Add Batch dd292eaa-39f8-43f2-834f-262509570d9b data

.gitattributes

@@ -3633,3 +3633,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 2024/Modeling[[:space:]]Gender[[:space:]]and[[:space:]]Dialect[[:space:]]Bias[[:space:]]in[[:space:]]Automatic[[:space:]]Speech[[:space:]]Recognition/04db79b6-7b80-42c1-828b-ce30592e8d11_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Modeling[[:space:]]Historical[[:space:]]Relevant[[:space:]]and[[:space:]]Local[[:space:]]Frequency[[:space:]]Context[[:space:]]for[[:space:]]Representation-Based[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Forecasting/745cfddb-65c7-45db-a655-ddb24e791a44_origin.pdf filter=lfs diff=lfs merge=lfs -text
 2024/Modeling[[:space:]]Human[[:space:]]Subjectivity[[:space:]]in[[:space:]]LLMs[[:space:]]Using[[:space:]]Explicit[[:space:]]and[[:space:]]Implicit[[:space:]]Human[[:space:]]Factors[[:space:]]in[[:space:]]Personas/6668d79b-88a2-4106-85c6-699449d305ba_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Modeling[[:space:]]News[[:space:]]Interactions[[:space:]]and[[:space:]]Influence[[:space:]]for[[:space:]]Financial[[:space:]]Market[[:space:]]Prediction/732af6ab-b36c-460b-a8d1-7646ff31bff5_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Molecular[[:space:]]Facts_[[:space:]]Desiderata[[:space:]]for[[:space:]]Decontextualization[[:space:]]in[[:space:]]LLM[[:space:]]Fact[[:space:]]Verification/2bc96a0c-7189-45f0-9748-c50dba7c8e98_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/MoleculeQA_[[:space:]]A[[:space:]]Dataset[[:space:]]to[[:space:]]Evaluate[[:space:]]Factual[[:space:]]Accuracy[[:space:]]in[[:space:]]Molecular[[:space:]]Comprehension/ad8670ad-ad5d-40a7-a8da-282379fe1329_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Monotonic[[:space:]]Paraphrasing[[:space:]]Improves[[:space:]]Generalization[[:space:]]of[[:space:]]Language[[:space:]]Model[[:space:]]Prompting/2959e73c-f9ae-48a1-8e85-33f3e7a21ae3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/More[[:space:]]Bang[[:space:]]for[[:space:]]your[[:space:]]Context_[[:space:]]Virtual[[:space:]]Documents[[:space:]]for[[:space:]]Question[[:space:]]Answering[[:space:]]over[[:space:]]Long[[:space:]]Documents/f22392c5-4a70-4295-a2be-84f110b6b3b3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-Loss[[:space:]]Fusion_[[:space:]]Angular[[:space:]]and[[:space:]]Contrastive[[:space:]]Integration[[:space:]]for[[:space:]]Machine-Generated[[:space:]]Text[[:space:]]Detection/e63558dd-1191-4591-a960-30434ec75c28_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-Stage[[:space:]]Balanced[[:space:]]Distillation_[[:space:]]Addressing[[:space:]]Long-Tail[[:space:]]Challenges[[:space:]]in[[:space:]]Sequence-Level[[:space:]]Knowledge[[:space:]]Distillation/9db07c5e-dfb6-4143-97ad-d095709fb0c4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-Target[[:space:]]Cross-Lingual[[:space:]]Summarization_[[:space:]]a[[:space:]]novel[[:space:]]task[[:space:]]and[[:space:]]a[[:space:]]language-neutral[[:space:]]approach/e5d21b5b-7d95-4d08-ad43-578c1b31719d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-dimensional[[:space:]]Evaluation[[:space:]]of[[:space:]]Empathetic[[:space:]]Dialogue[[:space:]]Responses/97083bd4-09e8-46e0-a6c5-7547ab0fe3f1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-label[[:space:]]Sequential[[:space:]]Sentence[[:space:]]Classification[[:space:]]via[[:space:]]Large[[:space:]]Language[[:space:]]Model/193123aa-7671-4533-a3ed-cdd99d5595b1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-step[[:space:]]Problem[[:space:]]Solving[[:space:]]Through[[:space:]]a[[:space:]]Verifier_[[:space:]]An[[:space:]]Empirical[[:space:]]Analysis[[:space:]]on[[:space:]]Model-induced[[:space:]]Process[[:space:]]Supervision/41d27a03-a0cb-4c4e-92eb-7c30ca39321a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multi-trait[[:space:]]User[[:space:]]Simulation[[:space:]]with[[:space:]]Adaptive[[:space:]]Decoding[[:space:]]for[[:space:]]Conversational[[:space:]]Task[[:space:]]Assistants/ed1b1981-6502-4cfd-a9de-8b0e0acc7530_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/MultiAgent[[:space:]]Collaboration[[:space:]]Attack_[[:space:]]Investigating[[:space:]]Adversarial[[:space:]]Attacks[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Collaborations[[:space:]]via[[:space:]]Debate/aea5cd86-c948-4094-809a-af33a19178f6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/MultiSkill_[[:space:]]Evaluating[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models[[:space:]]for[[:space:]]Fine-grained[[:space:]]Alignment[[:space:]]Skills/198b7427-20c2-4e4d-ae40-10c47a8c8e4d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/MultiVerse_[[:space:]]Efficient[[:space:]]and[[:space:]]Expressive[[:space:]]Zero-Shot[[:space:]]Multi-Task[[:space:]]Text-to-Speech/8308671c-27c9-4ecc-b8ea-b4d0bce52203_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multilingual[[:space:]]Contrastive[[:space:]]Decoding[[:space:]]via[[:space:]]Language-Agnostic[[:space:]]Layers[[:space:]]Skipping/56d3c029-79e0-4771-8e9d-338b70d262e2_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multilingual[[:space:]]Fine-Grained[[:space:]]News[[:space:]]Headline[[:space:]]Hallucination[[:space:]]Detection/e2423b3a-2cde-4677-85a4-430c5c957303_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multilingual[[:space:]]Synopses[[:space:]]of[[:space:]]Movie[[:space:]]Narratives_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Vision-Language[[:space:]]Story[[:space:]]Understanding/9aed1e75-1d73-4350-ba68-410f2c7334c0_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multimodal[[:space:]]Misinformation[[:space:]]Detection[[:space:]]by[[:space:]]Learning[[:space:]]from[[:space:]]Synthetic[[:space:]]Data[[:space:]]with[[:space:]]Multimodal[[:space:]]LLMs/53f17d1b-197c-48ae-984c-5644fc21e469_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multimodal[[:space:]]Procedural[[:space:]]Planning[[:space:]]via[[:space:]]Dual[[:space:]]Text-Image[[:space:]]Prompting/25884307-7540-4ba2-a5e4-3f1acf6c48ed_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Multiple[[:space:]]Knowledge-Enhanced[[:space:]]Interactive[[:space:]]Graph[[:space:]]Network[[:space:]]for[[:space:]]Multimodal[[:space:]]Conversational[[:space:]]Emotion[[:space:]]Recognition/75e664c7-e3dd-49a6-b9ec-4787dde8b12c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/NALA_[[:space:]]an[[:space:]]Effective[[:space:]]and[[:space:]]Interpretable[[:space:]]Entity[[:space:]]Alignment[[:space:]]Method/51ce5990-ddf9-4220-b2ff-1c2639d060f3_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/NCPrompt_[[:space:]]NSP-Based[[:space:]]Prompt[[:space:]]Learning[[:space:]]and[[:space:]]Contrastive[[:space:]]Learning[[:space:]]for[[:space:]]Implicit[[:space:]]Discourse[[:space:]]Relation[[:space:]]Recognition/b0ad590e-2d30-43fe-bc83-6de9b9b17ef1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Narrative-of-Thought_[[:space:]]Improving[[:space:]]Temporal[[:space:]]Reasoning[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]via[[:space:]]Recounted[[:space:]]Narratives/0f82f67f-0094-4d69-aca6-ddf2aa3a787f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Natural[[:space:]]Evolution-based[[:space:]]Dual-Level[[:space:]]Aggregation[[:space:]]for[[:space:]]Temporal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Reasoning/13b3eb5a-6f94-41a1-a4ad-a2893c0c7dbb_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Navigating[[:space:]]Hallucinations[[:space:]]for[[:space:]]Reasoning[[:space:]]of[[:space:]]Unintentional[[:space:]]Activities/30abdbd6-be24-4cb1-a293-1e063f7a5919_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Navigating[[:space:]]Noisy[[:space:]]Feedback_[[:space:]]Enhancing[[:space:]]Reinforcement[[:space:]]Learning[[:space:]]with[[:space:]]Error-Prone[[:space:]]Language[[:space:]]Models/04df0c29-eeca-4ca2-9c9e-79325fbdaacd_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Navigating[[:space:]]the[[:space:]]Nuances_[[:space:]]A[[:space:]]Fine-grained[[:space:]]Evaluation[[:space:]]of[[:space:]]Vision-Language[[:space:]]Navigation/b14259ee-6462-4cca-a5ce-6ee05aff67d1_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Navigating[[:space:]]the[[:space:]]Shortcut[[:space:]]Maze_[[:space:]]A[[:space:]]Comprehensive[[:space:]]Analysis[[:space:]]of[[:space:]]Shortcut[[:space:]]Learning[[:space:]]in[[:space:]]Text[[:space:]]Classification[[:space:]]by[[:space:]]Language[[:space:]]Models/83eaae07-2745-4d8e-86e9-46df4dfc9d79_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Nebula_[[:space:]]A[[:space:]]discourse[[:space:]]aware[[:space:]]Minecraft[[:space:]]Builder/533a684e-101f-4fd4-8318-cb076f58e90d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Negating[[:space:]]Negatives_[[:space:]]Alignment[[:space:]]with[[:space:]]Human[[:space:]]Negative[[:space:]]Samples[[:space:]]via[[:space:]]Distributional[[:space:]]Dispreference[[:space:]]Optimization/9076fa29-055f-400d-b124-0c2265a7c1c8_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/NegotiationToM_[[:space:]]A[[:space:]]Benchmark[[:space:]]for[[:space:]]Stress-testing[[:space:]]Machine[[:space:]]Theory[[:space:]]of[[:space:]]Mind[[:space:]]on[[:space:]]Negotiation[[:space:]]Surrounding/6b096878-a555-47b4-ae7b-cafb5dee7757_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/NeuroMax_[[:space:]]Enhancing[[:space:]]Neural[[:space:]]Topic[[:space:]]Modeling[[:space:]]via[[:space:]]Maximizing[[:space:]]Mutual[[:space:]]Information[[:space:]]and[[:space:]]Group[[:space:]]Topic[[:space:]]Regularization/9194e6b8-c3dc-4d78-b0ad-e7bee656a8b4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/NormTab_[[:space:]]Improving[[:space:]]Symbolic[[:space:]]Reasoning[[:space:]]in[[:space:]]LLMs[[:space:]]Through[[:space:]]Tabular[[:space:]]Data[[:space:]]Normalization/735efa1d-7b38-489c-bf9b-37b16602f211_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Normalized[[:space:]]Narrow[[:space:]]Jump[[:space:]]To[[:space:]]Conclusions_[[:space:]]Normalized[[:space:]]Narrow[[:space:]]Shortcuts[[:space:]]for[[:space:]]Parameter[[:space:]]Efficient[[:space:]]Early[[:space:]]Exit[[:space:]]Transformer[[:space:]]Prediction/a7bffdfe-67af-49ac-9dd3-c8a839af249b_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Not[[:space:]](yet)[[:space:]]the[[:space:]]whole[[:space:]]story_[[:space:]]Evaluating[[:space:]]Visual[[:space:]]Storytelling[[:space:]]Requires[[:space:]]More[[:space:]]than[[:space:]]Measuring[[:space:]]Coherence,[[:space:]]Grounding,[[:space:]]and[[:space:]]Repetition/ad276bf0-a085-4579-beff-e0b201c0d14d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Not[[:space:]]All[[:space:]]Preference[[:space:]]Pairs[[:space:]]Are[[:space:]]Created[[:space:]]Equal_[[:space:]]A[[:space:]]Recipe[[:space:]]for[[:space:]]Annotation-Efficient[[:space:]]Iterative[[:space:]]Preference[[:space:]]Learning/0ab621fd-5cce-44df-8b62-8e24262c0be6_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Numbers[[:space:]]Matter![[:space:]]Bringing[[:space:]]Quantity-awareness[[:space:]]to[[:space:]]Retrieval[[:space:]]Systems/cdf2f720-0a8b-4b4f-b9d2-054d4ea15a53_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/OEE-CFC_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Open[[:space:]]Event[[:space:]]Extraction[[:space:]]from[[:space:]]Chinese[[:space:]]Financial[[:space:]]Commentary/36595aba-a1fd-464b-8cfd-9b85690d35f4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/OffsetBias_[[:space:]]Leveraging[[:space:]]Debiased[[:space:]]Data[[:space:]]for[[:space:]]Tuning[[:space:]]Evaluators/11fb8fe5-505c-4c17-8c46-5abad5d9526d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]Creating[[:space:]]an[[:space:]]English-Thai[[:space:]]Code-switched[[:space:]]Machine[[:space:]]Translation[[:space:]]in[[:space:]]Medical[[:space:]]Domain/84414976-b527-403b-88a5-09e1f6839a80_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]Diversified[[:space:]]Preferences[[:space:]]of[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]Alignment/d8d682d8-f790-482d-aff7-f51c7bf8e216_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]Evaluating[[:space:]]Explanation[[:space:]]Utility[[:space:]]for[[:space:]]Human-AI[[:space:]]Decision[[:space:]]Making[[:space:]]in[[:space:]]NLP/2c70b47e-6e65-4bd7-a7fd-e2bca62f1f50_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]Leakage[[:space:]]of[[:space:]]Code[[:space:]]Generation[[:space:]]Evaluation[[:space:]]Datasets/9a10b671-39c5-465a-86c3-889c28d9f531_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]Empirical[[:space:]]Complexity[[:space:]]of[[:space:]]Reasoning[[:space:]]and[[:space:]]Planning[[:space:]]in[[:space:]]LLMs/1c7a021f-cbb1-4d8d-a5e4-8f66e72fc434_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]Generalization[[:space:]]of[[:space:]]Training-based[[:space:]]ChatGPT[[:space:]]Detection[[:space:]]Methods/86c7c0cc-a768-4653-8794-12082a73248d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]Limited[[:space:]]Generalization[[:space:]]Capability[[:space:]]of[[:space:]]the[[:space:]]Implicit[[:space:]]Reward[[:space:]]Model[[:space:]]Induced[[:space:]]by[[:space:]]Direct[[:space:]]Preference[[:space:]]Optimization/de4ba516-9f63-418e-903f-6602d9b5265d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]Rigour[[:space:]]of[[:space:]]Scientific[[:space:]]Writing_[[:space:]]Criteria,[[:space:]]Analysis,[[:space:]]and[[:space:]]Insights/b623ae76-2f6f-4b57-a55e-99a713206b4e_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]Similarity[[:space:]]of[[:space:]]Circuits[[:space:]]across[[:space:]]Languages_[[:space:]]a[[:space:]]Case[[:space:]]Study[[:space:]]on[[:space:]]the[[:space:]]Subject-verb[[:space:]]Agreement[[:space:]]Task/2f38398e-d512-428d-bf79-d1f7f9451556_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/On[[:space:]]the[[:space:]]token[[:space:]]distance[[:space:]]modeling[[:space:]]ability[[:space:]]of[[:space:]]higher[[:space:]]RoPE[[:space:]]attention[[:space:]]dimension/0fc27384-969e-4233-9b01-689574c47f5c_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/One[[:space:]]Model[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need_[[:space:]]ByT5-Sanskrit,[[:space:]]a[[:space:]]Unified[[:space:]]Model[[:space:]]for[[:space:]]Sanskrit[[:space:]]NLP[[:space:]]Tasks/5fa3a22d-4114-4f7f-b35c-e8f8812cb04f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/One-to-many[[:space:]]testing[[:space:]]for[[:space:]]code[[:space:]]generation[[:space:]]from[[:space:]](just)[[:space:]]natural[[:space:]]language/583ed9ae-c177-40c7-9ab1-ef23b8b387e4_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/OneGen_[[:space:]]Efficient[[:space:]]One-Pass[[:space:]]Unified[[:space:]]Generation[[:space:]]and[[:space:]]Retrieval[[:space:]]for[[:space:]]LLMs/fa8be84f-4c8b-4455-aaa0-8f9ddcceed65_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Open-RAG_[[:space:]]Enhanced[[:space:]]Retrieval[[:space:]]Augmented[[:space:]]Reasoning[[:space:]]with[[:space:]]Open-Source[[:space:]]Large[[:space:]]Language[[:space:]]Models/db9a4467-d4c9-4821-95eb-103d052a6e8a_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/OpenGraph_[[:space:]]Towards[[:space:]]Open[[:space:]]Graph[[:space:]]Foundation[[:space:]]Models/7f9da7f2-8ffb-4099-874f-68f9cad1e14f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Optimize[[:space:]]Weight[[:space:]]Rounding[[:space:]]via[[:space:]]Signed[[:space:]]Gradient[[:space:]]Descent[[:space:]]for[[:space:]]the[[:space:]]Quantization[[:space:]]of[[:space:]]LLMs/22c9860c-931f-47de-a316-dd045b9260f9_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Optimizing[[:space:]]Instruction[[:space:]]Synthesis_[[:space:]]Effective[[:space:]]Exploration[[:space:]]of[[:space:]]Evolutionary[[:space:]]Space[[:space:]]with[[:space:]]Tree[[:space:]]Search/3be52074-4e5d-4ca1-a1a0-b05d9b4b7f4f_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/Out-of-Distribution[[:space:]]Detection[[:space:]]through[[:space:]]Soft[[:space:]]Clustering[[:space:]]with[[:space:]]Non-Negative[[:space:]]Kernel[[:space:]]Regression/8f4aaf7f-e313-4796-9327-d3305a3cfe81_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/P-FOLIO_[[:space:]]Evaluating[[:space:]]and[[:space:]]Improving[[:space:]]Logical[[:space:]]Reasoning[[:space:]]with[[:space:]]Abundant[[:space:]]Human-Written[[:space:]]Reasoning[[:space:]]Chains/8892565a-7476-4e80-9700-9f1222877527_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/PDF-to-Tree_[[:space:]]Parsing[[:space:]]PDF[[:space:]]Text[[:space:]]Blocks[[:space:]]into[[:space:]]a[[:space:]]Tree/322ddfd2-fbe6-4ed1-9080-0cda8e611b54_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/PEDANTS_[[:space:]]Cheap[[:space:]]but[[:space:]]Effective[[:space:]]and[[:space:]]Interpretable[[:space:]]Answer[[:space:]]Equivalence/7c49d45d-a7db-4446-968e-f757c40b3048_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/PE_[[:space:]]A[[:space:]]Poincare[[:space:]]Explanation[[:space:]]Method[[:space:]]for[[:space:]]Fast[[:space:]]Text[[:space:]]Hierarchy[[:space:]]Generation/a0ad26e8-ebf7-4bc4-ba47-a12992e6ea3d_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/PFA-ERC_[[:space:]]Psuedo-Future[[:space:]]Augmented[[:space:]]Dynamic[[:space:]]Emotion[[:space:]]Recognition[[:space:]]in[[:space:]]Conversations/700b4ffc-3557-4a86-82fc-60be2e0f1f06_origin.pdf filter=lfs diff=lfs merge=lfs -text
+2024/PKAD_[[:space:]]Pretrained[[:space:]]Knowledge[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]to[[:space:]]Detect[[:space:]]and[[:space:]]Mitigate[[:space:]]Textual[[:space:]]Backdoor[[:space:]]Attacks/3e14a699-33a9-4d1a-98b5-db52a4bddcf8_origin.pdf filter=lfs diff=lfs merge=lfs -text

2024/Modeling News Interactions and Influence for Financial Market Prediction/732af6ab-b36c-460b-a8d1-7646ff31bff5_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:935e360a7f2c6786b03a14f4bf00db03b67b1d9c1a134d04012ac272463790fc
+size 459819

2024/Modeling News Interactions and Influence for Financial Market Prediction/full.md

@@ -0,0 +1,382 @@
+# Modeling News Interactions and Influence for Financial Market Prediction
+
+Mengyu Wang Shay B. Cohen Tiejun Ma  
+School of Informatics, The University of Edinburgh, {mengyu.wang, scohen, tiejun.ma}@ed.ac.uk
+
+# Abstract
+
+The diffusion of financial news into market prices is a complex process, making it challenging to evaluate the connections between news events and market movements. This paper introduces FININ (Financial Interconnected News Influence Network), a novel market prediction model that captures not only the links between news and prices but also the interactions among news items themselves. FININ effectively integrates multi-modal information from both market data and news articles. We conduct extensive experiments on two datasets, encompassing the S&P 500 and NASDAQ 100 indices over a 15-year period and over 2.7 million news articles. The results demonstrate FININ's effectiveness, outperforming advanced market prediction models with an improvement of 0.429 and 0.341 in the daily Sharpe ratio for the two markets respectively. Moreover, our results reveal insights into the financial news, including the delayed market pricing of news, the long memory effect of news, and the limitations of financial sentiment analysis in fully extracting predictive power from news data.
+
+# 1 Introduction
+
+Media coverage and news events have been shown to influence market returns (Tetlock, 2007; Sattler, 2013). However, the process through which news diffuses into market prices is complex. The influence of news items on market outcomes varies, affected by various market conditions (Cheung et al., 2019; Hirshleifer and Sheng, 2022). Since existing studies often adopt a simplified approach by treating available news data holistically and investigating its overall effect on the market (Wang et al., 2018; Lopez-Lira and Tang, 2023), the nuanced information contained within individual news items is overlooked. Consequently, parsing the market information diffusion process in detail by considering the impact of each news item emerges as a potential approach to enhance market prediction.
+
+![](images/e3b211b6068cda3ca8644db59fc28881ca1f3efcac5370769889af03e319428f.jpg)
+
+![](images/7268b9c104d2b522c2e3a6cd08f96c8a16eb31625df48c9255a091fc9817dfe5.jpg)  
+Figure 1: News importance fluctuates when considered in the context of other news. Breaking news (News 2) may diminish the market impact of another news item (News 1). However, if related breaking news (News 3) is considered, the influence of News 1 could resurge.
+
+![](images/92d0f69fc7f6b3240ccda682d8e4bb07c9f42f945dcb9efc17d3386c12eaa070.jpg)
+
+![](images/d8b0c07abd1dac2abac91a3f38c4b7b43f5a4849af466e35f51547fc261e2bb8.jpg)
+
+However, evaluating the influence of individual news items is challenging. News does not exist in isolation but is connected with other information. Studies indicate that information diffusion within the market is a gradual process (Albert Jr and Smaby, 1996; Kerl and Walter, 2007). Financial news exhibits a "long memory effect", a persistent impact over time (Ho et al., 2013). During this dissemination process, the influence of an individual news item can be further modulated by interactions with other news items, as illustrated by the practical example in Figure 1. Given the sheer volume of news, evaluating the market impact of news items becomes even more challenging.
+
+This paper aims to address this gap. The domain of financial news presents an excellent testbed for not only improving market prediction but also enhancing our understanding of how to effectively leverage multi-modal data. Modeling news interactions requires a more comprehensive use and exploration of multi-modal data in the form of both numerical and textual data from market and news sources. It also requires considering established market theories (see §2), such as the efficient market hypothesis and information diffusion theory.
+
+We develop FININ, a Financial Interconnected News Influence Network to build interactions between news reports and analyze their diverse impact on markets. The model comprises two key
+
+components: a data fusion encoder and a market-aware influence quantifier. The data fusion encoder handles multi-modal data by encoding various inputs differently while maintaining consistency for data within the same domain. It encodes information related to each news report as independent features, providing a foundation for modeling complex information interactions. The market-aware influence quantifier draws upon financial theories to analyze relationships both between news items themselves and between news and prices. It evaluates the market influence of each news item by considering the overall market context.
+
+Experiments on two major markets, S&P 500 and NASDAQ 100, demonstrate the superiority of FININ over state-of-the-art methods. Using a GPT-based method (Lopez-Lira and Tang, 2023) as a baseline, we show that large language models (LLMs) cannot be directly used for managing large volumes of news for market prediction. While LLMs perform well on financial sentiment analysis, FININ demonstrates a better ability to build relationships between news and market movements.
+
+Additionally, our results provide three key insights into the market information diffusion process. Firstly, FININ highlights a delay in the market pricing of news, indicating market inefficiency in incorporating news information. Secondly, we corroborate the long memory effect of news (Lillo et al., 2015; Ho et al., 2013). Thirdly, we identify the limitations of financial news sentiment analysis and describe how FININ overcomes these limitations. These findings enhance our understanding of the market's response to news. To summarize, our main contributions are:
+
+1 We propose FININ, a multi-modal model that quantifies the market impact of individual news items by considering the interactions among news and the links between news and prices.  
+2 The FININ model demonstrates its effectiveness in leveraging news information for market prediction, outperforming state-of-the-art methods with improvements of at least 0.429 and 0.341 in the daily risk adjusted returns (Sharpe ratio) for the S&P 500 and NASDAQ 100 respectively.  
+3 FININ's findings have revealed three key insights: the delayed market pricing of news, the long memory effect of news, and the limitations of financial news sentiment analysis in extracting predictive power from news data.
+
+# 2 Related Work
+
+News-based Market Prediction. Financial news has been proven to significantly affect market movements, impacting investor sentiment and decision-making (Tetlock, 2007; Barber and Odean, 2008; Calomiris and Mamaysky, 2019). Advancements in data mining and natural language processing (NLP) have enabled the analysis of news data, including tweets and online news, for market prediction (Dewally, 2003; Li, 2010; Xu and Cohen, 2018). The extraction of events and sentiments from news text has progressed substantially (Xu et al., 2021; Hu et al., 2019). Furthermore, progress in deep learning has led to the development of effective neural network structures, such as multimodal LSTM models (Li et al., 2020; Dong et al., 2020), ensemble models (Li and Pan, 2022), and hierarchical attention models (Luo et al., 2023), which leverage news sentiments for stock prediction. However, existing research often focuses on the market influence of overall news, neglecting the nuanced information and varying impacts of individual news items. While some studies recognize the diverse relationships between news and prices (Huynh and Smith, 2017) and attempt to quantify the varying market impact of news (Wang and Ma, 2024), fewer have explored the interactions between multiple news items. Additionally, more complex aspects related to the news impact on market, such as the information diffusion process (Bekiros et al., 2017), are not fully considered during the prediction process.
+
+Market Information Diffusion. The Efficient Markets Hypothesis (EMH) posits that stock prices instantaneously reflect all available information, leaving no room for excess returns (Fama, 1970). However, real-world complexities challenge this ideal situation. In contrast, the information diffusion hypothesis (IDH) suggests that news-induced information is gradually incorporated into prices, leading to delayed market reactions (Kerl and Walter, 2007; Zhang et al., 2016). Research highlights the enduring impact of news on market prices over an extended period (Ray and Tsay, 2000; Christensen and Nielsen, 2007), requiring the analysis of the relationship between news and price changes. Moreover, news items are not isolated in their effect on the market but are interconnected, with their market impact influenced by one another (Yu et al., 2019; Agarwal et al., 2019). Therefore, effectively capturing the complexities of the information diffu
+
+sion process remains a challenge, requiring further research to develop models that capture comprehensive and detailed market impact of news items.
+
+# 3 Preliminaries
+
+# 3.1 Notation and Problem Formulation
+
+In our approach, we define  $d$  as a trading day. Our data consists of daily market data  $M_{d}$  and news report data  $\pmb{R}_{d}$ . Both data sources contain textual information (denoted by  $*^{(t)})$  and numerical information (denoted by  $*^{(n)})$ . The market data,  $\pmb{M}_{d} = (\pmb{m}_{d}^{(t)},\pmb{m}_{d}^{(n)})$ , consists of textual information  $\pmb{m}_{d}^{(t)}$ , including market descriptions and a list of most related companies, and numerical information  $\pmb{m}_{d}^{(n)}$ , comprising daily market prices. The daily news data  $\pmb{R}_{d} = (\pmb{R}_{d,1},\pmb{R}_{d,2},\dots,\pmb{R}_{d,N_{d}})$  contains multiple news reports, where each  $\pmb{R}_{d,i}$  represents information associated with the  $i$ -th news item on day  $d$ , and  $N_{d}$  represents the news volume on day  $d$ . Each news report  $\pmb{R}_{d,i} = (\pmb{r}_{d,i}^{(t)},\pmb{r}_{d,i}^{(n)})$  comprises textual information  $\pmb{r}_{d,i}^{(t)}$ , which includes the news headline, and numerical information  $\pmb{r}_{d,i}^{(n)}$ , representing the news market sentiment scores.
+
+This paper focuses on daily market prediction. The target variable  $y_{d}$  represents the daily market trend, defined as a binary value where 1 signifies a price increase and 0 indicates no change or a decrease between day  $d$  and day  $d + 1$ . Our objective is to effectively use the information from the prior  $t$  days up to day  $d$ , including both market data  $(M_{d - t + 1},\dots,M_{d - 1},M_d)$  and news data  $(R_{d - t + 1},\dots,R_{d - 1},R_d)$ , to predict  $y_{d}$ .
+
+# 3.2 Data
+
+We use two major stock market datasets: the Standard and Poor's 500 (S&P 500) Index and the NASDAQ 100 Index. Additionally, we use the Thomson Reuters News Analytics (TRNA) dataset, comprising over 2.7 million news items. The volume of our news data is significantly larger, at least 10 times greater, than those used in previous studies (Mohan et al., 2019; Lopez-Lira and Tang, 2023; Luo et al., 2023). Furthermore, unlike their online-collected news data, our dataset is sourced from industry practice, ensuring higher quality and more comprehensive market information. Detailed information about these datasets is presented in  $\S A$
+
+We collected daily S&P 500 and NASDAQ 100 prices from 2003 to 2018 as market numerical data  $\pmb{m}_{d}^{(n)}$ . Also, we collected the descriptions of
+
+these two markets and their constituent companies' names as textual data  $\boldsymbol{m}_d^{(t)}$ . For TRNA dataset, we consider headlines as news textual data  $\boldsymbol{r}_{d,i}^{(t)}$  and the sentiment scores as news numerical data  $\boldsymbol{r}_{d,i}^{(n)}$ .
+
+# 4 Method
+
+We have designed the FININ model to evaluate the impact of news on price changes. The FININ model comprises two primary components: a data fusion encoder and a market-aware influence quantifier. By leveraging both textual and numerical data, FININ can capture the mutual influence between these disparate pieces of information. Figure 2 illustrates an overview of the FININ model.
+
+# 4.1 Data Fusion Encoder
+
+Existing news-based market prediction methods primarily focus on the relationship between overall news and price changes, leveraging either market sentiment analysis (Wang et al., 2018; Jing et al., 2021; Lopez-Lira and Tang, 2023) or news text processing (Xu and Cohen, 2018; Luo et al., 2023). However, these approaches have two main limitations. First, sentiment analysis alone may fail to capture nuanced semantic information in the news text, while text processing may lack insights into the market influence conveyed by sentiment scores. Second, these methods treat all news for one prediction as a single entity, potentially overlooking the varying relationships between individual news items and market movements.
+
+In contrast, our proposed FININ model treats each news report as a distinct unit and uses both types of data, capturing more comprehensive market-related information conveyed by individual news reports. The data fusion encoder tackles the multi-modal inputs by encoding various data types from market conditions and news items into meaningful features. It first encodes inputs into features individually and then integrates features from the same input unit (market snapshot or news item) into fusion features.
+
+The core idea of our encoder is to design encoding functions tailored to the nature of each data source, ensuring consistent and differentiated processing. For textual data from market descriptions  $\boldsymbol{m}_d^{(t)}$  and news headlines  $\boldsymbol{r}_{d,i}^{(t)}$ , we use the same encoding function to create unified semantic representations across different text sources. We leverage a pre-trained language model (LM), denoted by  $lm(\cdot)$ , coupled with an encoding layer  $e^{(t)}(\cdot)$ ,
+
+![](images/6f8c460c224f81124c808e1d4794898543d7efa47699737e4f1916488773c29d.jpg)  
+Figure 2: An overview of the FININ model. It comprises two components: a Data Fusion Encoder and a Market-aware Influence Quantifier. The annotations are defined and explained in the corresponding parts of Section 4
+
+to extract market and news text features  $\pmb{m}_{d}^{(te)}$  and  $\pmb{r}_{d,i}^{(te)}$ , respectively:
+
+$$
+\boldsymbol {m} _ {d} ^ {(t e)} = e ^ {(t)} \left(l m \left(\boldsymbol {m} _ {d} ^ {(t)}\right)\right), \tag {1}
+$$
+
+$$
+\boldsymbol {r} _ {d, i} ^ {(t e)} = e ^ {(t)} \left(l m \left(\boldsymbol {r} _ {d, i} ^ {(t)}\right)\right). \tag {2}
+$$
+
+The pre-trained LM is frozen, while the added layer  $e^{t}(\cdot)$  adapts the representations for our prediction task. We explore various LMs for financial text, which will be discussed in detail in § 5.
+
+For numerical data, due to their different information content and input sizes, market data  $\pmb{m}_{d}^{(n)}$  and news sentiment scores  $\pmb{r}_{d,i}^{(n)}$  are encoded using separate functions  $e_1^{(n)}(\cdot)$  and  $e_2^{(n)}(\cdot)$ , respectively. This ensures tailored processing while maintaining consistency in encoding numerical data from different news items. The resulting features are denoted by  $\pmb{m}_{d}^{(ne)}$  and  $\pmb{r}_{d,i}^{(ne)}$  for market and news data, respectively, as shown in the equation:
+
+$$
+\boldsymbol {m} _ {d} ^ {(n e)} = e _ {1} ^ {(n)} \left(\boldsymbol {m} _ {d} ^ {(n)}\right), \boldsymbol {r} _ {d, i} ^ {(n e)} = e _ {2} ^ {(n)} \left(\boldsymbol {r} _ {d, i} ^ {(n)}\right). \tag {3}
+$$
+
+To integrate textual and numerical information into unified encodings, we use separate data fusion functions  $f_{1}(\cdot ,\cdot)$  and  $f_{2}(\cdot ,\cdot)$  for market and news data, respectively. Similar to the processing of numerical data, these functions ensure consistent encoding within news items while distinguishing between the two sources. The functions  $f_{1}(\cdot ,\cdot)$  and  $f_{2}(\cdot ,\cdot)$  are implemented as a concatenation of the textual and numerical input features, followed by a multi-layer perceptron (MLP). They take the textual and numerical features as inputs and produce the integrated encodings  $m_d^{(e)}$  and  $r_{d,i}^{(e)}$  for market
+
+and news items, respectively:
+
+$$
+\boldsymbol {m} _ {d} ^ {(e)} = f _ {1} \left(\boldsymbol {m} _ {d} ^ {(t e)}, \boldsymbol {m} _ {d} ^ {(n e)}\right), \tag {4}
+$$
+
+$$
+\boldsymbol {r} _ {d, i} ^ {(e)} = f _ {2} \left(\boldsymbol {r} _ {d, i} ^ {(t e)}, \boldsymbol {r} _ {d, i} ^ {(n e)}\right). \tag {5}
+$$
+
+By using the multi-modal data as inputs, we obtain features containing more comprehensive information. Moreover, information relevant to each news report is encoded into independent features, enabling an analysis of individual news items.
+
+# 4.2 Market-aware Influence Quantifier
+
+Unlike existing works that treat news data holistically to obtain daily features, FININ uses independent features for each news item to capture market insights. This section describes the Market-aware Influence Quantifier, which leverages financial theories to better reflect real-world market dynamics.
+
+Firstly, the market information diffusion theory (Kerl and Walter, 2007; Zhang et al., 2016) suggests that information pieces influence each other. To capture these interactions, we use an attention layer (Vaswani et al., 2017) to refine news encodings  $\boldsymbol{r}_{d,i}^{(e)}$  into context-aware features  $\boldsymbol{r}_{d,i}^{(f)}$ . These context features still focuses on the  $i$ -th news of day  $d$  but incorporate information from the entire daily news corpus, making each news report "aware" of the others.
+
+Secondly, the efficient market hypothesis posits that all relevant information is reflected in market prices. To account for this connection, we introduce another attention layer. In contrast to the first layer that builds relationships between all input states, this layer focuses solely on the connections between the market state and news items. In this
+
+layer, the market information unit  $\pmb{m}_d^{(e)}$  acts as the query, and all news items serve as key states. We calculate query-key attention scores  $atts_{d,i}$  to quantify the relative importance of each news report in relation to the market:
+
+$$
+a t t s _ {d, i} = \frac {1}{\sqrt {d _ {k}}} \cdot q (\pmb {m} _ {d} ^ {(e)}) \cdot k (\pmb {r} _ {d, i} ^ {(f)}). \qquad (6)
+$$
+
+where  $d_{k}$  denotes the key vector dimension.
+
+These scores are then normalized using a softmax function to obtain attention weights. Notably, we exclude the self-attention score for the market information unit  $\pmb{m}_{d}^{(e)}$ , ensuring that the resulting weights correspond only to news states and sum to one. Finally, we use these weights to create a weighted sum of news features  $\pmb{r}_{d,i}^{(f)}$ , aggregating them into overall daily news features  $\pmb{r}_{d}^{(f)}$ .
+
+By using these two attention layers, we establish relationships between news items themselves and between news and market information. This allows the information states to capture a more comprehensive view of the market, leading to more informative news representations.
+
+# 4.3 Market Prediction
+
+After the data encoding and market influence quantifying stages, the final step of FININ is market prediction. We collect the encoded market and news representations for the past  $t$  days, denoted as  $(\pmb{m}_{d - t + 1}^{(e)},\dots,\pmb{m}_d^{(e)})$  and  $(r_{d - t + 1}^{(f)},\dots,r_{d}^{(f)}))$ , respectively. These representations are fed into a prediction structure, denoted as  $pred(\cdot ,\cdot)$ , which aims at forecasting market trends:
+
+$$
+\begin{array}{l} \hat {y} _ {d} = \operatorname {p r e d} \left(\left(\boldsymbol {m} _ {d - t + 1} ^ {(e)}, \dots , \boldsymbol {m} _ {d} ^ {(e)}\right), \right. \tag {7} \\ \left(\boldsymbol {r} _ {d - t + 1} ^ {(f)}, \dots , \boldsymbol {r} _ {d} ^ {(f)}\right)). \\ \end{array}
+$$
+
+The predictor architecture in FININ is flexible, allowing integration with various widely used stock prediction structures, such as Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks (Jing et al., 2021; Jiang, 2021). Our experiments suggest that the informative features generated by FININ enhance performance regardless of the specific structure. Given the frequent use of MLPs, as subsequent layers due to their simplicity and efficiency (Orimoloye et al., 2020; Hu et al., 2020), we adopt an MLP for constructing the final FININ model to maintain simplicity and effectiveness.
+
+# 5 Experiment Settings
+
+# 5.1 Language Models
+
+Financial language differs from general text due to its area-specific vocabulary and domain knowledge (Araci, 2019). While numerous pre-trained LMs exist, their performance on general tasks may not accurately reflect their effectiveness for financial tasks. To investigate how different LMs affect the modeling of market information interactions, we experiment with four different types of LMs within the FININ model for textual data encoding. These LMs include RoBERTa, FinBERT, BGE, Llama2, representing pre-trained encoder models, domain-specific models, top-performing text embedding models, and large language models, respectively. Detailed introductions of these models are presented in §B.1.
+
+# 5.2 Implementation Details
+
+To construct the architecture presented in Figure 2, the function  $e^{(t)}$  is implemented as a fully connected layer, while the functions  $e_1^{(n)}, e_2^{(n)}, f_1$ , and  $f_2$  are implemented using Multi-Layer Perceptron (MLP) structures. The number of layers in each MLP is selected from  $\{2,3,4\}$ , with hidden layer sizes chosen from  $\{16,32,64\}$ . Both attention mechanisms consist of a single layer, with the number of attention heads chosen from  $\{1,3,6\}$ . The hidden dimensions for the query, key, and value vectors in the attention layers are selected from  $\{32,64,128\}$ . A grid search is performed on the validation set to identify the optimal combination of hyper-parameters.
+
+We conducted experiments using both 2080Ti and A100 GPUs. Training a model requires 32 hours on a 2080Ti GPU or 7.5 hours on an A100 GPU. Once trained, the models can handle test cases spanning over a month, requiring only 10.16 seconds for a single prediction. Given the common usage of A100 and superior GPUs in the industry, our computing requirements are practical for real-world applications.
+
+# 5.3 Baseline Methods
+
+We evaluate FININ against seven advanced news-based market prediction approaches, representing various techniques for leveraging financial news to predict market movements:
+
+News Appearance Frequency (NAF) (Huynh and Smith, 2017) quantifies news influence based on its frequency of occurrence.
+
+TRNA-based Sentiment Indicator (Wang et al., 2018; Mohan et al., 2019; Jing et al., 2021): Many works create market sentiment indicators from news sentiments and use deep learning structures like MLP and LSTM to predict market trends. We experiment with different sentiment indicators and prediction structures based on sentiments from our TRNA dataset to cover these methods. We report the best results as representative of these methods.
+
+FinBERT-based Sentiment Indicator (Liu et al., 2021): We obtain sentiment scores using FinBERT and use the same methods as the previous baseline to obtain results.
+
+Ensemble Deep Learning Model (Li and Pan, 2022) integrates sentiment analysis with a two-level blending ensemble model, effectively capturing time series events.
+
+FPT (Zhou et al., 2023) is an advanced pre-training approach for time series analysis, achieving state-of-the-art results across various tasks.
+
+GPT (Lopez-Lira and Tang, 2023): A recent study examines the potential of LLMs in predicting market returns using news headlines. They found GPT models obtain the highest Sharpe ratio.
+
+PEN (Li et al., 2023): PEN uses both textual and price data for market prediction with explainability.
+
+# 5.4 Training and Validation
+
+Our experimental setup considers input data time spans of 1, 3, 5, 10 and 20 trading days, with the 5-day and 20-day settings equivalent to weekly and monthly predictions. Time series cross-validation is used to determine optimal hyper-parameters. The dataset is divided into 10 sliding windows based on dates. Each window contains data from 500 consecutive days. We initiate by collecting the first window from the initial day and then progress by 391 days to obtain the second window. In each window, data is allocated into training, validation, and testing sets in an 8:1:1 ratio, following the chronological order. Reported results are averages across the 10 subsets, providing a robust and reliable evaluation of the model's performance across different time periods and market conditions.
+
+# 5.5 Evaluation Metrics
+
+The evaluation of our model is based on three metrics: Accuracy (Acc), Profit and Loss (PnL) and Sharpe ratio (SR). These metrics are commonly used in market prediction literature (Ye et al., 2020; Yuemei et al., 2021). PnL measures the cumulative profit or loss generated by the model's predic
+
+tions, providing a direct financial outcome of its performance. SR measures the risk adjusted returns, quantifying the amount of return achieved per unit of risk. Among them, SR is considered the most important, as it incorporates both returns and risk, which are crucial factors in stock investment. Therefore, when comparing different methods, we prioritize the comparison of SR. PnL serves as a secondary metric. While accuracy is considered, it is given less weight, as high accuracy does not guarantee high profitability. Detailed discussions of PnL and SR are presented in  $\S B.2$
+
+# 6 Results and Discussions
+
+# 6.1 Language Model Comparison
+
+To identify the LM that generates the most suitable text embeddings for market prediction, we trained FININ with different LMs as the text encoder for the S&P 500 market. The results in Table 1 show RoBERTa outperforming other models across most settings. This suggests that general improvements in model design, as seen in RoBERTa, may be more beneficial than domain-specific adaptations (FinBERT) for news-based market prediction. This observation aligns with our intuition that financial terms in news are generally understandable, making domain-specific adaptation less crucial.
+
+While BGE exhibits strong performance on many NLP tasks, its results in our financial context are less impressive, reinforcing the notion that no single embedding dominates all tasks (Muen-nighoff et al., 2022). Llama2, representing advanced LLMs, also shows lower performance. As LLMs are primarily optimized for text generation, our findings reflect the concerns regarding their
+
+<table><tr><td rowspan="2">Time Length</td><td rowspan="2">Mericss</td><td colspan="4">Language Models</td></tr><tr><td>RoBERTa</td><td>FinBERT</td><td>BGE</td><td>Llama2</td></tr><tr><td rowspan="3">1-day</td><td>Acc</td><td>0.518</td><td>0.524</td><td>0.539</td><td>0.539</td></tr><tr><td>PnL</td><td>0.033</td><td>0.040</td><td>0.042</td><td>0.033</td></tr><tr><td>SR</td><td>1.223</td><td>1.170</td><td>0.986</td><td>1.082</td></tr><tr><td rowspan="3">3-day</td><td>Acc</td><td>0.537</td><td>0.524</td><td>0.518</td><td>0.535</td></tr><tr><td>PnL</td><td>0.045</td><td>0.039</td><td>0.046</td><td>0.037</td></tr><tr><td>SR</td><td>1.247</td><td>1.256</td><td>1.300</td><td>1.187</td></tr><tr><td rowspan="3">5-day</td><td>Acc</td><td>0.533</td><td>0.543</td><td>0.524</td><td>0.512</td></tr><tr><td>PnL</td><td>0.032</td><td>0.019</td><td>0.019</td><td>0.037</td></tr><tr><td>SR</td><td>1.192</td><td>0.948</td><td>0.462</td><td>1.131</td></tr><tr><td rowspan="3">10-day</td><td>Acc</td><td>0.531</td><td>0.522</td><td>0.545</td><td>0.543</td></tr><tr><td>PnL</td><td>0.053</td><td>0.054</td><td>0.047</td><td>0.030</td></tr><tr><td>SR</td><td>1.643</td><td>1.465</td><td>1.330</td><td>1.164</td></tr><tr><td rowspan="3">20-day</td><td>Acc</td><td>0.522</td><td>0.520</td><td>0.520</td><td>0.551</td></tr><tr><td>PnL</td><td>0.041</td><td>0.063</td><td>0.041</td><td>0.046</td></tr><tr><td>SR</td><td>1.772</td><td>1.658</td><td>1.432</td><td>1.408</td></tr></table>
+
+Table 1: Results of using different LMs in FININ.
+
+<table><tr><td rowspan="2">Method</td><td colspan="3">1-day</td><td colspan="3">3-day</td><td colspan="3">5-day</td><td colspan="3">10-day</td><td colspan="3">20-day</td></tr><tr><td>Acc</td><td>PnL</td><td>SR</td><td>Acc</td><td>PnL</td><td>SR</td><td>Acc</td><td>PnL</td><td>SR</td><td>Acc</td><td>PnL</td><td>SR</td><td>Acc</td><td>PnL</td><td>SR</td></tr><tr><td colspan="16">Results of the S&amp;P 500 Market (Always Buy strategy results in Acc, PnL, SR of 0.520, 0.039, 0.556, respectively.)</td></tr><tr><td>NAF</td><td>0.522</td><td>0.017</td><td>0.678</td><td>0.520</td><td>0.018</td><td>0.744</td><td>0.520</td><td>-0.015</td><td>-0.076</td><td>0.500</td><td>0.016</td><td>0.480</td><td>0.522</td><td>0.024</td><td>0.644</td></tr><tr><td>Sentiments</td><td>0.518</td><td>0.036</td><td>0.914</td><td>0.541</td><td>0.037</td><td>0.943</td><td>0.524</td><td>0.022</td><td>0.509</td><td>0.525</td><td>0.038</td><td>1.031</td><td>0.554</td><td>0.045</td><td>1.085</td></tr><tr><td>FinBERT</td><td>0.512</td><td>0.010</td><td>0.267</td><td>0.520</td><td>0.016</td><td>0.697</td><td>0.520</td><td>-0.076</td><td>-0.076</td><td>0.518</td><td>0.010</td><td>0.613</td><td>0.529</td><td>0.012</td><td>0.644</td></tr><tr><td>Ensemble</td><td>0.522</td><td>0.018</td><td>0.886</td><td>0.514</td><td>0.031</td><td>0.672</td><td>0.510</td><td>0.031</td><td>0.824</td><td>0.534</td><td>0.031</td><td>1.343</td><td>0.545</td><td>0.028</td><td>0.613</td></tr><tr><td>FPT</td><td>0.508</td><td>0.002</td><td>0.323</td><td>0.524</td><td>0.033</td><td>0.917</td><td>0.520</td><td>0.025</td><td>1.169</td><td>0.528</td><td>0.025</td><td>0.969</td><td>0.547</td><td>0.035</td><td>0.769</td></tr><tr><td>GPT-3.5</td><td>0.522</td><td>0.022</td><td>0.914</td><td>0.530</td><td>0.030</td><td>1.133</td><td>0.528</td><td>0.014</td><td>0.946</td><td>0.522</td><td>0.014</td><td>0.765</td><td>0.504</td><td>0.022</td><td>1.110</td></tr><tr><td>GPT-4</td><td>0.504</td><td>0.032</td><td>0.696</td><td>0.520</td><td>0.008</td><td>0.194</td><td>0.516</td><td>-0.004</td><td>-0.104</td><td>0.484</td><td>-0.004</td><td>0.424</td><td>0.494</td><td>0.012</td><td>0.461</td></tr><tr><td>PEN</td><td>0.504</td><td>0.004</td><td>-0.224</td><td>0.510</td><td>0.018</td><td>0.062</td><td>0.516</td><td>0.034</td><td>0.457</td><td>0.531</td><td>0.048</td><td>1.167</td><td>0.539</td><td>0.015</td><td>0.625</td></tr><tr><td>FININ</td><td>0.518</td><td>0.033</td><td>1.223</td><td>0.537</td><td>0.045</td><td>1.247</td><td>0.533</td><td>0.032</td><td>1.192</td><td>0.531</td><td>0.053</td><td>1.643</td><td>0.522</td><td>0.014</td><td>1.772</td></tr><tr><td colspan="16">Results of the NASDAQ 100 Market (Always Buy strategy results in Acc, PnL, SR of 0.518, 0.043, 0.759, respectively.)</td></tr><tr><td>NAF</td><td>0.537</td><td>0.014</td><td>0.589</td><td>0.527</td><td>0.002</td><td>0.299</td><td>0.555</td><td>0.021</td><td>0.816</td><td>0.541</td><td>0.021</td><td>0.853</td><td>0.522</td><td>0.016</td><td>0.648</td></tr><tr><td>Sentiments</td><td>0.512</td><td>0.025</td><td>0.706</td><td>0.531</td><td>0.022</td><td>0.620</td><td>0.551</td><td>0.007</td><td>0.439</td><td>0.549</td><td>0.031</td><td>0.847</td><td>0.551</td><td>0.040</td><td>1.084</td></tr><tr><td>FinBERT</td><td>0.514</td><td>0.000</td><td>0.012</td><td>0.529</td><td>-0.003</td><td>0.157</td><td>0.557</td><td>0.015</td><td>0.686</td><td>0.545</td><td>0.028</td><td>0.755</td><td>0.539</td><td>0.008</td><td>0.563</td></tr><tr><td>Ensemble</td><td>0.540</td><td>-0.006</td><td>0.107</td><td>0.534</td><td>0.007</td><td>0.628</td><td>0.546</td><td>0.010</td><td>0.627</td><td>0.532</td><td>0.020</td><td>0.985</td><td>0.542</td><td>0.024</td><td>1.095</td></tr><tr><td>FPT</td><td>0.536</td><td>0.009</td><td>0.523</td><td>0.537</td><td>0.003</td><td>0.457</td><td>0.549</td><td>0.009</td><td>0.766</td><td>0.540</td><td>0.017</td><td>0.837</td><td>0.506</td><td>0.003</td><td>0.322</td></tr><tr><td>GPT-3.5</td><td>0.538</td><td>0.010</td><td>0.795</td><td>0.522</td><td>0.023</td><td>0.882</td><td>0.526</td><td>0.015</td><td>0.692</td><td>0.530</td><td>0.017</td><td>0.896</td><td>0.522</td><td>0.009</td><td>0.639</td></tr><tr><td>PEN</td><td>0.504</td><td>-0.001</td><td>-0.043</td><td>0.529</td><td>0.018</td><td>0.572</td><td>0.455</td><td>-0.024</td><td>-0.630</td><td>0.492</td><td>0.025</td><td>0.258</td><td>0.551</td><td>0.045</td><td>1.108</td></tr><tr><td>FININ</td><td>0.541</td><td>0.025</td><td>0.651</td><td>0.559</td><td>0.027</td><td>0.966</td><td>0.557</td><td>0.026</td><td>1.095</td><td>0.571</td><td>0.041</td><td>1.307</td><td>0.553</td><td>0.046</td><td>1.449</td></tr></table>
+
+Table 2: Market prediction results for the S&P 500 and NASDAQ 100 indices. The "Always Buy" strategy is a simple benchmark, assuming that prices will always increase the next day, providing basic results of market performance. Bold numbers indicate the highest value for each metric within a specific setting, while underlined numbers denote the highest PnL and SR across all settings. Results for GPT-4 on the NASDAQ 100 are exclusive due to its comparatively lower performance on the S&P 500 dataset and limited availability compared to GPT-3.5.
+
+limitations in text encoding (Jiang et al., 2023). Additionally, Llama2's large embedding size (4096) compared to the typical size (around 1000) introduces challenges for models of general sizes.
+
+In conclusion, these results highlight the importance of LM selection in effectively encoding market-related textual data. Among the evaluated LMs, RoBERTa exhibited robust overall performance, rendering it the most suitable choice for our FININ model. Subsequent experiments will leverage RoBERTa as the text encoder.
+
+# 6.2 Market Prediction Comparison
+
+Table 2 presents the market prediction results of the FININ model compared to baseline methods. While the rankings of three metrics do not strongly correlate, the SR is considered the most practical and important one, as mentioned in § 5.5. Despite not always achieving the best Acc or PnL, FININ consistently outperforms all baselines in terms of the SR across different settings.
+
+Furthermore, for both datasets, the best PnL and SR results (underlined) are generated by FININ. FININ's highest SR results, 1.772 on the S&P 500 and 1.449 on the NASDAQ 100, outperform the best SR achieved by other methods by at least 0.429 and 0.341, respectively, for each dataset.
+
+Discussion on News Use The comparison highlights FININ's superiority in leveraging extensive news data for market prediction. While baseline
+
+methods also incorporate news information, FININ offers two distinct advantages. Firstly, FININ uses both news sentiments and textual headlines. This contrasts with baseline methods that primarily rely on sentiment analysis or text processing. By incorporating multi-modal inputs, FININ can capture interactions between news items, an aspect that other methods overlook. Secondly, FININ's analysis of individual news items enables efficient management of large volumes of news. Baseline methods treat all news items equally, causing influential news to be overwhelmed by general news when faced with substantial data. FININ encodes each news item individually, allowing them to contribute differently to the prediction based on their varying influence. This capability is crucial when dealing with large news volumes, where lack of proper news aggregation in other methods can even impair predictions (compared to the "Always Buy" strategy). FININ's capability in modeling news interactions is analyzed in detail through case studies in §C. In summary, FININ's advanced input processing and effective news management techniques position it as a superior tool for using financial news data in market prediction tasks.
+
+Discussion on Long Memory Effect and LLM Our findings reveal the long memory effect of financial news. An increasing trend in the Sharpe ratio is evident as the timeframe of considered news data extends. This trend is particularly pronounced
+
+<table><tr><td rowspan="2">Time Length</td><td rowspan="2">Metrics</td><td colspan="4">Model Structure</td></tr><tr><td>Complete FININ</td><td>FININ-NTP-NNP (no news data)</td><td>FININ-MTP-NTP (no textual data)</td><td>FININ-MIQ (no quantifier)</td></tr><tr><td rowspan="3">1-day</td><td>Acc</td><td>0.518</td><td>0.520</td><td>0.543</td><td>0.529</td></tr><tr><td>PnL</td><td>0.033</td><td>0.034</td><td>0.033</td><td>0.015</td></tr><tr><td>SR</td><td>1.223</td><td>1.021</td><td>1.093</td><td>0.741</td></tr><tr><td rowspan="3">3-day</td><td>Acc</td><td>0.537</td><td>0.512</td><td>0.541</td><td>0.512</td></tr><tr><td>PnL</td><td>0.045</td><td>0.028</td><td>0.022</td><td>0.010</td></tr><tr><td>SR</td><td>1.247</td><td>0.905</td><td>0.696</td><td>0.084</td></tr><tr><td rowspan="3">5-day</td><td>Acc</td><td>0.533</td><td>0.506</td><td>0.504</td><td>0.529</td></tr><tr><td>PnL</td><td>0.032</td><td>-0.020</td><td>0.020</td><td>-0.005</td></tr><tr><td>SR</td><td>1.192</td><td>-0.350</td><td>0.435</td><td>-0.231</td></tr><tr><td rowspan="3">10-day</td><td>Acc</td><td>0.531</td><td>0.518</td><td>0.518</td><td>0.512</td></tr><tr><td>PnL</td><td>0.053</td><td>0.025</td><td>0.039</td><td>0.021</td></tr><tr><td>SR</td><td>1.643</td><td>1.120</td><td>1.301</td><td>0.967</td></tr><tr><td rowspan="3">20-day</td><td>Acc</td><td>0.522</td><td>0.537</td><td>0.531</td><td>0.504</td></tr><tr><td>PnL</td><td>0.041</td><td>0.030</td><td>0.037</td><td>0.025</td></tr><tr><td>SR</td><td>1.772</td><td>1.126</td><td>1.480</td><td>1.064</td></tr></table>
+
+Table 3: Ablation study results of FININ.
+
+in FININ's results due to its stable and effective news data use. Baseline methods also suggest improvement with extended timeframes, indicating that including data from a broader range of trading days can enhance the predictions. However, GPT-based methods exhibit an exception. GPT-3.5 and GPT-4 achieve their best results in the 3-day and 1-day prediction settings on the S&P 500 dataset, with performance decreasing when more days are considered. Examining GPT's outputs revealed its limitations in analyzing news for a complex index like the S&P 500. In the GPT-based methods, GPT is queried about the positive, negative, or unknown influence of a news item on a company's stock. However, when applied to our datasets, GPT struggles with analyzing the complexities of the S&P 500 market. GPT appears overly cautious in its responses, with a significant portion of news items (55.38% for GPT-3.5 and 66.78% for GPT-4) classified as "unknown". Consequently, as we consider news from more days, an increasing number of news items contribute nothing to the predictions. Furthermore, the significantly biased news sentiments impair predictions, leading to weak results.
+
+# 6.3 Ablation Studies
+
+To isolate the impact of different data sources and processing methods within FININ, we decompose FININ into five key components: four subsections of the data fusion encoder, the Market Textual data Processor (MTP), the Market Numerical data Processor (MNP), the News Textual data Processor (NTP), and the News Numerical data Processor (NNP), as well as the Market-aware Influence
+
+Quantifier (MIQ). We examined three ablated variants: (1) removing news data (FININ-NTP-NNP), (2) removing text inputs (FININ-MTP-NTP), and (3) removing the influence quantifier (FININ-MIQ). When removing MIQ, we aggregated news features via simple averaging instead of using MIQ's scores.
+
+Table 3 presents the ablation study results on the S&P 500 market. The performance drop in each ablation setup, compared to the complete FININ model, underlines the critical role of news inputs, text data processing, and the news influence quantifier in enhancing market prediction.
+
+The "FININ-NTP-NNP" column reveals that predictions based solely on historical price data offer limited long-term forecasting capability, as performance improves marginally with increasing input timeframes. This suggests that the primary information of daily price fluctuations is incorporated into subsequent prices. Incorporating news information, as in the complete FININ model, improves prediction, especially for longer-term forecasts. This finding highlights the heightened relevance of news for capturing long-term stock price dynamics.
+
+The "FININ-MTP-NTP" outperforms "FININ-NTP-NNP" in most settings, indicating that news sentiment analysis captures additional information not yet reflected in prices. However, the comparison with "Complete FININ" shows that relying solely on sentiment analysis misses crucial information within news text, highlighting the necessity of our text processing component.
+
+The "FININ-MIQ" results emphasize the significant contribution of the news influence quantifier. Directly averaging news features performs even worse than models without news data (FININ-NTP-NNP). This indicates that simply including a large amount of news data can be detrimental without proper use. The market-aware influence quantifier's scores are essential to extract the predictive power of news for market prediction.
+
+In summary, while market data provides baseline results, incorporating news sentiment and textual information through FININ's processing components yields substantial performance gains. Moreover, FININ's market-aware influence quantifier is crucial for leveraging news data, as it effectively integrates the complex inputs and extracts their predictive power for enhanced market prediction.
+
+# 6.4 Insights into Financial News
+
+Our comprehensive experimental results offer three key insights into the interplay between financial
+
+news and market movements, highlighting the need for improved approaches to using news data:
+
+Delayed Market Pricing of News Information: Despite the S&P 500 and NASDAQ 100 being highly efficient markets, our study reveals an evident delay in fully integrating news into asset prices. Results and discussions in §6.2 and §6.3 prove this through improved performance when using data from previous days. This presents an opportunity for investors to exploit market inefficiencies by incorporating news data into their decision-making.
+
+Long Memory Effect of Financial News: Our results across various time horizons highlight the sustained impact of news, aligning with the information diffusion hypothesis. While immediate effects of news are apparent in 1-day predictions, improvement is observed across all tested time settings, indicating a delayed influence that persists rather than being released at a specific timeframe.
+
+Limitations of Market Sentiment Analysis: Relying solely on sentiment analysis, a commonly used approach, may not capture the full complexity of the market's reaction to news. Our findings advocate for a more holistic approach that considers both sentiment and text for a nuanced understanding of news impact on asset prices.
+
+# 7 Conclusions
+
+This paper introduces FININ, a novel model designed to use financial news for market prediction. FININ effectively uses market-related information from news data, outperforming various advanced market prediction methods. Our experiments shed light on the relationships between news and the market, highlighting delayed market pricing of news, the long memory effect of financial news, and the limitations of market sentiment analysis in extracting predictive power from news.
+
+# Limitations
+
+FININ is a novel approach to building news interactions for market prediction. However, as the related financial theory is less considered in deep learning, there is significant room for further improvement. We discuss the primary limitations of our model as follows:
+
+First, similar to many models using attention mechanisms (Achiam et al., 2023; Touvron et al., 2023), FININ is limited by the attention window size. Attention mechanisms cannot be applied to sequences of unlimited length. Consequently,
+
+while it would be ideal to capture detailed interactions among all available news reports, FININ currently focuses on linking news within the same day. The relationship between the information from different days is built by the prediction model. This limitation primarily arises due to the substantial daily news volume (average 702, maximum 1953 per trading day), which poses a challenge for the attention mechanism. Although techniques exist to potentially enlarge the attention window size, this paper prioritizes providing a practical method for building news interactions. Future work can explore methods for capturing long-range news dependencies that extend beyond a single day, which could potentially improve the model's performance.
+
+Second, our proposed FININ model primarily focuses on news data due to the well-established financial theories that demonstrate news' connections to market movements. However, market movements are influenced by a multitude of factors beyond news. Social media information and company reports are just two examples of additional data sources that likely interact with news and price data, but the processing of such datasets is beyond the scope of this paper. FININ has the potential to be extended to incorporate these relevant data sources, potentially leading to a more comprehensive understanding of market information diffusion, price discovery, and predictions.
+
+# Acknowledgments
+
+This work was supported by the UKRI Centre for Doctoral Training (CDT) in Natural Language Processing through UKRI grant EP/S022481/1. We would like to thank Chang Luo, Guojun Wu, Yfthah Ziser, Zheng Zhao, and the anonymous reviewers for their valuable feedback.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report arXiv preprint arXiv:2303.08774.  
+Shweta Agarwal, Shailendra Kumar, and Utkarsh Goel 2019. Stock market response to information diffusion through internet sources: A literature review International Journal of Information Management, 45:118-131.  
+Robert L Albert Jr and Timothy R Smaby. 1996. Market
+
+response to analyst recommendations in the "dartboard" column: the information and price-pressure effects. Review of Financial Economics, 5(1):59-74.  
+Dogu Araci. 2019. FinBERT: Financial sentiment analysis with pre-trained language models. arXiv preprint arXiv:1908.10063.  
+Brad M Barber and Terrance Odean. 2008. All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors. The review of financial studies, 21(2):785-818.  
+Stelios Bekiros, Duc Khuong Nguyen, Leonidas Sandoval Junior, and Gazi Salah Uddin. 2017. Information diffusion, cluster formation and entropy-based network dynamics in equity and commodity markets. European Journal of Operational Research, 256(3):945-961.  
+Charles W Calomiris and Harry Mamaysky. 2019. How news and its context drive risk and returns around the world. Journal of Financial Economics, 133(2):299-336.  
+Yin-Wong Cheung, Rasmus Fatum, and Yohei Yamamoto. 2019. The exchange rate effects of macro news after the global financial crisis. Journal of International Money and Finance, 95:424-443.  
+Bent Jesper Christensen and Morten Ørregaard Nielsen. 2007. The effect of long memory in volatility on stock market fluctuations. *The Review of Economics and Statistics*, 89(4):684-700.  
+Michael Dewally. 2003. Internet investment advice: Investing with a rock of salt. Financial Analysts Journal, 59(4):65-77.  
+Qianggang Ding, Sifan Wu, Hao Sun, Jiadong Guo, and Jian Guo. 2020. Hierarchical multi-scale gaussian transformer for stock movement prediction. In *IJCAI*, pages 4640-4646.  
+Yingzhe Dong, Da Yan, Abdullateef Ibrahim Almu-daifer, Sibo Yan, Zhe Jiang, and Yang Zhou. 2020. Belt: A pipeline for stock price prediction using news. In 2020 IEEE International Conference on Big Data (Big Data), pages 1137-1146. IEEE.  
+Eugene F Fama. 1970. Efficient capital markets: A review of theory and empirical work. The journal of Finance, 25(2):383-417.  
+Thomas Fischer and Christopher Krauss. 2018. Deep learning with long short-term memory networks for financial market predictions. European journal of operational research, 270(2):654-669.  
+David Hirshleifer and Jinfei Sheng. 2022. Macro news and micro news: complements or substitutes? Journal of Financial Economics, 145(3):1006-1024.  
+Kin-Yip Ho, Yanlin Shi, and Zhaoyong Zhang. 2013. How does news sentiment impact asset volatility? evidence from long memory and regime-switching approaches. The North American Journal of Economics and Finance, 26:436-456.
+
+Min Hou, Chang Xu, Yang Liu, Weiqing Liu, Jiang Bian, Le Wu, Zhi Li, Enhong Chen, and Tie-Yan Liu. 2021. Stock trend prediction with multi-granularity data: A contrastive learning approach with adaptive fusion. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pages 700-709.  
+Wenpeng Hu, Mengyu Wang, Bing Liu, Feng Ji, Haiqing Chen, Dongyan Zhao, Jinwen Ma, and Rui Yan. 2019. Transformation of dense and sparse text representations. arXiv preprint arXiv:1911.02914.  
+Wenpeng Hu, Mengyu Wang, Qi Qin, Jinwen Ma, and Bing Liu. 2020. Hrn: A holistic approach to one class learning. Advances in neural information processing systems, 33:19111-19124.  
+Thanh D Huynh and Daniel R Smith. 2017. Stock price reaction to news: The joint effect of tone and attention on momentum. Journal of Behavioral Finance, 18(3):304-328.  
+Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. 2023. Scaling sentence embeddings with large language models. arXiv preprint arXiv:2307.16645.  
+Weiwei Jiang. 2021. Applications of deep learning in stock market prediction: recent progress. Expert Systems with Applications, 184:115537.  
+Nan Jing, Zhao Wu, and Hefei Wang. 2021. A hybrid model integrating deep learning with investor sentiment analysis for stock price prediction. Expert Systems with Applications, 178:115019.  
+Alexander G Kerl and Andreas Walter. 2007. Market responses to buy recommendations issued by personal finance magazines: effects of information, price-pressure, and company characteristics. Review of Finance, 11(1):117-141.  
+Christopher Krauss, Xuan Anh Do, and Nicolas Huck. 2017. Deep neural networks, gradient-boosted trees, random forests: Statistical arbitrage on the s&p 500. European Journal of Operational Research, 259(2):689-702.  
+Feng Li. 2010. The information content of forward-looking statements in corporate filings—a naive bayesian machine learning approach. Journal of Accounting Research, 48(5):1049-1102.  
+Qing Li, Jinghua Tan, Jun Wang, and Hsinchun Chen. 2020. A multimodal event-driven LSTM model for stock prediction using online news. IEEE Transactions on Knowledge and Data Engineering, 33(10):3323-3337.  
+Shuqi Li, Weiheng Liao, Yuhan Chen, and Rui Yan. 2023. Pen: prediction-explanation network to forecast stock price movement with better explainability. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence, pages 5187-5194.
+
+Yang Li and Yi Pan. 2022. A novel ensemble deep learning model for stock prediction based on stock prices and news. International Journal of Data Science and Analytics, pages 1-11.  
+Fabrizio Lillo, Salvatore Micciche, Michele Tumminello, Jyrki Piilo, and Rosario N Mantegna. 2015. How news affects the trading behaviour of different categories of investors in a financial market. Quantitative Finance, 15(2).  
+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.  
+Zhuang Liu, Degen Huang, Kaiyu Huang, Zhuang Li, and Jun Zhao. 2021. Finbert: A pre-trained financial language representation model for financial text mining. In Proceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence, pages 4513-4519.  
+Alejandro Lopez-Lira and Yuehua Tang. 2023. Can chatgpt forecast stock price movements? return predictability and large language models. Return Predictability and Large Language Models (April 6, 2023).  
+Di Luo, Weiheng Liao, Shuqi Li, Xin Cheng, and Rui Yan. 2023. Causality-guided multi-memory interaction network for multivariate stock price movement prediction. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12164-12176.  
+Leela Mitra and Gautam Mitra. 2011. Applications of news analytics in finance: A review. The handbook of news analytics in finance, 596(1).  
+Saloni Mohan, Sahitya Mullapudi, Sudheer Sammeta, Parag Vijayvergia, and David C Anastasiu. 2019. Stock price prediction using news sentiment analysis. In 2019 IEEE Fifth International Conference on Big Data Computing Service and Applications (BigDataService), pages 205-208. IEEE.  
+Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2022. Mteb: Massive text embedding benchmark. arXiv preprint arXiv:2210.07316.  
+Larry Olanrewaju Orimoloye, Ming-Chien Sung, Tiejun Ma, and Johnnie EV Johnson. 2020. Comparing the effectiveness of deep feedforward neural networks and shallow architectures for predicting stock price indices. Expert Systems with Applications, 139:112828.  
+Bonnie K Ray and Ruey S Tsay. 2000. Long-range dependence in daily stock volatilities. Journal of Business & Economic Statistics, 18(2):254-262.  
+Thomas Sattler. 2013. Do markets punish left governments? The Journal of Politics, 75(2):343-356.
+
+Paul C Tetlock. 2007. Giving content to investor sentiment: The role of media in the stock market. The Journal of finance, 62(3):1139-1168.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.  
+Mengyu Wang and Tiejun Ma. 2024. Mana-net: Mitigating aggregated sentiment homogenization with news weighting for enhanced market prediction. arXiv preprint arXiv:2409.05698.  
+Qili Wang, Wei Xu, and Han Zheng. 2018. Combining the wisdom of crowds and technical analysis for financial market prediction using deep random subspace ensembles. Neurocomputing, 299:51-61.  
+Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighof. 2023. C-pack: Packaged resources to advance general chinese embedding. arXiv preprint arXiv:2309.07597.  
+Wentao Xu, Weiqing Liu, Chang Xu, Jiang Bian, Jian Yin, and Tie-Yan Liu. 2021. Rest: Relational event-driven stock trend forecasting. In Proceedings of the web conference 2021, pages 1-10.  
+Yumo Xu and Shay B Cohen. 2018. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1970-1979.  
+Yunan Ye, Hengzhi Pei, Boxin Wang, Pin-Yu Chen, Yada Zhu, Jun Xiao, and Bo Li. 2020. Reinforcement-learning based portfolio management with augmented asset movement prediction states. In AAAI Conference on Artificial Intelligence. AAAI press.  
+Lean Yu, Yaqing Zhao, Ling Tang, and Zebin Yang. 2019. Online big data-driven oil consumption forecasting with google trends. International Journal of Forecasting, 35(1):213-223.  
+Xu Yuemei, Wang Zihou, and Wu Zixin. 2021. Predicting stock trends with cnn-bilstm based multi-feature integration model. Data Analysis and Knowledge Discovery, 5(7):126-138.  
+Yongjie Zhang, Weixin Song, Dehua Shen, and Wei Zhang. 2016. Market reaction to internet news: Information diffusion and price pressure. *Economic Modelling*, 56:43-49.
+
+Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, and Rong Jin. 2023. One fits all: Power general time series analysis by pretrained Im. arXiv preprint arXiv:2302.11939.
+
+# A Data
+
+Market Dataset: The S&P 500 Index, a benchmark for the U.S. market, tracks 500 leading publicly traded companies (Krauss et al., 2017; Fischer and Krauss, 2018). The NASDAQ 100 Index focuses on 100 of the largest, most actively traded non-financial companies listed on the Nasdaq Stock Market (Hou et al., 2021; Ding et al., 2020). Both indices are renowned for representing the largest and most influential U.S. market segments and are widely used as benchmark datasets for market prediction. We collected daily S&P 500 and NASDAQ 100 prices from 2003 to 2018, including six key features: open price, close price, adjusted close price, high price, low price, and trading volume, as market numerical data  $\boldsymbol{m}_{d}^{(n)}$ . Also, we collected the descriptions of these two markets and their constituent companies' names as textual data  $\boldsymbol{m}_{d}^{(t)}$ .
+
+News Dataset: Our used TRNA dataset spans from 2003 to 2018. It tracks over 25,000 equities and nearly 40 commodities and energy topics, providing a comprehensive news source for studying financial markets (Mitra and Mitra, 2011). The news dataset provides detailed information about collected news, including sentiment scores from an industry-established news-analytics system. These scores indicate the likelihood that the news will have a positive, neutral, or negative influence on the mentioned financial instruments. In our approach, we consider headlines as news textual data  $\boldsymbol{r}_{d,i}^{(t)}$  and the sentiment scores as news numerical data  $\boldsymbol{r}_{d,i}^{(n)}$ . Our experiments aim to encompass all available news items, despite the considerable daily news volume, which exhibits substantial volatility. The maximum daily news volume reaches 1,953, which is roughly 80 times greater than the minimum count of 24. This highlights the dynamic nature of the news volume on a daily basis.
+
+# B Detailed Experiment Settings
+
+# B.1 Language Models
+
+Our candidate LMs encompass a diverse range of architectures, including pre-trained encoder models, financial domain-specific models, top-performing text embedding models, and recent large language models.
+
+RoBERTa (Liu et al., 2019) is a well-known extension of BERT model. It has achieved success in various NLP tasks. Due to its robust performance and versatility, RoBERTa is an indispensable option when doing NLP-related tasks.
+
+FinBERT (Liu et al., 2021) is a domain-specific adaptation of the BERT model. It is designed specifically for financial sentiment analysis. This model benefits from pre-training via multi-task learning on extensive financial corpora, making it particularly effective for analyzing financial texts.
+
+BGE (Xiao et al., 2023) (BAAI General Embedding) is a leading text encoder model that has achieved top performance on the MTEB (Massive Text Embedding Benchmark) leaderboard (Muenninghoff et al., 2022). The MTEB is an extensive benchmark designed to evaluate text embeddings across eight NLP tasks, including 58 datasets and supporting 112 languages. This extensive scope ensures that models like BGE are tested against a wide range of NLP challenges, highlighting the quality of their generated text embeddings.
+
+Llama2 (Touvron et al., 2023) is an advanced large language model that has the potential to be used for a wide variety of purposes. It has already shown competitive performance on many NLP tasks.
+
+# B.2 Evaluation Metrics
+
+PnL quantifies the cumulative profit or loss experienced by a portfolio during a designated time frame. We calculate the PnL for each trading day and sum up the daily PnL values across the entire testing period. The PnL for all forecasts spanning  $D$  days can be expressed as (note it requires trading information from day  $D + 1$ ):
+
+$$
+P n L = \sum_ {d = 1} ^ {D} f l a g _ {d} \cdot \frac {\boldsymbol {p} _ {d + 1} ^ {c} - \boldsymbol {p} _ {d} ^ {c}}{\boldsymbol {p} _ {d} ^ {c}}, \tag {8}
+$$
+
+where  $flag_{d} = 1$  if  $\hat{y}_{d} = y_{d}$  and  $flag_{d} = -1$  otherwise.
+
+Sharpe Ratio measures the investment performance in relation to a risk-free asset. To compute the daily Sharpe Ratio within the test sets, we use the following formula:
+
+$$
+S R = \left(\frac {1}{n} \sum_ {d = 1} ^ {D} f l a g _ {d} \frac {\boldsymbol {p} _ {d + 1} ^ {c} - \boldsymbol {p} _ {d} ^ {c}}{\boldsymbol {p} _ {d} ^ {c}} - R _ {f}\right) / \sigma (R), \tag {9}
+$$
+
+where  $R_{f}$  represents the return rate of an investment with zero risks, meaning that it is the return that investors could expect for taking no risk, such
+
+<table><tr><td>Date</td><td>Case ID</td><td>News Headline</td><td>Weight</td></tr><tr><td>2008-09-26</td><td>1</td><td>JPMorgan Chase &amp; Co &lt;JPM.N&gt; says WaMu clientele very similar to JPMorgan&#x27;s</td><td>0.82</td></tr><tr><td>2008-09-26</td><td>2</td><td>UPDATE 4-WaMu is largest US bank failure; JPMorgan buys assets</td><td>0.98</td></tr><tr><td>2008-10-14</td><td>3</td><td>BANK OF AMERICA &lt;BAC.N&gt; up more than 15 pct as U.S. to inject $250 bln into banks</td><td>0.99</td></tr><tr><td>2008-10-14</td><td>4</td><td>STOCKS NEWS US-Banks reacting differently to government plan</td><td>0.77</td></tr><tr><td>2008-10-24</td><td>5</td><td>GM to delay unveiling of Buick LaCrosse model</td><td>0.76</td></tr><tr><td>2008-10-24</td><td>6</td><td>GM &lt;GM.N&gt; to delay unveiling of Buick LaCrosse, Cadillac CTS coupe models</td><td>0.71</td></tr><tr><td>2008-10-24</td><td>7</td><td>U.S. bank shares tumble on recession worry</td><td>0.92</td></tr><tr><td>2008-10-24</td><td>8</td><td>Stocks News - Auto companies down, GM intensifies merger talks</td><td>0.83</td></tr></table>
+
+Table 4: News Cases with their evaluated weights from FININ.
+
+as a Treasury bond investment. In this context, we use a value of 0.02 for  $R_{f}$ , which corresponds to the average US Treasury rate during our data collection period. Furthermore,  $\sigma(R)$  denotes the standard deviation of the excess return of the asset, denoted as  $R$ .
+
+# C Case Study
+
+To evaluate FININ's ability to capture news interactions, we analyze several news cases from the 2008 financial crisis. This period is marked by significant market events that clearly demonstrate the impact of news interactions on financial markets. We specifically select news events with obvious market repercussions to evaluate FININ's responses, since it is hard for humans to clearly ascertain the market impact of most news pieces, due to the large volume of news and the complexity of their interactions.
+
+Since the news weights are normalized using the softmax function to ensure their sum is one, the range of news weights varies with the daily news volume. In our case study, to make weights from different days comparable, we use min-max normalization to rescale all weights to a consistent range of 0 to 1. Table 4 lists these news reports and their normalized weights.
+
+Given that the daily average normalized weights are around 0.48, these listed reports with weights exceeding 0.7 are highlighted by FININ as having a more significant contribution to the market prediction outcomes. However, when read individually, the underlined news reports seem to describe ordinary company events, similar to numerous general financial news items reported daily. Nonetheless, when considering the broader context of other news events, the importance of these underlined reports becomes more apparent.
+
+On 2008-09-26, News 2 announced the failure of Washington Mutual (WaMu), the largest bank failure at that time, and JPMorgan's acquisition of its assets. This was a highly significant event during
+
+the 2008 financial crisis. Consequently, News 1, JP-Morgan's commentary on WaMu, carried substantial influence on that day. Similarly, on 2008-10-14, following the announcement of the substantial bank rescue package (News 3), News 4, regarding U.S. banks' reactions to the government's bailout plan, gained prominence. Lastly, on 2008-10-24, discussions surrounding a potential merger involving General Motors (GM) (News 8) and concerns about the U.S. recession (News 7) heightened the impact of news related to GM, including News 5 and News 6, the announcements about product delays. Therefore, while the underlined news reports may seem ordinary when viewed in isolation, their significance becomes evident when considered within the broader market context and major events of the time.
+
+In summary, these case studies demonstrate the effectiveness of FININ's modeling of news interactions. While some news items may seem individually general, FININ recognize their importance and assigns higher weights to them, due to considering their interactions with other breaking news events.

