillation". For a better adaption to real-world scenarios, we concatenate the instruction prompt and the instruction prompt into one instruction using a line break.  
+
+<table><tr><td>system content</td><td>You are a helpful and precise assistant for checking the quality of the answer.</td></tr><tr><td>user content</td><td>[Instruction]
+{instruction}
+[The Start of Assistant 1&#x27;s Answer]
+{answer_1}
+[The End of Assistant 1&#x27;s Answer]
+[The Start of Assistant 2&#x27;s Answer]
+{answer_2}
+[The End of Assistant 2&#x27;s Answer]
+[System]
+We would like to request your feedback on the performance of two AI assistants in response to the user instruction and input displayed above.
+Please rate the helpfulness, relevance, accuracy, and level of detail of their responses. Each assistant receives an overall score on a scale of 1 to 10, where a higher score indicates better overall performance.
+Please first provide a comprehensive explanation of your evaluation, avoiding any potential bias and ensuring that the order in which the responses were presented does not affect your judgment. Then, output two lines indicating the scores for Assistant 1 and 2, respectively.
+Output with the following format:
+Evaluation evidence: &lt;your evaluation explanation here&gt;
+Score of the Assistant 1: &lt;score&gt;
+Score of the Assistant 2: &lt;score&gt;</td></tr></table>
+
+Table 11: Prompt template of gpt-3.5-turbo for comparing the quality of two responses generated by two AI assistants.  
+
+<table><tr><td>system content</td><td>You are a helpful assistant.</td></tr><tr><td>user content</td><td>I want you to act as an Instruction Creator.Your goal is to draw inspiration from the #Given Instruction# to create a brand new instruction.This new instruction should belong to the same domain and the same task type as the #Given Instruction#.The LENGTH and difficulty level of the #Created Instruction# should be similar to that of the #Given Instruction#.The #Created Instruction# must be reasonable and must be understood and responded to by humans.&#x27;#Given Instruction#, &#x27;#Created Instruction#, &#x27;given instruction&#x27; and &#x27;created instruction&#x27; are not allowed to appear in #Created Instruction#.#Given Instruction#: {instruction}#Created Instruction#:</td></tr></table>
+
+Table 12: Prompt template of gpt-3.5-turbo for generating new hard instructions.
+
+<table><tr><td>system content</td><td>You are a helpful assistant.</td></tr><tr><td>user content</td><td>I want you to act as an Instruction Creator.
+Your goal is to draw inspiration from the #Given Instruction# to create a brand new instruction.
+This new instruction should belong to the same domain as the #Given Instruction# but be even more rare.
+The LENGTH and difficulty level of the #Created Instruction# should be similar to that of the #Given Instruction#.
+The #Created Instruction# must be reasonable and must be understood and responded to by humans.
+&#x27;#Given Instruction#&#x27;, &#x27;#Created Instruction#&#x27;, &#x27;given instruction&#x27; and &#x27;created instruction&#x27; are not allowed to appear in #Created Instruction#.
+#Given Instruction#: {instruction}
+#Created Instruction#:</td></tr></table>
+
+Table 13: Prompt template of gpt-3.5-turbo for generating new easy instructions.
+
+<table><tr><td colspan="2">Instruction: If the endpoints of a line segment are (2, -2) and (10, 4), what is the length of the segment?</td></tr><tr><td>Lion</td><td>To find the length of the line segment with endpoints (2, -2) and (10, 4), we can use the distance formula. The distance formula is derived from the Pythagorean theorem and is given as: d = √((x2 - x1)2 + (y2 - y1)2) where (x1, y1) and (x2, y2) are the coordinates of the endpoints of the line segment. Substituting the given values, we get: d = √((10 - 2)2 + (4 - (-2))2) Simplifying this expression, we get: d = √((8)2 + 62) d = √(64 + 36) d = √(100) d = 10 Therefore, the length of the line segment with endpoints (2, -2) and (10, 4) is 10 units. ✓</td></tr><tr><td>LLaMA</td><td>The length of the line segment is 10 units. ✓</td></tr><tr><td>Alpaca</td><td>The length of the line segment is 8 units. X</td></tr><tr><td>Vicuna</td><td>The length of the line segment is the distance between the two points. To find the length, you can use the Pythagorean theorem, which states that the sum of the squares of the distances from a point to the two endpoints is equal to the square of the distance between the two endpoints. In this case, the distance from the point (2, -2) to the point (10, 4) is given by the formula: √((x2 - x1)2 + (y2 - y1)2) = √((10 - 2)2 + (4 - (-2))2) = √(182 + 42) = √(324 + 16) = √(340) = 17.973 So the length of the line segment is 17.973. X</td></tr><tr><td>ChatGPT</td><td>To find the length of a line segment with two endpoints (x1, y1) and (x2, y2), we use the distance formula: distance = √((x2 - x1)2 + (y2 - y1)2) Plugging in the coordinates of the given endpoints, we get: distance = √((10 - 2)2 + (4 - (-2))2) distance = √(82 + 62) distance = √(64 + 36) distance = √(100) distance = 10 Therefore, the length of the line segment is 10 units. ✓</td></tr></table>
+
+Table 14: A math case from Vicuna-Instructions.
+
+<table><tr><td colspan="2">Instruction: What if Isaac Newton had focused on biology instead of physics?</td></tr><tr><td>Lion</td><td>Isaac Newton is renowned for his contributions to physics, particularly his laws of motion and universal gravitation. However, what if Newton had instead chosen to focus on biology? 
+Newton&#x27;s work in physics was heavily influenced by his fascination with the natural world and his desire to understand how it worked. Similarly, had he chosen to study biology, he may have made significant contributions to the field. Newton&#x27;s mathematical abilities and his interest in the natural world would have made him a formidable biologist. 
+One area where Newton may have made a significant impact in biology is in the field of tax-onomy. Newton was a keen observer of nature and had a particular interest in plants. He spent much of his time studying and cataloguing the various species of plants around him. He even developed a system for classifying plants based on their physical characteristics. 
+In addition to his work in taxonomy, Newton may have also made important contributions to the field of physiology. He was interested in the workings of the human body and may have studied the mechanics of muscles and the circulatory system. 
+Overall, while Isaac Newton is most well-known for his work in physics, his interests and abilities would have made him a formidable biologist as well. Who knows what discoveries he might have made had he chosen to focus on the natural world in a different way.</td></tr><tr><td>LLaMA</td><td>I&#x27;m not sure. If he did so, he wouldn&#x27;t have discovered the law of gravity.</td></tr><tr><td>Alpaca</td><td>If Isaac Newton had focused on biology instead of physics, he might have made important contributions to the field of biology, such as developing a scientific theory of evolution or explaining the processes of cell growth and development. He might also have studied and discovered new principles of genetics and developed methods to study and treat diseases.</td></tr><tr><td>Vicuna</td><td>Isaac Newton is widely regarded as one of the most influential scientists in history for his work in physics, particularly for his development of the laws of motion and universal gravitation. However, it is worth noting that Newton also made significant contributions to other fields, including mathematics, optics, and natural philosophy. 
+If Newton had decided to focus on biology instead of physics, it is possible that he would have made significant contributions to the field. However, it is difficult to say exactly what those contributions would have been without knowing more about his specific interests and areas of focus. 
+Newton was a prolific scientist and wrote extensively on a wide range of topics, including natural philosophy, mathematics, and theology. He also conducted extensive research in optics, including the study of light and color, and made important contributions to the field of mathematics, including the development of calculus. 
+It is possible that if Newton had focused on biology, he would have made significant contributions to the understanding of living organisms and their behavior. However, it is also possible that his work in physics and mathematics would have had a greater impact on the world.</td></tr><tr><td>ChatGPT</td><td>If Isaac Newton had focused on biology instead of physics, the scientific world may be vastly different today. His advancements in mathematics, such as calculus, would have certainly been useful in the field of biology. Newton&#x27;s curiosity, analytical mind, and dedication to scientific inquiry would have been equally effective in biology as they were in physics. He may have made significant contributions to the field of genetics or evolutionary theory. 
+On the other hand, his contributions to physics were so groundbreaking that the subject may have been hindered without his contributions. It&#x27;s difficult to say exactly what would have happened, but it&#x27;s likely that the course of scientific history would have been different.</td></tr></table>
+
+Table 15: A counterfactual case from Vicuna-Instructions.
+
+<table><tr><td colspan="2">Instruction: How can I develop my critical thinking skills?</td></tr><tr><td>Lion</td><td>Developing critical thinking skills is an important aspect of personal growth and professional development. Critical thinking enables us to analyze and evaluate information, ideas, and arguments objectively, make informed decisions, and solve complex problems. Here are some strategies that can help you develop your critical thinking skills:
+1. Read widely and deeply: Reading widely exposes you to different perspectives and ideas, while reading deeply helps you to understand and analyze complex concepts. Choose books that are challenging and outside your comfort zone, and take notes as you read.
+2. Ask questions: Asking questions is an essential part of critical thinking. Start by asking why and how questions to gain a deeper understanding of a topic. Then, ask what if questions to explore possible outcomes and consequences.
+3. Practice analysis and evaluation: Critical thinking involves analyzing and evaluating information and arguments. Start by breaking down complex ideas into smaller parts, and then evaluate each part based on evidence, reasoning, and credibility.
+4. Seek diverse perspectives: Surround yourself with people who have different backgrounds, experiences, and perspectives. Listen to their ideas and opinions, and challenge your own assumptions and biases.
+5. Practice decision-making: Critical thinking is essential in decision-making. Practice making decisions by weighing the pros and cons, considering potential outcomes, and seeking advice from others.
+6. Take courses and workshops: Enrolling in courses and workshops that focus on critical thinking, problem-solving, and decision-making can provide you with the tools and techniques to develop your critical thinking skills.
+By implementing these strategies, you can develop your critical thinking skills and become better equipped to analyze, evaluate, and make informed decisions in your personal and professional life.</td></tr><tr><td>LLaMA</td><td>Critical thinking is a vital skill to have, but it can be difficult to develop. Here are some ways you can start developing your critical thinking skills:
+1. Analyze and interpret: When you&#x27;re presented with information, try to analyze and interpret it on your own terms. When you start to analyze and interpret information on your own, it will become easier to do so over time.
+2. Assess your information: When you receive new information, try to assess whether it&#x27;s valid or reliable. This is important for evaluating and forming your own opinions.
+3. Use your intuition: Use your intuition to complement your critical thinking skills. When you pay attention to your gut feelings, they&#x27;ll usually tell you when you need to double-check your information.
+4. Learn from your mistakes: When you make a mistake, try not to beat yourself up about it. Instead, learn from your mistakes and use them to your advantage.</td></tr><tr><td>Alpaca</td><td>Developing critical thinking skills involves actively questioning assumptions, weighing evidence, and thinking analytically. It also involves being open-minded and considering various perspectives, as well as thinking logically and systematically. Practicing problem-solving and creative thinking, working through scenarios and dilemmas, and engaging in thoughtful conversations are all ways to improve critical thinking. Additionally, reading, reflecting, and writing can help increase your capacity for critical thinking.</td></tr><tr><td>Vicuna</td><td>Developing your critical thinking skills can help you to better analyze information, make more informed decisions, and solve complex problems. Here are some tips for improving your critical thinking skills:
+1. Practice active listening: Pay attention to what others are saying and try to understand their perspective.
+2. Ask questions: Be curious and ask questions to clarify information and better understand a situation.
+3. Analyze information: Take time to carefully analyze information and consider different viewpoints.
+4. Seek out diverse viewpoints: Expose yourself to a variety of viewpoints and ideas to challenge your own assumptions and biases.
+5. Practice logic and reasoning: Use logic and reasoning to evaluate arguments and make decisions.
+6. Learn from your mistakes: Be open to learning from your mistakes and using them as opportunities for growth.
+7. Read and watch a variety of sources: Read and watch a variety of sources to expand your knowledge and expose yourself to different perspectives.
+8. Engage in debates and discussions: Engage in debates and discussions with others to practice critical thinking and develop your skills.
+9. Take breaks and reflect: Take breaks and reflect on your thinking process to improve your critical thinking skills.
+10. Practice mindfulness: Mindfulness can help you to be more aware of your thoughts and biases, which can help you to think more critically.</td></tr><tr><td>ChatGPT</td><td>Sure, here are a few practical tips to develop critical thinking skills:
+1. Ask questions: Don&#x27;t simply accept information at face value. Ask questions that allow you to evaluate the reason, perspective and logic behind the information.
+2. Analyze: Take the time to analyze and evaluate the information that you receive. This will help you to identify assumptions, biases, strengths and weaknesses of an argument.
+3. Consider different perspectives: When analyzing information, try to consider different perspectives to gain a wider understanding of any topic.
+4. Research and read extensively: Read a wide range of sources to get a better understanding of different viewpoints and opinions.
+5. Practice logic and reasoning: Practice thinking through problems logically by breaking them down into smaller parts.
+6. Establish a point of view: Consider developing an opinion or thesis statement, and then collect evidence to support that position.
+7. Reflect on your own thinking: Consider how you arrived at a conclusion, and whether you have any biases or assumptions that may have influenced your thinking.
+With time, and practicing these skills consistently, you can become a better critical thinker.</td></tr></table>
+
+Table 16: A generic case from Vicuna-Instructions.
+
+<table><tr><td colspan="2">Instruction: Q: If 120 is reduced to 96, what is the reduction percent? Answer Choices: (A)30% (B)40% (C)20% (D)10% (E)5% A: Among A through E, the answer is</td></tr><tr><td>Lion</td><td>To calculate the reduction percent, we need to divide the difference between the original number and the reduced number by the original number and then multiply by 100. 
+Here, the original number is 120 and the reduced number is 96. So, the difference between the two numbers is 120 - 96 = 24. 
+To calculate the percentage reduction, we divide the difference by the original number and then multiply by 100: 
+(24 / 120) x 100 = 20% 
+Therefore, the correct answer is option (C) 20%. 
+Option (A) 30%, option (B) 40%, option (D) 10%, and option (E) 5% are all incorrect. ✓</td></tr><tr><td>Vicuna</td><td>A. 30% X</td></tr><tr><td>ChatGPT</td><td>To calculate the reduction percentage, we need to find the difference between the original value and the reduced value, and then divide it by the original value. 
+The difference between 120 and 96 is 120 - 96 = 24. 
+Now, divide the difference by the original value and multiply by 100 to get the percentage reduction: 
+(24 / 120) * 100 = 20% 
+Therefore, the correct answer is (C) 20%. ✓</td></tr></table>
+
+Table 17: An AQuA-RAT case from AGIEval.  
+
+<table><tr><td colspan="2">Instruction: Q: not ( True ) and ( True ) is? Answer Choices: ( A) True ( B) False A: Among A through B, the answer is</td></tr><tr><td>Lion</td><td>The expression &quot;not (True) and (True)&quot; can be simplified as &quot;(not True) and (True)&quot; or &quot;False and True&quot;. The &quot;and&quot; operator has a higher precedence than the &quot;not&quot; operator, so the first expression is evaluated as &quot;False and True&quot;, which is false. Therefore, the correct answer is option (B) False. ✓</td></tr><tr><td>Vicuna</td><td>(B) False
+&gt;&gt;&gt; Explanation:
+This statement is false.
+If the given statement is &quot;not (True)&quot; and &quot;True&quot;, it would be equivalent to &quot;not (True) and True&quot;, which would be true.
+(This is because negation of &quot;True&quot; is &quot;False&quot;, and &quot;True&quot; and &quot;False&quot; evaluate to &quot;True&quot;). 
+Therefore, the correct answer is (A) True.✗</td></tr><tr><td>ChatGPT</td><td>(A) True.✗</td></tr></table>
+
+Table 18: A boolean expressions case from BIGBench Hard.  
+
+<table><tr><td colspan="2">Instruction: Q: If you follow these instructions, do you return to the starting point? Take 8 steps. Turn around.
+Take 8 steps. Answer Choices: (A) Yes (B) No A: Among A through B, the answer is</td></tr><tr><td>Lion</td><td>According to the instructions given, if you follow them, you will return to the starting point. Therefore, the correct answer is Option A, &quot;Yes&quot;. By taking 8 steps and turning around, you will end up back where you started. It is important to follow instructions carefully to achieve the desired outcome. In this case, following these instructions will ensure that you return to the starting point. ✓</td></tr><tr><td>Vicuna</td><td>(A) Yes ✓</td></tr><tr><td>ChatGPT</td><td>(B) No X</td></tr></table>
+
+Table 19: A navigate case from BIGBench Hard.
+
+<table><tr><td colspan="2">Instruction: Q: Today is Christmas Eve of 1937. What is the date tomorrow in MM/DD/YYYY? Answer Choices: (A) 12/11/1937 (B) 12/25/1937 (C) 01/04/1938 (D) 12/04/1937 (E) 12/25/2006 (F) 07/25/1937 A: Among A through F, the answer is</td></tr><tr><td rowspan="7">Lion</td><td>The correct answer is (B) 12/25/1937.</td></tr><tr><td>To arrive at this answer, we need to consider the fact that Christmas Eve is the evening before Christmas Day. Therefore, the date of Christmas Day is one day after Christmas Eve. In this case, the date of Christmas Eve is 12/24/1937, and the date of Christmas Day is 12/25/1937. Therefore, the date tomorrow in MM/DD/YYYY format is 12/25/1937.</td></tr><tr><td>Option (A) is incorrect because it gives the date as 12/11/1937, which is incorrect.</td></tr><tr><td>Option (C) is incorrect because it gives the date as 01/04/1938, which is incorrect.</td></tr><tr><td>Option (D) is incorrect because it gives the date as 12/04/1937, which is incorrect.</td></tr><tr><td>Option (E) is incorrect because it gives the date as 12/25/2006, which is incorrect.</td></tr><tr><td>Option (F) is incorrect because it gives the date as 07/25/1937, which is incorrect. ✓</td></tr><tr><td rowspan="3">Vicuna</td><td>(A) 12/11/1937</td></tr><tr><td>### Explanation:</td></tr><tr><td>The date tomorrow, on Christmas Eve of 1937, would be 12/11/1937. X</td></tr><tr><td>ChatGPT</td><td>(C) 01/04/1938. X</td></tr></table>
+
+Table 20: A data understanding case from BIGBench Hard.  
+
+<table><tr><td colspan="2">Instruction: Q: Which statement is sarcastic? Answer Choices: (A) Hey just be happy then you won’t be depressed anymore (B) Hey just be happy that you won’t be depressed anymore A: Among A through B, the answer is</td></tr><tr><td>Lion</td><td>The correct answer is (B). The statement &quot;Hey just be happy that you won’t be depressed anymore&quot; is sarcastic because it implies that being happy is an easy and straightforward solution to depression, which is a complex mental health issue that requires professional help and support. The use of the word &quot;just&quot; also adds a dismissive tone to the statement, suggesting that the speaker doesn’t take depression seriously. It’s important to avoid using sarcasm when talking about mental health issues, as it can be hurtful and invalidating to those who are struggling. Instead, it’s important to offer compassion and support and to encourage individuals to seek professional help if needed. X</td></tr><tr><td>Vicuna</td><td>(B) Hey just be happy that you won’t be depressed anymore X</td></tr><tr><td>ChatGPT</td><td>The sarcastic statement is (A) &quot;Hey just be happy then you won’t be depressed anymore.&quot; ✓</td></tr></table>
+
+Table 21: A snarks case from BIGBench Hard.  
+
+<table><tr><td colspan="2">Instruction: Q: In the following sentences, explain the antecedent of the pronoun (which thing the pronoun refers to), or state that it is ambiguous. Sentence: The scientist collaborated with the artist, and he shared a story. Answer Choices: (A) The scientist shared a story (B) The artist shared a story (C) Ambiguous A: Among A through C, the answer is</td></tr><tr><td>Lion</td><td>In the given sentence, the antecedent of the pronoun &quot;he&quot; is ambiguous. &quot;He&quot; could refer to either the scientist or the artist, as both of them shared a story. Therefore, option (C) is the correct answer. ✓</td></tr><tr><td>Vicuna</td><td>(C) Ambiguous
+&gt;&gt;&gt; Explanation:
+The pronoun &quot;he&quot; is ambiguous, as it could refer to either the scientist or the artist. Without more context, it is impossible to determine which person is being referred to. Therefore, the answer is &quot;C - Ambiguous.&quot; ✓</td></tr><tr><td>ChatGPT</td><td>(C) Ambiguous.
+In the given sentence, the antecedent of the pronoun &quot;he&quot; is unclear. It could refer to either the scientist or the artist because there is no clear indication of which person shared the story. Therefore, the sentence is ambiguous. ✓</td></tr></table>
+
+Table 22: A disambiguation QA case from BIGBench Hard.

2023/Lion_ Adversarial Distillation of Proprietary Large Language Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85c895804f83b08458eda999d9f24657aea540bd13ee99c3e17ba879d3a81dba
+size 2481673

2023/Localizing Active Objects from Egocentric Vision with Symbolic World Knowledge/2bad6413-23a3-4118-b2c2-882a5a9baf49_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f91d8d50f965a1c6ec0529ba7419e18f4a03e0a1fec0ed18f81c7391b2827ac
+size 18701787

2023/Localizing Active Objects from Egocentric Vision with Symbolic World Knowledge/full.md

@@ -0,0 +1,412 @@
+# Localizing Active Objects from Egocentric Vision with Symbolic World Knowledge
+
+Te-Lin Wu* Yu Zhou* Nanyun Peng
+
+University of California, Los Angeles
+
+{telinwu,violetpeng}@cs.ucla.edu，yu.zhou@ucla.edu
+
+# Abstract
+
+The ability to actively ground task instructions from an egocentric view is crucial for AI agents to accomplish tasks or assist humans. One important step towards this goal is to localize and track key active objects that undergo major state change as a consequence of human actions/interactions in the environment (e.g., localizing and tracking the 'sponge' in video from the instruction "Dip the sponge into the bucket.") without being told exactly what/where to ground. While existing works approach this problem from a pure vision perspective, we investigate to which extent the language modality (i.e., task instructions) and their interaction with visual modality can be beneficial. Specifically, we propose to improve phrase grounding models' (Li* et al., 2022) ability in localizing the active objects by: (1) learning the role of objects undergoing change and accurately extracting them from the instructions, (2) leveraging pre- and post-conditions of the objects during actions, and (3) recognizing the objects more robustly with descriptional knowledge. We leverage large language models (LLMs) to extract the aforementioned action-object knowledge, and design a per-object aggregation masking technique to effectively perform joint inference on object phrases with symbolic knowledge. We evaluate our framework on Ego4D (Grauman et al., 2022) and Epic-Kitchens (Dunnhofer et al., 2022) datasets. Extensive experiments demonstrate the effectiveness of our proposed framework, which leads to  $>54\%$  improvements in all standard metrics on the TREK-150-OPE-Det localization + tracking task,  $>7\%$  improvements in all standard metrics on the TREK-150-OPE tracking task, and  $>3\%$  improvements in average precision (AP) on the Ego4D SCOD task.
+
+# 1 Introduction
+
+Recent technological advancements in smart glasses (and headsets) from industry leaders such
+
+![](images/74e920ca2ff8387ea705e8305203db403db156ae1bdfd7e06caf37ac458b7f03.jpg)  
+Figure 1: Active object grounding is the task of localizing the active objects undergoing state change (OUC). In this example action instruction "cut the pawpaw into half with the knife", the AI assistant is required to firstly infer the OUC (pawpaw) and the Tool (knife) from the instruction, and then localize them in the egocentric visual scenes throughout the action trajectories. Symbolic knowledge including pre/post conditions and object descriptions can bring additional information to facilitate the grounding.
+
+as Meta, Google, and Apple have attracted growing research in on-device AI that can provide just-in-time assistance to human wearers. While giving (or receiving) instructions during task execution, the AI assistant should co-observe its wearer's first-person (egocentric) viewpoint to comprehend the visual scenes and provide appropriate assistance. To accomplish this, it is crucial for AI to first be able to localize and track the objects that are undergoing significant state change according to the instruction and/or actions performed. For example in Figure 1, it can be inferred from the instruction that the object undergoing change which should be actively grounded and tracked is the pawpaw.
+
+Existing works have focused on the visual modality alone for such state change object localization tasks, including recognizing hand-object interac
+
+tions (Shan et al., 2020a) and object visual state changes (Alayrac et al., 2017). However, it remains under-explored whether the visual modality by itself is sufficient for providing signals to enable robust state change object localizing/tracking without enhanced signals from the textual modality. While utilizing a phrase grounding model (Liu et al., 2022, 2023) is presumably a straightforward alternative, it leaves unanswered questions of which mentioned objects/entities in the instruction are supposedly the one(s) that undergo major state changes, e.g., the pawpaw in Figure 1 instead of the knife is the correct target-object. Furthermore, how visual appearances of the objects can help such multimodal grounding is yet to be investigated.