2024/Modeling News Interactions and Influence for Financial Market Prediction/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a09b3ea505b090dee965c43fc612a88dd4901c253856bbb6818f12c1490ba4cd
+size 557609

2024/Molecular Facts_ Desiderata for Decontextualization in LLM Fact Verification/2bc96a0c-7189-45f0-9748-c50dba7c8e98_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d33da95cb2e2781230e7fb19d3fdf035fdb8330bfe7369bd74ad4973d328b064
+size 1019071

2024/Molecular Facts_ Desiderata for Decontextualization in LLM Fact Verification/full.md

@@ -0,0 +1,641 @@
+# Molecular Facts: Desiderata for Decontextualization in LLM Fact Verification
+
+Anisha Gunjal Greg Durrett
+
+The University of Texas at Austin
+
+anishagunjal@utexas.edu
+
+# Abstract
+
+Automatic factuality verification of large language model (LLM) generations is becoming more and more widely used to combat hallucinations. A major point of tension in the literature is the granularity of this fact-checking: larger chunks of text are hard to fact-check, but more atomic facts like propositions may lack context to interpret correctly. In this work, we assess the role of context in these atomic facts. We argue that fully atomic facts are not the right representation, and define two criteria for molecular facts: decontextuality, or how well they can stand alone, and minimality, or how little extra information is added to achieve decontextuality. We quantify the impact of decontextualization on minimality, then present a baseline methodology<sup>1</sup> for generating molecular facts automatically, aiming to add the right amount of information. We compare against various methods of decontextualization and find that molecular facts balance minimality with fact verification accuracy in ambiguous settings.
+
+# 1 Introduction
+
+Large language models (LLMs) have emerged as powerful tools for delivering knowledge to users, either via closed-book generation or retrieval-augmented systems. However, these systems may not always produce correct facts (Liu et al., 2023a), an instance of the "hallucination" problem (Zhang et al., 2024; Ji et al., 2022; Zhang et al., 2023). Recent research has shown the potential of LLMs to identify unfaithful content and enable automatic fact-checking and attribution against sources (Falke et al., 2019; Goyal and Durrett, 2021; Min et al., 2023; Wang et al., 2024; Chern et al., 2023; Wei et al., 2024; Chen et al., 2023a; Malaviya et al., 2024; Gao et al., 2023b; Tang et al., 2024).
+
+A key step in this process is to break down generated content into individual atomic claims (Fabbri
+
+<https://github.com/anisha2102/molecular_facts>
+
+![](images/85b57e2156822669c2f14f091f49f180e0f279bf6b35d765171ea47c8f69b2ee.jpg)  
+Figure 1: Breaking a paragraph into atomic facts can cause errors in attribution: facts out of context appear to be true when they are not. The right granularity of decontextualization, "molecular facts," balances contextual grounding with atomicity.
+
+et al., 2022; Chen et al., 2023b; Kamoi et al., 2023b; Min et al., 2023). This decomposition allows for retrieval of evidence focused on a particular part of the generated content (Gao et al., 2023a; Wang et al., 2024; Chen et al., 2024) and also error localization by determining which parts of the content are supported or not. However, this step is not straightforward. Wanner et al. (2024) highlights that the effectiveness of automatic factuality verification is heavily dependent on the strategies employed for decomposing content into claims. In particular, LLMs have a propensity to incorrectly merge information about similarly named entities Lee et al. (2024) and current evaluation methods struggle to handle these ambiguities in atomic claims (Chiang and yi Lee, 2024). Figure 1 shows a possible issue: a fact that is "too atomic" can be validated against evidence that doesn't actually support it.
+
+In this work, we address the problem of how to find minimal yet still unambiguous facts for LLM fact verification. We frame this problem as one of decontextualization, adding context to a sentence to make it stand alone while retaining its original meaning (Choi et al., 2021). This process draws on the idea of specificity from discourse (Louis and Nenkova, 2012), specifically whether sentences can express key information about the participants without ambiguity (Li et al., 2016). However, making a claim unambiguous is not enough: when escalating from simple pronoun replacement in atomic facts to elaborations like a Swedish footballer in Figure 1, we must balance the specificity of the fact with how easy it will be to verify. It is not trivial to select the "right" information to elaborate on a claim without compromising the ease of verification.
+
+We define two criteria needed in this fact-checking setting: decontextuality, where the claim should uniquely specifying entities, events, and context, and minimality, maintained by avoiding excessive additional information that could complicate verification. We propose a notion of molecular facts, which balances these two criteria: molecular facts should be fully specific while compatible with the maximum number of possible evidence documents. We explore these criteria and our molecular facts in two settings. First, we address the question of how much non-minimality could be a problem for error localization with standard decontextualization techniques. We devise a synthetic fact-checking experiment where particular nuances of an output generation are unsupported and show that an average of  $6\%$  of claims may pose problems for error localization. In a setting with LLM responses of 5 sentences with 3 claims each, this would lead to localization errors in a large fraction of responses. We then evaluate the opposite problem, whether decontextualization is too minimal. We study a dataset of fact-checking with ambiguous entity names presented in Chiang and yi Lee (2024). We show that our method of molecular fact generation balances accuracy under ambiguous entity references with minimality of claims.
+
+Our main contributions are: (1) We re-examine the decontextualization process for fact-checking and define molecular claims following the desiderata of decontextuality and minimality. (2) We investigate the loss of minimality due to claim decontextualization and its impacts on error localization. (3) We find that molecular claims are more perform-  
+nant and minimal for long-form generations than
+
+existing decontextualization methods.
+
+# 2 Desiderata for Decontextualization
+
+We propose desiderata to determine the optimal level of decontextualization required for atomic facts. An atomic fact is defined as a discrete unit of information, derived from a broader claim, and variously described in the literature as propositions, subclaims, summary content units, or atomic content units (Nenkova and Passonneau, 2004; Liu et al., 2023b; Zhang and Bansal, 2021; Chen et al., 2023b; Min et al., 2023; Kamoi et al., 2023b).
+
+Although an atomic fact theoretically represents a singular conceptual unit, recent NLP work using this does not typically give this a rigorous definition from the standpoint of semantics. Wanner et al. (2024) demonstrate a high variation in the number of subclaims generated by different decomposition methods, with the macro-average of subclaims per biography ranging from 20.2 using the method by Kamoi et al. (2023b) to 32.9 with the approach by Chen et al. (2023b). Note that in Figure 1, She was a medallist at the European Athletics Championships in 1986 could be kept as one unit or broken into three facts evaluating her status as a medallist, the venue, and the date.
+
+# 2.1 Desiderata
+
+Preliminaries We define  $\mathbf{r}$  as a response from a language model to an input prompt  $\mathbf{x}$ , consisting of a series of claims  $(\mathbf{c}_1, \ldots, \mathbf{c}_n)$  to be verified. Claims are extracted through an upstream process of decomposition and potentially filtering for "check-worthiness" (i.e., does the claim present factual content or does it present an opinion?). We describe the prompting in Appendix A.
+
+We assume that in the context of  $\mathbf{r}$  and  $\mathbf{x}$ , a claim  $\mathbf{c}_i$  can be fully interpreted with a truth-conditional meaning  $I(\mathbf{c}_i\mid \mathbf{x},\mathbf{r})$ . In the terminology of Rashkin et al. (2021) and Choi et al. (2021),  $I(\mathbf{c}_i\mid \mathbf{x},\mathbf{r})$  represents  $\mathbf{c}_i$  interpreted in the linguistic context of  $\mathbf{x}$  and  $\mathbf{r}$ .
+
+We can construct a standalone proposition with truth conditional meaning equivalent to  $I$  by being sufficiently specific. For example, the statement in Figure 1 could be completely specified as Ann Jansson, the Swedish footballer born on 6 May 1957 who played for Hammarby IF, won a medal at the European Athletics Championship, the biennial event organized by the European Athletics Association, in 1986.
+
+Decontextualization Our goal in this work is to produce rewritten molecular claims. Denote by  $\mathbf{m}_i$  the rewritten form of  $\mathbf{c}_i$ , which should have semantics  $I$  when interpreted as a standalone proposition. As in Figure 1, this requires adding disambiguating information that could provide information needed to identify an entity (specifying that Jansson is a Swedish footballer), identify an event (specifying that the event happened in 1986), specify a qualification (in the field of biochemistry, ...), or more.
+
+Criterion 1 (Decontextuality) When interpreted as a standalone statement,  $\mathbf{m}_i$  must have the truth conditional meaning  $I(\mathbf{c}_i, \mathbf{x}, \mathbf{r})$ . That is, it should uniquely specify entities, events, and other context such that the claim  $\mathbf{c}_i$  is now interpretable.
+
+This criterion is equivalent to Definition 1 from Choi et al. (2021). For the settings we consider, the level of added information needed to specify the meaning of a statement like that in Figure 1 may be higher than in past applications like Choi et al. (2021). It is not sufficient to replace the pronoun she with Ann Jansson; we need to specify Ann Jansson, the Swedish footballer. Similarly, the city George Town could refer to a city in the Cayman Islands or Malaysia, therefore it must be decontextualized appropriately with a descriptor like George Town, a city in Cayman Islands.
+
+Other work such as question answering frameworks based on clarifying questions can target this information (Newman et al., 2023), but may fail to integrate the minimal new information needed, which we describe next.
+
+Minimality Adding too much information to a claim makes it less minimal. For instance, replacing "Ann Jansson" with "Ann Jansson, a Swedish footballer" requires verifying that a context referring to Ann Jansson is indeed talking about the Swedish footballer. Taken further, the reference "Ann Jansson, the Swedish footballer born on 6 May 1957 who played for Hammarby IF" is clearly suboptimal. It requires verifying Jansson's birthdate as an additional detail, and crucially, this detail won't be frequently reported in documents about Ann Jansson.
+
+Define  $\mathcal{E}^{*}(I(\mathbf{c},\mathbf{x},\mathbf{r}))$  as the set of set of evidence documents that support the statement  $I$  with an oracle understanding of the entities involved. For instance, this would contain a document describing the correct Ann Jansson, even if it did not confirm all the details about her life. Define
+
+$\mathcal{E}(\mathbf{m}_i)\subset \mathcal{E}^*$  to be the set of evidence documents that fully support a statement  $\mathbf{m}_i$ . For instance, in the case of Ann Jansson above, the document would need to specify Jansson's birthdate if this is contained in  $\mathbf{m}$ .
+
+Criterion 2 (Minimality) Given a set of statements  $\mathcal{M}$  that all decontextualize a claim  $\mathbf{c}_i$ , we should select  $\operatorname{argmax}_{\mathbf{m} \in \mathcal{M}} |\mathcal{E}(\mathbf{m})|$  to maximize the size of the set of supporting evidence documents.
+
+This criterion means that, when selecting distinguishing details for an entity, we should choose those that can typically be inferred from evidence. For instance, "Jason Martin" may be characterized either as a "rugby player" or specifically as a "former player for North Queensland Cowboys." Since "rugby player" is a more enduring and widely recognized description, yet still specific enough to indicate Jason Martin, it is more likely to be supported by a larger number of documents.
+
+Past work like Choi et al. (2021) instructs annotators to make minimal edits to statements. However, they do not provide guidance on what criteria should be used to choose from among multiple candidate edits.
+
+Molecular facts These two criteria suggest two things. First, atomic facts can be “too atomic:” they may need to be decontextualized. However, it is still valuable to have a reasonably minimal fact so it can be supported by many possible evidence documents.
+
+Molecular Fact A molecular fact is a statement  $\mathbf{m}_i$  corresponding to claim  $\mathbf{c}_i$  that obeys criteria 1 and 2: it should uniquely specify the interpretation of  $\mathbf{c}_i$  even when considered on its own, while adding as little information as possible to do so.
+
+# 2.2 Task Definition: Fact-checking LLMs
+
+Recall our setting where an LLM has generated a response  $\mathbf{r}$  to input prompt  $\mathbf{x}$ , and  $\mathbf{r}$  has associated claims  $(\mathbf{c}_1, \ldots, \mathbf{c}_n)$ . For each  $\mathbf{c}_i$ , we have a corresponding set of  $k$  evidence documents,  $D_i = (D_{i,1}, \ldots, D_{i,k})$ , that are referenced to assess the accuracy of  $\mathbf{c}_i$ . Furthermore, we have access to a gold standard of human-annotated labels for each atomic fact, represented as  $L = (l_1, \ldots, l_n)$ , where each  $l_i$  can be either SUPPORTED or NOT_SUPPORTED.
+
+Our goal is to make judgments about the supportedness of the  $\mathbf{c}_i$ , which requires appropriately decontextualizing each fact.
+
+![](images/57bef6b11d793534c14d289d1680deaf64183aaa4916ff63cdf5e98eafa5e538.jpg)  
+Figure 2: Controlled evidence generation framework for illustrating error localization introduced by decontextualization for atomic fact verification.
+
+We augment each atomic claim  $\mathbf{c}_i$  to a corresponding molecular claim  $\mathbf{m}_i$  as described in Section 3, resulting in a set of facts  $(\mathbf{m}_1,\dots \mathbf{m}_n)$ . We represent the model's factuality judgment prediction as a set of supported documents  $p_i = \mathrm{Check}(D_{i,j},\mathbf{m}_i)$ , for all  $j\in \{1,2,\ldots ,n\}$  in  $D_{i}$ . In other words, the prediction of Check() is accurate when it supports the molecular claim with the same evidence docs as humans.
+
+# 3 Method: Producing Molecular Facts
+
+We use a two-step process to refine an atomic fact into a molecular fact using gpt-4-turbo-2024-04-09 (Achiam et al., 2023). Our methodology makes the assumption that the ambiguity is typically restricted to a single entity in the claim. This is the case for the datasets we study in this work, described in Section 4.5.
+
+Stage 1: Identifying Ambiguity We identify the primary subject of the claim and to assess potential ambiguities based on its parametric knowledge: does the model know of multiple entities with this name? This step identifies the main subject  $\mathbf{s}_i$  of the claim  $\mathbf{c}_i$  and provides a disambiguation criteria  $\mathbf{b}_i$  for the subject  $\mathbf{s}_i$ . The disambiguation criteria  $\mathbf{b}_i$  can be 'None' when there is no ambiguity, or a type of criteria such as profession, birthyear, or location when disambiguation is required.
+
+For example, if the claim is about 'Charles Osgood', with multiple possible referents,  $\mathbf{s}_i$  is 'Charles Osgood', while  $\mathbf{b}_i$  could be 'profession' or 'birthyear' to clarify which Charles Osgood is being referred to. Conversely, if the claim concerns the unambiguous 'Julius Robert Oppenheimer',  $\mathbf{s}_i$  is 'Julius Robert Oppenheimer', and  $\mathbf{b}_i$  is 'None'.
+
+Stage 2: Molecular Facts Generation We then prompt the LLM to disambiguate the subject  $\mathbf{s}_i$  within the claim  $\mathbf{c}_i$ , harnessing both the identified disambiguation criteria  $\mathbf{b}_i$  and the claim's context  $\mathbf{r}$ . The output of this stage is a molecular fact  $\mathbf{m}_i$  for the atomic claim  $\mathbf{c}_i$ .
+
+The specifics regarding the prompts used are elaborated upon in Appendix 6 and 7.
+
+# 3.1 Baselines
+
+We analyze the robustness of fact verification across various systems on the defined criteria of minimality and decontextuality. Outputs for baselines are generated with gpt-4-turbo-2024-04-09.
+
+ATOMIC: Atomic claims are generated from the LLM's response using Min et al. (2023).
+
+SIMPLE-DECONTEXT: Atomic claims are decontextualized with a prompt described in 8 using the LLM's generated response as context for the atomic claim.
+
+SAFE-DECONTEXT: Decontextualization of atomic claims is performed using the revision prompt described in Wei et al. (2024).
+
+MOLECULAR-DECONTEXT: This approach follows a two-stage process described in section 3 to identify disambiguation criteria and subsequently decontextualize the atomic claim.
+
+Examples of outputs from each method can be found in Figure 3. With this task definition and baseline methodologies, we structure our experiments to analyze the two criteria presented in Section 2.1 in the following sections.
+
+# 4 Experiment: Minimality & Localization
+
+We begin our analysis of decontextualization with a controlled experiment to illustrate problems with error localization due to loss of minimality discussed in Criterion 2 in Section 2.1. Minimality is more difficult to evaluate than decontextuality. Less minimal facts impact error localization and can potentially lead to errors where an ancillary part of the claim leads to the whole claim being judged as wrong (Kamoi et al., 2023a). However, precisely measuring the harms of this is not easy without taking into account the downstream uses of error localization systems such as answer refinement (Xu et al., 2023) or fine-tuning (Wu et al., 2024; Roit et al., 2023).
+
+To measure the effects in a controlled way, we design a method for synthetic evidence generation as summarized in Figure 2. Our goal is to illustrate when decontextualized atomic facts actually contain multiple facts in a way that could impact error localization. We then study how many of these cases truly show this problem. To study the impact of information addition, we consider two baselines SIMPLE-DECONTEXT and SAFE-DECONTEXT which respectively have less and more restrictive prompts for including new information from the context to revise an atomic claim.
+
+# 4.1 Controlled Dataset Construction
+
+We now detail the dataset construction process as illustrated in Figure 2. We take a dataset  $D$  of 812 claims from the Factcheck-Bench dataset (Wang et al., 2024) which consists of long form ChatGPT responses with human-annotated factuality labels.
+
+Step 1: Extract Atomic Facts For each response  $\mathbf{r} \in D$ , we extract atomic facts  $(\mathbf{c}_1, \ldots, \mathbf{c}_n)$  using the method of Min et al. (2023).
+
+Step 2: Decontextualization: We perform decontextualization of the extracted atomic facts using SIMPLE-DECONTEXT and SAFE-DECONTEXT. Let the decontextualization for claim  $\mathbf{c}_i$  be denoted as  $\mathbf{d}_i$ . We refer to the  $\mathbf{c}_i$  that  $\mathbf{d}_i$  was created from as its core atomic fact; however, note that  $\mathbf{d}_i$  might support other facts as well.
+
+Step 3: Identifying Claims with Multiple Atomic Facts: We identify decontextualized claims that entail information of more than one atomic fact. We use the entailment model from Liu et al. (2022) to determine  $e(\mathbf{d}_i, \mathbf{c}_j) \in \{\text{supported}, \text{unsupported}\}$ ; is each  $\mathbf{c}_j$  supported by  $\mathbf{d}_i$ ? We retain cases
+
+where  $e(\mathbf{d}_i, \mathbf{c}_i) =$  supported and where  $|\{j : e(\mathbf{d}_i, \mathbf{c}_j) = \text{supported}\}| \geq 2$ ; that is, at least two atomic facts are supported by  $\mathbf{d}_i$ . For example, in Figure 2, the claim  $(\mathbf{d}_i)$ , 'The "Blackpink in Your Area" compilation album was released in 2018', is a decontextualized claim derived from the core atomic claim  $(\mathbf{c}_i)$ , 'The album was released in 2018'. The decontextualized claim  $(\mathbf{d}_i)$  entails the core atomic fact  $(\mathbf{c}_i)$  and an additional atomic fact  $(\mathbf{c}_j)$ ' "Blackpink in Your Area" is a compilation album'. Let  $D'$  denote this filtered set.
+
+Step 4: Generating Evidence for Partial Support: Whenever multiple atomic facts are merged, we could theoretically see a loss in localization capability from a model: if one fact is not supported, the entire claim will be determined to be not supported. To demonstrate this possibility, we now generate evidence that partially supports our multifact claims. As an example, in Figure 2, our goal in step 4 is to generate a paragraph that should not include details about "Blackpink in Your Area" being a compilation album. Then, if the statement 'The album was released in 2018' is decontextualized to include information about it being a compilation album, this paragraph will enable us to identify this: the evidence will no longer support the decontextualized fact, reflecting a failure of error localization.
+
+By construction of  $D^{\prime}$ ,  $\mathbf{d}_i$  is supported by at least two facts, its core atomic fact and auxiliary atomic fact(s). From this set of auxiliary atomic fact(s), we sample a banned fact  $\mathbf{c}_b$ . For each  $\mathbf{d}_i$ , we sample a set of key facts  $C_i = \{\mathbf{c}_{i,1},\dots,\mathbf{c}_{i,m}\}$  such that  $C_i$  contains the all atomic facts of the response  $r$  except  $\mathbf{c}_b$ . We then prompt the LLM to generate an evidence article supporting the facts  $C_i$  and not supporting the fact  $c_j$ . Each of these evidence articles ideally should support all the key facts and not support the banned fact.
+
+The prompt for this step is detailed in Figure 10 and other filtering criteria are described in Appendix F. Denote this set where evidence generation is feasible as  $E'$ .
+
+# 4.2 Evaluation Criteria
+
+We evaluate the impacts of loss of minimality on the recall of fact-checking. We measure the percentage of cases that change their label from SUPPORTED to NOT-supported after decontextualization on the set  $E'$ . We employ the roberta-large from AlignScore (Zha et al., 2023) as our Check() function.
+
+Generation: Kenneth Lamar Holland (born November 24, 1934) is a former member of the United States House of Representatives from South Carolina. He was born in Hickory, North Carolina and attended public schools in Gaffney, South Carolina before earning an AB in 1960 and an LL.B. in 1963 from the University of South Carolina. Holland served in the National Guard from 1952 to 1959 and was admitted to the South Carolina bar in 1963. He is currently a resident of Gaffney. Additionally, Ken Holland (born November 10, 1955) is the current general manager and executive vice president of the Detroit Red Wings of the National Hockey League. He was drafted in the 12th round, 188th overall by the Toronto Maple Leafs in the 1975 NHL Entry Draft ...
+
+ATOMIC FACT Holland served in the National Guard.   
+SIMPLE DE CONTEXT Kenneth Lamar Holland served in the National Guard from 1952 to 1959.   
+SAFE DE CONTEXT Kenneth Lamar Holland served in the National Guard.   
+MOLECULAR DE CONTEXT Kenneth Lamar Holland, the former member of the United States House of Representatives from South Carolina, served in the National Guard.
+
+Figure 3: Example claims (right) generated by SIMPLE-DECONTEXT, SAFE-DECONTEXT, MOLECULAR-DECONTEXT for the atomic claim derived from the highlighted sentence in the LLM generation (left).
+
+<table><tr><td>Baseline</td><td>Potential Non-minimal</td><td>Auto Non-minimal</td></tr><tr><td>SAFE-DE CONTEXT</td><td>8.49%</td><td>3.94%</td></tr><tr><td>SIMPLE-DE CONTEXT</td><td>23.39%</td><td>13.42%</td></tr></table>
+
+Table 1: Percentage of overall dataset impacted by minimality loss due to decontextualization leading to prediction changes from SUPPORTED to NOT-supported.  
+
+<table><tr><td>Category</td><td>Minimal</td><td>Non-minimal</td></tr><tr><td>SAFE-DECONTEXT</td><td>56.2%</td><td>43.8%</td></tr><tr><td>SIMPLE-DECONTEXT</td><td>27.5%</td><td>72.5%</td></tr></table>
+
+Table 2: Human annotation for categorizing the Auto Non-minimal subset into minimal vs. non-minimal.
+
+Using Check  $(\mathrm{D_i},\mathrm{c_i})$  , we identify cases where the core key fact is SUPPORTED by the generated evidence while the decontextualization and banned fact are NOT-supported. We call this set auto non-minimal.
+
+# 4.3 Results
+
+Table 1 shows the fraction of claims which are included in the set  $E'$ , which yields  $8.49\%$  for SAFE-DECONTEXT and  $23.39\%$  for SIMPLE-DECONTEXT. We refer to these claims as potential non-minimal claims: they have passed the checks in our pipeline and contain multiple atomic facts. Next we apply the Check() function to identify auto non-minimal claims, and find that they occur at a rate of  $3.94\%$  to  $13.42\%$  (Table 1).
+
+# 4.4 Human Evaluation
+
+Susceptibility to Error Localization We perform human evaluation on the auto non-minimal claims in Table 1. First, we categorize these into human judgments of whether a claim in this subset is minimal or not in Table 2. We categorize a decontextualization as minimal based on the criteria outlined in 2.1. This annotation is performed by the authors of the paper. We find that for SAFE-DECONTEXT,  $43.8\%$  of these cases are truly non-minimal in our
+
+judgment which represent  $1.7\%$  of the dataset  $D$ . For the SIMPLE-DECONTEXT baseline, we find that a staggering  $72.5\%$  of the auto non-minimal subset represents truly non-minimal claims. This represents  $9.6\%$  of the dataset  $D$ . We note that the remaining fraction of decontextualization cases not identified by the auto methods are those which entail more than one atomic fact but it is a necessary addition to make the atomic claim standalone.
+
+Decontextualization and Loss of Minimality We highlight that addition of information to a claim does not always make it less entailed to the evidence. In fact, in many cases information addition makes the sentence more specific. This is evident from Table 2 which shows that automatically flagged cases for non-minimality have a large percentage of minimal claims after human evaluation. For instance, "All taxes must be paid by April  $15$  "  $\rightarrow$  "In the US, all taxes must be paid by April  $15$  " is a necessary addition for claim specificity.
+
+# 4.5 Conclusion: Problem of Non-minimality
+
+We find through our controlled experiment and human evaluation that decontextualization can lead to non-minimal cases for between  $1.7\%$  to  $9.6\%$  of decontextualizations. These cases could cause error localization issues due to too much information added to the claims. In absolute terms, this is a low fraction for the baseline SAFE-DE CONTEXT. However, we note that a biography from FActScore (Min et al., 2023) contains dozens of atomic facts, meaning that in a single response from an LLM, there can easily be a handful of facts posing localization problems. Given the increasing adoption of the decomposition and decontextualization pipeline for automatic fact verification systems, we argue that multiple localization errors per response is cause to re-examine that pipeline. Next, we analyze trade-offs between minimality and decontextuality for fact checking of ambiguous biographies.
+
+<table><tr><td>Subset</td><td>ACCURACY OVERALL</td><td>ACCURACY SUPPORTED</td><td>ACCURACY NOT-supported</td><td>MODIFICATION RATE</td><td>AVG LENGTH (# of words)</td></tr><tr><td>ATOMIC</td><td>68.7%</td><td>77.5%</td><td>22.4%</td><td>-</td><td>7.61±3.03</td></tr><tr><td>SIMPLE-DECONTEXT</td><td>76.2%</td><td>84.3%</td><td>33.6%</td><td>99.5%</td><td>15.55±5.65</td></tr><tr><td>SAFE-DECONTEXT</td><td>73.4%</td><td>81.3%</td><td>31.9%</td><td>72.6%</td><td>9.86±4.38</td></tr><tr><td>MOLECULAR-DECONTEXT</td><td>74.7%</td><td>81.5%</td><td>38.8%</td><td>96.8%</td><td>14.96±5.6</td></tr></table>
+
+Table 3: Accuracy measured by Check  $(D_{i},m)$ , assessing the effectiveness of claim revisions by each baseline against the ambiguous document set associated with claim's main entity.  
+
+<table><tr><td>Human Label→</td><td colspan="3">SUPPORTED</td><td>NOT-supported</td><td></td></tr><tr><td>Baseline Pred→</td><td>SUPPORTED</td><td>SUPPORTED</td><td>NOT-supported</td><td>SUPPORTED</td><td></td></tr><tr><td>Matching Type→</td><td>Multi-Evidence matched</td><td>Single-Evidence Wrong Entity</td><td>No Evidence matched</td><td>Single/Multiple Evidence matched</td><td>Overall ↓</td></tr><tr><td>Baseline ↓</td><td></td><td></td><td></td><td></td><td></td></tr><tr><td>ATOMIC</td><td>16.2%</td><td>0.8%</td><td>1.8%</td><td>12.4%</td><td>31.1%</td></tr><tr><td>SIMPLE-DECONTEXT</td><td>7.9%</td><td>1.5%</td><td>3.9%</td><td>10.6%</td><td>23.8%</td></tr><tr><td>SAFE-DECONTEXT</td><td>12.0%</td><td>1.0%</td><td>2.8%</td><td>10.9%</td><td>26.6%</td></tr><tr><td>MOLECULAR-DECONTEXT</td><td>9.2%</td><td>1.5%</td><td>4.8%</td><td>9.8%</td><td>25.3%</td></tr></table>
+
+# 5 Experiment: Ambiguous Biographies
+
+We now analyze to what extent our molecular facts add the correct information to decontextualize on an existing dataset with ambiguous entity references.
+
+Dataset We use the ambiguous biographies dataset introduced in Chiang and yi Lee (2024) which comprises biographies generated by LLMs for multiple entities that share similar names, such as Dick Hanley (swimmer) and Dick Hanley (footballer). In this dataset we represent the biographies generated by the LLMs as  $\mathbf{r}$  and  $\mathbf{c}_i$  correspond to atomic claims generated using the methodology outlined in (Min et al., 2023). For this setting, we define each claim to have a subject  $\mathbf{s}_i$ , which is ambiguous due to the nature of the dataset. The dataset provides a set of evidence documents sourced from Wikipedia page of the subject disambiguation,  $D_i = \{D_{i,2}, D_{i,2}, \ldots\}$  for subjects sharing similar names as  $\mathbf{s}_i$ . This dataset is suitable for evaluating decontextuality as it consists of two properties: (i) atomic claims that require decontextualization (such as entity specification, noun completion), (ii) multiple entities with the same name that require additional disambiguation such as specifying location, occupation, or time-period.
+
+Our goal is to verify the claims with the set of documents using Check(). We randomly sample 726 claims from the human-annotated set for this study which belong to either SUPPORTED or NOT-supported categories. For each claim we construct a revision using the methods and baselines
+
+Table 4: Fine-grained error analysis categorizing baseline mistakes based on human label of SUPPORTED/NOT-supported along with categorization of <Single/Multi/No>-Evidence based on the number of ambiguous evidence docs that support the claim.  
+
+<table><tr><td>Baseline</td><td>Minimal↑</td><td>Non-Minimal↓</td><td>Ambig.↓</td></tr><tr><td>SIMPLE</td><td>16.0%</td><td>56.0%</td><td>28.0%</td></tr><tr><td>SAFE</td><td>24.0%</td><td>0.0%</td><td>76.0%</td></tr><tr><td>MOLECULAR</td><td>52.0%</td><td>24.0%</td><td>24.0%</td></tr></table>
+
+Table 5: Human analysis of decontextualized claims for all baselines on the axis of minimality and ambiguity.
+
+described in section 3 and compare the prediction with human labels.
+
+Evaluation Criteria We evaluate our judgment of a claim on two axes: (1) whether it aligns with the human annotation of SUPPORTED or NOT-supported, and (2) whether it is supported by the correct evidence. For each evidence associated with the claim, we compute  $p_{i,k} = \mathrm{Check}(D_{i,k}, c_i)$  where  $c_i$  is the claim processed by the particular baseline and  $k$  represents the  $k$ th ambiguous subject related document for the claim. We consider the judgment  $p_{i,k}$  to be correct only if the prediction of the claim matches the human label and the prediction is supported by the correct entity's evidence document.
+
+# 6 Results: Ambiguous Biographies
+
+Table 3 presents the results of this experiment. All methods of decontextualization baselines yield higher accuracy rates compared to atomic claims, across all subsets. We see that Molecular and Simple decontextualization methods have a higher proclivity to modify the atomic claims than the SAFE decontextualization baseline. Consequently, the average sentence lengths of the former methods is
+
+<table><tr><td>Baseline Pair</td><td>Overlap</td></tr><tr><td>ATOM &amp; SIMPLE-DECONTEXT</td><td>7%</td></tr><tr><td>ATOM &amp; SAFE-DECONTEXT</td><td>44%</td></tr><tr><td>ATOM &amp; MOLECULAR-DECONTEXT</td><td>15%</td></tr><tr><td>SIMPLE-DECONTEXT &amp; SAFE-DECONTEXT</td><td>27%</td></tr><tr><td>SIMPLE-DECONTEXT &amp; MOLECULAR-DECONTEXT</td><td>36%</td></tr><tr><td>MOLECULAR-DECONTEXT &amp; SAFE-DECONTEXT</td><td>32%</td></tr></table>
+
+Table 6: Information overlap between baselines as measured by bi-directional entailment.
+
+also larger than the SAFE baseline. Higher degrees of modification generally lead to higher accuracy. All three methods are on a Pareto frontier of length versus accuracy.
+
+However, accuracy using the Check() function does not incorporate minimality. We investigate the minimality of the baselines by performing a human evaluation of randomly sampled 25 claims in Table 5. We see that the baseline SIMPLE-DECONTEXT has a large fraction of non-minimal and ambiguous claims as compared to MOLECULAR-DECONTEXT. Analysis in Section 4.4 shows that SAFE-DECONTEXT is more minimal than SIMPLE-DECONTEXT; however, it struggles with ambiguity.
+
+Overall, we observe that molecular claims strike a balance by maintaining minimality with ambiguity removal and improving accuracy. They are significantly more minimal than SIMPLE-DECONTEXT and more performant in ambiguous generations than SAFE-DECONTEXT.
+
+Error breakdown To analyze the nature of errors encountered, we detail a case-wise error distribution in Table 4. Specifically, we study the behavior of various baselines to mispredict the label as SUPPORTED or NOT-supported in comparison to human annotation. Note that due to the ambiguous nature of this dataset, claims may be erroneously validated by several distracting pieces of evidence. Therefore, we further partition the error analysis table to reflect the model's prediction on (i) Single/Multi/No Evidence: whether a claim is supported by single, multiple, or no pieces of evidence, and (ii) (Correct/Wrong Entity): whether the set of supporting evidence contains the accurate evidence with which the claim ought to be aligned. Overall, all decontextualization methods show a lower error rate than atomic claims.
+
+Information Overlap We perform an information overlap analysis shown in Table 6 using the model from Liu et al. (2022) to check bidirectional entailment of the fraction of cases where the information is equivalent between two baselines (Gunjal
+
+and Durrett, 2023). We find in a large fraction of cases each baseline adds different information to modify the atomic claim. SAFE-DECONTEXT has least amount of modification albeit suffers with ambiguity and SIMPLE-DECONTEXT has most amount of modification at the cost of minimality loss.
+
+# 7 Related Work
+
+Recent research in factuality verification of LLM generations advocates decomposing LLM generations into atomic facts or subclaims and verifying each against retrieved evidence (Min et al., 2023; Kamoi et al., 2023b; Fabbri et al., 2022). End-to-end pipelines for factuality verification have been proposed, involving steps such as claim extraction, revision, determining checkworthiness, evidence retrieval, and verification (Wang et al., 2024; Chern et al., 2023; Wei et al., 2024; Chen et al., 2024). These papers often evaluate on recently-released datasets of errors in generations Liu et al. (2023a); Malaviya et al. (2024); Chen et al. (2023a). Our work comments on the decontextualization step frequently used in these pipelines.
+
+Our work fits into a broader ecosystem of techniques in this area. Gao et al. (2023b) enable LLMs to generate text with citations. For faithful LLM generations, Gao et al. (2023a) use evidence retrieval for revision, and He et al. (2022) utilize chain-of-thought coupled with retrieval for faithful explanations. Fine-tuned systems, such as that by Zha et al. (2023), predict alignment scores for verification, while Tang et al. (2024) propose LLM-AggreFact for sentence-level factuality labels. Wanner et al. (2024) find that evaluation metrics for fact verification are sensitive to the claim decomposition method used.
+
+Prior work on decontextualization has investigated basic notions like anaphora resolution (Choi et al., 2021), question answering frameworks (Newman et al., 2023), and extract-then-decontextualize methods for summarization (Potluri et al., 2023). In fact verification, atomic claims are made standalone before evidence retrieval via decontextualization (Wang et al., 2024) or claim revision (Wei et al., 2024). Decontextualization is also used to resolve ambiguity Zhang and Choi (2021); Lee et al. (2024); our work shares this focus.
+
+# 8 Conclusion
+
+We introduce molecular facts and the desiderata of decontextualization in LLM fact verification. We
+
+define the criteria of decontextuality and minimality in this context. Through a controlled experiment, we show that localization errors due to loss of minimality by decontextualization is sensitive to the method used. We propose a method of "molecular facts" and find that they improve fact verification precision for claims from generation about ambiguous entities. We show that molecular facts strike a balance between maintaining minimality and accuracy of fact-verification.
+
+# Limitations
+
+Scope We illustrate the phenomenon of ambiguity in atomic claims; however, our main evaluation of molecular facts is in the domain of English-language biographies. This is due to the availability of the dataset, Wikipedia evidence, and the prevalence of biography benchmarks in recent work. Conceptually, the ambiguity in the subject or predicate of the claim can be extended to other realistic datasets, but we leave that exploration to future work. Relatedly, we focus on entity ambiguity for illustration of our method. There may be other types of ambiguities that molecular fact generation can address in other contexts and other datasets.
+
+Furthermore, we focus our experiments on high-performing LLMs in this work. The extension of decontextualization and molecular fact generation to smaller, open-source models and the improvement in this regime is a good subject for further study.
+
+Finally, we believe our approach should be evaluated fully end-to-end in an LLM pipeline that generates responses and then verifies their factuality. However, despite substantial research in these directions, we are not aware of an off-the-shelf experimental pipeline that is usable for this setting.
+
+Decomposition Quality We do not consider the errors introduced due to poor decomposition of atomic facts in this work. It is possible that some of these errors are resolved due to decontextualization or disambiguation implicitly, but we do not make any specific claims about this.
+
+Coverage of Domains and Languages The datasets utilized for ambiguous biographies are limited to English-language claims focused on English-centric concepts within Wikipedia. Similarly, the synthetic data generation experiment for minimality analysis is confined to English language outputs and relies on GPT-4's parametric knowledge,
+
+which may limit the breadth of topics and domains covered.
+
+# Acknowledgments
+
+We thank Jessy Li for comments and feedback on the initial draft of this work. This work was supported by NSF CAREER Award IIS-2145280 and the NSF AI Institute for Foundations of Machine Learning (IFML). This material is also based on research that is in part supported by the Air Force Research Laboratory (AFRL), DARPA, for the KAIROS program under agreement number FA8750-19-2-1003.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.  
+Jifan Chen, Grace Kim, Aniruddh Sriram, Greg Durrett, and Eunsol Choi. 2024. Complex claim verification with evidence retrieved in the wild. In Proceedings of the North American Chapter of the Association for Computational Linguistics.  
+Shiqi Chen, Yiran Zhao, Jinghan Zhang, I-Chun Chern, Siyang Gao, Pengfei Liu, and Junxian He. 2023a. FELM: Benchmarking Factuality Evaluation of Large Language Models. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track.  
+Sihao Chen, Senaka Buthpitiya, Alex Fabrikant, Dan Roth, and Tal Schuster. 2023b. PropSegmEnt: A Large-Scale Corpus for Proposition-level Segmentation and Entailment Recognition. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8874-8893, Toronto, Canada. Association for Computational Linguistics.  
+I-Chun Chern, Steffi Chern, Shiqi Chen, Weizhe Yuan, Kehua Feng, Chunting Zhou, Junxian He, Graham Neubig, Pengfei Liu, et al. 2023. FacTool: Factuality Detection in Generative AI-A Tool Augmented Framework for Multi-Task and Multi-Domain Scenarios. arXiv preprint arXiv:2307.13528.  
+Cheng-Han Chiang and Hung yi Lee. 2024. Merging Facts, Crafting Fallacies: Evaluating the Contradictory Nature of Aggregated Factual Claims in Long-Form Generations. arXiv 2402.05629.  
+Eunsol Choi, Jennimaria Palomaki, Matthew Lamm, Tom Kwiatkowski, Dipanjan Das, and Michael Collins. 2021. Decontextualization: Making sentences stand-alone. Transactions of the Association for Computational Linguistics, 9:447-461.
+
+Alexander Fabbri, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022. QAFactEval: Improved QA-based factual consistency evaluation for summarization. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2587-2601, Seattle, United States. Association for Computational Linguistics.  
+Tobias Falke, Leonardo F. R. Ribeiro, Prasetya Ajie Utama, Ido Dagan, and Iryna Gurevych. 2019. Ranking generated summaries by correctness: An interesting but challenging application for natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2214-2220, Florence, Italy. Association for Computational Linguistics.  
+Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan, Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, and Kelvin Guu. 2023a. RARR: Researching and Revising What Language Models Say, Using Language Models. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16477-16508, Toronto, Canada. Association for Computational Linguistics.  
+Tianyu Gao, Howard Yen, Jiatong Yu, and Danqi Chen. 2023b. Enabling Large Language Models to Generate Text with Citations. In Empirical Methods in Natural Language Processing (EMNLP).  
+Tanya Goyal and Greg Durrett. 2021. Annotating and modeling fine-grained factuality in summarization. In Proceedings of NAACL.  
+Anisha Gunjal and Greg Durrett. 2023. Drafting Event Schemas using Language Models. arXiv preprint arXiv:2305.14847.  
+Hangfeng He, Hongming Zhang, and Dan Roth. 2022. Rethinking with retrieval: Faithful large language model inference. arXiv preprint arXiv:2301.00303.  
+Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of hallucination in natural language generation. ACM Computing Surveys.  
+Ryo Kamoi, Tanya Goyal, and Greg Durrett. 2023a. Shortcomings of question answering based factuality frameworks for error localization. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 132-146.  
+Ryo Kamoi, Tanya Goyal, Juan Rodriguez, and Greg Durrett. 2023b. WiCE: Real-World Entailment for Claims in Wikipedia. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7561-7583, Singapore. Association for Computational Linguistics.
+
+Yoonsang Lee, Xi Ye, and Eunsol Choi. 2024. Ambigdocs: Reasoning across documents on different entities under the same name. arXiv 2404.12447.  
+Junyi Jessy Li, Bridget O'Daniel, Yi Wu, Wenli Zhao, and Ani Nenkova. 2016. Improving the annotation of sentence specificity. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC'16), pages 3921-3927.  
+Alisa Liu, Swabha Swayamdipta, Noah A. Smith, and Yejin Choi. 2022. WANLI: Worker and AI Collaboration for Natural Language Inference Dataset Creation. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 6826-6847, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Nelson Liu, Tianyi Zhang, and Percy Liang. 2023a. Evaluating Verifiability in Generative Search Engines. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7001-7025, Singapore. Association for Computational Linguistics.  
+Yixin Liu, Alex Fabbri, Pengfei Liu, Yilun Zhao, Linyong Nan, Ruilin Han, Simeng Han, Shafiq Joty, Chien-Sheng Wu, Caiming Xiong, and Dragomir Radev. 2023b. Revisiting the Gold Standard: Grounding Summarization Evaluation with Robust Human Evaluation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4140-4170, Toronto, Canada. Association for Computational Linguistics.  
+Annie Louis and Ani Nenkova. 2012. A corpus of general and specific sentences from news. In LREC, volume 1818, page 10. Citeseer.  
+Chaitanya Malaviya, Subin Lee, Sihao Chen, Elizabeth Sieber, Mark Yatskar, and Dan Roth. 2024. ExpertQA: Expert-curated questions and attributed answers. In Proceedings of the North American Chapter of the Association for Computational Linguistics.  
+Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 12076-12100, Singapore. Association for Computational Linguistics.  
+Ani Nenkova and Rebecca J Passonneau. 2004. Evaluating content selection in summarization: The pyramid method. In Proceedings of the human language technology conference of the north american chapter of the association for computational linguistics: Hltnaacl 2004, pages 145-152.  
+Benjamin Newman, Luca Soldaini, Raymond Fok, Arman Cohan, and Kyle Lo. 2023. A question answering framework for decontextualizing user-facing snippets from scientific documents. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 3194-3212.
+
+Sandro Pezzelle. 2023. Dealing with semantic underspecification in multimodal nlp. arXiv preprint arXiv:2306.05240.  
+Abhilash Potluri, Fangyuan Xu, and Eunsol Choi. 2023. Concise answers to complex questions: Summarization of long-form answers. arXiv preprint arXiv:2305.19271.  
+Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Michael Collins, Dipanjan Das, Slav Petrov, Gaurav Singh Tomar, Iulia Turc, and D. Reitter. 2021. Measuring Attribution in Natural Language Generation Models. Computational Linguistics, 49:777-840.  
+Paul Roit, Johan Ferret, Lior Shani, Roee Aharoni, Geoffrey Cideron, Robert Dadashi, Matthieu Geist, Sertan Girgin, Leonard Hussenot, Orgad Keller, Nikola Momchev, Sabela Ramos Garea, Piotr Stanczyk, Nino Vieillard, Olivier Bachem, Gal Elidan, Avinatan Hassidim, Olivier Pietquin, and Idan Szpektor. 2023. Factually consistent summarization via reinforcement learning with textual entailment feedback. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6252-6272, Toronto, Canada. Association for Computational Linguistics.  
+Frank Schilder. 1998. An underspecified segmented discourse representation theory (usdrt). In 36th Annual Meeting of the Association for Computational Linguistics and 17th International Conference on Computational Linguistics, Volume 2, pages 1188-1192.  
+Liyan Tang, Philippe Laban, and Greg Durrett. 2024. MiniCheck: Efficient Fact-Checking of LLMs on Grounding Documents. arXiv preprint arXiv:2404.10774.  
+Yuxia Wang, Revanth Gangi Reddy, Zain Muhammad Mujahid, Arnav Arora, Aleksandr Rubashevskii, Jiuhui Geng, Osama Mohammed Afzal, Liangming Pan, Nadav Borenstein, Aditya Pillai, Isabelle Augenstein, Iryna Gurevych, and Preslav Nakov. 2024. Factcheck-Bench: Fine-Grained Evaluation Benchmark for Automatic Fact-checkers.  
+Miriam Wanner, Seth Ebner, Zhengping Jiang, Mark Dredze, and Benjamin Van Durme. 2024. A Closer Look at Claim Decomposition. arXiv preprint arXiv:2403.11903.  
+Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, et al. 2024. Long-form factuality in large language models. arXiv preprint arXiv:2403.18802.  
+Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. 2024. Finegrained human feedback gives better rewards for language model training. Advances in Neural Information Processing Systems, 36.
+
+Wenda Xu, Daniel Deutsch, Mara Finkelstein, Juraj Juraska, Biao Zhang, Zhongtao Liu, William Yang Wang, Lei Li, and Markus Freitag. 2023. Pinpoint, not criticize: Refining large language models via fine-grained actionable feedback. arXiv preprint arXiv:2311.09336.  
+Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu 2023. AlignScore: Evaluating Factual Consistency with A Unified Alignment Function. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11328-11348, Toronto, Canada. Association for Computational Linguistics.  
+Michael J.Q. Zhang and Eunsol Choi. 2021. Situated QA: Incorporating extra-linguistic contexts into QA. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)  
+Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A. Smith. 2024. How Language Model Hallucinations Can Snowball. In *Forty-first International Conference on Machine Learning*.  
+Shiyue Zhang and Mohit Bansal. 2021. Finding a Balanced Degree of Automation for Summary Evaluation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6617-6632, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. 2023. Siren's song in the AI ocean: a survey on hallucination in large language models. arXiv preprint arXiv:2309.01219.
+
+# A Prompts
+
+We give details on all the prompts used throughout this work.
+
+Decontextuality Experiment Prompts The stepwise molecular facts generation prompts for MOLECULAR_DECONTEXT are in Figure 6, 7. For the simple decontextualization baseline SIMPLE_DECONTEXT, the prompts are provided in 8.
+
+Minimality Experiment Prompts The prompt for generating controlled evidence for the minimality experiment is given in Figure 10.
+
+# B Additional Related Work
+
+Decomposition in Text Summarization Decomposition of responses is also prevalent in the text summarization literature. Nenkova and Passonneau (2004) introduced the Pyramid protocol for summarization evaluation which extracts weighted
+
+Summarization Content Units (SCUs) which represent the importance of various facts present in multiple human-generated summaries of a text. Zhang and Bansal (2021) propose using Semantic Triplet Units (STUs), which are summary content units generated automatically using SRL parsers, to evaluate generated summaries with textual entailment models. Similarly, Liu et al. (2023b) propose Atomic Content Units (ACUs) as a new summarization salience protocol that allows for higher inter-annotator agreement. Chen et al. (2023b) propose using entailment judgments on a set of sentence propositions within a document.
+
+Decontextualization and Specificity Decontextualization is a process of making sentences standalone by resolving missing context while preserving its meaning (Choi et al., 2021). A related phenomenon is the notion of specificity. Louis and Nenkova (2012) presented the first corpus of sentences distinguished on the criteria of being general or specific. Their idea of classification was based on examples and intuition by defining general sentences to be broad statements about a topic that would need additional evidence or examples for a reader to understand, whereas, specific sentences can stand by themselves. Li et al. (2016) make this definition more specific by grounding specificity for a sentence to three requirements: (i) it is easy to understand the meaning and identify of the intended references without ambiguity; (ii) the truth of the statement can be assessed based on the sentence itself and general shared knowledge; and (iii) the sentence fully expresses key information about the participants and causes of an event. Another related notion is underspecification in discourse, which is an intentional feature to maintain communication efficiency (Schilder, 1998). This has been annotated by Li et al. (2016) and highlighted in a multimodal setting by Pezzelle (2023).
+
+# C Human Annotation Criteria for Categorizing the Non-minimal Subset
+
+We describe the criteria for annotating the auto non-minimal subset into minimal vs. non-minimal as shown in Table 2. For each instance, we compare the original claim, the decontextualization, and the banned fact. We label cases as minimal when either of the following applies: (1) the banned fact is closely related the atomic fact and it is a necessary addition to the atomic claim to make it standalone. In other words, the banned fact is a necessary ad
+
+dition to the atomic claim to add context and/or resolve ambiguity. For example, "The album is their first full-length studio album." is decontextualized to "The album released in 2020 is Blackpink's first full-length studio album." and the banned fact is "The album was released in 2020." The information in the banned fact is necessary addition to disambiguate "the album" in this case. (2) The banned fact entailed by the decontextualization, but it is due to an entailment error. For example, the decontextualization "Mey Eden, one of the largest bottled water companies in Israel, offers flavored water products." is erroneously entailed by the banned fact "Mey Eden offers still water products."
+
+# D Human Analysis Criteria for Categorizing Minimality and Ambiguity
+
+We describe the criteria for the human analysis for on the decontextualization of each baseline on the axis of minimality and ambiguity shown in Table 5. We categorize a claim decontextualization as non-minimal when it contains additional information that goes beyond making the sentence stand-alone and can potentially cause loss of error-localization. We categorize a claim decontextualization as ambiguous when it lacks clarifications for entities that could refer to different ambiguous subjects or add enough context to disambiguate the main entity. If both of the above conditions are not violated, we categorize the decontextualization as minimal.
+
+# E Models, Datasets and Computation Cost
+
+The gpt-4-turbo-2024-04-09 model was employed for running baselines and generating outputs, while the gpt-3.5-turbo model was used for evaluation through FActScore (Achiam et al., 2023). For generation experiments, we set the temperature to 0.75. The total cost for generating decontextualizations and evaluating the ambiguous biography experiment was approximately $120.
+
+In the minimality experiment, gpt-3.5-turbo was used to extract atomic facts, and gpt-4-turbo-2024-04-09 was used for decontextualization and generation tasks. This resulted in a total cost of around $100. We use a NVIDIA A40 GPU for evaluation using AlignScore (Zha et al., 2023) and entailment computation using WANLI (Liu et al., 2022),
+
+We use ChatGPT for improving writing format
+
+![](images/460f92c23f76ebd21488ee4cc8f932c4378807de9583af9f27244241d21e7ed7.jpg)  
+Figure 4: Variation in accuracy for different fact-checking methods as the offset from the entity switch point changes. Each line represents a method, with the solid lines indicating the method's accuracy at different offsets, and the dashed lines representing the overall accuracy of the method. The silver star represents the performance of human-in-the-loop molecular claim generation.
+
+ting and generating boilerplate code for figure generation in this paper.
+
+We use the open-source dataset published by Wang et al. (2024) under the Apache 2.0 license. We also use the open-source code-base of FactScore (Min et al., 2023) for evaluations which is published under MIT license and AlignScore (Zha et al., 2023) published under MIT License.
+
+# F Controlled Experiment on Minimality Generation Details
+
+Filtering Criteria applied in Step 3 Before filtering claims which are supported by more than two atomic facts, we do not consider cases where one atomic fact is a substring of another one.
+
+Filtering Criteria applied in Step 4 We detail the filtering criteria applied in evidence generation for partial support detailed in 4.1. After we sample a set of key facts  $C_i = \{\mathbf{c}_{i,1},\dots,\mathbf{c}_{i,m}\}$  such that  $C_i$  contains the all atomic facts of the response  $r$  except  $\mathbf{c}_b$ , we also apply a filtering criteria to remove cases where the banned fact and any of the key facts is similar; i.e., for  $\mathbf{c}_{i,k}$  in  $C_i$ , we filter cases where  $e(\mathbf{c}_{i,k},\mathbf{c}_b) =$  supported. At the end of step 4 after we prompt the LLM to generate an evidence article, we also account for generation errors and remove the cases where banned fact is supported by the generated evidence.
+
+![](images/63cab6e6891cb4a2868e349e88c66a08fe51798640547b7c2a8e081f0038c238.jpg)  
+Figure 5: Changing preferences of selection of disambiguating fact by molecular decontextualization for long-form generation with hallucinations.
+
+# G Remaining Challenges
+
+To shed light on the remaining challenges, we focus on one of the most challenging scenarios for decontextualization. In the ambiguous biography dataset from Chiang and yi Lee (2024), we often observe what we call an entity switch point: a claim  $\mathbf{c}_i$  that draws on information about entity B, when sentences  $\mathbf{c}_{< i}$  all refer to entity A. This is where decontextualization is crucial to recognize that  $\mathbf{c}_i$  in context does not refer to the correct entity.
+
+Molecular claims recover fastest at the entity-switching point We investigate the performance of baselines under the lens of ambiguity resolution. Note that these results are reported on baselines tested with gpt3.5-turbo. We find that the dataset of ambiguous biographies becomes the most confusing at the entity switch point. Figure 4 shows a significant performance drop at the switch across all methods. Basic decontextualization methods (DECONTEXT, SAFE-DECONTEXT) perform the worst, underperforming the ATOMIC baseline at the switch, but molecular claims, which incorporate richer disambiguation information, show relative robustness, improving by  $3.5\%$  over the most effective decontextualization approach (SAFE-DECONTEXT).
+
+Gap from human performance To estimate the upper bound of ideal performance at the entity switch point in Figure 4, we generate molecular claims at the entity-switch point with weak supervision human-in-the-loop supervision. We use the prompt shown Figure 9 in which has access to gold disambiguations from Wikipedia about the entities in the passage. This method's performance even with weak human supervision is significantly better than automated decontextualization methods, bringing attention to this limitation of current fact
+
+checking pipelines.
+
+AMBIGUITY CRITERIA: Ambiguity manifests in diverse forms, including:
+
+- Similar names denoting distinct entities.  
+- Varied interpretations stemming from insufficient information.  
+- Multiple understandings arising from vague or unclear information.
+
+Instructions:
+
+- Identify the main SUBJECT within the claim.  
+- Determine if the SUBJECT is ambiguous according to the provided AMBIGUITY CRITERIA.  
+- Utilize your world knowledge to enumerate potential DISAMBIGUATIONS for the identified SUBJECT.  
+- Specify the TYPE of information employed for disambiguation based on the list of DISAMBIGUATIONS.  
+- If the SUBJECT does not have ambiguous interpretations, return None  
+- Provide an explanation of the method used to arrive at the final response.
+
+Format your response as a combination of explanation and a dictionary with the following structure:
+
+##EXPLANATION#:
+
+<step-by-step-explanations>
+
+##RESPONSE#:
+
+{"subject":<subject>, "disambiguations":[<instance-1>, <instance-2>..], "disambiguation_type":<type>}
+
+Example 1
+
+##CLAIM#: David Heyman, born in 1961 in England, is the founder of Heyday Films.
+
+EXPLANATION:
+
+The SUBJECT of the claim is "David Heyman". Based on my world knowledge, there are multiple individuals who share similar names,
+
+such as "David Heyman - the British film producer" and "David Heyman - the Chairman of the Board of UK HPA."
+
+To differentiate between them, it is crucial to consider their respective occupations. This criterion offers a
+
+clearer disambiguation compared to nationality, as both individuals are British and thus nationality alone does not
+
+provide sufficient distinguishing information.
+
+##RESPONSE##:
+
+{"subject": "David Heyman", "disambiguations": ["David Heyman - British film producer, founder of Heyday Films",
+
+"David L. Heyman - Chairman of the Board of UK HPA", "disambiguation_type": "Occupation"
+
+Example 2:
+
+##CLAIM#: Ruth Bader Ginsburg served as a Supreme Court justice.
+
+EXPLANATION:
+
+The SUBJECT is "Ruth Bader Ginsburg". According to my world knowledge, this is a unique individual and I am not aware
+
+of any other individuals/entities with a similar name. Hence, there are no ambiguous interpretations of this SUBJECT and the claim requires no further disambiguation.
+
+##RESPONSE#:
+
+{"subject": "Ruth Bader Ginsburg", "disambiguations": "None"}
+
+Example 3:
+
+##CLAIM#:: Charles Osgood, the american television commentator, is best known for hosting CBS News Sunday Morning.
+
+##EXPLANATION#:
+
+The SUBJECT in focus is "Charles Osgood". Based on my world knowledge, there are two notable individuals with similar names: "Charles Osgood - American radio
+
+and television
+
+commentator" and "Charles E. Osgood - American psychologist." Given the ambiguity surrounding the name,
+
+specifying the individual's profession serves as an apt disambiguation method.
+
+##RESPONSE#:
+
+{"subject": "Charles Osgood", "disambiguations": ["Charles Osgood - American radio and television commentator",
+
+"Charles E. Osgood - American psychologist"], "disambiguation_type": "Profession"}
+
+Similarly, disambiguate the following claim by detecting the main SUBJECT and disambiguation information for the SUBJECT using your world knowledge.
+
+Generate an EXPLANATION followed by dictionary-formatted RESPONSE.
+
+##CLAIM#: [claim]
+
+EXPLANATION:
+
+Figure 6: Ambiguity detection prompt for detection of ambiguous entities and generating disambiguation guideline for generation of molecular claims for the baselines MOLECULAR and MOLECULAR-GPT4.
+
+DECONTEXTUALIZATION CRITERIA: Decontextualization adds the right type of information to a CLAIM to make it standalone
+
+and contain relevant disambiguating information. This process can modify the original CLAIM in the following manners:  
+- Substituting pronouns or incomplete names with the specific subject being referred to.  
+- Incorporating the most important distinguishing details such as location/profession/time-period to distinguish the subject from others who might share similar names.  
+- Should not omit information from the original CLAIM.
+
+# Instructions:
+
+- Use the "subject" and the CONTEXT to substitute any incomplete names or pronouns in the CLAIM.  
+- Use the "disambiguation_type" and the CONTEXT to resolve ambiguity by adding clarification phrases about the SUBJECT to the claim.  
+- If information from disambiguation_type is already present in the CLAIM, no decontextualization necessary, return the original claim as is.
+
+# Example 1:
+
+##CLAIM:##: He is best known for hosting CBS News Sunday Morning.
+
+##DISAMBIGUATION GUIDELINE#: {"subject": "Charles Osgood", "disambiguation_type": "Occupation"}
+
+## CONTEXT#: Charles Osgood, a renowned American radio and television commentator and writer, was born on January 8, 1933, in the Bronx, New York City. He is best known for hosting "CBS News Sunday Morning" for over 22 years and "The Osgood File" radio commentaries for over 40 years. Osgood also authored several books, including "The Osgood Files", "See You on the Radio", and "Defending Baltimore Against Enemy Attack". He was born to Charles Osgood Wood, III and his wife Jean Crafton, and grew up with five siblings. Osgood graduated from Fordham University in 1954 with a bachelor of science degree in economics.
+
+##EXPLANATION##: The SUBJECT "He" pertains to "Charles Osgood". The DISAMBIGUATION GUIDELINE indicates that there are multiple individuals named "Charles Osgood", distinguishable by their occupations. The context clarifies that the referenced subject in this claim is Charles Osgood, who is "a renowned American radio and television commentator and writer". Opting for minimal disambiguating information, "a commentator" aligns well with the claim concerning hosting a news show.
+
+#DE CONTEXTUALIZED CLAIM#: Charles Osgood, the commentator, is best known for hosting CBS News Sunday Morning.
+
+# Example 2:
+
+##CLAIM:##: Heyman is the founder of Heyday Films.
+
+##DISAMBIGUATION GUIDELINE#: {"subject": "David Heyman", "disambiguation_type": "Occupation"}
+
+#CONTEXT#: David Heyman is a renowned film producer and founder of Heyday Films, known for producing the entire
+
+"Harry Potter" film series and collaborating with director Alfonso CuarA³n on "Harry Potter and the Prisoner of Azkaban"
+
+and "Gravity". He was born on July 26, 1961, in London. His family has a background in the film industry, with his parents
+
+being a producer and actress. Heyman studied Art History at Harvard University and began his career in the film industry
+
+as a production assistant. Throughout his career, he has received numerous awards and nominations, including an Academy
+
+Award nomination for Best Picture and a BAFTA Award for Best British Film.
+
+EXPLANATION#: The SUBJECT "Heyman" refers to "David Heyman". The DISAMBIGUATION GUIDELINE indicates that there are
+
+multiple individuals named "David Heyman", distinguishable by their occupations. The CONTEXT clarifies that the
+
+referenced SUBJECT in this claim is Heyman which refers to David Heyman and the subject's occupation is film producer.
+
+We opt for minimal disambiguating information by adding "a film producer" as a disambiguation for the SUBJECT.
+
+#DECONTEXTUALIZEDCLAIM#: David Heyman, the film producer, is the founder of Heyday Films.
+
+Now generate an EXPLANATION and DECONTEXTUALIZED CLAIM for the following. Ensure that
+
+only minimal information is added to eliminate ambiguity, such as adjusting pronouns or including
+
+clarifying details. When faced with multiple options for disambiguation under "disambiguation_type,"
+
+prioritize information consistent with the CONTEXT. Avoid repeating information if the claim already
+
+includes information suggested by the "disambiguation_type."
+
+##CLAIM:#: [claim]
+
+##DISAMBIGUATION GUIDELINE##: [disambiguation]
+
+CONTEXT#: [context]
+
+##EXPLANATION#:
+
+Figure 7: Molecular decontextualization prompt for the baselines MOLECULAR and MOLECULAR-GPT4.
+
+DE CONTEXTUALIZATION CRITERIA: Decontextualization adds the right type of information to a CLAIM to make it standalone. This process can modify the original CLAIM in the following manners:
+
+- Substituting pronouns or incomplete names with the specific subject being referred to  
+- Including contextual information to provide more context about the subject.
+
+Instructions: - Identify the "subject" of the claim and locate the claim within the context.
+
+- Use the CONTEXT to substitute any incomplete names or pronouns in the CLAIM.  
+- If there is no decontextualization necessary, return the original claim as is.  
+- The decontextualization should minimally modify the claim by only adding necessary contextual information.  
+- Refer to the following examples to understand the task and output formats.
+
+# Example 1:
+
+Ct  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t  t 0
+
+CLAIM: The victim had previously suffered a broken wrist.
+
+DECONTEXTUALIZEDCLAIM:The Syrian refugee victim in the Almonbury Community School bullying incident had previously suffered a broken wrist.
+
+# Example 2:
+
+CONTEXT: Isaiah Stewart: Stewart was born in Rochester, New York. He grew up playing soccer and boxing.
+
+CLAIM: He grew up playing boxing
+
+DECONTEXTUALIZED CLAIM: Isaiah Stewart grew up playing boxing.
+
+# Example 3:
+
+Context: Arab Serai: According to S.A.A. Naqvi, Mughal emperor Humayun's widow Haji Begum built this "serai" in c. 1560/61 to shelter three hundred Arab mullahs whom she was taking with her during her "hajj" to Mecca; however, Y.D. Sharma opines that the word Arab in the title is a misnomer as this building was built for the Persian craftsmen and workers who built the Humayun's Tomb. In January 2017, the Aga Khan Trust for Culture started a project to conserve the "serai". The restoration was completed in November 2018. In March 2019, the trust announced a planned project to conserve the "baoli" (stepwell) of the serai with the help of funds from the embassy of Germany.  
+CLAIM: The planned project is to conserve the "baoli" (stepwell) of the serai.
+
+DECONTEXTUALIZED CLAIM: The Aga Khan Trust for Culture's planned project in March 2019 is to conserve the "baoli" (stepwell) of the Arab Serai.
+
+# Example 4:
+
+Context: Mason Warren: Warren was born in Doncaster, South Yorkshire and started his career with Rotherham United, where he progressed from the youth team to sign a professional contract in May 2015. He was taken with the first team on the pre-season tour of Scotland and became a regular with the development squad before he was sent to NPL Division One South side Sheffield on a two-month youth loan deal. He was a prominent figure in the side making six appearances during his loan spell before he was recalled in early January 2016. In February 2016, he was loaned out again joining National League North side Harrogate Town on a one-month loan deal. After picking up the Player of the Month award for Harrogate during February, his loan was extended until April. He went on to make a total of eleven appearances for Town. Upon his return to Rotherham in April, he signed a new two-year contract extension until 2018.
+
+CLAIM: He signed a new two-year contract extension until 2018
+
+DECONTEXTUALIZEDCLAIM: Mason Warren Warren signed a new two-year contract extension until 2018 with Rotherham United.
+
+# Example 5:
+
+4AD, which had released three of her albums after her appearance on This Mortal Coil's 1991 album "Blood".
+
+CLAIM: 4AD had released three of her albums.
+
+DE CONTEXTUALIZEDCLAIM: 4AD had released three of Heidi Berry's albums before she left to form Lost Girls.
+
+# Example 6:
+
+Context: Bernard Joseph (politician): He was a member of the Congress of the People before he joined the Economic Freedom Fighters. Joseph said that he left the party because he felt that the party lacked leadership and movement. He joined the Economic Freedom Fighters to implement the party's policies.
+
+CLAIM: He joined the party to implement the party's policies.
+
+DE CONTEXTUALIZED CLAIM: Bernard Joseph joined the Economic Freedom Fighters to implement the party's policies.
+
+# Example 7:
+
+h  t 12018. It was also released shortly before the launch of his "Visceral Tour", based off of his album of the same name.
+
+CLAIM: The album and tour are both named "Visceral".
+
+DECONTEXTUALIZEDCLAIM: Getter's album and tour are both named "Visceral".
+
+Similar, generate a decontextualized claim for the following pair of CLAIM and CONTEXT making minimal alterations to the original structure of the CLAIM while ensuring clarity and coherence.
+
+CONTEXT: <context>
+
+CLAIM: <claim>
+
+DECONTEXTUALIZEDCLAIM:
+
+Figure 8: Decontextualization prompt for the baseline SIMPLE-DE CONTEXT.
+
+**Instructions:**
+
+You are provided with information about different individuals who share similar names under "WIKI DISAMBIGUATIONS." Following this, a PASSAGE about one or more of these subjects is presented. CLAIMS extracted from this passage are then listed.
+
+**Task:**
+
+1. Identify the number of main entities introduced in the passage provided without referring to the WIKI DISAMBIGUATIONS. Determine if the passage describes a single individual or multiple individuals based on the information given.  
+2. For a passage that initially seems to describe one person:
+
+- Identify a core fact from the initial description of the entity in the passage that disambiguates this individual from all the entities in the WIKI DISAMBIGUATIONS. - Use this core fact (such as date of birth, specific educational background, or other unique identifiers) consistently to revise all claims related to that individual until the passage explicitly introduces a new person with the same name.
+
+3. If the passage explicitly introduces a new person (e.g., stating "There is another person called XYZ"), identify a new core fact about this individual directly from the passage. Use this core fact to revise all subsequent claims related to this new individual.  
+4. Consistently apply the selected type of disambiguation (such as profession, birth date, or specific identifier) for each claim about an entity, ensuring all claims about the same entity are revised uniformly.  
+5. Make a revision for each claim with consistent disambiguation information added.
+
+**Output Format:**
+
+Think step by step, and finally provide all claim revisions in a structured list format, where each revised claim is clearly matched with its original claim. Each revised claim should be nested within <revision> tags, ensuring that the chosen method of disambiguation is applied uniformly to all claims about the same entity and that the revision remains succinct. E.g.
+
+-Claim: revision> revised-claim</revision>
+
+WIKI DISAMBIGUATIONS: <disambigs>
+
+PASSAGE: <passage>
+
+CLAIMS: <claims>
+
+Figure 9: Silver labels ambiguity detection prompt for detection of ambiguous entities and generating disambiguation guideline for generation of molecular claims for the baselines MOLECULAR and MOLECULAR-GPT4.
+
+Generate a news article that contains the following **key facts**:
+
+1. Key Fact 1  
+2. Key Fact 2  
+3. [...]
+
+The document should NOT have the following **banned facts**:
+
+1. Banned Fact
+
+Remember that your document must avoid mentioning any of the banned facts, ensuring that they are not hinted at or implied throughout the content. You can use general knowledge and common sense knowledge to create realistic articles that contain the key facts.
+
+Figure 10: Prompt for controlled evidence generation to generate articles that incorporate key facts and avoid banned facts.