+
+In light of this, we tackle the active object grounding task by first extracting target object mentions from the instructions using large language models (ChatGPT (OpenAI, 2023)) with a specifically designed prompting pipeline, and then finetuning an open-vocabulary detection model, GLIP (Li* et al., 2022), for visual grounding.
+
+We further hypothesize that additional action- and object-level symbolic knowledge could be helpful. As shown in Figure 1, state conditions prior to (pre-conditions: which indicate pre-action states) and after (post-conditions: which suggest at past state changes) the execution of the action are often considered when locating the objects, especially when the state changes are more visually significant. Furthermore, generic object knowledge including visual descriptions (e.g., "yellow-greenish flesh"), are helpful for uncommon objects. Based on this hypothesis, we prompt the LLM to obtain pre- and post-conditions on the extracted object mentions, along with a brief description focusing on specific object attributes.
+
+To improve the grounding models by effectively using all the aforementioned action-object knowledge, we design an object-dependent mask to separately attribute the symbolic knowledge to its corresponding object mentions for training. During inference time, a pre-/post-condition dependent scoring mechanism is devised to aggregate the object and the corresponding knowledge logit scores to produce a joint inference prediction.
+
+We evaluate our proposed framework on two
+
+![](images/cbad6f359d4f0fe9c0008f46af4dbfacb876a376ca5482a45339c324bef5f65b.jpg)  
+Use the spray nozzle(TOOL) to spray the car windshield(COUc).  
+Figure 2: Ego4D SCOD active grounding: Example object undergo change (OUC) due to the instructed actions and associated Tools, spanning: the pre-condition, point-of-no-return (PNR) and post-condition frames.
+
+![](images/28b054fdaf388baf89532124aadc6832a61c65fcea29232a7920f7685398c81c.jpg)
+
+![](images/d82eeb2e0f1833fc711429ea5a29f8157adbaa1d2b54ddeefad98a2a4662c109.jpg)
+
+narrated egocentric video datasets, Ego4D (Grauman et al., 2022) and Epic-Kitchens (Damen et al., 2022) and demonstrate strong gains. Our main contributions are two folds: (1) We design a sophisticated prompting pipeline to extract useful symbolic knowledge for objects undergoing state change during an action from instructions. (2) We propose a joint inference framework with a per-object knowledge aggregation technique to effectively utilize the extracted knowledge for improving multimodal grounding models.
+
+# 2 Tasks and Terminologies
+
+Active Object Grounding. For both robotics and assistant in virtual or augmented reality, the AI observes (or co-observes with the device wearer) the visual scene in the first-person (egocentric) point of view, while receiving (or giving) the task instructions of what actions to be performed next. To understand the context of the instructions as well as engage in assisting the task performer's actions, it is crucial to closely follow the key objects/entities that are involved in the actions undergoing major state change. We term these actively involved objects as objects undergoing change (OUC), and what facilitate such state change as Tools.
+
+Tasks. As there is not yet an existing resource that directly studies such active instruction grounding problem in real-world task-performing situations, we re-purpose two existing egocentric video datasets that can be seamlessly transformed into such a setting: Ego4D (Grauman et al., 2022) and Epic-Kitchens (Damen et al., 2018). Both come with per-time-interval annotated narrations transcribing the main actions occurred in the videos. $^{4}$
+
+Ego4D: SCOD. According to Ego4D's definition,
+
+object state change can encapsulate both spatial and temporal aspects. There is a timestamp that the state change caused by certain actions start to occur, i.e., the point-of-no-return (PNR). Ego4D's state change object detection (SCOD) subtask then defines, chronologically, three types of frames: the pre-condition (Pre), the PNR, and the post-condition (Post) frames, during a performed action. Pre-frames capture the prior (visual) states where a particular action is allowed to take place, while post-frames depict the outcomes caused by the action, and hence also record the associated object state change. Each frame annotated with its corresponding frame-type is further annotated with bounding boxes of the OUC (and Tools, if applicable), that is required to be regressed by the models. Figure 2 shows an exemplar SCOD data point.
+
+Our re-purposed active grounding task is thus as follows: Given an instructed action and one of a Pre/PNR/Post-typed frames, localize (ground) both the OUC(s) and Tool(s) in the visuals. While the official SCOD challenge only evaluates the PNR frame predictions, we consider all (Pre, PNR, and Post) frames for both training and inference.
+
+Epic-Kitchens: TREK-150. TREK-150 object tracking challenge (Dunnhofer et al., 2022, 2021) enriches a subset of 150 videos from the Epic-Kitchens (Damen et al., 2018, 2022) dataset, with densely annotated per-frame bounding boxes for tracking a target object. Since the Epic-Kitchens also comprises egocentric videos capturing human performing (specifically kitchen) tasks, the target objects to track are exactly the OUCs per the terminology defined above. Hence, given an instructed action, the model is required to ground and track the OUC in the egocentric visual scenes.[5]
+
+It is worth noting that some OUCs may occasionally go "in-and-out" of the egocentric point of view (PoV), resulting in partial occlusion and/or full occlusion frames where no ground truth annotations for the OUCs are provided. Such frames are excluded from the final evaluation. And in Section 4.2.3 we will show that our proposed model is very successful in predicting the objects when they come back due to the robustness of our symbolic joint inference grounding mechanism.
+
+![](images/fab224a9ddc31e0b69afa81f9684b9f35a3f6e16a7b7ab6f69f0f9cc082bbb92.jpg)  
+Figure 3: Overview of proposed framework that comprises a base multimodal phrase grounding model (GLIP), a frametype predictor, a knowledge extractor leveraging LLMs (GPT), and predictions supervised by both bounding box regression of the objects and their ranked scores.
+
+# 3 Method
+
+Figure 3 overviews the proposed framework, consisted of: (1) A base multimodal grounding architecture, where we adopt a strong open vocabulary object detection module, GLIP (Li* et al., 2022). (2) A frame-type prediction subcomponent which adds output projection layers on top of GLIP to utilize both image (frame) and text features to predict of which frame-type (Pre/PN-R/Post) is currently observed. (Section 3.2.1) (3) A prompting pipeline that is engineered to extract useful action-object knowledge from an LLM (GPT). (Section 3.2) (4) A per-object knowledge aggregation technique is applied to GLIP's word-region alignment contrastive training. (Section 3.3)
+
+# 3.1 Adapting GLIP
+
+GLIP (Li* et al., 2022; Zhang* et al., 2022) achieves open vocabulary object detection by pretraining on a contrastive phrase grounding objective. Specifically, GLIP extends the text(capTION)-to-image dot product matching objective from CLIP (Radford et al., 2021) to a word-region-level alignment objective. For some (tokenized) words of the textual description of an image, there are certain image region(s) that could be grounded to, while other regions are viewed as the negative samples for the CLIP-like alignment contrastive learning. During pretraining, GLIP utilizes both phrase grounding datasets (Ordonez et al., 2011; Plummer et al., 2015; Sharma et al., 2018) and object detection datasets (Krishna et al., 2017; Krasin et al., 2017; Shao et al., 2019).
+
+# Contrastive Learning. We illustrate the GLIP
+
+![](images/044346a19f0a9a974bb8d4a7b1d8fefc47c8914f0d45d82abb74b0fb6ff85839.jpg)  
+Figure 4: Model architecture (knowledge-enhanced grounding): On the left depicts the word-region alignment (contrastive) learning of the base GLIP architecture, where the model is trained to align the encoded latent word and image features with their dot-product logits being supervised by the positive and negative word-region pairs. On the right illustrates the enhanced object-knowledge grounding. During training we apply an object-type dependent mask to propagate the positive alignment supervisions; while during inference time the frame-type predictor (offline trained by the encoded textual and image features) acts as a combinator to fuse dot product-logit scores from both (extracted) object phrases and corresponding knowledge. (Note that for simplicity we do not fully split some phrases into individual words.)
+
+training adapted to our task in Figure 4 left half. Notice that for simplicity we do not fully expand the tokenized word blocks, e.g., "fish fillet" should span two words where each word (fish" and "fillet") and its corresponding region is all regarded as the positive matching samples. The model is trained to align the encoded latent word and image features with their dot-product logits being supervised by the positive and negative word-region pairs. The alignment scores will then be used to score (and rank) the regressed bounding boxes produced by the image features, and each box will feature an object-class prediction score. Concretely, for  $j$ th regressed box, its grounding score to a phrase  $W = \{w\}_{1:T}$  is a mean pooling of the dot-products between the  $j$ th region feature and all the word features that compose such a phrase:  $\mathbf{S}_j^{box} = \frac{1}{T}\sum_i^T\mathbf{I}_j\cdot \mathbf{W}_i$ . In this work, we mainly focus on the OUC and Tool object classes, i.e., each textually-grounded region will further predict whether it is an OUC or Tool class.
+
+# 3.2 LLM for Action-Object Knowledge
+
+Pipeline. As illustrated in Figure 5, we implement an LLM query pipeline to extract active entities and relevant symbolic knowledge from an instructional caption. To account for GPT's verbose tendency, we forcibly instruct GPT to produce the active objects (OUC and/or Tool) following a specifically designed format and then apply heuristic-based
+
+post-processing to further refine the extractions. Conditioned on the extracted OUC (and Tool), two additional queries are made to generate: (1) the symbolic pre- and post-conditions of such objects induced by the actions, and (2) brief descriptions characterizing the objects and their attributes. Interestingly, we empirically find it beneficial to situate GPT with a role, e.g., "From the first-person view."
+
+GPT Intrinsic Evaluation. In Table 2, we automatically evaluate the OUC/Tool extraction of GPT against the labelled ground truth entities in both datasets. We report both exact (string) match and word overlapping ratio (as GPT often extracts complete clauses of entities), to quantify the robustness of our GPT active entity extractions.
+
+Table 3 reports human evaluation results of GPT symbolic knowledge, including pre-/post-conditions and descriptions. Evaluation is based on two binary metrics, namely: (1). Textual Correctness: "Based on text alone, does the knowledge make sense?" and (2). Visual Correctness: "Does the conds./desc. match the image?" Despite impressive performance on both intrinsic evaluations, we qualitatively analyze in Table 1 some representative cases where GPT mismatches with annotations or humans, including cases where GPT's answer is actually more reasonable than the annotations.
+
+# 3.2.1 Incorporating Knowledge
+
+Adding Knowledge. We use the following schema to enrich the instruction with the obtained knowl-
+
+![](images/5e36b018de66228ffe31178a4d7416d63b21947b72b85ebfe934c30cda760799.jpg)  
+Figure 5: The GPT knowledge extraction pipeline. Demonstrated through an example from the Ego4D SCOD Dataset.
+
+![](images/7559ac7635aae52b639cceec9a7ededf5d18e8e6c0d1eafe9879fc01ab7f12e5.jpg)
+
+![](images/baddc87c86ea25983d4dcab8fa39c476999b6fd27b9a69d41f0214682b255314.jpg)
+
+<table><tr><td>Examples</td><td>Explanations</td></tr><tr><td>GPT: Pick up some green papers(OUC) from the table. [No Tool] 
+Desc.: &quot;Green papers are consultation documents issued by government.&quot;</td><td>Without visual knowledge input, GPT is not robust to phrase ambiguity, leading to undesirable definition.</td></tr><tr><td>GPT: Cut the fish fillet(OUC) with a knife(TOOL) 
+OUC: [Pre-cond]: &quot;fresh, raw&quot; [Post-cond]: &quot;sliced, cut&quot; 
+Tool: [Pre-cond]: &quot;Sharp, metallic&quot; [Post-cond]: &quot;Blunt, distorted&quot;</td><td>LLMs may hallucinate exaggerated state changes, in this case claiming the knife to be &quot;blunt, distorted&quot; after a single use, which is unreasonable.</td></tr><tr><td>GPT: Hold the iron(OUC) on the ironing board with your hand. [No Tool] 
+Ego4D gt-label: [OUC]: &quot;pants&quot; [Tool]: &quot;iron&quot;</td><td>&quot;Pants&quot; is not mentioned in the narration, GPT fails to capture OUC due to text narration reporting bias.</td></tr><tr><td>GPT: Spin the mop(OUC) in the mop bucket spinner(Tool). 
+Ego4D gt-label: [OUC]: &quot;mop&quot; [Tool]: &quot;mop&quot;</td><td>GPT prediction is more reasonable compared to the Ego4D ground truth label.</td></tr></table>
+
+Table 1: Qualitative Analysis of GPT Knowledge Extraction: Examples of cases where GPT-extracted symbolic knowledge are wrong or conflict with Ego4D annotations. Here the GPT-extracted or dataset-annotated knowledge are displayed in GREEN if they match human analysis and RED otherwise. Explanations for each example are provided on the right.  
+
+<table><tr><td rowspan="2">Object Type</td><td colspan="2">Ego4D</td><td colspan="2">TREK-150</td></tr><tr><td>EM (%)</td><td>Overlap.</td><td>EM (%)</td><td>Overlap.</td></tr><tr><td>OUC</td><td>77.8</td><td>88.6</td><td>76.0</td><td>94.3</td></tr><tr><td>Tool</td><td>60.3</td><td>88.5</td><td>—</td><td>—</td></tr></table>
+
+Table 2: Automatic evaluation of GPT entity extraction. Abbreviations: EM: exact string matching; Overlap: The ratio of GPT extractions fully covering the ground truth phrases  
+
+<table><tr><td rowspan="2">Knowledge Type</td><td colspan="2">Ego4D SCOD</td><td colspan="2">TREK-150</td></tr><tr><td>Textual</td><td>Visual</td><td>Textual</td><td>Visual</td></tr><tr><td>Pre-Cond.</td><td>86.5</td><td>81.6</td><td>83.0</td><td>79.9</td></tr><tr><td>Post-Cond.</td><td>75.2</td><td>70.3</td><td>76.6</td><td>73.5</td></tr><tr><td>Desc.</td><td>98.9</td><td>91.4</td><td>99.2</td><td>95.3</td></tr></table>
+
+Table 3: Human evaluation of GPT symbolic knowledge extraction. Abbreviations: Textual: i.e. "textual correctness" "Based on text alone, does the GPT cond./desc. make sense?"; Visual: i.e. "visual correctness": "Does the GPT cond./desc. match what is shown in the image?"
+
+edge: {"instr.} [SEP] object/tool (pre/post)-state is {conds.} [SEP] object/tool description is {desc.}", where [SEP] is the separation special token; {conds.} and {desc.} are the pre-/post-condition and object definition knowledge to be filled-in. Empirically, we find diffusing the post-condition knowl
+
+edge to PNR frame yield better results. As Figure 4 illustrates (omitting some prefixes for simplicity), we propagate the positive matching labels to object/tool's corresponding knowledge. In the same training mini-batch, we encourage the contrastiveness to focus on more detailed visual appearance changes grounded to the symbolic condition statements and/or descriptions, by sampling frames from the same video clips with higher probability.
+
+Frame-Type Prediction. Using both the encoded textual and image features, we learn an additional layer to predict the types of frames conditioned on the associated language instruction. Note that the frame-type definition proposed in Ego4D should be generalizable outside of the specific task, i.e., these frame types could be defined on any kinds of action videos. In addition to the annotated frames in SCOD, we randomly sub-sample nearby frames within 0.2 seconds (roughly 5-6 frames) to expand the training data. The frame-type prediction achieves a  $64.38\%$  accuracy on our SCOD test-set, which is then directly applied to the TREK-150 task for deciding the amount of pre- and post-condition
+
+knowledge to use given the multimodal inputs.
+
+# 3.3 Object-Centric Joint Inference
+
+Masking. As illustrated in Figure 4, a straightforward way to assign symbolic knowledge to its corresponding object type respectively is to construct a per-object-type mask. For example, an OUC mask  $\mathbf{M}_{OUC}$  will have 1s spanning the positions of the words from condition (e.g., "fresh, raw") of the OUC "fish fillet" in Figure 4) and descriptive knowledge, and 0s everywhere else. We omit the knowledge prefixes in Section 3.2.1 (e.g., the phrase "object state is") so that the models can concentrate on grounding the meaningful words. Such mask for each object type can be deterministically constructed to serve as additional word-region alignment supervisions, and can generalize to object types outside of OUC and Tool (beyond the scope of this work) as the GPT extraction can clearly indicate the object-to-knowledge correspondences. In other words, we enrich the GLIP's phrase grounding to additionally consider symbolic knowledge during the contrastive training. Note that the mask is frame-type dependent, e.g.,  $\mathbf{M}_{OUC}^{Pre}$  and  $\mathbf{M}_{OUC}^{Post}$  will focus on their corresponding conditional knowledge.
+
+Aggregation. During the inference time, we combine the frame-type prediction scores  $\mathbf{S}^{fr}$  with the per-object mask to aggregate the dot-product logit scores for ranking the regressed boxes. Specifically, we have  $\mathbf{S}_{OUC}^{box} = \sum_{fr} \mathbf{S}^{fr} * \mathbf{M}_{OUC}^{fr}$ , where  $\mathbf{S}$  is a 3-way logit and  $fr \in \{\mathrm{Pre}, \mathrm{PNR}, \mathrm{Post}\}$ .
+
+# 4 Experiments and Analysis
+
+We adopt the GLIP-L variant (and its pretrained weights) for all of our experiments, where its visual encoder uses the Swin-L transformer (Liu et al., 2021). We train the GLIP-L with our framework primarily on the SCOD dataset, and perform a zero-shot transfer to the TREK-150 task.
+
+# 4.1 Ego4D SCOD
+
+# 4.1.1 Experimental Setups
+
+Data Splits. We split the official SCOD train set following a 90-10 train-validation ratio and use the official validation set as our primary test set. $^{9}$
+
+Evaluation Metrics. Following the original SCOD task's main settings, we adopt average precision
+
+(AP) as the main evaluation metric, and utilize the COCO API (Lin et al., 2014) for metric computation. Specifically, we report AP, AP50, (AP at  $\mathrm{IOU}\geq 0.5$ ) and AP75 (AP at  $\mathrm{IOU}\geq 0.75$ ).
+
+# 4.1.2 Baselines
+
+We evaluate three categories of baselines: (1) Pure object detection models, where the language instructions are not utilized. (2) (Pseudo) referential grounding, where certain linguistic heuristics are used to propose the key OUCs. (3) GPT with symbolic knowledge, where GPT is used to extract both the OUCs and Tools, with additional symbolic knowledge available to utilize.
+
+Pure Object Detection (OD). We finetune the state-of-the-art model of the SCOD task from Chen et al. (2022) (VidIntern) on all types of frames (Pre, PNR, and Post) to serve as the pure object detection model baseline, which learns to localize the OUC from a strong hand-object-interaction prior in the training distribution. We also train an OD version of GLIP providing a generic instruction, "Find the object of change.", to quantify its ability to fit the training distribution of plausible OUCs.
+
+Pseudo Grounding (GT/SRL). We experiment four types of models utilizing the instructions and certain linguistic patterns as heuristics: (1) We extract all the nouns using Spacy NLP tool (Honnibal and Montani, 2017) and randomly assign OUC to one of which (Random Entity). (2) A simple yet strong baseline is to ground the full sentence of the instruction if the only object class to be predicted is the OUC type (Full-Instr.). (3) Following (2), we hypothesize that the first argument type (ARG1) of the semantic-role-labelling (SRL) parses (Shi and Lin, 2019; Gardner et al., 2017) of most simple instructions is likely regarded as the OUC ((SRL-ARG1)). (4) Lastly, to quantify a possible upper bound of simple grounding methods, we utilize the annotated ground truth object class labels from SCOD task and perform a simple pattern matching to extract the OUCs and Tools. For those ground truth words are not easily matched, we adopt the ARG1 method from (3) (GT-SRL-ARG1).
+
+GPT-based. For our main methods leveraging LLMs (GPT) and its generated action-object symbolic knowledge, we consider four types of combinations: (1) GPT with its extracted OUCs and Tools. (2) The model from (1) with additional utilization of object definitions  $(\mathbf{GPT} + \mathbf{Desc}.)$ . (3) Similar to (2) but condition on generated pre- and
+
+<table><tr><td rowspan="2">Base</td><td rowspan="2">Type</td><td rowspan="2">Method</td><td rowspan="2">Objects</td><td colspan="3">Pre-Frame↑</td><td colspan="3">PNR-Frame↑</td><td colspan="3">Post-Frame↑</td></tr><tr><td>AP</td><td>AP50</td><td>AP75</td><td>AP</td><td>AP50</td><td>AP75</td><td>AP</td><td>AP50</td><td>AP75</td></tr><tr><td rowspan="2">VidIntern</td><td rowspan="2">OD</td><td rowspan="2">—</td><td>OUC</td><td>32.73</td><td>49.17</td><td>34.05</td><td>37.49</td><td>57.04</td><td>38.59</td><td>29.68</td><td>44.43</td><td>30.94</td></tr><tr><td>Tool</td><td>16.39</td><td>23.43</td><td>17.25</td><td>16.53</td><td>24.51</td><td>17.14</td><td>14.03</td><td>21.70</td><td>14.44</td></tr><tr><td rowspan="16">GLIP-L</td><td rowspan="7">OD</td><td>—</td><td>OUC</td><td>26.91</td><td>42.83</td><td>27.86</td><td>29.74</td><td>47.70</td><td>30.47</td><td>24.13</td><td>38.74</td><td>24.71</td></tr><tr><td>Zero-Shot on GTs</td><td>OUC</td><td>20.18</td><td>32.97</td><td>20.63</td><td>19.51</td><td>32.34</td><td>19.39</td><td>19.34</td><td>31.07</td><td>19.88</td></tr><tr><td>Random Entity</td><td>OUC</td><td>25.90</td><td>42.17</td><td>26.20</td><td>26.85</td><td>44.21</td><td>26.80</td><td>24.45</td><td>39.17</td><td>25.10</td></tr><tr><td>Full-Instr.</td><td>OUC</td><td>32.45</td><td>51.62</td><td>33.34</td><td>33.78</td><td>54.44</td><td>34.49</td><td>31.30</td><td>49.23</td><td>32.42</td></tr><tr><td>SRL-ARG1</td><td>OUC</td><td>36.41</td><td>54.93</td><td>37.65</td><td>38.32</td><td>58.07</td><td>39.41</td><td>33.59</td><td>49.99</td><td>34.90</td></tr><tr><td rowspan="2">GT-SRL-ARG1</td><td>OUC</td><td>37.87</td><td>56.35</td><td>39.55</td><td>39.64</td><td>59.41</td><td>40.73</td><td>34.97</td><td>51.34</td><td>36.69</td></tr><tr><td>Tool</td><td>45.53</td><td>71.22</td><td>46.27</td><td>43.70</td><td>68.96</td><td>44.54</td><td>43.76</td><td>69.56</td><td>44.04</td></tr><tr><td rowspan="9">Instr.</td><td rowspan="2">GPT</td><td>OUC</td><td>37.46</td><td>56.05</td><td>38.96</td><td>39.07</td><td>59.17</td><td>40.13</td><td>34.77</td><td>51.34</td><td>36.35</td></tr><tr><td>Tool</td><td>38.41</td><td>60.66</td><td>39.33</td><td>37.64</td><td>60.26</td><td>39.29</td><td>37.67</td><td>59.73</td><td>38.24</td></tr><tr><td rowspan="2">GPT+Desc.</td><td>OUC</td><td>36.97</td><td>56.16</td><td>38.35</td><td>38.49</td><td>59.38</td><td>39.41</td><td>34.09</td><td>51.18</td><td>35.56</td></tr><tr><td>Tool</td><td>42.26</td><td>64.37</td><td>44.59</td><td>41.30</td><td>64.46</td><td>43.53</td><td>40.20</td><td>63.92</td><td>41.60</td></tr><tr><td rowspan="2">GPT+Conds.</td><td>OUC</td><td>38.65</td><td>57.55</td><td>40.16</td><td>40.19</td><td>60.39</td><td>41.56</td><td>35.40</td><td>52.15</td><td>37.11</td></tr><tr><td>Tool</td><td>43.48</td><td>65.78</td><td>45.58</td><td>42.37</td><td>64.97</td><td>44.77</td><td>41.08</td><td>63.26</td><td>42.07</td></tr><tr><td>w/o obj.-mask</td><td>OUC</td><td>37.59</td><td>56.28</td><td>39.19</td><td>39.09</td><td>59.31</td><td>40.58</td><td>33.93</td><td>50.80</td><td>35.38</td></tr><tr><td rowspan="2">GPT+Conds.+Desc.</td><td>OUC</td><td>38.27</td><td>57.79</td><td>39.65</td><td>39.96</td><td>60.91</td><td>41.35</td><td>35.37</td><td>52.82</td><td>36.95</td></tr><tr><td>Tool</td><td>44.00</td><td>66.49</td><td>46.12</td><td>42.77</td><td>66.06</td><td>44.82</td><td>42.12</td><td>65.44</td><td>42.45</td></tr></table>
+
+post-conditions of the objects (GPT+Conds.).(4) Combining both (2) and (3) (GPT+Conds.+Desc.).