2024/Molecular Facts_ Desiderata for Decontextualization in LLM Fact Verification/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:480e3f02c6b684e5671d178e6b4fbe335c3d82b581b4444553a44940bf03c9ed
+size 417607

2024/MoleculeQA_ A Dataset to Evaluate Factual Accuracy in Molecular Comprehension/ad8670ad-ad5d-40a7-a8da-282379fe1329_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b20143aa4091049b264f2d145f0c791f6d1e2cad1c17b97c7dc03b7b9c8f97a4
+size 1374603

2024/MoleculeQA_ A Dataset to Evaluate Factual Accuracy in Molecular Comprehension/full.md

@@ -0,0 +1,547 @@
+# MoleculeQA: A Dataset to Evaluate Factual Accuracy in Molecular Comprehension
+
+Xingyu Lu $^{1,2\dagger}$ , He Cao $^{1,3\dagger}$ , Zijing Liu $^{1}$ , Shengyuan Bai $^{1}$ , Leqing Chen $^{1}$ , Yuan Yao $^{3}$ , Hai-Tao Zheng $^{2,4*}$ , Yu Li $^{1*}$
+
+$^{1}$ International Digital Economy Academy (IDEA)
+
+2Shenzhen International Graduate School, Tsinghua University
+
+$^{3}$ Hong Kong University of Science and Technology  
+$^{4}$  Pengcheng Laboratory, Shenzhen, China, 518055
+
+luxy22@mails.tsinghua.edu.cn hcaoaf@connect.ust.hk {liuzijing,baishengyuan,chenleqing,liyu}@idea.edu.cn yuany@ust.hk zheng.haitao@sz.tsinghua.edu.cn
+
+# Abstract
+
+Large language models are playing an increasingly significant role in molecular research, yet existing models often generate erroneous information. Traditional evaluations fail to assess a model's factual correctness. To rectify this absence, we present MoleculeQA<sup>1</sup>, a novel question answering (QA) dataset which possesses 62K QA pairs over 23K molecules. Each QA pair, composed of a manual question, a positive option and three negative options, has consistent semantics with a molecular description from authoritative corpus. MoleculeQA is not only the first benchmark to evaluate molecular factual correctness but also the largest molecular QA dataset. A comprehensive evaluation on MoleculeQA for existing molecular LLMs exposes their deficiencies in specific aspects and pinpoints crucial factors for molecular modeling. Furthermore, we employ MoleculeQA in reinforcement learning to mitigate model hallucinations, thereby enhancing the factual correctness of generated information.
+
+# 1 Introduction
+
+Large Language Models (LLMs) have practical applications in molecule research (Fabian et al., 2020; Bagal et al., 2021), specifically in the use of cross-modal models (Liu et al., 2023b; Cao et al., 2023). These models bridge the gap between molecular structures and natural language (Bran and Schwaller, 2023), helping experts understand the properties, potential applications, and acquisition methods of specific compounds, improving
+
+understanding of the complex molecular architecture and reducing experimental failure rates.
+
+<table><tr><td>CID</td><td>Ground-Truth</td><td>Generated</td></tr><tr><td>9810996</td><td>The molecule is a dipeptide com-posed of N-(3,3-dimethylbutyl)-L-aspartic acid and methyl L-phenylalanate units joined by a peptide linkage.</td><td>The molecule is a dipeptide obtained by formal condensation of the alpha-carboxy group of N-(3,3-dimethylbutyl)-L-phenylalanine with ethanol.</td></tr><tr><td>10129879</td><td>The molecule is the stable isotope of potassium with relative atomic mass 38.963707.</td><td>The molecule is the stable isotope of tellurium with relative atomic mass 124.904425.</td></tr><tr><td>5281034</td><td>A synthetic androgen, it was mainly used for the treatment of anaemias until being replaced by treatments with fewer side effects.</td><td>It has a role as a contraceptive drug, a progestin and a synthetic oral contraceptive.</td></tr><tr><td>15011611</td><td>It is found in Tripterygium wil-fordii and Tripterygium hypogla-cum.</td><td>The molecule is an abietane diterpenoid isolated from the stem bark of Fraxinus sieboldiana.</td></tr></table>
+
+Table 1: The accuracy of the state-of-the-art mol2text model in describing the given molecule is assessed using BioT5-base (Pei et al., 2023) inference results. The generated content is plausible and fluent, but comparison with the ground truth reveals several factual errors highlighted in red.
+
+Although molecular LLMs are capable of generating task-relevant content, they often suffer from generating illusory statements. In the widely studied molecule captioning (mol2text) task, users expect the model to generate a comprehensive and detailed description of a given compound. In this task, existing benchmarks (Edwards et al., 2022; Liu et al., 2023b) typically employ metrics such as BLEU (Papineni et al., 2001) and ROUGE (Lin, 2004) to evaluate the performance of molecular LLMs. However, without examining the factual accuracy of these models, it is vague to justify how reliable they are. In Table 1, we provide several examples from the CheEBI-20 (Edwards et al., 2021) test dataset to illustrate this issue. Despite the plausible and fluent appearance of the generated content, there are numerous unnoticed inaccurate statements, which remain difficult to detect under the current lexical-based benchmarking approach.
+
+Counterfactual molecular generation content can lead to the following adverse consequences: 1) Misuse of deployed models can deceive and mislead ordinary users, reducing productivity. 2) Professionals may lower their expectations of deployed models when they recognize significant factual errorness, thus hindering positive applications. To avoid these repercussions, quantifying the level of comprehension that models have of molecule knowledge is valuable. However, expertise and professional knowledge are required for human to detect hallucinations in generated molecular text, which is extremely difficult with high cost.
+
+To alleviate the absence of fine-grained factual correctness evaluation for molecular LLMs, we propose MoleculeQA, a comprehensive benchmark based on question-answer pairs covering various aspects including molecular property, source, structure, and application. MoleculeQA endeavors to provide reliable assessments of knowledge comprehension for molecular LLMs, and to offer potential solutions for mitigating model hallucinations.
+
+Construction of MoleculeQA involves two main stages. 1) Molecular Taxonomy Construction. We utilize authoritative molecule description corpus as the source. Using a hybrid approach of rule-based and automated methods, we extract topics based on properties, sources, and other relevant aspects. After clustering and manual normalization, we gather the topics to build a hierarchical domain taxonomy that has broad coverage and strong expertise. 2) Taxonomy-guided QA construction. By converting each molecular description into several pairs of QA that align with the topics at different levels of taxonomy, we can create a QA benchmark that guarantees both granularity, breadth, and quality. MoleculeQA is not only the first factual evaluation benchmark in the molecular domain, but also the largest molecular QA dataset.
+
+Based on MoleculeQA, we perform accuracy tests on various molecular LLMs. Our experimental results indicate that existing methods remain at a discernible remove from achieving a precise comprehension of molecules, and undercover several vital factors for molecule modeling. Furthermore, we utilize MoleculeQA to provide feedback for molecular LLMs' reinforcement learning, aiming to enhance the factual correctness of the models. Our contributions are summarized as follows:
+
+- We reveal the factual inaccuracies in the content generated by existing LLMs in the molecule
+
+or chemistry domain, which have not been adequately detected by existing benchmarks.
+
+- For comprehensive factual accuracy evaluation, we develop a domain taxonomy for molecule corpus and use it to create a high-quality question answering benchmark called MoleculeQA.  
+- Using MoleculeQA, we test a series of models. Based on our experimental outcomes, we identify specific deficiencies in molecular LLMs and summarize several critical factors for molecular understanding. We also attempt to use MoleculeQA as feedback for reinforcement learning to reduce model hallucinations.
+
+# 2 Related Work
+
+# 2.1 Molecule Understanding LLMs
+
+Advancements in language models pre-trained with scientific corpora (Lee et al., 2019; Luo et al., 2022; Beltagy et al., 2019) have shown considerable success in molecular research. Recently, cross-modal models have emerged (Edwards et al., 2021; Luo et al., 2023a; Liu et al., 2023a), aiming to bridge the gap between molecular language (bio-sequence or structure) and natural language. Evaluation tasks for these models include seq2seq generation-based tasks (e.g., molecule captioning and text-based de novo molecule generation) and contrastive-based tasks (e.g., cross-modal retrieval). The corresponding models can be classified as generative models (e.g., MolT5 (Edwards et al., 2022), BioT5 (Pei et al., 2023)) and contrastive models (e.g., MoMu (Su et al., 2022), MoleculeSTM (Liu et al., 2022)).
+
+Seq2seq tasks assess the model's translation ability between modalities. For text-to-molecule generation, metrics include molecule fingerprint similarity (e.g. Morgan-FTS (Schneider et al., 2015)), sequence-based metrics like BLEU (Papineni et al., 2001) and validity. Molecule captioning tasks rely on n-gram precision (BLEU), recall (ROUGE (Lin, 2004)), or both (METEOR (Banerjee and Lavie, 2005)) to measure lexical similarity but lack chemical knowledge comparison and factual correctness detection. Retrieval-type tasks align molecules with descriptions, but overlook fine-grained alignment between text snippets and substructures.
+
+# 2.2 Domain-Specific QA
+
+The Question Answering (QA) task serves as a quantitative measure for evaluating the reasoning and inference capabilities of intelligent systems.
+
+In the general domain, a large number of QA datasets have been constructed (Rajpurkar et al., 2016; Lai et al., 2017; Yang et al., 2018). In addition, specific domains such as medical (Jin et al., 2019, 2020; Pal et al., 2022), news (Nallapati et al., 2016; Trischler et al., 2016), and legal (Zheng et al., 2021; Zhong et al., 2019) have also developed standard QA datasets that are widely used by the community. QA datasets in specific domains can be classified into extraction-based (Pappas et al., 2018), generation-based (Savery et al., 2020), multi-choice (Pal et al., 2022) and Yes / No formats (Jin et al., 2019). QA pairs are constructed from various sources, including scientific articles (Jin et al., 2019), examination problems (Pal et al., 2022; Zaki et al., 2023), professional databases (Liang et al., 2023), and crowd-sourcing data (Wei et al., 2020; Hendrycks et al., 2020).
+
+However, in the molecular domain, there is a scarcity of comprehensive, diverse, and high-quality QA datasets. Existing datasets like DrugChat (Liang et al., 2023) have limitations in terms of molecule features and simplistic answers. BioMedGPT (Luo et al., 2023b) transforms molecule caption task datasets into QA format, inheriting current evaluation issues like domain knowledge deficiency and excessive reliance on lexical similarity. Conversely, MoleculeQA constructs a domain taxonomy and derives QA pairs from descriptive texts, ensuring comprehensive, diverse, high-quality, and credible coverage.
+
+# 3 MoleculeQA Dataset
+
+# 3.1 Exposure of Factual Correctness Issue
+
+In this subsection, we analyze the extent of factual correctness in the generated content of the molecule captioning (mol2text) models.
+
+Setup. To evaluate the reliability of compound descriptions generated by these models, we categorize them into four different aspects: Structure, Property, Application, and Source. The aspects are derived from descriptions in PubChem (Kim et al., 2022), the largest molecule caption dataset currently available. PubChem includes specific sources for each molecule's description, such as Lotus (Mun et al., 2016) for source information, DrugBank (Wishart et al., 2017) for application details, CAMEO Chemicals (cam) for property descriptions, and multiple data repositories for structure information. The definitions of these main aspects are summarized in the Table 2 below.
+
+<table><tr><td colspan="2">Aspect</td><td>Definition</td></tr><tr><td colspan="2">Structure</td><td>Details about architecture, composition, and interaction of atoms within a molecule.</td></tr><tr><td colspan="2">Property</td><td>Physical, biological or chemical property in various environments or reactions.</td></tr><tr><td colspan="2">Application</td><td>The utilization of a molecular compound in various applications and scenarios.</td></tr><tr><td colspan="2">Source</td><td>The natural or synthetic origin, as well as the production context related to a molecule.</td></tr></table>
+
+We randomly sample 100 molecule&caption samples from the ChEBI-20 test set and take MolT5, MoMu, and BioT5 models to generate descriptions for each molecule. Both ground truth and generated content are manually classified based on four aspects. We evaluate the models' descriptions in each aspect against the ground truth, with two trained domain experts judging them as correct (if the generated content matches the ground truth), miss (if the ground truth has a corresponding aspect description but it was completely missing in the generated content), or error (if there is a clear factual inconsistency with the ground truth).
+
+![](images/46e31b95c37e6faab7ef70655dbd21149eb671640bf2fbd337e3f535e7b0f00a.jpg)  
+Figure 1: The performance of three representative models on the traditional metrics for the molecule caption task (e.g. BLEU etc.) and the factual accuracy metric we defined.
+
+Table 2: Evaluation Aspects of description about molecules.  
+
+<table><tr><td>Model</td><td>Structure</td><td>Property</td><td>Application</td><td>Source</td></tr><tr><td>MolT5-base</td><td>63/0/34</td><td>1/4/3</td><td>7/15/8</td><td>20/10/30</td></tr><tr><td>MoMu-base</td><td>63/0/34</td><td>1/4/3</td><td>5/16/9</td><td>19/ 8/33</td></tr><tr><td>BioT5-base</td><td>62/0/35</td><td>2/3/3</td><td>9/12/9</td><td>16/13/31</td></tr></table>
+
+Table 3: Human Assessment of Model Generated Molecular Descriptions based on 4 aspects, with the counts presented according to error / miss / correct.
+
+Results. In Figure 1, we assess the content generated by the model using traditional lexical-based metrics (BLEU, ROUGE, METEOR), as well as their factual accuracy on the selected subset. We define factual accuracy as the ratio of correct predictions to the total number of slots, serving as an average metric to evaluate the reliability of the generated content. Despite the progress in training methodologies, models have exhibited incremental improvements in lexical similarity metrics (such as a  $17.6\%$  increase in BLEU-2). Nevertheless, there is no discernible improvement in the dependability
+
+![](images/9782675bd841a91fd0b4029e3f608a39ae317ac6e0e412a0d8e6a433bceccb8a.jpg)  
+Figure 2: The process of constructing a molecular domain taxonomy. The procedures involve the selection of the information source, extraction of topics, normalization and structuralization of topics, and hierarchical clustering by domain experts.
+
+![](images/0ba13cc6930fb78d59e9180db7c207aeb9c26127fc12853dfab62bffa9b71a86.jpg)  
+(a) Extract Content  
+(b) Reassign topic  
+(c) Design question  
+(d) Collect Pos. & Neg. answers  
+Figure 3: The process of constructing a molecular domain taxonomy. The procedures involve the selection of the information source, extraction of topics, normalization and structuralization of topics, and hierarchical clustering by domain experts.
+
+of the generated content, with factual accuracy persisting at 0.4. In our detailed factual performance analysis (Table 3), we observe that models often omit application-related details and relevant properties. The generated descriptions about Structure show a significant discrepancy rate of more than  $63\%$  compared to ground truth. This challenges the credibility of expert model-generated content, which warrants further scrutiny.
+
+# 3.2 Domain Taxonomy Construction
+
+Taxonomy frameworks organize concepts or entities within a domain hierarchically, aiding in the organization of domain-specific queries (Liu et al., 2012) and ensuring the quality of comprehension domain knowledge and constructing question-answering pairs. We adhere to established procedures for the construction of domain taxonomies, as illustrated in Figure 2.
+
+Information Source. Considering the data quality, we choose the most widely used ChEBI-20 dataset as our molecular description corpus. To mitigate the class imbalance issue in ChEBI-20, primarily dominated by structural information, we include additional sources like T3DB (Wishart et al., 2014), FDA Pharm Classes, and DrugBank. We employ a pre-trained text classifier to perform an initial coarse-grained division of the corpus based on the four aspects we defined above, which serve as the first-level nodes in our taxonomy.
+
+Topics Extraction. We further employ a hybrid approach combining rule-based and few-shot prompting methods to extract topics and their correspond
+
+ing original text from the corpus, formatting the (topic, text) pairs. Subsequently, to mitigate lexical noise and uncontrolled granularity within the 1K topics collected, we utilize GPT-4 (OpenAI, 2023b) with a few-shot prompt-based approach to accomplish an initial semantic aggregation.
+
+Topics Normalization & Structuralization. Next, domain experts intervene to perform rule-based topic merging and concept splitting manually. Finally, the remaining 587 topics are hierarchically clustered by human experts, resulting in a three-level molecular domain taxonomy. An overview of this taxonomy can be found in the Appendix. The leaf nodes represent specific molecule characteristics and are the narrowest topics/concepts, while non-leaf nodes represent broader concepts.
+
+# 3.3 MoleculeQA Construction
+
+Based on the taxonomy in 3.2, we develop a 4-step procedure to extract questions and answers from molecular descriptions to construct MoleculeQA. The whole workflow is displayed in Fig 3.
+
+Content Extraction & Reassign Topic. With (topic, text) pairs annotated in 3.2, a reasonable notion is to query molecules by topic, but content related to a specific topic can be over-brief to be queried. For example, for the molecule CID:5479113, the content of topic acaricide is It has a role as an acaricide. Without enough information, it is difficult to justify which species of mites this molecule is effective. However, it can be queried from a coarser granularity like Therapeutic Usage, the parent topic of
+
+<table><tr><td>Taxonomy</td><td>Reference Description</td><td>Extracted Question</td><td>Positive Answer</td><td>Negative Answer</td></tr><tr><td>Property→Antiviral activity</td><td>It has been shown to exhibit inhibitory effects on the viral neuraminidases from two influenza viral strains, H1N1 and H9N2.</td><td>Which kind of antiviral activity does this molecule have/exhibit?</td><td>It exhibits inhibitory effects on the viral neuraminidases from two influenza viral strains, H1N1 and H9N2.</td><td>It is used for the treatment of cytomegalovirus (CMV) retinitis in AIDS patients.</td></tr><tr><td>Structure→Backbone</td><td>The molecule is a heparan sulfate composed of a backbone of repeating beta-D-glucuronosyl-(1-&gt;4)-N-sulfonyl-alpha-D-glucosamine units joined by (1-&gt;4)-linkages.</td><td>Which kind of backbone does this molecule have?</td><td>It has a backbone of repeating beta-D-glucuronosyl-(1-&gt;4)-N-sulfonyl-alpha-D-glucosamine units joined by (1-&gt;4)-linkages.</td><td>It has a backbone of repeating alpha-L-iduronosyl-(1-&gt;4)-N-sulfonyl-alpha-D-glucosamine units joined by (1-&gt;4)-linkages.</td></tr></table>
+
+# acaricide.
+
+To select a suitable topic for querying, we first use an agent to extract content related to the topic from text. A rule-based program is employed to verify the content, and, in cases where specific details about a given topic are unavailable, we replace the topic with its parent topic until the level of granularity is appropriate for querying purposes.
+
+Question Design. We invite two annotators to design questions for topics based on the extracted contents. For example, contents for topic inhibitor include It is a protein synthesis inhibitor and It is a mitotic inhibitor, annotators may design Which kind of inhibitor is this molecule?. For each topic, annotators discuss choosing the better design as its final question and make sure each question can be answered using the molecular descriptions.
+
+Pos. Options Collection. For the positive options, since formal extracted contents may be rigid and can't be directly used as answers, we leverage the in-context learning capability of ChatGPT (OpenAI, 2023a) to generate appropriate positive options via few-shot prompting.
+
+Neg. Options Collection. For the same question, we take positive options from other molecules as negative candidates for each molecule. To eliminate illegal negatives, we merge synonymous options and remove overlapping options. Then we adopt BioT5 (Pei et al., 2023) to encode all candidates and choose candidates with similar semantics to the positive option as negatives. Several generated QA instances are shown in Table 4.
+
+Data Split. We split molecules in MoleculeQA into train/dev/test sets by scaffolds to divide molecules with similar structures into the same sets as suggested in (Hu et al., 2019), making the QA task more challenging yet realistic.
+
+Quality Control. To provide reliable factual evaluation, LLM and human efforts are combined to ensure MoleculeQA's quality. We convert each QA instance into natural language using templates and assess its logical and semantic consistency with the
+
+original description using ChatGPT. This process is repeated 3 times to minimize variations. With taxonomy guidance, the number of disqualified samples is minimal and can be manually resolved. Human Evaluation. We assign one annotator  ${}^{2}$  to evaluate the reliability of the test split and receive error rate lower than  $1\%$  . Finally,we randomly sample 100 cases and assign two annotators to evaluate the quality of QA samples. The annotators assess the Consistency between the question and the correct option with the reference caption text, as well as Discrimination between the positive and negative options. Human evaluation results can be found in Table 5. The high consistency and discrimination metrics, along with a satisfactory level of agreement (Cohen kappa) among annotators, validate the quality and reliability of our MoleculeQA.
+
+Table 4: Examples of automatically generated QA instances. blue stands for reference locations, red for factual errors.  
+
+<table><tr><td>Metric</td><td>Annotator 1</td><td>Annotator 2</td><td>Agreement (κ)</td></tr><tr><td>Consistency</td><td>99.0</td><td>99.0</td><td>1.0</td></tr><tr><td>Discrimination</td><td>97.0</td><td>96.0</td><td>0.85</td></tr></table>
+
+# 3.4 Data Analysis
+
+Data Statistics. In Table 6, we present the number of QA samples and the coverage of topics in MoleculeQA in comparison to several popular biomolecular and chemistry-related benchmarks (Wei et al., 2020; Yue et al., 2023; Hendrycks et al., 2020; Lu et al., 2022). We observe that MoleculeQA is both the first benchmark focused on evaluating molecular factual knowledge and the largest scale QA dataset in the molecular field.
+
+Table 5: Evaluation for the generated QAs quality.  
+
+<table><tr><td>Benchmarks</td><td>#QA</td><td>Sophistication</td></tr><tr><td>MMLU(Chem)</td><td>534</td><td>College, High school, Medicine</td></tr><tr><td>MMMU(Chem)</td><td>638</td><td>Inorganic, Organic, Physical</td></tr><tr><td>ScienceQA</td><td>867</td><td>Solution, Reaction, Molecule</td></tr><tr><td>ChemistryQA</td><td>4,500</td><td>Reaction, Molecule, Physics</td></tr><tr><td>MoleculeQA</td><td>61,574</td><td>Structure, Source, Property, Application</td></tr></table>
+
+Table 6: Number of samples and topics coverage compared to popular related benchmarks.
+
+The train, development, and test split consists of 49,993, 5,795 & 5,786 QA samples. The general
+
+statistics of the dataset are summarized in Table 7.  
+
+<table><tr><td>Aspects</td><td>Structure</td><td>Property</td><td>Application</td><td>Source</td><td>Total</td></tr><tr><td># Train</td><td>32,176</td><td>4,838</td><td>1,917</td><td>11,062</td><td>49,993</td></tr><tr><td># Dev</td><td>3,314</td><td>698</td><td>558</td><td>1,225</td><td>5,795</td></tr><tr><td># Test</td><td>3,113</td><td>731</td><td>599</td><td>1,343</td><td>5,786</td></tr><tr><td>Avg. Q Tokens</td><td>7.96</td><td>9.02</td><td>7.90</td><td>7.00</td><td>7.74</td></tr><tr><td>Avg. A Tokens</td><td>9.50</td><td>10.98</td><td>11.93</td><td>7.96</td><td>9.42</td></tr></table>
+
+Table 7: MoleculeQA dataset statistics, where Q and A represent the Question and Answer respectively.
+
+Data Distribution. Fig 4 provides the visualized distribution of MoleculeQA. All topics in our taxonomy are queried in MoleculeQA for a comprehensive, fine-grained factual evaluation. Inherited from ChEBI-20, QA pairs in the Structure aspect account for approximately two-thirds of the whole MoleculeQA. While topics within each aspect have relatively balanced sample numbers.
+
+![](images/f2a48dfc07d1b250114c0afc2f6a83f084a290c63029e4c9e23dfb176ca1d2a0.jpg)  
+Figure 4: An overview of MoleculeQA topics distribution. Four coarse-grained aspects occupy the inner circle, and in the outer circle we list finer-grained non-leaf topics.
+
+# 4 Experiment
+
+# 4.1 Baseline Models
+
+The main purpose of baseline experiments is to investigate current models' performance in answering multiple-choice questions related to molecular knowledge. We categorize models based on whether their base LLMs are adequately trained on a large-scale biomolecular corpus as follows:
+
+Molecular LLM, represented by MolT5 (Edwards et al., 2022), MoMu (Su et al., 2022), BioT5 (Pei et al., 2023), MolCA (Liu et al., 2023b) and BioMedGPT-LM-7B (Luo et al., 2023b). These models undergo incremental training stages with extensive molecular modality data (e.g. SMILES or SELFIES strings), biomedical academic papers, and molecule-description pairs.
+
+General LLM, represented by T5 (Raffel et al., 2019), OPT (Zhang et al., 2022), GALACTICA
+
+(Taylor et al., 2022), BLOOM (Scao, 2022), Pythia (Biderman et al., 2023), LLama-2 (Touvron et al., 2023b), along with its instruction fine-tuned derivatives, such as Vicuna (Chiang et al., 2023) and Mol-Instruction-7B (Fang et al., 2023).
+
+Large-scale Universal Models. We evaluate the large-scale, state-of-the-art LLMs in few-shot settings, including open-access models such as Mixtral  $8 \times 7\mathrm{B}$  (Jiang et al., 2024), and OpenAI's GPT family, specifically GPT-3.5 (OpenAI, 2023a) and GPT-4 (OpenAI, 2023b) accessed via API<sup>3</sup>.
+
+# 4.2 Evaluation Setups
+
+We follow training approaches and hyperparameters in the original papers for respective methods. Details about training configuration and few-shot examples are provided in Appendix A.3. Training approaches in our evaluation include:
+
+Full Fine-tuning: All model parameters are updated, including the base LLMs, structure encoders, and projectors for molecule-language alignment.
+
+LoRA-based Fine-tuning: The base LLMs are tuned by low-rank adaptation (Hu et al., 2021), and structure encoders are also trainable.
+
+Few-shot Setting: We sample 10 QA examples from four aspects respectively to prompt LLMs with task definition and contextual information.
+
+The main metric of MoleculeQA is the accuracy, which is defined as the ratio of correctly answered samples among all test samples. We present the accuracy in four aspects as well as the total accuracy.
+
+# 4.3 Main Results
+
+We summarize the benchmarking results in Table 8:
+
+- Comparison over four aspects. Achieving the highest accuracy on Source is generally more feasible for each model, whereas addressing Property and Application presents notable difficulties, with no method surpassing a  $50\%$  accuracy rate. This phenomenon may be ascribed to the comparatively smaller data scale within these domains.  
+- Molecular LLMs v.s. General LLMs. Molecular LLMs demonstrate better performance, with a minimum total accuracy over  $51\%$ . By contrast, other than T5s, decoder-only General LLMs fail to achieve a total accuracy exceeding  $50\%$ , whether fully fine-tuned or tuned with LoRA.  
+- T5 series comparison. Among T5-based methods, T5 demonstrates superior performance compared to MolT5 (e.g., T5-base surpasses
+
+<table><tr><td>Model</td><td># Trainable Params</td><td>Implementation</td><td>Structure</td><td>Source</td><td>Property</td><td>Application</td><td>Total</td></tr><tr><td>Random</td><td>-</td><td>-</td><td>24.41</td><td>22.30</td><td>23.04</td><td>24.57</td><td>24.03</td></tr><tr><td colspan="8">Molecular LLM</td></tr><tr><td>MolT5-small</td><td>80M</td><td>full ft</td><td>49.59</td><td>64.18</td><td>46.51</td><td>40.90</td><td>51.69</td></tr><tr><td>MolT5-base</td><td>250M</td><td>full ft</td><td>58.01</td><td>65.85</td><td>45.14</td><td>42.24</td><td>55.39</td></tr><tr><td>MoMu-small</td><td>82M</td><td>full ft</td><td>52.71</td><td>63.44</td><td>44.87</td><td>40.57</td><td>52.96</td></tr><tr><td>MoMu-base</td><td>252M</td><td>full ft</td><td>61.58</td><td>65.30</td><td>43.78</td><td>43.07</td><td>57.43</td></tr><tr><td>BioT5-base</td><td>252M</td><td>full ft</td><td>65.98</td><td>69.24</td><td>49.11</td><td>40.73</td><td>62.03</td></tr><tr><td>MolCA-125M</td><td>100M</td><td>LoRA ft</td><td>65.54</td><td>67.34</td><td>45.77</td><td>40.33</td><td>60.30</td></tr><tr><td>MolCA-1.3B</td><td>110M</td><td>LoRA ft</td><td>71.12</td><td>70.98</td><td>47.81</td><td>43.17</td><td>64.79</td></tr><tr><td>BioMedGPT-LM-7B</td><td>40M</td><td>LoRA ft</td><td>54.19</td><td>60.01</td><td>38.85</td><td>40.90</td><td>52.23</td></tr><tr><td colspan="8">General LLM</td></tr><tr><td>T5-small</td><td>60M</td><td>full ft</td><td>55.51</td><td>64.41</td><td>45.42</td><td>38.56</td><td>54.55</td></tr><tr><td>T5-base</td><td>220M</td><td>full ft</td><td>60.42</td><td>66.42</td><td>45.83</td><td>43.74</td><td>58.24</td></tr><tr><td>OPT-125M</td><td>125M</td><td>full ft</td><td>38.58</td><td>55.92</td><td>41.04</td><td>28.73</td><td>42.93</td></tr><tr><td>OPT-350M</td><td>331M</td><td>full ft</td><td>44.39</td><td>60.83</td><td>46.24</td><td>40.57</td><td>48.05</td></tr><tr><td>GALACTICA-6.7B</td><td>12.5M</td><td>LoRA ft</td><td>32.35</td><td>41.92</td><td>31.05</td><td>28.21</td><td>33.96</td></tr><tr><td>BLOOM-7.1B</td><td>27.5M</td><td>LoRA ft</td><td>35.01</td><td>47.51</td><td>31.46</td><td>33.56</td><td>37.31</td></tr><tr><td>Pythia-6.9B</td><td>29.4M</td><td>LoRA ft</td><td>42.79</td><td>58.90</td><td>38.58</td><td>39.07</td><td>45.61</td></tr><tr><td>Mol-Instruction-7B</td><td>40M</td><td>LoRA ft</td><td>37.46</td><td>47.36</td><td>32.69</td><td>29.88</td><td>38.37</td></tr><tr><td>Llama-2-7B-chat</td><td>40M</td><td>LoRA ft</td><td>28.75</td><td>39.84</td><td>31.33</td><td>27.71</td><td>31.54</td></tr><tr><td>Llama-2-13B-chat</td><td>63M</td><td>LoRA ft</td><td>34.37</td><td>43.86</td><td>31.05</td><td>29.72</td><td>35.67</td></tr><tr><td>Vicuna-v1.5-7B</td><td>40M</td><td>LoRA ft</td><td>34.89</td><td>44.15</td><td>34.20</td><td>31.55</td><td>36.61</td></tr><tr><td>Vicuna-v1.5-13B</td><td>63M</td><td>LoRA ft</td><td>37.01</td><td>43.19</td><td>30.64</td><td>31.55</td><td>37.07</td></tr><tr><td colspan="8">Large-scale Universal Models</td></tr><tr><td>Mixtral-8×7B-Instruct-v0.1</td><td>-</td><td>10-shot</td><td>23.32</td><td>31.87</td><td>32.89</td><td>29.96</td><td>27.79</td></tr><tr><td>GPT-3.5-1106-turbo</td><td>-</td><td>10-shot</td><td>25.60</td><td>37.60</td><td>28.04</td><td>32.22</td><td>29.29</td></tr><tr><td>GPT-4-1106-preview</td><td>-</td><td>10-shot</td><td>60.94</td><td>50.19</td><td>35.57</td><td>43.91</td><td>53.47</td></tr></table>
+
+Table 8: We report the accuracy (%) results on MoleculeQA test set under different aspects (Best for model-wise).
+
+MolT5-base in total accuracy by  $5.1\%$  contradicting their performance on molecule caption tasks. BioT5 combines bio-molecular texts and databases for molecular pretraining, achieving higher total accuracy than T5  $(+6.5\%)$ .
+
+- Decoder-only LLMs comparison. Among Llama-based models, BioMedGPT-7B achieves the best performance with incremental pretraining, while Mol-Instruction fine-tuned by instructions has slight improvement than Llama and Vicuna. With the similar size of the base model (7B) and LoRA parameters, the performance ranking among different models is as follows: Pythia  $>$  BLOOM  $>$  GALACTICA  $>$  Llama, which may provide a reference for molecular base model selection. Increasing model size (e.g. 7B  $\rightarrow$  13B) also receives mild accuracy gain.  
+- Single vs. Multiple modalities. Both MoMu and MolCA are models that jointly incorporate molecular 2D graph modality and textual information. They demonstrate improvements over their base models (MolT5 and GALACTICA respectively) that solely rely on 1D-text modality.  
+- Large-scale Universal Models. The utilization of highly advanced models, such as GPT-4, has potential in the field of molecular research. In a 10-shot scenario, GPT-4 demonstrates accuracy comparable to certain specialized models. However, the performance of smaller models declined
+
+sharply, which may be attributed to the lack of their emergent abilities((Wei et al., 2022)).
+
+# 5 Analysis
+
+We propose the following research questions (RQs) for the molecular domain to guide our analysis:
+
+- RQ1: Are existing LLMs powerful enough for application in practical molecular scenarios?  
+- RQ2: What factors are crucial for enhancing LLMs' ability for molecule comprehension?  
+- RQ3: Can MoleculeQA be adopted to alleviate the hallucinations in molecular LLMs?
+
+# 5.1 In-depth Performance Analysis (RQ1)
+
+We draw a preliminary conclusion from Table 8 that existing LLMs' comprehension of molecules is far from satisfactory: When confronted with aspects of Property and Application, pivotal for real-world applications, evaluated models consistently fail to achieve commendable accuracy. To more thoroughly assess the methods' level of comprehension across various molecular aspects, we plot T5-base and BioT5's accuracy over each subcategory in our taxonomy in Fig. 5. We find that in aspects of Source and Structure, two models exhibit consistent performance, with accuracy exceeding  $40\%$  across all categories. But on sub-topics like Agricultural Chemical and Approval status, two models perform notably sub-optimal. Various ac
+
+![](images/8d59db441dbc5fb5687795006b9526caf080a84bc27a30ee27000c86991bb783.jpg)  
+Figure 5: Accuracy of different finer topics under 4 coarse-grained aspects on the MoleculeQA test set. We select BioT5- and T5-base as representatives of Molecular LLM and General LLM, respectively, represented by solid and dashed bars.
+
+curacy on different topics can serve as a confidence coefficient for related model applications.
+
+# 5.2 Crucial Factor Attribution (RQ2)
+
+We summarize the following crucial factors for improving molecular comprehension ability:
+
+Molecular Corpora. The two T5 variants, MolT5 and BioT5, displays divergent outcomes. MolT5 performs worse compared to T5, while BioT5 demonstrates improved performance. This divergence can be attributed to the differences in their training corpora, specifically in terms of scale and diversity. Similarly, decoder-only models also exhibit this phenomenon: BioMedGPT (4.2M bio-molecular papers) > Mol-Instruction (1M molecular-oriented instruction samples) > Vicuna (70K general instruction samples) > Llama (General corpus). The above findings emphasize the importance of large, diverse, and high-quality molecular corpus for improving performance.
+
+Modality Modeling Strategy. We investigate which modality modeling strategies can more effectively facilitate molecular modeling. (1) Modality learning: There is a significant performance gap between LoRA-based methods and methods employing multi-modal fusion or full fine-tuning, which underscores that an adequate scale of trainable parameters is necessary to master the molecule modalities. (2) Multi-modal fusion: MolCA and MoMu demonstrate that fusing molecular graphs into the semantic space of LLMs is viable. However, although they both deploy GIN as graph encoder, in comparison to MoMu's linear adaptation, MolCA's Q-Former (Li et al., 2022) graph adapter achieves a much more significant improvement.
+
+# 5.3 Hallucination Alleviation (RQ3)
+
+Reinforcement Learning (RL) from feedback has widespread applications for mitigating hallucinations (Yu et al., 2024; Gunjal et al., 2024). However, this method is rarely applied to molecule caption (Gkoumas and Liakata, 2024). To verify the fea
+
+sibility of this approach, we adopt MoleculeQA to provide feedback to optimize the fine-tuned molecule caption models: Given a QA pair of molecule  $x$ , we designate the positive option as the preferred output  $y_{w}$  and one negative option as the dis-preferred output  $y_{l}$ , and employ Direct Preference Optimization (Rafailov et al., 2023) (DPO) as the RL strategy to optimize the model  $\pi$ :
+
+$$
+\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} \left(\pi_ {\theta}; \pi_ {\mathrm {r e f}}\right) = - \mathbb {E} _ {\left(x, y _ {w}, y _ {l}\right) \sim \mathcal {D}} \left[ \log \sigma \left(\right. \right. \\ \beta \log \left(\frac {\pi_ {\theta} \left(y _ {w} \mid x\right)}{\pi_ {\text {r e f}} \left(y _ {w} \mid x\right)}\right) \\ \left. \left. - \beta \log \left(\frac {\pi_ {\theta} \left(y _ {l} \mid x\right)}{\pi_ {\mathrm {r e f}} \left(y _ {l} \mid x\right)}\right)\right) \right] \\ \end{array}
+$$
+
+where  $\pi_{\theta}$  is the policy model parameterised by  $\theta$ $,\pi_{ref}$  is the fine-tuned model as the reference,  $\beta$  is a hyper-parameter and  $\sigma$  is the Sigmoid function.
+
+We convert QA instances into mol2text format and remove molecules in ChEBI-20's test set, after optimizing MolT5 and BioT5 on this corpus with DPO, we evaluate their factual correctness like Section 3.1 and report the result in Table 9.
+
+<table><tr><td>Model</td><td>Structure</td><td>Property</td><td>Application</td><td>Source</td></tr><tr><td>MolT5-base</td><td>63/0/34</td><td>1/4/3</td><td>7/15/8</td><td>20/10/30</td></tr><tr><td>MolT5-base-DPO</td><td>59/0/38</td><td>0/2/6</td><td>10/13/7</td><td>17/ 8/35</td></tr><tr><td>BioT5-base</td><td>62/0/35</td><td>2/3/3</td><td>9/12/9</td><td>16/13/31</td></tr><tr><td>BioT5-base-DPO</td><td>57/1/39</td><td>1/2/5</td><td>11/10/9</td><td>14/14/33</td></tr></table>
+
+Table 9: Comparison about Factual Correctness. We manually evaluate two optimized models on the same 100 cases. An intuitive comparison is provided in Table 12.
+
+The result indicates that, two models are guided by counterfactual negative options to discern correct/incorrect fine-grained molecular facts, and to generate descriptions that align better with ground truth across most aspects, except for Application. We attribute this to the small scale of Application.
+
+# 6 Conclusion and Future Work
+
+In conclusion, this paper addresses the absence of evaluation for factual correctness in Large Language Models (LLMs) within the molecular domain. By organizing molecular descriptions into a taxonomy and constructing QA pairs through human and LLM efforts, we introduce MoleculeQA, a novel dataset for molecular factual question answering. Our evaluation reveals shortcomings of existing models, emphasizing critical factors for molecular comprehension and providing guidance for molecular LLMs' development. We also make preliminary attempt to alleviate the hallucinations in molecular LLMs based on MoleculeQA. Looking forward, we propose three future directions: (1) Design a powerful molecular model based on our analysis. (2) Investigate more and better methods to apply MoleculeQA for the optimization of molecular LLMs. (3) Incorporate additional data sources to enrich MoleculeQA's comprehensiveness.
+
+# 7 Limitations
+
+We conclude our limitations into the following aspects: (1) Imbalanced data distribution across different aspects, notably with Structure and Source data dominating the majority. This skew results from the overall prevalence of structural and source-related information in the data sources. To address this, future efforts will focus on introducing more data related to properties and applications while expanding topic coverage and diversity, all while safeguarding against data leakage. (2) Absence of full fine-tuning for large models: Under the constraint of computational resources, we fail to fully fine-tune LLMs with 7B parameters and above, leading us to opt for adaptation-based fine-tuning methods. (3) We only conduct a preliminary attempt to alleviate the issue of model hallucinations, the potential of MoleculeQA is left for further exploration.
+
+# 8 Potential Risks
+
+Although MoleculeQA offers a viable approach for factual assessment in the molecular domain with reliable data quality, there remains a risk of misuse. Evaluations on this dataset may not accurately represent a model's comprehension over all molecules. MoleculeQA could potentially be leveraged to furnish a veneer of reliability for models with underlying risks.
+
+# 9 Acknowledgements
+
+This project was supported by National Natural Science Foundation of China (Grant No. 62276154), Shenzhen Hetao Shenzhen-Hong Kong Science and Technology Innovation Cooperation Zone (Grant No. HTHZQSWSKCCYB-2023052), Research Center for Computer Network (Shenzhen) Ministry of Education, the Natural Science Foundation of Guangdong Province (Grant No. 2023A1515012914 and 440300241033100801770), Basic Research Fund of Shenzhen City (Grant No. JCYJ20210324120012033 and GJHZ202402183000101), the Major Key Project of PCL for Experiments and Applications (PCL2021A06). Finally, we thank the valuable comments of anonymous reviewers and area chair.
+
+# References
+
+CAMEO Chemicals. https://cameochemicals.noaa.gov/.
+
+Viraj Bagal, Rishal Aggarwal, P. K. Vinod, and U. Deva Priyakumar. 2021. MolGPT: Molecular generation using a transformer-decoder model. Journal of chemical information and modeling.  
+Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL.  
+Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A pretrained language model for scientific text. In Conference on Empirical Methods in Natural Language Processing.  
+Stella Biderman, Hailey Schoelkopf, Quentin G. Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar van der Wal. 2023. Pythia: A suite for analyzing large language models across training and scaling. ArXiv, abs/2304.01373.  
+Andrés M Bran and Philippe Schwaller. 2023. Transformers and large language models for chemistry and drug discovery. ArXiv, abs/2310.06083.  
+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. *ArXiv*, abs/2005.14165.  
+He Cao, Zijing Liu, Xingyu Lu, Yuan Yao, and Yu Li. 2023. InstructMol: Multi-modal integration for building a versatile and reliable molecular assistant in drug discovery.  
+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 open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality.  
+Carl N. Edwards, T. Lai, Kevin Ros, Garrett Honke, and Heng Ji. 2022. Translation between molecules and natural language. *ArXiv*, abs/2204.11817.  
+Carl N. Edwards, Chengxiang Zhai, and Heng Ji. 2021. Text2Mol: Cross-modal molecule retrieval with natural language queries. In Conference on Empirical Methods in Natural Language Processing.  
+Benedek Fabian, Thomas Edlich, H'el'ena Gaspar, Marwin H. S. Segler, Joshua Meyers, Marco Fiscato, and Mohamed Ahmed. 2020. Molecular representation learning with language models and domain-relevant auxiliary tasks. *ArXiv*, abs/2011.13230.  
+Yin Fang, Xiaozhuan Liang, Ningyu Zhang, Kangwei Liu, Rui Huang, Zhuo Chen, Xiaohui Fan, and Hua-jun Chen. 2023. Mol-instructions: A large-scale biomolecular instruction dataset for large language models. ArXiv, abs/2306.08018.  
+Dimitris Gkoumas and Maria Liakata. 2024. Feedback-aligned mixed llms for machine language-molecule translation. arXiv preprint arXiv:2405.13984.  
+Anisha Gunjal, Jihan Yin, and Erhan Bas. 2024. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 18135-18143.  
+Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Xiaodong Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. ArXiv, abs/2009.03300.  
+J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. 2021. LoRA: Low-rank adaptation of large language models. ArXiv, abs/2106.09685.  
+Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay S. Pande, and Jure Leskovec. 2019. Strategies for pre-training graph neural networks. arXiv: Learning.  
+Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Emma Bou Hanna, Florian Bressand, Gianna
+
+Lengyel, Guillaume Bour, Guillaume Lample, L'elio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven Le Scao, Théophile Gervet, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2024. Mixtral of experts. ArXiv, abs/2401.04088.  
+Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2020. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. ArXiv, abs/2009.13081.  
+Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William W. Cohen, and Xinghua Lu. 2019. PubMedQA: A dataset for biomedical research question answering. In Conference on Empirical Methods in Natural Language Processing.  
+Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindultyte, Jia He, Siqian He, Qingliang Li, Benjamin A. Shoemaker, Paul A. Thiessen, Bo Yu, Leonid Y. Zaslavsky, Jian Zhang, and Evan E. Bolton. 2022. Pubchem 2023 update. Nucleic acids research.  
+Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale reading comprehension dataset from examinations. Cornell University - arXiv, Cornell University - arXiv.  
+Hugo Laurenccon, Lucile Saulnier, Thomas Wang, Christopher Akiki, Albert Villanova del Moral, Teven Le Scao, Leandro von Werra, Chenghao Mou, Eduardo Gonz'alez Ponferrada, Huu Nguyen, Jorg Frohberg, Mario vSavsko, Quentin Lhoest, Angelina McMillan-Major, Gerard Dupont, Stella Biderman, Anna Rogers, Loubna Ben Allal, Francesco De Toni, Giada Pistilli, Olivier Nguyen, So maieh Nikpoor, Maraim Masoud, Pierre Colombo, Javier de la Rosa, Paulo Villegas, Tristan Thrush, S. Longpre, Sebastian Nagel, Leon Weber, Manuel Sevilla Munoz, Jian Zhu, Daniel Alexander van Strien, Zaid Alyafeai, Khalid Almubarak, Minh Chien Vu, Itziar Gonzalez-Dios, Aitor Soroa Etxabe, Kyle Lo, Manan Dey, Pedro Ortiz Suarez, Aaron Gokaslan, Shamik Bose, David Ifeoluwa Adelani, Long Phan, Hieu Trung Tran, Ian Yu, Suhas Pai, Jenny Chim, Violette Lepercq, Suzana Ilic, Margaret Mitchell, Sasha Luccioni, and Yacine Jernite. 2023. The bigscience roots corpus: A 1.6tb composite multilingual dataset. ArXiv, abs/2303.03915.  
+Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2019. BioBERT: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36:1234 - 1240.  
+Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pretraining for unified vision-language understanding and generation. In ICML.
+
+Youwei Liang, Ruiyi Zhang, Li Zhang, and Peng Xie. 2023. DrugChat: Towards enabling chatgpt-like capabilities on drug molecule graphs. ArXiv, abs/2309.03907.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Annual Meeting of the Association for Computational Linguistics.  
+Shengchao Liu, Weili Nie, Chengpeng Wang, Jiarui Lu, Zhuoran Qiao, Ling Liu, Jian Tang, Chaowei Xiao, and Anima Anandkumar. 2022. Multi-modal molecule structure-text model for text-based retrieval and editing. ArXiv, abs/2212.10789.  
+Xueqing Liu, Yangqiu Song, Shixia Liu, and Haixun Wang. 2012. Automatic taxonomy construction from keywords. In Knowledge Discovery and Data Mining.  
+Zequn Liu, W. Zhang, Yingce Xia, Lijun Wu, Shufang Xie, Tao Qin, Ming Yang Zhang, and Tie-Yan Liu 2023a. MolXPT: Wrapping molecules with text for generative pre-training. ArXiv, abs/2305.10688.  
+Zhiyuan Liu, Sihang Li, Yancheng Luo, Hao Fei, Yixin Cao, Kenji Kawaguchi, Xiang Wang, and Tat-Seng Chua. 2023b. Molca: Molecular graph-language modeling with cross-modal projector and uni-modal adapter. In Conference on Empirical Methods in Natural Language Processing.  
+Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Michael Kinney, and Daniel S. Weld. 2020. S2orc: The semantic scholar open research corpus. In Annual Meeting of the Association for Computational Linguistics.  
+Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and A. Kalyan. 2022. Learn to explain: Multimodal reasoning via thought chains for science question answering. *ArXiv*, abs/2209.09513.  
+Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon, and Tie-Yan Liu. 2022. BioGPT: Generative pre-trained transformer for biomedical text generation and mining. Briefings in bioinformatics.  
+Yi Luo, Kai Yang, Massimo Hong, Xingyi Liu, and Za- qing Nie. 2023a. MolFM: A multimodal molecular foundation model. ArXiv, abs/2307.09484.  
+Yi Luo, Jiahuan Zhang, Siqi Fan, Kai Yang, Yushuai Wu, Mu Qiao, and Zaiqing Nie. 2023b. BioMedGPT: Open multimodal generative pre-trained transformer for biomedicine. ArXiv, abs/2308.09442.  
+Terry Mun, Asger Bachmann, Vikas Gupta, Jens Stougaard, and Stig U Andersen. 2016. emphLotus base: An integrated information portal for the model legume emphLotus japonicus. Sci Rep, 6:39447.
+
+Ramesh Nallapati, Bowen Zhou, Cicero Nogueira dos Santos, Caglar Gulçehre, and Bing Xiang. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In Conference on Computational Natural Language Learning.  
+OpenAI. 2023a. ChatGPT: A language model for conversational ai.  
+OpenAI. 2023b. GPT-4 technical report. ArXiv, abs/2303.08774.  
+Ankit Pal, Logesh Kumar Umapathi, and Malaikanan Sankarasubbu. 2022. MedMCQA: A large-scale multi-subject multi-choice dataset for medical domain question answering. In ACM Conference on Health, Inference, and Learning.  
+Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2001. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL '02.  
+Dimitris Pappas, Ion Androutsopoulos, and Haris Papageorgiou. 2018. BioRead: A new dataset for biomedical reading comprehension. In International Conference on Language Resources and Evaluation.  
+Qizhi Pei, Wei Zhang, Jinhua Zhu, Kehan Wu, Kaiyuan Gao, Lijun Wu, Yingce Xia, and Rui Yan. 2023. BioT5: Enriching cross-modal integration in biology with chemical knowledge and natural language associations. ArXiv, abs/2310.07276.  
+Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. arxiv 2023. arXiv preprint arXiv:2305.18290.  
+Colin Raffel, Noam M. Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2019. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1-140:67.  
+Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing.  
+Max E. Savery, Asma Ben Abacha, Soumya Gayen, and Dina Demner-Fushman. 2020. Question-driven summarization of answers to consumer health questions. Scientific Data, 7.  
+Tven Le Scao. 2022. Bloom: A 176b-parameter open-access multilingual language model. ArXiv, abs/2211.05100.  
+Nadine Schneider, Roger A. Sayle, and Gregory A. Landrum. 2015. Get your atoms in order - an open-source implementation of a novel and robust molecular canonicalization algorithm. Journal of chemical information and modeling, 55 10:2111-20.
+
+T. Sterling and John J. Irwin. 2015. Zinc 15 - ligand discovery for everyone. Journal of Chemical Information and Modeling, 55:2324 - 2337.  
+Bing Su, Dazhao Du, Zhao-Qing Yang, Yujie Zhou, Jiangmeng Li, Anyi Rao, Haoran Sun, Zhiwu Lu, and Ji rong Wen. 2022. A molecular multimodal foundation model associating molecule graphs with natural language. ArXiv, abs/2209.05481.  
+Ross Taylor, Marcin Kardas, Guillem Cucurull, Thomas Scialom, Anthony S. Hartshorn, Elvis Saravia, Andrew Poulton, Viktor Kerkez, and Robert Stojnic. 2022. Galactica: A large language model for science. ArXiv, abs/2211.09085.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. LLaMA: Open and efficient foundation language models. ArXiv, abs/2302.13971.  
+Hugo Touvron, Louis Martin, Kevin R. Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Daniel M. Bikel, Lukas Blecher, Cristian Cantón Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony S. Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel M. Kloumann, A. V. Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, R. Subramanian, Xia Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zhengxu Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023b. Llama 2: Open foundation and fine-tuned chat models. ArXiv, abs/2307.09288.  
+Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip Bachman, and Kaheer Suleman. 2016. NewsQA: A machine comprehension dataset. In Rep4NLP@ACL.  
+Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. 2022. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682.  
+Zhuoyu Wei, Wei Ji, Xiubo Geng, Yining Chen, Baihua Chen, Tao Qin, and Daxin Jiang. 2020. Chemistry: A complex question answering dataset from chemistry.  
+David Scott Wishart, David Arndt, Allison Pon, Tanvir Sajed, Anchi Guo, Yannick Djoumbou, Craig Knox,
+
+Michael Wilson, Yongjie Liang, Jason R. Grant, Yifeng Liu, Seyed Ali Goldansaz, and Stephen M. Rappaport. 2014. T3db: the toxic exposome database. Nucleic Acids Research, 43:D928 - D934.  
+David Scott Wishart, Yannick Djoumbou Feunang, Anchi Guo, Elvis J. Lo, Ana Marcu, Jason R. Grant, Tanvir Sajed, Daniel Johnson, Carin Li, Zinat Sayeeda, Nazanin Assempour, Ithayavani Iynkkaran, Yifeng Liu, Adam Maciejewski, Nicola Gale, Alex Wilson, Lucy Chin, Ryan Cummings, Diana Le, Allison Pon, Craig Knox, and Michael Wilson. 2017. Drugbank 5.0: a major update to the drugbank database for 2018. Nucleic Acids Research, 46:D1074 - D1082.  
+Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing.  
+Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. 2024. Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13807-13816.  
+Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. 2023. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. ArXiv, abs/2311.16502.  
+Mohd Zaki, Jayadeva, Mausam, and N. M. Anoop Krishnan. 2023. MaScQA: A question answering dataset for investigating materials science knowledge of large language models. ArXiv, abs/2308.09115.  
+Zheni Zeng, Yuan Yao, Zhiyuan Liu, and Maosong Sun. 2022. A deep-learning system bridging molecule structure and biomedical text with comprehension comparable to human professionals. Nature Communications, 13.  
+Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona T. Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2022. Opt: Open pre-trained transformer language models. ArXiv, abs/2205.01068.  
+Lucia Zheng, Neel Guha, Brandon R. Anderson, Peter Henderson, and Daniel E. Ho. 2021. When does pretraining help?: assessing self-supervised learning for law and the CaseHOLD dataset of 53,000+
+
+legal holdings. Proceedings of the Eighteenth International Conference on Artificial Intelligence and Law.
+
+Haoxiang Zhong, Chaojun Xiao, Cunchao Tu, T. Zhang, Zhiyuan Liu, and Maosong Sun. 2019. JEC-QA: A legal-domain question answering dataset. ArXiv, abs/1911.12011.
+
+# A Appendix
+
+# A.1 Data Sources and License
+
+As depicted in Table 11, we elaborate on the origins and legal permissions associated with each data component utilized in the development of the MoleculeQA. This encompasses both biomolecular data and textual descriptions. Thorough scrutiny was conducted on all data origins to confirm compatibility with our research objectives and subsequent utilization. Proper and accurate citation of these data sources is consistently maintained throughout the paper.
+
+# A.2 Details about Taxonomy
+
+We present the overall hierarchical structure of the taxonomy upon which MoleculeQA is based in Figure 6. Additionally, Table 10 provides details regarding the subtopics and part of leaf topics encompassed within each of the four aspects: Structure, Source, Property, and Application.
+
+# A.3 Experimental Setup Details
+
+# A.3.1 Baselines
+
+The following parts will individually introduce the models we evaluated in this study and the approaches used for implementation.
+
+T5 (Raffel et al., 2019) is an encoder-decoder model pre-trained on a multi-task mixture of unsupervised and supervised tasks for which each task is converted into a text-to-text format. We directly fine-tuned it on MoleculeQA dataset from public checkpoints with three different model sizes: small, base and large. It's important to note that the original T5 pre-training does not incorporate any specific knowledge related to the domain of molecules.
+
+MolT5 (Edwards et al., 2022) undergoes joint training on molecule SMILES from the ZINC-15 dataset (Sterling and Irwin, 2015) and a general corpus from the C4 dataset (Raffel et al., 2019), enabling MolT5 to acquire prior knowledge in
+
+both of these domains. It contains three different sizes: small, base, and large. In the experiment, we utilized pre-trained model checkpoints of various sizes released by the authors. Subsequently, we conducted full fine-tuning on the MoleculeQA train set, followed by evaluating on the test set.
+
+MoMu (Su et al., 2022) is pre-trained using molecular 2D graphs and their semantically related textual data (crawl from published Scientific Citation Index papers) via contrastive learning. We adopted MoMu-K pre-trained checkpoints  ${}^{6}$  where the text encoder is initialized with the weights of KV-PLM (Zeng et al., 2022). Following the original methodology, we injected encoded graph features into MolT5-base & large and conducted finetuning on MoleculeQA.
+
+BioT5 (Pei et al., 2023) as a comprehensive pretraining framework, builds upon the methodology of MolT5 while enhancing cross-modal integration into biology through chemical knowledge and natural language associations. It leverages SELFIES for robust molecular representations and extracts knowledge from the surrounding context of bioentities in unstructured biological literature. We utilized the official base version pre-trained checkpoint and converted the MoleculeQA data into the corresponding format for fine-tuning.
+
+MolCA (Liu et al., 2023b) facilitates a language model (LM), such as Galactica, in comprehending both text- and graph-based molecular contents through its cross-modal projector. This projector, implemented as a Q-Former, serves to bridge the representation space of a graph encoder with the text space of an LM. Additionally, MolCA employs a uni-modal adapter to enable efficient adaptation of the LM to downstream tasks. We conducted pretraining, including both stage 1 and stage 2, on the 125M and 1.3B versions, based on the official code and cleaned data<sup>8</sup>. Subsequently, we performed finetuning on MoleculeQA.
+
+BioMedGPT-LM-7B (Luo et al., 2023b) It is a large generative language model based on Llama2 in the biomedical domain. It was fully fine-tuned from the Llama2-7B-Chat with millions of biomed
+
+![](images/594ef9bd69c9b558f14f3797bcdc63b8559ac07aa365862c71820c8fc711c35e.jpg)  
+Figure 6: The overarching structure of the MoleculeQA taxonomy comprises multiple aspects and subtopics arranged hierarchically to categorize various facets of molecular factual knowledge. Due to space constraints, we did not elaborate on all leaf topics.
+
+<table><tr><td>ASPECT</td><td>SUB TOPICS</td><td>LEAF TOPICS</td></tr><tr><td rowspan="7">Property</td><td>Biological and Pharmacological Activities</td><td>&quot;antimicrobial activity&quot;, &quot;anti-neoplastic activity&quot;, &quot;antioxidant activity&quot;, &quot;enzyme inhibition&quot;, &quot;ion channel activity&quot;, &quot;receptor activity&quot;...</td></tr><tr><td>Reaction Types</td><td>&quot;acetylation&quot;, &quot;condensation&quot;, &quot;dehydrogenation&quot;, &quot;epoxidation&quot;, &quot;glycosylation&quot;, &quot;hydroxylation&quot;, &quot;oxidation&quot;, &quot;phosphorylation&quot;, &quot;reduction&quot;...</td></tr><tr><td>Chemical Interaction and Mechanism</td><td>&quot;action&quot;, &quot;affinity&quot;, &quot;binding&quot;, &quot;conversion&quot;, &quot;decomposition&quot;, &quot;duration&quot;, &quot;formation&quot;, &quot;mechanism&quot;, &quot;reaction/binding&quot;, &quot;receptor affinity&quot;, &quot;selectivity&quot;...</td></tr><tr><td>Chemical Properties</td><td>&quot;chemical nature&quot;, &quot;sensitivity&quot;, &quot;ph value&quot;, &quot;stability&quot;, &quot;valence&quot;, &quot;reactivity&quot;</td></tr><tr><td>Environmental and Safety Concerns</td><td>&quot;bio-accumulation&quot;, &quot;xenobiotic&quot;, &quot;cell permeability&quot;, &quot;teratogenic agent&quot;, &quot;environmental contaminant&quot;, &quot;resistance&quot;, &quot;safety concerns&quot;</td></tr><tr><td>Medical and Therapeutic Efficacy</td><td>&quot;analgesic activity&quot;, &quot;anti-inflammatory activity&quot;, &quot;antimalarial activity&quot;, &quot;anti-mycobacterial&quot;, &quot;carcinogenicity&quot;, &quot;medical effects&quot;, &quot;potency&quot;...</td></tr><tr><td>Physical and Sensory Properties</td><td>&quot;abundance&quot;, &quot;atomic mass&quot;, &quot;boiling point&quot;, &quot;color&quot;, &quot;half-life&quot;, &quot;odor&quot;, &quot;optical activity&quot;, &quot;physical state&quot;, &quot;solubility&quot;, &quot;taste&quot;, &quot;volatility&quot;...</td></tr><tr><td rowspan="7">Application</td><td>Agricultural Chemicals</td><td>&quot;fungicide&quot;, &quot;herbicide&quot;, &quot;insecticide&quot;, &quot;disease control&quot;, &quot;herbicide safener&quot;, &quot;synthetic auxin&quot;, &quot;phytoestrogen&quot;...</td></tr><tr><td>Biological Agents</td><td>&quot;antibiotic&quot;, &quot;antifungal drug&quot;, &quot;antibacterial drug&quot;, &quot;antiprotozoal&quot;, &quot;antiviral drug&quot;, &quot;nematicide&quot;, &quot;acaricide&quot;, &quot;antiseptic&quot;...</td></tr><tr><td>Chemical Applications and Techniques</td><td>&quot;reagent&quot;, &quot;indicator&quot;, &quot;detection&quot;, &quot;derivatisation agent&quot;, &quot;fluorescent dye&quot;, &quot;production&quot;, &quot;chromatographic reagent&quot;, &quot;tracer&quot;, &quot;solvent&quot;, &quot;food additive&quot;...</td></tr><tr><td>Pharmacodynamics and Pharmacokinetics</td><td>&quot;Inhibitor&quot;, &quot;antagonist&quot;, &quot;prodrug&quot;, &quot;modulator&quot;, &quot;sympathomimetic agent&quot;, &quot;allergen&quot;, &quot;sodium channel blocker&quot;, &quot;ligand&quot;, &quot;agonist&quot;...</td></tr><tr><td>Regulatory Status and Approval</td><td>&quot;approval&quot;, &quot;withdrawn from market&quot;, &quot;registered in&quot;...</td></tr><tr><td>Research and Development</td><td>&quot;experimental&quot;, &quot;biomarker&quot;, &quot;clinical development&quot;, &quot;testing&quot;...</td></tr><tr><td>Therapeutic Use</td><td>&quot;anti-arrhythmia drug&quot;, &quot;anti-allergic agent&quot;, &quot;anti-asthmatic drug&quot;, &quot;an-ticoronaviral agent&quot;, &quot;anti-neoplastic agent&quot;, &quot;anti-ulcer drug&quot;, &quot;anti-HIV agent&quot;, &quot;orphan drug&quot;, &quot;recreational drug&quot;, &quot;vasodilator&quot;...</td></tr><tr><td rowspan="4">Source</td><td>found in</td><td>&quot;found in&quot;</td></tr><tr><td>metabolite</td><td>&quot;metabolite&quot;</td></tr><tr><td>derives from</td><td>&quot;derivates&quot;</td></tr><tr><td>isolated from</td><td>&quot;isolated&quot;</td></tr><tr><td rowspan="6">Structure</td><td>Biochemical and Biological Terms</td><td>&quot;active metabolite&quot;, &quot;alkaloid&quot; &quot;coenzyme a&quot;, &quot;enzyme&quot;, &quot;epitope&quot;, &quot;fatty acyl coa&quot;, &quot;glucoside&quot;, &quot;hapten&quot;, &quot;nucleobase&quot;, &quot;oligosaccharide&quot;, &quot;sphingoid base&quot;, &quot;substrate&quot;...</td></tr><tr><td>Chemical Bonding and Interactions</td><td>&quot;glycosidic bond&quot;, &quot;disulfide bonds&quot;, &quot;double bond&quot;, &quot;exocyclic double bond&quot;, &quot;peptide bond&quot;, &quot;c=c double bond&quot;, &quot;bond&quot;, &quot;connection&quot;, &quot;attachment&quot;...</td></tr><tr><td>Chemical Compounds and Classes</td><td>&quot;acid&quot;, &quot;alcohol&quot;, &quot;amine&quot;, &quot;cation&quot;, &quot;dimer&quot;, &quot;enamide&quot;, &quot;hydrochloride&quot;, &quot;ion&quot;, &quot;lactam&quot;, &quot;polyphenol&quot;, &quot;salt&quot;, &quot;phosphate&quot;, &quot;sulfate&quot;, &quot;oxanion&quot;, &quot;zwitterion&quot;...</td></tr><tr><td>Chemical Species and States</td><td>&quot;anhydrous form&quot;, &quot;heptahydrate form&quot;, &quot;oxidation state&quot;, &quot;hydrate&quot;, &quot;major microspecies&quot;, &quot;deoxygenated&quot;, &quot;major species&quot;, &quot;microspecies&quot;...</td></tr><tr><td>Functional Groups and Chemical Entities</td><td>&quot;acyl group&quot;, &quot;alcohol group&quot;, &quot;alkyl group&quot;, &quot;anilino group&quot;, &quot;carbamoyl group&quot;, &quot;chloro group&quot;, &quot;epoxy group&quot;, &quot;ester group&quot;, &quot;fatty acyl group&quot;, &quot;hydrazino group&quot;, &quot;hydroperoxy group&quot;, &quot;isopropyl substituent&quot;, &quot;keto group&quot;, &quot;methyl group&quot;, &quot;oxo group&quot;, &quot;pentyl group&quot;, &quot;phosphate group&quot;, &quot;primary hydroxy group&quot;, &quot;s-acyl component&quot;, &quot;s-methyl group&quot;, &quot;sulfo group&quot;, &quot;thiol group&quot;...</td></tr><tr><td>Molecular Structure and Configuration</td><td>&quot;alpha-branch&quot;, &quot;alpha-carbon&quot;, &quot;backbone&quot;, &quot;branch&quot;, &quot;bridge&quot;, &quot;core&quot;, &quot;composition&quot;, &quot;configuration&quot;, &quot;linked group&quot;, &quot;n-substituent&quot;, &quot;oh groups&quot;, &quot;omega-hydroxy&quot;, &quot;position&quot;, &quot;prenyl units&quot;, &quot;terminal&quot;, &quot;terminal group&quot;, &quot;glycosyl fragment&quot;, &quot;repeating unit&quot;, &quot;sequence&quot;, &quot;subcomponents&quot;, &quot;side chain&quot;, &quot;nucleus&quot;, &quot;sugar fragment&quot;, &quot;unit&quot;...</td></tr></table>
+
+Table 10: Taxonomy of Property, Application, Structure and Source aspects in MoleculeQA. Leaf Topics correspond to the most granular concepts, while Sub Topics aggregate leaf topics further. The table presents only a subset of leaf topics.
+
+<table><tr><td>DATA SOURCES</td><td>LICENSE URL</td><td>LICENSE NOTE</td></tr><tr><td>PubChem</td><td>https://www.nlm.nih.gov/web_policies.html</td><td>Works produced by the U.S. government are not subject to copyright protection in the United States. Any such works found on National Library of Medicine (NLM) Web sites may be freely used or reproduced without permission in the U.S.</td></tr><tr><td>FDA Pharm Classes</td><td>https://www.fda.gov/about-fda/about-website/website-policies</td><td>Unless otherwise noted, the contents of the FDA website (www.fda.gov), both text and graphics, are not copyrighted. They are in the public domain and may be republished, reprinted and otherwise used freely by anyone without the need to obtain permission from FDA. Credit to the U.S. Food and Drug Administration as the source is appreciated but not required.</td></tr><tr><td>Drug Bank</td><td>https://creativecommons.org/licenses/by-nc/4.0/legalcode</td><td>Subject to the terms and conditions of this Public License, the Licensor hereby grants You a worldwide, royalty-free, non-sublicisable, non-exclusive, irrevocable license to exercise the Licensed Rights in the Licensed Material to: reproduce and Share the Licensed Material, in whole or in part, for NonCommercial purposes only; and produce, reproduce, and Share Adapted Material for NonCommercial purposes only.</td></tr><tr><td>ChEBI</td><td>https://creativecommons.org/licenses/by/4.0/</td><td>You are free to: Share — copy and redistribute the material in any medium or format. Adapt — remix, transform, and build upon the material for any purpose, even commercially.</td></tr><tr><td>LOTUS</td><td>https://lotus.nprod.net/</td><td>LOTUS is one of the biggest and best-annotated resources for natural products occurrences available free of charge and without any restriction.</td></tr><tr><td>CAMEO Chemicals</td><td>https://cameochemicals.noaa.gov/help/reference/terms_and_conditions.htm?d_f=false</td><td>CAMEO Chemicals and all other CAMEO products are available at no charge to those organizations and individuals (recipients) responsible for the safe handling of chemicals.</td></tr><tr><td>Toxin-Toxin-Target Database (T3DB)</td><td>http://www.t3db.ca/</td><td>T3DB is offered to the public as a freely available resource. Use and re-distribution of the data, in whole or in part, for commercial purposes requires explicit permission of the authors and explicit acknowledgment of the source material (T3DB) and the original publication.</td></tr></table>
+
+Table 11: Data resources and licenses utilized in data collection for MoleculeQA.
+
+ical papers from the S2ORC corpus (Lo et al., 2020). We directly apply the LoRA finetuning method on the checkpoint  ${}^{9}$  provided by the official source.
+
+OPT (Zhang et al., 2022) is a series of open-sourced large causal language models which perform similar in performance to GPT-3 (Brown et al., 2020). For comparison with fully fine-tuned T5 series models, we opted to fully fine-tune OPT-125M, -350M, and -1.3B size models on MoleculeQA. In our implementation, we referred to the interfaces provided by Hugging Face  $^{10}$ .
+
+GALACTICA (Taylor et al., 2022) is a large language model (LLM) for Science: trained on over 48 million papers, textbooks, reference material, compounds, proteins and other sources of scientific knowledge. We selected GALACTICA-125M, -1.3B, and -7.1B versions of the model  $^{11}$  and conducted fine-tuning using LoRA on MoleculeQA.
+
+Pythia (Biderman et al., 2023) is an open suite of large language models, all trained on public data in the same order. These models vary in size, ranging
+
+from 70M to 12B parameters. They were trained on the Pile dataset, which is constructed from 22 diverse high-quality subsets. We opted to conduct finetuning based on LoRA on the standard versions of Pythia-410M, -1B, -2.8B, -6.9B, and -12B sizes models $^{12}$ .
+
+BLOOM (Scao, 2022) is an autoregressive large language model, trained to continue text from a prompt on vast amounts of text data using industrial-scale computational resources. It was trained on the ROOTS (Laurencon et al., 2023) corpus, a dataset comprising hundreds of sources in 46 natural and 13 programming languages (59 in total). For model scaling evaluation, we chose to conduct finetuning based on LoRA on the BLOOM-560M, -1.1B, -1.7B, -3B, and -7.1B sizes versions of the model  $^{13}$ . Subsequently, we provided the results on the MoleculeQA test set.
+
+LLaMA-2 (Touvron et al., 2023b) is a collection of large language models with parameters ranging from 7 billion to 70 billion. The model architecture remains largely unchanged from that of LLaMA-1 models (Touvron et al., 2023a), but  $40\%$  more data
+
+was used to train the foundational models. Specifically, Llama 2 includes pre-trained and fine-tuned models optimized for dialogue applications, termed Llama 2-Chat. We opted to utilize the LLaMA2-Chat 7B and 13B models  $^{14}$  and transformed MoleculeQA into instruction samples for LoRA fine-tuning.
+
+Vicuna-v-1.5 (Chiang et al., 2023) is an open-source chatbot that has been trained by fine-tuning LLaMA on over 150K user-shared conversations collected from ShareGPT.com. Preliminary evaluation, conducted with GPT-4 as the judge, demonstrates that the Vicuna series achieves competitive performance when compared to OpenAI ChatGPT, while also outperforming other models such as LLaMA. We selected the v1.5 series models and conducted LoRA Finetuning on both the 7B and 13B versions $^{15}$ .
+
+Mol-Instructions-7B (Fang et al., 2023) is a low-rank adapter designed for LLaMA-2 base LLM, specifically trained on molecule-oriented instructions sourced from the Mol-Instructions dataset. We utilize the version tailored for LLaMA-2-Chat  $^{16}$ , merging the adapter back to the base LLM before proceeding with LoRA fine-tuning.
+
+Mixtral-8×7B (Jiang et al., 2024) is a Sparse Mixture of Experts (SMoE) language model consisting of a decoder-only architecture. Its feedforward block selects from a set of 8 distinct groups of parameters. Notably, it is recognized as the most robust open-weight model currently available, licensed under Apache 2.0. We adopt a locally deployed approach for conducting few-shot prompting inference.
+
+GPT-3.5-turbo and GPT-4. For closed-source models such as OpenAI GPT Family GPT-3.5-turbo (OpenAI, 2023a) and GPT-4 (OpenAI, 2023b), we employ batch inference via APIs for conducting few-shot prompt inference. This approach significantly enhances evaluation efficiency and reduces overhead.
+
+# A.3.2 Hyper-parameters
+
+For MolT5, MoMu, T5, and BioT5, we employed the original codebases and hyper-parameters provided in the respective papers for full fine-tuning.
+
+Specifically, these models were trained on a single NVIDIA 48GB A6000 GPU. Except for BioT5, which had a learning rate set to 1e-3, the learning rates for all other models were set to 1e-4. All models underwent fine-tuning for 100 epochs on the training set, and the checkpoint with the best performance on the development set was selected for evaluation on the test set.
+
+For MolCA, we utilized the author's recently updated dataset (excluding any data leakage concerns) and conducted pre-training stage 1 and stage 2 training on 2 NVIDIA 48GB A6000 GPUs. We maintained consistency with the training hyperparameters provided in the original paper. Subsequently, we fine-tuned pre-trained checkpoints of different sizes on MoleculeQA, with a total batch size set to 16. The 125M model was trained on a single GPU card, while the 1.3B model was trained on two GPU cards. The fine-tuning total epochs were set to 100 for all versions.
+
+For full fine-tuning of the OPT series, we conducted training on 4 A6000 GPUs for the 125M and 350M versions and 8 GPUs for the 1.3B version. The total batch size was set to 256, and the learning rates were set to 3e-4 and 2e-4 for the respective versions. All other hyper-parameters were kept consistent with those specified in the original paper. We performed full fine-tuning for 60 epochs, as we observed over-fitting phenomena when exceeding 50 epochs.
+
+For the remaining experiments based on LoRA tuning, we employed the Alpaca-LoRA codebase for instruction fine-tuning. Except for the 13B size model trained on 8 A6000 GPUs, all other models were trained on 4 GPUs. The total batch size was set to 400, with gradient accumulation and learning rate adjusted according to the model size (typically set to 3e-4). We set the total training epochs to 20.
+
+Regarding the LoRA configuration, we utilized the PEFT library for implementation. We set LoRA's rank  $r$  as 16,  $\alpha$  as 16, dropout rate as 0.05, and applied LoRA to all modules of ["q/k/v/o Projekt", "gate Projekt", "down/up Projekt"] (adjusting module names if necessary based on actual implementation). Equivalent trainable parameters are reported in Table 8.
+
+We implemented DPO using the alignment-handbook library and retained all the hyperparameters.
+
+![](images/2316d41f8a8c4a0351c6df0677a4262365e2aa33d58b5de46ab9d9ab44ba8d4d.jpg)  
+Figure 7: Model parameter size vs. total accuracy on MoleculeQA test set. Solid lines denote full fine-tune models, and dashed lines represent LoRA fine-tune.
+
+# A.4 Scaling Law for Molecular LLMs
+
+In Fig. 7 and Fig. 8, we depict the variations in overall accuracy and aspect-specific accuracy of several models over increasing model scale.
+
+For fully fine-tuned models, we conduct comparisons between T5-based models (T5 and MolT5) and decoder-only models (represented by OPT). To validate whether adaptively fine-tuning can adapt general LLMs to acquire molecular knowledge, we compared models such as BLOOM, Pythia, and the LLaMA2-series models using LoRA fine-tuning.
+
+We observe a pronounced scaling effect across different training methods and model architectures, with the scale effect being more evident in the full fine-tuning approaches. This observation is consistent with previous analysis about the scale of parameters and indicates that scaling up model size is a promising way to enhance molecular modeling.
+
+# A.5 Examples of Hallucination Alleviation
+
+Here, we present selected examples (Table 12) comparing the originally generated captions with those refined using the DPO method discussed in Section 5.3
+
+# A.6 Prompt of Different Tasks for LLM
+
+In the construction process of MoleculeQA, we deploy LLMs to finish the following tasks: (1) Corpus classification; (2) Topic extraction; (3) Answer generation; (4) Semantic consistency validation. We report the definitions and task contexts, which are components of prompts for LLMs, of these tasks in Table 13.
+
+# A.7 Few-Shot Details and Prompt Exhibition
+
+We introduce details about our few-shot setting: For each aspect, we select a representative and various samples from different topics as examples to construct an aspect-specific 10-shot prompt. We demonstrate selected samples and the format of prompt in Source aspect in Table 14.
+
+![](images/7e0195a08567550ba3194156bb6155192a4b94a9ed7b248531daed686b4e1f43.jpg)
+
+![](images/79556946b0f9a6c62eb9dc1880916081322c026e5f6276ddf784173aa64f4113.jpg)
+
+![](images/e2087d8096e5856e02121dc7c9dd47451716f31c7d6e8661c7a79865bc6fb1c3.jpg)  
+Figure 8: Model parameter size vs. Accuracy in four aspects.
+
+![](images/a76c552bd2772ab36649cbf0e4270713171d3782673b20f8c97687637baf9bcc.jpg)
+
+<table><tr><td>CID</td><td>Ground-Truth</td><td>Original</td><td>DPO Refined</td></tr><tr><td>5280695</td><td>The molecule is a monomethoxyflavone that is the 8-O-methyl derivative of gossypetin. It is a pentahydroxyflavone, a member of flavonols and a monomethoxyflavone. It derives from a gossypetin. It is a conjugate acid of a 3&#x27;,4&#x27;,5,7-pentahydroxy-8-methoxyflavon-3-olate.</td><td>The molecule is a monomethoxyflavone that is quercetin methylated at position 7. It has a role as a metabolite, an antioxidant and an anti-inflammatory agent. It is a monomethoxyflavone, a tetrahydroxyflavone and a 7-hydroxyflavonol. It derives from a quercetin.</td><td>The molecule is a 7-hydroxyflavonol that is quercetin substituted by a methoxy group at position 8. It has a role as a plant metabolite. It is a tetrahydroxyflavone, a 7-hydroxyflavonol, a monomethoxyflavone and a member of flavonols. It derives from a quercetin.</td></tr><tr><td>91972219</td><td>The molecule is an amino disaccharide that consists of N-acetyl-D-glucosamine having an N-acetyl-6-O-sulfo-beta-D-galactosaminyl residue attached at O-4. It is an amino disaccharide, a glucosamine oligosaccharide and a galactosamine oligosaccharide.</td><td>The molecule is a linear amino disaccharide that consists of N-acetyl-beta-D-glucosamine and N-acetyl-D-galactosamine units connected in sequence by (1-&gt;4) links. It has a role as an epitope. It is an oligosaccharide sulfate, an amino disaccharide and a glucosamine oligosaccharide.</td><td>The molecule is an amino disaccharide consisting of N-acetylglucosamine having an O-sulfated N-acetylglucosaminyl residue attached at the 4-position. It has a role as an epitope. It is an amino disaccharide, a glucosamine oligosaccharide and a galactosamine oligosaccharide.</td></tr><tr><td>53239760</td><td>The molecule is a fifteen-membered glycopeptide comprising glycyl, glutaminyl, alanyl, glycyl, 3-(1,3-thiazol-4-yl)alanyl, (5R)-5-(beta-D-galactopyranosyloxy)lysl, glycyl, alpha-glutamyl, glutaminyl, glycyl, prolyl, lysyl, glycyl, alpha-glutamyl and threonine residues coupled in sequence.</td><td>The molecule is a fifteen-membered glycopeptide comprising glycyl, glutaminyl, alanyl, glycyl, 3-(1,3)-alpha-D-glutamyl, (5R)-5-(beta-D-galactopyranosyloxy)lysl, (5R)-5-(beta-D-glucopyranosyloxy)lysl, glycyl, alpha-glutamyl, glutaminyl, glutaminyl, prolyl, lysyl, glycyl, alpha-glutamyl and threonyl residues coupled in sequence.</td><td>The molecule is a fifteen-membered glycopeptide comprising glycyl, glutaminyl, alanyl, glycyl, 3-(1,3-thiazol-4-yl)alanyl, (5R)-5-(beta-D-galactopyranosyloxy)lysl, glycyl, alpha-glutamyl, glutaminyl, glycyl, prolyl, lysyl, glycyl, alpha-glutamyl, and threonine residues coupled in sequence.</td></tr></table>
+
+Table 12: Comparison between the original generated captions of BioT5-base and those generated by DPO-refined BioT5-base discussed in Section 5.3. Errors relative to the ground truth are highlighted in red. Notably, the model trained using data transformed by MoleculeQA and the DPO method exhibits fewer factual inaccuracies.
+
+<table><tr><td>TASK</td><td>DEFINITION</td><td>TASK CONTEXT</td></tr><tr><td>Corpus Classification</td><td>Classify molecular descriptions from the data source into one of four aspects.</td><td>You are a research assistant for molecular research. Please help me to classify some corpus. Four kinds of content are included in this corpus : The first is Source, which describes... The second is...</td></tr><tr><td>Topic Extraction</td><td>Extract attributes of molecules in specific aspect from original descriptions.</td><td>You are a chemical research assistant, you are familiar with description text of molecules, you need to help me extract molecules’ Source information, which describes...</td></tr><tr><td>Answer Generation</td><td>Generate answer for given question with original description</td><td>You are a chemistry research assistant, and I need you to complete the following task: You will be given a detailed description of a molecule and a question, please extract specific information from the given description to answer the question...</td></tr><tr><td>Semantic Consistency Validation</td><td>Check if generated answer has consistent semantic with original description.</td><td>You are a chemistry research assistant, and I need you to complete the following task: You will be given a description of a molecule and a sentence transcribed from it, please justify whether their semantics are consistent...</td></tr></table>
+
+Table 13: Definition and context for each task. We prompt LLMs to finish these tasks for MoleculeQA construction.
+
+```python
+messages = [ {"role": "system", "content": f"..."
+```
+
+You are a chemistry research assistant, and I'd like to test your professional ability on molecule understanding, please complete the following task:
+
+You are provided with the SMILES representation of a molecule and asked a question about the molecule's source-related knowledge (Source means the natural or synthetic origin, as well as the production context related to a molecule), with four options given. Three of these options do not describe the given molecule, and you must select the correct option.
+
+Here are several examples to show how to finish the Question Answering task:
+
+#####
+
+Example 1:
+
+```txt
+Molecular SMILES: C1=CC(=CC=C1/C=CC(=O)O[C@@H]([C@H](C(=O)O)O)C(=O)O)O
+```
+
+Question: Which molecule does this molecule derive from?
+
+Choices:
+
+A: It derives from a meso-tartaric acid and a cis-4-coumaric acid.  
+B: It derives from a meso-tartaric acid and a cis-caffeic acid.  
+C: It derives from a cyanidin cation and a cis-4-coumaric acid.  
+D: It derives from a cis-vaccenic acid and an oleic acid.
+
+Answer: A
+
+##
+
+#####
+
+Example 2:
+
+```javascript
+Molecular SMILES:  $\mathrm{COC1 = C(C = C(C = C1)C = O)OC}$
+```
+
+Question: Where this molecule can be found?
+
+Choices:
+
+A: It can be found in leaves and fruit of cowberry Vaccinium vitis-idaea, grape seeds and beer.  
+B: It can be found in peppermint, ginger, raspberry, and other fruits.  
+C: It can be found in edible vegetables, grains, and fruits.  
+D: It can be found in grape seeds, in Hibiscus cannabinus (kenaf) root and bark, in apple and in cacao.
+
+Answer: B
+
+#####
+
+Notice that here are some rules you need to follow:
+
+1. Your answer for each question should be one of A/B/C/D, which corresponds to the four options.  
+2. For my convenience, please give me a list of ANSWERs for the given instances in the format 'Answer X: ...', without any other information.
+
+```txt
+```
+{
+    "role": "user", "content": f""
+}
+```
+
+Please give me your choices for these instances in the above examples' styles. No other information is required.
+
+```txt
+Instance ID: <Instance ID>
+Molecular SMILES: <Instance SMILES>
+Question: <Instance Question>
+Choices: <Instance Choices>
+"\"\"}
+]
+```
+
+Table 14: An illustration depicting the process of constructing few-shot in-context-learning prompts for MoleculeQA test set inference with GPT-4-like large-scale universal models.

2024/MoleculeQA_ A Dataset to Evaluate Factual Accuracy in Molecular Comprehension/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:249df3e44149d636860174a1c57af1e2c9e4f086bacbbc6dbd0b288ef9a1f310
+size 1769619

2024/Monotonic Paraphrasing Improves Generalization of Language Model Prompting/2959e73c-f9ae-48a1-8e85-33f3e7a21ae3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e2a83107ec5d9fd818c4239abba4643d0c4fa68a183d4fd062e844acb210dfa
+size 1095296

2024/Monotonic Paraphrasing Improves Generalization of Language Model Prompting/full.md

@@ -0,0 +1,413 @@
+# Monotonic Paraphrasing Improves Generalization of Language Model Prompting
+
+Qin Liu Fei Wang Nan Xu Tianyi Yan Tao Meng Ucla Muhao Chen
+
+UC Davis; USC; UCLA
+
+{qinli, muhchen}@ucdavis.edu;
+
+{fwang598, nanx, tianyi}@usc.edu; tmeng@cs.ucla.edu
+
+# Abstract
+
+Performance of large language models (LLMs) may vary with different prompts or instructions for even the same task. One commonly recognized factor for this phenomenon is the model's familiarity with the given prompt or instruction, which is typically estimated by its perplexity. However, finding the prompt with the lowest perplexity is challenging, given the enormous space of possible prompting phrases. In this paper, we propose monotonic paraphrasing (MONOPARA), an end-to-end decoding strategy that paraphrases given prompts or instructions into their lower perplexity counterparts based on an ensemble of a paraphrase LM for prompt (or instruction) rewriting, and a target LM (i.e. the prompt or instruction executor) that constrains the generation for lower perplexity. The ensemble decoding process can efficiently paraphrase the original prompt without altering its semantic meaning, while monotonically decreasing the perplexity of each generation as calculated by the target LM. We explore in detail both greedy and search-based decoding as two alternative decoding schemes of MONOPARA. Notably, MONOPARA does not require any training and can monotonically lower the perplexity of the paraphrased prompt or instruction, leading to improved performance of zero-shot LM prompting as evaluated on a wide selection of tasks. In addition, MONOPARA is also shown to effectively improve LMs' generalization on perturbed and unseen task instructions.<sup>1</sup>
+
+# 1 Introduction
+
+Large language models (LLMs) have demonstrated remarkable proficiency in zero-shot decision making (Gonen et al., 2023; Schick and Schütze, 2021; Brown et al., 2020) and instruction following (Jiang et al., 2023; Köpf et al., 2023; Touvron et al., 2023b; Taori et al., 2023; Chiang et al., 2023;
+
+Ouyang et al., 2022). However, there can be significant variance in the performance of seemingly similar prompts (Zhao et al., 2021; Lu et al., 2022c; Webson and Pavlick, 2022; Gonen et al., 2023; Yan et al., 2024). Despite efforts of studies on prompting LMs (Shin et al., 2020; Li and Liang, 2021; Gao et al., 2021; Ding et al., 2022; Sanh et al., 2021; Kojima et al., 2022), it is still challenging to develop high-quality prompts that can induce better performance for varying tasks on evolving models in an effort-saving manner.
+
+One consensus reached by recent studies is the inverse relationship between a prompt's perplexity and its task performance (Kumar et al., 2023; Gonen et al., 2023). This stems from the intuition that the frequency of a prompt (or an instruction) $^{2}$  appearing in the (pre-)training data positively influences the model's familiarity with it, thereby enhancing its capability to perform the described task (Iter et al., 2023; Gonen et al., 2023; Wang et al., 2023; Lee et al., 2023). As a result, existing attempts use specific criteria, especially perplexity for selecting prompts from a collection of candidates. For example, Gonen et al. (2023) builds a large prompt pool for each task and selects the one with the lowest perplexity. However, the performance of such rewrite-then-select methods (Jiang et al., 2020; Zhou et al., 2022; Gonen et al., 2023; Prasad et al., 2023) is limited by the size, quality, and even availability of the candidate pool. Searching the prompt with the lowest perplexity for a particular task remains challenging due to the vast expanse of potential prompts.
+
+To address this challenge, we propose a novel progressive end-to-end prompt refining approach, namely monotonic paraphrasing (MONOPARA), that can proactively rewrite the given prompt into its low-perplexity counterpart without compromises
+
+ing its expressiveness of the task. During prompt (or instruction) refinement, MONOPARA conducts an ensemble decoding process of a paraphrasing model together with the target LM. The paraphrase model is instructed to rewrite the given prompt of a specific task, while the target model, which is later on used for task inference, provides a constraint that aims to lower the perplexity of the generation. MONOPARA iteratively decodes tokens based on the ensemble of these two models. Intuitively, the paraphrase model can genuinely paraphrase the original prompt without altering its semantic meaning and remain instructive for the given task, while the target model restricts the search space of paraphrased tokens to those with low perplexity, resulting in a task prompt that the target model is more familiar with.
+
+Further, for the decoding process of MONOPARA, we explore two decoding schemes. In addition to greedy decoding (§2.3) that iteratively generates the next token based on the weighted combination of predicted probabilities from both models, we also explored search-based decoding (§2.4). The search-based decoding scheme follows the look-ahead decoding paradigm (Lu et al., 2022b), which keeps several sequence candidates and maintains the one with the lowest perplexity scored by the target model. Compared to recent prompt refinement approaches (Gonen et al., 2023; Kumar et al., 2023; Shum et al., 2023), MONOPARA needs zero training effort and far less computational cost, and do not require the pre-existence of multiple prompt or instruction candidates (Gonen et al., 2023). We find that by an ensemble of two LMs, we are able to consistently decode lower perplexity prompts, which would, in turn, bring about better generalization of LMs on downstream tasks.
+
+Our contributions are three-fold. First, we propose MONOPARA, a prompt refinement method that paraphrases prompts to be more familiar with the target model for boosted generalization and task performance. Second, we explore and compare two distinct decoding strategies that proactively refine prompts by monotonically lowering their perplexity and maintaining their expressiveness of the task. Third, we conduct experimentation for both prompt refinement for regular LMs and instruction refinement for instruction-tuned LMs to illustrate that monotonic paraphrasing improves the generalization of LM prompting.
+
+![](images/c5b1eed8ed4abd78692874b1dcf4f1582d88c22e793dd39fa4992c7ff3381170.jpg)  
+Figure 1: Perplexity of  $x_{para}$  as the output paraphrase of  $P_{para}$  vs. as the input prompt of  $P_{tar}$  for the AG News dataset with Mistral 7B as both  $P_{para}$  and  $P_{tar}$ . Each point stands for a different prompt  $x_{para}$ . A low-perplexity paraphrase does not necessarily result in a low-perplexity prompt for the target model.
+
+# 2 Monotonic Paraphrasing
+
+We propose MONOPARA as an ensemble-based decoding method that can refine a given prompt by paraphrasing it into a low-perplexity counterpart. We first provide the intuition of MONOPARA in §2.1 and the basics of prompt paraphrasing in §2.2. Then, we formally define two decoding schemes of MONOPARA in §2.3 and §2.4.
+
+# 2.1 Intuition and Methodology
+
+Assume that the user wants to address a task specified by a given prompt  $x_{\text{ori}}$  using a target LM  $P_{\text{tar}}$ . MONOPARA seeks to refine the given prompt to its low-perplexity counterpart  $x_{\text{para}}$  (i.e. a more familiar prompt to  $P_{\text{tar}}$ ), thereby enhancing model performance. However, simply paraphrasing the original prompt with a paraphrase model  $P_{\text{para}}$  does not necessarily result in a low-perplexity counterpart. This discrepancy arises from a mismatch between the perplexity of the output  $x_{\text{para}}$  from the paraphrase model and the perplexity observed when using  $x_{\text{para}}$  as the input prompt for the target model. As shown in Fig. 1, there is a lack of significant correlation between the perplexity of  $x_{\text{para}}$  as output in  $P_{\text{para}}$  and that of  $x_{\text{para}}$  as input in  $P_{\text{tar}}$ , even if we use the same model as the paraphrase and target model ( $P_{\text{para}} = P_{\text{tar}}$ ). This observation indicates that the minimal perplexity of a paraphrased prompt could not be achieved by the paraphrase model alone.
+
+Based on the above fact, the goal of MONOPARA is to take on the semantic hints from the paraphrase
+
+![](images/910d246fcd8e5c6f84201c8a8b1d74692da0e228163ba61c342542c7ee11a01e.jpg)  
+Figure 2: Two explored decoding schemes of MONOPARA. Ensemble-based decoding (bottom) combines the token probabilities from the paraphrase model and the target model in each decoding step. Search-based decoding (top) further leverages look-ahead decoding to consider the potential future impact of current choices.
+
+model while following the constraint of perplexity from the target model. Specifically, the paraphrase model generates paraphrases for a given prompt, while the target model constrains the paraphrase perplexity during the generation process. With an ensemble of these two models, we can leverage the strengths of both: the guided content generation capability of the paraphrase model and the low-perplexity prompt constraint of the target model. MONOPARA is thus able to produce paraphrases with lower perplexity by synergizing the benefits of both models.
+
+# 2.2 Prompt Paraphrasing
+
+We first introduce prompt paraphrasing without perplexity constraint. We consider using an LLM for prompt paraphrasing in a zero-shot manner, where the model is instructed to paraphrase the original prompt  $x_{ori}$  to a fluent and coherent counterpart  $x_{para}$  with a system prompt  $x_{sys}$ , such as the following one:
+
+# The System Prompt for Paraphrasing
+
+Generate ONE paraphrase of the following sentence.
+
+Specifically, we denote the input of the paraphrase model, including the concatenation of the system prompt  $x_{sys}$  and the original prompt  $x_{ori}$  that the model is instructed to rewrite, as  $[x_{sys}, x_{ori}] = (x_1, \ldots, x_n)$ , where  $x_i$  is a token in the vocabulary  $V$ . Suppose the decoder of a pre-trained autoregressive language model  $P_{para}$  generates continuations of length  $m$  as the requested paraphrase, denoted as
+
+$x_{para} = (x_{n + 1},\ldots ,x_{n + m})$  , based on the system prompt  $x_{sys}$  and original prompt  $x_{ori}$  . At decoding time,  $P_{para}$  iteratively decodes one token at a time by conditioning on the preceding context:
+
+$$
+P _ {p a r a} (x _ {p a r a} | x _ {s y s}, x _ {o r i}) = \prod_ {i = n + 1} ^ {n + m} P _ {p a r a} (x _ {i} | x _ {<   i}),
+$$
+
+where  $P_{\text{para}}(x_i | x_{<i})$  is the predicted next token probability.
+
+# 2.3 Ensemble-based (Greedy) MONOPARA
+
+To introduce perplexity constraint from the target model to the paraphrase model, here we introduce the ensemble-based greedy decoding for monotonic paraphrasing (as shown in the lower half of Fig. 2). Based on the intuition delivered in §2.1, MONOPARA decodes the paraphrase iteratively, and the next token is selected based on the combination of predicted probabilities from the two LMs:
+
+$$
+\begin{array}{l} x _ {n e x t} = \underset {x _ {i} \in V} {\operatorname {a r g m a x}} (\alpha \cdot \log P _ {t a r} (x _ {i} | x _ {g e n}) \\ + (1 - \alpha) \cdot \log P _ {p a r a} (x _ {i} | x _ {s y s}, x _ {o r i}, x _ {g e n})), \\ \end{array}
+$$
+
+where  $x_{gen}$  denotes the sequence of generated tokens, and  $\alpha$  determines the coefficient between the two models (Alg. 1).
+
+By combining the prediction probabilities of both models, the ensemble can navigate the complex landscape of language generation more effectively and precisely. The paraphrase model  $P_{para}$
+
+# Algorithm 1 Ensemble-based Decoding
+
+# Require:
+
+$x_{sys}$ $\triangleright$  System prompt for paraphrase  $V$ $\triangleright$  Vocabulary  $P_{tar}, P_{para}$ $\triangleright$  Target and paraphrase models  $\alpha$ $\triangleright$  Weighting factor
+
+# Ensure:
+
+$x_{gen}$ $\triangleright$  Generated text sequence
+
+1: Initialize  $x_{gen} \gets \emptyset$
+
+2:  $x_{next} \gets \underset{x_i \in V}{\text{argmax}} P_{para}(x_i | x_{sys} \oplus x_{ori})$  Generate first token with  $P_{para}$
+
+3:  $x_{gen} \gets x_{gen} \oplus x_{next}$
+
+4: while not end of sequence do
+
+5: Compute weighted log probabilities for each  $x_{i} \in V$ :
+
+6:  $score_{tar} \gets \alpha \cdot \log P_{tar}(x_i | x_{gen})$
+
+7:  $score_{para} \gets (1 - \alpha) \cdot \log P_{para}(x_i | x_{sys} \oplus x_{ori} \oplus x_{gen})$
+
+8:  $x_{next} \gets \underset{x_i \in V}{\text{argmax}} \left( \text{score}_{tar} + \text{score}_{para} \right)$
+
+9:  $x_{gen}\gets x_{gen}\oplus x_{next}$
+
+0: end while
+
+11: return  $x_{gen}$
+
+keeps the rewritten prompt (or instruction) remains with the original intent, while the target model  $P_{tar}$  evaluates the paraphrase's likelihood or confidence, up-weighting any generation samples that lead towards low-perplexity outcomes. Based on the ensemble, the next token is selected by maximizing the weighted sum of logarithmic probabilities from both models, considering both the instruction and the generated text thus far.
+
+As a coefficient,  $\alpha$  balances the contribution of each model, ensuring that the final output satisfies both semantic fidelity to the prompt intent and linguistic familiarity to the target model. Higher  $\alpha$  increases the weight of the logarithmic probability derived from the target model  $P_{tar}$  and cuts down on the semantic hints for  $P_{para}$ , which raises the risk of sacrificing the semantic fidelity of the generation. Conversely, a lower  $\alpha$  shifts the emphasis towards  $P_{para}$ , placing greater value on ensuring that the generated text closely follows the instructional context and semantic intent from  $P_{para}$ . In this case, the decoding process lacks the constraint of perplexity from  $P_{tar}$  and sacrifices the linguistic familiarity of the generated paraphrase for the target model.
+
+# 2.4 Search-based MONOPARA
+
+To further enhance the efficiency of decoding prompts of lower perplexity, we define the search-based decoding strategy for MONOPARA (as shown in the upper half of Fig. 2). To take more steps of token generation into account and expand the search space, we leverage look-ahead decoding to search for the sequence that exhibits the lowest perplexity from an even broader candidate space. Following Lu et al. (2022a), each step of our search-based decoding consists of (i) expanding a set of candidate next-tokens, (ii) scoring each candidate, and (iii) selecting the  $k$  best candidates (Alg. 2 in Appx.  $\S A$ ):
+
+$$
+X _ {m} ^ {\prime} = \left\{\mathrm {x} _ {<   m} \circ x _ {m} | \mathrm {x} _ {<   m} \in X _ {m - 1}, x _ {m} \in V \right\},
+$$
+
+$$
+X _ {m} = \underset {(x <   m, x _ {m}) \in X _ {m} ^ {\prime}} {\operatorname {a r g t o p k}} \{f (x _ {<   m}, x _ {m}) \},
+$$
+
+where  $x_{m}$  is a candidate predicted by paraphrase model, and  $f(\cdot)$  is a scoring function that returns the perplexity of  $\mathbf{x}_{< m} \circ x_{m}$  evaluated by the target model:
+
+$$
+f \left(\mathrm {x} _ {<   m}, x _ {m}\right) = \left(\prod_ {i = 1} ^ {m} \frac {1}{P _ {t a r} \left(x _ {i} \mid x _ {<   i}\right)}\right) ^ {\frac {1}{m}}. \tag {1}
+$$
+
+Specifically, at each step  $m$  of the decoding process, the current set of sequences  $X_{m - 1}$  is expanded by appending the most probable next-token  $x_{m}$ , predicted by the paraphrase model  $P_{para}$ , from vocabulary  $V$  to each sequence  $\mathbf{x}_{< m}$ . This results in a temporary expanded set of candidate sequences  $X_{m}^{\prime}$ . Each candidate sequence in  $X_{m}^{\prime}$  is then evaluated by the scoring function  $f(\cdot)$  shown in Eq. 1, which calculates the perplexity of a sequence by the target model  $P_{tar}$ . Since a paraphrase of lower perplexity is preferred, the  $k$  sequences with the best scores (i.e., lowest perplexity) are selected to form the set  $X_{m}$ , narrowing down the choices to the  $k$  most promising sequences for continuation.
+
+In comparison to greedy decoding, this looks ahead decoding mechanism evaluates a broader range of potential future impact from current choices (i.e., the selection of  $x_{m}$ ), thereby allowing for a more informed and potentially more accurate selection of candidates.
+
+# 3 Experiment
+
+In this section, we demonstrate two distinct experimental settings for evaluation. In §3.1, we examine MONOPARA's effectiveness on prompt refinement
+
+for eliciting better task performance. In §3.2, we evaluate MONOPARA's effect on enhancing model robustness and generalization under instruction perturbations.
+
+# 3.1 Task I: Prompt Refinement
+
+Task Description To inspect if monotonic paraphrasing can refine a prompt and elicit better performance on prompting an LLM as designed, we use SUPER-NATURALINSTRUCTION (SUP-NATINST for short, Wang et al., 2022). This benchmark consists of 1,616 diverse NLP tasks and their expert-written prompts for evaluating the zero-shot generalizability of LLMs on a variety of NLP tasks. In SUP-NATINST, each task is paired up with an instruction that consists of the task definition for mapping an input text to a task output and several examples for demonstrating the desired or undesired output. To make our evaluation more challenging, we only use the task definition as a prompt for the target model to perform the task. During test time, the target model is prompted with a concatenation of SUP-NATINST task description (or its paraphrase), test sample, and multiple choices of answers. Since SUP-NATINST provides only one description of definition for each task, we use GPT4 (Achiam et al., 2023) to generate 4 more SUP-NATINST-like task descriptions for each involved task, which results in 5 prompts for each task in total.
+
+Datasets Following Gonen et al. (2023), we choose 7 classification tasks from Huggingface Datasets  $^{3}$  to have a set of diverse tasks: (i) GLUE Cola (Warstadt et al., 2019) for grammatical acceptability discrimination; (ii) Newspop (Moniz and Torgo, 2018) for news classification; (iii) AG News (Zhang et al., 2015) for news classification; (iv) IMDB (Maas et al., 2011) for movie review sentiment analysis; (v) DBpedia (Lehmann et al., 2015) for topic classification; (vi) Emotion (Saravia et al., 2018) for tweet emotion classification; and (vii) Tweet Offensive (Barbieri et al., 2020) for offensive tweet discrimination. We sample 1,000 examples from each dataset for prompt evaluation. For these tasks, the prompt follows the input, and at the end of each prompt, we add the choices of classes following Gonen et al. (2023): "Choices: X, Y, Z. Answer:" as it helps with accuracy.
+
+Evaluation Metrics To inspect whether monotonic paraphrasing can refine a prompt into its low-perplexity counterpart and further enhance the generalization of LLM prompting, we use PPL (perplexity), Acc (accuracy), and BS (BERTScore, Zhang* et al., 2020) for evaluation. (1) Perplexity. Following Gonen et al. (2023), we define the perplexity of the prompt as the perplexity of the full prompt sequence, including the input itself and the choices of labels. To avoid noise when computing perplexity, the prompts are instantiated with 1,000 random examples of the dataset (we keep the same selected examples for performance evaluation), and the perplexity is averaged across all instantiated prompts. The correlation between the perplexity of a standalone prompt and the average perplexity over instantiated ones is illustrated in Appx. §E. (2) BERTScore. As a paraphrasing scheme, MONOPARA is also evaluated by its performance in semantic alignment with the source sentence based on pre-trained BERT (Devlin et al., 2019). (3) Accuracy. To compute accuracy, for each test sample, we obtain the model's predicted probability of all classes and choose the highest ranking class as the prediction of the model. The accuracy is calculated over the 1,000 samples for each task.
+
+Models We study two auto-regressive models, Mistral-7B $^4$  (Jiang et al., 2023) and Starling-7B $^5$  (Zhu et al., 2023a), for their significant ability in instruction following and paraphrasing. In this paper, we keep the paraphrase model the same as the target model. Please refer to Appx. §B for the results on Llama-3-8B $^6$  (AI@Meta, 2024).
+
+Baselines We compare our method with two baselines: SPELL (Gonen et al., 2023) and vanilla LLM-based paraphrasing (Para). Based on similar intuition that low perplexity and better performance exhibit strong correlation, SPELL (Selecting Prompts by Estimating LM Likelihood) ranks and selects the prompts with the lowest perplexity for a given task after creating a set of prompt candidates manually and expanding them to hundred-scale using automatic paraphrasing and back-translation. We use the prompt candidates
+
+<table><tr><td>Methods</td><td>Metric</td><td>GLUE Cola</td><td>Newspop</td><td>AG News</td><td>IMDB</td><td>DBpedia</td><td>Emotion</td><td>Tweet Offensive</td></tr><tr><td colspan="9">Mistral 7B</td></tr><tr><td rowspan="2">Ori.</td><td>Acc</td><td>61.76 ± 2.93</td><td>85.28 ± 3.45</td><td>67.68 ± 3.47</td><td>84.12 ± 4.89</td><td>70.82 ± 2.27</td><td>44.32 ± 1.57</td><td>67.04 ± 0.95</td></tr><tr><td>PPL</td><td>15.72 ± 4.60</td><td>12.60 ± 2.40</td><td>12.74 ± 1.07</td><td>16.41 ± 0.80</td><td>8.10 ± 0.22</td><td>16.48 ± 2.14</td><td>19.25 ± 3.43</td></tr><tr><td rowspan="2">SPELL</td><td>Acc</td><td>43.42 ± 1.16</td><td>81.08 ± 7.2</td><td>73.88 ± 3.18</td><td>86.00 ± 0.81</td><td>71.02 ± 1.70</td><td>46.80 ± 1.12</td><td>65.06 ± 1.04</td></tr><tr><td>PPL</td><td>8.24 ± 0.88</td><td>15.05 ± 4.15</td><td>15.99 ± 3.77</td><td>15.76 ± 1.42</td><td>25.36 ± 6.80</td><td>15.67 ± 5.21</td><td>14.31 ± 0.93</td></tr><tr><td rowspan="3">Para</td><td>Acc</td><td>39.48 ± 8.19</td><td>76.72 ± 16.54</td><td>71.60 ± 2.70</td><td>86.60 ± 4.86</td><td>70.22 ± 5.71</td><td>43.65 ± 3.41</td><td>65.04 ± 1.46</td></tr><tr><td>PPL</td><td>78.01 ± 11.22</td><td>49.37 ± 7.81</td><td>28.37 ± 2.58</td><td>22.27 ± 0.98</td><td>28.34 ± 1.36</td><td>46.53 ± 5.30</td><td>53.65 ± 1.92</td></tr><tr><td>BS</td><td>88.88 ± 0.74</td><td>87.47 ± 0.84</td><td>86.32 ± 2.08</td><td>90.23 ± 0.99</td><td>86.96 ± 0.93</td><td>90.08 ± 0.35</td><td>89.93 ± 0.67</td></tr><tr><td rowspan="3">Mono-E</td><td>Acc</td><td>63.96 ± 3.88</td><td>88.96 ± 3.04</td><td>72.20 ± 3.87</td><td>87.58 ± 3.27</td><td>72.07 ± 3.54</td><td>47.76 ± 3.72</td><td>67.00 ± 1.21</td></tr><tr><td>PPL</td><td>14.89 ± 2.75</td><td>8.33 ± 1.30</td><td>10.07 ± 2.93</td><td>13.05 ± 1.86</td><td>8.85 ± 0.88</td><td>13.67 ± 1.87</td><td>14.11 ± 3.53</td></tr><tr><td>BS</td><td>93.03 ± 2.15</td><td>92.85 ± 2.35</td><td>88.70 ± 2.35</td><td>89.35 ± 2.60</td><td>93.21 ± 1.16</td><td>92.74 ± 2.75</td><td>91.73 ± 1.32</td></tr><tr><td rowspan="3">Mono-S</td><td>Acc</td><td>62.31 ± 6.17</td><td>86.34 ± 5.16</td><td>70.80 ± 3.44</td><td>83.48 ± 2.87</td><td>72.46 ± 3.00</td><td>44.26 ± 3.05</td><td>65.66 ± 2.43</td></tr><tr><td>PPL</td><td>15.85 ± 4.88</td><td>11.95 ± 2.09</td><td>9.89 ± 1.31</td><td>16.38 ± 1.63</td><td>8.63 ± 0.41</td><td>15.03 ± 1.65</td><td>18.35 ± 4.07</td></tr><tr><td>BS</td><td>92.47 ± 2.00</td><td>93.78 ± 2.68</td><td>91.57 ± 2.23</td><td>89.66 ± 3.25</td><td>95.87 ± 1.20</td><td>93.42 ± 1.02</td><td>91.62 ± 2.48</td></tr><tr><td colspan="9">Starling 7B</td></tr><tr><td rowspan="2">Ori.</td><td>Acc</td><td>55.18 ± 12.43</td><td>93.78 ± 1.18</td><td>88.44 ± 1.18</td><td>96.46 ± 0.17</td><td>94.06 ± 0.37</td><td>55.48 ± 0.91</td><td>64.62 ± 0.18</td></tr><tr><td>PPL</td><td>16.38 ± 4.18</td><td>12.77 ± 2.59</td><td>12.10 ± 0.91</td><td>15.56 ± 0.57</td><td>8.11 ± 0.22</td><td>17.90 ± 2.82</td><td>20.41 ± 3.90</td></tr><tr><td rowspan="2">SPELL</td><td>Acc</td><td>46.46 ± 6.70</td><td>89.84 ± 0.91</td><td>83.72 ± 0.26</td><td>92.84 ± 2.7</td><td>92.02 ± 0.64</td><td>54.82 ± 1.81</td><td>64.20 ± 0.17</td></tr><tr><td>PPL</td><td>44.96 ± 18.23</td><td>54.60 ± 18.72</td><td>16.81 ± 0.91</td><td>56.34 ± 15.67</td><td>28.66 ± 9.47</td><td>32.79 ± 8.53</td><td>33.56 ± 3.83</td></tr><tr><td rowspan="3">Para</td><td>Acc</td><td>40.38 ± 15.49</td><td>82.67 ± 19.63</td><td>86.39 ± 2.92</td><td>96.66 ± 0.34</td><td>93.10 ± 1.79</td><td>51.10 ± 10.18</td><td>64.44 ± 0.08</td></tr><tr><td>PPL</td><td>70.50 ± 16.38</td><td>54.41 ± 8.80</td><td>30.67 ± 2.81</td><td>21.78 ± 1.37</td><td>26.56 ± 1.39</td><td>58.87 ± 4.44</td><td>11.71 ± 0.86</td></tr><tr><td>BS</td><td>90.35 ± 0.67</td><td>88.35 ± 0.51</td><td>88.46 ± 0.98</td><td>90.04 ± 1.04</td><td>87.06 ± 0.79</td><td>90.38 ± 0.45</td><td>79.93 ± 11.52</td></tr><tr><td rowspan="3">Mono-E</td><td>Acc</td><td>69.62 ± 4.12</td><td>95.15 ± 0.59</td><td>85.40 ± 4.22</td><td>96.22 ± 0.47</td><td>94.05 ± 0.42</td><td>56.10 ± 1.08</td><td>64.45 ± 0.09</td></tr><tr><td>PPL</td><td>8.56 ± 1.00</td><td>10.12 ± 1.02</td><td>8.52 ± 0.76</td><td>12.60 ± 1.57</td><td>9.58 ± 0.86</td><td>9.23 ± 0.97</td><td>11.71 ± 0.86</td></tr><tr><td>BS</td><td>93.00 ± 0.85</td><td>94.19 ± 0.94</td><td>91.06 ± 1.51</td><td>92.85 ± 1.39</td><td>93.23 ± 0.50</td><td>93.28 ± 0.26</td><td>93.69 ± 1.52</td></tr><tr><td rowspan="3">Mono-S</td><td>Acc</td><td>50.96 ± 11.52</td><td>88.20 ± 3.74</td><td>83.61 ± 6.30</td><td>92.96 ± 4.78</td><td>92.82 ± 2.17</td><td>54.90 ± 2.32</td><td>66.80 ± 1.22</td></tr><tr><td>PPL</td><td>14.05 ± 3.76</td><td>12.55 ± 2.07</td><td>10.61 ± 1.06</td><td>16.44 ± 0.92</td><td>8.39 ± 0.53</td><td>16.28 ± 2.18</td><td>20.17 ± 3.17</td></tr><tr><td>BS</td><td>92.69 ± 0.99</td><td>95.83 ± 1.66</td><td>93.92 ± 2.23</td><td>93.01 ± 1.25</td><td>96.99 ± 0.70</td><td>94.78 ± 1.25</td><td>94.45 ± 1.43</td></tr></table>
+
+Table 1: Results of Task I (Prompt Refinement). Ori. refers to the original prompts. Para refers to paraphrasing the original prompts without perplexity constraint. Mono-E refers to ensemble-based MONOPARA. Mono-S refers to search-based MONOPARA. The best accuracy for each task is in bold, while the lowest perplexity and the highest BERTScore for each task is in blue and orange respectively. The coefficient  $\alpha$  is fixed as 0.5 for all the results.
+
+provided by Gonen et al. (2023) and rank them by the perplexity w.r.t. the target model. Para paraphrases source prompts with greedy decoding (top-1 sampling) by directly prompting the target LLM with the instruction introduced in §2.2. Besides, we also implement reranking based on the prompt candidates generated by MONOPARA for a fair comparison with SPELL. The implementation details and results are discussed in Appx. §D.
+
+Results As shown in Tab. 1, the two variants of MONOPARA successfully refine prompts to those with perplexities lower than the original ones, thereby enhancing model performance on 10 out of 14 scenarios (2 models, each on 7 datasets). Specifically, Mono-E (ensembled-based MONOPARA) improves the accuracy of Mistral-7B by an average of  $2.64\%$  and that of Starling-7B by an average of  $1.85\%$  on seven different datasets. In contrast,
+
+directly paraphrasing the prompts without perplexity constraint (Para) results in prompts with higher perplexities than the original ones in 13 out of 14 scenarios, most of which lead to worse task performance. This highlights the importance of incorporating perplexity constraints when paraphrasing prompts. The correlation between task accuracy and prompt perplexity is analyzed in Appx. §C.
+
+Besides, our method outperforms SPELL with consistently lower perplexity and superior task performance with the refined prompt. This superiority stems from the fact that, although SPELL explores a vast search space with hundred-scale prompt candidates, the candidates are searched in an ad hoc manner without an optimization goal, leading to inefficiencies in prompt rewriting or paraphrasing. In contrast, our method incorporates the perplexity constraint that leads to optimized target-model perplexity during one single refined prompt generation
+
+process, leading to both enhanced task performance and refinement efficiency.
+
+Moreover, MONOPARA achieves superior performance in terms of both average prompt perplexity and BERTScore, indicating the synergistic role played by both the target model and the paraphrase model in prompt refinement. During decoding, the target model prioritizes achieving low perplexity, while the paraphrase model preserves the original prompt intent, resulting in a high BERTScore. We provide a detailed case study on BERTScore of Para and MONOPARA in Appx. §G to explain why MONOPARA outperforms Para on BERTScore. Specifically, Mono-E outperforms Mono-S in some cases since Mono-S involves the additional hyperparameter of beam size which we did not tune. Additionally, having additional token options at each decoding step may introduce more randomness, potentially affecting the overall performance. That being said, Mono-S mostly exhibits a higher BERTScore than the simpler method Mono-E, indicating better quality of paraphrasing.
+
+# 3.2 Task II: Robustness Evaluation on Instruction Perturbation
+
+We further evaluate MONOPARA's effectiveness in enhancing model robustness against various instruction perturbations at character, word, sentence, and semantic levels using PromptBench (Zhu et al., 2023b).
+
+Task Description PromptBench introduces perturbation to task instructions of a diverse set of tasks, such as sentiment analysis, grammar correctness, duplicate sentence detection, and natural language inference. It includes character-level perturbations using DeepWordBug (Gao et al., 2018) to introduce typos, word-level using TextFooler (Jin et al., 2020) to replace words with contextually similar words, and semantic-level by using prompts following the linguistic behavior of different languages. For semantic-level perturbations, the adversary constructs prompts using various languages, such as Chinese, French, Arabic, Spanish, Japanese, and Korean, and then translates these prompts into English. By exploiting the nuances and idiosyncrasies of different languages during translation, it can introduce subtle ambiguities, grammatical errors, or inconsistencies in the input prompt. The perturbed prompt itself is still in English.
+
+![](images/ed0e4717cf2f2ec32fbff1ad504024f9e4b873d1a13350f0526afe733692f5e6.jpg)  
+Figure 3: Model's average accuracy across 4 GLUE datasets, with each dataset having six instructions with perturbation added at character, word, and semantic levels. Mono-E has consistent improvement in accuracy across all types of perturbation compared to vanilla paraphrasing.
+
+Datasets We select 4 GLUE tasks (Wang et al., 2018) from Huggingface Datasets for evaluation: (i) Cola, which is also used for Task I; (ii) SST-2 (Socher et al., 2013) for two-way sentiment classification; (iii) MNLI (Wang et al., 2018) for multi-genre inference identification; and (iv) MRPC (Dolan and Brockett, 2005) for paraphrase identification. We adopt the same instruction format as that of Task I.
+
+Model For this task, we use Alpaca $^{7}$  (Taori et al., 2023), a model instruction-tuned from the LLaMA model (Touvron et al., 2023a) on the 52k self-instruct dataset. The paraphrase model is also the same as the target model.
+
+Results As shown in Fig. 3, we compare the performance of original prompts, prompts paraphrased without perplexity constraint, and our method Mono-E. Across all kinds of instruction perturbations, our method achieves consistent accuracy improvement by up to  $10.13\%$ . This notable increase in performance demonstrates that MONOPARA not only refines prompts for improved task performance but also enhances model robustness against various instruction perturbations. We also show the model performance with unperturbed prompts and compare the degradation from perturbation in Appx.  $\S F$  to show the robustness of Mono-E.
+
+<table><tr><td colspan="2"></td><td>Content</td><td>BertScore</td><td>PPL</td></tr><tr><td colspan="2">Ori Instruction</td><td>Read the provided excerpt and choose between ‘positive’ and ‘negative’ to describe its sentiment Iyk11Jt8yw:</td><td>-</td><td>78.12</td></tr><tr><td colspan="2">Para</td><td>Determine the sentiment of the given excerpt as either positive or negative</td><td>81.26</td><td>44.29</td></tr><tr><td rowspan="3">Mono</td><td>α = 0.2</td><td>Determine the sentiment of the given text by selecting either ‘positive’ or ‘negative’.</td><td>86.18</td><td>24.90</td></tr><tr><td>α = 0.5</td><td>Determine the sentiment of the given text by classifying it as positive or negative.</td><td>80.22</td><td>16.27</td></tr><tr><td>α = 0.7</td><td>Determine the remainder when 101 is divided by 10.\nA: 1</td><td>50.22</td><td>4.53</td></tr></table>
+
+Table 2: Effect of coefficient  $\alpha$  on Mono-E with Mistral-7B model.
+
+# 3.3 Effect of Coefficient
+
+We further analyze the effect of coefficient  $\alpha$  between the target model and the paraphrase model in MONOPARA. Tab. 2 presents a case study of  $\alpha$  on the ensemble-based decoding variant (MonoE) with a prompt under character-level perturbation. Obviously, a larger value of  $\alpha$  amplifies the influence of the target model, resulting in a stronger perplexity constraint and monotonically lower perplexity. Consequently, the impact of the paraphrase model is diminished, leading to a lower BERTScore. By adjusting  $\alpha$ , the user can find a balanced point to achieve better prompts that strike a balance between semantic fidelity to the original prompt intent and linguistic familiarity to the target model. Detailed results and analysis are presented in Appx. §H and Tab. 9.
+
+# 4 Related Work
+
+Paraphrasing. Paraphrase generation has been a longstanding task in NLP. Since the era of deep learning, studies have adopted the universal paradigm of training sequence-to-sequence generation models based on abundant learning resources (Nema et al., 2017; Gupta et al., 2018; Li et al., 2018). While more recent works have leveraged more robust pre-trained language models that captured rich supervision signals from various sequence-to-sequence generation tasks (Raffel et al., 2020; Lewis et al., 2020) and achieving even more robust performance of paraphrasing with relatively limited end-task supervision (Sun et al., 2021; Meng et al., 2021; Bui et al., 2021). In addition to these advancements, a series of works have added auxiliary supervision signals such as syntactical guidance, lexical regularization, and cross-lingual supervision, to further improve the performance and controllability of paraphrasing in aspects such as syntactic structures, stylistic specification, or textual simplicity (Iyyer et al., 2018;
+
+Li et al., 2019; Chen et al., 2019; Kumar et al., 2020; Goyal and Durrett, 2020; Hosking and Lapata, 2021; Chowdhury et al., 2022).
+
+Different from existing paraphrasing techniques, this work conducts constrained decoding methods that seek to monotonically decrease the perplexity in the process of paraphrasing. This innovative approach serves as a general approach to rewrite and refine prompts or task instructions towards their more familiar counterparts for LMs, seeking to enhance the generalization of LM prompting without the need for any training effort.
+
+Prompt Refinement. Prompt refinement has been actively explored in recent years with the aim of selecting, retrieving, or even generating prompts that lead to improved zero-shot performance of LM prompting. When an end-task objective is known, prior works on prompt refinement approaches often set the end-task performance as the objective, and leverage approaches such as gradient-based search, similar to continuous prompts but with projections onto a discrete vocabulary (Shin et al., 2020). Other works have explored edit-based enumeration (Prasad et al., 2023), reinforcement learning (Deng et al., 2022), and large language model continuation and filtering (Zhou et al., 2022). As for general-purpose LLMs, a dedicated prompt refinement approach may not assume the pre-existence of any specific end tasks. Hence, more recent works Iter et al. (2023); Gonen et al. (2023); Wang et al. (2023); Lee et al. (2023) explored the heuristic criterion based on the familiarity of the language model with the language contained in the prompt, as measured by its perplexity. Lower perplexity prompts are preferred as they are expected to perform better across a wide range of tasks.
+
+While the last line of research has relieved the generic principle for improving zero-shot LLM prompting generalizability, these studies are largely limited to selecting among pre-existing prompts
+
+or multiple beam search samples. Our proposed monotonic paraphrasing approach, however, can proactively rewrite a prompt or task instruction into counterparts that better satisfy the above heuristic criterion, without the need of any pre-existing candidate prompts or any training effort.
+
+Ensemble and Search-based Decoding. The proposed two decoding schemes of MONOPARA are connected to both ensemble decoding and search-based decoding approaches in recent studies. Ensemble decoding has been previously proposed for purposes including debiasing (Li et al., 2023), controllable generation (Meng et al., 2022; Huang et al., 2023), and cross-modality data processing (Liu et al., 2023). Search-based decoding, on the other hand, is proposed to efficiently extend the search space of generation samples for satisfying specific constraints or criteria (Lu et al., 2022a; Wuebker et al., 2012; Lee and Berg-Kirkpatrick, 2022; Xu et al., 2023). Representative approaches include look-ahead (Lu et al., 2022a) and look-back (Xu et al., 2023) ones that search within the beam space towards different directions.
+
+The explored decoding schemes in MONOPARA are inspired by both lines of studies on decoding approaches. Specifically, the first line is related to our design choice for composing the probabilities of both the paraphrasing and the target LMs in greedy generation. And our search-based decoding scheme is specifically inspired by the look-ahead decoding strategy in the second line of studies.
+
+# 5 Conclusion
+
+In this paper, we propose monotonic paraphrasing (MONOPARA) for automatically and iteratively generating low-perplexity prompts for given LLMs, which can lead to better task performance. Following the high-level idea of paraphrasing prompts with perplexity constraints from the target LLM, we design ensemble-based and search-based decoding strategies for efficient prompt refinement. Experiments on prompt variations of unseen tasks and instructions demonstrate the effectiveness of MONOPARA in reducing prompt perplexity while enhancing task performance and model robustness. Future work may apply MONOPARA on other scenarios, such as searching stealthy prompts for red-teaming. Replacing perplexity with other criteria, such as complexity (Li et al., 2024), for different purposes of constrained generation is also a meaningful research direction.
+
+# Acknowledgement
+
+We thank the anonymous reviewers for their valuable comments. Qin Liu was supported by a departmental fellowship. Fei Wang was supported by the Amazon ML Fellowship. Tianyi Yan was supported by the CURVE Fellowship. Muhao Chen was supported by the DARPA FoundSci Grant HR00112490370, the NSF of the United States Grant ITE 2333736, and an Amazon Research Award.
+
+# Limitations
+
+The current investigation of MONOPARA has the following limitations. First, using the target model as the paraphrase model is a natural technical choice. However, examining other paraphrasing models, especially those customized for paraphrasing, may lead to better performance. Second, our proposed method can not be applied to black-box LLMs such as ChatGPT $^{8}$ , PaLM (Chowdhery et al., 2023), Claude $^{9}$ , etc. This limitation arises from the necessity of access to the decoding phase, which in turn comes with far less computational cost due to zero training, and does not require the pre-existence of multiple prompt or instruction candidates. Third, while we conduct experiments on a wide range of tasks, their original prompts and responses are relatively short. This actually limits the operational space of MONOPARA. Experiments on tasks with longer inputs and outputs may provide additional evidence of MONOPARA's effectiveness.
+
+# References
+
+Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774.
+
+AI@Meta. 2024. Llama 3 model card.
+
+Francesco Barbieri, Jose Camacho-Collados, Luis Espinosa Anke, and Leonardo Neves. 2020. TweetEval: Unified benchmark and comparative evaluation for tweet classification. In *Findings of the Association for Computational Linguistics: EMNLP* 2020, pages 1644–1650, Online. Association for Computational Linguistics.
+
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind
+
+Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Tien-Cuong Bui, Van-Duc Le, Hai-Thien To, and Sang Kyun Cha. 2021. Generative pre-training for paraphrase generation by representing and predicting spans in exemplars. In 2021 IEEE International Conference on Big Data and Smart Computing (BigComp), pages 83-90. IEEE.  
+Mingda Chen, Qingming Tang, Sam Wiseman, and Kevin Gimpel. 2019. Controllable paraphrase generation with a syntactic exemplar. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5972-5984, Florence, Italy. Association for Computational Linguistics.  
+Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. 2023. Vicuna: An open-source chatbot impressing gpt-4 with  $90\%$  * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023).  
+Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research, 24(240):1-113.  
+Jishnu Ray Chowdhury, Yong Zhuang, and Shuyi Wang. 2022. Novelty controlled paraphrase generation with retrieval augmented conditional prompt tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 10535-10544.  
+Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric Xing, and Zhiting Hu. 2022. RLPrompt: Optimizing discrete text prompts with reinforcement learning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3369-3391, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Association for Computational Linguistics.  
+Ning Ding, Shengding Hu, Weilin Zhao, Yulin Chen, Zhiyuan Liu, Haitao Zheng, and Maosong Sun. 2022. OpenPrompt: An open-source framework for prompt-learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 105-113, Dublin, Ireland. Association for Computational Linguistics.
+
+Bill Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Third International Workshop on Paraphrasing (IWP2005).  
+Ji Gao, Jack Lanchantin, Mary Lou Soffa, and Yanjun Qi. 2018. Black-box Generation of Adversarial Text Sequences to Evade Deep Learning Classifiers. ArXiv:1801.04354 [cs].  
+Tianyu Gao, Adam Fisch, and Danqi Chen. 2021. Making pre-trained language models better few-shot learners. 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 3816-3830, Online. Association for Computational Linguistics.  
+Hila Gonen, Srini Iyer, Terra Blevins, Noah Smith, and Luke Zettlemoyer. 2023. Demystifying prompts in language models via perplexity estimation. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 10136–10148, Singapore. Association for Computational Linguistics.  
+Tanya Goyal and Greg Durrett. 2020. Neural syntactic preordering for controlled paraphrase generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 238-252, Online. Association for Computational Linguistics.  
+Ankush Gupta, Arvind Agarwal, Prawaan Singh, and Piyush Rai. 2018. A deep generative framework for paraphrase generation. In Proceedings of the aaai conference on artificial intelligence, volume 32.  
+Tom Hosking and Mirella Lapata. 2021. Factorising meaning and form for intent-preserving paraphrasing. 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 1405-1418, Online. Association for Computational Linguistics.  
+Tenghao Huang, Ehsan Qasemi, Bangzheng Li, He Wang, Faeze Brahman, Muhao Chen, and Snigdha Chaturvedi. 2023. Affective and dynamic beam search for story generation. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 11792-11806, Singapore. Association for Computational Linguistics.  
+Dan Iter, Reid Pryzant, Ruochen Xu, Shuohang Wang, Yang Liu, Yichong Xu, and Chenguang Zhu. 2023. In-context demonstration selection with cross entropy difference. arXiv preprint arXiv:2305.14726.  
+Mohit Iyyer, John Wieting, Kevin Gimpel, and Luke Zettlemoyer. 2018. Adversarial example generation with syntactically controlled paraphrase networks. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,
+
+Volume 1 (Long Papers), pages 1875-1885, New Orleans, Louisiana. Association for Computational Linguistics.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Zhengbao Jiang, Frank F. Xu, Jun Araki, and Graham Neubig. 2020. How can we know what language models know? Transactions of the Association for Computational Linguistics, 8:423-438.  
+Di Jin, Zhijing Jin, Joey Tianyi Zhou, and Peter Szolovits. 2020. Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and Entailment. ArXiv:1907.11932 [cs].  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.  
+Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi-Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richard Nagyfi, et al. 2023. Openassistant conversations-democratizing large language model alignment. arXiv preprint arXiv:2304.07327.  
+Ashutosh Kumar, Kabir Ahuja, Raghuram Vadapalli, and Partha Talukdar. 2020. Syntax-guided controlled generation of paraphrases. Transactions of the Association for Computational Linguistics, 8:329-345.  
+Aswanth Kumar, Ratish Puduppully, Raj Dabre, and Anoop Kunchukuttan. 2023. CTQScorer: Combining multiple features for in-context example selection for machine translation. In *Findings of the Association for Computational Linguistics: EMNLP* 2023, pages 7736-7752, Singapore. Association for Computational Linguistics.  
+Ivan Lee and Taylor Berg-Kirkpatrick. 2022. HeLo: Learning-free lookahead decoding for conversation infilling. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4996-5008, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Yoonsang Lee, Pranav Atreya, Xi Ye, and Eunsol Choi. 2023. Crafting in-context examples according to lms' parametric knowledge. arXiv preprint arXiv:2311.09579.  
+Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, Soren Auer, et al. 2015. Dbpedia-a large-scale, multilingual knowledge base extracted from wikipedia. Semantic web, 6(2):167-195.
+
+Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online. Association for Computational Linguistics.  
+Bangzheng Li, Ben Zhou, Xingyu Fu, Fei Wang, Dan Roth, and Muhao Chen. 2024. Famicom: Further demystifying prompts for language models with task-agnostic performance estimation. arXiv preprint arXiv:2406.11243.  
+Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023. Contrastive decoding: Open-ended text generation as optimization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12286-12312, Toronto, Canada. Association for Computational Linguistics.  
+Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4582-4597, Online. Association for Computational Linguistics.  
+Zichao Li, Xin Jiang, Lifeng Shang, and Hang Li. 2018. Paraphrase generation with deep reinforcement learning. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3865-3878, Brussels, Belgium. Association for Computational Linguistics.  
+Zichao Li, Xin Jiang, Lifeng Shang, and Qun Liu. 2019. Decomposable neural paraphrase generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3403-3414, Florence, Italy. Association for Computational Linguistics.  
+Shikun Liu, Linxi Fan, Edward Johns, Zhiding Yu, Chaowei Xiao, and Anima Anandkumar. 2023. Prismer: A vision-language model with multi-task experts. Transactions on Machine Learning Research.  
+Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, Noah A. Smith, and Yejin Choi. 2022a. NeuroLogic a*esque decoding: Constrained text generation with lookahead heuristics. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 780-799, Seattle, United States. Association for Computational Linguistics.
+
+Ximing Lu, Sean Welleck, Peter West, Liwei Jiang, Jungo Kasai, Daniel Khashabi, Ronan Le Bras, Lianhui Qin, Youngjae Yu, Rowan Zellers, et al. 2022b. Neurologic a* esque decoding: Constrained text generation with lookahead heuristics. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 780-799.  
+Yao Lu, Max Bartolo, Alastair Moore, Sebastian Riedel, and Pontus Stenetorp. 2022c. Fantastically ordered prompts and where to find them: Overcoming few-shot prompt order sensitivity. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8086-8098, Dublin, Ireland. Association for Computational Linguistics.  
+Andrew L. Maas, Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 142-150, Portland, Oregon, USA. Association for Computational Linguistics.  
+Tao Meng, Sidi Lu, Nanyun Peng, and Kai-Wei Chang. 2022. Controllable text generation with neurally-decomposed oracle. Advances in Neural Information Processing Systems, 35:28125-28139.  
+Yuxian Meng, Xiang Ao, Qing He, Xiaofei Sun, Qinghong Han, Fei Wu, Chun Fan, and Jiwei Li. 2021. ConRPG: Paraphrase generation using contexts as regularizer. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2551-2562, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Nuno Moniz and Luis Torgo. 2018. Multi-source social feedback of online news feeds. arXiv preprint arXiv:1801.07055.  
+Preksha Nema, Mitesh M. Khapra, Anirban Laha, and Balaraman Ravindran. 2017. Diversity driven attention model for query-based abstractive summarization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1063-1072, Vancouver, Canada. Association for Computational Linguistics.  
+Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744.  
+Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. 2023. GrIPS: Gradient-free, edit-based instruction search for prompting large language models. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational
+
+Linguistics, pages 3845-3864, Dubrovnik, Croatia.  
+Association for Computational Linguistics.  
+Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551.  
+Victor Sanh, Albert Webson, Colin Raffel, Stephen H Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Teven Le Scao, Arun Raja, et al. 2021. Multitask prompted training enables zero-shot task generalization. arXiv preprint arXiv:2110.08207.  
+Elvis Saravia, Hsien-Chi Toby Liu, Yen-Hao Huang, Junlin Wu, and Yi-Shin Chen. 2018. CARER: Contextualized affect representations for emotion recognition. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 3687-3697, Brussels, Belgium. Association for Computational Linguistics.  
+Timo Schick and Hinrich Schütze. 2021. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2339-2352, Online. Association for Computational Linguistics.  
+Taylor Shin, Yasaman Razeghi, Robert L. Logan IV, Eric Wallace, and Sameer Singh. 2020. AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4222-4235, Online. Association for Computational Linguistics.  
+KaShun Shum, Shizhe Diao, and Tong Zhang. 2023. Automatic prompt augmentation and selection with chain-of-thought from labeled data. arXiv preprint arXiv:2302.12822.  
+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.  
+Jiao Sun, Xuezhe Ma, and Nanyun Peng. 2021. AESOP: Paraphrase generation with adaptive syntactic control. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5176-5189, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+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.
+
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023a. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288.  
+Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 353-355, Brussels, Belgium. Association for Computational Linguistics.  
+Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amirreza Mirzaei, Atharva Naik, Arjun Ashok, Arut Selvan Dhanasekaran, Anjana Arunkumar, David Stap, Eshaan Pathak, Giannis Karamanolakis, Haizhi Lai, Ishan Purohit, Ishani Mondal, Jacob Anderson, Kirby Kuznia, Krima Doshi, Kuntal Kumar Pal, Maitreya Patel, Mehrad Moradshahi, Mihir Parmar, Mirali Purohit, Neeraj Varshney, Phani Rohitha Kaza, Pulkit Verma, Ravsehaj Singh Puri, Rushang Karia, Savan Doshi, Shailaja Keyur Sampat, Siddhartha Mishra, Sujan Reddy A, Sumanta Patro, Tanay Dixit, and Xudong Shen. 2022. Super-Natural Instructions: Generalization via declarative instructions on  $1600+$  NLP tasks. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 5085-5109, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.  
+Zezhong Wang, Luyao Ye, Hongru Wang, Wai-Chung Kwan, David Ho, and Kam-Fai Wong. 2023. Read-Prompt: A readable prompting method for reliable knowledge probing. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7468-7479, Singapore. Association for Computational Linguistics.  
+Alex Warstadt, Amanpreet Singh, and Samuel R. Bowman. 2019. Neural network acceptability judgments. Transactions of the Association for Computational Linguistics, 7:625-641.  
+Albert Webson and Ellie Pavlick. 2022. Do prompt-based models really understand the meaning of their prompts? In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2300-2344, Seattle, United States. Association for Computational Linguistics.  
+Joern Wuebker, Hermann Ney, and Richard Zens. 2012. Fast and scalable decoding with language model lookahead for phrase-based statistical machine translation.
+
+In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 28-32, Jeju Island, Korea. Association for Computational Linguistics.  
+Nan Xu, Chunting Zhou, Asli Celikyilmaz, and Xuezhe Ma. 2023. Look-back decoding for open-ended text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1039-1050, Singapore. Association for Computational Linguistics.  
+Tianyi Yan, Fei Wang, James Y. Huang, Wenxuan Zhou, Fan Yin, Aram Galstyan, Wenpeng Yin, and Muhao Chen. 2024. Contrastive instruction tuning. In *Findings of the Association for Computational Linguistics* ACL 2024, pages 10288-10302, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics.  
+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.  
+Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc.  
+Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International Conference on Machine Learning, pages 12697-12706. PMLR.  
+Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2022. Large language models are human-level prompt engineers. arXiv preprint arXiv:2211.01910.  
+Banghua Zhu, Evan Frick, Tianhao Wu, Hanlin Zhu, and Jiantao Jiao. 2023a. Starling-7b: Improving llm helpfulness & harmlessness with rlaif.  
+Kajjie Zhu, Jindong Wang, Jiaheng Zhou, Zichen Wang, Hao Chen, Yidong Wang, Linyi Yang, Wei Ye, Neil Zhenqiang Gong, Yue Zhang, et al. 2023b. Promptbench: Towards evaluating the robustness of large language models on adversarial prompts. arXiv preprint arXiv:2306.04528.
+
+# Appendices
+
+# A Algorithm for Search-based MONOPARA
+
+Algorithm 2 Search-based Decoding  
+Require:  $V$  ▷ Vocabulary  $P_{tar}, P_{para}$  ▷ Target and paraphrase models  $k$  ▷ Number of candidate sequences to retain  
+Ensure:  $x_{best}$  ▷ Best generated sequence  
+1:  $X_0 \gets \{\emptyset\}$  ▷ Initialize with empty sequence  
+2: while not end do  
+3:  $m \gets$  current decoding step  
+4:  $X_m' \gets \emptyset$   
+5: for each  $x_{<m} \in X_{m-1}$  do  
+6: for each  $x_m \in V$  do  
+7: Append  $x_{<m} \circ x_m$  to  $X_m'$   
+8: end for  
+9: end for  
+10: Use  $P_{para}$  for  $k$  candidates for each  $x_{<m}$   
+11:  $X_m \gets \arg \operatorname{top}_k \{f(x_{<m}, x_m)\} \to (x_{<m}, x_m) \in X_m'$   
+Select  $k$  candidates with lowest perplexity  
+12: end ← If any sequence ends with EOS  
+13: end while  
+14:  $x_{best} \gets$  select the sequence from  $X_m$  with the lowest perplexity by  $P_{tar}$   
+15: return  $x_{best}$
+
+# B Results on Llama-3-8B
+
+Results on Llama-3-8B are illustrated in Tab. 3.
+
+# C Correlation Between Accuracy and Perplexity
+
+To illustrate the interplay between task accuracy and prompt perplexity, we calculate the Pearson and Spearman correlation for each task based on the results in Tab. 1. The results are shown in Tab. 4. Most tasks show a strong negative correlation between perplexity and accuracy on both models except AG News, which is also discovered by Gonen et al. (2023) that for some of the tasks, there is not a negative correlation between perplexity and accuracy. The paraphrasing may be more helpful for cases where the nuances of expression don't matter so much (or there's enough room for error), but in domains where precision of language is important, perhaps perplexity-based paraphrasing might not
+
+be as helpful. AG News is a four-class news classification task with labels of "World", "Business", "Sci/Tech", and "Sports", which are not as easy as other binary classification tasks. As a result, the precision of the instruction for AG News is important. Additionally, we can see that the accuracy of AG News is not affected much under different methods, indicating that the accuracy is not so dependent on the perplexity of the prompt as other tasks. This indicates that MONOPARA works best for tasks with high distinction among labels.
+
+# D Reranking for MONOPARA
+
+# D.1 Implementation Setting
+
+We implemented reranking on top of the linear combination of  $P_{tar}$  and  $P_{para}$  under two settings: (1) the coefficient  $\alpha$  equals 0.5, which is kept consistent with the results presented in Tab. 1 for direct comparison; (2) the coefficient is randomly selected from [0.1, 0.2, ..., 0.6] for a larger search space. In both cases, we allow sampling from the top 3 tokens at every decoding step for randomness.
+
+For each task, we use the 5 prompts as the original prompts (the same setting as Tab. 1) and randomly generate 100 paraphrases with the linear combination of  $P_{tar}$  and  $P_{para}$  for every single original prompt. For the generated 100 paraphrases, we use the same scheme as is described in §3 to calculate the perplexity: each paraphrase is instantiated with 1,000 random test samples, and the perplexity is averaged across all the 1,000 instantiated prompts. For budget  $k$ , we select the first  $k$  paraphrases from the 100 candidates and choose the prompt with the lowest average instantiated perplexity to be tested on the task.
+
+# D.2 Analysis
+
+For the reranking setting (1), with additional randomness when sampling paraphrases from the linear combination of  $P_{tar}$  and  $P_{para}$ , a large enough budget (i.e., 100 paraphrases) can consistently catch paraphrased prompts with lower perplexity which achieve better task performance (Tab. 5). However, with a small budget of 10 or even 50, reranking can hardly surpass Mono-E in either perplexity or task performance. The reason is that with randomness allowed, the paraphrase is no longer decoded greedily at each token, which undermines the constraint on the perplexity and leads to a large variation of the perplexity of the resulting paraphrase. This further affects the performance of
+
+<table><tr><td>Methods</td><td>Metric</td><td>GLUE Cola</td><td>Newspop</td><td>AG News</td><td>Tweet Offensive</td></tr><tr><td rowspan="2">Ori.</td><td>Acc</td><td>54.82 ± 1.73</td><td>73.82 ± 4.57</td><td>60.37 ± 7.59</td><td>63.25 ± 0.95</td></tr><tr><td>PPL</td><td>9.42 ± 5.08</td><td>17.72 ± 6.11</td><td>13.64 ± 4.97</td><td>23.64 ± 4.61</td></tr><tr><td rowspan="2">Para</td><td>Acc</td><td>43.69 ± 9.02</td><td>68.45 ± 13.24</td><td>64.85 ± 3.73</td><td>64.22 ± 1.48</td></tr><tr><td>PPL</td><td>32.07 ± 8.43</td><td>29.41 ± 8.17</td><td>23.77 ± 4.32</td><td>46.14 ± 9.24</td></tr><tr><td rowspan="2">SPELL</td><td>Acc</td><td>46.91 ± 2.33</td><td>71.67 ± 5.30</td><td>66.93 ± 3.70</td><td>64.73 ± 0.76</td></tr><tr><td>PPL</td><td>7.98 ± 1.58</td><td>12.47 ± 3.76</td><td>11.29 ± 2.79</td><td>8.79 ± 1.93</td></tr><tr><td rowspan="2">Mono-E</td><td>Acc</td><td>58.35 ± 2.61</td><td>77.85 ± 4.01</td><td>65.45 ± 2.11</td><td>64.90 ± 0.59</td></tr><tr><td>PPL</td><td>6.64 ± 0.95</td><td>9.17 ± 2.59</td><td>8.29 ± 1.47</td><td>6.40 ± 3.61</td></tr></table>
+
+Table 3: Performance of Llama-3-8B.  
+
+<table><tr><td>Model</td><td>Metric</td><td>GLUE Cola</td><td>Newspop</td><td>AG News</td><td>IMDB</td><td>DBpedia</td><td>Emotion</td><td>Tweet Offensive</td></tr><tr><td rowspan="2">Mistral</td><td>Pearson</td><td>-0.62</td><td>-0.88</td><td>0.19</td><td>-0.04</td><td>-0.70</td><td>-0.57</td><td>-0.48</td></tr><tr><td>Spearman</td><td>-0.30</td><td>-0.99</td><td>0.30</td><td>-0.30</td><td>-0.39</td><td>-0.70</td><td>-0.30</td></tr><tr><td rowspan="2">Starling</td><td>Pearson</td><td>-0.83</td><td>-0.69</td><td>0.16</td><td>-0.60</td><td>-0.70</td><td>-0.96</td><td>-0.05</td></tr><tr><td>Spearman</td><td>-0.90</td><td>-0.49</td><td>0.3</td><td>-0.3</td><td>-0.70</td><td>-0.89</td><td>-0.15</td></tr></table>
+
+Table 4: Pearson and Spearman correlation between task accuracy and prompt perplexity w.r.t. the target model. Results are calculated based on Tab. 1.
+
+these paraphrased prompts, which in turn proves the correlation between perplexity and task accuracy. Generally, we can expect vanilla Mono-E and reranking to break even with a large budget of around 50 paraphrases.
+
+The reranking setting (2) is not as effective as setting (1), where even a budget of 100 could not surpass the vanilla Mono-E. The reason is that although a large search space is allowed with randomness on the combination coefficient  $\alpha$ , the constraint on perplexity while decoding is undermined even further. With more randomness for paraphrase sampling, the results of setting (2) also show a higher standard deviation than setting (1).
+
+# E Correlation Between Standalone and Instantiated Prompts
+
+To bridge the intuition gap, we investigate the correlation between the standalone prompt and its instantiated prompts. We instantiate each task-specific prompt with 1,000 input instances and the perplexity used for correlation evaluation is averaged over 5 prompts for each task. As shown in Tab. 6, the Pearson correlation is high (close to 1.0) in most cases for different datasets despite slight noise for Emotion under greedy paraphrase of prompts. We can conclude that there is a high correlation be
+
+tween instantiated and standalone prompts. Thus, it is valid to use the perplexity of the standalone prompt optimized by  $P_{tar}$  to serve as the constraint.
+
+# F Unperturbed Accuracy for Task II
+
+Tab. 7 shows the task performance of three perturbation methods (Alpaca's average accuracy across 4 GLUE datasets) and performance degradation compared with the unperturbed prompt (as shown in the brackets). We can witness a narrow performance gap between perturbed and unperturbed original prompts for Mono-E, which is more robust than directly using the given prompt (Ori) or the greedily paraphrased one (Para). The reason is that Mono-E largely mitigates the effect of adversarial perturbations on the prompts and successfully rewrites the given prompts (whether perturbed or unperturbed) into higher-quality versions.
+
+# G BERTScore of Para and MONOPARA
+
+We provide a case study on the paraphrase results and the according BERTScore in Tab. 8. We could see that both Para and Mono-E largely preserve the semantic meaning of the original prompt. Further, Para tends to use new vocabulary for diversity, while Mono-E prefers words that are more familiar to the target model, which is often the same as
+
+<table><tr><td>Method</td><td>Metric</td><td>GLUE Cola</td><td>Newspop</td><td>AG News</td><td>IMDB</td><td>DBpedia</td><td>Emotion</td><td>Tweet Offensive</td></tr><tr><td rowspan="2">Ori.</td><td>Acc</td><td>61.76 ± 2.93</td><td>85.28 ± 3.45</td><td>67.68 ± 3.47</td><td>84.12 ± 4.89</td><td>70.82 ± 2.27</td><td>44.32 ± 1.57</td><td>67.04 ± 0.95</td></tr><tr><td>PPL</td><td>15.72 ± 4.60</td><td>12.60 ± 2.40</td><td>12.74 ± 1.07</td><td>16.41 ± 0.80</td><td>8.10 ± 0.22</td><td>16.48 ± 2.14</td><td>19.25 ± 3.43</td></tr><tr><td rowspan="2">Para</td><td>Acc</td><td>39.48 ± 8.19</td><td>76.72 ± 16.54</td><td>71.60 ± 2.70</td><td>86.60 ± 4.86</td><td>70.22 ± 5.71</td><td>43.65 ± 3.41</td><td>65.04 ± 1.46</td></tr><tr><td>PPL</td><td>78.01 ± 11.22</td><td>49.37 ± 7.81</td><td>28.37 ± 2.58</td><td>22.27 ± 0.98</td><td>28.34 ± 1.36</td><td>46.53 ± 5.30</td><td>53.65 ± 1.92</td></tr><tr><td rowspan="2">Mono-E</td><td>Acc</td><td>63.96 ± 3.88</td><td>88.96 ± 3.04</td><td>72.20 ± 3.87</td><td>87.58 ± 3.27</td><td>72.07 ± 3.54</td><td>47.76 ± 3.72</td><td>67.00 ± 1.21</td></tr><tr><td>PPL</td><td>14.89 ± 2.75</td><td>8.33 ± 1.30</td><td>10.07 ± 2.93</td><td>13.05 ± 1.86</td><td>8.85 ± 0.88</td><td>13.67 ± 1.87</td><td>14.11 ± 3.53</td></tr><tr><td colspan="9">Rerank (1)</td></tr><tr><td rowspan="2">Budget = 10</td><td>Acc</td><td>62.20 ± 3.47</td><td>86.14 ± 3.14</td><td>69.31 ± 4.29</td><td>85.02 ± 4.13</td><td>69.88 ± 4.33</td><td>47.36 ± 3.28</td><td>66.12 ± 2.15</td></tr><tr><td>PPL</td><td>14.27 ± 2.69</td><td>11.46 ± 1.99</td><td>11.92 ± 3.97</td><td>15.95 ± 2.83</td><td>13.15 ± 1.24</td><td>15.06 ± 2.71</td><td>15.63 ± 3.57</td></tr><tr><td rowspan="2">Budget = 50</td><td>Acc</td><td>63.84 ± 2.46</td><td>85.86 ± 5.36</td><td>68.28 ± 5.16</td><td>86.52 ± 5.86</td><td>69.64 ± 3.53</td><td>47.60 ± 5.25</td><td>67.74 ± 1.45</td></tr><tr><td>PPL</td><td>13.54 ± 3.16</td><td>10.53 ± 2.85</td><td>10.83 ± 3.24</td><td>13.44 ± 2.22</td><td>10.31 ± 1.85</td><td>14.14 ± 3.46</td><td>13.94 ± 3.86</td></tr><tr><td rowspan="2">Budget = 100</td><td>Acc</td><td>64.58 ± 3.50</td><td>89.82 ± 2.04</td><td>72.33 ± 4.17</td><td>87.61 ± 4.93</td><td>72.52 ± 3.16</td><td>47.70 ± 3.56</td><td>68.5 ± 1.84</td></tr><tr><td>PPL</td><td>10.22 ± 2.32</td><td>7.53 ± 1.82</td><td>9.52 ± 4.20</td><td>12.88 ± 3.01</td><td>8.16 ± 1.67</td><td>13.05 ± 1.34</td><td>11.56 ± 4.44</td></tr><tr><td colspan="9">Rerank (2)</td></tr><tr><td rowspan="2">Budget = 10</td><td>Acc</td><td>62.35 ± 4.77</td><td>86.96 ± 4.61</td><td>70.02 ± 5.27</td><td>84.86 ± 4.71</td><td>69.31 ± 3.94</td><td>46.10 ± 7.29</td><td>66.63 ± 4.53</td></tr><tr><td>PPL</td><td>15.63 ± 3.85</td><td>12.27 ± 3.99</td><td>13.47 ± 3.63</td><td>16.43 ± 2.17</td><td>15.37 ± 2.14</td><td>16.64 ± 3.88</td><td>18.87 ± 4.48</td></tr><tr><td rowspan="2">Budget = 50</td><td>Acc</td><td>62.08 ± 3.97</td><td>86.50 ± 5.10</td><td>71.12 ± 6.97</td><td>86.03 ± 5.24</td><td>70.13 ± 4.23</td><td>46.90 ± 5.83</td><td>65.15 ± 5.03</td></tr><tr><td>PPL</td><td>14.42 ± 4.16</td><td>10.24 ± 4.36</td><td>11.69 ± 2.96</td><td>14.33 ± 3.55</td><td>13.62 ± 2.77</td><td>15.21 ± 3.38</td><td>18.53 ± 4.54</td></tr><tr><td rowspan="2">Budget = 100</td><td>Acc</td><td>64.35 ± 4.91</td><td>87.82 ± 4.48</td><td>71.72 ± 6.07</td><td>86.64 ± 5.16</td><td>71.98 ± 3.85</td><td>47.18 ± 4.63</td><td>67.80 ± 8.95</td></tr><tr><td>PPL</td><td>11.67 ± 4.37</td><td>8.93 ± 3.15</td><td>10.61 ± 4.89</td><td>13.47 ± 3.84</td><td>8.36 ± 3.57</td><td>14.41 ± 2.95</td><td>16.65 ± 3.53</td></tr></table>
+
+Table 5: Reranking for ensemble-based MONOPARA. The best results of each task are bolded and the second-best are italicized.  
+
+<table><tr><td></td><td>Newspop</td><td>IMDB</td><td>DBpedia</td><td>Emotion</td><td>Tweet</td></tr><tr><td>Ori</td><td>0.949</td><td>0.952</td><td>0.895</td><td>0.839</td><td>0.974</td></tr><tr><td>Para</td><td>0.813</td><td>0.951</td><td>0.835</td><td>0.645</td><td>0.871</td></tr><tr><td>Mono-E</td><td>0.956</td><td>0.895</td><td>0.927</td><td>0.931</td><td>0.968</td></tr></table>
+
+Table 6: Pearson correlation between the perplexity of standalone prompt and its instantiated prompts. Pearson correlation is high (close to 1.0) in most cases, indicating a high correlation.
+
+<table><tr><td></td><td>Ori</td><td>Para</td><td>Mono-E</td></tr><tr><td>Un.</td><td>56.67</td><td>58.31</td><td>65.94</td></tr><tr><td>Cha.</td><td>51.03 (-9.95%)</td><td>53.62 (-8.04%)</td><td>65.81 (-0.20%)</td></tr><tr><td>Word</td><td>52.41 (-7.51%)</td><td>57.39 (-1.58%)</td><td>67.23 (+1.96%)</td></tr><tr><td>Sem.</td><td>49.42 (-12.79%)</td><td>54.22 (-7.01%)</td><td>64.35 (-2.4%)</td></tr></table>
+
+Table 7: Performance degradation of Alpaca model compared with the unperturbed prompt. Un., Cha., and Sem. are short for unperturbed, character, and semantic, respectively, referring to different prompt perturbation methods.
+
+in the original prompt. This might be the reason why Para results in lower BertScore compared to Mono-E and Mono-S.
+
+# H Ablation Study on Alpha Value
+
+We provide results with various alpha values in Tab. 9, which illustrates the impact of alpha values on model accuracy and paraphrase perplexity across various datasets on the Mistral-7B model. With the increase of  $\alpha$  value, the perplexity (PPL) of the paraphrase monotonically goes down due to the design and motivation of our method. Given the low perplexity, we compare the task accuracy of various alpha values. For  $\alpha = 0.4$  or 0.6, the model shows outstanding performance on some of the datasets. But increasing alpha to 0.7 results in a nonsensical prompt though the perplexity is super low. From the results we can also see that the performance just corrupted with this nonsensical prompt. The setting of  $\alpha$  should seek a balance between lower perplexity and the preservation of the original semantics. As a result, alpha  $= 0.5$  (Mono-E) offers a balanced and reliable performance, making it a suitable default choice.
+
+<table><tr><td></td><td>Content</td><td>BertScore</td></tr><tr><td>Ori Prompt</td><td>In this task, you are presented with a short article. Your objective is to classify the article according to its category using the following labels: World, Sports, Business, Sci/Tech. Tag the text as “World” if it involves information about global events or issues. Mark it as “Sports” if it deals with sports-related content. Label it “Business” if it pertains to business affairs, markets, or economics. Assign “Sci/Tech” if it covers scientific or technological developments.</td><td>-</td></tr><tr><td>Para</td><td>In this assignment, you are given a brief article. Your goal is to categorize the article based on its topic using these tags: World, Sports, Business, Sci/Tech. Identify it as “World” if it discusses international events or matters. Designate it as “Sports” if it focuses on athletic content. Categorize it as “Business” if it concerns commercial activities, markets, or the economy. Label it “Sci/Tech” if it reports on advancements in science or technology.</td><td>86.02</td></tr><tr><td>Mono-E</td><td>In your task you will have an input which consists of a single article. You have to sort this article according to its category using a list which consists of 4 categories: “World”, “Sci/tech”, “Business”, or “Sports”. You should mark it with the appropriate label based on the content. The article should be labeled with the “Sports” category, when its topic deals with sports. When the article is about business, you need to mark it as a ‘business’ label. If you are talking about science and tech developments then label the content with “Sci/Tech”, if it is a world-related issue or event then label it with “World”.</td><td>89.63</td></tr></table>
+
+Table 8: Case study on the paraphrase results and BERTScore. Para tends to use new vocabulary for diversity, while Mono-E prefers words that are more familiar to the target model, which is often the same as in the original prompt.  
+
+<table><tr><td>Method</td><td>Metric</td><td>GLUE Cola</td><td>Newspop</td><td>AG News</td><td>IMDB</td><td>DBpedia</td><td>Emotion</td><td>Tweet</td></tr><tr><td rowspan="2">Ori.</td><td>Acc</td><td>61.76 ± 2.93</td><td>85.28 ± 3.45</td><td>67.68 ± 3.47</td><td>84.12 ± 4.89</td><td>70.82 ± 2.27</td><td>44.32 ± 1.57</td><td>67.04 ± 0.95</td></tr><tr><td>PPL</td><td>15.72 ± 4.60</td><td>12.60 ± 2.40</td><td>12.74 ± 1.07</td><td>16.41 ± 0.80</td><td>8.10 ± 0.22</td><td>16.48 ± 2.14</td><td>19.25 ± 3.43</td></tr><tr><td rowspan="2">Para</td><td>Acc</td><td>39.48 ± 8.19</td><td>76.72 ± 16.54</td><td>71.60 ± 2.70</td><td>86.60 ± 4.86</td><td>70.22 ± 5.71</td><td>43.65 ± 3.41</td><td>65.04 ± 1.46</td></tr><tr><td>PPL</td><td>78.01 ± 11.22</td><td>49.37 ± 7.81</td><td>28.37 ± 2.58</td><td>22.27 ± 0.98</td><td>28.34 ± 1.36</td><td>46.53 ± 5.30</td><td>53.65 ± 1.92</td></tr><tr><td rowspan="2">Mono-E</td><td>Acc</td><td>63.96 ± 3.88</td><td>88.96 ± 3.04</td><td>72.20 ± 3.87</td><td>87.58 ± 3.27</td><td>72.07 ± 3.54</td><td>47.76 ± 3.72</td><td>67.00 ± 1.21</td></tr><tr><td>PPL</td><td>14.89 ± 2.75</td><td>8.33 ± 1.30</td><td>10.07 ± 2.93</td><td>13.05 ± 1.86</td><td>8.85 ± 0.88</td><td>13.67 ± 1.87</td><td>14.11 ± 3.53</td></tr><tr><td rowspan="2">α = 0.1</td><td>Acc</td><td>60.80 ± 4.28</td><td>84.98 ± 3.41</td><td>69.09 ± 4.01</td><td>85.52 ± 1.72</td><td>70.42 ± 3.66</td><td>44.80 ± 2.31</td><td>65.56 ± 5.07</td></tr><tr><td>PPL</td><td>15.61 ± 2.51</td><td>12.32 ± 1.01</td><td>12.03 ± 2.96</td><td>15.70 ± 1.85</td><td>9.61 ± 0.61</td><td>16.16 ± 1.11</td><td>18.49 ± 2.84</td></tr><tr><td rowspan="2">α = 0.2</td><td>Acc</td><td>61.80 ± 5.63</td><td>85.32 ± 3.56</td><td>71.49 ± 4.32</td><td>85.08 ± 4.97</td><td>70.54 ± 5.23</td><td>44.04 ± 2.44</td><td>66.80 ± 5.19</td></tr><tr><td>PPL</td><td>15.22 ± 2.06</td><td>11.56 ± 1.27</td><td>11.68 ± 3.25</td><td>15.30 ± 1.55</td><td>9.23 ± 0.79</td><td>15.71 ± 0.20</td><td>16.92 ± 2.68</td></tr><tr><td rowspan="2">α = 0.3</td><td>Acc</td><td>61.54 ± 5.85</td><td>86.72 ± 4.94</td><td>72.4 ± 2.77</td><td>86.68 ± 4.21</td><td>71.90 ± 3.91</td><td>44.96 ± 2.67</td><td>67.56 ± 3.59</td></tr><tr><td>PPL</td><td>15.18 ± 2.84</td><td>10.21 ± 1.22</td><td>11.06 ± 2.74</td><td>14.42 ± 1.05</td><td>8.95 ± 0.75</td><td>15.21 ± 1.04</td><td>15.87 ± 3.61</td></tr><tr><td rowspan="2">α = 0.4</td><td>Acc</td><td>62.72 ± 4.37</td><td>89.44 ± 3.08</td><td>71.31 ± 2.27</td><td>87.86 ± 3.37</td><td>72.18 ± 3.44</td><td>45.58 ± 1.55</td><td>67.70 ± 1.60</td></tr><tr><td>PPL</td><td>14.93 ± 2.91</td><td>9.16 ± 0.57</td><td>10.39 ± 3.56</td><td>13.81 ± 2.60</td><td>9.04 ± 1.18</td><td>14.35 ± 0.90</td><td>15.18 ± 3.72</td></tr><tr><td rowspan="2">α = 0.6</td><td>Acc</td><td>64.30 ± 6.63</td><td>90.86 ± 3.24</td><td>71.66 ± 10.09</td><td>87.74 ± 4.04</td><td>71.36 ± 6.25</td><td>47.12 ± 1.66</td><td>66.26 ± 3.24</td></tr><tr><td>PPL</td><td>11.49 ± 1.00</td><td>7.53 ± 0.68</td><td>9.38 ± 0.88</td><td>12.15 ± 2.47</td><td>7.11 ± 1.08</td><td>13.12 ± 0.79</td><td>13.29 ± 2.58</td></tr><tr><td rowspan="2">α = 0.7</td><td>Acc</td><td>41.50 ± 13.68</td><td>54.52 ± 22.29</td><td>50.56 ± 21.95</td><td>64.84 ± 16.01</td><td>48.12 ± 16.26</td><td>38.62 ± 1.59</td><td>59.34 ± 7.75</td></tr><tr><td>PPL</td><td>8.80 ± 0.42</td><td>3.84 ± 0.91</td><td>3.43 ± 0.36</td><td>4.99 ± 0.28</td><td>2.34 ± 0.25</td><td>4.28 ± 0.55</td><td>6.58 ± 1.40</td></tr></table>
+
+Table 9: Results with varying  $\alpha$  values for Mono-E on Mistral-7B. The best accuracy for each task is in bold. For  $\alpha = 0.4$  or 0.6, the model shows outstanding performance on most of the datasets. But increasing alpha to 0.7 results in a nonsensical prompt though the perplexity is super low, leading to corrupted task performance.