+
+# 4.1.3 Results
+
+Table 4 summarizes the overall model performance on Ego4D SCOD task. Even using the ground truth phrases, GLIP's zero-shot performance is significantly worse than pure OD baselines, implying that many of the SCOD objects are uncommon to its original training distribution. Generally, the instruction grounded performance (Instr.) are all better than the pure OD baselines, even with using the whole instruction sentence as the grounding phrase. The significant performance gaps between our models and the VidIntern baseline verifies that visual-only models can be much benefit from incorporation of textual information (should they be available) for the active object grounding task.
+
+Particularly for OUC, with vanilla GPT extractions we can almost match the performance using the ground truth phrases, where the both the conditional and definition symbolic knowledge further improve the performance. Notice that condition knowledge by itself is more useful than the definition, and would perform better when combined. We also ablate a row excluding the per-object aggregation mechanism so that the conditional knowledge is simply utilized as a contextualized suffix for an instruction, which indeed performs worse, especially for the post-frames. As implied in Table 4, best performance on Tool is achieved using the ground truth phrases, leaving room for improve
+
+ment on more accurate extractions and search of better suited symbolic knowledge.
+
+Table 4: Model performance on Ego4D SCOD. OD: pure object detection. Instr: grounding with instructions. We highlight best OUC performance in RED for and best Tool performance in GREEN.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Top-K</td><td colspan="2">Post-Frame↑</td><td>OD Metric</td></tr><tr><td>AP</td><td>AP50</td><td>AP75</td></tr><tr><td>Track from GT PNR</td><td>1</td><td>20.36</td><td>41.15</td><td>17.78</td></tr><tr><td>Track from Pred. PNR</td><td>1</td><td>10.21</td><td>21.27</td><td>8.63</td></tr><tr><td>GPT+Conds.+Desc.</td><td>1</td><td>29.85</td><td>43.53</td><td>31.45</td></tr></table>
+
+Table 5: PNR to Post OUC tracking ablation study. Since tracking module only produce a single box for each frame, we report the top-1 performance of our grounding model. (Normally COCO API reports max 100 detection boxes.)
+
+However, one may raise a natural question: if the OUC/Tool can be more robustly localized in the PNR frame, would a tracker improve the post-frame performance over our grounding framework? We thus conduct an ablation study using the tracker in Section 4.2 to track from PNR-frames using either the ground truth box and our model grounded box to the post-frames. Results in Table 5 contradict this hypothesis, where we find that, due to viewpoint variations and appearance differences induced by the state change, our grounding model is significantly more robust than using tracking.
+
+# 4.1.4 Qualitative Inspections
+
+Figure 6 shows six different examples for in-depth qualitative inspections. It mainly shows that, generally, when the models grounding with the symbolic knowledge outperforms the ones without, the provided symbolic knowledge, especially the conditional knowledge, plays an important role
+
+<table><tr><td rowspan="2">FPV OD</td><td rowspan="2">BBox Rank.</td><td rowspan="2">Method</td><td colspan="2">OPE-Det↑</td><td colspan="2">OPE↑</td><td colspan="3">Std. OD Metric↑</td></tr><tr><td>SS</td><td>NPS</td><td>SS</td><td>NPS</td><td>AP</td><td>AP50</td><td>AP75</td></tr><tr><td>HiC</td><td>MDNet</td><td>LTMU-H</td><td>0.267</td><td>0.261</td><td>0.505</td><td>0.520</td><td>—</td><td>—</td><td>—</td></tr><tr><td>HiC</td><td>MDNet</td><td>TbyD-H</td><td>0.047</td><td>0.018</td><td>0.433</td><td>0.455</td><td>—</td><td>—</td><td>—</td></tr><tr><td>Swin-L + DINO</td><td>VidIntern</td><td>—</td><td>0.341</td><td>0.340</td><td>0.526</td><td>0.541</td><td>29.49</td><td>41.47</td><td>30.85</td></tr><tr><td rowspan="7">GLIP-L</td><td rowspan="7">GLIP-L</td><td>Full-Instr.</td><td>0.355</td><td>0.361</td><td>0.521</td><td>0.537</td><td>38.51</td><td>60.06</td><td>40.17</td></tr><tr><td>SRL-ARG1</td><td>0.373</td><td>0.377</td><td>0.528</td><td>0.544</td><td>40.00</td><td>60.52</td><td>40.96</td></tr><tr><td>GT-SRL-ARG1</td><td>0.383</td><td>0.390</td><td>0.531</td><td>0.548</td><td>42.35</td><td>61.41</td><td>44.27</td></tr><tr><td>GPT</td><td>0.379</td><td>0.389</td><td>0.529</td><td>0.545</td><td>41.85</td><td>61.89</td><td>43.46</td></tr><tr><td>GPT+Desc.</td><td>0.402</td><td>0.409</td><td>0.528</td><td>0.543</td><td>41.40</td><td>60.70</td><td>43.17</td></tr><tr><td>GPT+Conds.</td><td>0.412</td><td>0.422</td><td>0.541</td><td>0.557</td><td>45.90</td><td>67.26</td><td>47.94</td></tr><tr><td>GPT+Desc.+Conds.</td><td>0.413</td><td>0.424</td><td>0.539</td><td>0.557</td><td>43.49</td><td>64.34</td><td>45.43</td></tr></table>
+
+Table 6: Model performance on TREK-150. OPE denotes One-Pass Evaluation (Dunnhofer et al., 2022) and OPE-Det is a variant to OPE where each tracker is initialized with its corresponding object detector prediction on the first frame. Success Score (SS) and Normalized Precision Score (NPS) are standard tracking metrics.
+
+# 4.2 TREK-150
+
+# 4.2.1 Experimental Setups
+
+Protocols. TREK-150's official evaluation protocol is One-Pass Evaluation (OPE), where the tracker is initialized with the ground-truth bounding box of the target in the first frame; and then ran on every subsequent frame until the end of the video. Tracking predictions and the ground-truth bounding boxes are compared based on IOU and distance of box centers. However, as the premise of having ground-truth bounding box initialization can be generally impractical, a variant of OPE-Det is additionally conducted, where the first-frame bounding box is directly predicted by our trained grounding model (grounded to the instructions).
+
+Evaluation Metrics. Following Dunnhofer et al. (2022), we use common tracking metrics, i.e., Success Score (SS) and Normalized Precision Score (NSP), as the primary evaluation metrics. In addition, we also report standard OD metric (APs) simply viewing each frame to be tracked as the localization task, as an alternative reference.
+
+# 4.2.2 Baselines.
+
+We adopt the best performing framework, the LTMU-H, in the original TREK-150 paper as the major baseline. LTMU-H integrates an fpv (first-person-view) object detector (HiC (Shan et al., 2020b)) into a generic object tracker (LTMU (Dai et al., 2020)), which is able to re-focus the tracker on the active objects after the (tracked) focus is lost (i.e., identified by low tracker confidence scores).
+
+Following the convention of utilizing object detection models to improve tracking (Feichtenhofer et al., 2017), we focus on improving object tracking performance by replacing the HiC-detector with our knowledge-enhanced GLIP models. We substi
+
+tute the HiC-detector for all 8 GLIP-based models and the VideoIntern baseline trained on the SCOD task and perform a zero-shot knowledge transfer (directly from Ego4D SCOD).<sup>10</sup>
+
+# 4.2.3 Results
+
+Table 6 summarizes the performance on TREK-150. Our best GLIP model trained using GPT-extracted objects and symbolic knowledge outperforms the best HiC baseline by over  $54\%$  relative gains in the SS metric and over  $62\%$  relative gains in the NPS scores for the OPE-D task. It also outperforms the VideoIntern baseline by  $16 - 18\%$  relative gains in SS/NPS and even the GLIP-(GT-SRL-ARG1) model by  $7 - 9\%$  relative gains on both metrics. This demonstrates the transferability of our OUC grounding model in fpv-tracking. For the OPE task with ground-truth initializations, the gains provided by our GLIP-GPT models over LTMU-H narrow to  $7 - 8\%$  relative gains across both metrics while still maintaining a lead over all other methods. This shows that the model is still able to better help the tracker re-focus on the OUCs although the overall tracking performance is more empirically bounded by the tracking module.
+
+# 5 Related Works
+
+Egocentric Vision. Egocentric vision has recently attracted research attentions thanks to advancements in smart wearable devices and robotics. Datasets used in this work, Ego4D (Grauman et al., 2022) and Epic-Kitchens (Damen et al., 2022, 2018; Dunnhofer et al., 2022) are two representative large-scale collections of egocentric videos
+
+![](images/8b5f2185342fe318516bb98e78af3d212ff8dfd6950491d44f9a73088b163b8f.jpg)  
+(A)
+
+![](images/748ad0aa68684972a8547effb9ac881574611c3e4b64aa0d8f063f215d0313c9.jpg)  
+(B)
+
+![](images/31c4091a458a8f71067ee2e145743a49357a86a6f77ab8dc0e1bd1a3f6874a44.jpg)  
+(c)
+
+![](images/c37cc55ed2853e2e9d05bbcb2a0f4804704fc3e0373d100c591991c45fc5c18a.jpg)  
+(D)
+
+![](images/b12d132ef2950778287eff497144c62b78ab59c1d1867cd7f6539a30d7fc065b.jpg)  
+(E)  
+Figure 6: Qualitative inspections, mainly on the effectiveness of the GPT generated symbolic knowledge. Bounding box color code: Ground truth boxes, Models with uses of symbolic knowledge (MD-1), i.e. the GPT+Conds.+Desc.; Models without uses of symbolic knowledge (MD-2), i.e., the vanilla GPT.
+
+![](images/9480c8dc5fd442ba6304bd7a5ddacb851fe57c6bd6805f2ef21d1ea3b4ca261e.jpg)  
+(F)
+
+recording tasks performed by the camera wearers. Other existing works have also investigated egocentric vision in audio-visual learning (Kazakos et al., 2019), object detection with EgoNet (Bertasius et al., 2017; Furnari et al., 2017), object segmentation with eye-gazes (Kirillov et al., 2023) and videos (Darkhalil et al., 2022).
+
+Action-Object Knowledge. The knowledge of objects are often at the center of understanding human actions. Prior works in both NLP and vision communities, have studied problems such as tracking visual object state changes (Alayrac et al., 2017; Isola et al., 2015; Yang et al., 2022), understanding object manipulations and affordances (Shan et al., 2020a; Fang et al., 2018), tracking textual entity state changes (Branavan et al., 2012; Bosselut et al., 2018; Mishra et al., 2018; Tandon et al., 2020), and understanding textual pre-/post-conditions from action instructions (Wu et al., 2023). While handobject interactions (Shan et al., 2020a; Fu et al., 2022) are perhaps one of the most common object manipulation schemes, the objects undergoing change may not be directly in contact with the hands (see Figure 2). Here additional textual information can aid disambiguating the active object during localization and tracking. In this spirit, our work marries the merits from both modalities to tackle the active object grounding problem according to specific task instructions, and utilize action-object knowledge to further improve the models.
+
+Multimodal Grounding. In this work, we adopt the GLIP model (Li* et al., 2022; Zhang* et al., 2022) for its compatibility with our settings and the joint inference framework, which indeed demonstrate significant improvements for the active object grounding task. There are many related works for multimodal grounding and/or leveraging language (LLMs) to help with vision tasks, including (but not limited to) Grounding-DINO (Liu et al., 2023), DQ-DETR (Liu et al., 2022), ELEVATER (Li* et al., 2022), phrase segmentation (Zou* et al., 2022), visually-enhanced grounding (Yang et al., 2023), video-to-text grounding (Zhou et al., 2023), LLM-enhanced zero-shot novel object classification (Naeem et al., 2023), and multimodal object description generations (Li et al., 2022, 2023).
+
+# 6 Conclusions
+
+In this work, we approach the active object grounding task leveraging two narrated egocentric video datasets, Ego4D and Epic-Kitchens. We propose a carefully designed prompting scheme to obtain useful action-object knowledge from LLMs (GPT), with specific focuses on object pre-/post-conditions during an action and its attributional descriptions. Enriching the GLIP model with the aforementioned knowledge as well as the proposed perobject knowledge aggregation technique, our models outperforms various strong baselines in both active object localization and tracking tasks.
+
+# 7 Limitations
+
+We hereby discuss the potential limitations of our work:
+
+(1) While we make our best endeavours to engineer comprehensive and appropriate prompts for obtaining essential symbolic action-object knowledge from large language models (LLMs) such as GPT, there are still few cases where the extracted objects are not ideal (see Table 1). Hence, our model performance could potentially be bounded by such limitation inherited from the LLM ability to fully and accurately comprehend the provided instructions. Future works can explore whether more sophisticated in-context learning (by providing examples that could be tricky to the LLM) would be able to alleviate this issue. Alternatively, we may utilize LLM-self-constructed datasets to fine-tune another strong language models (such as Alpaca (Taori et al., 2023)) for the object extraction task. On the other hand, incorporating high-level descriptions of the visual contexts using off-the-shelf captioning models could also be explored to make the LLM more situated to further improve the efficacy of the extracted knowledge.
+
+(2) As this work focuses on action frames in first-person egocentric videos (from both Ego4D (Grauman et al., 2022) and Epic-Kitchens dataset (Damen et al., 2018; Dunnhofer et al., 2022) datasets), the underlying learned model would obviously perform better on visual observations/scenes from first-person viewpoints. While we hypothesize that, unless heavy occlusion and drastic domain shifts (of the performed tasks and/or objects involved) occur, the learned models in this work should be able to transfer to third-person viewpoints, we have not fully tested and verified such hypothesis. However, if applied properly, the overall framework as well as the utilization of LLMs for action-object knowledge should be well-generalizable regardless of the viewpoints.
+
+(3) There is more object- and action-relevant knowledge that could be obtained from LLMs, such as spatial relations among the objects, size difference between the objects, and other subtle geometrical transitions of the objects. During experiments, we attempted to incorporate spatial and size information to our models. However, experimental results on the given datasets did not show significant improvement. Thus we omitted them from this work. We hope to inspire future relevant research along this line to further exploit other potentially
+
+useful knowledge.
+
+# 8 Ethics and Broader Impacts
+
+We hereby acknowledge that all of the co-authors of this work are aware of the provided ACL Code of Ethics and honor the code of conduct. This work is mainly about understanding and localizing the key objects undergoing major state changes during human performing an instructed or guided action, which is a crucial step for application such as ondevice AI assistant for AR or VR devices/glasses.
+
+Datasets. We mainly use the state-change-object-detection (SCOD) subtask provided by Ego4D (Grauman et al., 2022) grand challenge for training and (testing on their own tasks), and transfer the learned knowledge to TREK-150 (Dunnhofer et al., 2022) from Epic-Kitchens dataset (Damen et al., 2018). These datasets is not created to have intended biases towards any population where the videos are collected spanning across multiple regions around the world. The main knowledge learned from the dataset is mainly physical knowledge, which should be generally applicable to any social groups when conducting daily tasks.
+
+Techniques. We propose to leverage both the symbolic knowledge from large-language models (LLMs) such as GPT to guide the multimodal grounding model and the visual appearances and relations among objects for localizing the object undergo changes. The technique should be generally transferable to similar tasks even outside of the domain concerned in this work, and unless misused intentionally to harmful data (or trained on), should not contain harmful information for its predictions and usages.
+
+# Acknowledgments
+
+Many thanks to I-Hung Hsu for his valuable feedback during our paper preparation, and to Lianian Harold Li for his tips on using his great work, GLIP models. This material is based on research supported by the Machine Common Sense (MCS) program under Cooperative Agreement N66001-19-2-4032 and the ECOLE program under Cooperative Agreement HR00112390060, both with the US Defense Advanced Research Projects Agency (DARPA). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing DARPA, or the U.S. Government.
+
+# References
+
+Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. 2017. Joint discovery of object states and manipulation actions. In International Conference on Computer Vision (ICCV).  
+Gedas Bertasius, Hyun Soo Park, Stella X Yu, and Jianbo Shi. 2017. First person action-object detection with egonet. In Robotics: Science and Systems (RSS).  
+Antoine Bosselut, Omer Levy, Ari Holtzman, Corin Ennis, Dieter Fox, and Yejin Choi. 2018. Simulating action dynamics with neural process networks. In International Conference on Learning Representations (ICLR).  
+S.R.K. Branavan, Nate Kushman, Tao Lei, and Regina Barzilay. 2012. Learning high-level planning from text. In Association for Computational Linguistics (ACL).  
+Guo Chen, Sen Xing, Zhe Chen, Yi Wang, Kunchang Li, Yizhuo Li, Yi Liu, Jiahao Wang, Yin-Dong Zheng, Bingkun Huang, et al. 2022. Internvideo-ego4d: A pack of champion solutions to ego4d challenges. arXiv preprint arXiv:2211.09529.  
+Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. 2020. High-performance long-term tracking with meta-updater. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6297-6306.  
+Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. 2018. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV).  
+Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonio Furnari, Jian Ma, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, and Michael Wray. 2022. Rescaling egocentric vision: Collection, pipeline and challenges for epic-kitchens-100. International Journal of Computer Vision (IJCV).  
+Ahmad Darkhalil, Dandan Shan, Bin Zhu, Jian Ma, Amlan Kar, Richard Higgins, Sanja Fidler, David Fouhey, and Dima Damen. 2022. Epic-kitchens visor benchmark: Video segmentations and object relations. In Proceedings of the Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks.  
+Matteo Dunnhofer, Antonino Furnari, Giovanni Maria Farinella, and Christian Micheloni. 2021. Is first person vision challenging for object tracking? In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops.  
+Matteo Dunnhofer, Antonino Furnari, Giovanni Maria Farinella, and Christian Micheloni. 2022. Visual
+
+object tracking in first person vision. In International Journal of Computer Vision (IJCV).  
+Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J. Lim. 2018. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. 2017. Detect to track and track to detect. In International Conference on Computer Vision (ICCV).  
+Qichen Fu, Xingyu Liu, and Kris Kitani. 2022. Sequential voting with relational box fields for active object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Antonino Furnari, Sebastiano Battiato, Kristen Grauman, and Giovanni Maria Farinella. 2017. Next-active-object prediction from egocentric videos. volume 49, pages 401-411. Elsevier.  
+Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson F. Liu, Matthew Peters, Michael Schmitz, and Luke S. Zettlemoyer. 2017. Alennlp: A deep semantic natural language processing platform.  
+Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolar, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbelaez, David Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. 2022. Ego4d: Around the World in 3,000 Hours of Egocentric Video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language understanding with Bloom embeddings, convolutional neural networks and incremental parsing. To appear.
+
+Phillip Isola, Joseph J. Lim, and Edward H. Adelson. 2015. Discovering states and transformations in image collections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Evangelos Kazakos, Arsha Nagrani, Andrew Zisserman, and Dima Damen. 2019. Epic-fusion: Audiovisual temporal binding for egocentric action recognition. In International Conference on Computer Vision (ICCV).  
+Diederik P Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR).  
+Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dolkar, and Ross Girshick. 2023. Segment anything. arXiv:2304.02643.  
+Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, et al. 2017. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages.  
+Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. 2017. Visual genome: Connecting language and vision using crowdsourced dense image annotations. In International Journal of Computer Vision (IJCV).  
+Chunyuan Li*, Haotian Liu*, LIunian Harold Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Yong Jae Lee, Houdong Hu, Zicheng Liu, et al. 2022. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. In Track on Datasets and Benchmarks, Neural Information Processing Systems (NeurIPS).  
+Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, and Steven C. H. Hoi. 2022. Lavis: A library for language-vision intelligence.  
+Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597.  
+Lianian Harold Li*, Pengchuan Zhang*, Haotian Zhang*, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. 2022. Grounded language-image pre-training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV).
+
+Shilong Liu, Yaoyuan Liang, Feng Li, Shijia Huang, Hao Zhang, Hang Su, Jun Zhu, and Lei Zhang. 2022. Dq-detr: Dual query detection transformer for phrase extraction and grounding. In Association for the Advancement of Artificial Intelligence (AAAI).  
+Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. 2023. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499.  
+Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision (ICCV).  
+Bhavana Dalvi Mishra, Lifu Huang, Niket Tandon, Wen-tau Yih, and Peter Clark. 2018. Tracking state changes in procedural text: a challenge dataset and models for process paragraph comprehension. In North American Chapter of the Association for Computational Linguistics (NAACL-HLT).  
+Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian, Muhammad Zeshan Afzal, Didier Stricker, Luc Van Gool, and Federico Tombari. 2023. I2mvformer: Large language model generated multi-view document supervision for zero-shot image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+OpenAI. 2023. Chatgpt.  
+Vicente Ordonez, Girish Kulkarni, and Tamara Berg. 2011. Im2text: Describing images using 1 million captioned photographs.  
+Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. 2015. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In International Conference on Computer Vision (ICCV).  
+Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020.  
+Dandan Shan, Jiaqi Geng, Michelle Shu, and David Fouhey. 2020a. Understanding human hands in contact at internet scale. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Dandan Shan, Jiaqi Geng, Michelle Shu, and David F. Fouhey. 2020b. Understanding human hands in contact at internet scale. pages 9866-9875.
+
+Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. 2019. Objects365: A large-scale, high-quality dataset for object detection. In International Conference on Computer Vision (ICCV).  
+Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. 2018. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Association for Computational Linguistics (ACL), pages 2556-2565.  
+Peng Shi and Jimmy Lin. 2019. Simple bert models for relation extraction and semantic role labeling. ArXiv, abs/1904.05255.  
+Niket Tandon, Keisuke Sakaguchi, Bhavana Dalvi, Dheeraj Rajagopal, Peter Clark, Michal Guerquin, Kyle Richardson, and Eduard Hovy. 2020. A dataset for tracking entities in open domain procedural text. In Empirical Methods in Natural Language Processing (EMNLP), pages 6408-6417.  
+Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. 2023. Alpaca: A strong, replicable instruction-following model. Stanford Center for Research on Foundation Models. https://crfm.stanford.edu/2023/03/13/alpaca.html, 3(6):7.  
+Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, 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. 2020. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online. Association for Computational Linguistics.  
+Te-Lin Wu, Caiqi Zhang, Qingyuan Hu, Alex Spangher, and Nanyun Peng. 2023. Learning action conditions from instructional manuals for instruction understanding.  
+Guang Yang, Manling Li, Jiajie Zhang, Xudong Lin, Shih-Fu Chang, and Heng Ji. 2022. Video event extraction via tracking visual states of arguments. In AAAI Conference on Artificial Intelligence.  
+Ziyan Yang, Kushal Kafle, Franck Dernoncourt, and Vicente Ordonez. 2023. Improving visual grounding by encouraging consistent gradient-based explanations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).  
+Haotian Zhang*, Pengchuan Zhang*, Xiaowei Hu, YenChun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. 2022. Glipv2: Unifying localization and vision-language understanding. In Neural Information Processing Systems (NeurIPS).
+
+Yu Zhou, Sha Li, Li Manling, Lin Xudong, Shih-Fu Chang, Mohit Bansal, and Heng Ji. 2023. Non-sequential graph script induction via multimedia grounding. In Proc. the 61th Annual Meeting of the Association for Computational Linguistics (ACL2023).  
+Xueyan Zou*, Zi-Yi Dou*, Jianwei Yang*, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Jianfeng Wang, Lu Yuan, Nanyun Peng, Lijuan Wang, Yong Jae Lee*, and Jianfeng Gao*. 2022. Generalized decoding for pixel, image and language. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
+
+# A GPT Prompt Engineering Details
+
+# A.1 GPT Prompts
+
+Below are the prompts used in the GPT entity and symbolic knowledge extraction pipeline as described in Figure 5.
+
+```txt
+GPT Prompt Set
+```
+
+```txt
+Identify OUC & Tool:
+```
+
+```txt
+GPT Prompt 1 = "From the action [{}], please extract the follow objects exactly as they are written: (a) The single object being manipulated (b) The one single tool used to manipulate (a). Please do not explain anything in you answer. Please only output one object for each part. If you cannot find either (a) or (b), please simply output [ None] for that field and print nothing else."
+```
+
+```txt
+Precise Semantic Grounding (for both OUC and Tool):
+```
+
+```txt
+GPT Prompt 2 = "Please output an exact subpart of the sentence [{}] where the object [{}] is referred to. Please remove leading verbs in your answer. Please output only the exact sentence subpart and nothing else. Please make sure that you answer can be exactly found in the sentence {}. If question is not valid, please simply output [None] and nothing else. Please do not output any explanation of your answer."
+```
+
+```txt
+State Change Forcast
+```
+
+```txt
+GPT Prompt 3 = "From the first person view of the action [{}], would the object [{}] undergo significant change in its visual appearance? If so, please output [yes] in the first line and If not, please output [no] in the first line. If the answer is yes, then on the second line, please simply print one or two visually recognizable adjectives to describe the appearance of the object before the state change, on the third line, please simply print a few (<7) words that visually describe the object after the state change"
+```
+
+```txt
+Object Description
+```
+
+```txt
+GPT Prompt 4 = "In one sentence, please define the object[{}] visually"
+```
+
+# A.2 GPT Pipeline
+
+The GPT prompts are used in the order above. First, we plug in the instructional caption to GPT Prompt 1 and generate a tentative OUC-Tool pair. If either entity is not found, GPT would return None and
+
+future queries will ignore that entity.
+
+Due to GPT's tendency to paraphrase and provide unwanted chain-of-thought explanations, we further pass GPT outputs for OUC and/or tool through an additional query aimed to produce exact grounding if it is not contained in the caption. Emperically this query has a very high chance of producing a caption-groundable answer if exist. In the unlikely event that no groundable OUC is found after running GPT Prompt 2, we resort to using SRL [ARG1] as the OUC. On the other hand, if no groundable Tool is found after running GPT Prompt 2, we set the Tool to None.
+
+At this point, we run GPT Prompts 3 and 4 on the OUC and Tool (if exist). GPT results are parsed and stored if following the designated formats.
+
+# A.3 GPT Situated Role
+
+For a small randomly selected subset of Ego4d prompts (220 samples) in the human evaluation of GPT generated answers for pre/post conditions (Table 2), we evaluated results both with and without the situated role specification "From the first-person view." in Table 7.
+
+<table><tr><td rowspan="2">Prompt Style</td><td colspan="2">Pre Cond.</td><td colspan="2">Post Cond.</td></tr><tr><td>Textual</td><td>Visual</td><td>Textual</td><td>Visual</td></tr><tr><td>wo/ Situated Role</td><td>70.9</td><td>75.0</td><td>61.4</td><td>64.5</td></tr><tr><td>w/ Situated Role</td><td>86.4</td><td>82.3</td><td>75.5</td><td>71.8</td></tr></table>
+
+Table 7: GPT prompt comparison on Ego4D subset. Here "Textual" refers to "textual correctness": i.e. "based on text and common sense alone, does the GPT cond./desc. make sense?"; whereas "Visual" refers to "visual correctness": i.e. "does the GPT cond./desc. match what is shown in the image?"
+
+We observed that when provided with additional situated viewpoint specification, GPT results displayed a significant increase in alignment with human judgment, thus it was incorporated into our final prompt.
+
+# B Details of Modeling & Learning
+
+# B.1 Narration Processing
+
+Ego4D. Ego4D's original annotated narrations are of third-person descriptive tone, e.g., "#C cuts the vegetables with the knife.", where the symbol of #C indicates the camera wearer. Since this pattern is universally applicable to all the available narrations of the videos, we perform a deterministic string transformation stripping these special indi
+
+cators of camera wearers or other human characters, followed by a simple conversion to change the third-person singular verb to first-person verb. The rest of the pronouns are also deterministically converted, so that the narration becomes a first-person imperative tone. Although we did not empirically find such a transformation make any significant impacts to the model performance, we conduct this transformation for GPT to more easily grasp the situated roles (given to it) as well as conform to the main motivation of this work better.
+
+Epic-Kitchens. For TREK-150 where its videos basically belong to Epic-Kitchens, the original released narrations are already in an imperative instructional tone. In light of this, we simply adopt them as the task instructions seamlessly without needing to perform any modifications.
+
+# B.2 Training & Implementation Details
+
+Training Details. For GLIP series of models, we simply truncate the textual inputs (the action instruction texts) at maximum 256 tokens, ensuring all of the textual inputs are below such maximum length even with additional symbolic knowledge is incorporated. The hyperparameters are manually tuned against an Ego4D SCOD validation set, and the checkpoints used for testing are selected by the best performing ones on such set.
+
+All the models in this work are trained on 2-4 Nvidia A100/A6000 GPUs<sup>1112</sup> on a Ubuntu 20.04.2 operating system. The hyperparameters for each model are manually tuned against the development datasets, and the checkpoints used for testing are selected by the best performing ones on such held-out development sets.
+
+Implementation Details. The implementations of the transformer-based models are extended from the HuggingFace $^{13}$  code base (Wolf et al., 2020), and our entire code-base is implemented in PyTorch. $^{14}$
+
+# B.3 Hyperparameters
+
+We train our models until performance convergence is observed on the held-out development set (split from the original train set from Ego4D SCOD subtask). The training time is roughly 12-14 hours, spanning 10-15 training epochs. We list all the
+
+hyperparameters used in Table 8. The basic hyperparameters such as learning rate, batch size, and gradient accumulation steps, are kept consistent for models based off the same architecture (for baselines and our GLIP-L model series). All of our models adopt the same search bounds and ranges of trials as in Table 9.
+
+<table><tr><td>Models</td><td>Batch Size</td><td>Initial LR</td><td># Training Epochs</td><td>Gradient Accu-mulation Steps</td><td># Params</td></tr><tr><td>VideoIntern Baselines</td><td>8</td><td>1 × 10-3</td><td>10</td><td>1</td><td>218M</td></tr><tr><td>GLIP-L Series</td><td>4</td><td>1 × 10-3</td><td>10</td><td>1</td><td>429M</td></tr></table>
+
+Table 8: Hyperparameters in this work: Initial LR denotes the initial learning rate. All the models are trained with Adam optimizers (Kingma and Ba, 2015). We include number of learnable parameters of each model in the column of # params.  
+
+<table><tr><td>Type</td><td>Batch Size</td><td>Initial LR</td><td># Training Epochs</td><td>Gradient Accumulation Steps</td></tr><tr><td>Bound (lower-upper)</td><td>2-8</td><td>1 × 10-3-1 × 10-5</td><td>5-15</td><td>1</td></tr><tr><td>Number of Trials</td><td>2-4</td><td>2-3</td><td>2-4</td><td>1</td></tr></table>
+
+Table 9: Search bounds for the hyperparameters of all the models.

2023/Localizing Active Objects from Egocentric Vision with Symbolic World Knowledge/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba8765d7daa57b88a0c000f42f7a7ef25d501128e8857102c2ef757ec64b0e41
+size 868801

2023/Location-Aware Visual Question Generation with Lightweight Models/0e4f617e-b625-4a64-9c41-5b1c70b50c7f_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01dffe56bee7883bc83a0e38a8ae0f4f651ce33870ee447114390aa782920652
+size 5210666

2023/Location-Aware Visual Question Generation with Lightweight Models/full.md

@@ -0,0 +1,659 @@
+# Location-Aware Visual Question Generation with Lightweight Models
+
+Nicholas Collin Suwono<sup>1,2</sup> Chih Yao Chen<sup>1</sup> Tun Min Hung<sup>1</sup>
+
+Ting-Hao 'Kenneth' Huang<sup>3</sup> I-Bin Liao<sup>4</sup> Yung-Hui Li<sup>4</sup> Lun-Wei Ku<sup>1</sup> Shao-Hua Sun<sup>2</sup>
+
+Institute of Information Science, Academia Sinica $^{1}$  National Taiwan University $^{2}$
+
+Pennsylvania State University $^{3}$  Hon Hai Research Institute $^{4}$
+
+r10946021@ntu.edu.tw cyaochen@cs.unc.edu hungtungming@gmail.com
+
+txh710@psu.edu yunghui.li@foxconn.com ibin.liao@foxconn.com
+
+lwku@iis.sinica.edu shaohuas@ntu.edu.tw
+
+# Abstract
+
+This work introduces a novel task, location-aware visual question generation (LocaVQG), which aims to generate engaging questions from data relevant to a particular geographical location. Specifically, we represent such location-aware information with surrounding images and a GPS coordinate. To tackle this task, we present a dataset generation pipeline that leverages GPT-4 to produce diverse and sophisticated questions. Then, we aim to learn a lightweight model that can address the LocaVQG task and fit on an edge device, such as a mobile phone. To this end, we propose a method which can reliably generate engaging questions from location-aware information. Our proposed method outperforms baselines regarding human evaluation (e.g., engagement, grounding, coherence) and automatic evaluation metrics (e.g., BERTScore, ROUGE-2). Moreover, we conduct extensive ablation studies to justify our proposed techniques for generating the dataset and solving the task.
+
+# 1 Introduction
+
+Driving is an integral part of our daily routines, playing a significant role in our lives. Whether commuting to work, running errands, or embarking on exciting adventures, we heavily rely on automobiles to get us from one place to another. Despite its undeniable convenience, driving requires constant focus on the road, the need to remain alert, and the mental strain of navigating through traffic. Hence, staying behind the wheel after long working hours or during a long-distance trip can give rise to hazardous circumstances. To combat this, passengers often engage in conversation to keep the driver awake and attentive.
+
+Can we develop a system running on a lightweight device that automatically engages the
+
+# Location-aware Information
+
+![](images/593325c224bd8881c504e47d4cc5f5a95d234e20d136289509bbdb67f4df0a22.jpg)
+
+![](images/d1616b244cde0bd2e93d9dd213b4cb19fc03d5c94aa45d3637467be7c60ac451.jpg)  
+Figure 1: Location-aware Visual Question Generation (LocaVQG) involves generating engaging questions from a specified location, represented by a GPS coordinate of a vehicle and a set of street view images captured by on-car cameras.
+
+driver in a conversation? While initiating a conversation with general questions may not interest the driver, delving into driver-specific inquiries raises privacy concerns since it requires personal information. Our key insight is to engage the driver in a conversation by posing questions based on the location-aware information, composed of both the geographical coordinate of the car and surrounding visual perception represented by pictures captured by on-car cameras. Such rich location-aware information allows for producing diverse and relevant questions, enabling a system to initiate an engaging conversation.
+
+In this work, we introduce a novel task, Location-aware Visual Question Generation (LocaVQG), which aims to produce engaging questions from a GPS coordinate of a vehicle and a set of street-view images captured by on-car cameras, as illustrated in Figure 1. We make the first attempt to tackle this task by developing a data generation pipeline that can create a dataset containing high-quality samples for the LocaVQG task. To this end, we leverage the recent advances in large language models (LLMs) (Liu et al., 2023; Touvron et al., 2023).
+
+Specifically, we collect data from Google Street View and design a prompt according to the address obtained by reverse geocoding the GPS coordinate and the captions of street-view images provided by an off-the-shelf image captioning model. While LLMs can generate questions relevant to the provided location-aware information, the produced questions may not always be engaging. Therefore, we further propose to train an engaging question classifier that can filter out non-engaging questions. Our proposed dataset generation pipeline is illustrated in Figure 2.
+
+We present a method, FDT5, that can learn a lightweight model and reliably address the LocaVQG task. We compare our proposed method to various small and mid-size language models learning from the generated dataset. The experimental results demonstrate that our proposed FDT5 outperforms the baselines regarding human evaluation (e.g., engagement, coherence, grounding) and automatic evaluation metrics, e.g., BERTScore (Zhang et al., 2019), ROUGE-2 (Lin, 2004). Our FDT5 with only 15M parameters achieves competitive performance even compared to a large language model (i.e., GPT-4). This highlights the effectiveness of the proposed dataset generation pipeline as well as the proposed training techniques.
+
+The main contributions of this work are three-fold as follows:
+
+- Task. We propose Location-aware Visual Question Generation (LocaVQG), a novel task that aims to produce engaging questions from a GPS coordinate of a vehicle and a set of street-view images captured by on-car cameras. This will lead to the development of more intelligent in-car assistant systems.  
+- Dataset. To address LocaVQG, we introduce a dataset generation pipeline that can produce diverse and engaging questions from a specified location by leveraging pre-trained LLMs.  
+- Method. We present a method FDT5 that outperforms all the lightweight baselines regarding human evaluation (e.g., engagement, coherence, grounding) and automatic evaluation metrics (e.g., BERTScore, ROUGE-2).
+
+# 2 Related Works
+
+Self-driving cars. Despite the recent advances in developing self-driving cars (Parekh et al., 2022),
+
+most current commercialized autonomous vehicles are categorized as SAE (the Society of Automotive Engineers) (International, 2018) Level 2 (e.g., Tesla, Hyundai, Kia) or Level 3 (e.g., Mercedes). When driving an SAE Level 2 vehicle, the driver must always hold the steering wheel. With an SAE Level 3 vehicle, the driver must still be ready to take control of the vehicle at all times when prompted by the vehicle. That being said, driving in modern days still requires the driver's attention and therefore can be assisted with the task and the system proposed in this work.
+
+In-car intelligent assistant system. Developing in-car intelligent assistant systems, such as a voice assistant (Lin et al., 2018; Braun et al., 2019), is an emerging research area. Large et al. (2017) discovered that engaging drivers in a conversation could effectively reduce driver fatigue. In contrast, this work focuses on raising drivers' attention by formulating a task and devising a system to produce location-aware engaging questions.
+
+Visual question generation (VQG). VQG concerns generating questions from visual inputs (Mostafazadeh et al., 2016). Compared to this work, recent works (Lu et al., 2021; Yeh et al., 2022) that explore VQG do not leverage geographical information (e.g., GPS). On the other hand, Zhang and Choi (2021) presented a dataset with questions respecting geographical and temporal contexts; yet, it does not utilize visual inputs. In contrast, the task and the dataset proposed in this work leverage images captured by on-car cameras.
+
+Large language models (LLMs). Recent advances in LLMs (Liu et al., 2023; Touvron et al., 2023) have led to promising results in various domains (Wei et al., 2022; Zhang et al., 2023). However, these gigantic LLMs with billions of parameters (Liu et al., 2023; Touvron et al., 2023) cannot be deployed on lightweight devices and therefore are not ideal for in-car intelligent assistant systems. This work aims to develop lightweight models that can run on edge devices like mobile phones.
+
+Lightweight language models. Existing mobile-friendly language models (Sun et al., 2020; Mehta and Rastegari, 2021) struggle at language generation tasks. This work aims to devise lightweight models that can address VQG tasks and achieve competitive results even compared to LLMs.
+
+![](images/c1b883ef925a3dacd5a96892853e9b5510eb963a90e90719957083d8f7fd7729.jpg)  
+Figure 2: Data Generation Pipeline. This pipeline produces questions from a given LocaVQG task tuple consisting of four street view images  $V_{N}, V_{E}, V_{S}, V_{W}$  and a GPS coordinate  $X$ . We caption each image using an image captioning model and infer the address by reverse geocoding the GPS coordinate. Then, we construct a prompt that describes in detail the location-aware information and leverage GPT-4 to generate questions. We further employ an engaging question classifier to filter out non-engaging questions. Finally, the remained questions are included in the dataset with the given LocaVQG task tuple.
+
+# 3 LocaVQG: Location-aware Visual Question Generation
+
+We introduce a novel task, Location-aware Visual Question Generation (LocaVQG). This section formally defines this task and describes how we collect data to construct the LocaVQG task tuples.
+
+# 3.1 Location-aware Information
+
+Location-aware Information includes data or content specifically relevant to, or influenced by a particular geographical location. With such information, applications can offer location-specific recommendations, directions, local weather updates, nearby points of interest, targeted advertisements, etc. Since our goal is to produce engaging questions with an in-car device based on location-aware information, we limit it to the information that is easily accessible even without the internet. Specifically, we consider the surrounding visual perception and the geographical coordinate.
+
+# 3.2 LocaVQG Task Tuple
+
+To collect the surrounding visual perception and the geographical coordinate of diverse locations, we propose to leverage Google Street View Dataset (Zamir and Shah, 2014). The dataset contains 10,343 coordinates, and each coordinate comes with 5 corresponding directional images (North, East, South, West, and Upper/Sky view).
+
+To ensure the location diversity, we select 3,759 coordinates with their 4 directional images, excluding the upper/sky view, which is usually not observable by the driver. We denote the geographical coordinate of each location as  $X$  and its surrounding images as  $V_{d}$  with  $d = [N,E,S,W]$ , standing for each direction. We define our LocaVQG task tuple  $T$  as:  $T = [V_N,V_E,V_S,V_W,X]$ . Given a LocaVQG task tuple  $T$ , our goal is to produce an engaging question  $Q$  with a model  $f$ :  $f(T) = Q$ .
+
+# 4 Generating LocaVQG Dataset
+
+Our goal is to train lightweight models to address the LocaVQG task. Therefore, we aim to "label" the task tuples described in Section 3. Annotating the task tuples with engaging questions requires creativity and location-specific domain knowledge, which can be challenging for human annotators. In this work, we propose automatically generating questions from LocaVQG task tuples by leveraging the recent development of LLMs. An overview of the proposed dataset generation pipeline is depicted in Figure 2.
+
+# 4.1 Prompting GPT-4
+
+This section describes how we utilize GPT-4 (OpenAI, 2023) to produce questions from LocaVQG task tuples by processing task tuples and designing LocaVQG prompts.
+
+<table><tr><td>Engaging Questions</td></tr><tr><td>The city of Pittsburgh is known for its numerous bridges. How many bridges do you think are in the city, and why do you think there are so many?What types of events or festivals might take place in this park throughout the year?As we look towards the south, can you guess the purpose of this brick building with cars parked in front? Perhaps an office building, a restaurant, or something else?</td></tr><tr><td></td></tr><tr><td>Non-Engaging Questions</td></tr><tr><td>Speaking of the hospital, does anyone know the range of medical services provided at Prince George’s Hospital?What are some ways that city planners might improve traffic flow at busy intersections?Noticing the mixture of architectural styles, can you guess which era had the most significant influence on the city’s development?</td></tr></table>
+
+Street view images  $\rightarrow$  captions. While GPT-4 is a multimodal model, its feature of taking image inputs is not yet publicly accessible as of May 2023. Hence, to inform GPT-4 with the street view images, we caption street view images using an off-the-shelf image captioning model (Wang et al., 2022).
+
+GPS coordinate  $\rightarrow$  address. To leverage the GPS coordinate, we reverse geocode it using Google's Reverse Geocoding API (Google), translating the coordinate into a street address. We found that with the decoded street address, GPT-4 can often infer nearby famous landmarks, or information, and generate relevant questions.
+
+Constructing prompts. We aim to prompt GPT-4 with the processed location-aware input and produce engaging questions. We first design a system prompt that infuses GPT-4 with a tour guide role, enforcing it to engage users. Then, we design a chat prompt that encapsulates processed location-aware information and requires GPT-4 to generate engaging questions. The two prompts are presented as follows.
+
+- System prompt: You are a tour guide and you are driving in a car with your tourists. You want to engage with them with any kind of information you have around you.  
+- Chat prompt: Here are some descriptions of your surroundings. You are currently driving on [Street Address]. On your North, [Image Caption]. On your East, [Image Caption]. On your South, [Image Caption]. On your West, [Image Caption]. Based on those descriptions, please ask 10 engaging questions.
+
+# 4.2 Filtering GPT-Generated Questions
+
+While GPT-4 can generate numerous diverse questions from our designed prompts, we empirically find that some generated questions are not particularly engaging (e.g., requiring domain knowledge), as shown in Table 1. To combat this, we propose to learn a BERT-based (Devlin et al., 2019) engaging question classifier to filter out non-engaging questions. We construct the training data for this classifier with non-engaging questions from SQuaD (Rajpurkar et al., 2016) and engaging questions from MVQG (Yeh et al., 2022). The key insight is that SQuaD questions are made for question-answering tasks, thus, solely revolves around a passage, while MVQG questions are collected with engagement in mind.
+
+With this trained engaging question classifier, for each LocaVQG task tuple  $T$ , we filter out non-engaging questions generated by GPT-4, and the remained questions are included in the dataset as the "labels" for this task tuple.
+
+# 4.3 Dataset Statistics
+
+Applying the procedures described in Section 4.1 and Section 4.2 results in a dataset with 3759 task tuples and 35K questions. The basic statistics of the dataset are described in Table 2.
+
+Table 1: GPT-Generated Questions. We provide examples of GPT-4 generated questions that are classified as engaging and non-engaging by the engaging question classifier. Answering these non-engaging questions often requires specific domain knowledge and therefore may interest only limited audience. blue-colored text indicate visual cues, red-colored text indicate directional cues, teal-colored text indicate location-specific information.  
+
+<table><tr><td># of LocaVQG Task Tuples</td><td>3759</td></tr><tr><td>- # of Task Tuples from Pittsburgh</td><td>919</td></tr><tr><td>- # of Task Tuples from Orlando</td><td>611</td></tr><tr><td>- # of Task Tuples from New York</td><td>2217</td></tr><tr><td># of Questions After Filtering</td><td>35551</td></tr><tr><td>Average Sentence Length</td><td>16.6</td></tr><tr><td>Average Question Length</td><td>30.8</td></tr></table>
+
+Table 2: Dataset Statistics. We present the statistics of our location-aware visual question generation dataset
+
+![](images/d5350164ac4bc0f5a443f05525c1177db8f35cb02a32933f378c1271c17f1756.jpg)  
+Figure 3: Question Length. We present the question length statistics in # and tokens and # of sentences.
+
+![](images/937fac3ed27379a02a74c139e1869131998a71788f1e6b2c2c708185f106975a.jpg)
+
+# 4.3.1 Question Length
+
+We present the histograms of question lengths in terms of # of tokens and # of sentences in Figure 3.
+
+# 4.3.2 Frequent Trigrams and Words
+
+We present the top 15 frequent trigrams in the dataset in Table 3. The questions often start by trying to intrigue the respondent (e.g., Did you know, What do you). Also, open-ended questions (e.g., have you noticed, have you ever) appear quite frequently. Most frequent words, presented in Table 4, require or lead the attention of the respondent (e.g., considering, looking, notice).
+
+<table><tr><td>Did you know</td><td>What do you</td><td>Can you spot</td></tr><tr><td>Can you guess</td><td>As we drive</td><td>How do you</td></tr><tr><td>Have any of</td><td>Have you noticed</td><td>Can anyone guess</td></tr><tr><td>Have you ever</td><td>Can you identify</td><td>What are your</td></tr><tr><td>Are you familiar</td><td>What are some</td><td>As we continue</td></tr></table>
+
+Table 3: Top 15 Frequent Trigram of Questions. The frequent trigrams appearing in the dataset show that the questions aim to intrigue or engage the respondent.  
+
+<table><tr><td>Speaking</td><td>Considering</td><td>Based</td><td>Looking</td><td>Notice</td></tr><tr><td>Since</td><td>Perhaps</td><td>Observing</td><td>Residential</td><td>New</td></tr><tr><td>Look</td><td>Pittsburgh</td><td>Given</td><td>Orlando</td><td>See</td></tr><tr><td>Let</td><td>Turning</td><td>Judging</td><td>Noticing</td><td>Besides</td></tr></table>
+
+# 4.3.3 Question Quality
+
+We compare our proposed dataset to an engaging question dataset, the MVQG dataset (Yeh et al., 2022), regarding the following criteria and report the results in Table 5.
+
+- Vocabulary Size (Vocab Size) measures the number of distinct words in a dataset.
+
+- Average Sentence Length (Avg Sent. Length) computes the average length of sentences across the whole dataset, representing how rich and detailed a dataset is.  
+- Syntactic Complexity calculates the degree of variation, sophistication, and elaboration of the questions in a dataset (Ferraro et al., 2015). We report the mean of Yngve Score normalized by the sentence length.  
+- Percentage of Abstract Terms (% Abstract Term) computes the ratio of visual and nonvisual concepts covered by a dataset, based on the abstract terms defined by Vanderwende et al. (2015).  
+- Average Term Depth is calculated based on WordNet, where noun words with a lower depth indicate higher-level concepts (Lu et al., 2021).
+
+Compared to the MVQG dataset, the results show that our proposed LocaVQG dataset contains significantly more diverse and sophisticated questions. In fact, the questions included in the MVQG dataset are collected from human annotators. This highlights the effectiveness of generating questions by leveraging the recent advances in LLMs (e.g., GPT-4), as proposed in this work. Further evaluations on the generated questions can be found in Section D and Section E.
+
+Table 4: Top 20 Frequent Words. This table presents 20 frequent words appearing in the dataset aside from commonly used words, such as "have", "can".  
+
+<table><tr><td>Criteria</td><td>MVQG</td><td>LocaVQG (Ours)</td></tr><tr><td>Vocabulary Size ↑</td><td>608</td><td>3046</td></tr><tr><td>Average Sentence Length ↑</td><td>12.341</td><td>17.168</td></tr><tr><td>Yngve Score ↑</td><td>2.271</td><td>3.761</td></tr><tr><td>% Abstract Terms ↑</td><td>0.127</td><td>0.167</td></tr><tr><td>Average Term Depth ↓</td><td>7.906</td><td>7.259</td></tr></table>
+
+Table 5: Question Quality Comparison with MVQG.