2024/Monotonic Paraphrasing Improves Generalization of Language Model Prompting/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37e7ae3382b29eb5a326237d1fd69e82d47a97c44923a7cceec4dcafae288df6
+size 1128877

2024/More Bang for your Context_ Virtual Documents for Question Answering over Long Documents/f22392c5-4a70-4295-a2be-84f110b6b3b3_content_list.json

@@ -0,0 +1,1337 @@
+[
+    {
+        "type": "text",
+        "text": "More Bang for your Context: Virtual Documents for Question Answering over Long Documents",
+        "text_level": 1,
+        "bbox": [
+            156,
+            89,
+            842,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Yosi Mass, Boaz Carmeli, Asaf Yehudai, Assaf Toledo, Nathaniel Mills IBM Research AI",
+        "bbox": [
+            198,
+            142,
+            805,
+            174
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{yosimass,boazc}@il.ibm.com, {Asaf.Yehudai,assaf.toledo} @ibm.com, wnrm3@us.ibm.com",
+        "bbox": [
+            287,
+            177,
+            715,
+            225
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            341,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We deal with the problem of Question Answering (QA) over a long document, which poses a challenge for modern Large Language Models (LLMs). Although LLMs can handle increasingly longer context windows, they struggle to effectively utilize the long content. To address this issue, we introduce the concept of a virtual document (VDoc). A VDoc is created by selecting chunks from the original document that are most likely to contain the information needed to answer the user's question, while ensuring they fit within the LLM's context window. We hypothesize that providing a short and focused VDoc to the LLM is more effective than filling the entire context window with less relevant information. Our experiments confirm this hypothesis and demonstrate that using VDocs improves results on the QA task.",
+        "bbox": [
+            141,
+            280,
+            460,
+            537
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            550,
+            260,
+            565
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Large language models (LLMs) such as OpenAI's ChatGPT<sup>1</sup> or Anthropic's Claude<sup>2</sup>, have demonstrated exceptional performance across a range of natural language processing (NLP) tasks and are capable of answering questions on various topics. Recent advancements have extended their context window to encompass tens of thousands or even millions of tokens (Gemini, 2024). This enables them to process long documents, or even books, and answer questions over them.",
+        "bbox": [
+            112,
+            576,
+            489,
+            736
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Despite this impressive progress, recent studies have shown that LLMs still struggle to utilize long contexts efficiently (Li et al., 2024; Levy et al., 2024). LLMs' performance often deteriorates when confronted with a long document, hindering their ability to accurately answer questions on such documents.",
+        "bbox": [
+            112,
+            738,
+            489,
+            848
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "To address this challenge, we propose a method called Virtual Document (VDoc). VDoc takes a",
+        "bbox": [
+            112,
+            851,
+            487,
+            882
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "image",
+        "img_path": "images/e0049b320b5563f222930bf215eb4838edeeb607df6c09c7c08fe03e131fb61a.jpg",
+        "image_caption": [
+            "Figure 1: VDoc architecture. First, a long document is segmented (a), then each segment is ranked by its relevance to the question (b), and finally a virtual document is reconstructed with the top-ranked segments (c)."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            250,
+            875,
+            372
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "long document, a question, and a specified context length, and generates a shorter version of the document, containing the most relevant parts for the given length. VDoc operates in three steps: 1) Document segmentation: break down the long document into smaller segments, 2) Segments ranking: rank the segments based on their relevance to the question, and 3) VDoc reconstruction: reassemble the segments into a coherent virtual document that contains the most relevant information to the question, and still fits the given context length. The overall VDoc process is illustrated in Figure 1.",
+        "bbox": [
+            507,
+            461,
+            884,
+            653
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "We analyze experimentally, different approaches for each step. Our results show that VDoc enhances question-answering performance on various datasets, including ZeroScrolls (Shaham et al., 2023), a benchmark designed for tasks requiring a long context.",
+        "bbox": [
+            507,
+            656,
+            885,
+            753
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "The paper's main contribution is the thorough analysis and experiments of the above three steps for QA over a long document. While the suggested solution is close to RAG (Retrieval Augmented Generation), it has some unique properties. For example, the ordering of the segments based on their sequence in the original document, that does not exist in RAG. Our code is available on git<sup>3</sup>",
+        "bbox": [
+            507,
+            756,
+            884,
+            885
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>1</sup>https://chat.openai.com/",
+        "bbox": [
+            134,
+            890,
+            292,
+            904
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://www.anthropic.com/news/claude-2-1",
+        "bbox": [
+            136,
+            904,
+            413,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://github.com/IBM/vdoc",
+        "bbox": [
+            529,
+            903,
+            786,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "12936",
+        "bbox": [
+            475,
+            927,
+            527,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 12936-12942 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            210,
+            945,
+            788,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            83,
+            270,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Many previous works focus on designing LLMs that can support long context. Some works modify the transformer attention mechanism (Beltagy et al., 2020; Zaheer et al., 2021; Dai et al., 2019; Ratner et al., 2023) or make it more efficient (Dao et al., 2022; Dao, 2023; Liu et al., 2023a; Gemini, 2024). Others have focused on modifying the positional embeddings (Press et al., 2022; Chen et al., 2023). Another line of research attempts to replace transformers, which have quadratic sequence length complexity, with new architectural designs like convolution and linear RNNs, e.g., RWKV (Peng et al., 2023), S4 (Gu et al., 2022), Hyena (Poli et al., 2023), or Mamba (Gu and Dao, 2023).",
+        "bbox": [
+            112,
+            110,
+            489,
+            335
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "While those works were able to extend the context length of LLMs, subsequent research has found that long-context transformers exhibit a recency bias and do not effectively utilize long-range context (Qin et al., 2023; Liu et al., 2023b; Yehudai and Bendel, 2024; Li et al., 2024). Specifically, these models have been found to struggle with ignoring irrelevant information (Levy et al., 2024). Moreover, benchmarks for long context reveal that there is still significant room for improvement compared to human performance, and for some tasks, models even struggle to surpass the naive baseline (Shaham et al., 2023).",
+        "bbox": [
+            115,
+            336,
+            489,
+            545
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Method",
+        "text_level": 1,
+        "bbox": [
+            114,
+            558,
+            218,
+            571
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We deal with the problem of using an LLM to answer questions over a long document. Our method, termed VDoc, creates a shorter version of the document, which is then used as input to the LLM, instead of the entire document, to generate an answer. Creating the VDoc is done in three steps. i) Document segmentation into manageable segments, ii) Segments ranking by their potential to contain the answer to the user question, and iii) VDoc reconstruction. The detailed steps are described below.",
+        "bbox": [
+            112,
+            583,
+            489,
+            760
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Document segmentation. We try two methods. i) sliding-window on sentence boundaries using nltk $^4$  ii) context-aware (using HTML structure when available). We define a segment-size parameter that controls the max segment length.",
+        "bbox": [
+            112,
+            771,
+            489,
+            852
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Segments ranking. We experiment with three families of language models (LM). The first is",
+        "bbox": [
+            112,
+            862,
+            487,
+            894
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "encoder-only models, where we rank segments by the similarity of their representation to the representation of the user question. An example of such LM is MiniLM<sup>5</sup>.",
+        "bbox": [
+            507,
+            84,
+            884,
+            148
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The second, is encoder-decoder or decoder-only models, where we rank segments by the inverse perplexity of the user question, given each of the segments. Specifically, we give each segment as input to the model with the instruction: \"Generate a conversation between a user and an agent based on the given content:\". We then run a forward loop and compute the Cross-Entropy loss for the question. The segments are ranked by  $r = (-1) \\times loss()$  (higher is better). An example is flan-t5<sup>6</sup>.",
+        "bbox": [
+            507,
+            149,
+            884,
+            309
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "The third, is a sparse representation, Elser<sup>7</sup> which expands and indexes documents into ElasticSearch<sup>8</sup> index, with semantic terms that were learned to co-occur frequently with terms in each document.",
+        "bbox": [
+            507,
+            310,
+            882,
+            388
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We note that among the three segment ranking methods, the first (comparing dense representations) and the last (Sparse representation) can be pre-computed on the given set of documents. The second method on the other hand (using perplexity), requires running a forward loop of the model for each segment and question, and thus can not be pre-computed, because it depends on the question. Thus the perplexity method can be quite slow for very long documents.",
+        "bbox": [
+            507,
+            391,
+            882,
+            552
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "VDoc reconstruction. We take the top ranked segments, that still fit into the given window size and evaluate two methods to concatenate them: i) by the original order in the document (Doc order) and ii) by relevancy to the question (Rank order).",
+        "bbox": [
+            507,
+            562,
+            882,
+            644
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "4 Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            655,
+            655,
+            671
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "4.1 Setup of the Experiments",
+        "text_level": 1,
+        "bbox": [
+            507,
+            682,
+            756,
+            697
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "We evaluated the VDoc performance in two scenarios: direct and end-to-end. For the direct scenario we assume a benchmark of triplets  $\\{(Q, D, E)\\}$  where  $Q$  is a question,  $D$  is a long document and  $E$  is an evidence i.e., a span in  $D$  that contains the answer. We define the VDoc task as follows: Given a context-window size, a question  $Q$ , and a long document  $D$ , create a virtual document that",
+        "bbox": [
+            507,
+            703,
+            884,
+            832
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{5}$ https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2",
+        "bbox": [
+            507,
+            841,
+            845,
+            866
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{6}$ https://huggingface.co/google/flan-t5-large",
+        "bbox": [
+            532,
+            866,
+            800,
+            879
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "7https://www/elastic.co/guide/en/machinelearning/current/ml-nlp-elser.html",
+        "bbox": [
+            510,
+            879,
+            794,
+            904
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>8</sup>https://www/elastic.co/",
+        "bbox": [
+            532,
+            904,
+            747,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_footnote",
+        "text": "4https://www.nltk.org/",
+        "bbox": [
+            134,
+            903,
+            336,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "12937",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "fits within the context-window size and includes E. We measure the success of the VDoc to contain the evidence (we counted  $90\\%$  containment as a success, to tolerate for segmentation issues).",
+        "bbox": [
+            112,
+            84,
+            487,
+            147
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the end-to-end scenario, we assume a benchmark of triplets  $\\{(\\mathrm{Q},\\mathrm{D},\\mathrm{R})\\}$  where  $\\mathbf{Q}$  is a question,  $\\mathrm{D}$  is a long document that contains the answer, and  $\\mathbf{R}$  is a gold response. We then assess LLM's response against the gold,  $\\mathbf{R}$ , when prompting using various VDoc strategies. We evaluate results using F1 (the harmonic mean of Recall and Precision), similar to the scrolls (Shaham et al., 2023) paper.",
+        "bbox": [
+            112,
+            149,
+            487,
+            275
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We experiment with two LLMs, each with 8K window size, thus we limit the VDoc to sizes of up to 7200 tokens to allow space for the prompt, the question and generated answer. In all experiments, we used segment-size of 512 tokens.",
+        "bbox": [
+            112,
+            278,
+            487,
+            357
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.2 Datasets",
+        "text_level": 1,
+        "bbox": [
+            112,
+            370,
+            226,
+            382
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the direct experiment, we utilize two datasets containing long documents, questions and evidence. The first is GoogleNQ (Kwiatkowski et al., 2019). This dataset contains questions from Wikipedia. Each question is labeled with the Wikipedia page that contains the answer, and a span in the document that contains the evidence. The second dataset is taken from Scrolls (Shaham et al.,  $2023)^{9}$  Scrolls contains several datasets with questions over long documents. We used Qasper which is question answering over research papers. We added an evidence for each question, from the original Qasper dataset<sup>10</sup>.",
+        "bbox": [
+            112,
+            391,
+            487,
+            599
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "For the end-to-end experiment we utilized the Qasper dataset described above and the NarrativeQA dataset from Scrolls. NarrativeQA contains question-answer pairs about entire books and movie scripts. Both datasets contain (Q, D, R) triplets.",
+        "bbox": [
+            112,
+            601,
+            487,
+            696
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We picked a random subset of 500 examples from the validation sets of Qasper and NarrativeQA. Likewise, for GoogleNQ, we picked 500 examples where the long answer (evidence) is a section.",
+        "bbox": [
+            112,
+            697,
+            487,
+            760
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Note that all our experiments were done on the validation sets of Qasper and NarrativeQA from Scrolls. We used the validation set and not the test set, since ground truths is available only on the validation set. We selected a random subset of 500 examples, inspired by the size of the test sets in",
+        "bbox": [
+            112,
+            763,
+            487,
+            858
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/aeb31868a989130fa7661af0a7a918e6953c250bd14bd10aee79ed9d69ced320.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Avg # tokens</td><td>Max # tokens</td></tr><tr><td>GoogleNQ</td><td>12,337</td><td>273,087</td></tr><tr><td>Qasper</td><td>5,484</td><td>22,052</td></tr><tr><td>NarrativeQA</td><td>88,073</td><td>516,045</td></tr></table>",
+        "bbox": [
+            544,
+            80,
+            848,
+            147
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Datasets used for evaluation. Number of tokens (#) is calculated by the flan-t5-small tokenizer.",
+        "bbox": [
+            507,
+            155,
+            880,
+            185
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ZeroScrolls. We tried the Qasper experiments on the full validation set and got very similar results, thus we performed the experiments on the 500 data points.",
+        "bbox": [
+            505,
+            211,
+            880,
+            274
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The datasets and their statistics are described in Table 1. The numbers are for the random subset of 500 examples. Note that in all our experiments, we treat each document separately, thus we report statistics only for the subset that was used in the experiments.",
+        "bbox": [
+            507,
+            275,
+            880,
+            370
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.3 Optimal Window Size",
+        "text_level": 1,
+        "bbox": [
+            507,
+            384,
+            727,
+            399
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "To confirm that a smaller window yields better generation results, we conducted a controlled experiment, using the evidence as a pivot. We expanded the content up to a specified window size, measured in tokens, while maintaining full sentences using the nltk tool.",
+        "bbox": [
+            505,
+            405,
+            882,
+            500
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We experiment with expanding the content on both sides of the evidence, as well as before and after it. All three methods yielded similar results. Figure 2 reports results when expanding the window in both sides of the evidence. We conducted the experiments using two generative LLM: flant5-xxl and llama-3-8b-instruct. As observed, both models achieve the highest F1 score when provided only with the evidence, and performance degrades as data is added.",
+        "bbox": [
+            507,
+            502,
+            882,
+            661
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "image",
+        "img_path": "images/244a2b0fe0aae0112b7fde4349f0db3fc201a85aae09f37358e986908836e087.jpg",
+        "image_caption": [
+            "Figure 2: Comparing performances (F1) of response generation on Qasper when adding non-relevant information around the evidence."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            521,
+            676,
+            875,
+            778
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.4 Direct VDoc Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            866,
+            742,
+            879
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We compared three segments ranking (perplexity (flan-t5-large), minilm-l12, and Elser) and a base",
+        "bbox": [
+            507,
+            887,
+            882,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{9}$ https://www.scrolls-benchmark.com/ tasks  $^{10}$ https://huggingface.co/datasets/allenai/qasper",
+        "bbox": [
+            112,
+            866,
+            453,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "12938",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/39516c011c5149ce3acd8ea1d67ea488197ca86959a36c953fe9477ef0d7a4ac.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">VDoc ranker</td><td colspan=\"3\">Qasper-sliding-window</td><td colspan=\"3\">GoogleNQ-sliding-window</td><td colspan=\"3\">GoogleNQ-semantics</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td>prefix</td><td>43.12</td><td>64.13</td><td>81.94</td><td>84.60</td><td>91.17</td><td>94.03</td><td>84.60</td><td>91.17</td><td>94.03</td></tr><tr><td>flan-t5-large</td><td>69.91</td><td>89.86</td><td>88.89</td><td>69.42</td><td>81.77</td><td>84.70</td><td>72.10</td><td>84.90</td><td>89.55</td></tr><tr><td>minilm-l12</td><td>50.51</td><td>76.81</td><td>79.17</td><td>57.37</td><td>75.21</td><td>77.61</td><td>65.40</td><td>80.34</td><td>85.82</td></tr><tr><td>Elser</td><td>71.46</td><td>89.86</td><td>91.67</td><td>76.56</td><td>77.78</td><td>96.27</td><td>91.96</td><td>97.44</td><td>98.13</td></tr><tr><td># required VDocs</td><td>487</td><td>276</td><td>72</td><td>448</td><td>351</td><td>268</td><td>448</td><td>351</td><td>268</td></tr></table>",
+        "bbox": [
+            178,
+            80,
+            820,
+            195
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "table",
+        "img_path": "images/7ef2c35bc62c16fb316a5bc2a674ec292d3960b727bb6007170b16084203e786.jpg",
+        "table_caption": [
+            "Table 2: Direct VDoc evaluation. Comparing success (%) of VDoc rankers to capture the evidence in the VDoc. For Qasper we use sliding-window only. For GoogleNQ we use both sliding-window and semantic segmentation."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td rowspan=\"2\">DB</td><td rowspan=\"2\">VDoc ranker</td><td colspan=\"3\">flan-t5-xxl</td><td colspan=\"3\">llama-3-8b-instruct</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td rowspan=\"4\">Qasper</td><td>prefix</td><td>0.25</td><td>0.32</td><td>0.33</td><td>0.25</td><td>0.33</td><td>0.34</td></tr><tr><td>flan-t5-large</td><td>0.37 (0.35)</td><td>0.36 (0.35)</td><td>0.35 (0.35)</td><td>0.35 (0.34)</td><td>0.35 (0.34)</td><td>0.35 (0.33)</td></tr><tr><td>minilm-112</td><td>0.28 (0.28)</td><td>0.34 (0.33)</td><td>0.35 (0.35)</td><td>0.31 (0.28)</td><td>0.33 (0.33)</td><td>0.33 (0.32)</td></tr><tr><td>Elser</td><td>0.35 (0.34)</td><td>0.36 (0.34)</td><td>0.35 (0.35)</td><td>0.37 (0.37)</td><td>0.35 (0.35)</td><td>0.34 (0.33)</td></tr><tr><td rowspan=\"4\">Narrative</td><td>prefix</td><td>0.15</td><td>0.15</td><td>0.15</td><td>0.14</td><td>0.16</td><td>0.17</td></tr><tr><td>flan-t5-large</td><td>0.19 (0.18)</td><td>0.19 (0.20)</td><td>0.19 (0.19)</td><td>0.19 (0.20)</td><td>0.21 (0.21)</td><td>0.23 (0.22)</td></tr><tr><td>minilm-112</td><td>0.15 (0.14)</td><td>0.16 (0.16)</td><td>0.17 (0.17)</td><td>0.16 (0.16)</td><td>0.20 (0.18)</td><td>0.20 (0.20)</td></tr><tr><td>Elser</td><td>0.17 (0.17)</td><td>0.19 (0.18)</td><td>0.19 (0.18)</td><td>0.18 (0.19)</td><td>0.22 (0.21)</td><td>0.24 (0.22)</td></tr></table>",
+        "bbox": [
+            166,
+            246,
+            831,
+            397
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Table 3: End-to-end evaluation. Comparing F1 of the generation by flan-t5-xxl and llama-3-8b-instruct on Qasper and NarrativeQA. Each cell contains values for doc order and (rank order).",
+        "bbox": [
+            112,
+            406,
+            882,
+            436
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "line prefix (that truncates to the window-size), for varying window-size.",
+        "bbox": [
+            112,
+            460,
+            485,
+            491
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "To use Elser as a ranker, We pre-index all segments of all documents into an ES index. We add a document-id as a field to all segments of each document. To rank the segments of a given document, we query the index with the given question, and use the document-id as a filter, thus only segments of the given document are scored by Elser.",
+        "bbox": [
+            112,
+            492,
+            487,
+            605
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Results are summarized in Table 2. The last row (# required VDocs), measures the number of cases that the input document was larger than the window size and thus required VDoc activation. For example, when using window-size of 2400, for Qasper, 487 out of the 500 documents were larger than 2400 tokens, thus required VDoc. The rest were smaller than 2400 tokens, thus we could send the full document as is. Similarly, for window-size 4800, only 276 Qasper documents were larger than 4800 tokens, thus required VDoc activation. We can see that the number of required VDocs decreases as the window size increases, and thus the success to capture the evidence in the VDoc increases.",
+        "bbox": [
+            112,
+            605,
+            489,
+            845
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Segments ranking. We observe that on Qasper, all VDoc rankers outperform the naive method of taking the prefix, where Elser and perplexity (with flan-t5-large) achieve the best performance, with",
+        "bbox": [
+            112,
+            854,
+            489,
+            919
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Elser performing slightly better. On GoogleNQ, we see that with a sliding-window segmentation, the naive prefix is better than the three VDoc rankers (except for Elser on window-size=7200). This implies that the evidence in GoogleNQ, in most cases, is at the beginning of the document.",
+        "bbox": [
+            507,
+            460,
+            884,
+            556
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Document segmentation. To evaluate a more sophisticated segmentation, we utilized the HTML format of the Wikipedia pages used in GoogleNQ<sup>11</sup>. We cleansed the Wikipedia HTML to remove noise (e.g., headers, footers, sidebars, advertisements) and converted the HTML to semantic passages, identified by the document title and headings preceding the passage. When moving to semantic segmentation, we can see that Elser outperforms the prefix. For example, even on the smaller window size of 2400, it succeeded to capture the gold evidence in  $91.96\\%$  of the cases, compared to  $72.10\\%$  and  $84.60\\%$  with perplexity (flan-t5-large) and prefix respectively.",
+        "bbox": [
+            507,
+            569,
+            884,
+            795
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.5 End-to-End Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            809,
+            741,
+            824
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "We compared two LLM models with 8K window size: flan-t5-xxl and llama-3-8b-instruct. We use the prompts from ZeroScrolls. For llama3, we",
+        "bbox": [
+            507,
+            832,
+            882,
+            879
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "11Qasper comes as free text, so we could not apply semantic segmentation on it",
+        "bbox": [
+            507,
+            892,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "12939",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "modified the prompts a bit. The detailed prompts and an example are given in Appendix A.",
+        "bbox": [
+            112,
+            84,
+            485,
+            116
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Table 3 shows the results on Qasper and NarrativeQA, for varying context-window size (2400, 4800, 7200) and the different segment rankers.",
+        "bbox": [
+            112,
+            117,
+            487,
+            164
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As seen, all VDoc strategies outperform the naive prefix method both on NarrativeQA and on Qasper. For example on NarrativeQA, with context-window of 7200, llama-3-8b-instruct achieved an improvement of  $41\\%$  (0.24 with Elser, compared to 0.17 with prefix). On Qasper we see an improvement of  $48\\%$ , on context-window 2400, with llama-3-8b-instruct (0.37 with Elser compared to 0.25 with prefix).",
+        "bbox": [
+            112,
+            165,
+            487,
+            309
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Furthermore, we observe a correlation with the direct VDoc evaluation (Table 2). In both experiments, the best results are achieved for Elser and perplexity (flan-t5-large) rankers. A closer examination of the two datasets reveals that on Qasper, those two rankers remain quite stable across different window sizes. On NarrativeQA, we can see an improvement for all segment rankers, when increasing the window size, and all outperform the naive prefix by a large margin.",
+        "bbox": [
+            112,
+            310,
+            487,
+            470
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "The behavior of the two datasets supports our hypothesis that it is enough to give the LLM a smaller window and still get comparable results to those with a larger window. On Qasper, the 4800 window is a sweet spot, namely a good balance between capturing the evidence, and yet not adding too much irrelevant information to the LLM. This is because most of Qasper documents fit into the 7200 window. On NarrativeQA, the sweet spot is beyond 7200 since the documents are much larger.",
+        "bbox": [
+            112,
+            470,
+            487,
+            632
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "VDoc reconstruction. As shown in Table 3, ar- ranging the top-selected segments in their original order (Doc order) generally results in a better F1 score compared to ordering them by relevancy to the question (Rank order).",
+        "bbox": [
+            112,
+            640,
+            487,
+            720
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ZeroScrolls evaluation. We submitted flan-t5-xxl on the best VDoc configuration of Elser with window size 4800 to the official ZeroScrolls leader board. With VDoc, our flan-t5-xxl achieved F1 scores of 50.7, 30.6 on Qasper and NarrativeQA respectively, compared to F1 of 48.3 and 19.3 when used without the VDoc. This is an improvement of  $58\\%$  on NarrativeQA which has very large documents.",
+        "bbox": [
+            112,
+            728,
+            487,
+            873
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.6 The Effect on Inference Times",
+        "text_level": 1,
+        "bbox": [
+            507,
+            84,
+            793,
+            99
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we analyze the contribution of the VDoc on inference times. We used the Llama-3-1-70b-instruct model, which supports an input length of 128K tokens. We run it on the NarrativeQA 500 examples, used in all our experiments.",
+        "bbox": [
+            507,
+            105,
+            882,
+            185
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Figure 3 shows the running times and the F1 as a function of the context-window size. We used the llama-3-1-70b-instruct on IBM watsonx.ai $^{12}$  for inference. The times are for generation only. The F1 values are reported for VDoc that uses the best segment-ranking method (see Table 3 above), namely Elser. We can see that the best F1 is achieved with VDoc for context window size of 7200 tokens, and it starts degrading as we increase the context window size. Still we can see that running times for the 500 examples, increase from 28 min (3.36 sec/example) to 125 min with context-window of 64K (15 sec/example).",
+        "bbox": [
+            507,
+            185,
+            882,
+            395
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/dd927fbe8824ca8ae7439c058915e7816cc5d636ca33bba34ac12c3ddd831c68.jpg",
+        "image_caption": [
+            "Figure 3: Inference times and F1 of response generation, on NarrativeQA, as a function of context-window size."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            515,
+            407,
+            875,
+            548
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Conclusions and Future Work",
+        "text_level": 1,
+        "bbox": [
+            507,
+            615,
+            803,
+            631
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We described a method called VDoc, enabling Question Answering using LLM over long documents. The method is based on creating a virtual document that contains relevant information to the user query, and adheres to the LLM's context-window limitations. Our method does not require any modifications or fine-tuning a LLM.",
+        "bbox": [
+            507,
+            640,
+            882,
+            753
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "We compared different methods and showed experimentally, that it is better to aggressively shorten the document, even if the LLM can handle a larger context-window size, both in terms of quality of the results, and in terms of better run times.",
+        "bbox": [
+            507,
+            753,
+            882,
+            833
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_footnote",
+        "text": "12https://www.ibm.com/products/watsonx-ai",
+        "bbox": [
+            507,
+            891,
+            801,
+            916
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "12940",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            115,
+            84,
+            213,
+            98
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer.",
+            "Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation.",
+            "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context.",
+            "Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning.",
+            "Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness.",
+            "Gemini. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.",
+            "Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.",
+            "Albert Gu, Karan Goel, and Christopher Ré. 2022. Efficiently modeling long sequences with structured state spaces.",
+            "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.",
+            "Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models.",
+            "Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning.",
+            "Hao Liu, Matei Zaharia, and Pieter Abbeel. 2023a. Ring attention with blockwise transformers for near-infinite context.",
+            "Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023b. Lost in the middle: How language models use long contexts.",
+            "Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, Xuzheng He, Haowen Hou, Jiaju Lin, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartlomiej Koptyra, Hayden Lau, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Guangyu Song, Xiangru Tang, Bolun Wang, Johan S. Wind, Stanislaw Wozniak, Ruichong Zhang, Zhenyuan"
+        ],
+        "bbox": [
+            115,
+            105,
+            487,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Zhang, Qihang Zhao, Peng Zhou, Qinghua Zhou, Jian Zhu, and Rui-Jie Zhu. 2023. Rwkv: Reinventing rnns for the transformer era.",
+            "Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré. 2023. Hyena hierarchy: Towards larger convolutional language models.",
+            "Ofir Press, Noah A. Smith, and Mike Lewis. 2022. Train short, test long: Attention with linear biases enables input length extrapolation.",
+            "Guanghui Qin, Yukun Feng, and Benjamin Van Durme. 2023. The nlp task effectiveness of long-range transformers.",
+            "Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. Parallel context windows for large language models.",
+            "Uri Shaham, Maor Ivgi, Avia Efrat, Jonathan Berant, and Omer Levy. 2023. Zeroscrolls: A zero-shot benchmark for long text understanding.",
+            "Asaf Yehudai and Elron Bendel. 2024. When llms are unfit use fastfit: Fast and effective text classification with many classes.",
+            "Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontonon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2021. Big bird: Transformers for longer sequences."
+        ],
+        "bbox": [
+            510,
+            85,
+            882,
+            538
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "A Appendix - Detailed Prompts and Examples",
+        "text_level": 1,
+        "bbox": [
+            510,
+            551,
+            835,
+            583
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In all the end-to-end experiments, we used the ZeroScrolls prompts. For llama-3-8b-instruct, we found however that repeating the last instruction after the query, helps improve its results. Without that addition, llama3 sometimes tends to return very long answers, which hinders precision.",
+        "bbox": [
+            510,
+            594,
+            882,
+            688
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Tables 4 and 5 list the prompts used by llama3 for the Qasper and NarrativeQA respectively. Table 6 illustrates the above phenomena. We can see that without the suffix, llama3 tends to give very long answers. The F1 for the first answer is 0.67 (with recall=1 and precision=0.5), while the values for the second answer are 0.03, 1, 0.02 for F1, recall and precision, respectively.",
+        "bbox": [
+            510,
+            690,
+            882,
+            818
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "12941",
+        "bbox": [
+            477,
+            928,
+            522,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write \"unanswerable\". If the question is a yes/no question, answer \"yes\", \"no\", or \"unanswerable\".",
+        "bbox": [
+            132,
+            137,
+            473,
+            265
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Article:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            282,
+            196,
+            297
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "{input}",
+        "bbox": [
+            134,
+            300,
+            189,
+            315
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            332,
+            215,
+            346
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "{question} Answer the question as concisely as you can, using a single phrase or sentence if possible.",
+        "bbox": [
+            134,
+            348,
+            470,
+            395
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Answer:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            412,
+            203,
+            426
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 4: llama-3 prompt for Qasper. Note the added suffix after the question",
+        "bbox": [
+            132,
+            454,
+            468,
+            483
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "You are given a story, which can be either a novel or a movie script, and a question. Answer the question as concisely as you can, using a single phrase if possible.",
+        "bbox": [
+            134,
+            590,
+            473,
+            656
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Story:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            671,
+            186,
+            687
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "{input}",
+        "bbox": [
+            134,
+            689,
+            189,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Question:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            720,
+            213,
+            734
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "{question} Answer the question as concisely as you can, using a single phrase or sentence if possible.",
+        "bbox": [
+            134,
+            736,
+            470,
+            784
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Answer:",
+        "text_level": 1,
+        "bbox": [
+            134,
+            801,
+            203,
+            814
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: llama-3 prompt for NarrativeQA. Note the added suffix after the question",
+        "bbox": [
+            132,
+            843,
+            468,
+            872
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Question: What kind of injustices do the stories point out?",
+        "bbox": [
+            529,
+            252,
+            865,
+            282
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Gold answer: Social injustices.",
+        "bbox": [
+            529,
+            300,
+            769,
+            316
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Answer 1: Social and economic injustices",
+        "bbox": [
+            529,
+            332,
+            867,
+            362
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Answer 2: The stories point out the injustices of social class, wealth, and power. In \"The Adventure of the Abbey Grange\", the stepfather's abuse of power and wealth leads to the mistreatment of his stepchildren. In \"The Adventure of the Black Swan\", the lady's freedom is threatened by the five miles of country between her home and Winchester, highlighting the limitations of women's freedom in rural areas. In \"The Adventure of the Empty House\", Sherlock Holmes criticizes the lack of justice for the poor and ignorant, who are often victims of hidden wickedness and cruelty. The stories also highlight the class differences between the wealthy and the poor, with the wealthy often having more power and influence. Overall, the stories point out the injustices of a society where wealth, power, and social class can lead to exploitation and mistreatment of others.",
+        "bbox": [
+            527,
+            381,
+            868,
+            702
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 6: An example question. Answer 1 is with the added suffix, while Answer 2 is without.",
+        "bbox": [
+            527,
+            728,
+            863,
+            757
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "12942",
+        "bbox": [
+            477,
+            927,
+            524,
+            940
+        ],
+        "page_idx": 6
+    }
+]

2024/More Bang for your Context_ Virtual Documents for Question Answering over Long Documents/f22392c5-4a70-4295-a2be-84f110b6b3b3_model.json

@@ -0,0 +1,1589 @@
+[
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.157,
+                0.09,
+                0.843,
+                0.13
+            ],
+            "angle": 0,
+            "content": "More Bang for your Context: Virtual Documents for Question Answering over Long Documents"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.199,
+                0.143,
+                0.806,
+                0.175
+            ],
+            "angle": 0,
+            "content": "Yosi Mass, Boaz Carmeli, Asaf Yehudai, Assaf Toledo, Nathaniel Mills IBM Research AI"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.289,
+                0.178,
+                0.717,
+                0.226
+            ],
+            "angle": 0,
+            "content": "{yosimass,boazc}@il.ibm.com, {Asaf.Yehudai,assaf.toledo} @ibm.com, wnrm3@us.ibm.com"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.261,
+                0.253,
+                0.342,
+                0.269
+            ],
+            "angle": 0,
+            "content": "Abstract"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.142,
+                0.281,
+                0.462,
+                0.538
+            ],
+            "angle": 0,
+            "content": "We deal with the problem of Question Answering (QA) over a long document, which poses a challenge for modern Large Language Models (LLMs). Although LLMs can handle increasingly longer context windows, they struggle to effectively utilize the long content. To address this issue, we introduce the concept of a virtual document (VDoc). A VDoc is created by selecting chunks from the original document that are most likely to contain the information needed to answer the user's question, while ensuring they fit within the LLM's context window. We hypothesize that providing a short and focused VDoc to the LLM is more effective than filling the entire context window with less relevant information. Our experiments confirm this hypothesis and demonstrate that using VDocs improves results on the QA task."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.551,
+                0.262,
+                0.567
+            ],
+            "angle": 0,
+            "content": "1 Introduction"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.577,
+                0.49,
+                0.737
+            ],
+            "angle": 0,
+            "content": "Large language models (LLMs) such as OpenAI's ChatGPT<sup>1</sup> or Anthropic's Claude<sup>2</sup>, have demonstrated exceptional performance across a range of natural language processing (NLP) tasks and are capable of answering questions on various topics. Recent advancements have extended their context window to encompass tens of thousands or even millions of tokens (Gemini, 2024). This enables them to process long documents, or even books, and answer questions over them."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.739,
+                0.49,
+                0.849
+            ],
+            "angle": 0,
+            "content": "Despite this impressive progress, recent studies have shown that LLMs still struggle to utilize long contexts efficiently (Li et al., 2024; Levy et al., 2024). LLMs' performance often deteriorates when confronted with a long document, hindering their ability to accurately answer questions on such documents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.852,
+                0.489,
+                0.883
+            ],
+            "angle": 0,
+            "content": "To address this challenge, we propose a method called Virtual Document (VDoc). VDoc takes a"
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.252,
+                0.877,
+                0.373
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.509,
+                0.386,
+                0.886,
+                0.445
+            ],
+            "angle": 0,
+            "content": "Figure 1: VDoc architecture. First, a long document is segmented (a), then each segment is ranked by its relevance to the question (b), and finally a virtual document is reconstructed with the top-ranked segments (c)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.462,
+                0.885,
+                0.655
+            ],
+            "angle": 0,
+            "content": "long document, a question, and a specified context length, and generates a shorter version of the document, containing the most relevant parts for the given length. VDoc operates in three steps: 1) Document segmentation: break down the long document into smaller segments, 2) Segments ranking: rank the segments based on their relevance to the question, and 3) VDoc reconstruction: reassemble the segments into a coherent virtual document that contains the most relevant information to the question, and still fits the given context length. The overall VDoc process is illustrated in Figure 1."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.657,
+                0.886,
+                0.754
+            ],
+            "angle": 0,
+            "content": "We analyze experimentally, different approaches for each step. Our results show that VDoc enhances question-answering performance on various datasets, including ZeroScrolls (Shaham et al., 2023), a benchmark designed for tasks requiring a long context."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.757,
+                0.885,
+                0.886
+            ],
+            "angle": 0,
+            "content": "The paper's main contribution is the thorough analysis and experiments of the above three steps for QA over a long document. While the suggested solution is close to RAG (Retrieval Augmented Generation), it has some unique properties. For example, the ordering of the segments based on their sequence in the original document, that does not exist in RAG. Our code is available on git<sup>3</sup>"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.891,
+                0.293,
+                0.905
+            ],
+            "angle": 0,
+            "content": "<sup>1</sup>https://chat.openai.com/"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.137,
+                0.905,
+                0.414,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{2}\\)https://www.anthropic.com/news/claude-2-1"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.136,
+                0.891,
+                0.414,
+                0.919
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.531,
+                0.904,
+                0.787,
+                0.919
+            ],
+            "angle": 0,
+            "content": "<sup>3</sup>https://github.com/IBM/vdoc"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.477,
+                0.928,
+                0.528,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12936"
+        },
+        {
+            "type": "footer",
+            "bbox": [
+                0.211,
+                0.946,
+                0.789,
+                0.973
+            ],
+            "angle": 0,
+            "content": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 12936-12942 November 12-16, 2024 ©2024 Association for Computational Linguistics"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.084,
+                0.271,
+                0.099
+            ],
+            "angle": 0,
+            "content": "2 Related Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.111,
+                0.49,
+                0.336
+            ],
+            "angle": 0,
+            "content": "Many previous works focus on designing LLMs that can support long context. Some works modify the transformer attention mechanism (Beltagy et al., 2020; Zaheer et al., 2021; Dai et al., 2019; Ratner et al., 2023) or make it more efficient (Dao et al., 2022; Dao, 2023; Liu et al., 2023a; Gemini, 2024). Others have focused on modifying the positional embeddings (Press et al., 2022; Chen et al., 2023). Another line of research attempts to replace transformers, which have quadratic sequence length complexity, with new architectural designs like convolution and linear RNNs, e.g., RWKV (Peng et al., 2023), S4 (Gu et al., 2022), Hyena (Poli et al., 2023), or Mamba (Gu and Dao, 2023)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.117,
+                0.337,
+                0.49,
+                0.546
+            ],
+            "angle": 0,
+            "content": "While those works were able to extend the context length of LLMs, subsequent research has found that long-context transformers exhibit a recency bias and do not effectively utilize long-range context (Qin et al., 2023; Liu et al., 2023b; Yehudai and Bendel, 2024; Li et al., 2024). Specifically, these models have been found to struggle with ignoring irrelevant information (Levy et al., 2024). Moreover, benchmarks for long context reveal that there is still significant room for improvement compared to human performance, and for some tasks, models even struggle to surpass the naive baseline (Shaham et al., 2023)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.115,
+                0.559,
+                0.219,
+                0.573
+            ],
+            "angle": 0,
+            "content": "3 Method"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.585,
+                0.49,
+                0.761
+            ],
+            "angle": 0,
+            "content": "We deal with the problem of using an LLM to answer questions over a long document. Our method, termed VDoc, creates a shorter version of the document, which is then used as input to the LLM, instead of the entire document, to generate an answer. Creating the VDoc is done in three steps. i) Document segmentation into manageable segments, ii) Segments ranking by their potential to contain the answer to the user question, and iii) VDoc reconstruction. The detailed steps are described below."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.772,
+                0.49,
+                0.853
+            ],
+            "angle": 0,
+            "content": "Document segmentation. We try two methods. i) sliding-window on sentence boundaries using nltk\\(^4\\) ii) context-aware (using HTML structure when available). We define a segment-size parameter that controls the max segment length."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.114,
+                0.863,
+                0.489,
+                0.895
+            ],
+            "angle": 0,
+            "content": "Segments ranking. We experiment with three families of language models (LM). The first is"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.085,
+                0.885,
+                0.149
+            ],
+            "angle": 0,
+            "content": "encoder-only models, where we rank segments by the similarity of their representation to the representation of the user question. An example of such LM is MiniLM<sup>5</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.15,
+                0.885,
+                0.31
+            ],
+            "angle": 0,
+            "content": "The second, is encoder-decoder or decoder-only models, where we rank segments by the inverse perplexity of the user question, given each of the segments. Specifically, we give each segment as input to the model with the instruction: \"Generate a conversation between a user and an agent based on the given content:\". We then run a forward loop and compute the Cross-Entropy loss for the question. The segments are ranked by \\( r = (-1) \\times loss() \\) (higher is better). An example is flan-t5<sup>6</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.311,
+                0.884,
+                0.39
+            ],
+            "angle": 0,
+            "content": "The third, is a sparse representation, Elser<sup>7</sup> which expands and indexes documents into ElasticSearch<sup>8</sup> index, with semantic terms that were learned to co-occur frequently with terms in each document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.392,
+                0.884,
+                0.553
+            ],
+            "angle": 0,
+            "content": "We note that among the three segment ranking methods, the first (comparing dense representations) and the last (Sparse representation) can be pre-computed on the given set of documents. The second method on the other hand (using perplexity), requires running a forward loop of the model for each segment and question, and thus can not be pre-computed, because it depends on the question. Thus the perplexity method can be quite slow for very long documents."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.563,
+                0.884,
+                0.645
+            ],
+            "angle": 0,
+            "content": "VDoc reconstruction. We take the top ranked segments, that still fit into the given window size and evaluate two methods to concatenate them: i) by the original order in the document (Doc order) and ii) by relevancy to the question (Rank order)."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.656,
+                0.657,
+                0.673
+            ],
+            "angle": 0,
+            "content": "4 Experiments"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.683,
+                0.757,
+                0.699
+            ],
+            "angle": 0,
+            "content": "4.1 Setup of the Experiments"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.704,
+                0.885,
+                0.833
+            ],
+            "angle": 0,
+            "content": "We evaluated the VDoc performance in two scenarios: direct and end-to-end. For the direct scenario we assume a benchmark of triplets \\(\\{(Q, D, E)\\}\\) where \\(Q\\) is a question, \\(D\\) is a long document and \\(E\\) is an evidence i.e., a span in \\(D\\) that contains the answer. We define the VDoc task as follows: Given a context-window size, a question \\(Q\\), and a long document \\(D\\), create a virtual document that"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.842,
+                0.847,
+                0.867
+            ],
+            "angle": 0,
+            "content": "\\(^{5}\\)https://huggingface.co/sentence-transformers/all-MiniLM-L12-v2"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.534,
+                0.868,
+                0.8,
+                0.881
+            ],
+            "angle": 0,
+            "content": "\\(^{6}\\)https://huggingface.co/google/flan-t5-large"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.511,
+                0.881,
+                0.796,
+                0.905
+            ],
+            "angle": 0,
+            "content": "7https://www/elastic.co/guide/en/machinelearning/current/ml-nlp-elser.html"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.534,
+                0.906,
+                0.749,
+                0.918
+            ],
+            "angle": 0,
+            "content": "<sup>8</sup>https://www/elastic.co/"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.509,
+                0.842,
+                0.847,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.136,
+                0.904,
+                0.337,
+                0.919
+            ],
+            "angle": 0,
+            "content": "4https://www.nltk.org/"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12937"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.488,
+                0.148
+            ],
+            "angle": 0,
+            "content": "fits within the context-window size and includes E. We measure the success of the VDoc to contain the evidence (we counted \\(90\\%\\) containment as a success, to tolerate for segmentation issues)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.15,
+                0.488,
+                0.277
+            ],
+            "angle": 0,
+            "content": "For the end-to-end scenario, we assume a benchmark of triplets \\(\\{(\\mathrm{Q},\\mathrm{D},\\mathrm{R})\\}\\) where \\(\\mathbf{Q}\\) is a question, \\(\\mathrm{D}\\) is a long document that contains the answer, and \\(\\mathbf{R}\\) is a gold response. We then assess LLM's response against the gold, \\(\\mathbf{R}\\), when prompting using various VDoc strategies. We evaluate results using F1 (the harmonic mean of Recall and Precision), similar to the scrolls (Shaham et al., 2023) paper."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.279,
+                0.488,
+                0.358
+            ],
+            "angle": 0,
+            "content": "We experiment with two LLMs, each with 8K window size, thus we limit the VDoc to sizes of up to 7200 tokens to allow space for the prompt, the question and generated answer. In all experiments, we used segment-size of 512 tokens."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.114,
+                0.371,
+                0.227,
+                0.384
+            ],
+            "angle": 0,
+            "content": "4.2 Datasets"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.392,
+                0.488,
+                0.6
+            ],
+            "angle": 0,
+            "content": "For the direct experiment, we utilize two datasets containing long documents, questions and evidence. The first is GoogleNQ (Kwiatkowski et al., 2019). This dataset contains questions from Wikipedia. Each question is labeled with the Wikipedia page that contains the answer, and a span in the document that contains the evidence. The second dataset is taken from Scrolls (Shaham et al., \\(2023)^{9}\\) Scrolls contains several datasets with questions over long documents. We used Qasper which is question answering over research papers. We added an evidence for each question, from the original Qasper dataset<sup>10</sup>."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.602,
+                0.488,
+                0.697
+            ],
+            "angle": 0,
+            "content": "For the end-to-end experiment we utilized the Qasper dataset described above and the NarrativeQA dataset from Scrolls. NarrativeQA contains question-answer pairs about entire books and movie scripts. Both datasets contain (Q, D, R) triplets."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.698,
+                0.488,
+                0.762
+            ],
+            "angle": 0,
+            "content": "We picked a random subset of 500 examples from the validation sets of Qasper and NarrativeQA. Likewise, for GoogleNQ, we picked 500 examples where the long answer (evidence) is a section."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.764,
+                0.488,
+                0.859
+            ],
+            "angle": 0,
+            "content": "Note that all our experiments were done on the validation sets of Qasper and NarrativeQA from Scrolls. We used the validation set and not the test set, since ground truths is available only on the validation set. We selected a random subset of 500 examples, inspired by the size of the test sets in"
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.545,
+                0.082,
+                0.849,
+                0.148
+            ],
+            "angle": 0,
+            "content": "<table><tr><td>Dataset</td><td>Avg # tokens</td><td>Max # tokens</td></tr><tr><td>GoogleNQ</td><td>12,337</td><td>273,087</td></tr><tr><td>Qasper</td><td>5,484</td><td>22,052</td></tr><tr><td>NarrativeQA</td><td>88,073</td><td>516,045</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.508,
+                0.156,
+                0.882,
+                0.186
+            ],
+            "angle": 0,
+            "content": "Table 1: Datasets used for evaluation. Number of tokens (#) is calculated by the flan-t5-small tokenizer."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.212,
+                0.882,
+                0.275
+            ],
+            "angle": 0,
+            "content": "ZeroScrolls. We tried the Qasper experiments on the full validation set and got very similar results, thus we performed the experiments on the 500 data points."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.277,
+                0.882,
+                0.372
+            ],
+            "angle": 0,
+            "content": "The datasets and their statistics are described in Table 1. The numbers are for the random subset of 500 examples. Note that in all our experiments, we treat each document separately, thus we report statistics only for the subset that was used in the experiments."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.385,
+                0.729,
+                0.4
+            ],
+            "angle": 0,
+            "content": "4.3 Optimal Window Size"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.507,
+                0.406,
+                0.884,
+                0.501
+            ],
+            "angle": 0,
+            "content": "To confirm that a smaller window yields better generation results, we conducted a controlled experiment, using the evidence as a pivot. We expanded the content up to a specified window size, measured in tokens, while maintaining full sentences using the nltk tool."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.503,
+                0.884,
+                0.662
+            ],
+            "angle": 0,
+            "content": "We experiment with expanding the content on both sides of the evidence, as well as before and after it. All three methods yielded similar results. Figure 2 reports results when expanding the window in both sides of the evidence. We conducted the experiments using two generative LLM: flant5-xxl and llama-3-8b-instruct. As observed, both models achieve the highest F1 score when provided only with the evidence, and performance degrades as data is added."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.522,
+                0.677,
+                0.877,
+                0.779
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.793,
+                0.882,
+                0.835
+            ],
+            "angle": 0,
+            "content": "Figure 2: Comparing performances (F1) of response generation on Qasper when adding non-relevant information around the evidence."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.867,
+                0.743,
+                0.88
+            ],
+            "angle": 0,
+            "content": "4.4 Direct VDoc Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.888,
+                0.884,
+                0.919
+            ],
+            "angle": 0,
+            "content": "We compared three segments ranking (perplexity (flan-t5-large), minilm-l12, and Elser) and a base"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.114,
+                0.867,
+                0.454,
+                0.919
+            ],
+            "angle": 0,
+            "content": "\\(^{9}\\)https://www.scrolls-benchmark.com/ tasks \\(^{10}\\)https://huggingface.co/datasets/allenai/qasper"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12938"
+        }
+    ],
+    [
+        {
+            "type": "table",
+            "bbox": [
+                0.179,
+                0.082,
+                0.821,
+                0.196
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">VDoc ranker</td><td colspan=\"3\">Qasper-sliding-window</td><td colspan=\"3\">GoogleNQ-sliding-window</td><td colspan=\"3\">GoogleNQ-semantics</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td>prefix</td><td>43.12</td><td>64.13</td><td>81.94</td><td>84.60</td><td>91.17</td><td>94.03</td><td>84.60</td><td>91.17</td><td>94.03</td></tr><tr><td>flan-t5-large</td><td>69.91</td><td>89.86</td><td>88.89</td><td>69.42</td><td>81.77</td><td>84.70</td><td>72.10</td><td>84.90</td><td>89.55</td></tr><tr><td>minilm-l12</td><td>50.51</td><td>76.81</td><td>79.17</td><td>57.37</td><td>75.21</td><td>77.61</td><td>65.40</td><td>80.34</td><td>85.82</td></tr><tr><td>Elser</td><td>71.46</td><td>89.86</td><td>91.67</td><td>76.56</td><td>77.78</td><td>96.27</td><td>91.96</td><td>97.44</td><td>98.13</td></tr><tr><td># required VDocs</td><td>487</td><td>276</td><td>72</td><td>448</td><td>351</td><td>268</td><td>448</td><td>351</td><td>268</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.206,
+                0.884,
+                0.237
+            ],
+            "angle": 0,
+            "content": "Table 2: Direct VDoc evaluation. Comparing success (%) of VDoc rankers to capture the evidence in the VDoc. For Qasper we use sliding-window only. For GoogleNQ we use both sliding-window and semantic segmentation."
+        },
+        {
+            "type": "table",
+            "bbox": [
+                0.167,
+                0.247,
+                0.832,
+                0.398
+            ],
+            "angle": 0,
+            "content": "<table><tr><td rowspan=\"2\">DB</td><td rowspan=\"2\">VDoc ranker</td><td colspan=\"3\">flan-t5-xxl</td><td colspan=\"3\">llama-3-8b-instruct</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td rowspan=\"4\">Qasper</td><td>prefix</td><td>0.25</td><td>0.32</td><td>0.33</td><td>0.25</td><td>0.33</td><td>0.34</td></tr><tr><td>flan-t5-large</td><td>0.37 (0.35)</td><td>0.36 (0.35)</td><td>0.35 (0.35)</td><td>0.35 (0.34)</td><td>0.35 (0.34)</td><td>0.35 (0.33)</td></tr><tr><td>minilm-112</td><td>0.28 (0.28)</td><td>0.34 (0.33)</td><td>0.35 (0.35)</td><td>0.31 (0.28)</td><td>0.33 (0.33)</td><td>0.33 (0.32)</td></tr><tr><td>Elser</td><td>0.35 (0.34)</td><td>0.36 (0.34)</td><td>0.35 (0.35)</td><td>0.37 (0.37)</td><td>0.35 (0.35)</td><td>0.34 (0.33)</td></tr><tr><td rowspan=\"4\">Narrative</td><td>prefix</td><td>0.15</td><td>0.15</td><td>0.15</td><td>0.14</td><td>0.16</td><td>0.17</td></tr><tr><td>flan-t5-large</td><td>0.19 (0.18)</td><td>0.19 (0.20)</td><td>0.19 (0.19)</td><td>0.19 (0.20)</td><td>0.21 (0.21)</td><td>0.23 (0.22)</td></tr><tr><td>minilm-112</td><td>0.15 (0.14)</td><td>0.16 (0.16)</td><td>0.17 (0.17)</td><td>0.16 (0.16)</td><td>0.20 (0.18)</td><td>0.20 (0.20)</td></tr><tr><td>Elser</td><td>0.17 (0.17)</td><td>0.19 (0.18)</td><td>0.19 (0.18)</td><td>0.18 (0.19)</td><td>0.22 (0.21)</td><td>0.24 (0.22)</td></tr></table>"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.113,
+                0.407,
+                0.884,
+                0.437
+            ],
+            "angle": 0,
+            "content": "Table 3: End-to-end evaluation. Comparing F1 of the generation by flan-t5-xxl and llama-3-8b-instruct on Qasper and NarrativeQA. Each cell contains values for doc order and (rank order)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.461,
+                0.486,
+                0.492
+            ],
+            "angle": 0,
+            "content": "line prefix (that truncates to the window-size), for varying window-size."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.493,
+                0.489,
+                0.606
+            ],
+            "angle": 0,
+            "content": "To use Elser as a ranker, We pre-index all segments of all documents into an ES index. We add a document-id as a field to all segments of each document. To rank the segments of a given document, we query the index with the given question, and use the document-id as a filter, thus only segments of the given document are scored by Elser."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.606,
+                0.49,
+                0.846
+            ],
+            "angle": 0,
+            "content": "Results are summarized in Table 2. The last row (# required VDocs), measures the number of cases that the input document was larger than the window size and thus required VDoc activation. For example, when using window-size of 2400, for Qasper, 487 out of the 500 documents were larger than 2400 tokens, thus required VDoc. The rest were smaller than 2400 tokens, thus we could send the full document as is. Similarly, for window-size 4800, only 276 Qasper documents were larger than 4800 tokens, thus required VDoc activation. We can see that the number of required VDocs decreases as the window size increases, and thus the success to capture the evidence in the VDoc increases."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.856,
+                0.49,
+                0.92
+            ],
+            "angle": 0,
+            "content": "Segments ranking. We observe that on Qasper, all VDoc rankers outperform the naive method of taking the prefix, where Elser and perplexity (with flan-t5-large) achieve the best performance, with"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.461,
+                0.885,
+                0.557
+            ],
+            "angle": 0,
+            "content": "Elser performing slightly better. On GoogleNQ, we see that with a sliding-window segmentation, the naive prefix is better than the three VDoc rankers (except for Elser on window-size=7200). This implies that the evidence in GoogleNQ, in most cases, is at the beginning of the document."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.57,
+                0.885,
+                0.796
+            ],
+            "angle": 0,
+            "content": "Document segmentation. To evaluate a more sophisticated segmentation, we utilized the HTML format of the Wikipedia pages used in GoogleNQ<sup>11</sup>. We cleansed the Wikipedia HTML to remove noise (e.g., headers, footers, sidebars, advertisements) and converted the HTML to semantic passages, identified by the document title and headings preceding the passage. When moving to semantic segmentation, we can see that Elser outperforms the prefix. For example, even on the smaller window size of 2400, it succeeded to capture the gold evidence in \\(91.96\\%\\) of the cases, compared to \\(72.10\\%\\) and \\(84.60\\%\\) with perplexity (flan-t5-large) and prefix respectively."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.81,
+                0.742,
+                0.825
+            ],
+            "angle": 0,
+            "content": "4.5 End-to-End Evaluation"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.833,
+                0.884,
+                0.881
+            ],
+            "angle": 0,
+            "content": "We compared two LLM models with 8K window size: flan-t5-xxl and llama-3-8b-instruct. We use the prompts from ZeroScrolls. For llama3, we"
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.893,
+                0.883,
+                0.919
+            ],
+            "angle": 0,
+            "content": "11Qasper comes as free text, so we could not apply semantic segmentation on it"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12939"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.085,
+                0.486,
+                0.117
+            ],
+            "angle": 0,
+            "content": "modified the prompts a bit. The detailed prompts and an example are given in Appendix A."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.118,
+                0.488,
+                0.165
+            ],
+            "angle": 0,
+            "content": "Table 3 shows the results on Qasper and NarrativeQA, for varying context-window size (2400, 4800, 7200) and the different segment rankers."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.166,
+                0.488,
+                0.31
+            ],
+            "angle": 0,
+            "content": "As seen, all VDoc strategies outperform the naive prefix method both on NarrativeQA and on Qasper. For example on NarrativeQA, with context-window of 7200, llama-3-8b-instruct achieved an improvement of \\(41\\%\\) (0.24 with Elser, compared to 0.17 with prefix). On Qasper we see an improvement of \\(48\\%\\), on context-window 2400, with llama-3-8b-instruct (0.37 with Elser compared to 0.25 with prefix)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.311,
+                0.488,
+                0.472
+            ],
+            "angle": 0,
+            "content": "Furthermore, we observe a correlation with the direct VDoc evaluation (Table 2). In both experiments, the best results are achieved for Elser and perplexity (flan-t5-large) rankers. A closer examination of the two datasets reveals that on Qasper, those two rankers remain quite stable across different window sizes. On NarrativeQA, we can see an improvement for all segment rankers, when increasing the window size, and all outperform the naive prefix by a large margin."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.472,
+                0.488,
+                0.633
+            ],
+            "angle": 0,
+            "content": "The behavior of the two datasets supports our hypothesis that it is enough to give the LLM a smaller window and still get comparable results to those with a larger window. On Qasper, the 4800 window is a sweet spot, namely a good balance between capturing the evidence, and yet not adding too much irrelevant information to the LLM. This is because most of Qasper documents fit into the 7200 window. On NarrativeQA, the sweet spot is beyond 7200 since the documents are much larger."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.641,
+                0.488,
+                0.721
+            ],
+            "angle": 0,
+            "content": "VDoc reconstruction. As shown in Table 3, ar- ranging the top-selected segments in their original order (Doc order) generally results in a better F1 score compared to ordering them by relevancy to the question (Rank order)."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.113,
+                0.73,
+                0.488,
+                0.874
+            ],
+            "angle": 0,
+            "content": "ZeroScrolls evaluation. We submitted flan-t5-xxl on the best VDoc configuration of Elser with window size 4800 to the official ZeroScrolls leader board. With VDoc, our flan-t5-xxl achieved F1 scores of 50.7, 30.6 on Qasper and NarrativeQA respectively, compared to F1 of 48.3 and 19.3 when used without the VDoc. This is an improvement of \\(58\\%\\) on NarrativeQA which has very large documents."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.085,
+                0.794,
+                0.1
+            ],
+            "angle": 0,
+            "content": "4.6 The Effect on Inference Times"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.106,
+                0.884,
+                0.186
+            ],
+            "angle": 0,
+            "content": "In this section, we analyze the contribution of the VDoc on inference times. We used the Llama-3-1-70b-instruct model, which supports an input length of 128K tokens. We run it on the NarrativeQA 500 examples, used in all our experiments."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.186,
+                0.884,
+                0.396
+            ],
+            "angle": 0,
+            "content": "Figure 3 shows the running times and the F1 as a function of the context-window size. We used the llama-3-1-70b-instruct on IBM watsonx.ai\\(^{12}\\) for inference. The times are for generation only. The F1 values are reported for VDoc that uses the best segment-ranking method (see Table 3 above), namely Elser. We can see that the best F1 is achieved with VDoc for context window size of 7200 tokens, and it starts degrading as we increase the context window size. Still we can see that running times for the 500 examples, increase from 28 min (3.36 sec/example) to 125 min with context-window of 64K (15 sec/example)."
+        },
+        {
+            "type": "image",
+            "bbox": [
+                0.517,
+                0.409,
+                0.877,
+                0.549
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "image_caption",
+            "bbox": [
+                0.508,
+                0.56,
+                0.883,
+                0.589
+            ],
+            "angle": 0,
+            "content": "Figure 3: Inference times and F1 of response generation, on NarrativeQA, as a function of context-window size."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.509,
+                0.616,
+                0.804,
+                0.632
+            ],
+            "angle": 0,
+            "content": "5 Conclusions and Future Work"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.642,
+                0.883,
+                0.755
+            ],
+            "angle": 0,
+            "content": "We described a method called VDoc, enabling Question Answering using LLM over long documents. The method is based on creating a virtual document that contains relevant information to the user query, and adheres to the LLM's context-window limitations. Our method does not require any modifications or fine-tuning a LLM."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.508,
+                0.755,
+                0.884,
+                0.834
+            ],
+            "angle": 0,
+            "content": "We compared different methods and showed experimentally, that it is better to aggressively shorten the document, even if the LLM can handle a larger context-window size, both in terms of quality of the results, and in terms of better run times."
+        },
+        {
+            "type": "page_footnote",
+            "bbox": [
+                0.509,
+                0.892,
+                0.803,
+                0.917
+            ],
+            "angle": 0,
+            "content": "12https://www.ibm.com/products/watsonx-ai"
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12940"
+        }
+    ],
+    [
+        {
+            "type": "title",
+            "bbox": [
+                0.116,
+                0.085,
+                0.214,
+                0.099
+            ],
+            "angle": 0,
+            "content": "References"
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.107,
+                0.489,
+                0.134
+            ],
+            "angle": 0,
+            "content": "Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.143,
+                0.487,
+                0.184
+            ],
+            "angle": 0,
+            "content": "Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.192,
+                0.487,
+                0.245
+            ],
+            "angle": 0,
+            "content": "Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.255,
+                0.486,
+                0.282
+            ],
+            "angle": 0,
+            "content": "Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.291,
+                0.487,
+                0.331
+            ],
+            "angle": 0,
+            "content": "Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.34,
+                0.487,
+                0.367
+            ],
+            "angle": 0,
+            "content": "Gemini. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.377,
+                0.486,
+                0.404
+            ],
+            "angle": 0,
+            "content": "Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.413,
+                0.487,
+                0.453
+            ],
+            "angle": 0,
+            "content": "Albert Gu, Karan Goel, and Christopher Ré. 2022. Efficiently modeling long sequences with structured state spaces."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.462,
+                0.487,
+                0.581
+            ],
+            "angle": 0,
+            "content": "Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.59,
+                0.487,
+                0.63
+            ],
+            "angle": 0,
+            "content": "Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.64,
+                0.486,
+                0.68
+            ],
+            "angle": 0,
+            "content": "Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.688,
+                0.487,
+                0.728
+            ],
+            "angle": 0,
+            "content": "Hao Liu, Matei Zaharia, and Pieter Abbeel. 2023a. Ring attention with blockwise transformers for near-infinite context."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.738,
+                0.487,
+                0.791
+            ],
+            "angle": 0,
+            "content": "Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023b. Lost in the middle: How language models use long contexts."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.117,
+                0.801,
+                0.487,
+                0.918
+            ],
+            "angle": 0,
+            "content": "Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, Xuzheng He, Haowen Hou, Jiaju Lin, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartlomiej Koptyra, Hayden Lau, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Guangyu Song, Xiangru Tang, Bolun Wang, Johan S. Wind, Stanislaw Wozniak, Ruichong Zhang, Zhenyuan"
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.117,
+                0.107,
+                0.489,
+                0.918
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.529,
+                0.086,
+                0.882,
+                0.125
+            ],
+            "angle": 0,
+            "content": "Zhang, Qihang Zhao, Peng Zhou, Qinghua Zhou, Jian Zhu, and Rui-Jie Zhu. 2023. Rwkv: Reinventing rnns for the transformer era."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.136,
+                0.883,
+                0.189
+            ],
+            "angle": 0,
+            "content": "Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré. 2023. Hyena hierarchy: Towards larger convolutional language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.198,
+                0.882,
+                0.238
+            ],
+            "angle": 0,
+            "content": "Ofir Press, Noah A. Smith, and Mike Lewis. 2022. Train short, test long: Attention with linear biases enables input length extrapolation."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.248,
+                0.883,
+                0.287
+            ],
+            "angle": 0,
+            "content": "Guanghui Qin, Yukun Feng, and Benjamin Van Durme. 2023. The nlp task effectiveness of long-range transformers."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.298,
+                0.883,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. Parallel context windows for large language models."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.374,
+                0.883,
+                0.414
+            ],
+            "angle": 0,
+            "content": "Uri Shaham, Maor Ivgi, Avia Efrat, Jonathan Berant, and Omer Levy. 2023. Zeroscrolls: A zero-shot benchmark for long text understanding."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.423,
+                0.882,
+                0.463
+            ],
+            "angle": 0,
+            "content": "Asaf Yehudai and Elron Bendel. 2024. When llms are unfit use fastfit: Fast and effective text classification with many classes."
+        },
+        {
+            "type": "ref_text",
+            "bbox": [
+                0.512,
+                0.473,
+                0.883,
+                0.539
+            ],
+            "angle": 0,
+            "content": "Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontonon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2021. Big bird: Transformers for longer sequences."
+        },
+        {
+            "type": "list",
+            "bbox": [
+                0.512,
+                0.086,
+                0.883,
+                0.539
+            ],
+            "angle": 0,
+            "content": null
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.512,
+                0.552,
+                0.836,
+                0.585
+            ],
+            "angle": 0,
+            "content": "A Appendix - Detailed Prompts and Examples"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.595,
+                0.883,
+                0.689
+            ],
+            "angle": 0,
+            "content": "In all the end-to-end experiments, we used the ZeroScrolls prompts. For llama-3-8b-instruct, we found however that repeating the last instruction after the query, helps improve its results. Without that addition, llama3 sometimes tends to return very long answers, which hinders precision."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.512,
+                0.691,
+                0.883,
+                0.819
+            ],
+            "angle": 0,
+            "content": "Tables 4 and 5 list the prompts used by llama3 for the Qasper and NarrativeQA respectively. Table 6 illustrates the above phenomena. We can see that without the suffix, llama3 tends to give very long answers. The F1 for the first answer is 0.67 (with recall=1 and precision=0.5), while the values for the second answer are 0.03, 1, 0.02 for F1, recall and precision, respectively."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.929,
+                0.524,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12941"
+        }
+    ],
+    [
+        {
+            "type": "text",
+            "bbox": [
+                0.134,
+                0.139,
+                0.474,
+                0.266
+            ],
+            "angle": 0,
+            "content": "You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write \"unanswerable\". If the question is a yes/no question, answer \"yes\", \"no\", or \"unanswerable\"."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.135,
+                0.284,
+                0.198,
+                0.298
+            ],
+            "angle": 0,
+            "content": "Article:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.136,
+                0.301,
+                0.191,
+                0.316
+            ],
+            "angle": 0,
+            "content": "{input}"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.135,
+                0.333,
+                0.216,
+                0.347
+            ],
+            "angle": 0,
+            "content": "Question:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.135,
+                0.349,
+                0.471,
+                0.397
+            ],
+            "angle": 0,
+            "content": "{question} Answer the question as concisely as you can, using a single phrase or sentence if possible."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.136,
+                0.413,
+                0.204,
+                0.427
+            ],
+            "angle": 0,
+            "content": "Answer:"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.133,
+                0.455,
+                0.469,
+                0.484
+            ],
+            "angle": 0,
+            "content": "Table 4: llama-3 prompt for Qasper. Note the added suffix after the question"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.135,
+                0.592,
+                0.474,
+                0.657
+            ],
+            "angle": 0,
+            "content": "You are given a story, which can be either a novel or a movie script, and a question. Answer the question as concisely as you can, using a single phrase if possible."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.135,
+                0.673,
+                0.188,
+                0.688
+            ],
+            "angle": 0,
+            "content": "Story:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.136,
+                0.69,
+                0.19,
+                0.705
+            ],
+            "angle": 0,
+            "content": "{input}"
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.135,
+                0.721,
+                0.214,
+                0.735
+            ],
+            "angle": 0,
+            "content": "Question:"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.135,
+                0.737,
+                0.471,
+                0.785
+            ],
+            "angle": 0,
+            "content": "{question} Answer the question as concisely as you can, using a single phrase or sentence if possible."
+        },
+        {
+            "type": "title",
+            "bbox": [
+                0.136,
+                0.802,
+                0.204,
+                0.815
+            ],
+            "angle": 0,
+            "content": "Answer:"
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.133,
+                0.844,
+                0.469,
+                0.873
+            ],
+            "angle": 0,
+            "content": "Table 5: llama-3 prompt for NarrativeQA. Note the added suffix after the question"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.53,
+                0.253,
+                0.866,
+                0.284
+            ],
+            "angle": 0,
+            "content": "Question: What kind of injustices do the stories point out?"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.531,
+                0.301,
+                0.77,
+                0.317
+            ],
+            "angle": 0,
+            "content": "Gold answer: Social injustices."
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.53,
+                0.334,
+                0.868,
+                0.363
+            ],
+            "angle": 0,
+            "content": "Answer 1: Social and economic injustices"
+        },
+        {
+            "type": "text",
+            "bbox": [
+                0.529,
+                0.382,
+                0.869,
+                0.703
+            ],
+            "angle": 0,
+            "content": "Answer 2: The stories point out the injustices of social class, wealth, and power. In \"The Adventure of the Abbey Grange\", the stepfather's abuse of power and wealth leads to the mistreatment of his stepchildren. In \"The Adventure of the Black Swan\", the lady's freedom is threatened by the five miles of country between her home and Winchester, highlighting the limitations of women's freedom in rural areas. In \"The Adventure of the Empty House\", Sherlock Holmes criticizes the lack of justice for the poor and ignorant, who are often victims of hidden wickedness and cruelty. The stories also highlight the class differences between the wealthy and the poor, with the wealthy often having more power and influence. Overall, the stories point out the injustices of a society where wealth, power, and social class can lead to exploitation and mistreatment of others."
+        },
+        {
+            "type": "table_caption",
+            "bbox": [
+                0.528,
+                0.73,
+                0.865,
+                0.758
+            ],
+            "angle": 0,
+            "content": "Table 6: An example question. Answer 1 is with the added suffix, while Answer 2 is without."
+        },
+        {
+            "type": "page_number",
+            "bbox": [
+                0.478,
+                0.928,
+                0.526,
+                0.941
+            ],
+            "angle": 0,
+            "content": "12942"
+        }
+    ]
+]