+
+# 5 Learning and Evaluating Lightweight Models
+
+We aim to train and evaluate lightweight models learning from the proposed LocaVQG dataset.
+
+# 5.1 Baselines
+
+We compare our method to the following baselines. Text-To-Text Transfer Transformer (T5). We experiment with a family of T5 pre-trained language models (Raffel et al., 2020; Tay et al., 2021), which includes T5-Large (770M), T5-Base (220M),
+
+<table><tr><td>Model</td><td>#Parameters</td><td>Engagement</td><td>Naturalness</td><td>Coherence</td><td>Common Sense</td><td>Grounding</td><td>Overall</td></tr><tr><td>MVQG-VL-T5</td><td>254M</td><td>3.84</td><td>3.64</td><td>3.65</td><td>3.81</td><td>3.84</td><td>3.76</td></tr><tr><td>MVQG-VL-T5fine-tuned</td><td>254M</td><td>3.96</td><td>3.82</td><td>3.82</td><td>3.99</td><td>3.66</td><td>3.85</td></tr><tr><td>T5-Large</td><td>770M</td><td>3.92</td><td>3.81</td><td>3.78</td><td>4.03</td><td>3.83</td><td>3.87</td></tr><tr><td>T5-Base</td><td>220M</td><td>3.92</td><td>3.81</td><td>3.73</td><td>3.97</td><td>3.78</td><td>3.84</td></tr><tr><td>T5-Tiny</td><td>15.6M</td><td>3.96</td><td>3.79</td><td>3.67</td><td>4.01</td><td>3.81</td><td>3.85</td></tr><tr><td>FDT5 (Ours)</td><td>15.6M</td><td>4.03</td><td>3.83</td><td>3.96</td><td>4.05</td><td>4.03</td><td>3.98</td></tr><tr><td>GPT-4</td><td>1T*</td><td>4.12</td><td>3.99</td><td>4.01</td><td>4.05</td><td>4.01</td><td>4.04</td></tr><tr><td>Human Annotator</td><td>-</td><td>4.06</td><td>3.87</td><td>3.90</td><td>4.06</td><td>3.88</td><td>3.95</td></tr></table>
+
+Table 6: Human Evaluations. Each question is rated by three AMT workers. Among all the light-weight models, our proposed FDT5 achieves the best overall performance and has the fewest parameters. Note that while the exact number of parameters GPT-4 is not revealed, many believe it is at least 6 times larger than GPT-3 (Brown et al., 2020) (175B).
+
+and T5-Tiny (15.6M). We fine-tune the pre-trained T5 models on our LocaVQG dataset. Specifically, for each LocaVQG task, the models learn to map the prompt presented in Section 4.1 to one of the ground truth questions generated by GPT-4.
+
+MVQG-VL-T5. Cho et al. (2021) introduced Vision-and-Language T5 (VLT5) for vision-and-language tasks. Yeh et al. (2022) adapted it for generating questions from a set of images. We adopt this method, dubbed MVQG-VL-T5, and fine-tune the pre-trained model on our LocaVQG dataset. The input of MVQG-VL-T5 consists of 4 street view images and the street address.
+
+More details can be found in Section F.
+
+# 5.2 Our Approach
+
+We propose Filtered Distilled T5-Tiny (FDT5).
+
+Distillation. While T5-Tiny has the fewest of parameters and can fit on mobile phones, its capacity might be limited to a complex task like LoCAVQG. Therefore, we propose to learn a T5-Tiny model by distilling a learned T5-Large model. Inspired by Chen et al. (2020), during training, we utilize both the questions generated by GPT-4 from the dataset and the questions generated by the T5-Large model, resulting in the objective:
+
+$$
+\mathcal {L} (\theta) = \alpha \cdot \mathcal {L} _ {h a r d} (\theta) + (1 - \alpha) \cdot \mathcal {L} _ {s o f t} (\theta), \quad (1)
+$$
+
+where  $\alpha$  balances the relative importance of learning from each loss and  $\theta$  parameterizes the model. The hard-label loss  $\mathcal{L}_{\text {hard }}$  (ground truth target) optimizes cross-entropy, while the soft-label loss  $\mathcal{L}_{\text {soft }}$  (teacher model) optimizes KL Divergence.
+
+Filtering. To further improve the engagingness of the questions produced by our method, we propose to utilize the engaging questions classifier described in Section 4.2 to filter out non-engaging
+
+questions. Specifically, given a LocaVQG task, our method keeps generating questions until accepted (i.e., classified as "engaging") by the classifier.
+
+Our proposed method FDT5 combines the two techniques described above.
+
+# 5.3 Human Evaluation
+
+We provide human evaluation of the questions generated by all the methods.
+
+# 5.3.1 Evaluation Metrics
+
+We randomly sampled 100 LocaVQG task tuples and the questions produced by all the models. Each question is evaluated by three Amazon Mechanical Turk (AMT) workers according to the following metrics. We adopt a 5-point Likert scale for all the evaluation metrics.
+
+- Engagement: You find the question engaging and you would want to answer the question.  
+- Naturalness: It is natural to ask this question given the information you have.  
+- Coherence: The question is coherent with the information you have.  
+Common Sense: It makes sense to ask these questions given the information you have.  
+- Grounding: The question asked about things related to the information you have.
+
+We also provide the evaluation of the questions generated by GPT-4, which can be considered as an upper bound as GPT-4 has an unparalleled number of parameters compared to the lightweight models. Furthermore, to compare the performance of these LMs against humans, we crowdsourced and collect 100 questions on AMT based on the same set of
+
+<table><tr><td>Model</td><td>Engagement</td><td>Naturalness</td><td>Coherence</td><td>Common Sense</td><td>Grounding</td><td>Overall</td></tr><tr><td>Filtered Dataset (Ours)</td><td>3.92</td><td>3.81</td><td>3.73</td><td>3.97</td><td>3.78</td><td>3.84</td></tr><tr><td>Unfiltered Dataset</td><td>3.89</td><td>3.76</td><td>3.78</td><td>3.85</td><td>3.78</td><td>3.81</td></tr></table>
+
+Table 8: Engaging Question Classifier for Dataset Generation. Employing the engaging question classifier in the dataset generation process to filter out unengaging questions improves the quality of generated questions.  
+
+<table><tr><td>Model</td><td>Engagement</td><td>Naturalness</td><td>Coherence</td><td>Common Sense</td><td>Grounding</td><td>Overall</td></tr><tr><td>Filtered Inference (Ours)</td><td>4.03</td><td>3.83</td><td>3.96</td><td>4.05</td><td>4.03</td><td>3.98</td></tr><tr><td>Unfiltered Inference</td><td>3.96</td><td>3.78</td><td>3.82</td><td>4.04</td><td>3.79</td><td>3.88</td></tr></table>
+
+Table 9: Engaging Question Classifier for Inference. Our proposed FDT5 employs the classifier during inference to filter out unengaging questions. Excluding the filtering phase results in significantly worse performance.
+
+LocaVQG task tuples. More details on AMT can be found in Section G.
+
+# 5.3.2 Results
+
+The human evaluations are presented in Table 6.
+
+FDT5 outperforms all the lightweight models. Our proposed method FDT5 achieves the best overall score with the fewest parameters. This justifies the effectiveness of our adopted distillation scheme. Furthermore, an average score of 3.98 indicates that our model can reliably generate satisfactory questions from location-aware information.
+
+MVQG-VL-T5. MVQG-VL-T5, without learning from our dataset, achieves the worst performance, demonstrating the importance of constructing and learning from a dataset dedicated to the LocaVQG task. Alternatively, the MVQG-VL-T5 model fine-tuned on our dataset (MVQG-VL-T5fine-tuned) struggles at grounding, aligning with the findings discussed in (Yeh et al., 2022).
+
+GPT-4 asks better questions than humans. The questions produced by GPT-4 are preferred by the workers compared to those provided by human annotators on all the metrics, except for common sense. This justifies our proposed dataset generation pipeline, which collects questions from GPT-4 instead of humans.
+
+5.4 Automatic Evaluation Metrics  
+
+<table><tr><td>Model</td><td>BLEU-4</td><td>ROUGE-2</td><td>BERTScore</td><td>BLEURT</td></tr><tr><td>VLT5</td><td>0.2712</td><td>0.0342</td><td>0.5093</td><td>-0.7208</td></tr><tr><td>T5-Large</td><td>0.2756</td><td>0.0380</td><td>0.5165</td><td>-0.7336</td></tr><tr><td>T5-Base</td><td>0.2746</td><td>0.0388</td><td>0.5163</td><td>-0.7305</td></tr><tr><td>T5-Tiny</td><td>0.2635</td><td>0.0371</td><td>0.5164</td><td>-0.7419</td></tr><tr><td>FDT5 (Ours)</td><td>0.2661</td><td>0.0393</td><td>0.5190</td><td>-0.7073</td></tr></table>
+
+Table 7: Automatic Evaluation. Our proposed FDT5 achieves the best performance on 3 out of 4 metrics (i.e., ROUGE-2, BERTScore, and BLEURT).
+
+We further evaluate the questions generated by all the models with some automatic evaluation metrics. To compare two questions based on exact wording, we are using BLEU-4 (Papineni et al., 2002), ROUGE-2 (Lin, 2004). To compare the questions based on semantic similarity, we are using ML-Based evaluation: BERTScore (Zhang et al., 2019), and BLEURT (Sellam et al., 2020). The results are presented in Table 7. Our proposed method FDT5 achieves the best performance on ROUGE-2, BERTScore, and BLEURT. T5-Large outperforms others in terms of BLEU-4. This verifies the effectiveness of the filtering and distillation technique employed in FDT5.
+
+# 5.5 Ablation Study
+
+We conduct extensive ablation studies to investigate the effectiveness of employing the filtering classifier (Section 5.5.1), the impact of incorporating GPS coordinates into the question generation process (Section 5.5.2), and the effect of varying dataset sizes (Section 5.5.3).
+
+# 5.5.1 Employing Engaging Question Classifier
+
+We propose to learn an engaging question classifier for (1) filtering out non-engaging questions generated by GPT-4 during the dataset generation process (Section 4.2), and (2) filtering out non-engaging questions produced by our model during inference (Section 5.2). This section examines the effect of employing this classifier.
+
+Dataset Generation. To verify the effectiveness of filtering out questions from GPT-4 with the classifier, we train a T5-Base model to learn from an unfiltered dataset that contains all the questions produced by GPT-4 (Unfiltered Dataset). We compare the performance of this model to the T5-Base model trained on our proposed filtered dataset (Filtered Dataset) and report the human evaluation re
+
+![](images/d9158c215485d3bd4f995bddd77313d217ddda35f75215593aca1e687a05d8ea.jpg)  
+Figure 4: Qualitative Results. We present sampled generated questions from T5-Large, T5-Tiny, our proposed method FDT5, and human annotators, together with corresponding streetview images and addresses. With only 15M parameters, FDT5 can reliably generate engaging location-aware questions.
+
+sults in Table 8. The results demonstrate that the model learning from the filtered dataset achieves better performance, justifying the efficacy of employing the classifier.
+
+Inference. We propose to filter out non-engaging questions generated during inference, adopted in our method FDT5. We conduct human evaluations on filtered generation questions (Filtered Inference) and unfiltered questions (Unfiltered Inference), reported in Table 9. The results show that filtering non-engaging generated questions with the classifier can significantly improve the question quality on all the metrics. This justifies the effectiveness of employing the classifier during inference.
+
+# 5.5.2 Incorporating GPS Coordinates
+
+While Yeh et al. (2022) explored generating questions from a set of images, our work further incorporates addresses (reverse geocoded from GPS coordinates) into the question generation process. This section investigates the effect of employing such information. We compare the questions generated by GPT-4 with or without the address in the prompt and report the results in Table 10. The results show that incorporating the address leads to richer and more diverse questions, verifying the unique value of the proposed LocaVQG task.
+
+<table><tr><td>Criteria</td><td>w/o address</td><td>w/ address (ours)</td></tr><tr><td>Vocabulary Size ↑</td><td>450</td><td>525</td></tr><tr><td>Average Question Length ↑</td><td>25.02</td><td>30.18</td></tr><tr><td>Yngve Score ↑</td><td>3.531</td><td>3.693</td></tr></table>
+
+# 5.5.3 Varying Dataset Sizes
+
+We investigate the impact of varying dataset sizes with T5-Tiny and our proposed FDT5, and report the results in Table 11. FDT5 achieves better performance with fewer data points and performs comparably to T5-Tiny when dataset size increases. This indicates that our method is more data efficient.
+
+Table 10: Effect of Leveraging Street Address The questions generated with street addresses are richer and more diverse.  
+
+<table><tr><td>Model</td><td>#Samples</td><td>BLEU-4</td><td>ROUGE-2</td><td>BERTScore</td><td>BLEURT</td></tr><tr><td rowspan="5">T5-Tiny</td><td>0.7K</td><td>0.2566</td><td>0.0366</td><td>0.5160</td><td>-0.7666</td></tr><tr><td>1.7K</td><td>0.2629</td><td>0.0341</td><td>0.5139</td><td>-0.7530</td></tr><tr><td>2.7k</td><td>0.2604</td><td>0.0374</td><td>0.5164</td><td>-0.7589</td></tr><tr><td>3.7K</td><td>0.2635</td><td>0.0371</td><td>0.5156</td><td>-0.7419</td></tr><tr><td>4.7K</td><td>0.2639</td><td>0.0361</td><td>0.5145</td><td>-0.7398</td></tr><tr><td rowspan="5">FDT5 (Ours)</td><td>0.7k</td><td>0.2565</td><td>0.0361</td><td>0.5201</td><td>-0.7149</td></tr><tr><td>1.7k</td><td>0.2700</td><td>0.0402</td><td>0.5214</td><td>-0.7245</td></tr><tr><td>2.7k</td><td>0.2675</td><td>0.0422</td><td>0.5211</td><td>-0.7126</td></tr><tr><td>3.7k</td><td>0.2661</td><td>0.0393</td><td>0.5190</td><td>-0.7073</td></tr><tr><td>4.7k</td><td>0.2706</td><td>0.0386</td><td>0.5180</td><td>-0.7256</td></tr></table>
+
+Table 11: Effect of Dataset Size. From the results, FDT5 is more data efficient as it could achieve better performance with smaller sample size
+
+# 5.6 Qualitative Results
+
+As human evaluations can be subjective, we present qualitative results in Figure 4 for the readers to better understand the generated questions. The results show that FDT5 with only 15M parameters can reliably generate engaging location-aware questions.
+
+# 6 Conclusion
+
+In this work, we propose a novel task, location-aware visual question generation (LocaVQG), which aims to generate engaging questions from data relevant to a particular geographical location. Specifically, we represent location-aware information using four directional street view images and a GPS coordinate. To address this task, we introduce a dataset generation pipeline that leverages the recent advances of large language models (i.e.,
+
+GPT-4) to generate diverse and sophisticated questions. To ensure the engagingness of the questions produced by GPT-4, we employ an engaging question classifier to filter out non-engaging questions. Our proposed dataset contains richer and various questions compared to existing datasets.
+
+To learn from the proposed LocaVQG dataset with lightweight models, we present Filtered Distilled T5-Tiny (FDT5) method. We extensively evaluate the proposed method and the baselines with human evaluation and automatic evaluation metrics. Our proposed FDT5, with the fewest parameters, demonstrates superior performance on most metrics. We conduct extensive ablation studies to verify the effect of employing the filtering classifier, the effectiveness of incorporating GPS coordinates into the question generation process, and the impact of varying dataset sizes. We hope this work will encourage researchers to explore the LocaVQG task and its applications.
+
+# Limitations
+
+We discuss the limitations and how we can potentially address them in this section.
+
+Biases in AMT workers. We notice that the AMT workers involved in human evaluation might be biased due to their demographic. This can potentially be addressed by ensuring the diversity of their background.
+
+Location-aware information. Aiming to develop an in-car intelligent assistant, this work proposes representing location-aware information as a GPS coordinate and a set of images captured by on-car cameras. Incorporating more detailed information, such as local news and weather, can potentially lead to more diverse and engaging questions, and is left for future research.
+
+Address-aware LLMs. Our proposed dataset generation pipeline heavily relies on GPT-4. This partially limits the generated questions to locations/addresses that are known by GPT-4 and therefore this pipeline might not produce coherent questions given locations that are less known by GPT-4. We can potentially address this by employing a more sophisticated external information retrieval system to extract information from a location
+
+Human evaluation setup. While our motivation is to develop an in-car intelligent system that can engage a driver in a conversation to keep the driver awake, this work falls short from the following perspectives. First, our work solely focuses on gen
+
+erating a question without considering continuing a conversation. Second, we evaluate the generated questions with an AMT interface where the AMT workers read and evaluate the questions. However, in a driving scenario, interacting with a virtual assistant by reading a question is impractical. Hence, evaluating the generated questions by connecting to a text-to-speech system and requiring the annotators to rate the questions by listening to them would align better with the application.
+
+Distractingness of generated questions. This work makes the very first attempt to develop an in-car visual question generation system that can ask engaging questions to initiate conversations with drivers. However, such engaging questions can potentially distract drivers and lead to dangerous situations in the worst case. To address such a concern, we encourage future works along this line to consider the "distractingness" of generated questions. In particular, developing evaluation metrics to determine if a question would distract a driver from the road and devising methods to produce engaging yet non-distracting questions are promising and interesting research directions.
+
+# Ethics Statement
+
+Since our proposed dataset generation pipeline involves collecting questions from GPT-4, the data inherits any biases of GPT-4. Moreover, our proposed method learns from the dataset, and therefore will also be biased. Therefore, inheriting biases can lead to generating inappropriate, sexist, racist questions or descriptions. Fortunately, addressing ethical issues has been an active research area (Liang et al., 2021; Baldini et al., 2021; Yan et al., 2023; Kasneci et al., 2023). We wish to incorporate the advances in the field to alleviate ethical concerns.
+
+# Acknowledgement
+
+This work is supported by the National Science and Technology Council of Taiwan under grants 111-2634-F-002-022- and by the Academia Sinica and Hong Hai Research Institute collaborative project 05T-1110523-1Q. Shao-Hua Sun was partially supported by the National Taiwan University and its Department of Electrical Engineering, Graduate Institute of Communication Engineering, and College of Electrical Engineering and Computer Science, and the Yushan Fellow Program by the Ministry of Education, Taiwan. We thank the online crowd workers for participating in the study.
+
+# References
+
+Ioana Baldini, Dennis Wei, Karthikeyan Natesan Ramamurthy, Mikhail Yurochkin, and Moninder Singh. 2021. Your fairness may vary: Pretrained language model fairness in toxic text classification. arXiv preprint arXiv:2108.01250.  
+Michael Braun, Anja Mainz, Ronee Chadowitz, Bastian Pfleging, and Florian Alt. 2019. At your service: Designing voice assistant personalities to improve automotive user interfaces. In CHI Conference on Human Factors in Computing Systems.  
+Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners.  
+Yen-Chun Chen, Zhe Gan, Yu Cheng, Jingzhou Liu, and Jingjing Liu. 2020. Distilling the knowledge of BERT for text generation. In ACL.  
+Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. 2021. Unifying vision-and-language tasks via text generation. In ICML.  
+Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. QLoRA: Efficient finetuning of quantized LLMs. arXiv preprint arXiv:2305.14314.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding.  
+Francis Ferraro, Nasrin Mostafazadeh, Ting-Hao Huang, Lucy Vanderwende, Jacob Devlin, Michel Galley, and Margaret Mitchell. 2015. A survey of current datasets for vision and language research. In EMNLP.  
+Google. Reverse geocoding.  
+HuggingFace. sentence-transformers/multi-qa-minilm16-cos-v1.  
+Sae International. 2018. Taxonomy and definitions for terms related to driving automation systems for on-road motor vehicles. SAE international.  
+Enkelejda Kasneci, Kathrin Seßler, Stefan Küchemann, Maria Bannert, Daryna Dementieva, Frank Fischer, Urs Gasser, Georg Groh, Stephan Gunnemann, Eyke Hüllermeier, et al. 2023. Chatgpt for good? on opportunities and challenges of large language models for education. Learning and Individual Differences.
+
+David Large, Gary Burnett, Vicki Antrobus, and Lee Skrypchuk. 2017. Stimulating conversation: Engaging drivers in natural language interactions with an autonomous digital driving assistant to counteract passive task-related fatigue. In International Conference on Driver Distraction and Inattention.  
+Paul Pu Liang, Chiyu Wu, Louis-Philippe Morency, and Ruslan Salakhutdinov. 2021. Towards understanding and mitigating social biases in language models. In ICML.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out. ACL.  
+Shih-Chieh Lin, Chang-Hong Hsu, Walter Talamonti, Yunqi Zhang, Steve Oney, Jason Mars, and Lingjia Tang. 2018. Adasa: A conversational in-vehicle digital assistant for advanced driver assistance features. In ACM Symposium on User Interface Software and Technology. Association for Computing Machinery.  
+Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, Zihao Wu, Dajiang Zhu, Xiang Li, Ning Qiang, Dingang Shen, Tianming Liu, and Bao Ge. 2023. Summary of chatgpt/gpt-4 research and perspective towards the future of large language models. arXiv preprint arXiv:2304.01852.  
+Zexin Lu, Keyang Ding, Yuji Zhang, Jing Li, Baolin Peng, and Lemao Liu. 2021. Engage the public: Poll question generation for social media posts. In ACLIJCNLP.  
+Sachin Mehta and Mohammad Rastegari. 2021. Mobilevit: light-weight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178.  
+Nasrin Mostafazadeh, Ishan Misra, Jacob Devlin, Margaret Mitchell, Xiaodong He, and Lucy Vanderwende. 2016. Generating natural questions about an image. In ACL.  
+OpenAI. 2023. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In ACL.  
+Darsh Parekh, Nishi Poddar, Aakash Rajpurkar, Manisha Chahal, Neeraj Kumar, Gyanendra Prasad Joshi, and Woong Cho. 2022. A review on autonomous vehicles: Progress, methods and challenges. *Electronics*.  
+Sarah Pratt, Mark Yatskar, Luca Weihs, Ali Farhadi, and Aniruddha Kembhavi. 2020. Grounded situation recognition. In ECCV.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits
+
+of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In EMNLP.  
+Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. 2015. Faster R-CNN: towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems.  
+Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. 2022. A-OKVQA: A benchmark for visual question answering using world knowledge. arXiv preprint arXiv:2206.01718.  
+Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. BLEURT: learning robust metrics for text generation. arXiv preprint arXiv:2004.04696.  
+Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou. 2020. Mobilebert: a compact task-agnostic bert for resource-limited devices. arXiv preprint arXiv:2004.02984.  
+Yi Tay, Mostafa Dehghani, Jinfeng Rao, William Fedus, Samira Abnar, Hyung Won Chung, Sharan Narang, Dani Yogatama, Ashish Vaswani, and Donald Metzler. 2021. Scale efficiently: Insights from pretraining and fine-tuning transformers. arXiv preprint arXiv:2109.10686.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Lucy Vanderwende, Arul Menezes, and Chris Quirk. 2015. An AMR parser for English, French, German, Spanish and Japanese and a new AMR-annotated corpus. In NAACL.  
+Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. Emergent abilities of large language models. TMLR.  
+Lixiang Yan, Lele Sha, Linxuan Zhao, Yuheng Li, Roberto Martinez-Maldonado, Guanliang Chen, Xinyu Li, Yueqiao Jin, and Dragan Gašević. 2023. Practical and ethical challenges of large language models in education: A systematic literature review. arXiv preprint arXiv:2303.13379.
+
+Min-Hsuan Yeh, Vincent Chen, Ting-Hao Huang, and Lun-Wei Ku. 2022. Multi-VQG: Generating engaging questions for multiple images. In EMNLP.  
+Amir Roshan Zamir and Mubarak Shah. 2014. Image geo-localization based on multiplenearest neighbor feature matching using generalized graphs. PAMI.  
+Jesse Zhang, Jiahui Zhang, Karl Pertsch, Ziyi Liu, Xiang Ren, Minsuk Chang, Shao-Hua Sun, and Joseph J Lim. 2023. Bootstrap your own skills: Learning to solve new tasks with large language model guidance. In Conference on Robot Learning.  
+Michael J.Q. Zhang and Eunsol Choi. 2021. SituatedQA: Incorporating extra-linguistic contexts into QA. EMNLP.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. BERTScore: Evaluating text generation with BERT. arXiv preprint arXiv:1904.09675.
+
+# A Additional Diversity Analyses on GPT-4 Generated Questions
+
+We perform further diversity analyses on the questions provided in our LocaVQG dataset.
+
+- Question type analysis: While Table 3 shows the top 15 most frequent trigrams of generated questions, we have performed an additional trigram analysis during the rebuttal period to examine the diversity of the generated questions. In particular, we followed Yeh et al. (2022) and identified 2437 question types among our  $35\mathrm{K}$  generated questions. This highlights the diversity of the generated questions.  
+- Pairwise cosine similarity: Inspired by Schwenk et al. (2022), which computes the average pairwise cosine similarity between each pair of generated questions encoded by a sentence transformer (multi-qa-MiniLM-L6-cos-v1 provided by HuggingFace) in a dataset, we have performed this evaluation on our generated dataset. We obtained an average cosine similarity of 0.1698, indicating that the generated questions are not highly correlated and therefore ensuring the diversity of our proposed dataset.
+
+# B Latency Analysis
+
+Our goal is to develop lightweight models that can run on mobile devices. To examine the applicability of our proposed model FDT5 and the baselines, we measure and report the latency of MVQG-VLT5, T5-Large, T5-Base, T5-Tiny, and our proposed FDT5 in Table 15. Each inference and post-filtering time is computed by averaging over 300 trials to reduce the variance.
+
+<table><tr><td>Latency (sec)</td><td>Loading Model</td><td>Inference</td><td>Post-Filtering</td></tr><tr><td>MVQG-VL-T5</td><td>7.09</td><td>6.38</td><td>N/A</td></tr><tr><td>T5-Large</td><td>12.79</td><td>10.04</td><td>N/A</td></tr><tr><td>T5-Base</td><td>10.34</td><td>5.9</td><td>N/A</td></tr><tr><td>T5-Tiny</td><td>3.89</td><td>2.02</td><td>N/A</td></tr><tr><td>FDT5</td><td>4.25</td><td>2.27</td><td>3.92</td></tr></table>
+
+Table 15: Latency Testing of the trained models.
+
+The results show that FDT5 and T5-Tiny, with the same model architecture and the same number of parameters, enjoy a significantly reduced time for loading models and running inference. The post-filtering phase of FDT5 takes 3.92 seconds
+
+on average, indicating that the engaging question classifier requires FDT5 to perform 1.73 additional inference trials for each LocaVQG task on average. Note that this post-filtering phase can be shut down for latency-critical scenarios, and FDT5 without post-filtering still outperforms T5-Tiny in human evaluation, according to Table 6 and Table 9.
+
+# C Filtering Out Non-engaging Questions Generated by GPT-4 Using GPT-4
+
+This work proposes to train an engaging question classifier to filter out non-engaging questions generated by GPT-4; alternatively, we can use GPT-4 to evaluate and filter out non-engaging questions that it generates. To investigate this possibility, we feed the questions generated by GPT-4 back into GPT-4 for scoring (i.e., determining if each generated question is engaging or not).
+
+Specifically, we provide GPT-4 with 10 questions generated by itself and asked it to determine if each question is engaging or not. The prompt and the response of GPT-4 are shown in Table 12. As a result, GPT-4 classifies all the questions it generated as engaging. This suggests that GPT-4 may not be useful to score its own generated questions, and motivates learning an external model to score the questions, such as our proposed engaging question classifier.
+
+Note that this finding aligns with the statement, "we observe that GPT-4 assigns significantly higher scores to its own outputs compared to human ratings," stated in Dettmers et al. (2023).
+
+# D Effect of Incorporating Address Information
+
+We are using the address as the main cue for GPT-4 to retrieve some information regarding those places. While it is possible to produce general, non-location-specific questions based on hand-crafted templates, we found that GPT-4 can certainly produce location-specific questions that require knowledge of specific locations, which can potentially be more engaging. Some examples are presented as follows.
+
+- Did you know that Fort Duquesne Boulevard is named after the historic Fort Duquesne, which was a key location during the French and Indian War? Have any of you studied that period in history?
+
+# Prompting GPT-4 for Self-Evaluation
+
+You will be given 10 questions. Please classify this question into engaging and non-engaging questions.
+
+1. Did you know that Pittsburgh is known as the City of Bridges due to its 446 bridges?  
+Can you spot any nearby bridges from this location?  
+2. The tall buildings on the south side are a part of Pittsburgh's beautiful skyline. Can anyone name some of the prominent buildings that make up the skyline?  
+3. As we pass by these parked cars on the north side, can you guess the most popular car brand in Pittsburgh based on what you see?  
+4. Pittsburgh has a rich history of steel production, earning it the nickname Steel City. Can you see any indications of this industrial past in the architecture of the buildings around us?  
+5. The traffic light at the empty intersection on our east might make you wonder about the city's traffic patterns. What do you think contributes to the current flow of traffic in this area?  
+6. Pittsburgh is known for its vibrant arts scene. Can anyone spot a nearby art gallery or theater as we drive along these city streets?  
+7. As we pass by the buildings on the west, let's talk about the city's diverse architectural styles. Can you identify different styles just by looking at their exteriors?  
+8. Pittsburgh is also famous for its rich sports culture, with successful teams in football, baseball, and hockey. Can you spot any sports fans or sports-themed establishments in our surroundings?  
+9. The city has a growing tech scene, with companies like Google and Uber opening offices here. Can you see any tech companies or startups as we drive through these streets?  
+10. Pittsburgh is known for its friendly locals, often called yinzers due to their unique accent. Have you encountered any friendly yinzers during your visit so far, and what has been your favorite interaction with them?
+
+# GPT-4 Response
+
+1. Engaging  
+2. Engaging  
+3. Engaging  
+4. Engaging  
+5. Engaging  
+6. Engaging  
+7. Engaging  
+8. Engaging  
+9. Engaging  
+10. Engaging
+
+Table 12: GPT-4 Self Evaluation Results.
+
+- The city of Pittsburgh is known for its numerous bridges. How many bridges do you think are in the city, and why do you think there are so many?  
+- The 59th Street Bridge, also known as the Ed Koch Queensboro Bridge, connects Manhattan to Queens. Can you identify any famous movies or TV shows that have featured this iconic bridge?
+
+We cannot obtain these questions by simply replacing the city name in other questions, nor should these questions be asked at a different location.
+
+Moreover, while generating a question with an address can increase the vocabulary size and average question length by simply inserting the address into the question, we still observe some questions that are generated based on the knowledge extracted by GPT-4 according to the address. We provide some example questions as follows.
+
+- Address: 1250 Penn Ave, Pittsburgh, PA 15222, USA
+
+- Generated question: "As we drive along 1250 Penn Ave, are there any upcoming events, festivals, or celebrations in the area that you'd like to learn more about?"  
+- Observation: GPT-4 knows this location has hosted several events in the past and therefore asks about upcoming events.
+
+- Address: 333 Boulevard of the Allies, Pittsburgh, PA 15222, USA
+
+- Generated question: Did you know that the Boulevard of the Allies is named to honor the Allies of World War I? What do you think about the significance of this historical connection?  
+- Observation: Based on knowing the history of the Boulevard of the Allies, GPT-4 asks about World War I.
+
+# E Importance of Incorporating Visual Input and Learning to Generate Questions
+
+Since it is possible to generate questions solely based on the fetched address, we aim to further analyze the effect of employing visual inputs to produce questions. Also, as discussed in Section D, we aim to quantitatively compare generating questions by our learned model and producing questions using general hand-crafted templates. To this end, we labeled 100 questions generated by FDT5 based on the following two criteria:
+
+- The generated questions contain visual information (w/ vis) or not (w/o vis)  
+- The generated questions are based on some.   
+templated (templated) or a learned language model (learned)
+
+Specifically, we went through each question and
+
+- Determined if it contains visual information (e.g., describing surroundings). If so, this question is labeled as w/ vis; otherwise, it is labeled as w/o vis.  
+- Decided if it can be generated based on some templates (e.g., the city's name can be replaced with another city and the question still makes sense). If so, this question is labeled as templated; otherwise, it is labeled as learned.
+
+Then, we analyze the engagement score and diversity of each group of questions. Regarding engagement scores, the questions containing visual information (w/ vis) achieve an average score of 4.007, slightly outperforming the questions without visual information (w/o vis) with an average score of 3.915, indicating that the visual-related questions may be more engaging. On the other hand, as the reviewer anticipated, the templated questions have a higher average score of 4.002 compared to learned questions (3.901).
+
+Then, we analyzed the diversity of each group of questions and found that the learned questions are more diverse (with a pairwise Cosine similarity score of 0.3614), outperforming the templated questions with a pairwise Cosine similarity score of 0.3995.
+
+In conclusion, we believe that generating engaging and diverse questions still requires incorporating visual inputs with a well-learned language model.
+
+# F Implementation Details
+
+# F.1 GPT-4 Parameters and Expenses
+
+Setup. When generating our proposed dataset using GPT-4, we use the model "gpt4" listed in the OpenAI API, with 0.7 temperature and 0.1 presence penalty.
+
+Expenses. On average, each request of our task uses up around 500 tokens, costing us around $0.001 USD. In total, generating the dataset and experimenting with it cost around $150 – 200 USD.
+
+# F.2 T5 and FDT5
+
+Input. Similar to GPT-4, the T5 models and FDT5 take image captions as input. The text input to the T5 models and FDT5 is modified from the chat prompt provided to GPT-4. Specifically, we pretend "generate questions:" prefix to each input, resulting in the model input as: generate questions: You are currently driving on [Street Address]. On your North, [Image Caption]. On your East, [Image Caption]. On your South, [Image Caption]. On your West, [Image Caption].
+
+Implementation. We adopt the basic pre-trained T5 models available on the Hugging Face platform. Training. During the training, we use 5 questions for each LocaVQG task tuple. We train each model for 20 epochs with a learning rate of  $10^{-4}$ .
+
+# F.3 VL-T5
+
+Input. The input of VL-T5 contains the following prompt prefix, street address, visual embeddings, and visual semantic groundings.
+
+- Prompt prefix. The prompt prefix is generate question:, which guides the model to generate questions with the instruction.  
+- Street address. The street address is the specific street address of the pictures that is verbalized, e.g., You are currently driving in Penn Avenue, Pittsburgh.  
+- Visual embeddings and visual semantic groundings. We extract the visual embeddings from the whole image and the image regions with Faster-RCNN (Ren et al., 2015). Also, we adopted the grounded situation recognition (GSR) (Pratt et al., 2020) model to extract information on the sequence of images to understand the semantics. Prefix of the directions of the images is also added, e.g., North: [Visual embeddings]
+
+Training. During the training, we used the pretrained baselines presented in (Yeh et al., 2022), specifically the adapter. We train the model for 30 epochs with a learning rate of  $10^{-5}$ . We also employ gradient accumulation steps of 4 and warm-up steps of 10.
+
+# F.4 Engaging Question Classifier
+
+The engaging question classifier is trained on the questions from the two datasets: SQuaD (20239 questions) and MVQG (31098 questions). The engaging question classifier is a BERT-based classifier with 110M parameters. We train the classifier to classify the questions sampled from SQuaD as non-engaging and those sampled from MVQG as engaging for 10 epochs. We use the ADAM optimizer with a learning rate of  $10^{-5}$ .
+
+Some example questions are as follows:
+
+- Why is that man playing billiards by himself? (Engaging)  
+- How did you celebrate your last birthday party? (Engaging)  
+- What document was signed in 1999? (Non-engaging)  
+- What did John Milton do for world literature? (Non-engaging)
+
+We report the performance of the learned engaging question classifier on the training (train), validation (val), and testing (test) sets in Table 14. The accuracy evaluates if the classifier can correctly distinguish the questions in the MVQG dataset from those in the SQuaD dataset. The results show that the trained engaging question classifier can accurately distinguish the questions from the two datasets.
+
+<table><tr><td></td><td>train</td><td>valid</td><td>test</td></tr><tr><td>Accuracy</td><td>99.9%</td><td>98.9%</td><td>99.0%</td></tr></table>
+
+Table 14: Performance of the Learned Engaging Question Classifier.
+
+# G Amazon Mechanical Turk Details
+
+# G.1 Human Evaluation Details
+
+Section 5.3 conducts human evaluation on AMT to compare questions generated by all models, GPT-4, and humans. For each generated question, each
+
+worker is provided with the street address and 4 streetview images of this location. Then, the worker is required to rate the generated question according to 5 evaluation metrics: Engagement, Naturalness, Coherence, Common Sense, and Grounding. The descriptions of the metrics are as follows.
+
+- Engagement This is an engaging question for this set of photos. You would want to answer or respond to this question  
+- Naturalness Given the pictures and location that you are in, it is natural to ask this question.  
+- Coherence This question asks something about the description and information that could be found in the image or relevant to the location.  
+- Common Sense This question provides enough Common Sense. The question asks something that makes sense according to our common sense.  
+- Grounding This question mentions the essential objects or information of the images, and it is mentioned in the right direction or talking about the location.
+
+This AMT interface is shown in Figure 5
+
+# G.2 Collecting Human Generated Questions
+
+In order to compare GPT-4 to humans regarding the ability to produce engaging questions from location-aware information, we crowdsourced human-generated questions from human annotators on AMT. We use three-stage questions to collect questions as follows.
+
+- Question 1. Pick 1 (or more) pictures that you want to focus on, and write down the object (or event) that stands out the most for you.  
+- Question 2. Please describe the object (or event).  
+- Question 3. Please write a question based on it.
+
+This AMT interface is shown in Figure 6. For each LocaVQG task tuple, we require three AMT workers to write a question, and then take the best question from the three questions. We present some questions produced by the workers in Table 14.
+
+<table><tr><td>There is a big strong building around us. Do you want to try come and visit it?</td></tr><tr><td>Can you give me some tips to how to drive as like you?</td></tr><tr><td>what do you think this building is used for?</td></tr><tr><td>where are your coming from and which place you are going to visit?</td></tr><tr><td>Look at the building around us, what do you think about the building? I find it looking really sturdy</td></tr><tr><td>On the east side, you could see a cargo with few people around it. What do you think is inside the cargo?</td></tr><tr><td>Have you ever wondered what is beyond that colossal vertical gateway adorned with glass windows of elegance? I wonder if it could be an art gallery or perhaps a stock exchange?It is Pittsburgh ya know, the Šteel City/the possibilities are endless!</td></tr><tr><td>What do you think about Garland Avenue compared to other streets in Orlando?</td></tr></table>
+
+Table 14: Human-Generated Questions.
+
+# Instruction
+
+In this HIT, we will show you  
+1. A sequence of 4 photos.  
+2. A question that you will be asked regarding these photos while driving
+
+Please rate the quality of these questions by 1 (the worst) to 5 (the best) according to the box's instructions.
+
+Some notes about the HIT
+
+1. It might take some time to load all images, please start rating only after all images have been loaded.  
+2. You can click on each photo to see the original image.
+
+![](images/7c6ebe9f00460af781497c920b89b19ae505054493cf8d7eb7dc19c075f6dbae.jpg)  
+Imagine you are driving in a car, and these are your streetview of North - East - South - West respectively  
+You are driving in 451 West St, New York, NY 10014, USA
+
+1. Given the location that you are in, it is natural to ask this question.
+
+# Notice the truck parked in front of the building to our north. Can anyone guess how many trucks are operating in a city like Pittsburgh?
+
+Rate the quality of the question by 1 (the worst) to 5 (the best) according to the above instructions.  
+Note: Natural questions are not about what it is seen, but rather about what is inferred given these objects in the images.  
+Ex: Given an image containing a gift, "Is the gift for the boy's birthday?" is natural, while "What is the color of the gift?" is unnatural.
+
+![](images/b867e2c31230e7dd2e71f3a9d9a8a1b2d21e94b6f64a0ea115f0a6ed2407624c.jpg)
+
+2. This question mentioned the essential objects of the images and it is mentioned in the right direction.
+
+# Notice the truck parked in front of the building to our north. Can anyone guess how many trucks are operating in a city like Pittsburgh?
+
+Rate the quality of the question by 1 (the worst) to 5 (the best) according to the above instructions.
+
+Note: Given a bus is in the north image, the bus is mentioned in the north.
+
+![](images/311b5ab92f02a124c41783763254b4b9adec4ae2fb5162a83d87c6acba3399d2.jpg)
+
+3. This question is coherent to the description and information that could be found in the image or relevant to the location.
+
+# Notice the truck parked in front of the building to our north. Can anyone guess how many trucks are operating in a city like Pittsburgh?
+
+Rate the quality of the question by 1 (the worst) to 5 (the best) according to the above instructions.  
+Ex: Given the description is: "Look at that bus on the north side",  
+"Have you ever tried that kind of transportation?" is coherent compared to "What do you think about the weather"
+
+![](images/f62cb13a367aa1f31eab204d6972439ddd294d824e3153a127e176f571ea66a1.jpg)
+
+4. This question provide enough Common Sense. The question ask something that make sense according to our common sense
+
+# Notice the truck parked in front of the building to our north. Can anyone guess how many trucks are operating in a city like Pittsburgh?
+
+Rate the quality of the question by 1 (the worst) to 5 (the best) according to the above instructions.
+
+![](images/1ebb902b12de211346dfb07ce6a251f052e5d2e052271f2c377187b906defef3.jpg)
+
+5. This is an engaging question for this set of photos. You would want to answer or respond to this question
+
+# Notice the truck parked in front of the building to our north. Can anyone guess how many trucks are operating in a city like Pittsburgh?
+
+Rate the quality of the question by 1 (the worst) to 5 (the best) according to the above instructions.
+
+![](images/04e730dea8c924ad2a71e0669303321896198514bb0008df6a30faabbcd60003.jpg)  
+You can submit in 27 seconds  
+Figure 5: Human Evaluation AMT Interface.
+
+# Instructions
+
+In this HIT, you will see a sequence of photos of your surrounding while driving (Streetview)
+
+The first photo is your north view, second photo is your east view, third is your south view, and fourth is your west view
+
+Please take a look at the photo sequence carefully and answer three questions.
+
+You can submit the HIT after 30 seconds!
+
+This is the requirement for the third query
+
+1. Natural Question: Natural questions are not about what is seen, but rather about what is inferred given these objects.
+
+- Prompt: given an image of a building with a large garage.  
+- Prompt: Given a building, how many times has it been inside the house?
+
+- Good Question Example: What do you think is inside the garage? - Bad Question Example: What is the color of the buildings?
+
+2. Engaging Question: Questions that sparks conversation, or pique other interests.
+
+o Prompt: given an image with a beautiful riverside park. The images are taken from street in Pittsburgh
+
+Use the Location information to make more engaging question!
+
+- Good Question Example: What do you think about riverside park in Pittsburgh compared to other places?
+
+3. Non-generic and not event-centric: Generic questions can appear in most circumstances
+
+Bad Question Example: What is that building? What's the weather? What color is that?
+
+Note: Please think and read the description of the box carefully before answering. If your answering time is too short, your answer may be rejected
+
+- It will take some time to load all images, please start answering only after all 4 images have been loaded.
+
+- You can click on each photo to see the origin
+
+Please look at this 4 images of streetview images:
+
+Note: From left to right, North - East - South - West
+
+![](images/a070aad2ab1f688b7e437697fe7be52f04462a2dd57e73112a57324644a8d891.jpg)
+
+![](images/45338635fd12aad87d0c186df224dc94e1845cf5160c32cbe4f10078e88654f8.jpg)
+
+![](images/e309e7775c5410cb8ace9a6f961c74b19618e1442669408ed369460deb11645a.jpg)
+
+![](images/c90b4d16d494946bf7c2ab97b8f9dda7f30d1fcb5f54c74f7406cd9af788a1a4.jpg)
+
+You are viewing images from 5th Avenue + Grant, 27th St, Pittsburgh, PA 15222, USA
+
+# Please answer the following 3 questions.
+
+Q1. Pick 1 (or more) pictures that you want to focus on, and write down the object (or event) that stands out the most for you
+
+Please write it down in the format of "Picture", "Object". Example: 4th Picture, Building
+
+![](images/56890634793b709a4cf2dfecd50501f09529d64602908105f512c3c2c5e03331.jpg)  
+Figure 6: Location-aware Question Collection AMT Interface.
+
+Q2. Please describe the object (or event) in 1 - 3 sentences
+
+![](images/3315b2627c7e7ed2a8a3670ae3e4b1ed634a42be1f3ecf9c25c66a03b285337e.jpg)
+
+Q3. Please ask question based on your description in Q2
+
+Scenario:
+
+Imag that you are in passers seat, and your driver is getting tired. What kind of question you would ask to engage your driver in a conversation (to relieve his boredom)?