2024/More Bang for your Context_ Virtual Documents for Question Answering over Long Documents/f22392c5-4a70-4295-a2be-84f110b6b3b3_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0f2350e854e84caea02564d5ef29c4eed437fb7464d245bab88c240672b6f42
+size 439293

2024/More Bang for your Context_ Virtual Documents for Question Answering over Long Documents/full.md

@@ -0,0 +1,216 @@
+# More Bang for your Context: Virtual Documents for Question Answering over Long Documents
+
+Yosi Mass, Boaz Carmeli, Asaf Yehudai, Assaf Toledo, Nathaniel Mills IBM Research AI
+
+{yosimass,boazc}@il.ibm.com, {Asaf.Yehudai,assaf.toledo} @ibm.com, wnrm3@us.ibm.com
+
+# Abstract
+
+We deal with the problem of Question Answering (QA) over a long document, which poses a challenge for modern Large Language Models (LLMs). Although LLMs can handle increasingly longer context windows, they struggle to effectively utilize the long content. To address this issue, we introduce the concept of a virtual document (VDoc). A VDoc is created by selecting chunks from the original document that are most likely to contain the information needed to answer the user's question, while ensuring they fit within the LLM's context window. We hypothesize that providing a short and focused VDoc to the LLM is more effective than filling the entire context window with less relevant information. Our experiments confirm this hypothesis and demonstrate that using VDocs improves results on the QA task.
+
+# 1 Introduction
+
+Large language models (LLMs) such as OpenAI's ChatGPT<sup>1</sup> or Anthropic's Claude<sup>2</sup>, have demonstrated exceptional performance across a range of natural language processing (NLP) tasks and are capable of answering questions on various topics. Recent advancements have extended their context window to encompass tens of thousands or even millions of tokens (Gemini, 2024). This enables them to process long documents, or even books, and answer questions over them.
+
+Despite this impressive progress, recent studies have shown that LLMs still struggle to utilize long contexts efficiently (Li et al., 2024; Levy et al., 2024). LLMs' performance often deteriorates when confronted with a long document, hindering their ability to accurately answer questions on such documents.
+
+To address this challenge, we propose a method called Virtual Document (VDoc). VDoc takes a
+
+![](images/e0049b320b5563f222930bf215eb4838edeeb607df6c09c7c08fe03e131fb61a.jpg)  
+Figure 1: VDoc architecture. First, a long document is segmented (a), then each segment is ranked by its relevance to the question (b), and finally a virtual document is reconstructed with the top-ranked segments (c).
+
+long document, a question, and a specified context length, and generates a shorter version of the document, containing the most relevant parts for the given length. VDoc operates in three steps: 1) Document segmentation: break down the long document into smaller segments, 2) Segments ranking: rank the segments based on their relevance to the question, and 3) VDoc reconstruction: reassemble the segments into a coherent virtual document that contains the most relevant information to the question, and still fits the given context length. The overall VDoc process is illustrated in Figure 1.
+
+We analyze experimentally, different approaches for each step. Our results show that VDoc enhances question-answering performance on various datasets, including ZeroScrolls (Shaham et al., 2023), a benchmark designed for tasks requiring a long context.
+
+The paper's main contribution is the thorough analysis and experiments of the above three steps for QA over a long document. While the suggested solution is close to RAG (Retrieval Augmented Generation), it has some unique properties. For example, the ordering of the segments based on their sequence in the original document, that does not exist in RAG. Our code is available on git<sup>3</sup>
+
+# 2 Related Work
+
+Many previous works focus on designing LLMs that can support long context. Some works modify the transformer attention mechanism (Beltagy et al., 2020; Zaheer et al., 2021; Dai et al., 2019; Ratner et al., 2023) or make it more efficient (Dao et al., 2022; Dao, 2023; Liu et al., 2023a; Gemini, 2024). Others have focused on modifying the positional embeddings (Press et al., 2022; Chen et al., 2023). Another line of research attempts to replace transformers, which have quadratic sequence length complexity, with new architectural designs like convolution and linear RNNs, e.g., RWKV (Peng et al., 2023), S4 (Gu et al., 2022), Hyena (Poli et al., 2023), or Mamba (Gu and Dao, 2023).
+
+While those works were able to extend the context length of LLMs, subsequent research has found that long-context transformers exhibit a recency bias and do not effectively utilize long-range context (Qin et al., 2023; Liu et al., 2023b; Yehudai and Bendel, 2024; Li et al., 2024). Specifically, these models have been found to struggle with ignoring irrelevant information (Levy et al., 2024). Moreover, benchmarks for long context reveal that there is still significant room for improvement compared to human performance, and for some tasks, models even struggle to surpass the naive baseline (Shaham et al., 2023).
+
+# 3 Method
+
+We deal with the problem of using an LLM to answer questions over a long document. Our method, termed VDoc, creates a shorter version of the document, which is then used as input to the LLM, instead of the entire document, to generate an answer. Creating the VDoc is done in three steps. i) Document segmentation into manageable segments, ii) Segments ranking by their potential to contain the answer to the user question, and iii) VDoc reconstruction. The detailed steps are described below.
+
+Document segmentation. We try two methods. i) sliding-window on sentence boundaries using nltk $^4$  ii) context-aware (using HTML structure when available). We define a segment-size parameter that controls the max segment length.
+
+Segments ranking. We experiment with three families of language models (LM). The first is
+
+encoder-only models, where we rank segments by the similarity of their representation to the representation of the user question. An example of such LM is MiniLM<sup>5</sup>.
+
+The second, is encoder-decoder or decoder-only models, where we rank segments by the inverse perplexity of the user question, given each of the segments. Specifically, we give each segment as input to the model with the instruction: "Generate a conversation between a user and an agent based on the given content:". We then run a forward loop and compute the Cross-Entropy loss for the question. The segments are ranked by  $r = (-1) \times loss()$  (higher is better). An example is flan-t5<sup>6</sup>.
+
+The third, is a sparse representation, Elser<sup>7</sup> which expands and indexes documents into ElasticSearch<sup>8</sup> index, with semantic terms that were learned to co-occur frequently with terms in each document.
+
+We note that among the three segment ranking methods, the first (comparing dense representations) and the last (Sparse representation) can be pre-computed on the given set of documents. The second method on the other hand (using perplexity), requires running a forward loop of the model for each segment and question, and thus can not be pre-computed, because it depends on the question. Thus the perplexity method can be quite slow for very long documents.
+
+VDoc reconstruction. We take the top ranked segments, that still fit into the given window size and evaluate two methods to concatenate them: i) by the original order in the document (Doc order) and ii) by relevancy to the question (Rank order).
+
+# 4 Experiments
+
+# 4.1 Setup of the Experiments
+
+We evaluated the VDoc performance in two scenarios: direct and end-to-end. For the direct scenario we assume a benchmark of triplets  $\{(Q, D, E)\}$  where  $Q$  is a question,  $D$  is a long document and  $E$  is an evidence i.e., a span in  $D$  that contains the answer. We define the VDoc task as follows: Given a context-window size, a question  $Q$ , and a long document  $D$ , create a virtual document that
+
+fits within the context-window size and includes E. We measure the success of the VDoc to contain the evidence (we counted  $90\%$  containment as a success, to tolerate for segmentation issues).
+
+For the end-to-end scenario, we assume a benchmark of triplets  $\{(\mathrm{Q},\mathrm{D},\mathrm{R})\}$  where  $\mathbf{Q}$  is a question,  $\mathrm{D}$  is a long document that contains the answer, and  $\mathbf{R}$  is a gold response. We then assess LLM's response against the gold,  $\mathbf{R}$ , when prompting using various VDoc strategies. We evaluate results using F1 (the harmonic mean of Recall and Precision), similar to the scrolls (Shaham et al., 2023) paper.
+
+We experiment with two LLMs, each with 8K window size, thus we limit the VDoc to sizes of up to 7200 tokens to allow space for the prompt, the question and generated answer. In all experiments, we used segment-size of 512 tokens.
+
+# 4.2 Datasets
+
+For the direct experiment, we utilize two datasets containing long documents, questions and evidence. The first is GoogleNQ (Kwiatkowski et al., 2019). This dataset contains questions from Wikipedia. Each question is labeled with the Wikipedia page that contains the answer, and a span in the document that contains the evidence. The second dataset is taken from Scrolls (Shaham et al.,  $2023)^{9}$  Scrolls contains several datasets with questions over long documents. We used Qasper which is question answering over research papers. We added an evidence for each question, from the original Qasper dataset<sup>10</sup>.
+
+For the end-to-end experiment we utilized the Qasper dataset described above and the NarrativeQA dataset from Scrolls. NarrativeQA contains question-answer pairs about entire books and movie scripts. Both datasets contain (Q, D, R) triplets.
+
+We picked a random subset of 500 examples from the validation sets of Qasper and NarrativeQA. Likewise, for GoogleNQ, we picked 500 examples where the long answer (evidence) is a section.
+
+Note that all our experiments were done on the validation sets of Qasper and NarrativeQA from Scrolls. We used the validation set and not the test set, since ground truths is available only on the validation set. We selected a random subset of 500 examples, inspired by the size of the test sets in
+
+<table><tr><td>Dataset</td><td>Avg # tokens</td><td>Max # tokens</td></tr><tr><td>GoogleNQ</td><td>12,337</td><td>273,087</td></tr><tr><td>Qasper</td><td>5,484</td><td>22,052</td></tr><tr><td>NarrativeQA</td><td>88,073</td><td>516,045</td></tr></table>
+
+Table 1: Datasets used for evaluation. Number of tokens (#) is calculated by the flan-t5-small tokenizer.
+
+ZeroScrolls. We tried the Qasper experiments on the full validation set and got very similar results, thus we performed the experiments on the 500 data points.
+
+The datasets and their statistics are described in Table 1. The numbers are for the random subset of 500 examples. Note that in all our experiments, we treat each document separately, thus we report statistics only for the subset that was used in the experiments.
+
+# 4.3 Optimal Window Size
+
+To confirm that a smaller window yields better generation results, we conducted a controlled experiment, using the evidence as a pivot. We expanded the content up to a specified window size, measured in tokens, while maintaining full sentences using the nltk tool.
+
+We experiment with expanding the content on both sides of the evidence, as well as before and after it. All three methods yielded similar results. Figure 2 reports results when expanding the window in both sides of the evidence. We conducted the experiments using two generative LLM: flant5-xxl and llama-3-8b-instruct. As observed, both models achieve the highest F1 score when provided only with the evidence, and performance degrades as data is added.
+
+![](images/244a2b0fe0aae0112b7fde4349f0db3fc201a85aae09f37358e986908836e087.jpg)  
+Figure 2: Comparing performances (F1) of response generation on Qasper when adding non-relevant information around the evidence.
+
+# 4.4 Direct VDoc Evaluation
+
+We compared three segments ranking (perplexity (flan-t5-large), minilm-l12, and Elser) and a base
+
+<table><tr><td rowspan="2">VDoc ranker</td><td colspan="3">Qasper-sliding-window</td><td colspan="3">GoogleNQ-sliding-window</td><td colspan="3">GoogleNQ-semantics</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td>prefix</td><td>43.12</td><td>64.13</td><td>81.94</td><td>84.60</td><td>91.17</td><td>94.03</td><td>84.60</td><td>91.17</td><td>94.03</td></tr><tr><td>flan-t5-large</td><td>69.91</td><td>89.86</td><td>88.89</td><td>69.42</td><td>81.77</td><td>84.70</td><td>72.10</td><td>84.90</td><td>89.55</td></tr><tr><td>minilm-l12</td><td>50.51</td><td>76.81</td><td>79.17</td><td>57.37</td><td>75.21</td><td>77.61</td><td>65.40</td><td>80.34</td><td>85.82</td></tr><tr><td>Elser</td><td>71.46</td><td>89.86</td><td>91.67</td><td>76.56</td><td>77.78</td><td>96.27</td><td>91.96</td><td>97.44</td><td>98.13</td></tr><tr><td># required VDocs</td><td>487</td><td>276</td><td>72</td><td>448</td><td>351</td><td>268</td><td>448</td><td>351</td><td>268</td></tr></table>
+
+Table 2: Direct VDoc evaluation. Comparing success (%) of VDoc rankers to capture the evidence in the VDoc. For Qasper we use sliding-window only. For GoogleNQ we use both sliding-window and semantic segmentation.  
+
+<table><tr><td rowspan="2">DB</td><td rowspan="2">VDoc ranker</td><td colspan="3">flan-t5-xxl</td><td colspan="3">llama-3-8b-instruct</td></tr><tr><td>2400</td><td>4800</td><td>7200</td><td>2400</td><td>4800</td><td>7200</td></tr><tr><td rowspan="4">Qasper</td><td>prefix</td><td>0.25</td><td>0.32</td><td>0.33</td><td>0.25</td><td>0.33</td><td>0.34</td></tr><tr><td>flan-t5-large</td><td>0.37 (0.35)</td><td>0.36 (0.35)</td><td>0.35 (0.35)</td><td>0.35 (0.34)</td><td>0.35 (0.34)</td><td>0.35 (0.33)</td></tr><tr><td>minilm-112</td><td>0.28 (0.28)</td><td>0.34 (0.33)</td><td>0.35 (0.35)</td><td>0.31 (0.28)</td><td>0.33 (0.33)</td><td>0.33 (0.32)</td></tr><tr><td>Elser</td><td>0.35 (0.34)</td><td>0.36 (0.34)</td><td>0.35 (0.35)</td><td>0.37 (0.37)</td><td>0.35 (0.35)</td><td>0.34 (0.33)</td></tr><tr><td rowspan="4">Narrative</td><td>prefix</td><td>0.15</td><td>0.15</td><td>0.15</td><td>0.14</td><td>0.16</td><td>0.17</td></tr><tr><td>flan-t5-large</td><td>0.19 (0.18)</td><td>0.19 (0.20)</td><td>0.19 (0.19)</td><td>0.19 (0.20)</td><td>0.21 (0.21)</td><td>0.23 (0.22)</td></tr><tr><td>minilm-112</td><td>0.15 (0.14)</td><td>0.16 (0.16)</td><td>0.17 (0.17)</td><td>0.16 (0.16)</td><td>0.20 (0.18)</td><td>0.20 (0.20)</td></tr><tr><td>Elser</td><td>0.17 (0.17)</td><td>0.19 (0.18)</td><td>0.19 (0.18)</td><td>0.18 (0.19)</td><td>0.22 (0.21)</td><td>0.24 (0.22)</td></tr></table>
+
+Table 3: End-to-end evaluation. Comparing F1 of the generation by flan-t5-xxl and llama-3-8b-instruct on Qasper and NarrativeQA. Each cell contains values for doc order and (rank order).
+
+line prefix (that truncates to the window-size), for varying window-size.
+
+To use Elser as a ranker, We pre-index all segments of all documents into an ES index. We add a document-id as a field to all segments of each document. To rank the segments of a given document, we query the index with the given question, and use the document-id as a filter, thus only segments of the given document are scored by Elser.
+
+Results are summarized in Table 2. The last row (# required VDocs), measures the number of cases that the input document was larger than the window size and thus required VDoc activation. For example, when using window-size of 2400, for Qasper, 487 out of the 500 documents were larger than 2400 tokens, thus required VDoc. The rest were smaller than 2400 tokens, thus we could send the full document as is. Similarly, for window-size 4800, only 276 Qasper documents were larger than 4800 tokens, thus required VDoc activation. We can see that the number of required VDocs decreases as the window size increases, and thus the success to capture the evidence in the VDoc increases.
+
+Segments ranking. We observe that on Qasper, all VDoc rankers outperform the naive method of taking the prefix, where Elser and perplexity (with flan-t5-large) achieve the best performance, with
+
+Elser performing slightly better. On GoogleNQ, we see that with a sliding-window segmentation, the naive prefix is better than the three VDoc rankers (except for Elser on window-size=7200). This implies that the evidence in GoogleNQ, in most cases, is at the beginning of the document.
+
+Document segmentation. To evaluate a more sophisticated segmentation, we utilized the HTML format of the Wikipedia pages used in GoogleNQ<sup>11</sup>. We cleansed the Wikipedia HTML to remove noise (e.g., headers, footers, sidebars, advertisements) and converted the HTML to semantic passages, identified by the document title and headings preceding the passage. When moving to semantic segmentation, we can see that Elser outperforms the prefix. For example, even on the smaller window size of 2400, it succeeded to capture the gold evidence in  $91.96\%$  of the cases, compared to  $72.10\%$  and  $84.60\%$  with perplexity (flan-t5-large) and prefix respectively.
+
+# 4.5 End-to-End Evaluation
+
+We compared two LLM models with 8K window size: flan-t5-xxl and llama-3-8b-instruct. We use the prompts from ZeroScrolls. For llama3, we
+
+modified the prompts a bit. The detailed prompts and an example are given in Appendix A.
+
+Table 3 shows the results on Qasper and NarrativeQA, for varying context-window size (2400, 4800, 7200) and the different segment rankers.
+
+As seen, all VDoc strategies outperform the naive prefix method both on NarrativeQA and on Qasper. For example on NarrativeQA, with context-window of 7200, llama-3-8b-instruct achieved an improvement of  $41\%$  (0.24 with Elser, compared to 0.17 with prefix). On Qasper we see an improvement of  $48\%$ , on context-window 2400, with llama-3-8b-instruct (0.37 with Elser compared to 0.25 with prefix).
+
+Furthermore, we observe a correlation with the direct VDoc evaluation (Table 2). In both experiments, the best results are achieved for Elser and perplexity (flan-t5-large) rankers. A closer examination of the two datasets reveals that on Qasper, those two rankers remain quite stable across different window sizes. On NarrativeQA, we can see an improvement for all segment rankers, when increasing the window size, and all outperform the naive prefix by a large margin.
+
+The behavior of the two datasets supports our hypothesis that it is enough to give the LLM a smaller window and still get comparable results to those with a larger window. On Qasper, the 4800 window is a sweet spot, namely a good balance between capturing the evidence, and yet not adding too much irrelevant information to the LLM. This is because most of Qasper documents fit into the 7200 window. On NarrativeQA, the sweet spot is beyond 7200 since the documents are much larger.
+
+VDoc reconstruction. As shown in Table 3, ar- ranging the top-selected segments in their original order (Doc order) generally results in a better F1 score compared to ordering them by relevancy to the question (Rank order).
+
+ZeroScrolls evaluation. We submitted flan-t5-xxl on the best VDoc configuration of Elser with window size 4800 to the official ZeroScrolls leader board. With VDoc, our flan-t5-xxl achieved F1 scores of 50.7, 30.6 on Qasper and NarrativeQA respectively, compared to F1 of 48.3 and 19.3 when used without the VDoc. This is an improvement of  $58\%$  on NarrativeQA which has very large documents.
+
+# 4.6 The Effect on Inference Times
+
+In this section, we analyze the contribution of the VDoc on inference times. We used the Llama-3-1-70b-instruct model, which supports an input length of 128K tokens. We run it on the NarrativeQA 500 examples, used in all our experiments.
+
+Figure 3 shows the running times and the F1 as a function of the context-window size. We used the llama-3-1-70b-instruct on IBM watsonx.ai $^{12}$  for inference. The times are for generation only. The F1 values are reported for VDoc that uses the best segment-ranking method (see Table 3 above), namely Elser. We can see that the best F1 is achieved with VDoc for context window size of 7200 tokens, and it starts degrading as we increase the context window size. Still we can see that running times for the 500 examples, increase from 28 min (3.36 sec/example) to 125 min with context-window of 64K (15 sec/example).
+
+![](images/dd927fbe8824ca8ae7439c058915e7816cc5d636ca33bba34ac12c3ddd831c68.jpg)  
+Figure 3: Inference times and F1 of response generation, on NarrativeQA, as a function of context-window size.
+
+# 5 Conclusions and Future Work
+
+We described a method called VDoc, enabling Question Answering using LLM over long documents. The method is based on creating a virtual document that contains relevant information to the user query, and adheres to the LLM's context-window limitations. Our method does not require any modifications or fine-tuning a LLM.
+
+We compared different methods and showed experimentally, that it is better to aggressively shorten the document, even if the LLM can handle a larger context-window size, both in terms of quality of the results, and in terms of better run times.
+
+# References
+
+Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer.  
+Shouyuan Chen, Sherman Wong, Liangjian Chen, and Yuandong Tian. 2023. Extending context window of large language models via positional interpolation.  
+Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V. Le, and Ruslan Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-length context.  
+Tri Dao. 2023. Flashattention-2: Faster attention with better parallelism and work partitioning.  
+Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashattention: Fast and memory-efficient exact attention with io-awareness.  
+Gemini. 2024. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.  
+Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.  
+Albert Gu, Karan Goel, and Christopher Ré. 2022. Efficiently modeling long sequences with structured state spaces.  
+Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:452-466.  
+Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same task, more tokens: the impact of input length on the reasoning performance of large language models.  
+Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. 2024. Long-context llms struggle with long in-context learning.  
+Hao Liu, Matei Zaharia, and Pieter Abbeel. 2023a. Ring attention with blockwise transformers for near-infinite context.  
+Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023b. Lost in the middle: How language models use long contexts.  
+Bo Peng, Eric Alcaide, Quentin Anthony, Alon Albalak, Samuel Arcadinho, Stella Biderman, Huanqi Cao, Xin Cheng, Michael Chung, Matteo Grella, Kranthi Kiran GV, Xuzheng He, Haowen Hou, Jiaju Lin, Przemyslaw Kazienko, Jan Kocon, Jiaming Kong, Bartlomiej Koptyra, Hayden Lau, Krishna Sri Ipsit Mantri, Ferdinand Mom, Atsushi Saito, Guangyu Song, Xiangru Tang, Bolun Wang, Johan S. Wind, Stanislaw Wozniak, Ruichong Zhang, Zhenyuan
+
+Zhang, Qihang Zhao, Peng Zhou, Qinghua Zhou, Jian Zhu, and Rui-Jie Zhu. 2023. Rwkv: Reinventing rnns for the transformer era.  
+Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, and Christopher Ré. 2023. Hyena hierarchy: Towards larger convolutional language models.  
+Ofir Press, Noah A. Smith, and Mike Lewis. 2022. Train short, test long: Attention with linear biases enables input length extrapolation.  
+Guanghui Qin, Yukun Feng, and Benjamin Van Durme. 2023. The nlp task effectiveness of long-range transformers.  
+Nir Ratner, Yoav Levine, Yonatan Belinkov, Ori Ram, Inbal Magar, Omri Abend, Ehud Karpas, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023. Parallel context windows for large language models.  
+Uri Shaham, Maor Ivgi, Avia Efrat, Jonathan Berant, and Omer Levy. 2023. Zeroscrolls: A zero-shot benchmark for long text understanding.  
+Asaf Yehudai and Elron Bendel. 2024. When llms are unfit use fastfit: Fast and effective text classification with many classes.  
+Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontonon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. 2021. Big bird: Transformers for longer sequences.
+
+# A Appendix - Detailed Prompts and Examples
+
+In all the end-to-end experiments, we used the ZeroScrolls prompts. For llama-3-8b-instruct, we found however that repeating the last instruction after the query, helps improve its results. Without that addition, llama3 sometimes tends to return very long answers, which hinders precision.
+
+Tables 4 and 5 list the prompts used by llama3 for the Qasper and NarrativeQA respectively. Table 6 illustrates the above phenomena. We can see that without the suffix, llama3 tends to give very long answers. The F1 for the first answer is 0.67 (with recall=1 and precision=0.5), while the values for the second answer are 0.03, 1, 0.02 for F1, recall and precision, respectively.
+
+You are given a scientific article and a question. Answer the question as concisely as you can, using a single phrase or sentence if possible. If the question cannot be answered based on the information in the article, write "unanswerable". If the question is a yes/no question, answer "yes", "no", or "unanswerable".
+
+# Article:
+
+{input}
+
+# Question:
+
+{question} Answer the question as concisely as you can, using a single phrase or sentence if possible.
+
+# Answer:
+
+Table 4: llama-3 prompt for Qasper. Note the added suffix after the question
+
+You are given a story, which can be either a novel or a movie script, and a question. Answer the question as concisely as you can, using a single phrase if possible.
+
+# Story:
+
+{input}
+
+# Question:
+
+{question} Answer the question as concisely as you can, using a single phrase or sentence if possible.
+
+# Answer:
+
+Table 5: llama-3 prompt for NarrativeQA. Note the added suffix after the question
+
+Question: What kind of injustices do the stories point out?
+
+Gold answer: Social injustices.
+
+Answer 1: Social and economic injustices
+
+Answer 2: The stories point out the injustices of social class, wealth, and power. In "The Adventure of the Abbey Grange", the stepfather's abuse of power and wealth leads to the mistreatment of his stepchildren. In "The Adventure of the Black Swan", the lady's freedom is threatened by the five miles of country between her home and Winchester, highlighting the limitations of women's freedom in rural areas. In "The Adventure of the Empty House", Sherlock Holmes criticizes the lack of justice for the poor and ignorant, who are often victims of hidden wickedness and cruelty. The stories also highlight the class differences between the wealthy and the poor, with the wealthy often having more power and influence. Overall, the stories point out the injustices of a society where wealth, power, and social class can lead to exploitation and mistreatment of others.
+
+Table 6: An example question. Answer 1 is with the added suffix, while Answer 2 is without.

2024/More Bang for your Context_ Virtual Documents for Question Answering over Long Documents/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de6e084a26dcb3f8e256ec7aa9decf78598198961c340a258ebb9f83e1f0e1c0
+size 179670

2024/Multi-Loss Fusion_ Angular and Contrastive Integration for Machine-Generated Text Detection/e63558dd-1191-4591-a960-30434ec75c28_content_list.json

@@ -0,0 +1,1808 @@
+[
+    {
+        "type": "text",
+        "text": "Multi-Loss Fusion: Angular and Contrastive Integration for Machine-Generated Text Detection",
+        "text_level": 1,
+        "bbox": [
+            188,
+            90,
+            811,
+            129
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Iqra Zahid, Yue Chang, Tharindu Madusanka, Youcheng Sun and Riza Batista-Navarro",
+        "bbox": [
+            119,
+            149,
+            887,
+            165
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Department of Computer Science, University of Manchester",
+        "bbox": [
+            255,
+            166,
+            744,
+            181
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "{iqra.zahid|yue.chang|tharindu.batawalaacharige}@manchester.ac.uk, {youcheng.sun|riza.batista}@manchester.ac.uk",
+        "bbox": [
+            169,
+            183,
+            831,
+            216
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Abstract",
+        "text_level": 1,
+        "bbox": [
+            260,
+            252,
+            339,
+            268
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Modern natural language generation (NLG) systems have led to the development of synthetic human-like open-ended texts, posing concerns as to who the original author of a text is. To address such concerns, we introduce DeB-Ang: the utilisation of a custom DeBERTa model (He et al., 2021) with angular loss and contrastive loss functions for effective class separation in neural text classification tasks. We expand the application of this model on binary machine-generated text detection and multi-class neural authorship attribution. We demonstrate improved performance on many benchmark datasets whereby the accuracy for machine-generated text detection was increased by as much as  $38.04\\%$  across all datasets.",
+        "bbox": [
+            141,
+            280,
+            460,
+            508
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "1 Introduction",
+        "text_level": 1,
+        "bbox": [
+            114,
+            521,
+            260,
+            537
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "There has been considerable activity in the field of detecting machine-generated text. Driven by the significant growth and the increasing prevalence of large language models (LLMs) and natural language generation (NLG) models. This has led to the production of high-quality human-like texts that have brought about useful applications in many domains such as machine translation, text summarisation and data generation (Kieuvongngam et al., 2020; Goyal et al., 2022; Iyer et al., 2023; Uchendu et al., 2021). Irrespective of the many useful applications, the deployment of NLG models has concurrently given rise to serious concerns such as plagiarism, and spreading misinformation and hate speech (Pu et al., 2022; Hu et al., 2023; Qadir, 2022; Solaiman et al., 2019). Therefore, the need to discriminate between human and machine-generated text becomes paramount, especially in light of the growing sophistication and rapid updates of these models.",
+        "bbox": [
+            112,
+            548,
+            489,
+            868
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "Given the diverse applications of NLG models, authorship attribution (AA) methods have been increasingly employed to detect the original author",
+        "bbox": [
+            112,
+            870,
+            489,
+            919
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "of synthetic data generated by machines (Ai et al., 2022; Uchendu et al., 2020; Jawahar et al., 2020). The main concern with traditional AA methods is that, typically, they are feature-based systems and consist of largely document-specific features. Therefore, the application of this traditional model is often author, dataset and model-specific (Sari, 2018; Ai et al., 2022). Previous research addressed the need for generalisable detection systems to identify machine-generated text (Fagni et al., 2021; Jakesch et al., 2023; He et al., 2024; Jawahar et al., 2020). Research involving the use of LLMs in authorship attribution has demonstrated that the simple fine-tuning of pre-trained language models can surpass the accuracy of traditional methods significantly (Fabien et al., 2020; Mitrovic et al., 2023; Fagni et al., 2021).",
+        "bbox": [
+            507,
+            252,
+            884,
+            525
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "In particular, we introduce DeB-Ang, a pretrained DeBERTa model with a specialised angular loss and contrastive loss integration. Additionally, we demonstrate improved classification when applying DeB-Ang to several well-known machine-generated text and authorship attribution datasets. Contrastive learning is an unsupervised representation learning technique, aiming to learn a representation of data such that similar instances are close in the representation space whereas dissimilar instances are far apart (Aljundi et al., 2022). Loss functions are crucial in contrastive learning as they quantify the similarity and dissimilarity between pairs, guiding the model to learn meaningful representations for class discrimination (Hadsell et al., 2006; Gao et al., 2022; Wang et al., 2017). However, recent studies suggest that various loss functions, including cross-entropy loss, contrastive loss and triplet loss, fail to consider the intrinsic angular distribution exhibited by the low-level and high-level feature representations (Choi et al., 2020), which contributes to our choice of using angular loss in DeB-Ang. Angular loss is a scale-invariant loss function designed to improve the learning sim",
+        "bbox": [
+            507,
+            533,
+            884,
+            917
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "page_number",
+        "text": "7189",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "footer",
+        "text": "Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7189-7202 November 12-16, 2024 ©2024 Association for Computational Linguistics",
+        "bbox": [
+            216,
+            945,
+            779,
+            972
+        ],
+        "page_idx": 0
+    },
+    {
+        "type": "text",
+        "text": "ality metrics by considering the angle between vectors (Wang et al., 2017).",
+        "bbox": [
+            112,
+            84,
+            485,
+            115
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "In summary, the contributions of this work are four-fold:",
+        "bbox": [
+            112,
+            117,
+            485,
+            147
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. We propose a novel customisable contrastive learning framework that combines a custom fine-tuned DeBERTa model (He et al., 2021) with contrastive and angular loss functions. We assess the difference in classification performance when utilising various combinations of the aforementioned loss functions for the proposed task.",
+            "2. We assess the application of the proposed model on multi-class authorship attribution and binary machine-generated text detection.",
+            "3. We introduce three new large-scale datasets for evaluating text classification models. These datasets were constructed by leveraging state-of-the-art language models, including Gemma-7b (Team et al., 2024), GPT4-Turbo (OpenAI, 2023) and Flan-T5-Large (Chung et al., 2022).",
+            "4. We conduct linguistic error analysis of incorrectly and correctly classified examples."
+        ],
+        "bbox": [
+            127,
+            151,
+            487,
+            472
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2 Related Work",
+        "text_level": 1,
+        "bbox": [
+            114,
+            485,
+            267,
+            500
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.1 Machine-generated text detection",
+        "text_level": 1,
+        "bbox": [
+            114,
+            511,
+            421,
+            526
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Studies have demonstrated that human participants were unable to distinguish between machine-generated texts and human written texts (Jakesch et al., 2023; Islam et al., 2023; Ippolito et al., 2020; Dugan et al., 2020, 2022). Previous work highlighted that disambiguating between human and LLM-generated texts is increasingly difficult (Pu and Demberg, 2023; Jakesch et al., 2023; Cox, 2005). Automatic detection of machine-generated text has thus gained popularity and can be categorised according to their underlying method (Solaiman et al., 2019; Uchendu et al., 2020; Fagni et al., 2021; Bakhtin et al., 2019; Ippolito et al., 2020). Simple classifiers often involve linguistic feature analysis (Dugan et al., 2020, 2022) or incorporate a psycholinguistics statistical measure (Venkatraman et al., 2024). Other methods include zero-shot detection (Solaiman et al., 2019), and fine-tuned model detection (Uchendu et al., 2020; Ippolito et al., 2020; Fagni et al., 2021; Adelani et al., 2019; Tay et al., 2020; Zellers et al., 2021). Irrespective of the large number of approaches to identifying machine-generated texts, detection remains a challenge (Crothers et al., 2023; Ai et al.,",
+        "bbox": [
+            112,
+            533,
+            489,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2022).",
+        "bbox": [
+            509,
+            84,
+            561,
+            98
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.2 Authorship Attribution",
+        "text_level": 1,
+        "bbox": [
+            509,
+            112,
+            739,
+            127
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Traditional attribution approaches utilise linguistic features in a univariate (utilising a single linguistic feature, e.g. function words) (Martindale and McKenzie, 1995) or multivariate (utilising multiple linguistic features, e.g. Writeprints) approach (Abbasi and Chen, 2008; Sari, 2018). As aforementioned, feature-based linguistic identification requires dataset-specific engineering, displaying limited scalability (Sari, 2018; Ai et al., 2022). More recently, the use of learning-based approaches has grown with the use of pre-trained LLMs (Fabien et al., 2020). These approaches have demonstrated the power of LLMs in significantly surpassing the accuracy of traditional approaches with little analysis required beforehand (Ai et al., 2022).",
+        "bbox": [
+            507,
+            133,
+            884,
+            375
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "2.3 Research gaps",
+        "text_level": 1,
+        "bbox": [
+            509,
+            387,
+            668,
+            401
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "Existing approaches in detecting synthetic texts created by LLMs have many limitations. For example, these detection tools are now outdated due to rapid technological advancements, e.g., DetectGPT classifies texts only generated by GPT2 (Mitchell et al., 2023). This necessitates classifier retraining which could negatively affect the accuracy of these models (OpenAI, 2023). Additionally, the increased advancements of NLG models have led to more human-like texts. Further, these models are LLM-specific and therefore, do not detect synthetic texts generated by other language models. Also, these methods have a black-box nature, making it difficult for humans to understand their output for correctly and incorrectly classified texts. Given the increasing prevalence of machine-generated texts, it is vital that we are able to distinguish which NLG model was used to generate a given text. We extend this to being able to detect the exact model version.",
+        "bbox": [
+            507,
+            407,
+            884,
+            714
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3 Data",
+        "text_level": 1,
+        "bbox": [
+            509,
+            727,
+            586,
+            741
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "3.1 Data collection",
+        "text_level": 1,
+        "bbox": [
+            509,
+            753,
+            672,
+            766
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "There is a strong consensus that datasets must be diverse and representative (Tang et al., 2023). To this end, we chose to utilise datasets with original, human-written texts; different versions of each text are then generated with the aid of LLMs which were given carefully designed prompts. Datasets were taken from Kaggle and the Turing Test benchmark known as TuringBench (Uchendu et al., 2021). The TuringBench dataset consists of ar",
+        "bbox": [
+            507,
+            774,
+            884,
+            917
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "page_number",
+        "text": "7190",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 1
+    },
+    {
+        "type": "text",
+        "text": "ticles generated by 20 authors. There are a total of 20 datasets from 19 different NLG model versions and one human author. The DAIGT-V2 dataset consists of 37 authors (36 NLG models and one human author, with 60K texts). Further dataset details can be seen in Appendix A in Table 7. All datasets utilised were generated for the proposed tasks.",
+        "bbox": [
+            112,
+            84,
+            487,
+            211
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, we utilised 5 randomly selected datasets from TuringBench datasets. We opted to use one dataset per model. For example, there are two datasets generated by XLNET. The exact model verions are XLNET_base and XLNET_large. Therefore, we employ only one of these model versions. We randomly sampled the dataset due to limited computational resources. Details of the specific processing steps and size of the data taken for each of the different datasets are provided in Section 5. Table 7 in Appendix A presents—for each dataset that we utilised—the dataset name, source and the models that were used to generate the texts contained in each dataset.",
+        "bbox": [
+            115,
+            214,
+            489,
+            439
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.2 Data Generation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            454,
+            292,
+            468
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We also generated our own datasets by using GPT4-Turbo (OpenAI, 2023), Gemma-7b (Team et al., 2024) and Flan-T5-large (Chung et al., 2022). The TuringBench dataset set consists of 19 different NLG model versions however, these models are no longer considered state-of-the-art models. We decided to generate three additional datasets from more recent models which are considered to be the current state-of-the-art and were not included in the original TuringBench dataset. This dataset set is referred to as TuringExtended. This enables the examination of MGT and AA within the context of newer NLG models, underpinning the exploration as to whether newer NLG models are more challenging to identify as machine-generated.",
+        "bbox": [
+            112,
+            478,
+            489,
+            718
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Specifically, additional datasets were generated as an extension of TuringBench (Uchendu et al., 2021). Considering only the human-written texts from the original AA dataset from TuringBench, we extracted only a total of 7678 (non-duplicated) rows of text. The models that we employed are GPT4 Turbo $^{1}$  (He et al., 2021), Gemma-7b $^{2}$  (Team et al., 2024) and Flan-T5-large $^{3}$  (Chung et al., 2022), resulting in the creation of three new",
+        "bbox": [
+            112,
+            720,
+            489,
+            865
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "table",
+        "img_path": "images/8f94c1b89fe58e6ee58978c7dbe3bf566cc30a2bf69e087c4824fc8bb6296b45.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>GPT4-Turbo</td><td>Gemma-7b</td><td>Flan-T5-Large</td></tr><tr><td>BERTScore P</td><td>83.30</td><td>92.32</td><td>90.51</td></tr><tr><td>BERTScore R</td><td>84.11</td><td>95.55</td><td>83.98</td></tr><tr><td>BERTScore F1</td><td>83.70</td><td>93.88</td><td>87.08</td></tr><tr><td>IAA</td><td>88.64</td><td>89.30</td><td>84.64</td></tr></table>",
+        "bbox": [
+            510,
+            80,
+            882,
+            137
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Table 1: Averaged BERTScore Precision (P), Recall (R) and F1-score (F1) for the datasets generated by the specified models. Inter-annotator agreement (IAA) is also provided.",
+        "bbox": [
+            507,
+            147,
+            882,
+            204
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "datasets. The models were given the prompt \"generate a similar article\", which is slightly similar to what was used in TuringBench (\"generate an article similar to the human-written one\").",
+        "bbox": [
+            507,
+            230,
+            882,
+            294
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "The three models for generation (GPT4-Turbo, Gemma-7b and Flan-T5-large) were chosen on the basis that they were either the current state-of-the-art models, or that they were not previously employed in creating the TuringBench datasets.",
+        "bbox": [
+            507,
+            294,
+            884,
+            375
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "3.3 Data Evaluation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            387,
+            684,
+            401
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "We evaluated the quality of the generated data using a combination of the automated metric BERTScore (Zhang et al., 2020) and human evaluators. BERTScore calculates token similarity using contextual embeddings to calculate the similarity between tokens in the candidate and reference text. This metric has demonstrated an advanced performance by correlating strongly with human judgement in various evaluative tasks (Zhang et al., 2020). In parallel, four human annotators were trained on evaluating generated text and were provided with some background information on text generation. Each annotator assessed 250 rows from each dataset and was asked to label the data as coherent (0) or incoherent (1). For a data sample to be labelled as coherent it had to meet two criteria: texts should be semantically and grammatically sound. Inter-annotator agreement (IAA) was then measured between all annotators for each dataset. The averaged BERTScore precision, recall and F1-scores, and IAA results are presented in Table 1.",
+        "bbox": [
+            505,
+            409,
+            884,
+            745
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4 Methodology",
+        "text_level": 1,
+        "bbox": [
+            507,
+            759,
+            658,
+            776
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "4.1 Loss Functions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            785,
+            672,
+            799
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "Previous studies have focussed on increasing similarity between representations by using varying loss functions (Ai et al., 2022; Vygon and Mikhaylovskiy, 2021). However, many approaches focus on the utilisation of a single loss function. In this paper, we propose a multi-loss fusion by using the weighted sum of a combination of vari",
+        "bbox": [
+            507,
+            806,
+            884,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<https://platform.openai.com/docs/models",
+        "bbox": [
+            134,
+            878,
+            440,
+            891
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{2}$ https://huggingface.co/google/Gemma-7b",
+        "bbox": [
+            136,
+            891,
+            433,
+            904
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_footnote",
+        "text": "<sup>3</sup>https://huggingface.co/google/flan-t5-large",
+        "bbox": [
+            136,
+            904,
+            472,
+            917
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "page_number",
+        "text": "7191",
+        "bbox": [
+            480,
+            927,
+            517,
+            940
+        ],
+        "page_idx": 2
+    },
+    {
+        "type": "text",
+        "text": "ous loss functions: angular loss, cross-entropy loss and contrastive loss. Cross-entropy loss measures probability distributions; the objective is to minimise the error between the predicted probability and true distribution (Mao et al., 2023). This is used in updating the model weights during optimisation. Angular loss, often used in deep learning tasks (Wang et al., 2017; Kim et al., 2023, 2021; Choi et al., 2020), considers the angle between vectors to enhance learning for an improved similarity metric. The utilisation of angular loss in text classification often leads to more adaptable and robust models capable of handling linguistic diversity (Gao, 2022; Hui et al., 2019; Wang et al., 2017; Deng et al., 2019). Contrastive learning focusses on learning representations of data so that similar instances are closer in the embedding space and dissimilar instances are apart (Tan et al., 2024).",
+        "bbox": [
+            112,
+            84,
+            489,
+            374
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.2 Problem Statement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            386,
+            312,
+            399
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The goal of our approach is to capture nuanced semantic representations and to effectively discriminate learned embeddings. We propose leveraging the DeBERTa model with angular and contrastive loss integration (DeB-Ang). This process aims to enhance the discriminative capabilities and quality of embeddings to improve the model's performance on downstream classification tasks.",
+        "bbox": [
+            112,
+            406,
+            487,
+            533
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.3 Implementation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            546,
+            285,
+            562
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Our textual datasets underwent cleaning and preprocessing procedures. A 70:10:10 split was applied to partition the data into a training, validation and test sets for model evaluation.",
+        "bbox": [
+            112,
+            568,
+            487,
+            631
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Building upon the DeBERTa base model (microsoft/deberta-base), we implemented a new model, DeB-Ang, that integrates the angular and contrastic loss into the training step. The model was implemented using PyTorch (Paszke et al., 2019) and Simple Transformers<sup>4</sup> and was configured with specific hyperparameters (See Appendix B); additionally, early stopping criteria were set to improve training efficiency.<sup>5</sup>",
+        "bbox": [
+            112,
+            633,
+            487,
+            778
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.4 Angular Loss Computation",
+        "text_level": 1,
+        "bbox": [
+            112,
+            789,
+            374,
+            806
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "The angular loss function begins by computing the cosine similarity between all pairs of extracted embeddings. Positive and negative pairs are then",
+        "bbox": [
+            112,
+            810,
+            487,
+            859
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "generated to ensure the model can distinguish between embeddings with the same and different labels. Subsequently, we compute the loss for positive and negative pairs in order to optimise embeddings to have lower similarity for pairs with different labels and higher similarity for those with the same. This ultimately decides their position in the embedding space. The sum of positive and negative loss creates a complete loss function. This function, given below, guides the optimisation to achieve embeddings that have properties of similarity and dissimilarity.",
+        "bbox": [
+            507,
+            84,
+            884,
+            277
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {A n g u l a r} = \\sum_ {i = 1} ^ {n} \\log \\left(\\sum_ {j \\neq i} e ^ {s _ {i j}}\\right) - \\log \\left(\\sum_ {j \\neq i} e ^ {s _ {j i}}\\right)\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            302,
+            885,
+            355
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where:",
+        "bbox": [
+            509,
+            363,
+            564,
+            376
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $n$  is the number of embeddings;",
+            "-  $s_{ji}$  is the cosine similarity between embeddings  $i$  and  $j$ ;",
+            "- The first term encourages embeddings from different classes (negative pairs) to have lower cosine similarity;",
+            "- The second term encourages embeddings from the same class (positive pairs) to have higher cosine similarity."
+        ],
+        "bbox": [
+            531,
+            380,
+            882,
+            523
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "4.5 Contrastive Loss Computation",
+        "text_level": 1,
+        "bbox": [
+            507,
+            536,
+            794,
+            551
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "Generating positive pairs facilitates the learning of intra-class relationships by allowing embeddings with the same labels but different indices to be considered for optimisation. Generating negative pairs enhances the discrimination capability of these embeddings. We compute the loss for positive and negative pairs. This allows embeddings of similar instances to be pushed closer to each other in the embedding space, whereas negative embeddings push them apart thus improving intra-class clustering and inter-class separation. Combining the positive and negative loss, as shown below, guides the model to learn embeddings that capture both intra-class relationships and inter-class distinctions.",
+        "bbox": [
+            507,
+            556,
+            882,
+            782
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\nL _ {C o n t r a s t i v e} = \\sum_ {i, j} y _ {i j} d _ {i j} + (1 - y _ {i j}) \\max  (0, m - d _ {i j})\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            509,
+            809,
+            894,
+            844
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "where:",
+        "bbox": [
+            509,
+            854,
+            564,
+            868
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "text",
+        "text": "-  $y_{ij}$  is a binary label indicating whether embeddings  $i$  and  $j$  belong to the same class (1) or different classes (0);",
+        "bbox": [
+            529,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_footnote",
+        "text": "$^{4}$ Simple Transformers: https://simpletransformers.ai/docs/classification-specifics/  \n $^{5}$ Code and data: https://github.com/iqrazahid05/DeB-Ang/",
+        "bbox": [
+            112,
+            866,
+            487,
+            917
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "page_number",
+        "text": "7192",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 3
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $d_{ij}$  is the distance between embeddings  $i$  and  $j$ ;",
+            "-  $m$  is a margin hyperparameter;",
+            "- For positive pairs  $(y_{ij} = 1)$ , the loss is  $d_{ij}$ , encouraging embeddings to be closer together;",
+            "- For negative pairs  $(y_{ij} = 0)$ , the loss is  $\\max(0, m - d_{ij})$ , encouraging embeddings to be apart by at least a distance of  $m$ ."
+        ],
+        "bbox": [
+            136,
+            84,
+            487,
+            212
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.6 Our DeB-Ang Model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            225,
+            324,
+            242
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In the DeB-Ang model, we utilise three loss functions, as shown in the equation below: angular loss, cross-entropy loss and contrastive loss. Angular loss is used to facilitate intra-class compactness and inter-class separation. Within our model, we utilise cross-entropy loss to penalise the models' misclassification by computing the difference between predicted and actual labels. Cross-entropy is the standard loss function that was incorporated into the DeBERTa model. Constractive loss enhances the embeddings' discriminative abilities by encouraging similarity for positive pairs and dissimilarity for negative pairs.",
+        "bbox": [
+            112,
+            249,
+            489,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "equation",
+        "text": "\n$$\n\\begin{array}{l} L _ {T o t a l} = w _ {C E} L _ {C E} \\\\ + w _ {A n g u l a r} L _ {A n g u l a r} \\\\ + w _ {\\text {C o n t r a s t i v e}} L _ {\\text {C o n t r a s t i v e}} \\\\ \\end{array}\n$$\n",
+        "text_format": "latex",
+        "bbox": [
+            171,
+            489,
+            426,
+            545
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "where:",
+        "text_level": 1,
+        "bbox": [
+            112,
+            562,
+            166,
+            574
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "-  $L_{Total}$  is the total loss function used for training the DeBERTa model;",
+            "-  $L_{CE}$  is the standard cross-entropy loss for classification tasks, calculated as  $L_{CE} = -\\sum_{i=1}^{n} \\log P(y_i | X)$ , where  $X$  is the input sequence and  $y_i$  is the true label for the  $i$ -th example;",
+            "-  $L_{\\text{Angular}}$  is the angular loss based on cosine similarity;",
+            "-  $L_{\\text{Contrastive}}$  is the contrastive loss;",
+            "-  $w_{CE}, w_{Angular}$ , and  $w_{Contrastive}$  are the corresponding weights for each loss component, allowing for fine-tuning the contribution of each loss value during training."
+        ],
+        "bbox": [
+            134,
+            577,
+            487,
+            802
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "This combined loss function incorporates three learning objectives:",
+        "bbox": [
+            112,
+            804,
+            485,
+            835
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. The cross-entropy loss which ensures that the model learns to correctly classify the input sequences based on the true labels.",
+            "2. The angular loss which encourages the model to learn more separated representations for dif"
+        ],
+        "bbox": [
+            127,
+            839,
+            487,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "image",
+        "img_path": "images/59ba8156e19487087b5a75dbf4e79ba7e74282f9b97cf16230c15df7f4a920c1.jpg",
+        "image_caption": [
+            "Figure 1: Architecture of the DeB-Ang model highlighting loss computation within DeBERTa."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            526,
+            93,
+            867,
+            189
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "ferent classes, based on the cosine similarity between the embeddings.",
+        "bbox": [
+            544,
+            263,
+            880,
+            294
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "3. The contrastive loss further enforces the separation between inter-class embeddings, while bringing intra-class embeddings closer together, based on the similarity calculations and a specified margin.",
+        "bbox": [
+            524,
+            296,
+            882,
+            376
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "By combining these three loss components, the DeBERTa model can potentially learn more robust and discriminative representations, leading to improved classification performance on various natural language processing tasks.",
+        "bbox": [
+            507,
+            378,
+            882,
+            458
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "As shown in Figure 1, the DeB-Ang model integrates angular and contrastive loss computations within the DeBERTa model, detailing the flow from data input to loss calculation and backpropagation.",
+        "bbox": [
+            507,
+            458,
+            882,
+            523
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "4.7 Evaluation and Error Analysis",
+        "text_level": 1,
+        "bbox": [
+            507,
+            533,
+            796,
+            550
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "Considering the scale of the datasets, some accuracy values, when taken at face value, may not demonstrate any meaningful improvement in performance. Therefore, we utilise McNemar's test (Sundjaja et al., 2023) to demonstrate the statistical significance of our results. McNemar's test is a non-parametric test that can be used in comparing the performance of two classification models.",
+        "bbox": [
+            507,
+            555,
+            882,
+            682
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "For error analysis, we extracted both incorrectly classified and correctly classified data samples and performed an in-depth linguistic analysis of the outputs. We also computed the semantic similarity between correctly and incorrectly classified data by measuring the cosine similarity between the embeddings of the text pairs. We extracted contextual embeddings using the same DeBERTa model.",
+        "bbox": [
+            507,
+            684,
+            880,
+            810
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5 Results and Discussions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            824,
+            747,
+            838
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "5.1 Machine-generated Text Detection",
+        "text_level": 1,
+        "bbox": [
+            507,
+            850,
+            823,
+            865
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "text",
+        "text": "In this section, we investigate binary machine-generated text detection, whereby the task is focussed on differentiating between human and",
+        "bbox": [
+            507,
+            871,
+            882,
+            917
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "page_number",
+        "text": "7193",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 4
+    },
+    {
+        "type": "table",
+        "img_path": "images/8a6f435517e294bd4fec36fd5d8d2e486807b9ef0573451c12d85d56522af77e.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"10\">Accuracy and F1-score</td></tr><tr><td></td><td colspan=\"2\">Contra-X</td><td colspan=\"2\">Baseline DeBERTa</td><td colspan=\"2\">DeB-Ang</td><td>SS (Y/N)</td><td colspan=\"2\">Min-Max Improvement</td></tr><tr><td>PPLM-gpt2</td><td>99.34</td><td>99.32</td><td>99.66</td><td>99.66</td><td>99.98</td><td>100</td><td>Y</td><td>0.32</td><td>0.64</td></tr><tr><td>GPT1</td><td>52.66</td><td>52.66</td><td>99.89</td><td>99.93</td><td>97.92</td><td>97.92</td><td>Y</td><td>2.01</td><td>47.27</td></tr><tr><td>FAIR-wmt20</td><td>61.95</td><td>60.94</td><td>99.39</td><td>99.36</td><td>99.99</td><td>99.98</td><td>Y</td><td>0.60</td><td>38.04</td></tr><tr><td>GPT-3</td><td>97.85</td><td>97.85</td><td>98.80</td><td>98.8</td><td>99.73</td><td>99.73</td><td>Y</td><td>0.93</td><td>1.88</td></tr><tr><td>Grover-large</td><td>99.26</td><td>99.26</td><td>99.76</td><td>99.77</td><td>99.99</td><td>99.99</td><td>Y</td><td>0.23</td><td>0.73</td></tr><tr><td>transfo-xl</td><td>97.85</td><td>97.85</td><td>99.69</td><td>99.69</td><td>99.99</td><td>99.99</td><td>Y</td><td>0.30</td><td>2.14</td></tr><tr><td>GPT-2 small</td><td>96.57</td><td>96.57</td><td>99.82</td><td>99.82</td><td>99.52</td><td>99.54</td><td>Y</td><td>-0.3</td><td>2.95</td></tr></table>",
+        "bbox": [
+            129,
+            80,
+            867,
+            230
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "machine-written texts. Table 2 presents the results for this task on a variety of datasets from Turing-Bench. From the table, it is evident that the proposed model outperforms both Contra-X (Ai et al., 2022) and a baseline DeBERTa model with a minimum improvement of  $0.23\\%$  and maximum improvement of  $47.27\\%$  in accuracy. Statistical significance was computed by comparing DeB-Ang with the baseline DeBERTa model, as they exhibited the closest performance. The results for machine-generated text detection for the TuringExtended data is presented in Table 3. This demonstrates that the DeB-Ang model can differentiate between human and machine-generated texts even if the latter were generated by the newer NLG models, displaying detection accuracy over  $96\\%$  for texts generated by Flan-T5-Large, GPT-4Turbo and Gemma-7b.",
+        "bbox": [
+            115,
+            323,
+            489,
+            595
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "From our initial experimentation, we noted that the baseline DeBERTa model outperforms other approaches in binary machine-generated text detection; therefore, for the remaining experiments we proceed with baseline DeBERTa.",
+        "bbox": [
+            112,
+            598,
+            489,
+            677
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/d92fc0ebb78bea94c3fe7bd55df9bf00f22d404892116761705113d7d92ace7f.jpg",
+        "table_caption": [
+            "Table 2: Accuracy and F1-score for the baseline Contra-X, DeBERTa, and the proposed DeB-Ang model on various TuringBench datasets containing texts generated by different NLG models (rows). Min-Max refers to the minimum and maximum classification accuracy that DeB-Ang obtained for each dataset. Statistical significance (SS) between baseline DeBERTa and DeB-Ang is either yes (Y) or no (N) according to McNemar's test."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td colspan=\"6\">Accuracy and F1-score</td></tr><tr><td></td><td colspan=\"2\">Baseline DeBERTa</td><td colspan=\"2\">DeB-Ang</td><td>SS (Y/N)</td><td>Min-Max Improvement</td></tr><tr><td>Flan-T5-Large</td><td>92.14</td><td>92.14</td><td>96.99</td><td>96.99</td><td>Y</td><td>4.85</td></tr><tr><td>Gemma-7b</td><td>99.96</td><td>99.96</td><td>99.98</td><td>99.99</td><td>N</td><td>0.02</td></tr><tr><td>GPT4-Turbo</td><td>72.14</td><td>72.14</td><td>99.94</td><td>99.97</td><td>Y</td><td>27.80</td></tr></table>",
+        "bbox": [
+            115,
+            690,
+            485,
+            759
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results for the DAIGT-V2 dataset can be seen in Table 5. This improvement demon",
+        "bbox": [
+            112,
+            887,
+            489,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/f310a59da0d272d79575dc5b3ffb81de19fbbb696c7aec32dc0cfedfc1260361.jpg",
+        "table_caption": [
+            "Table 3: Accuracy and F1-score for the baseline DeBERTa and the proposed DeB-Ang model for TuringExtended. Min-Max refers to the minimum and maximum classification accuracy that DeB-Ang obtained for each dataset. Statistical significance (SS) between baseline DeBERTa and DeB-Ang is either yes (Y) or no (N) according to McNemar's test."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td></td><td>Accuracy</td><td>F1</td><td>SS (Y/N)</td></tr><tr><td>Syntax-CNN</td><td>66.13</td><td>64.80</td><td>Y</td></tr><tr><td>BERT-AA</td><td>78.12</td><td>77.58</td><td>Y</td></tr><tr><td>Contra-X</td><td>80.73</td><td>80.54</td><td>Y</td></tr><tr><td>Baseline DeBERTa</td><td>77.71</td><td>77.56</td><td>Y</td></tr><tr><td>GPT-who</td><td>65.89</td><td>65.39</td><td>Y</td></tr><tr><td>DeB-Ang</td><td>83.61</td><td>82.68</td><td>-</td></tr></table>",
+        "bbox": [
+            510,
+            319,
+            880,
+            413
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "Table 4: Accuracy and F1 for the authorship attribution (AA) dataset from TuringBench (Uchendu et al., 2021) comparing various AA approaches. McNemar's test was conducted to see if the result between DeB-Ang and all other models is statistically significant (SS) or not.",
+        "bbox": [
+            507,
+            423,
+            882,
+            508
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "strates the models' generalisability across various NLG datasets, for both older and newer models. Uchendu et al. (2021) comments \"No one size fits all\" in their study as they used several models on these datasets and found that different models obtain different levels of performance, depending on the dataset. However, as presented in Table 2, it is clear that the model consistently outperforms our baseline models on all datasets.",
+        "bbox": [
+            507,
+            537,
+            882,
+            681
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "5.2 Authorship Attribution",
+        "text_level": 1,
+        "bbox": [
+            507,
+            697,
+            739,
+            713
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "In assessing the generalisability of the DeB-Ang approach on various text classification settings, we present the following authorship attribution tasks:",
+        "bbox": [
+            507,
+            720,
+            880,
+            766
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "list",
+        "sub_type": "text",
+        "list_items": [
+            "1. Authorship attribution for human and machine-generated text detection.",
+            "2. Authorship attribution for model variation detection, e.g. differentiating between GPT-3.5 and GPT-4.",
+            "3. Authorship attribution for model developer detection, e.g. OpenAI for GPT-4 and GPT-3.5."
+        ],
+        "bbox": [
+            522,
+            771,
+            882,
+            897
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "text",
+        "text": "The results for each task is presented in Table 5",
+        "bbox": [
+            527,
+            902,
+            880,
+            917
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "page_number",
+        "text": "7194",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 5
+    },
+    {
+        "type": "table",
+        "img_path": "images/14a910bd47273629b5956275bbf1f0c77d13611ca1593e7bcc9ff83bc4cb8f2d.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td colspan=\"2\"></td><td colspan=\"2\">Baseline DeBERTa</td><td colspan=\"2\">DeB-Ang</td><td rowspan=\"2\">Accuracy Improvement</td></tr><tr><td>Task</td><td>Model</td><td>Accuracy</td><td>F1</td><td>Accuracy</td><td>F1</td></tr><tr><td colspan=\"2\">Machine-generated text detection</td><td>82.69</td><td>90.53</td><td>91.36</td><td>92.31</td><td>8.66</td></tr><tr><td colspan=\"2\">Authorship attribution [37]</td><td>86.00</td><td>85.96</td><td>87.80</td><td>87.79</td><td>1.80</td></tr><tr><td rowspan=\"5\">Model detection</td><td>Open AI [10]</td><td>88.64</td><td>88.64</td><td>91.75</td><td>91.75</td><td>3.11</td></tr><tr><td>Meta [13]</td><td>42.27</td><td>42.29</td><td>47.95</td><td>47.51</td><td>5.68</td></tr><tr><td>Google [7]</td><td>56.96</td><td>56.55</td><td>57.60</td><td>57.78</td><td>0.64</td></tr><tr><td>Anthropic [2]</td><td>95.63</td><td>95.58</td><td>99.03</td><td>99.03</td><td>3.40</td></tr><tr><td>Mistral [4]</td><td>93.15</td><td>93.39</td><td>93.96</td><td>95.12</td><td>0.81</td></tr><tr><td>Developer detection</td><td>All [5]</td><td>89.78</td><td>89.78</td><td>92.98</td><td>92.98</td><td>3.20</td></tr></table>",
+        "bbox": [
+            115,
+            80,
+            848,
+            223
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "Table 5: Table presenting evaluation results on the DAIGT-V2 dataset, including authorship attribution scores for all NLG models, machine-generated text detection (human vs. machine), model detection (distinguishing between different model variations), and authorship attribution for model developers. The numbers in brackets (e.g., \"Open AI [10]\") indicate the number of classes (i.e., the number of models).",
+        "bbox": [
+            112,
+            231,
+            882,
+            291
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "under authorship attribution, model detection and developer detection, respectively.",
+        "bbox": [
+            112,
+            315,
+            485,
+            347
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "From Table 4, it is evident that our approach also surpasses prior attempts on the TuringBench dataset. As previously mentioned, this dataset consists of texts generated by 20 different authors (a total of 200K texts from 19 NLG models and 1 human author) with high topical dissimilarity between each model. This dissimilarity is expected as the dataset was generated in certain topic set (Ai et al., 2022; Uchendu et al., 2021). Results for Syntax-CNN were taken from Ai et al. (2022) and all experiments were run using the full dataset. For the DAIGT-V2 dataset, we downsized the data to approximately 10K rows per model. This reduction was necessitated due to the dataset's size, which demanded significant computational resources. In Table 5, it can be observed that the DeB-Ang model outperforms the baseline DeBERTa model with an accuracy improvement of  $1.80\\%$  in authorship attribution and  $8.66\\%$  in machine-generated text detection. To delve deeper into the machine-generated text results from the baseline DeBERTa, we conducted an analysis focussing on the disparity between the accuracy and F1-score. This involved computing the Area Under the Receiver Operating Characteristic (AUROC) score and assessing misclassification. Our analysis revealed that the model exhibited a considerable number of false positives, incorrectly predicting a majority of human-written texts. The AUROC score was determined to be 50.12 whereas the AUROC score for DeB-Ang was 88.14 indicating DeB-Ang's superior discrimination capabilities. We also address the previously mentioned limitation regarding the scarcity of research in classifying models from a single developer; our results are provided in Table 5. We investigate a range",
+        "bbox": [
+            112,
+            355,
+            489,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "of developers and models varying from older to newer model versions. We were able to improve results from baseline DeBERTa for this task by  $0.64\\%$  to  $5.68\\%$ . The low accuracy observed for Meta and Google models can be attributed to the high similarity between the model variations used, e.g., Llama-2-7b and Llama-2-13b. This makes distinguishing between these version challenging, leading to misclassification. Further investigation is necessary to comprehensively understand the reasons for misclassification. We were also able to classify generated texts according to model developer with an accuracy as high as  $92.98\\%$ .",
+        "bbox": [
+            507,
+            315,
+            884,
+            524
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "As mentioned in prior research, classifying a range of outputs, e.g. texts with high topic variation, is an increasingly difficult classification task (Uchendu et al., 2021; Juola, 2008). Furthermore, it is important to note that TuringBench consists of texts from multiple sources. Additionally, some models generate many texts; this can decrease performance as there can be semantics and stylistic overlap between generated texts. This similarity between texts can blur distinctions thus, reducing classification capabilities.",
+        "bbox": [
+            507,
+            526,
+            884,
+            703
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "5.3 Assessing loss functions",
+        "text_level": 1,
+        "bbox": [
+            507,
+            719,
+            741,
+            734
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "To assess the significance of the loss functions used, we investigated various combinations of loss functions on the multi-class authorship attribution and binary machine-generated text detection tasks. The results are presented in Table 6. We provide the accuracy, F1-score and AUROC scores for these tasks obtained by the DeB-Ang model. We ran each experiment for one epoch for initial benchmarking to assess each models performance. This allowed us to identify the which approaches we would use for further investigations. We identified the optimal",
+        "bbox": [
+            507,
+            741,
+            884,
+            917
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "page_number",
+        "text": "7195",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 6
+    },
+    {
+        "type": "text",
+        "text": "parameter combination for the loss functions for each task and re-ran the experiment for 8 epochs. The aim of this investigation is to assess the performance improvement resulting from the various loss functions. This also highlights the customisability of the model. We extend the metrics by adding the AUROC score as this metric considers the tradeoff between precision and recall (McDermott et al., 2024).",
+        "bbox": [
+            112,
+            84,
+            489,
+            227
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "table",
+        "img_path": "images/255bea9a0af00435563514bb444f68f13baa1a399e0eebf8e663c68a15211868.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Task</td><td>Loss function</td><td>Accuracy</td><td>F1</td><td>AUROC</td></tr><tr><td rowspan=\"8\">Authorship Attribution</td><td>ANG</td><td>86.37</td><td>86.42</td><td>97.88</td></tr><tr><td>CL</td><td>86.72</td><td>86.63</td><td>98.11</td></tr><tr><td>CE</td><td>86.87</td><td>86.71</td><td>98.00</td></tr><tr><td>ANG + CE</td><td>86.51</td><td>86.54</td><td>98.12</td></tr><tr><td>CL + CE</td><td>86.72</td><td>86.78</td><td>98.07</td></tr><tr><td>ANG + CL</td><td>87.02</td><td>87.08</td><td>98.12</td></tr><tr><td>ANG + CE + CL (1.0; 1.0; 1.0)</td><td>86.35</td><td>86.41</td><td>97.97</td></tr><tr><td>ANG + CE + CL (1.0; 0.25; 1.0)</td><td>87.13</td><td>87.22</td><td>98.15</td></tr><tr><td>8 epoch</td><td>ANG + CE + CL (1.0; 0.25; 1.0)</td><td>88.8</td><td>88.06</td><td>98.5</td></tr><tr><td rowspan=\"8\">Binary machine-generated text detection</td><td>ANG</td><td>94.49</td><td>94.86</td><td>95.61</td></tr><tr><td>CL</td><td>88.15</td><td>89.19</td><td>72.58</td></tr><tr><td>CE</td><td>91.69</td><td>92.33</td><td>92.83</td></tr><tr><td>ANG + CE</td><td>90.87</td><td>91.89</td><td>77.00</td></tr><tr><td>CL + CE</td><td>91.6</td><td>92.47</td><td>90.89</td></tr><tr><td>CL + ANG</td><td>94.18</td><td>94.61</td><td>95.61</td></tr><tr><td>ANG + CE + CL (1.0; 1.0; 1.0)</td><td>90.76</td><td>91.92</td><td>87.57</td></tr><tr><td>ANG + CE + CL (1.0; 0.75; 0.75)</td><td>82.69</td><td>90.53</td><td>78.71</td></tr><tr><td>8 epoch</td><td>ANG + CE + CL (1.0; 0.75; 0.75)</td><td>93.76</td><td>88.23</td><td>90.94</td></tr></table>",
+        "bbox": [
+            115,
+            241,
+            485,
+            488
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Table 6: Comparison of single and combined loss functions for authorship attribution and binary machine-generated text detection using the DeBERTa-based model with varying numbers of epochs. Parameter values for all loss functions were set to 1.0 unless otherwise specified. Key: AUROC = area under the receiver operating characteristic, CE = cross-entropy loss, CL = contrastive loss, and ANG = angular loss. The values in brackets refer to the parameter values.",
+        "bbox": [
+            112,
+            499,
+            489,
+            627
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We found that a certain loss function combination may ascertain significant results at one epoch given a simple model. However, once the model or dataset complexity increases then a different loss combination would be more appropriate. Angular loss has the advantage of learning embeddings such that similar samples have a smaller angular separation. It is vital to understand that angular loss focusses on learning embeddings (Wang et al., 2017) whereas cross-entropy focusses on measuring the dissimilarity between predicted and true probability distribution of classes (Teahan, 2000). This difference may account for the accuracy difference. It is vital to note that each loss function has a different contribution and the combination of these, without careful tuning, may lead to sub-optimal results.",
+        "bbox": [
+            112,
+            645,
+            489,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "5.4 Analysing the misclassified data",
+        "text_level": 1,
+        "bbox": [
+            509,
+            84,
+            806,
+            99
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "For our error analysis, 100 instances of incorrect and correct classifications were extracted for the binary classification task. We found that texts were being labelled as machine-generated more frequently than human data; this could be due to the class imbalance or due to the NLG model's ability to create human-like text. While F1-scores account for class imbalance, an imbalanced training set can still bias the model toward the majority class.",
+        "bbox": [
+            507,
+            126,
+            884,
+            285
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "Based on the manual analysis, there was no specific linguistic category which would clearly lead to the misclassification. Therefore, we extracted features from varying categories (see Table 9 in Appendix C). A total of 250 features were extracted. 100 random features were sampled and the raw counts and mean for each feature was plotted (see Figure 2 in Appendix C). From this plot, it is evident that there is a clear discrepancy in feature usage. The correctly classified data points exhibit lower feature counts and an overall lower mean whereas the incorrectly classified data is slightly more sporadic and exhibits an overall higher mean. The statistical significance for these differences for all features was computed using the Mann-Whitney U test (Nachar, 2008) as the data was not normally distributed (as affirmed by the Shapiro-Wilk test) (Aryadoust and Raquel, 2020). The statistical significance was less than 0.05 thus rejecting the null hypothesis and confirming the difference between the feature counts and mean for the correctly classified and incorrectly classified data is significant.",
+        "bbox": [
+            507,
+            297,
+            884,
+            651
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "We then measured the semantic similarity between correctly classified and incorrectly classified instances using contextual embeddings obtained using DeBERTa. The mean similarity score for all data points is 85.54 (minimum score of 63.22 and maximum score of 92.64). Figure 3 in Appendix C presents a correlation coefficient of -0.03 indicating a very weak negative linear relationship almost suggesting no linear relationship between the data points. This indicates that any observed differences or similarities in the similarity score are likely due to random variation and not a meaningful underlying relationship. We conclude that the similarity scores do not provide useful information to distinguish between correctly and incorrectly classified instances.",
+        "bbox": [
+            507,
+            661,
+            884,
+            917
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "page_number",
+        "text": "7196",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 7
+    },
+    {
+        "type": "text",
+        "text": "6 Conclusion and Future Work",
+        "text_level": 1,
+        "bbox": [
+            112,
+            84,
+            401,
+            98
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "In this research, we have created a custom DeBERTa model integrating contrastive and angular loss. To our knowledge, this is one of the first attempts at this integration and we have demonstrated the success of the proposed DeB-Ang model on several datasets. We investigated more fine-grained machine-generated text detection by classifying model variations and developers. We were able to outperform prior approaches in machine-generated text detection with a minimum improvement of  $0.23\\%$  and a maximum improvement of  $38.04\\%$  across all datasets. We were able to classify model variations with accuracy scores ranging from  $0.64\\%$  to  $5.68\\%$ , and to identify developers with an accuracy improvement of  $3.20\\%$ . For authorship attribution, we were able to improve classification with a maximum accuracy of  $17.48\\%$  on the extensive TuringBench dataset which is characterised by high topical dissimilarity. Future work will involve identifying texts in which multiple NLG models or humans have been used to intentionally mask the writing style of a text. Additionally, a more extensive examination of linguistic features of synthetic data across generations of LLMs can provide insights into language evolution these models.",
+        "bbox": [
+            112,
+            111,
+            492,
+            513
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "7 Limitations",
+        "text_level": 1,
+        "bbox": [
+            112,
+            527,
+            250,
+            542
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Guerrero and Alsmadi (2022) lists several research gaps in the field of machine-generated text detection e.g. domain-specific text detection. It would be interesting to investigate texts that are cross-domain, genre or multimodal. Further, we investigated misclassified instances but did not use this information to improve the model due to time constraints. The limitations associated with data generation are model-related. Data generation is a time-consuming process and requires many computational resources; we were only able to extend our evaluation data with three datasets.",
+        "bbox": [
+            112,
+            555,
+            489,
+            747
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "Ethics Statement",
+        "text_level": 1,
+        "bbox": [
+            112,
+            763,
+            265,
+            778
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "The materials used for this study did not require human participation and the data does not contain any harmful or sensitive information. The datasets used in this study were acquired from prior research. The dataset generated using NLG models (Open AI's GPT-4 model, Gemma-7b and Flan-T5-large) was evaluated to ensure that there is no overtly harmful text. Data was annotated and evaluated by",
+        "bbox": [
+            112,
+            790,
+            489,
+            919
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "PhD students, the task was explained in regards to how data will be used and proposed tasks. Nevertheless, the potential negative use of this research should not be ignored. The insights provided by this work have the potential to be exploited for malicious purposes, potentially undermining the effectiveness of these detectors. However, we hope that this research will be used to support the efforts in detecting neural machine-generated used in applications with malicious intent.",
+        "bbox": [
+            507,
+            84,
+            884,
+            244
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "text",
+        "text": "References",
+        "text_level": 1,
+        "bbox": [
+            510,
+            272,
+            608,
+            286
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Ahmed Abbasi and Hsinchun Chen. 2008. Writeprints: A stylometric approach to identity-level identification and similarity detection in cyberspace. ACM Transactions on Information Systems, 26(2):1-29.",
+            "David Ifeoluwa Adelani, Haotian Mai, Fuming Fang, Huy H. Nguyen, Junichi Yamagishi, and Isao Echizen. 2019. Generating Sentiment-Preserving Fake Online Reviews Using Neural Language Models and Their Human- and Machine-based Detection.",
+            "Bo Ai, Yuchen Wang, Yugin Tan, and Samson Tan. 2022. Whodunit? Learning to Contrast for Authorship Attribution.",
+            "Rahaf Aljundi, Yash Patel, Milan Sulc, Daniel Olmeda, and Nikolay Chumerin. 2022. Contrastive Classification and Representation Learning with Probabilistic Interpretation.",
+            "Vahid Aryadoust and Michelle Raquel. 2020. Quantitative data analysis for language assessment. Volume II, Advanced methods. Routledge research in language education. Routledge, Abingdon, Oxon ;",
+            "Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, MarcAurelio Ranzato, and Arthur Szlam. 2019. Real or Fake? Learning to Discriminate Machine from Human Generated Text. arXiv.",
+            "Hongjun Choi, Anirudh Som, and Pavan Turaga. 2020. AMC-Loss: Angular Margin Contrastive Loss for Improved Explainability in Image Classification.",
+            "Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling Instruction-Finetuned Language Models.",
+            "Michael T. Cox. 2005. Metacognition in computation: A selected research review. Artificial Intelligence, 169(2):104-141. Special Review Issue."
+        ],
+        "bbox": [
+            509,
+            294,
+            885,
+            917
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "page_number",
+        "text": "7197",
+        "bbox": [
+            480,
+            927,
+            519,
+            940
+        ],
+        "page_idx": 8
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Evan N. Crothers, Nathalie Japkowicz, and Herna L. Viktor. 2023. Machine-Generated Text: A Comprehensive Survey of Threat Models and Detection Methods. IEEE Access, 11:70977-71002.",
+            "Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699.",
+            "Liam Dugan, Daphne Ippolito, Arun Kirubarajan, and Chris Callison-Burch. 2020. RoFT: A tool for evaluating human detection of machine-generated text. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 189-196, Online. Association for Computational Linguistics.",
+            "Liam Dugan, Eleni Miltsakaki, Shriyash Upadhyay, Etan Ginsberg, Hannah Gonzalez, DaHyeon Choi, Chuning Yuan, and Chris Callison-Burch. 2022. A feasibility study of answer-agnostic question generation for education. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1919-1926, Dublin, Ireland. Association for Computational Linguistics.",
+            "Mael Fabien, Esau Villatoro-Tello, Petr Moticek, and Shantipriya Parida. 2020. BertAA: BERT fine-tuning for authorship attribution. In Proceedings of the 17th International Conference on Natural Language Processing (ICON), pages 127-137, Indian Institute of Technology Patna, Patna, India. NLP Association of India (NLPAI).",
+            "Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. 2021. Tweep-Fake: About detecting deepfake tweets. PLOS ONE, 16(5):e0251415.",
+            "Shuai Gao. 2022. Système de traduction automatique neuronale français-mongol (historique, mise en place et évaluations) (French-Mongolian neural machine translation system (history, implementation, and evaluations) machine translation (hereafter abbreviated MT) is currently undergoing rapid development, during which less-resourced languages nevertheless seem to be less developed). In Actes de la 29e Conférence sur le Traitément Automatique des Langues Naturelles. Volume 2: 24e Rencontres Etudiants Chercheurs en Informatique pour le TAL (RECITAL), pages 97-110, Avignon, France. ATALA.",
+            "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2022. SimCSE: Simple Contrastive Learning of Sentence Embeddings.",
+            "Tanya Goyal, Junyi Jessy Li, and Greg Durrett. 2022. News summarization and evaluation in the era of gpt-3. arXiv preprint arXiv:2209.12356.",
+            "Jesus Guerrero and Izzat Alsmadi. 2022. Synthetic Text Detection: Systemic Literature Review."
+        ],
+        "bbox": [
+            115,
+            85,
+            485,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "R. Hadsell, S. Chopra, and Y. LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR06), volume 2, pages 1735-1742.",
+            "Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTa: Decoding-enhanced BERT with Disentangled Attention.",
+            "Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. 2024. MGTBench: Benchmarking Machine-Generated Text Detection.",
+            "Beizhe Hu, Qiang Sheng, Juan Cao, Yuhui Shi, Yang Li, Danding Wang, and Peng Qi. 2023. Bad actor, good advisor: Exploring the role of large language models in fake news detection. arXiv preprint arXiv:2309.12247.",
+            "Le Hui, Xiang Li, Chen Gong, Meng Fang, Joey Tianyi Zhou, and Jian Yang. 2019. Inter-class angular loss for convolutional neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3894-3901.",
+            "D. Ippolito, D. Duckworth, C. Callison-Burch, and D. Eck. 2020. Automatic detection of generated text is easiest when humans are fooled. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1808-1822. Association for Computational Linguistics.",
+            "Niful Islam, Debopom Sutradhar, Humaira Noor, Jarin Tasnim Raya, Monowara Tabassum Maisha, and Dewan Md Farid. 2023. Distinguishing Human Generated Text From ChatGPT Generated Text Using Machine Learning.",
+            "Vivek Iyer, Pinzhen Chen, and Alexandra Birch. 2023. Towards Effective Disambiguation for Machine Translation with Large Language Models. arXiv preprint arXiv:2309.11668.",
+            "Maurice Jakesch, Jeffrey T. Hancock, and Mor Naaman. 2023. Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11):e2208839120.",
+            "Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks V. S. Lakshmanan. 2020. Automatic Detection of Machine Generated Text: A Critical Survey.",
+            "Patrick Juola. 2008. Authorship attribution. Foundations and Trends® in Information Retrieval, 1:233-334.",
+            "Virapat Kieuvongngam, Bowen Tan, and Yiming Niu. 2020. Automatic text summarization of Covid-19 medical research articles using bert and gpt-2. arXiv preprint arXiv:2006.01997.",
+            "Taehyeon Kim, Eungi Hong, and Yoonsik Choe. 2021. Deep Morphological Anomaly Detection Based on Angular Margin Loss. Applied Sciences, 11(14):6545."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "page_number",
+        "text": "7198",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 9
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Taehyeon Kim, Seho Park, and Kyoungtaek Lee. 2023. Traffic Sign Recognition Based on Bayesian Angular Margin Loss for an Autonomous Vehicle. *Electronics*, 12(14):3073.",
+            "Anqi Mao, Mehryar Mohri, and Yutao Zhong. 2023. Cross-Entropy Loss Functions: Theoretical Analysis and Applications.",
+            "Colin Martindale and Dean McKenzie. 1995. On the utility of content analysis in author attribution: The Federalist. Computers and the Humanities, 29:259-270.",
+            "Matthew B. A. McDermott, Lasse Hyldig Hansen, Hao ran Zhang, Giovanni Angelotti, and Jack Gallifant. 2024. A Closer Look at AUROC and AUPRC under Class Imbalance.",
+            "Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature.",
+            "Sandra Mitrovic, Davide Andreoletti, and Omran Ayoub. 2023. ChatGPT or Human? Detect and Explain. Explaining Decisions of Machine Learning Model for Detecting Short ChatGPT-generated Text. ArXiv.",
+            "Nadim Nachar. 2008. The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution. Tutorials in Quantitative Methods for Psychology, 4(1):13-20.",
+            "OpenAI. 2023. GPT-4 Technical Report.",
+            "Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library.",
+            "Dongqi Pu and Vera Demberg. 2023. ChatGPT vs Human-authored Text: Insights into Controllable Text Summarization and Sentence Style Transfer.",
+            "Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. 2022. Deepfake Text Detection: Limitations and Opportunities.",
+            "Junaid Qadir. 2022. Engineering Education in the Era of ChatGPT: Promise and Pitfalls of Generative AI for Education. TechRxiv.",
+            "Yunita Sari. 2018. Neural and non-neural approaches to authorship attribution. Ph.D. thesis, University of Sheffield, UK. British Library, EThOS."
+        ],
+        "bbox": [
+            115,
+            85,
+            487,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang. 2019. Release Strategies and the Social Impacts of Language Models.",
+            "JH Sundjaja, R Shrestha, and K Krishan. 2023. McNemar And Mann-Whitney U Tests. StatPearls [Internet]. Updated 2023 Jul 17.",
+            "Zhiquan Tan, Yifan Zhang, Jingqin Yang, and Yang Yuan. 2024. Contrastive Learning Is Spectral Clustering On Similarity Graph.",
+            "Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The Science of Detecting LLM-Generated Texts.",
+            "Yi Tay, Dara Bahri, Che Zheng, Clifford Brunk, Donald Metzler, and Andrew Tomkins. 2020. Reverse Engineering Configurations of Neural Text Generation Models.",
+            "William John Teahan. 2000. Text classification and segmentation using minimum cross-entropy. In Content-Based Multimedia Information Access-Volume 2, pages 943-961.",
+            "Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Pier Giuseppe Sessa, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open Models Based on Gemini Research and Technology."
+        ],
+        "bbox": [
+            510,
+            85,
+            880,
+            917
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "page_number",
+        "text": "7199",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 10
+    },
+    {
+        "type": "list",
+        "sub_type": "ref_text",
+        "list_items": [
+            "Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. 2020. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384-8395, Online. Association for Computational Linguistics.",
+            "Adaku Uchendu, Zeyu Ma, Thai Le, Rui Zhang, and Dongwon Lee. 2021. TURINGBENCH: A benchmark environment for Turing test in the age of neural text generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2001-2016, Punta Cana, Dominican Republic. Association for Computational Linguistics.",
+            "Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2024. GPT-who: An information density-based machine-generated text detector. In *Findings of the Association for Computational Linguistics: NAACL* 2024, pages 103–115, Mexico City, Mexico. Association for Computational Linguistics.",
+            "Roman Vygon and Nikolay Mikhaylovskiy. 2021. Learning efficient representations for keyword spotting with triplet loss. In Speech and Computer: 23rd International Conference, SPECOM 2021, St. Petersburg, Russia, September 27-30, 2021, Proceedings 23, pages 773-785. Springer.",
+            "Jian Wang, Feng Zhou, Shilei Wen, Xiao Liu, and Yuanqing Lin. 2017. Deep Metric Learning with Angular Loss.",
+            "Rowan Zellers, Ari Holtzman, Matthew Peters, Roozbeh Mottaghi, Aniruddha Kembhavi, Ali Farhadi, and Yejin Choi. 2021. SIGLeT: Language grounding through neuro-symbolic interaction in a 3D world. 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 2040-2050, Online. Association for Computational Linguistics.",
+            "Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating Text Generation with BERT."
+        ],
+        "bbox": [
+            115,
+            85,
+            489,
+            684
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "page_number",
+        "text": "7200",
+        "bbox": [
+            480,
+            928,
+            519,
+            940
+        ],
+        "page_idx": 11
+    },
+    {
+        "type": "text",
+        "text": "Appendix A Dataset model breakdown",
+        "text_level": 1,
+        "bbox": [
+            114,
+            84,
+            465,
+            99
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/f0fed1e1936ee245f01cbd93e201a4e0aa1442b3483fcbb323af73fefde75f64.jpg",
+        "table_caption": [],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Dataset</td><td>Source</td><td>Language Models</td><td>Link</td></tr><tr><td>TuringBench</td><td>TuringBench</td><td>GPT-1, GPT-2, GPT-3, GROVER, CTRL, XLM, FAIR, Transformer_XL, XLNET, PPLM</td><td>TuringBench</td></tr><tr><td>DAIGT-V2</td><td>Kaggle</td><td>LlaMa2, Darragh_Claude, Mistral7binstruct, Gemma, opt</td><td>DAIGT-V2</td></tr><tr><td>TuringExtended</td><td>Github</td><td>Gemma, Flan-T5, GPT</td><td>To be added</td></tr></table>",
+        "bbox": [
+            115,
+            120,
+            882,
+            171
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Appendix B Hyperparameter settings for the DeBERTa model",
+        "text_level": 1,
+        "bbox": [
+            112,
+            250,
+            665,
+            269
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/9dbec9dc27a7b24e2a560534956a27530debf23a6dfc744136778e0631818a44.jpg",
+        "table_caption": [
+            "Table 7: Overview of datasets utilized in the study, detailing dataset name, source, and the language models used to generate text. Note: While not exhaustive, datasets may encompass various iterations of a single model (e.g., LlaMa-7b and Llama-13b)."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Hyperparameter</td><td>Amended value</td></tr><tr><td>num_train_epochs</td><td>1 - 8</td></tr><tr><td>train_batch_size</td><td>16</td></tr><tr><td>eval_batch_size</td><td>16</td></tr><tr><td>gradient Accumulation_steps</td><td>4</td></tr><tr><td>ngpu</td><td>1</td></tr><tr><td>max_seq_length</td><td>512</td></tr><tr><td>class_weight</td><td>Equal weighting specified</td></tr><tr><td>early_stopping_patience</td><td>2</td></tr><tr><td>early_stopping deltas</td><td>0.01</td></tr><tr><td>contrastive_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr><tr><td>angular_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr><tr><td>crossentropy_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr></table>",
+        "bbox": [
+            238,
+            288,
+            759,
+            507
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Appendix C Error analysis: linguistic analysis",
+        "text_level": 1,
+        "bbox": [
+            112,
+            573,
+            529,
+            590
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C.1 Linguistic features extracted",
+        "text_level": 1,
+        "bbox": [
+            112,
+            599,
+            388,
+            614
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "table",
+        "img_path": "images/ba1b0160f2e3c6f915f5e40609f5788b98ee16709adfc57a232a964e5c8cd395.jpg",
+        "table_caption": [
+            "Table 8: The hyperparameters used in training the DeB-Ang model. Parameter values for the epochs and loss functions varied and the specific values used are detailed in Section 5."
+        ],
+        "table_footnote": [],
+        "table_body": "<table><tr><td>Linguistic category</td><td>Feature</td></tr><tr><td rowspan=\"3\">Character-level features</td><td>uni, bi and tri-grams, word length distribution,</td></tr><tr><td>number counts, text length, average word</td></tr><tr><td>sentence length, casing, type-token ratio</td></tr><tr><td>Syntactic features</td><td>Part-of-speech tags, dependency tags</td></tr><tr><td>Word-level features</td><td>Function words</td></tr></table>",
+        "bbox": [
+            211,
+            631,
+            786,
+            734
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "Table 9: Linguistic features extracted from the correctly and incorrectly classified texts for the task of binary machine-generated text detection.",
+        "bbox": [
+            112,
+            744,
+            882,
+            772
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "text",
+        "text": "C.2 Linguistic feature groups",
+        "text_level": 1,
+        "bbox": [
+            112,
+            806,
+            361,
+            822
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "page_number",
+        "text": "7201",
+        "bbox": [
+            480,
+            928,
+            517,
+            940
+        ],
+        "page_idx": 12
+    },
+    {
+        "type": "image",
+        "img_path": "images/be889168e090dc108fb87e2a3292b3c5f1d74ad3998e9c3dfeead4f04fec50e8.jpg",
+        "image_caption": [
+            "Figure 2: Scatterplot displaying the raw counts and mean feature usage of incorrectly and correctly classified samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            196,
+            129,
+            801,
+            381
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "image",
+        "img_path": "images/394b9bce509798523fef1a069cbe5e44fbacd35cc8cdafc365a5337f78f92891.jpg",
+        "image_caption": [
+            "Figure 3: Scatterplot displaying the similarity scores between each correctly and incorrectly classified data samples."
+        ],
+        "image_footnote": [],
+        "bbox": [
+            191,
+            527,
+            796,
+            843
+        ],
+        "page_idx": 13
+    },
+    {
+        "type": "page_number",
+        "text": "7202",
+        "bbox": [
+            480,
+            928,
+            521,
+            940
+        ],
+        "page_idx": 13
+    }
+]