+
+![](images/8bd0c2c03b618c1a0d47856a10f0f7952c81707221ad07643adf5122e0554991.jpg)
+
+If you like additional good and bad question examples to refer, please click here
+
+![](images/480247213e023062fb65a03a0c3bfcf45b5b463c7cd28362e0263a53a85d787d.jpg)

2023/Location-Aware Visual Question Generation with Lightweight Models/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ab3b7809a29b0c4551e3e7bdc98a0ac3a04119f850a7e696723a5391b7a4384
+size 738350

2023/Log-FGAER_ Logic-Guided Fine-Grained Address Entity Recognition from Multi-Turn Spoken Dialogue/4166b175-fae8-4a02-8116-9769fd80da5b_content_list.json

@@ -0,0 +1,1682 @@
+[
+    {
+        "type": "text",
+        "text": "Log-FGAER: Logic-Guided Fine-Grained Address Entity Recognition from Multi-Turn Spoken Dialogue",
+        "text_level": 1,
+        "bbox": [
+            134,
+            87,
+            867,
+            130
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Xue Han $^{1\\star}$ , Yitong Wang $^{1\\star}$ , Qian Hu $^{1}$ , Pengwei Hu $^{2}$ , Chao Deng $^{1}$ , Junlan Feng $^{1}$ $^{1}$ JIUTIAN Team, China Mobile Research Institute, Beijing, China",
+        "bbox": [
+            149,
+            141,
+            850,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "$^{2}$ The Xinjiang Technical Institute of Physics and Chemistry, Urumqi, China {hanxueai, wangyitongyjy, huqianai, dengchao, fengjunlan}@chinamobile.com hupengwei@hotmail.com",
+        "bbox": [
+            131,
+            175,
+            873,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            342,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Fine-grained address entity recognition (FGAER) from multi-turn spoken dialogues is particularly challenging. The major reason lies in that a full address is often formed through a conversation process. Different parts of an address are distributed through multiple turns of a dialogue with spoken noises. It is nontrivial to extract by turn and combine them. This challenge has not been well emphasized by mainstream entity extraction algorithms. To address this issue, we propose in this paper a logic-guided fine-grained address recognition method (Log-FGAER), where we formulate the address hierarchy relationship as the logic rule and softly apply it in a probabilistic manner to improve the accuracy of FGAER. In addition, we provide an ontology-based data augmentation methodology that employs ChatGPT to augment a spoken dialogue dataset with labeled address entities. Experiments are conducted using datasets generated by the proposed data augmentation technique and derived from real-world scenarios. The results of the experiment demonstrate the efficacy of our proposal.",
+        "bbox": [
+            142,
+            281,
+            460,
+            636
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            649,
+            260,
+            665
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Fine-grained address entity recognition from calls is an important task in many applications (Wu and Juang, 2022; Yang et al., 2022), such as obtaining delivery addresses from E-commerce assistant or post-sale service (Eligüzel et al., 2020). To be specific, the scenario considered in this paper is the extraction of fine-grained address entities distributed through the multi-turn spoken dialogue contexts. Figure 1 depicts an example of such a dialogue between a customer and an after-service call center in which the fine-grained address entities are recognized and then typically combined to form a full standard address for further delivery.",
+        "bbox": [
+            112,
+            675,
+            490,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Existing name entity recognition (NER) methods usually combine the pre-trained language models (e.g. BERT) with supervised models such as BiLSTM/LSTM-CRF (Xu et al., 2019; Zhang et al., 2020; Mai et al., 2018) or with some expansive external knowledge base (Dogan et al., 2019). These methods, which are trained and evaluated on text from well-written sources, do not perform well in a spoken dialogue context, because the entities in the spoken scenario are distributed across a multi-turn context with a lot of noise. Other studies explore the NER problem in a spoken or dialogue context (Muralidharan et al., 2021; Chen et al., 2022; Luoma and Pyysalo, 2020; Hanh et al., 2021), but their labels are not applicable to the address domain. Because of the diversity and complexity of multi-turn spoken dialogue scenarios, annotating for the FGAER task is laborious and time-consuming (Shan et al., 2020).",
+        "bbox": [
+            507,
+            253,
+            885,
+            558
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address these issues, we propose a logic-guided fine-grained address entity recognition method (Log-FGAER). The Log-FGAER is based on the fact that address entities follow spatially constrained relationships (Zandbergen, 2008). As a result, the multi-turn spoken dialogue must maintain a discourse structure to get the full standard address for the duration of address acquiring in accordance with the spatial constraint relationship between the geographical address entities. Figure 1(a) shows an example of a 4-level spatial dependency tree structure, in which each node is a fine-grained address entity, each level represents an entity type (e.g. \"city\" and \"district\"), and each edge is a dependency relation between the entities. A full standard address is made up of nodes in the path from the root to the leaf node. We can improve entity recognition inference by presenting such an interdependent hierarchy relationship as the logic rule and then using probabilistic soft logic (PSL) to regulate the loss objective.",
+        "bbox": [
+            507,
+            562,
+            885,
+            900
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We further introduce an ontology-based data aug-",
+        "bbox": [
+            527,
+            903,
+            884,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "* These authors contributed to the work equally.",
+        "bbox": [
+            139,
+            892,
+            436,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "The corresponding author.",
+        "bbox": [
+            139,
+            904,
+            315,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "6988",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6988-6997 December 6-10, 2023 ©2023 Association for Computational Linguistics",
+        "bbox": [
+            159,
+            945,
+            836,
+            973
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/a8ea9ecff77141d7815f9fa6ab8ca47465afef7d7f67b13f7d1f7f859155958d.jpg",
+        "image_caption": [
+            "(a)"
+        ],
+        "image_footnote": [],
+        "bbox": [
+            136,
+            80,
+            500,
+            212
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "image",
+        "img_path": "images/251fc6e910d0eb241c8733920aa8a48994071755226f258cb0780088598bf82b.jpg",
+        "image_caption": [
+            "(b)",
+            "Figure 1: An example of location extraction task from a multi-turn dialogue. The path in red in (a) reflects the hierarchy relationship of the address entities extracted from (b)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            517,
+            80,
+            860,
+            229
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "mentation paradigm to build a labeled multi-turn dialogue dataset for the FGAER task, which could inspire future research. The paradigm first extracts utterance and response (UR) templates from publicly available dialogue datasets and then leverages the largely available full standard address to get labeled fine-grained address entities. The address entities in the UR templates could be altered using the technique we propose to expand current dialogues. To imitate real-world conversation circumstances, a noise injection approach and a ChatGPT augmentation method are also proposed.",
+        "bbox": [
+            112,
+            310,
+            487,
+            501
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We conduct experiments on both the created dataset and a small set of human-labeled datasets derived from real-world customer care conversations. Experimental results show that Log-FGAER improves baselines in both cases.",
+        "bbox": [
+            112,
+            504,
+            487,
+            581
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The contributions of this paper are summarized as follows:",
+        "bbox": [
+            112,
+            583,
+            485,
+            613
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- We make use of probabilistic soft logic to improve the FGAER tasks after utilizing logic rules to indicate the interdependent hierarchy relationship between fine-grained address entities, which reflects the discourse structure of an address-acquiring dialogue.",
+            "- We present an ontology-based and ChatGPT assisted data augmentation paradigm for generating a multi-turn spoken dialogue with labeled fine-grained address entities requiring little human effort, thereby addressing the issue of scarce labeled datasets and high labeling costs.",
+            "- Logic-guided method is a plug-and-play module compatible with all existing FGAER techniques. Experiments indicate that our Log-FGAER with PSL enhancement can obtain"
+        ],
+        "bbox": [
+            136,
+            625,
+            487,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "competitive or even superior performance.",
+        "bbox": [
+            544,
+            310,
+            858,
+            325
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Works",
+        "text_level": 1,
+        "bbox": [
+            507,
+            343,
+            672,
+            357
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Fine-grained Named Entity Recognition",
+        "text_level": 1,
+        "bbox": [
+            507,
+            372,
+            868,
+            387
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Fine-grained named entity recognition aims to resolve the restriction of coarse-grained NER that produces labels from only a small set of entity classes. To solve the lack of datasets where entity boundaries are properly annotated while covering a large spectrum of entity types, (Dogan et al., 2019) proposes to combine learning models (ELMo) with an expansive knowledge base (Wikidata). (Mai et al., 2018) improves the neural network-based Japanese FG-NER performance by removing the CNN layer and utilizing dictionary and type embeddings. The most similar work with us is DLocRL (Xu et al., 2019), which trains a BiLSTM-CRF based POI recognizer for fine-grained location recognition and linking in tweets. This work only considers the POI locations, which is one of the entity types in our scenario.",
+        "bbox": [
+            505,
+            395,
+            882,
+            669
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Named Entity Recognition for dialogue",
+        "text_level": 1,
+        "bbox": [
+            507,
+            686,
+            862,
+            701
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In dialogue systems, NER is a vital task for understanding user intent but suffers from the problems that NE labels are too coarse-grained or the entity types are not linked to a useful ontology. (Muralidharan et al., 2021) proposes a method of encoding the external knowledge of entities not directly using the toponym dictionary but by obtaining information from the knowledge graph. (Bowden et al., 2018) leverages external knowledge such as the Google Knowledge Graph API to solve the problems in the dialogue contexts. However, these methods have quite different entity types, which are not applicable for our scenario.",
+        "bbox": [
+            505,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "6989",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Probabilistic Soft Logic",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            346,
+            99
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "In recent years, PSL rules have been applicable to various machine learning topics. In the knowledge graph construction area, (Chen et al., 2019) infers the confidence score of unseen relations by introducing PSL to represent the uncertain knowledge graph. In the temporal relation extraction topic, (Zhou et al., 2021) leverages PSL regularization to jointly train a relation classifier with sentence embeddings from a deep language model and perform the global temporal inference with a time graph. As for causal inference, (Du et al., 2021) acquires additional evidence information from a large-scale causal event graph as logical rules for causal reasoning. (Cai et al., 2022) proposes a neural-logic based soft logic enhanced event temporal reasoning model for acquiring unbiased temporal common sense knowledge. (Mohler et al., 2020) models the procedural state changes over time with PSL. Most of the works inject the logic knowledge to neural networks, by introducing logic-driven loss functions. These works encourage us to apply PSL rules to the FGAER task.",
+        "bbox": [
+            112,
+            105,
+            489,
+            458
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            112,
+            470,
+            218,
+            483
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.1 Preliminary",
+        "text_level": 1,
+        "bbox": [
+            112,
+            495,
+            255,
+            511
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Problem Formulation. The FGAER task is formulated as a sequence labeling problem (Xu et al., 2019). A multi-turn spoken dialogue is denoted as several turns of sequences  $D = \\{S_1, \\dots, S_i\\}$ , where  $S_i$  represents one turn of question and answer in the dialogue. The FGAER's input  $S_i$  is defined as  $\\{s_1, \\dots, s_j\\}$ , where  $s_j$  denotes the  $j_{th}$  token. The fine-grained address entities are spread across in  $D$ . The output is a tag sequence  $Y_i = \\{y_1, \\dots, y_j\\}$  for each input  $S_i$ , where  $y_j \\in \\gamma$ .  $\\gamma$  is the set of predefined tags following the BIOE tagging scheme. Each tag (e.g., B-city and I-district) in  $\\gamma$  is indicated by B-(begin), I-(inside), and E-(end) of a fine-grained address entity with its type in a type set. For a character that is not inside of any entity, its tag is O-(outside). The FGAER task aims at discovering a list of address entities  $\\{E^1, \\dots, E^n\\}$ .  $E^n = \\{s_{start}, \\dots, s_{end}\\}$  is a substring of  $S_i$  satisfying start  $\\leq$  end.",
+        "bbox": [
+            112,
+            516,
+            489,
+            820
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Probabilistic Soft Logic. As mentioned before, the discourse structure of a multi-turn spoken dialogue usually reflects the spatial constraint relationship between the geographical address entities, which can be modeled with Probabilistic Soft Logic (PSL) (Kimmig et al., 2012). The PSL probabilistic",
+        "bbox": [
+            112,
+            822,
+            489,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "reasoning is incorporated as additional supervision signals to regulate the loss objective of FGAER task. We first introduce several concepts and notations in the PSL language following (Cai et al., 2022), and illustrate how the PSL rule is used to define templates for the hierarchical constraint dependency inference in FGAER task.",
+        "bbox": [
+            507,
+            84,
+            884,
+            197
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 1. The atom formula  $l$  is made up of a predicate  $p$  together with its arguments.  $l$  takes on continuous values in the unit interval [0, 1].",
+        "bbox": [
+            507,
+            198,
+            882,
+            246
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We define three kinds of predicates for our scenario: HIGH, LOW, and BETWEEN. HIGH and LOW both take two arguments to work: an input sequence and the type of address entity. BETWEEN takes three arguments: an input sequence and two address entity types. For example, the atom HIGH(S, city) represents that the highest level of address entity containing in the sequence  $S$  is city. The atom BETWEEN(S, city, town) means that the types of entities in  $S$  are between city and town in the hierarchy structure.",
+        "bbox": [
+            507,
+            247,
+            884,
+            425
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 2. Complex formulae are created by combining atom formulae with logical operations (e.g., conjunction  $(\\wedge)$  and disjunction  $(\\vee)$ ).",
+        "bbox": [
+            507,
+            426,
+            882,
+            475
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 3. A logic rule  $r: f_1 \\Rightarrow f_2$  is an implication formed by combining  $f_1$  and  $f_2$  with logical connectives.  $f_1$  and  $f_2$  can be atomic or complex formulae, respectively.",
+        "bbox": [
+            507,
+            475,
+            882,
+            539
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 4. The soft truth values of the formula  $f$  and the logic rule  $r$  are denoted by the interpretations  $I(f)$  and  $I(r)$ , which refer to the probability that  $f$  and  $r$  hold.",
+        "bbox": [
+            507,
+            541,
+            882,
+            604
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 5. Lukasiewicz t-norms (Klir and Yuan, 1995) adopted to relax logic in PSL are defined as below. With the t-norms to relax logic, we can turn the logic rule  $(r)$  and simple or complex formulas  $(f)$  into differentiable functions, which can then be used as learning goals to control the loss objective:  $I(f_{1} \\wedge f_{2}) = \\max(I(f_{1}) + I(f_{2}) - 1, 0)$  and  $I(r : f_{1} \\Rightarrow f_{2}) = \\min(1, 1 - I(f_{1}) + I(f_{2}))$",
+        "bbox": [
+            507,
+            606,
+            884,
+            751
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "With the above definitions, we can now define the logic rule that explain how the address entities in the current turn of a dialogue sequence relate to those in the previous and following turns.",
+        "bbox": [
+            507,
+            752,
+            882,
+            816
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Log-FGAER",
+        "text_level": 1,
+        "bbox": [
+            507,
+            831,
+            658,
+            847
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Definition 6. Since multi-turn spoken dialogues maintain a discourse structure to get the full standard address during address acquisition, we can assume that the address entity types mentioned in",
+        "bbox": [
+            507,
+            854,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "6990",
+        "bbox": [
+            480,
+            927,
+            521,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/64ee0639f795576920b33a33a5f4a9d98bdf784952365f83c4d72111861c2d02.jpg",
+        "image_caption": [
+            "Figure 2: The main framework of Log-FGAER. a) The inputs of the encoder are: the current turn of the sequence and sequences before and after it. b) The ERNIE encoder is fine-tuned using PSL regularization to learn the spatial hierarchy constraints from the contexts for the current input sequence. c) An example of how to calculate the logic regularization. d) The embedding from the trained (PSL)ERNIE is then sent to the CRF layer for the FGAER task."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            131,
+            80,
+            867,
+            273
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "the current sequence  $S_{i}$  are higher in the administrative division hierarchy than the highest level of address entity types in the next sequence  $S_{i + 1}$  but lower than the lowest level of types in the previous sequence  $S_{i - 1}$ . Such a kind of relationship could be formulated as logic rule  $\\mathbf{R}_{\\mathrm{addr}} \\triangleq \\mathbf{l}_1 \\wedge \\mathbf{l}_2 \\Rightarrow \\mathbf{f}$ :",
+        "bbox": [
+            112,
+            365,
+            487,
+            464
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} l _ {1}: L O W \\left(S _ {i - 1}, a\\right) \\wedge l _ {2}: H I G H \\left(S _ {i + 1}, b\\right) \\tag {1} \\\\ \\Rightarrow f: B E T W E E N (S _ {i}, a, b) \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            129,
+            479,
+            487,
+            514
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In Eq 1, the atom formulae  $l_{1}: LOW(S_{i-1}, a)$  indicates that the lowest-level address type of  $S_{i-1}$  is  $a$ . Similarly,  $l_{2}: HIGH(S_{i+1}, b)$  means that the highest-level address type of  $S_{i+1}$  is  $b$ . The address entity types contained in  $S_{i}$  are generally between  $a$  and  $b$  in the hierarchy structure. Because the hierarchical logic rule  $\\mathbf{R}_{\\mathrm{addr}}$  have the form  $l_{1} \\wedge l_{2} \\Rightarrow f$ , we can obtain a soft truth value by applying the Lukasiewicz t-norm to the interpretation of  $\\mathbf{R}_{\\mathrm{addr}}$ .",
+        "bbox": [
+            112,
+            530,
+            487,
+            676
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Figure 2 illustrates the framework of LogFGAER, which consists of two components: 1) a pre-trained language model (PLM) encoder that is fine-tuned with PSL regularization; and 2) an FGAER module that is composed of the fine-tuned PLM encoder and a CRF layer.",
+        "bbox": [
+            112,
+            677,
+            487,
+            772
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "1) Fine-tune the PLM encoder with PSL regularization. In this component, the PSL rule from Definition 6 is used to calculate the logic regulation loss, which is then combined with the original PLM model's cross-entropy loss to fine-tune the PLM encoder. In this manner, the embeddings generated by the fine-tuned PLM encoder would incorporate the spatial constraint dependency information in the dialogue context.",
+        "bbox": [
+            112,
+            774,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "More specifically, the PLM encoder accepts a sequence as input that may contain address entities and then outputs a token-level embedding of the sequence. The ERNIE (Sun et al., 2019) is proposed as the PLM backbone because it has been enhanced for named entity recognition by incorporating entity semantic information. Given  $S_{i}$  as input, we could get the embedding  $\\mathbf{H}_{\\mathbf{i}} = ERNIE(\\mathbf{S}_{\\mathbf{i}})$ , where  $H_{i} = \\{h_{1},\\dots,h_{j}\\}$  and  $h_j$  denotes the output of the pre-trained ERNIE that corresponds to the input token  $s_j \\in S_i$ . The predicted probability distribution  $\\hat{y}_j$  for  $s_j$  is then obtained by adding on top of the ERNIE a linear layer (Feed-Forward Network) with softmax, as shown below:",
+        "bbox": [
+            507,
+            365,
+            884,
+            589
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\hat {y _ {j}} = P \\left(y _ {j} \\mid s _ {j}\\right) = \\operatorname {s o f t m a x} \\left(W h _ {j} + b\\right) \\tag {2}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            569,
+            598,
+            880,
+            612
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "In the above equation,  $W$  and  $b$  are trainable parameters.  $y_{j}$  is the ground-truth label for  $s_{j}$ .",
+        "bbox": [
+            507,
+            625,
+            880,
+            657
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We compute the prediction loss  $L_{ce}$  using the cross-entropy objective as below:",
+        "bbox": [
+            507,
+            657,
+            880,
+            688
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {c e} = - \\frac {1}{\\left| S _ {i} \\right|} \\sum_ {j = 1} ^ {\\left| S _ {i} \\right|} C r o s s E n t r o p y \\left(y _ {j}, \\hat {y} _ {j}\\right) \\tag {3}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            563,
+            697,
+            880,
+            734
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Following that, we will show how to compute the logic regulation loss  $L_{logic}$  and then combine it with the  $L_{ce}$  to ensure that the predicted distribution  $\\hat{y}_j$  is compatible with the hierarchical logic rule  $\\mathbf{R}_{\\mathrm{addr}}$ , as shown in Definition 6.",
+        "bbox": [
+            507,
+            741,
+            880,
+            821
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To calculate  $L_{logic}$ , we should get  $\\mathrm{I}(\\mathbf{R}_{\\mathrm{addr}})$  first, which is the truth value of logic rule  $\\mathbf{R}_{\\mathrm{addr}}$ . Figure 2(c) illustrates an  $\\mathrm{I}(\\mathbf{R}_{\\mathrm{addr}})$  calculation example.  $\\mathrm{I}(\\mathrm{LOW}(S_i, city))$ , also known as  $\\mathrm{I}(l_1)$ , is intended to be the predicted probabilities of the characters \"Hangzhou city\" to be the entity type of",
+        "bbox": [
+            507,
+            822,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "6991",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "\"city\". Similarly, I(BETWEEN  $(S_{i},\\text{city},\\text{town}))$  also known as  $\\mathrm{I}(f)$ , represents the predicted probabilities of \"Jianggan District\" and \"Jiubao Town\" in  $S_{i}$ , whose entity types are between \"city\" and \"town\". After getting the values of  $\\mathrm{I}(l_1)$ ,  $\\mathrm{I}(l_2)$  and  $\\mathrm{I}(f)$ ,  $\\mathrm{I}(\\mathbf{R}_{\\mathbf{addr}})$  which denotes the degree that the rule  $\\mathbf{R}_{\\mathbf{addr}}$  are satisfied by the predicted distribution, can be derived by incorporating the Lukasiewicz t-norm.",
+        "bbox": [
+            112,
+            84,
+            487,
+            227
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The principle of PSL is: the larger the truth values are, the better the logic rules are satisfied. Based on this principle, we formulate the distance of  $\\mathrm{I}(\\mathbf{R}_{\\mathrm{addr}})$  to be true as a regularization term to penalize the distorted distribution that violates the logic rule  $\\mathbf{R}_{\\mathrm{addr}}$  (Cai et al., 2022). The logic regulation loss  $L_{logic}$  is calculated as below:",
+        "bbox": [
+            112,
+            229,
+            489,
+            342
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {\\text {l o g i c}} = 1 - I \\left(\\mathbf {R} _ {\\mathbf {a d d r}}\\right) \\tag {4}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            211,
+            355,
+            487,
+            370
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We finalize the loss function by applying a weighted sum of Eq 3 and Eq 4:",
+        "bbox": [
+            112,
+            384,
+            487,
+            416
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {E R N I E} = L _ {c e} + \\lambda L _ {\\text {l o g i c}} \\tag {5}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            201,
+            429,
+            487,
+            445
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where  $\\lambda$  is the weight for PSL regularization term. The objective is to minimize the  $L_{ERNIE}$  and to finetune the ERNIE encoder parameters.",
+        "bbox": [
+            112,
+            458,
+            487,
+            506
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "With the regularization of PSL rule on the outputs of the ERNIE encoder, our model can make embeddings capture more contextual information, which could further benefit the sequence labeling task described in the FGAER module.",
+        "bbox": [
+            112,
+            507,
+            487,
+            585
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "2) FGAER module is made up of the ERNIE encoder that has been fine-tuned with PSL regulation and a CRF layer (Lafferty et al., 2001), denoting as (PSL)ERNIE-CRF. The loss function  $L_{CRF}$  of FGAER is calculated by feeding  $\\mathbf{H}$  from Eq 2 into the CRF layer. The objective is to minimize  $L_{CRF}$ . The trained FGAER model could be used in the inference stage for the FGAER task.",
+        "bbox": [
+            112,
+            587,
+            489,
+            715
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4 Ontology Based Dialogue Data Augmentation Leveraging ChatGPT",
+        "text_level": 1,
+        "bbox": [
+            112,
+            727,
+            460,
+            760
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.1 Ontology-based dialogue enrichment",
+        "text_level": 1,
+        "bbox": [
+            112,
+            769,
+            448,
+            785
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The goal of ontology-based dialogue enrichment is to address the issue of a shortage of dialogue datasets with labeled fine-grained address entities. As previously stated, labeling the fine-grained entities is costly and time-consuming. Fortunately, large-scale, unlabeled, full standard addresses (for example, Ecommerce Park, Hanghai Road, Jiubao Town, Jianggan District, Hangzhou City) could be",
+        "bbox": [
+            112,
+            790,
+            489,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "simply crawled from web pages such as Google Maps<sup>1</sup>. Then leveraging the recently well-studied address entity segmentation methods(Chen et al., 2015), which refers to the process of splitting unstructured standard addresses into address entities, we were able to label the full standard address data with fine-grained entity types with minimal human effort.",
+        "bbox": [
+            507,
+            84,
+            884,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/a3b7aebfa66bd925cd4765e5928e6f0d7857178a86459dcbd24bc15369bb0002.jpg",
+        "image_caption": [
+            "Figure 3: Ontology-based dialogue enrichment."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            221,
+            878,
+            326
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By using the labeled full standard address, the ontology-based dialogue data enrichment method can expand the limited sampled UR template pairs that include address entities, extracting from current small-size labeled dialogue datasets. Take Figure 3 as an example to explain our enrichment method: \"Hangzhou city\", \"Jianggan District\" and \"Baiyang Town\" in the origin UR pair are replaced by new values in ontology  $②$ . About 578 groups of UR pair templates we used to create dialogue datasets are extracted from the CUCC call center dataset<sup>2</sup> and the Taobao E-commercial Conversation Corpus<sup>3</sup>. The candidate values for each of the ontologies are derived from Tianchi Competitions' full standard address dataset, which has 8.8K full standard addresses with 16 types. More full standard address data can be crawled from websites and labeled using address segmentation methods with limited efforts following (Chen et al., 2015).",
+        "bbox": [
+            507,
+            370,
+            882,
+            675
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.2 Noise injection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            687,
+            672,
+            702
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "To bridge the gap between actual customer support scenarios and ontology-based method-generated dialogues, we present the noise injection method, which captures the typical speaking habits of people in real-life situations. To be specific, we implement five strategies following the prior works (Tian et al., 2021; Feng et al., 2021; Liu et al., 2020): 1) Pause, 2) Repeat, 3) Restart, 4)Insert, and 5)Repair. The details could be referenced to the above mentioned work. To generate the in-",
+        "bbox": [
+            507,
+            708,
+            884,
+            869
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://www.google.com",
+        "bbox": [
+            529,
+            879,
+            690,
+            891
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://www.datafountain.cn/competitions/536/datasets",
+        "bbox": [
+            532,
+            891,
+            873,
+            904
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://github.com/coolelf/DeepUtteranceAggregation",
+        "bbox": [
+            532,
+            904,
+            863,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "6992",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/f18c866114931bfdfaff9aa909be76901f7659cf505039f60e8935decd0ec9ca.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\">Original texts</td><td>Altered texts</td></tr><tr><td rowspan=\"2\">Sentence level</td><td>U: Please provide the detailed address, where are you now?</td><td>U: Could you kindly furnish the specific address of your current location?\nU: Can you please share the precise whereabouts of your present position?\nU: May I request the elaborate address of your current whereabouts?</td></tr><tr><td>R: I am in Hangzhou, Zhejiang Province.</td><td>R:My current location is Hangzhou, located in Zhejiang Province.\nR:I find myself in Hangzhou, which is situated in Zhejiang Province.\nR:I am located in Hangzhou, specifically in Zhejiang Province.</td></tr><tr><td>Dialogue level</td><td>U: Repeat please, I didn’t hear you clearly.\nR: I am in the dental hospital.</td><td>U:Repeat please, I didn’t catch what you said clearly.\nR:I am currently situated in the dental hospital.\nU:Could you please repeat that? Your words weren’t clear to me.\nR:I am currently located at the dental hospital.\nU:I apologize for the inconvenience. Could you repeat your statement?\nR:I am currently in the dental hospital.</td></tr></table>",
+        "bbox": [
+            117,
+            80,
+            884,
+            212
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Generated examples from ChatGPT augmentation method.",
+        "bbox": [
+            267,
+            221,
+            724,
+            235
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "correct pronunciation, we employ xpinyin  $^{4}$  and Pinyin2Hanzi  $^{5}$  to mutually convert pinyin and Chinese characters.",
+        "bbox": [
+            112,
+            260,
+            489,
+            307
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "4.3 ChatGPT improves diversity",
+        "text_level": 1,
+        "bbox": [
+            112,
+            321,
+            384,
+            336
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We employ ChatGPT  $^{6}$ , a powerful data augmentation strategy (Ubani et al., 2023), to increase data diversity by improving current UR templates and transforming created dialogues (after Noise injection) into new forms that differ from the original but retain the same meaning.",
+        "bbox": [
+            112,
+            342,
+            487,
+            437
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "We have designed prompts for rephrasing at the sentence level (UR pair templates) and dialogue level (after Noise injection). The specific prompt for sentence-level rephrasing is \"Please rewrite the following sentence I give: {text}, and give me 3 rephrased results\". The number of UR pair templates is increased from 578 to 778 for Ontology-based dialogue enrichment. The prompt for dialogue-level rephrasing is \"Please rewrite the following dialogue I give: {text}, and give me 2 rephrased answers while remaining the number of sentences\". The amount of generated dataset exceeded 1k after ChatGPT enhancement. Table 1 shows the samples generated using ChatGPT. We also consider the degree of similarity between the new samples and the ones we already have, as illustrated in Figure 4. We are able to better capture data invariance and increase the sample size by leveraging ChatGPT's outstanding generating capabilities.",
+        "bbox": [
+            115,
+            439,
+            489,
+            760
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5 Experiments",
+        "text_level": 1,
+        "bbox": [
+            112,
+            772,
+            258,
+            789
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.1 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            800,
+            226,
+            814
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "FGAER-dia We create a labeled multi-turn dialogue dataset called FGAER-dia, which is constructed according to the paradigm in Section 4.",
+        "bbox": [
+            112,
+            821,
+            489,
+            869
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "image",
+        "img_path": "images/1342d349099931cddab67eb3007b86f9c21d24734cca025583e85e064b001dc8.jpg",
+        "image_caption": [
+            "Figure 4: The average Cosine Similarity, Levenshtein Ratio, Jaro Distance, Jaro-Winkler Distance, and Jaccard Similarity for 100 pairs of generated samples and original samples are listed individually at the sentence and dialogue levels."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            512,
+            256,
+            878,
+            369
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "FGAER-dia is a labeled multi-turn dialogue dataset of 9.8k groups of dialogue contexts with 16 types of fine-grained address entities. The generated dialogue's average length of characters is 167 and the maximum length is 366 in FGAER-dia. The maximum number of turns is 10. Our proposed dataset and the code for our Log-FGAER method are openly available<sup>7</sup>.",
+        "bbox": [
+            507,
+            476,
+            882,
+            604
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "CUCC-labeled The CUCC call-center dialogue dataset provided by DataFountain Competition<sup>2</sup> is transcribed by a real-world downstream speech recognition system, including plenty of transcription errors. We filter out 200 dialogues containing address entities and manually label the filtered data with 16 different types of address entities, to evaluate the Log-FGAER's performance on the multi-turn dialogues from real-life service environment.",
+        "bbox": [
+            507,
+            606,
+            882,
+            749
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 5 in Appendix contains more detailed examples of fine-grained address entity types.",
+        "bbox": [
+            507,
+            751,
+            882,
+            783
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Implementation details",
+        "text_level": 1,
+        "bbox": [
+            507,
+            796,
+            737,
+            810
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In our experiments, ERNIE is performed with BaiDu PaddlePaddle  $^{8}$ , which is a framework of deep learning. For hyperparameter settings, the max sequence length of RoBERTa-wwm-ext,",
+        "bbox": [
+            507,
+            816,
+            882,
+            881
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>4</sup>https://pypi.org/project/xpinyin/",
+        "bbox": [
+            134,
+            877,
+            339,
+            891
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>5</sup>https://www.cnpython.com/pypi/pinyin2hanzi",
+        "bbox": [
+            136,
+            891,
+            421,
+            904
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>6</sup>https://openai.com/product/chatgpt",
+        "bbox": [
+            136,
+            904,
+            357,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>7</sup>https://github.com/Devil0817/LOG-FGAER",
+        "bbox": [
+            529,
+            890,
+            806,
+            904
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{8}$ https://www.paddlepaddle.org",
+        "bbox": [
+            532,
+            904,
+            722,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "6993",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/593b99c653e0f609cea0e4f03c6bb9a6ab3a982068239224b742047427fdc0c9.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\" colspan=\"2\">Methods</td><td colspan=\"3\">FGAER-dia</td><td colspan=\"3\">CUCC-labeled</td></tr><tr><td>P</td><td>R</td><td>F1</td><td>P</td><td>R</td><td>F1</td></tr><tr><td rowspan=\"2\">BERT-based</td><td>Baseline-1</td><td>92.32</td><td>91.09</td><td>91.55</td><td>65.72</td><td>61.72</td><td>61.75</td></tr><tr><td>Baseline-2</td><td>92.26</td><td>91.62</td><td>91.83</td><td>72.43</td><td>67.74</td><td>68.53</td></tr><tr><td rowspan=\"3\">RoBERTa-based</td><td>Baseline-3</td><td>90.85</td><td>90.88</td><td>90.86</td><td>64.66</td><td>64.64</td><td>63.35</td></tr><tr><td>Baseline-4</td><td>92.79</td><td>91.68</td><td>92.21</td><td>69.59</td><td>65.77</td><td>66.86</td></tr><tr><td>Baseline-5</td><td>93.60</td><td>93.23</td><td>93.41</td><td>69.77</td><td>73.37</td><td>70.54</td></tr><tr><td rowspan=\"4\">ERNIE-based</td><td>Baseline-6</td><td>92.10</td><td>92.46</td><td>92.28</td><td>83.39</td><td>83.99</td><td>83.69</td></tr><tr><td>Baseline-7</td><td>93.28</td><td>92.78</td><td>93.03</td><td>87.47</td><td>82.91</td><td>85.13</td></tr><tr><td>Baseline-8</td><td>93.15</td><td>91.68</td><td>92.41</td><td>82.10</td><td>84.17</td><td>83.12</td></tr><tr><td>Log-FGAER</td><td>94.15</td><td>93.63</td><td>93.88</td><td>86.20</td><td>87.12</td><td>86.65</td></tr></table>",
+        "bbox": [
+            193,
+            80,
+            803,
+            263
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 2: Experiment results on FGAER-dia and CUCC-labeled datasets, and the best performance is highlighted in boldface.",
+        "bbox": [
+            112,
+            272,
+            882,
+            303
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "BERT, and ERNIE is 256. The hidden state size of BiLSTM is 128 and the dropout rate is set as 0.1 for BiLSTM output. We use the AdamW optimizer in a mini-batch size of 64 with learning rate  $\\gamma = 5 \\times 10^{-5}$  for each experiment. The weight decay of ERNIE is 0.01. The best  $\\lambda$  is 0.15. The split of  $80\\%$  for train,  $10\\%$  for validation and  $10\\%$  for test is used. The metrics used for evaluation are precision(P), recall(R) and Micro-F1 metrics. The models with the best validation loss that are selected from all epochs, are saved and utilized for testing. For hardware, we use a four-core CPU and an NVIDIA Tesla V100 GPU.",
+        "bbox": [
+            112,
+            326,
+            489,
+            536
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Baseline models.",
+        "text_level": 1,
+        "bbox": [
+            112,
+            550,
+            290,
+            564
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "We compare the performance of Log-FGAER with several strong baseline methods listed as below.",
+        "bbox": [
+            112,
+            571,
+            487,
+            604
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Baseline-1 uses the structure of Bert $^{9}$ -CRF.",
+            "- Baseline-2 employs BERT-BiLSTM-CRF as the structure.",
+            "- Baseline-3 adopts Chinese RoBERTa-wwm-ext $^{10}$ -CRF as the architecture.",
+            "- Baseline-4 adds a BiLSTM layer between RoBERTa-wwm-ext and CRF layers.",
+            "- Baseline-5 has the same structure as Baselined-4, using a distillation approach to transfer knowledge from labeled full standard addresses to dialogues (Wang et al., 2022).",
+            "- Baseline-6 only finetunes the ERNIE encoder."