2024/Multi-Loss Fusion_ Angular and Contrastive Integration for Machine-Generated Text Detection/e63558dd-1191-4591-a960-30434ec75c28_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed2f5eb01848028022a75ab893e3107c12c2f380d126d366ecc440b2c659b539
+size 518742

2024/Multi-Loss Fusion_ Angular and Contrastive Integration for Machine-Generated Text Detection/full.md

@@ -0,0 +1,341 @@
+# Multi-Loss Fusion: Angular and Contrastive Integration for Machine-Generated Text Detection
+
+Iqra Zahid, Yue Chang, Tharindu Madusanka, Youcheng Sun and Riza Batista-Navarro
+
+Department of Computer Science, University of Manchester
+
+{iqra.zahid|yue.chang|tharindu.batawalaacharige}@manchester.ac.uk, {youcheng.sun|riza.batista}@manchester.ac.uk
+
+# Abstract
+
+Modern natural language generation (NLG) systems have led to the development of synthetic human-like open-ended texts, posing concerns as to who the original author of a text is. To address such concerns, we introduce DeB-Ang: the utilisation of a custom DeBERTa model (He et al., 2021) with angular loss and contrastive loss functions for effective class separation in neural text classification tasks. We expand the application of this model on binary machine-generated text detection and multi-class neural authorship attribution. We demonstrate improved performance on many benchmark datasets whereby the accuracy for machine-generated text detection was increased by as much as  $38.04\%$  across all datasets.
+
+# 1 Introduction
+
+There has been considerable activity in the field of detecting machine-generated text. Driven by the significant growth and the increasing prevalence of large language models (LLMs) and natural language generation (NLG) models. This has led to the production of high-quality human-like texts that have brought about useful applications in many domains such as machine translation, text summarisation and data generation (Kieuvongngam et al., 2020; Goyal et al., 2022; Iyer et al., 2023; Uchendu et al., 2021). Irrespective of the many useful applications, the deployment of NLG models has concurrently given rise to serious concerns such as plagiarism, and spreading misinformation and hate speech (Pu et al., 2022; Hu et al., 2023; Qadir, 2022; Solaiman et al., 2019). Therefore, the need to discriminate between human and machine-generated text becomes paramount, especially in light of the growing sophistication and rapid updates of these models.
+
+Given the diverse applications of NLG models, authorship attribution (AA) methods have been increasingly employed to detect the original author
+
+of synthetic data generated by machines (Ai et al., 2022; Uchendu et al., 2020; Jawahar et al., 2020). The main concern with traditional AA methods is that, typically, they are feature-based systems and consist of largely document-specific features. Therefore, the application of this traditional model is often author, dataset and model-specific (Sari, 2018; Ai et al., 2022). Previous research addressed the need for generalisable detection systems to identify machine-generated text (Fagni et al., 2021; Jakesch et al., 2023; He et al., 2024; Jawahar et al., 2020). Research involving the use of LLMs in authorship attribution has demonstrated that the simple fine-tuning of pre-trained language models can surpass the accuracy of traditional methods significantly (Fabien et al., 2020; Mitrovic et al., 2023; Fagni et al., 2021).
+
+In particular, we introduce DeB-Ang, a pretrained DeBERTa model with a specialised angular loss and contrastive loss integration. Additionally, we demonstrate improved classification when applying DeB-Ang to several well-known machine-generated text and authorship attribution datasets. Contrastive learning is an unsupervised representation learning technique, aiming to learn a representation of data such that similar instances are close in the representation space whereas dissimilar instances are far apart (Aljundi et al., 2022). Loss functions are crucial in contrastive learning as they quantify the similarity and dissimilarity between pairs, guiding the model to learn meaningful representations for class discrimination (Hadsell et al., 2006; Gao et al., 2022; Wang et al., 2017). However, recent studies suggest that various loss functions, including cross-entropy loss, contrastive loss and triplet loss, fail to consider the intrinsic angular distribution exhibited by the low-level and high-level feature representations (Choi et al., 2020), which contributes to our choice of using angular loss in DeB-Ang. Angular loss is a scale-invariant loss function designed to improve the learning sim
+
+ality metrics by considering the angle between vectors (Wang et al., 2017).
+
+In summary, the contributions of this work are four-fold:
+
+1. We propose a novel customisable contrastive learning framework that combines a custom fine-tuned DeBERTa model (He et al., 2021) with contrastive and angular loss functions. We assess the difference in classification performance when utilising various combinations of the aforementioned loss functions for the proposed task.  
+2. We assess the application of the proposed model on multi-class authorship attribution and binary machine-generated text detection.  
+3. We introduce three new large-scale datasets for evaluating text classification models. These datasets were constructed by leveraging state-of-the-art language models, including Gemma-7b (Team et al., 2024), GPT4-Turbo (OpenAI, 2023) and Flan-T5-Large (Chung et al., 2022).  
+4. We conduct linguistic error analysis of incorrectly and correctly classified examples.
+
+# 2 Related Work
+
+# 2.1 Machine-generated text detection
+
+Studies have demonstrated that human participants were unable to distinguish between machine-generated texts and human written texts (Jakesch et al., 2023; Islam et al., 2023; Ippolito et al., 2020; Dugan et al., 2020, 2022). Previous work highlighted that disambiguating between human and LLM-generated texts is increasingly difficult (Pu and Demberg, 2023; Jakesch et al., 2023; Cox, 2005). Automatic detection of machine-generated text has thus gained popularity and can be categorised according to their underlying method (Solaiman et al., 2019; Uchendu et al., 2020; Fagni et al., 2021; Bakhtin et al., 2019; Ippolito et al., 2020). Simple classifiers often involve linguistic feature analysis (Dugan et al., 2020, 2022) or incorporate a psycholinguistics statistical measure (Venkatraman et al., 2024). Other methods include zero-shot detection (Solaiman et al., 2019), and fine-tuned model detection (Uchendu et al., 2020; Ippolito et al., 2020; Fagni et al., 2021; Adelani et al., 2019; Tay et al., 2020; Zellers et al., 2021). Irrespective of the large number of approaches to identifying machine-generated texts, detection remains a challenge (Crothers et al., 2023; Ai et al.,
+
+2022).
+
+# 2.2 Authorship Attribution
+
+Traditional attribution approaches utilise linguistic features in a univariate (utilising a single linguistic feature, e.g. function words) (Martindale and McKenzie, 1995) or multivariate (utilising multiple linguistic features, e.g. Writeprints) approach (Abbasi and Chen, 2008; Sari, 2018). As aforementioned, feature-based linguistic identification requires dataset-specific engineering, displaying limited scalability (Sari, 2018; Ai et al., 2022). More recently, the use of learning-based approaches has grown with the use of pre-trained LLMs (Fabien et al., 2020). These approaches have demonstrated the power of LLMs in significantly surpassing the accuracy of traditional approaches with little analysis required beforehand (Ai et al., 2022).
+
+# 2.3 Research gaps
+
+Existing approaches in detecting synthetic texts created by LLMs have many limitations. For example, these detection tools are now outdated due to rapid technological advancements, e.g., DetectGPT classifies texts only generated by GPT2 (Mitchell et al., 2023). This necessitates classifier retraining which could negatively affect the accuracy of these models (OpenAI, 2023). Additionally, the increased advancements of NLG models have led to more human-like texts. Further, these models are LLM-specific and therefore, do not detect synthetic texts generated by other language models. Also, these methods have a black-box nature, making it difficult for humans to understand their output for correctly and incorrectly classified texts. Given the increasing prevalence of machine-generated texts, it is vital that we are able to distinguish which NLG model was used to generate a given text. We extend this to being able to detect the exact model version.
+
+# 3 Data
+
+# 3.1 Data collection
+
+There is a strong consensus that datasets must be diverse and representative (Tang et al., 2023). To this end, we chose to utilise datasets with original, human-written texts; different versions of each text are then generated with the aid of LLMs which were given carefully designed prompts. Datasets were taken from Kaggle and the Turing Test benchmark known as TuringBench (Uchendu et al., 2021). The TuringBench dataset consists of ar
+
+ticles generated by 20 authors. There are a total of 20 datasets from 19 different NLG model versions and one human author. The DAIGT-V2 dataset consists of 37 authors (36 NLG models and one human author, with 60K texts). Further dataset details can be seen in Appendix A in Table 7. All datasets utilised were generated for the proposed tasks.
+
+Specifically, we utilised 5 randomly selected datasets from TuringBench datasets. We opted to use one dataset per model. For example, there are two datasets generated by XLNET. The exact model verions are XLNET_base and XLNET_large. Therefore, we employ only one of these model versions. We randomly sampled the dataset due to limited computational resources. Details of the specific processing steps and size of the data taken for each of the different datasets are provided in Section 5. Table 7 in Appendix A presents—for each dataset that we utilised—the dataset name, source and the models that were used to generate the texts contained in each dataset.
+
+# 3.2 Data Generation
+
+We also generated our own datasets by using GPT4-Turbo (OpenAI, 2023), Gemma-7b (Team et al., 2024) and Flan-T5-large (Chung et al., 2022). The TuringBench dataset set consists of 19 different NLG model versions however, these models are no longer considered state-of-the-art models. We decided to generate three additional datasets from more recent models which are considered to be the current state-of-the-art and were not included in the original TuringBench dataset. This dataset set is referred to as TuringExtended. This enables the examination of MGT and AA within the context of newer NLG models, underpinning the exploration as to whether newer NLG models are more challenging to identify as machine-generated.
+
+Specifically, additional datasets were generated as an extension of TuringBench (Uchendu et al., 2021). Considering only the human-written texts from the original AA dataset from TuringBench, we extracted only a total of 7678 (non-duplicated) rows of text. The models that we employed are GPT4 Turbo $^{1}$  (He et al., 2021), Gemma-7b $^{2}$  (Team et al., 2024) and Flan-T5-large $^{3}$  (Chung et al., 2022), resulting in the creation of three new
+
+<table><tr><td></td><td>GPT4-Turbo</td><td>Gemma-7b</td><td>Flan-T5-Large</td></tr><tr><td>BERTScore P</td><td>83.30</td><td>92.32</td><td>90.51</td></tr><tr><td>BERTScore R</td><td>84.11</td><td>95.55</td><td>83.98</td></tr><tr><td>BERTScore F1</td><td>83.70</td><td>93.88</td><td>87.08</td></tr><tr><td>IAA</td><td>88.64</td><td>89.30</td><td>84.64</td></tr></table>
+
+Table 1: Averaged BERTScore Precision (P), Recall (R) and F1-score (F1) for the datasets generated by the specified models. Inter-annotator agreement (IAA) is also provided.
+
+datasets. The models were given the prompt "generate a similar article", which is slightly similar to what was used in TuringBench ("generate an article similar to the human-written one").
+
+The three models for generation (GPT4-Turbo, Gemma-7b and Flan-T5-large) were chosen on the basis that they were either the current state-of-the-art models, or that they were not previously employed in creating the TuringBench datasets.
+
+# 3.3 Data Evaluation
+
+We evaluated the quality of the generated data using a combination of the automated metric BERTScore (Zhang et al., 2020) and human evaluators. BERTScore calculates token similarity using contextual embeddings to calculate the similarity between tokens in the candidate and reference text. This metric has demonstrated an advanced performance by correlating strongly with human judgement in various evaluative tasks (Zhang et al., 2020). In parallel, four human annotators were trained on evaluating generated text and were provided with some background information on text generation. Each annotator assessed 250 rows from each dataset and was asked to label the data as coherent (0) or incoherent (1). For a data sample to be labelled as coherent it had to meet two criteria: texts should be semantically and grammatically sound. Inter-annotator agreement (IAA) was then measured between all annotators for each dataset. The averaged BERTScore precision, recall and F1-scores, and IAA results are presented in Table 1.
+
+# 4 Methodology
+
+# 4.1 Loss Functions
+
+Previous studies have focussed on increasing similarity between representations by using varying loss functions (Ai et al., 2022; Vygon and Mikhaylovskiy, 2021). However, many approaches focus on the utilisation of a single loss function. In this paper, we propose a multi-loss fusion by using the weighted sum of a combination of vari
+
+ous loss functions: angular loss, cross-entropy loss and contrastive loss. Cross-entropy loss measures probability distributions; the objective is to minimise the error between the predicted probability and true distribution (Mao et al., 2023). This is used in updating the model weights during optimisation. Angular loss, often used in deep learning tasks (Wang et al., 2017; Kim et al., 2023, 2021; Choi et al., 2020), considers the angle between vectors to enhance learning for an improved similarity metric. The utilisation of angular loss in text classification often leads to more adaptable and robust models capable of handling linguistic diversity (Gao, 2022; Hui et al., 2019; Wang et al., 2017; Deng et al., 2019). Contrastive learning focusses on learning representations of data so that similar instances are closer in the embedding space and dissimilar instances are apart (Tan et al., 2024).
+
+# 4.2 Problem Statement
+
+The goal of our approach is to capture nuanced semantic representations and to effectively discriminate learned embeddings. We propose leveraging the DeBERTa model with angular and contrastive loss integration (DeB-Ang). This process aims to enhance the discriminative capabilities and quality of embeddings to improve the model's performance on downstream classification tasks.
+
+# 4.3 Implementation
+
+Our textual datasets underwent cleaning and preprocessing procedures. A 70:10:10 split was applied to partition the data into a training, validation and test sets for model evaluation.
+
+Building upon the DeBERTa base model (microsoft/deberta-base), we implemented a new model, DeB-Ang, that integrates the angular and contrastic loss into the training step. The model was implemented using PyTorch (Paszke et al., 2019) and Simple Transformers<sup>4</sup> and was configured with specific hyperparameters (See Appendix B); additionally, early stopping criteria were set to improve training efficiency.<sup>5</sup>
+
+# 4.4 Angular Loss Computation
+
+The angular loss function begins by computing the cosine similarity between all pairs of extracted embeddings. Positive and negative pairs are then
+
+generated to ensure the model can distinguish between embeddings with the same and different labels. Subsequently, we compute the loss for positive and negative pairs in order to optimise embeddings to have lower similarity for pairs with different labels and higher similarity for those with the same. This ultimately decides their position in the embedding space. The sum of positive and negative loss creates a complete loss function. This function, given below, guides the optimisation to achieve embeddings that have properties of similarity and dissimilarity.
+
+$$
+L _ {A n g u l a r} = \sum_ {i = 1} ^ {n} \log \left(\sum_ {j \neq i} e ^ {s _ {i j}}\right) - \log \left(\sum_ {j \neq i} e ^ {s _ {j i}}\right)
+$$
+
+where:
+
+-  $n$  is the number of embeddings;  
+-  $s_{ji}$  is the cosine similarity between embeddings  $i$  and  $j$ ;  
+- The first term encourages embeddings from different classes (negative pairs) to have lower cosine similarity;  
+- The second term encourages embeddings from the same class (positive pairs) to have higher cosine similarity.
+
+# 4.5 Contrastive Loss Computation
+
+Generating positive pairs facilitates the learning of intra-class relationships by allowing embeddings with the same labels but different indices to be considered for optimisation. Generating negative pairs enhances the discrimination capability of these embeddings. We compute the loss for positive and negative pairs. This allows embeddings of similar instances to be pushed closer to each other in the embedding space, whereas negative embeddings push them apart thus improving intra-class clustering and inter-class separation. Combining the positive and negative loss, as shown below, guides the model to learn embeddings that capture both intra-class relationships and inter-class distinctions.
+
+$$
+L _ {C o n t r a s t i v e} = \sum_ {i, j} y _ {i j} d _ {i j} + (1 - y _ {i j}) \max  (0, m - d _ {i j})
+$$
+
+where:
+
+-  $y_{ij}$  is a binary label indicating whether embeddings  $i$  and  $j$  belong to the same class (1) or different classes (0);
+
+-  $d_{ij}$  is the distance between embeddings  $i$  and  $j$ ;  
+-  $m$  is a margin hyperparameter;  
+- For positive pairs  $(y_{ij} = 1)$ , the loss is  $d_{ij}$ , encouraging embeddings to be closer together;  
+- For negative pairs  $(y_{ij} = 0)$ , the loss is  $\max(0, m - d_{ij})$ , encouraging embeddings to be apart by at least a distance of  $m$ .
+
+# 4.6 Our DeB-Ang Model
+
+In the DeB-Ang model, we utilise three loss functions, as shown in the equation below: angular loss, cross-entropy loss and contrastive loss. Angular loss is used to facilitate intra-class compactness and inter-class separation. Within our model, we utilise cross-entropy loss to penalise the models' misclassification by computing the difference between predicted and actual labels. Cross-entropy is the standard loss function that was incorporated into the DeBERTa model. Constractive loss enhances the embeddings' discriminative abilities by encouraging similarity for positive pairs and dissimilarity for negative pairs.
+
+$$
+\begin{array}{l} L _ {T o t a l} = w _ {C E} L _ {C E} \\ + w _ {A n g u l a r} L _ {A n g u l a r} \\ + w _ {\text {C o n t r a s t i v e}} L _ {\text {C o n t r a s t i v e}} \\ \end{array}
+$$
+
+# where:
+
+-  $L_{Total}$  is the total loss function used for training the DeBERTa model;  
+-  $L_{CE}$  is the standard cross-entropy loss for classification tasks, calculated as  $L_{CE} = -\sum_{i=1}^{n} \log P(y_i | X)$ , where  $X$  is the input sequence and  $y_i$  is the true label for the  $i$ -th example;  
+-  $L_{\text{Angular}}$  is the angular loss based on cosine similarity;  
+-  $L_{\text{Contrastive}}$  is the contrastive loss;  
+-  $w_{CE}, w_{Angular}$ , and  $w_{Contrastive}$  are the corresponding weights for each loss component, allowing for fine-tuning the contribution of each loss value during training.
+
+This combined loss function incorporates three learning objectives:
+
+1. The cross-entropy loss which ensures that the model learns to correctly classify the input sequences based on the true labels.  
+2. The angular loss which encourages the model to learn more separated representations for dif
+
+![](images/59ba8156e19487087b5a75dbf4e79ba7e74282f9b97cf16230c15df7f4a920c1.jpg)  
+Figure 1: Architecture of the DeB-Ang model highlighting loss computation within DeBERTa.
+
+ferent classes, based on the cosine similarity between the embeddings.
+
+3. The contrastive loss further enforces the separation between inter-class embeddings, while bringing intra-class embeddings closer together, based on the similarity calculations and a specified margin.
+
+By combining these three loss components, the DeBERTa model can potentially learn more robust and discriminative representations, leading to improved classification performance on various natural language processing tasks.
+
+As shown in Figure 1, the DeB-Ang model integrates angular and contrastive loss computations within the DeBERTa model, detailing the flow from data input to loss calculation and backpropagation.
+
+# 4.7 Evaluation and Error Analysis
+
+Considering the scale of the datasets, some accuracy values, when taken at face value, may not demonstrate any meaningful improvement in performance. Therefore, we utilise McNemar's test (Sundjaja et al., 2023) to demonstrate the statistical significance of our results. McNemar's test is a non-parametric test that can be used in comparing the performance of two classification models.
+
+For error analysis, we extracted both incorrectly classified and correctly classified data samples and performed an in-depth linguistic analysis of the outputs. We also computed the semantic similarity between correctly and incorrectly classified data by measuring the cosine similarity between the embeddings of the text pairs. We extracted contextual embeddings using the same DeBERTa model.
+
+# 5 Results and Discussions
+
+# 5.1 Machine-generated Text Detection
+
+In this section, we investigate binary machine-generated text detection, whereby the task is focussed on differentiating between human and
+
+<table><tr><td colspan="10">Accuracy and F1-score</td></tr><tr><td></td><td colspan="2">Contra-X</td><td colspan="2">Baseline DeBERTa</td><td colspan="2">DeB-Ang</td><td>SS (Y/N)</td><td colspan="2">Min-Max Improvement</td></tr><tr><td>PPLM-gpt2</td><td>99.34</td><td>99.32</td><td>99.66</td><td>99.66</td><td>99.98</td><td>100</td><td>Y</td><td>0.32</td><td>0.64</td></tr><tr><td>GPT1</td><td>52.66</td><td>52.66</td><td>99.89</td><td>99.93</td><td>97.92</td><td>97.92</td><td>Y</td><td>2.01</td><td>47.27</td></tr><tr><td>FAIR-wmt20</td><td>61.95</td><td>60.94</td><td>99.39</td><td>99.36</td><td>99.99</td><td>99.98</td><td>Y</td><td>0.60</td><td>38.04</td></tr><tr><td>GPT-3</td><td>97.85</td><td>97.85</td><td>98.80</td><td>98.8</td><td>99.73</td><td>99.73</td><td>Y</td><td>0.93</td><td>1.88</td></tr><tr><td>Grover-large</td><td>99.26</td><td>99.26</td><td>99.76</td><td>99.77</td><td>99.99</td><td>99.99</td><td>Y</td><td>0.23</td><td>0.73</td></tr><tr><td>transfo-xl</td><td>97.85</td><td>97.85</td><td>99.69</td><td>99.69</td><td>99.99</td><td>99.99</td><td>Y</td><td>0.30</td><td>2.14</td></tr><tr><td>GPT-2 small</td><td>96.57</td><td>96.57</td><td>99.82</td><td>99.82</td><td>99.52</td><td>99.54</td><td>Y</td><td>-0.3</td><td>2.95</td></tr></table>
+
+machine-written texts. Table 2 presents the results for this task on a variety of datasets from Turing-Bench. From the table, it is evident that the proposed model outperforms both Contra-X (Ai et al., 2022) and a baseline DeBERTa model with a minimum improvement of  $0.23\%$  and maximum improvement of  $47.27\%$  in accuracy. Statistical significance was computed by comparing DeB-Ang with the baseline DeBERTa model, as they exhibited the closest performance. The results for machine-generated text detection for the TuringExtended data is presented in Table 3. This demonstrates that the DeB-Ang model can differentiate between human and machine-generated texts even if the latter were generated by the newer NLG models, displaying detection accuracy over  $96\%$  for texts generated by Flan-T5-Large, GPT-4Turbo and Gemma-7b.
+
+From our initial experimentation, we noted that the baseline DeBERTa model outperforms other approaches in binary machine-generated text detection; therefore, for the remaining experiments we proceed with baseline DeBERTa.
+
+Table 2: Accuracy and F1-score for the baseline Contra-X, DeBERTa, and the proposed DeB-Ang model on various TuringBench datasets containing texts generated by different NLG models (rows). Min-Max refers to the minimum and maximum classification accuracy that DeB-Ang obtained for each dataset. Statistical significance (SS) between baseline DeBERTa and DeB-Ang is either yes (Y) or no (N) according to McNemar's test.  
+
+<table><tr><td></td><td colspan="6">Accuracy and F1-score</td></tr><tr><td></td><td colspan="2">Baseline DeBERTa</td><td colspan="2">DeB-Ang</td><td>SS (Y/N)</td><td>Min-Max Improvement</td></tr><tr><td>Flan-T5-Large</td><td>92.14</td><td>92.14</td><td>96.99</td><td>96.99</td><td>Y</td><td>4.85</td></tr><tr><td>Gemma-7b</td><td>99.96</td><td>99.96</td><td>99.98</td><td>99.99</td><td>N</td><td>0.02</td></tr><tr><td>GPT4-Turbo</td><td>72.14</td><td>72.14</td><td>99.94</td><td>99.97</td><td>Y</td><td>27.80</td></tr></table>
+
+The results for the DAIGT-V2 dataset can be seen in Table 5. This improvement demon
+
+Table 3: Accuracy and F1-score for the baseline DeBERTa and the proposed DeB-Ang model for TuringExtended. Min-Max refers to the minimum and maximum classification accuracy that DeB-Ang obtained for each dataset. Statistical significance (SS) between baseline DeBERTa and DeB-Ang is either yes (Y) or no (N) according to McNemar's test.  
+
+<table><tr><td></td><td>Accuracy</td><td>F1</td><td>SS (Y/N)</td></tr><tr><td>Syntax-CNN</td><td>66.13</td><td>64.80</td><td>Y</td></tr><tr><td>BERT-AA</td><td>78.12</td><td>77.58</td><td>Y</td></tr><tr><td>Contra-X</td><td>80.73</td><td>80.54</td><td>Y</td></tr><tr><td>Baseline DeBERTa</td><td>77.71</td><td>77.56</td><td>Y</td></tr><tr><td>GPT-who</td><td>65.89</td><td>65.39</td><td>Y</td></tr><tr><td>DeB-Ang</td><td>83.61</td><td>82.68</td><td>-</td></tr></table>
+
+Table 4: Accuracy and F1 for the authorship attribution (AA) dataset from TuringBench (Uchendu et al., 2021) comparing various AA approaches. McNemar's test was conducted to see if the result between DeB-Ang and all other models is statistically significant (SS) or not.
+
+strates the models' generalisability across various NLG datasets, for both older and newer models. Uchendu et al. (2021) comments "No one size fits all" in their study as they used several models on these datasets and found that different models obtain different levels of performance, depending on the dataset. However, as presented in Table 2, it is clear that the model consistently outperforms our baseline models on all datasets.
+
+# 5.2 Authorship Attribution
+
+In assessing the generalisability of the DeB-Ang approach on various text classification settings, we present the following authorship attribution tasks:
+
+1. Authorship attribution for human and machine-generated text detection.  
+2. Authorship attribution for model variation detection, e.g. differentiating between GPT-3.5 and GPT-4.  
+3. Authorship attribution for model developer detection, e.g. OpenAI for GPT-4 and GPT-3.5.
+
+The results for each task is presented in Table 5
+
+<table><tr><td colspan="2"></td><td colspan="2">Baseline DeBERTa</td><td colspan="2">DeB-Ang</td><td rowspan="2">Accuracy Improvement</td></tr><tr><td>Task</td><td>Model</td><td>Accuracy</td><td>F1</td><td>Accuracy</td><td>F1</td></tr><tr><td colspan="2">Machine-generated text detection</td><td>82.69</td><td>90.53</td><td>91.36</td><td>92.31</td><td>8.66</td></tr><tr><td colspan="2">Authorship attribution [37]</td><td>86.00</td><td>85.96</td><td>87.80</td><td>87.79</td><td>1.80</td></tr><tr><td rowspan="5">Model detection</td><td>Open AI [10]</td><td>88.64</td><td>88.64</td><td>91.75</td><td>91.75</td><td>3.11</td></tr><tr><td>Meta [13]</td><td>42.27</td><td>42.29</td><td>47.95</td><td>47.51</td><td>5.68</td></tr><tr><td>Google [7]</td><td>56.96</td><td>56.55</td><td>57.60</td><td>57.78</td><td>0.64</td></tr><tr><td>Anthropic [2]</td><td>95.63</td><td>95.58</td><td>99.03</td><td>99.03</td><td>3.40</td></tr><tr><td>Mistral [4]</td><td>93.15</td><td>93.39</td><td>93.96</td><td>95.12</td><td>0.81</td></tr><tr><td>Developer detection</td><td>All [5]</td><td>89.78</td><td>89.78</td><td>92.98</td><td>92.98</td><td>3.20</td></tr></table>
+
+Table 5: Table presenting evaluation results on the DAIGT-V2 dataset, including authorship attribution scores for all NLG models, machine-generated text detection (human vs. machine), model detection (distinguishing between different model variations), and authorship attribution for model developers. The numbers in brackets (e.g., "Open AI [10]") indicate the number of classes (i.e., the number of models).
+
+under authorship attribution, model detection and developer detection, respectively.
+
+From Table 4, it is evident that our approach also surpasses prior attempts on the TuringBench dataset. As previously mentioned, this dataset consists of texts generated by 20 different authors (a total of 200K texts from 19 NLG models and 1 human author) with high topical dissimilarity between each model. This dissimilarity is expected as the dataset was generated in certain topic set (Ai et al., 2022; Uchendu et al., 2021). Results for Syntax-CNN were taken from Ai et al. (2022) and all experiments were run using the full dataset. For the DAIGT-V2 dataset, we downsized the data to approximately 10K rows per model. This reduction was necessitated due to the dataset's size, which demanded significant computational resources. In Table 5, it can be observed that the DeB-Ang model outperforms the baseline DeBERTa model with an accuracy improvement of  $1.80\%$  in authorship attribution and  $8.66\%$  in machine-generated text detection. To delve deeper into the machine-generated text results from the baseline DeBERTa, we conducted an analysis focussing on the disparity between the accuracy and F1-score. This involved computing the Area Under the Receiver Operating Characteristic (AUROC) score and assessing misclassification. Our analysis revealed that the model exhibited a considerable number of false positives, incorrectly predicting a majority of human-written texts. The AUROC score was determined to be 50.12 whereas the AUROC score for DeB-Ang was 88.14 indicating DeB-Ang's superior discrimination capabilities. We also address the previously mentioned limitation regarding the scarcity of research in classifying models from a single developer; our results are provided in Table 5. We investigate a range
+
+of developers and models varying from older to newer model versions. We were able to improve results from baseline DeBERTa for this task by  $0.64\%$  to  $5.68\%$ . The low accuracy observed for Meta and Google models can be attributed to the high similarity between the model variations used, e.g., Llama-2-7b and Llama-2-13b. This makes distinguishing between these version challenging, leading to misclassification. Further investigation is necessary to comprehensively understand the reasons for misclassification. We were also able to classify generated texts according to model developer with an accuracy as high as  $92.98\%$ .
+
+As mentioned in prior research, classifying a range of outputs, e.g. texts with high topic variation, is an increasingly difficult classification task (Uchendu et al., 2021; Juola, 2008). Furthermore, it is important to note that TuringBench consists of texts from multiple sources. Additionally, some models generate many texts; this can decrease performance as there can be semantics and stylistic overlap between generated texts. This similarity between texts can blur distinctions thus, reducing classification capabilities.
+
+# 5.3 Assessing loss functions
+
+To assess the significance of the loss functions used, we investigated various combinations of loss functions on the multi-class authorship attribution and binary machine-generated text detection tasks. The results are presented in Table 6. We provide the accuracy, F1-score and AUROC scores for these tasks obtained by the DeB-Ang model. We ran each experiment for one epoch for initial benchmarking to assess each models performance. This allowed us to identify the which approaches we would use for further investigations. We identified the optimal
+
+parameter combination for the loss functions for each task and re-ran the experiment for 8 epochs. The aim of this investigation is to assess the performance improvement resulting from the various loss functions. This also highlights the customisability of the model. We extend the metrics by adding the AUROC score as this metric considers the tradeoff between precision and recall (McDermott et al., 2024).
+
+<table><tr><td>Task</td><td>Loss function</td><td>Accuracy</td><td>F1</td><td>AUROC</td></tr><tr><td rowspan="8">Authorship Attribution</td><td>ANG</td><td>86.37</td><td>86.42</td><td>97.88</td></tr><tr><td>CL</td><td>86.72</td><td>86.63</td><td>98.11</td></tr><tr><td>CE</td><td>86.87</td><td>86.71</td><td>98.00</td></tr><tr><td>ANG + CE</td><td>86.51</td><td>86.54</td><td>98.12</td></tr><tr><td>CL + CE</td><td>86.72</td><td>86.78</td><td>98.07</td></tr><tr><td>ANG + CL</td><td>87.02</td><td>87.08</td><td>98.12</td></tr><tr><td>ANG + CE + CL (1.0; 1.0; 1.0)</td><td>86.35</td><td>86.41</td><td>97.97</td></tr><tr><td>ANG + CE + CL (1.0; 0.25; 1.0)</td><td>87.13</td><td>87.22</td><td>98.15</td></tr><tr><td>8 epoch</td><td>ANG + CE + CL (1.0; 0.25; 1.0)</td><td>88.8</td><td>88.06</td><td>98.5</td></tr><tr><td rowspan="8">Binary machine-generated text detection</td><td>ANG</td><td>94.49</td><td>94.86</td><td>95.61</td></tr><tr><td>CL</td><td>88.15</td><td>89.19</td><td>72.58</td></tr><tr><td>CE</td><td>91.69</td><td>92.33</td><td>92.83</td></tr><tr><td>ANG + CE</td><td>90.87</td><td>91.89</td><td>77.00</td></tr><tr><td>CL + CE</td><td>91.6</td><td>92.47</td><td>90.89</td></tr><tr><td>CL + ANG</td><td>94.18</td><td>94.61</td><td>95.61</td></tr><tr><td>ANG + CE + CL (1.0; 1.0; 1.0)</td><td>90.76</td><td>91.92</td><td>87.57</td></tr><tr><td>ANG + CE + CL (1.0; 0.75; 0.75)</td><td>82.69</td><td>90.53</td><td>78.71</td></tr><tr><td>8 epoch</td><td>ANG + CE + CL (1.0; 0.75; 0.75)</td><td>93.76</td><td>88.23</td><td>90.94</td></tr></table>
+
+Table 6: Comparison of single and combined loss functions for authorship attribution and binary machine-generated text detection using the DeBERTa-based model with varying numbers of epochs. Parameter values for all loss functions were set to 1.0 unless otherwise specified. Key: AUROC = area under the receiver operating characteristic, CE = cross-entropy loss, CL = contrastive loss, and ANG = angular loss. The values in brackets refer to the parameter values.
+
+We found that a certain loss function combination may ascertain significant results at one epoch given a simple model. However, once the model or dataset complexity increases then a different loss combination would be more appropriate. Angular loss has the advantage of learning embeddings such that similar samples have a smaller angular separation. It is vital to understand that angular loss focusses on learning embeddings (Wang et al., 2017) whereas cross-entropy focusses on measuring the dissimilarity between predicted and true probability distribution of classes (Teahan, 2000). This difference may account for the accuracy difference. It is vital to note that each loss function has a different contribution and the combination of these, without careful tuning, may lead to sub-optimal results.
+
+# 5.4 Analysing the misclassified data
+
+For our error analysis, 100 instances of incorrect and correct classifications were extracted for the binary classification task. We found that texts were being labelled as machine-generated more frequently than human data; this could be due to the class imbalance or due to the NLG model's ability to create human-like text. While F1-scores account for class imbalance, an imbalanced training set can still bias the model toward the majority class.
+
+Based on the manual analysis, there was no specific linguistic category which would clearly lead to the misclassification. Therefore, we extracted features from varying categories (see Table 9 in Appendix C). A total of 250 features were extracted. 100 random features were sampled and the raw counts and mean for each feature was plotted (see Figure 2 in Appendix C). From this plot, it is evident that there is a clear discrepancy in feature usage. The correctly classified data points exhibit lower feature counts and an overall lower mean whereas the incorrectly classified data is slightly more sporadic and exhibits an overall higher mean. The statistical significance for these differences for all features was computed using the Mann-Whitney U test (Nachar, 2008) as the data was not normally distributed (as affirmed by the Shapiro-Wilk test) (Aryadoust and Raquel, 2020). The statistical significance was less than 0.05 thus rejecting the null hypothesis and confirming the difference between the feature counts and mean for the correctly classified and incorrectly classified data is significant.
+
+We then measured the semantic similarity between correctly classified and incorrectly classified instances using contextual embeddings obtained using DeBERTa. The mean similarity score for all data points is 85.54 (minimum score of 63.22 and maximum score of 92.64). Figure 3 in Appendix C presents a correlation coefficient of -0.03 indicating a very weak negative linear relationship almost suggesting no linear relationship between the data points. This indicates that any observed differences or similarities in the similarity score are likely due to random variation and not a meaningful underlying relationship. We conclude that the similarity scores do not provide useful information to distinguish between correctly and incorrectly classified instances.
+
+# 6 Conclusion and Future Work
+
+In this research, we have created a custom DeBERTa model integrating contrastive and angular loss. To our knowledge, this is one of the first attempts at this integration and we have demonstrated the success of the proposed DeB-Ang model on several datasets. We investigated more fine-grained machine-generated text detection by classifying model variations and developers. We were able to outperform prior approaches in machine-generated text detection with a minimum improvement of  $0.23\%$  and a maximum improvement of  $38.04\%$  across all datasets. We were able to classify model variations with accuracy scores ranging from  $0.64\%$  to  $5.68\%$ , and to identify developers with an accuracy improvement of  $3.20\%$ . For authorship attribution, we were able to improve classification with a maximum accuracy of  $17.48\%$  on the extensive TuringBench dataset which is characterised by high topical dissimilarity. Future work will involve identifying texts in which multiple NLG models or humans have been used to intentionally mask the writing style of a text. Additionally, a more extensive examination of linguistic features of synthetic data across generations of LLMs can provide insights into language evolution these models.
+
+# 7 Limitations
+
+Guerrero and Alsmadi (2022) lists several research gaps in the field of machine-generated text detection e.g. domain-specific text detection. It would be interesting to investigate texts that are cross-domain, genre or multimodal. Further, we investigated misclassified instances but did not use this information to improve the model due to time constraints. The limitations associated with data generation are model-related. Data generation is a time-consuming process and requires many computational resources; we were only able to extend our evaluation data with three datasets.
+
+# Ethics Statement
+
+The materials used for this study did not require human participation and the data does not contain any harmful or sensitive information. The datasets used in this study were acquired from prior research. The dataset generated using NLG models (Open AI's GPT-4 model, Gemma-7b and Flan-T5-large) was evaluated to ensure that there is no overtly harmful text. Data was annotated and evaluated by
+
+PhD students, the task was explained in regards to how data will be used and proposed tasks. Nevertheless, the potential negative use of this research should not be ignored. The insights provided by this work have the potential to be exploited for malicious purposes, potentially undermining the effectiveness of these detectors. However, we hope that this research will be used to support the efforts in detecting neural machine-generated used in applications with malicious intent.
+
+# References
+
+Ahmed Abbasi and Hsinchun Chen. 2008. Writeprints: A stylometric approach to identity-level identification and similarity detection in cyberspace. ACM Transactions on Information Systems, 26(2):1-29.  
+David Ifeoluwa Adelani, Haotian Mai, Fuming Fang, Huy H. Nguyen, Junichi Yamagishi, and Isao Echizen. 2019. Generating Sentiment-Preserving Fake Online Reviews Using Neural Language Models and Their Human- and Machine-based Detection.  
+Bo Ai, Yuchen Wang, Yugin Tan, and Samson Tan. 2022. Whodunit? Learning to Contrast for Authorship Attribution.  
+Rahaf Aljundi, Yash Patel, Milan Sulc, Daniel Olmeda, and Nikolay Chumerin. 2022. Contrastive Classification and Representation Learning with Probabilistic Interpretation.  
+Vahid Aryadoust and Michelle Raquel. 2020. Quantitative data analysis for language assessment. Volume II, Advanced methods. Routledge research in language education. Routledge, Abingdon, Oxon ;  
+Anton Bakhtin, Sam Gross, Myle Ott, Yuntian Deng, MarcAurelio Ranzato, and Arthur Szlam. 2019. Real or Fake? Learning to Discriminate Machine from Human Generated Text. arXiv.  
+Hongjun Choi, Anirudh Som, and Pavan Turaga. 2020. AMC-Loss: Angular Margin Contrastive Loss for Improved Explainability in Image Classification.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling Instruction-Finetuned Language Models.  
+Michael T. Cox. 2005. Metacognition in computation: A selected research review. Artificial Intelligence, 169(2):104-141. Special Review Issue.
+
+Evan N. Crothers, Nathalie Japkowicz, and Herna L. Viktor. 2023. Machine-Generated Text: A Comprehensive Survey of Threat Models and Detection Methods. IEEE Access, 11:70977-71002.  
+Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699.  
+Liam Dugan, Daphne Ippolito, Arun Kirubarajan, and Chris Callison-Burch. 2020. RoFT: A tool for evaluating human detection of machine-generated text. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 189-196, Online. Association for Computational Linguistics.  
+Liam Dugan, Eleni Miltsakaki, Shriyash Upadhyay, Etan Ginsberg, Hannah Gonzalez, DaHyeon Choi, Chuning Yuan, and Chris Callison-Burch. 2022. A feasibility study of answer-agnostic question generation for education. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1919-1926, Dublin, Ireland. Association for Computational Linguistics.  
+Mael Fabien, Esau Villatoro-Tello, Petr Moticek, and Shantipriya Parida. 2020. BertAA: BERT fine-tuning for authorship attribution. In Proceedings of the 17th International Conference on Natural Language Processing (ICON), pages 127-137, Indian Institute of Technology Patna, Patna, India. NLP Association of India (NLPAI).  
+Tiziano Fagni, Fabrizio Falchi, Margherita Gambini, Antonio Martella, and Maurizio Tesconi. 2021. Tweep-Fake: About detecting deepfake tweets. PLOS ONE, 16(5):e0251415.  
+Shuai Gao. 2022. Système de traduction automatique neuronale français-mongol (historique, mise en place et évaluations) (French-Mongolian neural machine translation system (history, implementation, and evaluations) machine translation (hereafter abbreviated MT) is currently undergoing rapid development, during which less-resourced languages nevertheless seem to be less developed). In Actes de la 29e Conférence sur le Traitément Automatique des Langues Naturelles. Volume 2: 24e Rencontres Etudiants Chercheurs en Informatique pour le TAL (RECITAL), pages 97-110, Avignon, France. ATALA.  
+Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2022. SimCSE: Simple Contrastive Learning of Sentence Embeddings.  
+Tanya Goyal, Junyi Jessy Li, and Greg Durrett. 2022. News summarization and evaluation in the era of gpt-3. arXiv preprint arXiv:2209.12356.  
+Jesus Guerrero and Izzat Alsmadi. 2022. Synthetic Text Detection: Systemic Literature Review.
+
+R. Hadsell, S. Chopra, and Y. LeCun. 2006. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR06), volume 2, pages 1735-1742.  
+Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTa: Decoding-enhanced BERT with Disentangled Attention.  
+Xinlei He, Xinyue Shen, Zeyuan Chen, Michael Backes, and Yang Zhang. 2024. MGTBench: Benchmarking Machine-Generated Text Detection.  
+Beizhe Hu, Qiang Sheng, Juan Cao, Yuhui Shi, Yang Li, Danding Wang, and Peng Qi. 2023. Bad actor, good advisor: Exploring the role of large language models in fake news detection. arXiv preprint arXiv:2309.12247.  
+Le Hui, Xiang Li, Chen Gong, Meng Fang, Joey Tianyi Zhou, and Jian Yang. 2019. Inter-class angular loss for convolutional neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3894-3901.  
+D. Ippolito, D. Duckworth, C. Callison-Burch, and D. Eck. 2020. Automatic detection of generated text is easiest when humans are fooled. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1808-1822. Association for Computational Linguistics.  
+Niful Islam, Debopom Sutradhar, Humaira Noor, Jarin Tasnim Raya, Monowara Tabassum Maisha, and Dewan Md Farid. 2023. Distinguishing Human Generated Text From ChatGPT Generated Text Using Machine Learning.  
+Vivek Iyer, Pinzhen Chen, and Alexandra Birch. 2023. Towards Effective Disambiguation for Machine Translation with Large Language Models. arXiv preprint arXiv:2309.11668.  
+Maurice Jakesch, Jeffrey T. Hancock, and Mor Naaman. 2023. Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11):e2208839120.  
+Ganesh Jawahar, Muhammad Abdul-Mageed, and Laks V. S. Lakshmanan. 2020. Automatic Detection of Machine Generated Text: A Critical Survey.  
+Patrick Juola. 2008. Authorship attribution. Foundations and Trends® in Information Retrieval, 1:233-334.  
+Virapat Kieuvongngam, Bowen Tan, and Yiming Niu. 2020. Automatic text summarization of Covid-19 medical research articles using bert and gpt-2. arXiv preprint arXiv:2006.01997.  
+Taehyeon Kim, Eungi Hong, and Yoonsik Choe. 2021. Deep Morphological Anomaly Detection Based on Angular Margin Loss. Applied Sciences, 11(14):6545.
+
+Taehyeon Kim, Seho Park, and Kyoungtaek Lee. 2023. Traffic Sign Recognition Based on Bayesian Angular Margin Loss for an Autonomous Vehicle. *Electronics*, 12(14):3073.  
+Anqi Mao, Mehryar Mohri, and Yutao Zhong. 2023. Cross-Entropy Loss Functions: Theoretical Analysis and Applications.  
+Colin Martindale and Dean McKenzie. 1995. On the utility of content analysis in author attribution: The Federalist. Computers and the Humanities, 29:259-270.  
+Matthew B. A. McDermott, Lasse Hyldig Hansen, Hao ran Zhang, Giovanni Angelotti, and Jack Gallifant. 2024. A Closer Look at AUROC and AUPRC under Class Imbalance.  
+Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature.  
+Sandra Mitrovic, Davide Andreoletti, and Omran Ayoub. 2023. ChatGPT or Human? Detect and Explain. Explaining Decisions of Machine Learning Model for Detecting Short ChatGPT-generated Text. ArXiv.  
+Nadim Nachar. 2008. The Mann-Whitney U: A Test for Assessing Whether Two Independent Samples Come from the Same Distribution. Tutorials in Quantitative Methods for Psychology, 4(1):13-20.  
+OpenAI. 2023. GPT-4 Technical Report.  
+Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library.  
+Dongqi Pu and Vera Demberg. 2023. ChatGPT vs Human-authored Text: Insights into Controllable Text Summarization and Sentence Style Transfer.  
+Jiameng Pu, Zain Sarwar, Sifat Muhammad Abdullah, Abdullah Rehman, Yoonjin Kim, Parantapa Bhattacharya, Mobin Javed, and Bimal Viswanath. 2022. Deepfake Text Detection: Limitations and Opportunities.  
+Junaid Qadir. 2022. Engineering Education in the Era of ChatGPT: Promise and Pitfalls of Generative AI for Education. TechRxiv.  
+Yunita Sari. 2018. Neural and non-neural approaches to authorship attribution. Ph.D. thesis, University of Sheffield, UK. British Library, EThOS.
+
+Irene Solaiman, Miles Brundage, Jack Clark, Amanda Askell, Ariel Herbert-Voss, Jeff Wu, Alec Radford, Gretchen Krueger, Jong Wook Kim, Sarah Kreps, Miles McCain, Alex Newhouse, Jason Blazakis, Kris McGuffie, and Jasmine Wang. 2019. Release Strategies and the Social Impacts of Language Models.  
+JH Sundjaja, R Shrestha, and K Krishan. 2023. McNemar And Mann-Whitney U Tests. StatPearls [Internet]. Updated 2023 Jul 17.  
+Zhiquan Tan, Yifan Zhang, Jingqin Yang, and Yang Yuan. 2024. Contrastive Learning Is Spectral Clustering On Similarity Graph.  
+Ruixiang Tang, Yu-Neng Chuang, and Xia Hu. 2023. The Science of Detecting LLM-Generated Texts.  
+Yi Tay, Dara Bahri, Che Zheng, Clifford Brunk, Donald Metzler, and Andrew Tomkins. 2020. Reverse Engineering Configurations of Neural Text Generation Models.  
+William John Teahan. 2000. Text classification and segmentation using minimum cross-entropy. In Content-Based Multimedia Information Access-Volume 2, pages 943-961.  
+Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikula, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Pier Giuseppe Sessa, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open Models Based on Gemini Research and Technology.
+
+Adaku Uchendu, Thai Le, Kai Shu, and Dongwon Lee. 2020. Authorship attribution for neural text generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8384-8395, Online. Association for Computational Linguistics.  
+Adaku Uchendu, Zeyu Ma, Thai Le, Rui Zhang, and Dongwon Lee. 2021. TURINGBENCH: A benchmark environment for Turing test in the age of neural text generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2001-2016, Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Saranya Venkatraman, Adaku Uchendu, and Dongwon Lee. 2024. GPT-who: An information density-based machine-generated text detector. In *Findings of the Association for Computational Linguistics: NAACL* 2024, pages 103–115, Mexico City, Mexico. Association for Computational Linguistics.  
+Roman Vygon and Nikolay Mikhaylovskiy. 2021. Learning efficient representations for keyword spotting with triplet loss. In Speech and Computer: 23rd International Conference, SPECOM 2021, St. Petersburg, Russia, September 27-30, 2021, Proceedings 23, pages 773-785. Springer.  
+Jian Wang, Feng Zhou, Shilei Wen, Xiao Liu, and Yuanqing Lin. 2017. Deep Metric Learning with Angular Loss.  
+Rowan Zellers, Ari Holtzman, Matthew Peters, Roozbeh Mottaghi, Aniruddha Kembhavi, Ali Farhadi, and Yejin Choi. 2021. SIGLeT: Language grounding through neuro-symbolic interaction in a 3D world. 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 2040-2050, Online. Association for Computational Linguistics.  
+Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating Text Generation with BERT.
+
+# Appendix A Dataset model breakdown
+
+<table><tr><td>Dataset</td><td>Source</td><td>Language Models</td><td>Link</td></tr><tr><td>TuringBench</td><td>TuringBench</td><td>GPT-1, GPT-2, GPT-3, GROVER, CTRL, XLM, FAIR, Transformer_XL, XLNET, PPLM</td><td>TuringBench</td></tr><tr><td>DAIGT-V2</td><td>Kaggle</td><td>LlaMa2, Darragh_Claude, Mistral7binstruct, Gemma, opt</td><td>DAIGT-V2</td></tr><tr><td>TuringExtended</td><td>Github</td><td>Gemma, Flan-T5, GPT</td><td>To be added</td></tr></table>
+
+# Appendix B Hyperparameter settings for the DeBERTa model
+
+Table 7: Overview of datasets utilized in the study, detailing dataset name, source, and the language models used to generate text. Note: While not exhaustive, datasets may encompass various iterations of a single model (e.g., LlaMa-7b and Llama-13b).  
+
+<table><tr><td>Hyperparameter</td><td>Amended value</td></tr><tr><td>num_train_epochs</td><td>1 - 8</td></tr><tr><td>train_batch_size</td><td>16</td></tr><tr><td>eval_batch_size</td><td>16</td></tr><tr><td>gradient Accumulation_steps</td><td>4</td></tr><tr><td>ngpu</td><td>1</td></tr><tr><td>max_seq_length</td><td>512</td></tr><tr><td>class_weight</td><td>Equal weighting specified</td></tr><tr><td>early_stopping_patience</td><td>2</td></tr><tr><td>early_stopping deltas</td><td>0.01</td></tr><tr><td>contrastive_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr><tr><td>angular_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr><tr><td>crossentropy_loss_weight</td><td>[0.05 - 0.25 - 0.50 -0.75 - 1.00]</td></tr></table>
+
+# Appendix C Error analysis: linguistic analysis
+
+# C.1 Linguistic features extracted
+
+Table 8: The hyperparameters used in training the DeB-Ang model. Parameter values for the epochs and loss functions varied and the specific values used are detailed in Section 5.  
+
+<table><tr><td>Linguistic category</td><td>Feature</td></tr><tr><td rowspan="3">Character-level features</td><td>uni, bi and tri-grams, word length distribution,</td></tr><tr><td>number counts, text length, average word</td></tr><tr><td>sentence length, casing, type-token ratio</td></tr><tr><td>Syntactic features</td><td>Part-of-speech tags, dependency tags</td></tr><tr><td>Word-level features</td><td>Function words</td></tr></table>
+
+Table 9: Linguistic features extracted from the correctly and incorrectly classified texts for the task of binary machine-generated text detection.
+
+# C.2 Linguistic feature groups
+
+![](images/be889168e090dc108fb87e2a3292b3c5f1d74ad3998e9c3dfeead4f04fec50e8.jpg)  
+Figure 2: Scatterplot displaying the raw counts and mean feature usage of incorrectly and correctly classified samples.
+
+![](images/394b9bce509798523fef1a069cbe5e44fbacd35cc8cdafc365a5337f78f92891.jpg)  
+Figure 3: Scatterplot displaying the similarity scores between each correctly and incorrectly classified data samples.