+        ],
+        "bbox": [
+            136,
+            617,
+            487,
+            879
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "- Baseline-7 is based on the ERNIE-CRF structure.",
+            "- Baseline-8 applies ERNIE-BiLSTM-CRF to deal with FGAER task."
+        ],
+        "bbox": [
+            531,
+            326,
+            882,
+            406
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.4 Results and Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            439,
+            717,
+            455
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "The following insights can be drawn from the results as listed in Table 2:",
+        "bbox": [
+            507,
+            462,
+            882,
+            493
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1) Compared baselines using the same model structures but different PLM encoders(e.g., Baseline-2, Baseline-4 and Baseline-8 all add BiLSTM-CRF layers on encoders), ERNIE significantly outperforms on FGAER task.",
+            "2) We compare the state-of-the-art (SOTA) structures (Baseline 1-8) for the FGAER task. On the FGAER-dia dataset, Baseline-7 achieves better F1-score performance. As a result, we combine the proposed Logic-guided method with Baseline-7 to obtain the Log-FGAER.",
+            "3) With the Logic-guided method enhanced, Log-FGAER outperforms other baselines on both FGAER-dia and CUCC-labeled datasets."
+        ],
+        "bbox": [
+            507,
+            495,
+            880,
+            722
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To further display the progress of Log-FGAER, we exclusively list all entity types with an F1-score below  $80\\%$ . As shown in Table 3 and Figure 5, Log-FGAER can efficiently enhance the F1-scores of almost all entity types that are prone to prediction errors. In particular, the Log-FGAER consistently outperforms Baseline-7 (same structure without PSL) by 3.28 percent over F1-score on entity type \"subpoi\". These findings demonstrate that the logic inference introduced in our method can capture and utilize geographical hierarchy information contained in the context.",
+        "bbox": [
+            507,
+            726,
+            882,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{9}$ https://huggingface.co/bert-base-chinese  \n $^{10}$ https://github.com/ymcui/Chinese-BERT-wwm",
+        "bbox": [
+            132,
+            890,
+            431,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "6994",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "table",
+        "img_path": "images/d02bc2c755fd62221692cd71230fcd0adfbb84d20085f56b1c2462c53744c5ec.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>DISTRICT</td><td>COMMUNITY</td><td>POI</td><td>SUBPOI</td><td>INTERSECTION</td><td>overall</td></tr><tr><td>Baseline-6</td><td>70.45</td><td>62.36</td><td>65.86</td><td>65.79</td><td>70.10</td><td>92.28</td></tr><tr><td>Baseline-7</td><td>70.88</td><td>62.91</td><td>66.43</td><td>67.18</td><td>70.48</td><td>93.03</td></tr><tr><td>Baseline-8</td><td>70.13</td><td>62.30</td><td>66.13</td><td>58.07</td><td>70.50</td><td>92.41</td></tr><tr><td>Log-FGAER</td><td>70.98</td><td>63.12</td><td>66.28</td><td>70.46</td><td>71.38</td><td>93.88</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            885,
+            168
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/6020dffb8669adca8059827cc357224b4d3b6cfbfaefd6ece00fcd55bb8a1e4f.jpg",
+        "image_caption": [
+            "Figure 5: F1-scores of Log-FGAER and ERNIE-based baselines on FGAER-dia."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            215,
+            485,
+            344
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.5 Ablation Study",
+        "text_level": 1,
+        "bbox": [
+            112,
+            404,
+            280,
+            420
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Additional experiments were conducted to demonstrate the efficacy of representing contextual information as logic rules that better illustrate the contextual information than comparable methods.",
+        "bbox": [
+            112,
+            425,
+            489,
+            489
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "1) To begin, we demonstrate how our proposed method can be utilized as an add-on module to improve existing sequence labeling models. We obtain (PSL)ERNIE and (PSL)ERNIE-BiLSTM-CRF after merging the proposed Probabilistic Soft Logic-guided methodology with Baseline-6 (ERNIE) and Baseline-8 (ERNIE-BiLSTM-CRF). As shown in Figure 6, both ERNIE and ERNIE-BiLSTM-CRF improve over various entity types with PSL enhancement. The PSL strategy has proven to be an independent and simple-to-implement module that can be applied to traditional sequence labeling models to make significant improvements.",
+        "bbox": [
+            112,
+            489,
+            489,
+            715
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/0b49a3020554a5d7f30098b9dd278cafd4f37380e39e192068ea7a11eece5cc9.jpg",
+        "image_caption": [
+            "Figure 6: F1 scores of (PSL)ERNIE and (PSL)ERNIE-BiLSTM-CRF models against the same structures without PSL regularization on FGAER-dia."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            115,
+            725,
+            295,
+            818
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "image",
+        "img_path": "images/11741b03b86c646291b19e94b079887066cea21aaa9720924ac9d92e13a764b2.jpg",
+        "image_caption": [],
+        "image_footnote": [],
+        "bbox": [
+            297,
+            725,
+            485,
+            819
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "2)Second, we attempt to demonstrate that the PSL strategy expresses contextual information",
+        "bbox": [
+            112,
+            887,
+            487,
+            919
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "better than comparable approaches in multi-turn dialogues. We compare to CMV(Luoma and Pyysalo, 2020) and Multi-head Attention methods. CMV(Luoma and Pyysalo, 2020) explores cross-sentence contexts for NER with BERT. Multi-head Attention employs the ERNIE-CRF structure to produce novel representations with context information. This is accomplished by adding address entities from the preceding and subsequent sequences to the current sequence. Take Figure 2(a) as an example: \"Hangzhou City\" is added at the front, while \"Jiubao Town\" is inserted at the end of  $S_{i}$ , forming a new representation as the input. In this way, the contextual information could be learned by Transformer multi-head attention mechanism. The comparisons are also evaluated on the FGAER-dia.",
+        "bbox": [
+            507,
+            217,
+            884,
+            475
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Log-FGAER outperforms the other two approaches, as shown in Table 4, demonstrating the efficiency of our logic-guided method in exploiting contextual address hierarchy information.",
+        "bbox": [
+            507,
+            475,
+            882,
+            539
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/e20da98569b8fb5999b7bca3bb7660dd03093219f13ed8f09162329685b06f6e.jpg",
+        "table_caption": [
+            "Table 3: F1-scores of Log-FGAER and ERNIE-based baselines on FGAER-dia."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Methods</td><td>P</td><td>R</td><td>F1</td></tr><tr><td>Multi-head attention</td><td>91.87</td><td>92.02</td><td>91.94</td></tr><tr><td>CMV</td><td>93.09</td><td>93.17</td><td>93.12</td></tr><tr><td>Log-FGAER</td><td>94.15</td><td>93.63</td><td>93.88</td></tr></table>",
+        "bbox": [
+            524,
+            549,
+            870,
+            620
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Results of models using different methods to obtain contextual information on FGAER-dia.",
+        "bbox": [
+            507,
+            627,
+            882,
+            657
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion",
+        "text_level": 1,
+        "bbox": [
+            507,
+            684,
+            640,
+            700
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "In this paper, we propose an effective method named Log-FGAER to extract address entities from spoken dialogue contexts using the PSL rule. Leveraging the spatially constrained relationship of address entities in the dialogues, we improve entity recognition inference by presenting the interdependent hierarchy relationship as the logic rule to regulate the loss objective. We also propose a data augmentation paradigm to construct the labeled multi-turn dialogue dataset for training and evaluation of our proposed method. Experiments on the multi-turn spoken dialogue dataset show the effectiveness of our method.",
+        "bbox": [
+            507,
+            709,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "6995",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Limitations",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            220,
+            99
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "While our model achieves satisfactory results, it still has several limitations. This study has some potential limitations based on the datasets. Due to the lack of large-scale bench datasets, we created the FGAER-dia and CUCC-labeled datasets, which were both originally constructed in Chinese. In addition, both of these datasets contain a small amount of data. This can lead to an increasing risk of biased or incomplete conclusions and generalizations when applied to real-world scenarios. These challenges and constraints will be addressed in future work.",
+        "bbox": [
+            112,
+            109,
+            492,
+            303
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            114,
+            313,
+            265,
+            330
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "This work adheres to ethical guidelines and responsible practices in ACL policies. The authors have no external conflicts of interest to declare and have not been required to seek any ethics clearances in order to undertake this work. The datasets used in this work are from previously published works or competitions and, in our view, do not have any attached privacy or ethical issues. We provided data augmentation approaches to improve the evaluation of our proposal. The augmentation-generated conversations aim to provide an optional way to inspire research rather than violate ethical guidelines and should be used in a strictly controlled manner. Transparency and responsible reporting are priorities, and we welcome community engagement to enhance ethical considerations in our work.",
+        "bbox": [
+            112,
+            340,
+            489,
+            598
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            114,
+            626,
+            213,
+            640
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Kevin K. Bowden, Jiaqi Wu, Shereen Oraby, Anita Misra, and Marilyn Walker. 2018. Slugnerds: A named entity recognition tool for open domain dialogue systems.",
+            "Bibo Cai, Xiao Ding, Bowen Chen, Li Du, and Ting Liu. 2022. Mitigating reporting bias for semi-supervised temporal commonsense inference with probabilistic soft logic. The Thirty-Sixth AAAI Conference on Artificial Intelligence.",
+            "Boli Chen, Guangwei Xu, Xiaobin Wang, Pengjun Xie, Meishan Zhang, and Fei Huang. 2022. Aishell-ner: Named entity recognition from chinese speech. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8352-8356.",
+            "Xinchi Chen, Xipeng Qiu, Chenxi Zhu, and Xuan-Jing Huang. 2015. Gated recursive neural network for Chinese word segmentation. In Proceedings of the 53rd"
+        ],
+        "bbox": [
+            115,
+            648,
+            489,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1744-1753.",
+            "Xuelu Chen, Muhao Chen, Weijia Shi, Yizhou Sun, and Carlo Zaniolo. 2019. Embedding uncertain knowledge graphs. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 3363-3370.",
+            "Cihan Dogan, Aimore Dutra, Adam Gara, Alfredo Gemma, Lei Shi, Michael Sigamani, and Ella Walters. 2019. Fine-grained named entity recognition using elmo and wikidata.",
+            "Li Du, Xiao Ding, Kai Xiong, Ting Liu, and Bing Qin. 2021. Excar: Event graph knowledge enhanced explainable causal reasoning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2354-2363.",
+            "Nazmiye Eligüzel, Cihan Çetinkaya, and TürKay Dereli. 2020. Comparison of different machine learning techniques on location extraction by utilizing geotagged tweets: A case study. Advanced Engineering Informatics, 46:101151.",
+            "Steven Y Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for nlp. arXiv preprint arXiv:2105.03075.",
+            "Tran Thi Hong Hanh, Antoine Douce, Nicolas Sidere, Jose G. Moreno, and Senja Pollak. 2021. Named entity recognition architecture combining contextual and global features. International Conference on Asian Digital Libraries, page 264-276.",
+            "Angelika Kimmig, Stephen Bach, Matthias Broecheler, Bert Huang, and Lise Getoor. 2012. A short introduction to probabilistic soft logic. In Proceedings of the NIPS workshop on probabilistic programming: foundations and applications, pages 1-4.",
+            "George Klir and Bo Yuan. 1995. Fuzzy sets and fuzzy logic, volume 4. Prentice hall New Jersey.",
+            "John Lafferty, Andrew McCallum, and Fernando CN Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data.",
+            "Jiexi Liu, Ryuichi Takanobu, Jiaxin Wen, Dazhen Wan, Hongguang Li, Weiran Nie, Cheng Li, Wei Peng, and Minlie Huang. 2020. Robustness testing of language understanding in task-oriented dialog. arXiv preprint arXiv:2012.15262.",
+            "Jouni Luoma and Sampo Pyysalo. 2020. Exploring cross-sentence contexts for named entity recognition with bert. In Proceedings of the 28th International Conference on Computational Linguistics, page 904-914."
+        ],
+        "bbox": [
+            510,
+            85,
+            884,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "6996",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Khai Mai, Thai-Hoang Pham, Minh Trung Nguyen, Tuan Duc Nguyen, Danushka Bollegala, Ryohei Sasano, and Satoshi Sekine. 2018. An empirical study on fine-grained named entity recognition. In Proceedings of the 27th International Conference on Computational Linguistics, pages 711-722.",
+            "Michael Mohler, Sean Monahan, and Marc Tomlinson. 2020. Modeling procedural state changes over time with probabilistic soft logic. In *The Thirty-Third International Flairs Conference*.",
+            "Deepak Muralidharan, Joel Ruben Antony Moniz, Weicheng Zhang, Stephen Pulman, Lin Li, Megan Barnes, Jingjing Pan, Jason Williams, and Alex Acero. 2021. Dexter: Deep encoding of external knowledge for named entity recognition in virtual assistants.",
+            "Shuangli Shan, Zhixu Li, Qiang Yang, An Liu, Lei Zhao, Guanfeng Liu, and Zhigang Chen. 2020. Geographical address representation learning for address matching. World Wide Web, 23(3):2005-2022.",
+            "Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. Ernie: Enhanced representation through knowledge integration. arXiv preprint arXiv:1904.09223.",
+            "Xin Tian, Xinxian Huang, Dongfeng He, Yingzhan Lin, Siqi Bao, Huang He, Liankai Huang, Qiang Ju, Xiyuan Zhang, Jian Xie, et al. 2021. Tod-da: Towards boosting the robustness of task-oriented dialogue modeling on spoken conversations. arXiv preprint arXiv:2112.12441.",
+            "Solomon Ubani, Suleyman Olcay Polat, and Rodney Nielsen. 2023. Zeroshotdataaug: Generating and augmenting training data with chatgpt. arXiv preprint arXiv:2304.14334.",
+            "Yitong Wang, Xue Han, Feng Zhou, Yiting Wang, Chao Deng, and Junlan Feng. 2022. Distill-aer: Fine-grained address entity recognition from spoken dialogue via knowledge distillation. In NLPCC 2022: Natural Language Processing and Chinese Computing, page 643-655.",
+            "Ting-Wei Wu and Biing-Hwang Juang. 2022. Knowledge augmented bert mutual network in multi-turn spoken dialogues. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7487-7491.",
+            "Canwen Xu, Jing Li, Xiangyang Luo, Jiaxin Pei, Chen-liang Li, and Donghong Ji. 2019. Dlocrl: A deep learning pipeline for fine-grained location recognition and linking in tweets. In The World Wide Web Conference, pages 3391-3397.",
+            "Jiudong Yang, Peiying Wang, Yi Zhu, Mingchao Feng, Meng Chen, and Xiaodong He. 2022. Gated multimodal fusion with contrastive learning for turn-taking prediction in human-robot dialogue. In ICASSP 2022 - 2022 IEEE International Conference on Acoustics,"
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Speech and Signal Processing (ICASSP), pages 7747-7751.",
+            "P.A. Zandbergen. 2008. A comparison of address point, parcel and street geocoding techniques. In Comput. Environ. Urban Syst., volume 32, page 214-232.",
+            "Haisong Zhang, Lemao Liu, Haiyun Jiang, Yangming Li, Enbo Zhao, Kun Xu, Linfeng Song, Suncong Zheng, Botong Zhou, Jianchen Zhu, et al. 2020. Texmart: A text understanding system for fine-grained ner and enhanced semantic analysis. arXiv preprint arXiv:2012.15639.",
+            "Yichao Zhou, Yu Yan, Rujun Han, J Harry Caufield, Kai-Wei Chang, Yizhou Sun, Peipei Ping, and Wei Wang. 2021. Clinical temporal relation extraction with probabilistic soft logic regularization and global inference. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 14647-14655."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            351
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "table",
+        "img_path": "images/2a9a66881ba3b19837bec99f6b2dd19a8990c8e24d659788c9474e089a876584.jpg",
+        "table_caption": [
+            "A Appendix"
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Label</td><td>Interpretation</td><td>Example</td></tr><tr><td>PROVINCE</td><td>name of a province</td><td>浙江\nZhejiang Province</td></tr><tr><td>CITY</td><td>name of a city</td><td>杭州市\nHangzhou City</td></tr><tr><td>DISTRICT</td><td>name of a district in a city</td><td>西湖区\nXihu District</td></tr><tr><td>TOWN</td><td>name of a town or a boulevard</td><td>九堡镇\nJiubao Town</td></tr><tr><td>ROAD</td><td>name of a road</td><td>航海路\nHanghai Road</td></tr><tr><td>INTERSECTION</td><td>road junction</td><td>交叉口\nIntersection</td></tr><tr><td>ROADNO</td><td>road number</td><td>5号\n#5</td></tr><tr><td>DEVZONE</td><td>a economic development zone</td><td>开发区\nDevelopment Zone</td></tr><tr><td>COMMUNITY</td><td>name of a community or a village</td><td>宁安社区\nNingan Community</td></tr><tr><td>POI</td><td>name of the point of interest</td><td>衢州人民医院\nQuzhou Hospital</td></tr><tr><td>SUBPOI</td><td>name of the second point of interest</td><td>放射科\nRadiology Department</td></tr><tr><td>ASSIST</td><td>a phrase indicating relative position</td><td>对面\nOpposite</td></tr><tr><td>DISTANCE</td><td>amount of space between two points</td><td>100米\n100m</td></tr><tr><td>HOUSENO</td><td>house number</td><td>3幢\nBlock #3</td></tr><tr><td>CELLNO</td><td>cell number</td><td>2单元\nUnit #2</td></tr><tr><td>FLOORNO</td><td>floor number</td><td>6层\nLevel 60</td></tr></table>",
+        "bbox": [
+            510,
+            401,
+            884,
+            731
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Interpretations and examples for each entity type used in our datasets, according to the address hierarchical order from top to bottom.",
+        "bbox": [
+            510,
+            741,
+            882,
+            784
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "6997",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    }
+]