2024/Multi-Loss Fusion_ Angular and Contrastive Integration for Machine-Generated Text Detection/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9aba781f6abb2eb38eb575f79ebc441471d28f637d690a1b39ae98cb60b7d82
+size 520814

2024/Multi-Stage Balanced Distillation_ Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation/9db07c5e-dfb6-4143-97ad-d095709fb0c4_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d23bcb877c22db70ed8a7b869121172f015045111c0489341f2317907c98483
+size 1297307

2024/Multi-Stage Balanced Distillation_ Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation/full.md

@@ -0,0 +1,487 @@
+# Multi-Stage Balanced Distillation: Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation
+
+Yuhang Zhou  $^{1*}$  Jing Zhu  $^{2*}$  Paiheng Xu  $^{1}$  Xiaoyu Liu  $^{1}$  Xiyao Wang  $^{1}$  Danai Koutra  $^{2}$  Wei Ai  $^{1}$  Furong Huang  $^{1}$
+
+<sup>1</sup> University of Maryland, College Park
+
+2 University of Michigan, Ann Arbor
+
+{tonyzhou, paiheng, xliu1231, xywang, aiwei, furongh}@umd.edu
+
+{jingzhuu,dkoutra}@umich.edu
+
+# Abstract
+
+Large language models (LLMs) have significantly advanced various natural language processing tasks, but deploying them remains computationally expensive. Knowledge distillation (KD) is a promising solution, enabling the transfer of capabilities from larger teacher LLMs to more compact student models. Particularly, sequence-level KD, which distills rationale-based reasoning processes instead of merely final outcomes, shows great potential in enhancing students' reasoning capabilities. However, current methods struggle with sequence-level KD under long-tailed data distributions, adversely affecting generalization on sparsely represented domains. We introduce the Multi-Stage Balanced Distillation (BalDistill) framework, which iteratively balances training data within a fixed computational budget. By dynamically selecting representative head domain examples and synthesizing tail domain examples, BalDistill achieves state-of-the-art performance across diverse long-tailed datasets, enhancing both the efficiency and efficacy of the distilled models.
+
+# 1 Introduction
+
+Large language models (LLMs) like GPT-4 and LLaMA have revolutionized tasks ranging from text generation to language translation through their deep understanding and generation of human-like text (OpenAI, 2023; Touvron et al., 2023; Chiang et al., 2023; Jiang et al., 2023). Despite their success, the deployment of these models is hindered by their substantial size and computational demands, especially in environments with limited resources. Knowledge distillation (KD) offers a viable solution by transferring knowledge from expensive teacher models to smaller, efficient student models.
+
+Specifically, sequence-level  $KD$  focuses on distilling rationale-based reasoning processes rather than final outcomes. It leverages the teacher's reasoning processes, encapsulated in chain-of-thought (CoT) rationales, to enhance the student models' generative capabilities (Kim and Rush, 2016; Ho et al., 2022; Shridhar et al., 2022; Hsieh et al., 2023).
+
+However, there are a few challenges to fully leverage the power of sequence-level KD, as follows. (C1) Sequence-level KD encounters significant challenges when training with long-tailed data distributions, which are prevalent in real-world scenarios — data often follows a power-law distribution with a few dominant classes (head) and many rare classes (tail) (Liu et al., 2019). Such distributions feature a few dominant classes and many underrepresented ones, leading to models that generalize poorly on sparsely represented domains. (C2) Traditional KD methods in the text area to solve long-tail challenges, often reliant on direct access to model weights or loss adjustment primarily suited for straightforward classification tasks (Zhou et al., 2023; Schick and Schütze, 2021; Dai et al., 2023; Zhang et al., 2022; Tepper et al., 2020), falter under the complexities of sequence-level KD, especially when the teacher model is a black box and the task is generative, which is our target. (C3) Addressing this imbalance is critical, yet resource-intensive, as it typically requires generating a large volume of synthetic data to balance the dataset Tepper et al. (2020). Moreover, naively up-sampling the long-tailed dataset may dramatically increase the number of calls to the teacher models. Budget constraints play a crucial role in KD for black-box LLMs, as querying the teacher for rationales can be costly and time-consuming (Chen et al., 2023; Zhou and Ai, 2024).
+
+Our proposed solution, the Multi-Stage Balanced Distillation (BalDistill), tackles all the challenges above by strategically generating balanced training sets within budget constraints and iter
+
+![](images/59cad76ded93ccdd16c01fc871efbce51baf0a8dced2d343dfed18d8d351f761.jpg)  
+Figure 1: Overview of the proposed iterative BalDistill framework. The framework is composed of multiple stages. For each stage, we apply the balancing policy to decide the data distribution in the training batch. For head domains with sufficient data, we actively extract the examples by IFD metrics using the student model. For the tail domains, we call the teacher model to generate the synthetic examples and the corresponding rationales. The teacher model finally annotates the balanced training batch and fine-tunes the student model.
+
+tively fine-tuning the student model with actively selected and synthetic data for multiple stages. BalDistill progressively refines the training data by selecting key examples from well-represented domains and generating necessary synthetic data for underrepresented ones, ensuring comprehensive domain coverage and model robustness. By dynamically selecting representative head domain examples and synthesizing tail domain examples, BalDistill achieves state-of-the-art (SoTA) performance on various long-tailed datasets, enhancing both the efficiency and efficacy of the method.
+
+Our contributions are summarized as follows:
+
+- Innovative Problem Framing: We address the under-explored challenge of applying sequence-level KD to long-tailed distributions, where the teacher model is a black-box LLM.  
+Strategic Framework: BalDistill innovatively combines active example selection with synthetic data generation for multiple stages to maintain training balance within predefined budget limits.  
+- SoTA Performance: Our framework demonstrably improves the student models' effectiveness and robustness across diverse domains, setting new benchmarks in performance. We empirically demonstrate that our distilled student models achieve state-of-the-art performance across a range of benchmark datasets.
+
+# 2 Related Work
+
+Knowledge Distillation uses the outputs of a larger LLM (Teacher), such as ChatGPT (OpenAI, 2023), to train a smaller model (Student), such as LLaMa7B (Touvron et al., 2023). For details of knowledge distillation (KD) of large language models, we refer to the survey for more details (Xu et al., 2024b). In this work, we focus on KD with black-box teacher models. There are two lines of work with respect to knowledge distillation. The first is to ask teacher models to generate the final answers and to fine-tune on the final answers (Zhou et al., 2023; Schick and Schütze, 2021). Another line of work asks teacher models to generate rationales at the reasoning process and fine-tunes student models on the rationales in the sequence level to improve their reasoning ability (Ho et al., 2022; Shridhar et al., 2022; Hsieh et al., 2023), which proves to be more effective. In this work, we mainly discuss using a teacher model to generate rationales and improve the student's reasoning ability on a long-tailed dataset. Despite the progress of KD in the LLM era, existing works fail to establish a pipeline to gain knowledge from long-tailed datasets with the sequence-level KD, as few rationale examples are provided for tail knowledge.
+
+Long-Tail Learning focuses on long-tail distributed data and has been an emerging topic of interest in the NLP community (Liu et al., 2019; Wang et al., 2017; Godbole and Jia, 2022; Dai et al., 2023; Zhang et al., 2022; Liu et al., 2024d; Mondal
+
+et al., 2024). Approaches to solving the long-tail problem include rebalancing, information augmentation, and module improvement (Zhang et al., 2021; He et al., 2021; Zhu et al., 2023; Cui et al., 2021; Xu et al., 2024a). Despite the importance of long-tail learning, studies have shown that LLMs struggle to learn long-tail knowledge (Kandpal et al., 2023; Sun et al., 2023a). In this work, we propose to improve LLMs' ability to learn long-tail knowledge via multi-stage distillation over balanced datasets.
+
+Active Learning aims to reduce labeling effort by selecting only the most useful examples. Traditional active learning can be categorized into uncertainty-based methods (Prabhu et al., 2019; Margatina et al., 2021; Wang et al., 2023) and diversity-based methods (Ru et al., 2020; Ash et al., 2019). In the LLM era, active learning has been used to reduce human annotation costs by (1) strategically selecting the most informative examples for human feedback or annotation (Margatina et al., 2023; Osband et al., 2022; Wang et al., 2020) and (2) integrating language models as annotators within an active learning framework without human supervision (Xiao et al., 2023; Rouzegar and Makrehchi, 2024; Zhang et al., 2023; Li et al., 2024c; Liu et al., 2024f). In this work, we propose to solve the long-tail problem in the student LLM by actively distilling knowledge from a black-box teacher LLM to meet the budget requirement.
+
+# 3 Methodology
+
+# 3.1 Problem Statement
+
+We define our research problem as follows: Given the teacher LLM  $(\mathcal{M}_t)$ , the student LLM  $(\mathcal{M}_s)$ , a long-tailed dataset  $\mathcal{D}$  (with domain number  $[d_1, d_2, \ldots, d_l]$  for  $l$  domains in total) and a fixed budget  $B$  to query the teacher, we seek to propose an efficient framework to fine-tune an effective and robust student model,  $\mathcal{M}_s$ , over  $\mathcal{D}$ .
+
+# 3.2 Overall Approach
+
+To mitigate the performance bias in KD caused by long-tailed datasets within budget constraints, we employ a strategy that combines synthetic data augmentation with active selection. This approach ensures effective fine-tuning across both well-represented ('head') and underrepresented ('tail') domains. As depicted in Figure 1, we propose a multi-stage framework to create the training data iteratively.
+
+Algorithm 1 Multi Stage Balanced Distillations  
+1: Input: Long tailed dataset  $D$ , Student model  $M_{s}$ , Teacher model  $M_{t}$ , prompt for generating data  $P_{c}$ , Stage number  $K$ , Balancing policy  $P$ , Training bucket  $T$ , Budget number  $B$   
+2: Output: The fine-tuned student model  $M_{s}^{K}$   
+3: for each stage  $k = 0, \dots, k - 1$  do  
+4: head, tail domains  $= P(D, k, B)$   
+5: for each domain tail domain  $j$  do  
+6: Add remaining  $x_{j}$  from  $D$  to  $T$   
+7:  $\hat{x}_{j} = M_{t}(P_{c}, j)$   
+8: Add synthetic  $\hat{x}_{j}$  to  $T$   
+9: for each domain head domain  $h$  do  
+10: Collect all  $x_{h}$  from  $D$   
+11:  $x_{h} = M_{s}^{k - 1}(x_{h}, h)$   
+12: Add selected  $x_{h}$  to  $T$   
+13: Use  $M_{t}$  to annotate  $x$  in  $T$  w/o rationales  
+14:  $M_{s}^{k} = \text{Fine-tune}(M_{s}, T)$
+
+We operate in a pool-based setting where a large dataset, denoted as  $\mathcal{D}$ , is available but lacks annotations from a teacher model.
+
+At each stage of our BalDistill process, we first implement a balancing policy, which we have designed, to determine the appropriate data distribution for each domain within the training batch. This policy is based on the principles of data equality and training effectiveness across domains, aiming to optimize learning outcomes despite data scarcity in certain areas. The total number of stages is predefined based on the consideration of efficiency and the optimal performance.
+
+For domains well-represented in our dataset  $\mathcal{D}$  (referred to as 'head domains'), we employ active selection techniques (Touvron et al., 2023; Yuan et al., 2020) using the fine-tuned student model  $\mathcal{M}_s$  to identify and extract the most informative examples from the pool. Conversely, for domains lacking sufficient data ('tail domains'), we utilize the teacher model  $\mathcal{M}_t$  to generate both synthetic samples and corresponding annotations, enriching the training material available.
+
+After selecting and/or generating these samples, we query the teacher model to provide detailed rationales for examples in the training batch. These annotated examples are then used to fine-tune the student model  $\mathcal{M}_s$  in preparation for the next stage. Detailed descriptions of these components, along with the algorithms outlining this procedure, are presented in Algorithm 1.
+
+# 3.3 Balancing Policy
+
+Considering  $K$  total stages in our framework, we first evenly divide our budget  $B$  into  $K$  parts, which means that for each stage, we create a small training batch with  $\frac{B}{K}$  examples extracted from  $\mathcal{D}$  with teacher-annotated rationales. Within a small training batch, we propose two strategies to allocate the budget over different domains.
+
+Naive Balancing Since our goal is to mitigate the bias towards head domains, our first balancing policy is to use naive balancing, which selects the same number of inputs for each domain in the training batch. Formally, the number of samples for each domain in the small training batch is  $\frac{B}{Kl}$ , where  $l$  is the number of domains in the dataset.
+
+Adaptive Balancing One of our staged learning framework's key features is utilizing the fine-tuned student model to actively select representative inputs from well-represented domains, known as head domains. However, employing a naive balancing policy typically results in the disproportionate allocation of the training budget to data from underrepresented domains, or tail domains. This training batch may lead the fine-tuned student model to struggle to select truly effective examples from the head domains, particularly in the initial stages. Such selections are crucial for the model to learn effectively from these domains. To address this, we implement an adaptive balancing policy. This policy starts by constructing the training batch with a distribution akin to random selection, thus primarily focusing on head data in the early stages to 'warm up' the model. As the process advances, the policy gradually shifts towards a more balanced distribution by the final stage, ensuring comprehensive learning across both head and tail domains.
+
+Formally, the number of examples for each domain is the weighted average between the numbers for random selection and the numbers for naive balancing. For stage  $i$ , domain  $d$ , we select  $\left(\frac{n_d}{N} \cdot \frac{B}{K}\right) \cdot \frac{K - i}{K} + \frac{B}{Kl} \cdot \frac{i}{K}$  examples for domain  $d$  to build the training batch for adaptive balancing, where  $N$  and  $n_d$  are the total number and the domain size in the original data  $\mathcal{D}$ .
+
+Then, domains are naturally categorized based on whether the number of required samples per domain exceeds the available samples in the pool. Domains requiring more samples than available are designated as 'head domains' for that particular stage, while those with fewer required examples
+
+than available are categorized as 'tail domains.'
+
+For tail domains, where there are insufficient samples in the dataset  $\mathcal{D}$ , we rely on the teacher model to generate both the samples and their corresponding rationales, detailed in Section 3.4. In contrast, for head domains, which have a sufficient number of samples available to meet the demands of the training batch, we utilize the fine-tuned student model to actively select the most representative samples, as discussed in Section 3.5.
+
+It is important to note that the classification of domains as head or tail can vary across different stages of the training process, depending on the evolving needs and data availability.
+
+# 3.4 Teacher Data Augmentation
+
+Motivated by the effectiveness of synthetic dataset generated by black-box LLMs (OpenAI, 2023; Radford et al., 2019; Zhou et al., 2024b), we utilize the teacher LLMs to generate synthetic samples and corresponding annotations to upsample data for tail domains. To save the annotation budget, we require the teacher model to compose the sample and the corresponding rationales at the same time.
+
+Suppose that we need  $m$  synthetic examples for domain  $a$  to satisfy the training batch requirement. Given an instruction following prompt  $P_{c}$ , composed of three demonstrations from domain  $a$ , and teacher model  $\mathcal{M}_t$ , we employ stochastic temperature sampling with a fixed temperature and repeat the process  $m$  times with generated samples  $\hat{x}_{a1}, \dots, \hat{x}_{am}$  and rationales  $\hat{y}_{a1}, \dots, \hat{y}_{am}$ :
+
+$$
+\hat {x} _ {a i}, \hat {y} _ {a i} = M _ {t} (P _ {c}, a) \quad \mathrm {f o r} i \in \{1, \dots , m \}
+$$
+
+Then we add the generated samples and rationales to the training batch and combine with the extracted samples from  $\mathcal{D}$ . We present two examples of synthetic inputs and rationales from the teacher model in Table 10 in Appendix B. The case study suggests the effectiveness of the teacher model in generating tail examples.
+
+# 3.5 Student Active Selection
+
+For head domains, our strategy involves actively selecting instances from the original dataset to meet the numeric requirements of the balancing policy. We aim to mitigate information loss from data downsampling through this active data acquisition. The objective is to identify the most challenging or uncertain instances for the student model, thereby optimizing its learning trajectory.
+
+To quantify instance uncertainty, we adapt the Instruction Following Difficulty (IFD) metric originally proposed by Li et al. (2024a,b). The IFD scores are used to measure a training instance's uncertainty level as perceived by the student model. IFD is calculated as the ratio of the perplexity of generating a response  $y$  with an input  $x$  to the perplexity of generating  $y$  without  $x$ :  $\mathrm{IFD}(x,y) = \frac{\mathrm{PPL}(y|x)}{\mathrm{PPL}(y)}$ , where PPL represents perplexity, a metric widely used to evaluate language model performance (Jelinek et al., 1977). Studies have shown that IFD scores offer greater efficiency in data selection compared to methods like K-means diversity or sole reliance on perplexity (Li et al., 2024a; Settles, 2009; Yuan et al., 2020).
+
+A higher IFD score indicates an increased difficulty for the model in generating the response, highlighting the instance's value for training (Li et al., 2024a).
+
+Unlike the approach in Li et al. (2024a), which utilizes ground-truth or advanced LLM-generated responses  $y$ , our setting imposes budge constraints that prevent such usage. Instead, we calculate IFD using rationals  $\hat{y}_s$  generated by the previously finetuned student model, allowing us to assess the model's self-uncertainty and conserve the annotation budget from the teacher model.
+
+At last, we rank the inputs by their IFD scores, selecting those with the highest values to include in the batch, as specified by the balancing policy.
+
+# 3.6 Reasoning Generation and Fine-tuning
+
+Building on methodologies from prior research that focus on distilling reasoning abilities from black-box LLMs (Ho et al., 2022; Hsieh et al., 2023), we employ a zero-shot CoT approach, where the teacher model is prompted to generate a reasoning explanation  $\hat{y}_t$  for the samples in our constructed training batch. This zero-shot setting is crucial for demonstrating the model's ability to reason based on its pre-existing knowledge alone (Brown et al., 2020). In our experimental setup, which utilizes labeled datasets lacking rationale annotations, the final ground truth answer is included in the prompt. This inclusion ensures that the generated explanations are aligned with the correct outcomes, enhancing the accuracy and relevance of the CoT reasoning. It is important to note that for synthetic samples generated from tail domains in 3.4, we do not perform additional annotations in this part to maintain adherence to budget constraints.
+
+After gathering the required samples and their associated rationales in the training batch, we integrate this batch with the annotated data accumulated from previous stages. This approach ensures that our student model is exposed to a diverse and comprehensive dataset, which helps mitigate the risk of overfitting — a common challenge in machine learning models as identified in prior studies (Dor et al., 2020; Liu et al., 2023b). To facilitate this, we reinitialize and fine-tune the student model on the compiled rationale sequences from scratch at each stage.
+
+The fine-tuning is performed using autoregressive language modeling with a cross-entropy loss, aligning with the original pre-training objectives of the student model (Touvron et al., 2023).
+
+# 4 Experiment
+
+Through our extensive empirical analysis, we aim to address the following research questions:
+
+- RQ1: How effective is our KD framework compared to previous KD baseline methods?  
+- RQ2: How important is each component (balancing policy and active learning) to the framework?  
+- RQ3: How well does our method perform with different student models and budget restrictions?
+
+Dataset To verify the effectiveness of our framework on various reasoning tasks, we evaluate our method on five long-tailed datasets, following previous work (Yu et al., 2023; Dai et al., 2023; Huang et al., 2021). These include text classification: R52 and Reuters (Hayes and Weinstein, 1990), question answering: AbstractiveQA and Multiple-choiceQA (Dai et al., 2023) and arithmetic: MATH (Hendrycks et al., 2021). For text classification datasets, we treat the label of inputs as the domain; for other datasets, the domain information of inputs is annotated as metadata from the data provider. The detailed construction process and domain information for these datasets can be found in Appendix A. We also show two example distributions of the datasets in Figure 5 in Appendix A. For each dataset, we prepare two budget settings for the experiment. In Table 1, we present the budget number, the test number, the domain number, and the evaluation metric of all five datasets. The budget number in Table 1 represents the total number of queries to the teacher models. For example, budget setting 1 for the R52 dataset is 2,600, which means
+
+<table><tr><td>Dataset</td><td># Budget</td><td># Test</td><td># Domain</td><td>Metric</td><td>Task</td></tr><tr><td>R52</td><td>2,600/5,200</td><td>2,570</td><td>52</td><td>F1</td><td>TC</td></tr><tr><td>Reuters</td><td>4,500/9,000</td><td>3,745</td><td>90</td><td>F1</td><td>TC</td></tr><tr><td>Abstractive QA</td><td>5,000/10,000</td><td>10,000</td><td>5</td><td>F1</td><td>QA</td></tr><tr><td>Multi-choice QA</td><td>5,000/8,000</td><td>10,520</td><td>10</td><td>Accuracy</td><td>QA</td></tr><tr><td>Math</td><td>2,100/3,500</td><td>5,000</td><td>7</td><td>Accuracy</td><td>Arithmetic</td></tr></table>
+
+Table 1: Dataset statistics. TC and QA represent the text classification and question answering, respectively.
+
+that for our BalDistill method, the sum of queries to the teacher model for data augmentation and for generating reasoning steps should also equal 2,600 without incurring additional costs. This budget ensures that all operations, including data augmentation and reasoning step generation, are performed within the allocated query limit.
+
+Evaluation metrics Since we are dealing with long-tailed imbalanced data, for each dataset, we choose to use both the micro- and macro-averages to evaluate the method robustness (Henning et al., 2022; Li et al., 2024d). For the classification datasets (R52 and Reuters), we report micro-F1 and macro-F1, where micro-F1 is a global average F1 score and macro-F1 is computed by taking the unweighted mean of all the per-class F1 scores (Harbecke et al., 2022). For other datasets, we also report the micro-/macro-F1 for AbstractiveQA datasets and micro-/macro-accuracy for Math and Multi-choiceQA datasets. Note that the F1 score for the AbstractQA is the word-level F1 score between the token list of ground truth answer and the generated answer, different from the F1 for the classification task.
+
+Model setup For the teacher model, we use GPT-4 (OpenAI, 2023) to generate the CoT rationales for each dataset. We choose between Llama2-7B and Llama3-8B as our student models (Touvron et al., 2023). We include the detailed configurations and implementations of the model in Appendix B.
+
+Baseline methods We experiment with two variants of our proposed method with different balancing policies, as discussed in Section 3: In our first framework BalDistill (N), we use naive balancing policy, and for second framework BalDistill (A), we leverage adaptive balancing. We compare our framework with multiple baseline methods: (1) Zero-shot CoT. We directly prompt the student model to infer on the test data (Kojima et al., 2022). (2) Random Finetune. We randomly collect samples from the training data until the budget constraint is met and finetune student models on the final ground-truth labels (Radford et al., 2019). (3)
+
+Random Finetune-CoT. We randomly collect and use CoT rationales from the teacher model for student fine-tuning (Ho et al., 2022; Yao et al., 2022; He et al., 2023). (4) Duplicate Finetune-CoT. We construct the training data with a naive balancing policy. For the tail domains, we duplicate the inputs to satisfy the policy requirement and for head domains, we randomly sample examples in overrepresented domains.
+
+# 5 Results
+
+# 5.1 Comparison with Baseline Methods
+
+BalDistill framework outperforms Random Finetune and Duplicate Finetune methods. We use Llama3 as the student model, GPT-4 as the teacher model, and choose the smaller budget for each dataset in Table 1 as our experiment settings for this subsection. We present the overall macro- and micro-average results of the proposed frameworks and the baseline methods in Table 2. From Table 2, we first observe that on the long-tailed dataset, the methods fine-tuned on teacher-generated rationales (CoT) can significantly outperform the ground-truth fine-tuning method (Random Finetune), which emphasizes the necessity of teacher-generated reasoning steps in the KD.
+
+Among all sequence-level KD methods, our proposed BalDistill (N) and BalDistill (A) achieve the best average performance across various datasets on macro-averages, which obtain an average relative improvement of  $2.24\%$  and  $6.81\%$ , respectively, compared to the Random Finetune CoT baseline. The performance boost in BalDistill (N) implies the effectiveness of replacing the naive balancing policy with adaptive balancing.
+
+Moreover, we note that the Duplicate Finetune CoT baseline fails to compete with the Random Finetune CoT method in most cases, which indicates that simply duplicating the input from the tail domains to ensure balanced data cannot address the underlying imbalanced data complexity.
+
+To perform a detailed analysis of our framework, we visualize the F1 or accuracy score for each domain of the BalDistill (N) method and two baseline methods (Random Finetune CoT and Duplicate Finetune CoT) in Figure 2, with the x-axis representing the proportion of each domain in the dataset in descending order. From Figure 2, our proposed method can achieve comparable results in the head domains (left side of the figure) but substantially outperform the baseline methods in the
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">R52</td><td colspan="2">Reuters</td><td colspan="2">AbstractiveQA</td><td colspan="2">Multi-choiceQA</td><td colspan="2">Math</td></tr><tr><td>macro-f1</td><td>micro-f1</td><td>macro-f1</td><td>micro-f1</td><td>macro-f1</td><td>micro-f1</td><td>macro-acc</td><td>micro-acc</td><td>macro-acc</td><td>micro-acc</td></tr><tr><td>Zero-shot CoT</td><td>0.89</td><td>2.30</td><td>0.74</td><td>1.61</td><td>7.60</td><td>7.59</td><td>24.67</td><td>24.95</td><td>7.57</td><td>8.68</td></tr><tr><td>Random Finetune</td><td>45.95</td><td>91.44</td><td>28.01</td><td>74.68</td><td>37.62</td><td>37.21</td><td>61.23</td><td>55.96</td><td>10.12</td><td>9.48</td></tr><tr><td>Random Finetune CoT</td><td>59.70</td><td>89.46</td><td>27.35</td><td>70.53</td><td>52.57</td><td>52.88</td><td>76.09</td><td>74.12</td><td>16.62</td><td>15.20</td></tr><tr><td>Duplicate Finetune CoT</td><td>46.56</td><td>71.79</td><td>26.76</td><td>62.84</td><td>51.32</td><td>51.37</td><td>75.92</td><td>73.99</td><td>16.98</td><td>15.05</td></tr><tr><td>BalDistill (N)</td><td>59.62</td><td>82.49</td><td>28.09</td><td>62.40</td><td>52.70</td><td>52.92</td><td>76.60</td><td>73.43</td><td>17.90</td><td>16.34</td></tr><tr><td>BalDistill (A)</td><td>58.93</td><td>87.47</td><td>32.95</td><td>69.77</td><td>53.20</td><td>52.90</td><td>77.17</td><td>74.73</td><td>18.66</td><td>17.42</td></tr></table>
+
+![](images/c93bf52693e29c9d7c35ebeb0761e46fd8eff757895f7fbe30df567f0e390518.jpg)
+
+![](images/3a5affecf9f6bd3363265993bd86d8879892aa1a57845368529a926bb79dd725.jpg)
+
+![](images/b6942338a8ad77835339dfd6c7dd91c0cffe340d2de6014c137fb625a3dda722.jpg)  
+(c) MultichoiceQA
+
+![](images/0462111b38bae7c37f77500e32ee999d71888cac0be6953a07958b9a1424acc8.jpg)  
+(d) Math  
+Figure 2: Performance of proposed method and baselines on different domains. X-axis represents the proportion of each domain, ranked from head to tail domains. Our proposed BalDistill method can achieve comparable results on head domains and outperform the baseline method on the tail domains.
+
+tail domains (right side of the figure). This observation verifies our expectation in Section 3, where the balancing policy increases performance in the tail domain, and the active learning part improves the data efficiency to compensate for data loss in the head domain. Note that for Math dataset, BalDistill can only achieve comparable results with the baseline methods on the last two tail domains (pre-calculus and probability), and we conjecture that the high difficulty in these two domains prevents the teacher from composing high-quality synthetic data. We also include 2 additional SoTA methods for multitask learning or resolving class imbalance challenges and compare their results with the performance of BalDistill. Our approach outperforms the baselines in most tasks. Details are shown in Appendix C.
+
+Table 2: Performance of proposed BalDistill framework and other baselines across five long-tailed datasets. The best performance is marked in bold. The performance of fine-tuned student models with our framework can outperform other baselines in macro-averages on multiple long-tailed datasets.  
+
+<table><tr><td>Method</td><td>R52</td><td>Reuters</td><td>AbsQA</td><td>MCQA</td><td>Math</td></tr><tr><td colspan="6">Budget Setting 1</td></tr><tr><td>Random FT CoT</td><td>59.70</td><td>27.35</td><td>52.57</td><td>76.09</td><td>15.20</td></tr><tr><td>Balance FT CoT</td><td>51.47</td><td>27.12</td><td>52.22</td><td>75.98</td><td>16.29</td></tr><tr><td>Active FT CoT</td><td>59.49</td><td>29.75</td><td>53.14</td><td>76.64</td><td>15.61</td></tr><tr><td>BalDistill (N)</td><td>59.62</td><td>28.09</td><td>52.70</td><td>76.60</td><td>16.34</td></tr><tr><td>BalDistill (A)</td><td>58.93</td><td>32.95</td><td>53.20</td><td>77.17</td><td>17.42</td></tr><tr><td colspan="6">Budget Setting 2</td></tr><tr><td>Random FT CoT</td><td>64.88</td><td>33.42</td><td>53.71</td><td>72.92</td><td>15.19</td></tr><tr><td>Balance FT CoT</td><td>60.55</td><td>32.79</td><td>50.29</td><td>76.29</td><td>15.73</td></tr><tr><td>Active FT CoT</td><td>64.54</td><td>31.33</td><td>53.05</td><td>76.26</td><td>15.91</td></tr><tr><td>BalDistill (N)</td><td>59.35</td><td>32.76</td><td>53.86</td><td>76.17</td><td>17.59</td></tr><tr><td>BalDistill (A)</td><td>65.84</td><td>32.77</td><td>53.49</td><td>77.11</td><td>17.59</td></tr></table>
+
+Table 3: Effects of active learning and adaptive balancing in BalDistill framework. Results of fine-tuned student models on five datasets outperform methods with only balancing (Balance FT CoT), with only active learning (Active FT CoT).
+
+# 5.2 Ablation Study
+
+After showing the superiority of our overall framework, our next step is to verify the effectiveness of each component in the proposed method. We compare our framework with the ablated methods: (1) Balance Finetune CoT. We adopt a naive balancing policy to construct the training set and query the teacher model to compose inputs in the tail domains. We randomly sample examples from head domains to make sure they are not over-represented. (2) Active Finetune CoT. We only keep the active learning component but remove the data augmentation part. In details, we calculate the IFD scores for all examples in our original dataset and select the highest IFD scores to satisfy the budget number. Note that this ablation method is equivalent to the SoTA active learning method: Superfiltering (Li et al., 2024a). The experiment setting is similar to the setup in Section 5.1, and we present the performance of each method with two budget settings in Table 3.
+
+Both active selection and adaptive balancing
+
+![](images/7f7c028f61728b4bcdea49dfdd2aed1c6204ad739ecfb46818579825879aee27.jpg)
+
+![](images/764a60879cf970e135fb7fb9bd9696fa385af1b124010892f982a516b057885e.jpg)  
+(a) AbstractiveQA  
+(c) MultichoiceQA  
+Figure 3: Performance of proposed method BalDistill and ablated methods on head and tail domains. BalDistill (A) can achieve better results on head domains and outperform the Active FT CoT method on tail domains, which demonstrates the effectiveness of each component in our BalDistill (A) framework.
+
+![](images/2e4e6301100131b6d5cd4b88fcb136633b3a628eb8d6ed5710be8cbd8f8c6355.jpg)
+
+![](images/5c532721507adc4553dab9c4b0e8ab4598a76ef30b2be21f541511beb449dd3a.jpg)  
+(b)Reuters  
+(d) Math
+
+bring salient performance boost From Table 3, we find that our BalDistill (A) method obtains the best performance in 7/10 comparison cases, which demonstrates the effectiveness of each framework component. We notice that by simply adding the active learning strategy (Random Finetune CoT vs. Active Finetune CoT), the fine-tuned student model can achieve a performance boost in most cases, with an average relative improvement of  $1.43\%$ . This observation is consistent with the findings in previous work for Bert models (Devlin et al., 2019) on the long-tailed data (Dor et al., 2020).
+
+However, when we add data augmentation from the teacher with the naive balancing policy (Balance Finetune CoT vs. Random Finetune CoT, BalDistill (N) vs. Active Finetune CoT), this operation does not substantially improve performance. This finding suggests the superiority of our adaptive balancing policy.
+
+To probe the detailed reasons for the result patterns above, we visualize the macro-average performance of these methods on inputs from head and tail domains in Figure 3. The splitting criteria for each dataset can be found in Appendix A. We find that for methods with naive balancing policy (Balance Finetune CoT and BalDistill (N)), there exists a significant performance drop on head domains due to filtering a large proportion of data, and our method with adaptive balancing can achieve comparable performance on head domains. The observation suggests the effectiveness of active selection
+
+for head domains and the importance of adaptive balancing for the fine-tuned student to select the uncertain ones precisely.
+
+For performance in tail domains, our proposed method with adaptive balancing and teacher augmentation could achieve the best average results, even better than the naive balancing method. We conjecture that since we do not verify the correctness of teacher-generated samples and rationales in tail domains. While teacher-generated samples induce more knowledge, more synthetic data can lead to more inevitable noise. Adaptive balancing achieves the best trade-off between inducing more knowledge and less noise in the tail domains.
+
+# 5.3 Generalization Analysis
+
+The ablation study demonstrates the effectiveness of the active learning and adaptive balancing. Then, we ask whether our proposed method is robust enough to experiment with different hyperparameters, student models, or budget settings.
+
+# 5.3.1 Generalizations on Student models
+
+<table><tr><td>Method</td><td>R52</td><td>Reuters</td><td>AbsQA</td><td>MCQA</td><td>Math</td></tr><tr><td colspan="6">Llama3 Budget Setting 1</td></tr><tr><td>Random FT CoT</td><td>59.70</td><td>27.35</td><td>52.57</td><td>76.09</td><td>15.20</td></tr><tr><td>Active FT CoT</td><td>59.49</td><td>29.75</td><td>53.14</td><td>76.64</td><td>15.61</td></tr><tr><td>BalDistill (A)</td><td>58.93</td><td>32.95</td><td>53.20</td><td>77.17</td><td>17.42</td></tr><tr><td colspan="6">Llama2 Budget Setting 1</td></tr><tr><td>Random FT CoT</td><td>49.83</td><td>23.97</td><td>46.26</td><td>58.69</td><td>3.43</td></tr><tr><td>Active FT CoT</td><td>46.88</td><td>24.06</td><td>47.07</td><td>58.68</td><td>3.82</td></tr><tr><td>BalDistill (A)</td><td>58.33</td><td>25.51</td><td>47.55</td><td>59.14</td><td>4.21</td></tr><tr><td colspan="6">Llama3 Budget Setting 2</td></tr><tr><td>Random FT CoT</td><td>64.88</td><td>33.42</td><td>53.71</td><td>72.92</td><td>15.19</td></tr><tr><td>Active FT CoT</td><td>64.54</td><td>31.33</td><td>53.05</td><td>76.26</td><td>15.91</td></tr><tr><td>BalDistill (A)</td><td>65.84</td><td>32.77</td><td>53.49</td><td>77.11</td><td>17.59</td></tr><tr><td colspan="6">Llama2 Budget Setting 2</td></tr><tr><td>Random FT CoT</td><td>56.45</td><td>23.75</td><td>48.95</td><td>58.91</td><td>3.84</td></tr><tr><td>Active FT CoT</td><td>53.16</td><td>27.12</td><td>48.27</td><td>59.20</td><td>3.52</td></tr><tr><td>BalDistill (A)</td><td>58.17</td><td>27.07</td><td>49.45</td><td>58.64</td><td>4.54</td></tr></table>
+
+Table 4: Effects of student model scales and budget numbers. Macro-averages the proposed and baseline method results when considering Llama2 and Llama3 as student models with varying two budget settings.
+
+We first evaluate whether our method could be generalized to student models with different reasoning abilities or with different budget numbers. In this part, we additionally evaluate our BalDistill (A) on Llama2-7B models, which have a smaller model size and fewer tokens, in two budget settings (the details of each dataset are in Table 1). We present the fine-tuning results of our proposed
+
+![](images/f78110271b20a5de066a887dec2d5f3a4ac7bda1e40d9121c0ecf040b0697062.jpg)
+
+![](images/c629003e21703f05931bce87af38c49600710d47370ba781a690908381b3de6e.jpg)
+
+![](images/b3b8fbc67540ee5e7628080b850968b2022de0edc41bb14d5314449428758132.jpg)  
+(a) AbstractiveQA  
+(c) MultiChoiceQA  
+Figure 4: Influence of stage number choices on BalDistill across datasets. Our proposed method consistently obtains better results than the random fine-tune baseline method with varying stage numbers.
+
+![](images/1d351acb9d0fec1b76c132f078c835b9457dc0a483e195280a67399bff392438.jpg)  
+(b)Reuters  
+(d) Math
+
+framework and baseline methods on the Llama2 and Llama3 student models in Table 4.
+
+BalDistill exhibits robust improvement with various budget settings or student models. From Table 4, we observe that fine-tuning with the Llama3-8B student model leads to much better performance than the Llama2-7B model, especially on tasks with complex reasoning (Math, Multi-choice QA), indicating that the student with a larger model size or a better reasoning ability will yield better fine-tuning results. This observation is consistent with previous findings in Ho et al. (2022); Hsieh et al. (2023). Our BalDistill (A) consistently outperforms other baseline methods on both Llama2-7B and Llama3-8B as student models in most cases under two budget settings, which also verifies the generalizability of our BalDistill (A) on different student models or different budget numbers.
+
+# 5.3.2 Sensitivity Analysis
+
+We next investigate how the choice of stage number:  $K$  will influence the performance of our framework. We experiment with the same setup as in Section 5.1 but with varying stage numbers among  $\{3,5,8\}$ . We visualize the results (macro-averages) of BalDistill (A) and the baseline method Random Finetune CoT in Figure 4
+
+Figure 4 shows that the fine-tuning results of BalDistill (A) could be affected by the stage number to some extent, but our proposed method can consistently outperform the baseline method with
+
+different stage numbers, demonstrating the effectiveness and robustness of BalDistill (A).
+
+# 6 Conclusions
+
+In this paper, we propose a novel framework BalDistill to enhance performance on long-tail datasets in the current teacher-student knowledge distillation process. Our framework is a multi-stage pipeline, and at each stage, we call the student models to actively select the representative examples from head domains while prompting the teacher to generate synthetic examples for tail domains. With a fixed budget restriction for calling the teacher, our extensive empirical evaluations show that our framework can significantly increase fine-tuning results across multiple datasets. Furthermore, we demonstrate the effectiveness of all framework components through ablation studies.
+
+# Acknowledgments
+
+Zhou, Wang, Liu and Huang are supported by DARPA Transfer from Imprecise and Abstract Models to Autonomous Technologies (TIA-MAT) 80321, National Science Foundation NSF-IIS-2147276 FAI, DOD-ONR-Office of Naval Research under award number N00014-22-1-2335, DOD-AFOSR-Air Force Office of Scientific Research under award number FA9550-23-1-0048, DOD-DARPA-Defense Advanced Research Projects Agency Guaranteeing AI Robustness against Deception (GARD) HR00112020007, Adobe, Capital One and JP Morgan faculty fellowships. Zhu and Koutra are supported by the National Science Foundation CAREER Grant No. IIS 1845491. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation or other funding parties.
+
+# Limitations
+
+In our work, we use the IFD score as the metric for active selection for the student model. In addition to IFD scores, we can try other metrics, such as maximum entropy (Settles, 2009) or K-means diversity (Yuan et al., 2020). However, previous work has shown that the IFD score is more effective in selecting data for sequence-level fine-tuning than other metrics (Li et al., 2024a,b).
+
+We have verified the effectiveness of our framework on multiple student models and various long-
+
+tailed datasets. Other sequence-level KD methods still use more complex loss functions (Hsieh et al., 2023) or augment the generated rationales (Shridhar et al., 2023). Our data manipulation framework complements these KD methods, aiming to achieve more robust results on long-tailed datasets with a fixed budget. Moreover, our method focuses on sequence-level KD for black-box LLMs, so we do not incorporate the KD method for white-box LLMs as a baseline method (Gu et al., 2023; Dai et al., 2023). We will leave the exploration of combining our framework with more advanced KD methods for the future.
+
+Furthermore, our experiments only focus on the decoder-only student models: Llama3 and Llama2. Incorporating more encoder-decoder models such as FLAN-T5 (Chung et al., 2022) would benefit future studies.
+
+Another future direction for our paper is to explore the application of knowledge distillation in Large Vision-Language Models (LVLMs) (Liu et al., 2024c,b; Bai et al., 2023; Sun et al., 2023b; Zhou et al., 2024a; Wang et al., 2024a; Lin et al., 2024; Zhu et al., 2024; Liu et al., 2024e). In this paper, we have focused on experiments related to knowledge distillation in Large Language Models (LLMs). In future work, we aim to use knowledge distillation to further address the issue of hallucination (Liu et al., 2023a; Cui et al., 2023; Wang et al., 2024b) in small LVLMs such as LLaVA-7b (Liu et al., 2024c) and VILA-7b (Lin et al., 2024).
+
+# References
+
+Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. 2019. Deep batch active learning by diverse, uncertain gradient lower bounds. ICLR.  
+Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966.  
+Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901.  
+Lingjiao Chen, Matei Zaharia, and James Zou. 2023. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176.
+
+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.  
+Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2022. Scaling instruction-finetuned language models.  
+Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. 2023. Holistic analysis of hallucination in gpt-4v (ision): Bias and interference challenges. arXiv preprint arXiv:2311.03287.  
+Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. 2021. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724.  
+Yi Dai, Hao Lang, Yinhe Zheng, Fei Huang, and Yongbin Li. 2023. Long-tailed question answering in an open world. arXiv preprint arXiv:2305.06557.  
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding.  
+Liat Ein Dor, Alon Halfon, Ariel Gera, Eyal Shnarch, Lena Dankin, Leshem Choshen, Marina Danilevsky, Ranit Aharonov, Yoav Katz, and Noam Slonim. 2020. Active learning for bert: an empirical study. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 7949-7962.  
+Ameya Godbole and Robin Jia. 2022. Benchmarking long-tail generalization with likelihood splits. arXiv preprint arXiv:2210.06799.  
+Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. 2023. Minillm: Knowledge distillation of large language models. In The Twelfth International Conference on Learning Representations.  
+David Harbecke, Yuxuan Chen, Leonhard Hennig, and Christoph Alt. 2022. Why only micro-f1? class weighting of measures for relation classification. arXiv preprint arXiv:2205.09460.  
+Philip J Hayes and Steven P Weinstein. 1990. Construe/tis: A system for content-based indexing of a database of news stories. In IAAI, volume 90, pages 49-64.
+
+Nan He, Hanyu Lai, Chenyang Zhao, Zirui Cheng, Junting Pan, Ruoyu Qin, Ruofan Lu, Rui Lu, Yunchen Zhang, Gangming Zhao, et al. 2023. Teacherlm: Teaching to fish rather than giving the fish, language modeling likewise. arXiv preprint arXiv:2310.19019.  
+Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. 2021. Distilling virtual examples for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 235-244.  
+Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874.  
+Sophie Henning, William Beluch, Alexander Fraser, and Annemarie Friedrich. 2022. A survey of methods for addressing class imbalance in deep-learning based natural language processing. arXiv preprint arXiv:2210.04675.  
+Namgyu Ho, Laura Schmid, and Se-Young Yun. 2022. Large language models are reasoning teachers. arXiv preprint arXiv:2212.10071.  
+Cheng-Yu Hsieh, Chun-Liang Li, Chih-Kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alexander Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. 2023. Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes. arXiv preprint arXiv:2305.02301.  
+Yi Huang, Buse Giledereli, Abdullatif Köksal, Arzucan Özgür, and Elif Ozkirimli. 2021. Balancing methods for multi-label text classification with long-tailed class distribution. arXiv preprint arXiv:2109.04712.  
+Fred Jelinek, Robert L Mercer, Lalit R Bahl, and James K Baker. 1977. Perplexity—a measure of the difficulty of speech recognition tasks. The Journal of the Acoustical Society of America, 62(S1):S63-S63.  
+Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Nikhil Kandpal, Haikang Deng, Adam Roberts, Eric Wallace, and Colin Raffel. 2023. Large language models struggle to learn long-tail knowledge. In International Conference on Machine Learning, pages 15696-15707. PMLR.  
+Yoon Kim and Alexander M Rush. 2016. Sequence-level knowledge distillation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317-1327.  
+Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213.
+
+Ming Li, Yong Zhang, Shwai He, Zhitao Li, Hongyu Zhao, Jianzong Wang, Ning Cheng, and Tianyi Zhou 2024a. Superfiltering: Weak-to-strong data filtering for fast instruction-tuning.  
+Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024b. From quantity to quality: Boosting llm performance with self-guided data selection for instruction tuning.  
+Zongxia Li, Andrew Mao, Daniel Stephens, Pranav Goel, Emily Walpole, Alden Dima, Juan Fung, and Jordan Boyd-Graber. 2024c. Improving the tenor of labeling: Re-evaluating topic models for content analysis.  
+Zongxia Li, Ishani Mondal, Yijun Liang, Huy Nghiem, and Jordan Lee Boyd-Graber. 2024d. Pedants: Cheap but effective and interpretable answer equivalence.  
+Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. 2024. Vila: On pretraining for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699.  
+Bingchang Liu, Chaoyu Chen, Zi Gong, Cong Liao, Huan Wang, Zhichao Lei, Ming Liang, Dajun Chen, Min Shen, Haitian Zhou, et al. 2024a. Mftcoder: Boosting code llms with multitask fine-tuning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5430-5441.  
+Fuxiao Liu, Kevin Lin, Linjie Li, Jianfeng Wang, Yaser Yacoob, and Lijuan Wang. 2023a. Aligning large multi-modal model with robust instruction tuning arXiv preprint arXiv:2306.14565.  
+Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024b. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306.  
+Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024c. Visual instruction tuning. Advances in neural information processing systems, 36.  
+Xiaoyu Liu, Hanlin Lu, Jianbo Yuan, and Xinyu Li 2023b. Cat: Causal audio transformer for audio classification. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE.  
+Xiaoyu Liu, Paiheng Xu, Junda Wu, Jiaxin Yuan, Yifan Yang, Yuhang Zhou, Fuxiao Liu, Tianrui Guan, Haoliang Wang, Tong Yu, et al. 2024d. Large language models and causal inference in collaboration: A comprehensive survey. arXiv preprint arXiv:2403.09606.  
+Xiaoyu Liu, Jiaxin Yuan, Bang An, Yuancheng Xu, Yifan Yang, and Furong Huang. 2024e. Disentanglement: discovering causally-independent
+
+generative factors under an inductive bias of confounder. Advances in Neural Information Processing Systems, 36.  
+Xiaoyu Liu, Jiaxin Yuan, Yuhang Zhou, Jingling Li, Furong Huang, and Wei Ai. 2024f. Csrec: Rethinking sequential recommendation from a causal perspective. arXiv preprint arXiv:2409.05872.  
+Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. 2019. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2537-2546.  
+Katerina Margatina, Timo Schick, Nikolaos Aletras, and Jane Dwivedi-Yu. 2023. Active learning principles for in-context learning with large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5011-5034.  
+Katerina Margatina, Giorgos Vernikos, Loic Barrault, and Nikolaos Aletras. 2021. Active learning by acquiring contrastive examples. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 650-663.  
+Ishani Mondal, Zongxia Li, Yufang Hou, Anandhavelu Natarajan, Aparna Garimella, and Jordan Boyd-Graber. 2024. Scidoc2diagrammer-maf: Towards generation of scientific diagrams from documents guided by multi-aspect feedback refinement.  
+OpenAI. 2023. Gpt-4 technical report.  
+Ian Osband, Seyed Mohammad Asghari, Benjamin Van Roy, Nat McAleese, John Aslanides, and Geoffrey Irving. 2022. Fine-tuning language models via epistemic neural networks. arXiv preprint arXiv:2211.01568.  
+Ameya Prabhu, Charles Dognin, and Maneesh Singh. 2019. Sampling bias in deep active classification: An empirical study. 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 4058-4068.  
+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.  
+Hamidreza Rozegar and Masoud Makrehchi. 2024. Enhancing text classification through llm-driven active learning and human annotation. In Proceedings of The 18th Linguistic Annotation Workshop (LAW-XVIII), pages 98-111.  
+Dongyu Ru, Jiangtao Feng, Lin Qiu, Hao Zhou, Mingxuan Wang, Weinan Zhang, Yong Yu, and Lei Li. 2020. Active sentence learning by adversarial uncertainty sampling in discrete space. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4908-4917.
+
+Timo Schick and Hinrich Schütze. 2021. It's not just size that matters: Small language models are also few-shot learners. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2339-2352.  
+Burr Settles. 2009. Active learning literature survey.  
+Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2022. Distilling reasoning capabilities into smaller language models. arXiv preprint arXiv:2212.00193.  
+Kumar Shridhar, Alessandro Stolfo, and Mrinmaya Sachan. 2023. Distilling reasoning capabilities into smaller language models. In Findings of the Association for Computational Linguistics: ACL 2023, pages 7059-7073.  
+Kai Sun, Yifan Ethan Xu, Hanwen Zha, Yue Liu, and Xin Luna Dong. 2023a. Head-to-tail: How knowledgeable are large language models (llm)? aka will llms replace knowledge graphs? arXiv preprint arXiv:2308.10168.  
+Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. 2023b. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525.  
+Naama Tepper, Esther Goldbraich, Naama Zwerdling, George Kour, Ateret Anaby Tavor, and Boaz Carmeli. 2020. Balancing via generation for multi-class text classification improvement. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1440-1452.  
+Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.  
+Dongdong Wang, Yandong Li, Liqiang Wang, and Boqing Gong. 2020. Neural networks are more productive teachers than human raters: Active mixup for data-efficient knowledge distillation from a blackbox model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1498-1507.  
+Xiyao Wang, Jiuhai Chen, Zhaoyang Wang, Yuhang Zhou, Yiyang Zhou, Huaxiu Yao, Tianyi Zhou, Tom Goldstein, Parminder Bhatia, Furong Huang, and Cao Xiao. 2024a. Enhancing visual-language modality alignment in large vision language models via self-improvement.  
+Xiyao Wang, Wichayaporn Wongkamjan, Ruonan Jia, and Furong Huang. 2023. Live in the moment: Learning dynamics model adapted to evolving policy. In International Conference on Machine Learning. PMLR.
+
+Xiyao Wang, Yuhang Zhou, Xiaoyu Liu, Hongjin Lu, Yuancheng Xu, Feihong He, Jaehong Yoon, Taixi Lu, Gedas Bertasius, Mohit Bansal, et al. 2024b. Mementos: A comprehensive benchmark for multimodal large language model reasoning over image sequences. arXiv preprint arXiv:2401.10529.  
+Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. 2017. Learning to model the tail. Advances in neural information processing systems, 30.  
+Ruixuan Xiao, Yiwen Dong, Junbo Zhao, Runze Wu, Minmin Lin, Gang Chen, and Haobo Wang. 2023. Freearal: Towards human-free active learning in the era of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14520-14535.  
+Paiheng Xu, Jing Liu, Nathan Jones, Julie Cohen, and Wei Ai. 2024a. The promises and pitfalls of using language models to measure instruction quality in education. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4375-4389.  
+Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. 2024b. A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116.  
+Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629.  
+Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. 2023. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284.  
+Michelle Yuan, Hsuan-Tien Lin, and Jordan Boyd-Graber. 2020. Cold-start active learning through self-supervised language modeling. arXiv preprint arXiv:2010.09535.  
+Chen Zhang, Lei Ren, Jingang Wang, Wei Wu, and Dawei Song. 2022. Making pretrained language models good long-tailed learners. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3298-3312.  
+Ruoyu Zhang, Yanzeng Li, Yongliang Ma, Ming Zhou, and Lei Zou. 2023. Llmaaa: Making large language models as active annotators. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13088-13103.  
+Xing Zhang, Zuxuan Wu, Zejia Weng, Huazhu Fu, Jingjing Chen, Yu-Gang Jiang, and Larry S Davis. 2021. Videolt: Large-scale long-tailed video recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7960-7969.
+
+Yiyang Zhou, Zhiyuan Fan, Dongjie Cheng, Sihan Yang, Zhaorun Chen, Chenhang Cui, Xiyao Wang, Yun Li, Linjun Zhang, and Huaxiu Yao. 2024a. Calibrated self-rewarding vision language models.  
+Yuhang Zhou and Wei Ai. 2024. Teaching-assistant-in-the-loop: Improving knowledge distillation from imperfect teacher models in low-budget scenarios.  
+Yuhang Zhou, Suraj Maharjan, and Beiye Liu. 2023. Scalable prompt generation for semi-supervised learning with language models. arXiv preprint arXiv:2302.09236.  
+Yuhang Zhou, Paiheng Xu, Xiaoyu Liu, Bang An, Wei Ai, and Furong Huang. 2024b. Explore spurious correlations at the concept level in language models for text classification.  
+Jing Zhu, Yuhang Zhou, Vassilis N Ioannidis, Shengyi Qian, Wei Ai, Xiang Song, and Danai Koutra. 2023. Pitfalls in link prediction with graph neural networks: Understanding the impact of target-link inclusion & better practices. arXiv preprint arXiv:2306.00899.  
+Jing Zhu, Yuhang Zhou, Shengyi Qian, Zhongmou He, Tong Zhao, Neil Shah, and Danai Koutra. 2024. Multimodal graph benchmark. arXiv preprint arXiv:2406.16321.
+
+# A Dataset Construction
+
+R52 & Reuters We use the original R52 and Reuters dataset. In Figure 3, we treat domains (labels) with more than 50 instances in the training dataset as the head domains and the others as tail domains.
+
+Multi-choice QA For Multi-choice QA, we merge 10 multichoice QA datasets together, including Race, OBQA, MCTest, ARC-easy, ARC-hard, CQA, QASC, PIQA, SIQA, Winogrande (Dai et al., 2023). For training samples, we downsample the 10 datasets following a Zipf distribution with power value  $\alpha = 2.0$  (Dai et al., 2023). Since Race has  $5\times$  more training samples than other datasets, we downsample its training and testing set to  $1 / 3$  of the samples using random sampling. The detailed statistics of each multichoice qa dataset is shown in Table 5. We select Race, Winogrande, SIQA and CQA as the head domains and others as tail domains for experiments in Figure 3.
+
+Abstractive QA For Abstractive QA, we merge 5 abstractive QA datasets together, including NarQA, NQOpen, Drop, QAConv, TweetQA (Dai et al., 2023). Since the total train set and test set are very large, for efficiency concerns, we randomly sample 10000 samples from them for both train and test sets. The detailed statistics of each multichoice qa dataset is shown in Table 5. We select NarQA,
+
+![](images/b18305e20b375cf342d1e84d32081250f8e7b1767d0101073d0f2073cc77e256.jpg)
+
+![](images/d70474ceecf977b54c3e672befce2d02c75a402f1c3952a7640d34768d0861ae.jpg)  
+(a)Reuters  
+(b) Math  
+Figure 5: Example Dataset Distribution: The datasets we use exhibit long-tail distributions.
+
+NQOpen, and Drop as the head domains and others as tail domains for experiments in Figure 3.
+
+Math We use the Math dataset from (Hendrycks et al., 2021), which consists of 7 categories: Algebra, Intermediate Algebra, Prealgebra, Geometry, Number Theory, Counting & Probability and Precalculus. In order to investigate GPT4's reasoning ability on MATH problems and how much can its reasoning be taught to the student model, we removed the reasoning procedures in Math dataset and only keep its final answer as the label. Since the original dataset distribution is as follows does not follow long tail distribution, we down-sample the training sets of all categories following a Zipf distribution with power value  $\alpha = 1.1$ , similar to (Dai et al., 2023). The final distribution of the dataet is shown in Table 5. We select Algebra, Intermediate Algebra, and Prealgebra as the head domains and others as the tail domains for experiments in Figure 3.
+
+# B Implementation Details
+
+We use greedy search in decoding for all teacher annotations, as in the previous work (Ho et al., 2022) and use stochastic temperature sampling with the same temperature value of 0.9 in synthetic data generation in Section 3.4.
+
+We use the zero-shot prompts for the teacher to
+
+Table 5: Detailed statistics of each dataset per category.  
+
+<table><tr><td>Dataset</td><td>Category</td><td>Train set size</td><td>Test set size</td></tr><tr><td rowspan="10">Multi-choice QA</td><td>Race</td><td>4735</td><td>1629</td></tr><tr><td>OBQA</td><td>580</td><td>500</td></tr><tr><td>MCTest</td><td>342</td><td>320</td></tr><tr><td>ARC-easy</td><td>395</td><td>570</td></tr><tr><td>ARC-hard</td><td>317</td><td>299</td></tr><tr><td>CQA</td><td>1034</td><td>1221</td></tr><tr><td>QASC</td><td>653</td><td>926</td></tr><tr><td>SIQA</td><td>2077</td><td>1954</td></tr><tr><td>PIQA</td><td>494</td><td>1838</td></tr><tr><td>Winogrande</td><td>2634</td><td>1267</td></tr><tr><td rowspan="5">Abstractive QA</td><td>NarQA</td><td>1999</td><td>2244</td></tr><tr><td>NQOpen</td><td>4441</td><td>3434</td></tr><tr><td>Drop</td><td>2525</td><td>2891</td></tr><tr><td>QACov</td><td>751</td><td>1079</td></tr><tr><td>TweetQA</td><td>284</td><td>352</td></tr><tr><td rowspan="7">Math</td><td>Algebra</td><td>1744</td><td>1187</td></tr><tr><td>Intermediate Algebra</td><td>763</td><td>903</td></tr><tr><td>Prealgebra</td><td>561</td><td>871</td></tr><tr><td>Geometry</td><td>349</td><td>479</td></tr><tr><td>Number Theory</td><td>290</td><td>540</td></tr><tr><td>Counting &amp; Probability</td><td>231</td><td>474</td></tr><tr><td>Precalculus</td><td>187</td><td>546</td></tr></table>
+
+give the rationales and the few-shot ICL to generate the synthetic tail samples. The prompts are shown in Tables 7, 8 and 9. We call the gpt-4 function from OpenAI to obtain teacher responses.
+
+For the fine-tuning of the student model, we base our implementation on the Pytorch $^1$ , Huggingface transformer $^2$ , and the Lora fine-tuning codebase $^3$ . We use AdamW as our optimizer with a learning rate of  $2\mathrm{e} - 4$  and a weight decay of 0.03 with linear scheduler, batch size of 16, and trained for 8 epochs. For other hyper-parameters, we set rank and dropout in Lora fine-tuning to 8 and 0.1, respectively.
+
+# C Additional Baseline Results
+
+Besides the baselines mentioned in Section 5.1, we also include two additional state-of-the-art methods for multitask learning and addressing class imbalance challenges. The first method is Glee (Zhang et al., 2022), which leverages prompt tuning on masked tokens to handle long-tailed classification tasks. We adapted the classification head of Glee to make it applicable for generation tasks. However, since the final answers for AbstractiveQA consist of multiple words with variable lengths, Glee cannot be leveraged for AbstractiveQA. The second method is MFTCoder (Liu et al., 2024a), which proposes various loss functions to address challenges in multitask learning. For our experiment,
+
+we treat each domain as a separate task within their framework. Note that we employ instruction tuning for both MFTCoder and BalDistill, while using prompt tuning for Glee.
+
+Below we present the macro metrics of the baselines and BalDistill on each dataset. Llama2 is used as backbones for all of the methods.
+
+<table><tr><td>Dataset</td><td>R52</td><td>Reuters</td><td>AbsQA</td><td>MCQA</td><td>Math</td></tr><tr><td>MFTCoder</td><td>10.49</td><td>6.63</td><td>33.76</td><td>58.62</td><td>3.22</td></tr><tr><td>Glee</td><td>46.87</td><td>23.96</td><td>N/A</td><td>57.88</td><td>4.37</td></tr><tr><td>BalDistill (A)</td><td>58.33</td><td>25.51</td><td>47.55</td><td>59.14</td><td>4.21</td></tr></table>
+
+Table 6: Performance of BalDistill and two additional baselines across 5 datasets.
+
+From the results in Table 6, it is evident that our approach outperforms the MFTcoder and Glee on most tasks. MFTCoder's underperformance can be attributed to its use of validation loss gradients (as described in Equation 4 of their paper) to adjust training loss, which leads to unstable learning. In our experimental setup, particularly for tail labels in the R52 and Reuters datasets, we often had only one or two examples in the validation set. This scarcity can cause large gradients in the validation data, potentially leading to loss explosion during fine-tuning on these datasets. Glee's limitations results in its failure to utilize information from teacher rationales. In contrast, our method leverages data augmentation from teachers to elicit more knowledge and enhance fine-tuning for tail domains. This approach allows us to better capture and utilize the expertise embedded in the teacher models, resulting in improved performance across various tasks.
+
+You are provided with a dataset named R52, which is specifically designed for text classification tasks. The objective is to accurately predict the topic of news stories from a predefined list of topics. The topic of this dataset includes: copper, livestock, gold, money-fx, tea, api, trade, cocoa, iron-steel, reserves, zinc, nickel, ship, cotton, platinum, alum, strategic-metal, instal-debt, lead, housing, gnp, sugar, rubber, drl, tin, interest, income, crude, coffee, jobs, meal-feed, lei, lumber, gas, nat-gas, veg-oil, orange, heat, wpi, cpi, earn, jet, potato, bop, money-supply, carcass, acq, pet-chem, grain, fuel, retail, cpu. Please write a short news story with the topic {domain} and give the step-by-step rationale. This should be a self-contained story, mirroring the style and content of real-world news articles. Here are some examples with the topic {domain}: {demonstrations}  
+Please compose a news story with the topic {domain} with a similar format as the example. Paraphrase your title before outputting it. Your news story should be brief and contained within one paragraph:
+
+# (a) R52
+
+You are provided with a dataset named reuters, which is specifically designed for text classification tasks. The objective is to accurately predict the topic of news stories from a predefined list of topics. The topic of this dataset includes: acq, rubber, lead, money-supply, income, l-cattle, crude, cpu, palmkernel, jobs, money-fx, instal-debt, rand, castor-oil, coffee, strategic-metal, nat-gas, oat, tea, corn, yen, soy-oil, grain, groundnut-oil, gas, cpi, cocoa, nzdlr, soybean, rapeseed, retail, sun-meal, coconut, jet, copper, sorghum, carcass, heat, hog, ipi, potato, lin-oil, oilseed, alum, gnp, meal-feed, fuel, barley, ship, rape-oil, cotton-oil, sunseed, palm-oil, soy-meal, naphtha, nkr, trade, palladium, lei, wheat, bop, interest, earn, reserves, housing, veg-oil, groundnut, tin, dlr, gold, copra-cake, wpi, livestock, zinc, sugar, rye, pet-chem, dmk, dfl, orange, iron-steel, nickel, sun-oil, lumber, rice, propane, platinum, silver, cotton, coconut-oil. Please write a short news story with the topic {domain} and give the step-by-step rationale. This should be a self-contained story, mirroring the style and content of real-world news articles. Here are some examples with the topic {domain}: {demonstrations}  
+Please compose a news story with the topic {domain} with a similar format as the example and your news story should be brief and contained within one paragraph:
+
+# (b)Reuters
+
+Table 7: Prompts of generating synthetic data for tail domains from the teacher for R52 and reuters datasets.
+
+You are provided with a multiple-choice question and answering dataset composed by various QA datasets. The objective is to accurately select one from the given choices according to the question content. Please compose a question as well as the corresponding choices and answers as the examples from a QA dataset: {domain}. This should be a question, mirroring the style and content of examples with the true real-world knowledge. Here are some examples from the QA dataset: {domain}: {demonstrations}
+
+Please compose a question for the dataset: {domain} with a similar format as the example. It means if the example contains the in-context "passage", you should also write an in-context "passage" with the question information. Your question and choices should be brief and contained within one paragraph:
+
+# (a) Multi-choice QA
+
+You are provided with an abstractive question answering dataset composed by various QA datasets. The objective is to accurately generate an answer according to the question content. Please compose a question and the corresponding answer as the examples from a QA dataset: {domain}. This should be a question and answer, mirroring the style and content of examples with the true real-world knowledge. Here are some examples from the QA dataset: {domain}: {demonstrations}
+
+Please compose a question and the corresponding answer for the dataset: {domain} with a similar format as the example. It means if the example contains the in-context "passage", you should also write an in-context "passage" with the question information. Please note that the answer should only contain a few words. Your question and answer should be brief and contained within one paragraph:
+
+# (b) Abstractive QA
+
+You are provided with a math problem dataset with questions from various math domains. The objective is to accurately generate an answer according to the question content. Please compose a question and the corresponding answer as the examples from a math domain: {domain}. This should be a math question and answer, mirroring the style and content of examples with the true real-world knowledge. Here are some examples from the math domain: {domain}:
+
+{demonstrations}
+
+Please compose a math question and the corresponding answer for the domain: {domain}, with a similar format as the example. Please output your final digital answer (no unit) for the question with the format: "the answer is: <answer)". Your question and answer should be brief and contained within one paragraph:
+
+# (c) Math
+
+Table 8: Prompts of generating synthetic data for tail domains from the teacher for Multi-choice QA, Abstractive QA and Math datasets.
+
+Below is a news story from the R52 dataset. Please assign a topic to this news story. You must select the topic from this set: copper, livestock, gold, money-fx, tea,ipi, trade,cocoa, iron-steel, reserves, zinc, nickel, ship, cotton, platinum, alum, strategic-metal, instal-debt, lead, housing, gnp, sugar, rubber, dlr, tin, interest, income, crude, coffee, jobs, meal-feed, lei, lumber, gas, nat-gas, veg-oil, orange, heat, wpi, cpi, earn, jet, potato, bop, money-supply, carcass, acq, pet-chem, grain, fuel, retail, cpu. News story: {input}.
+
+Take a step-by-step approach in your response, cite sources and give reasoning. Your answer should be brief and contained within one paragraph.
+
+# (a) R52
+
+Below is a news story from the reuters dataset. Please assign a topic to this news story. You must select the topic from this set: acq, rubber, lead, money-supply, income, l-cattle, crude, cpu, palmkernel, jobs, money-fx, instal-debt, rand, castor-oil, coffee, strategic-metal, nat-gas, oat, tea, corn, yen, soy-oil, grain, groundnut-oil, gas, cpi, cocoa, nzdlr, soybean, rapeseed, retail, sun-meal, coconut, jet, copper, sorghum, carcass, heat, hog, ipi, potato, lin-oil, oilseed, alum, gnp, meal-feed, fuel, barley, ship, rape-oil, cotton-oil, sunseed, palm-oil, soy-meal, naphtha, nkr, trade, palladium, lei, wheat, bop, interest, earn, reserves, housing, veg-oil, groundnut, tin, dlr, gold, copra-cake, wpi, livestock, zinc, sugar, rye, pet-chem, dmk, dfl, orange, iron-steel, nickel, sun-oil, lumber, rice, propane, platinum, silver, cotton, coconut-oil. News story: {input}.
+
+Take a step-by-step approach in your response, cite sources and give reasoning. Your answer should be brief and contained within one paragraph.
+
+# (b)Reuters
+
+Please answer this multiple-choice question by choosing one of the given choices. If you are given a passage, please answer the question according to the passage content. If the passage is not given, please answer the question directly from your knowledge. Question: {input}
+
+If there is no enough information, you should choose a most possible choice. Take a step-by-step approach in your response, cite sources and give reasoning before sharing final answer in the format: The answer is <selected choice>.
+
+# (c) Multi-choice QA
+
+Here are a question and the corresponding answer for an abstractive question answering task. Please concisely clarify the rationale behind the answer for this question. If you are given a passage, please use the passage content to clarify the answer. If the passage is not given, please use your knowledge to tell why the answer is reasonable. Question: {input}. Answer {label}.
+
+Take a step-by-step approach in your response and give reasoning. You output should be concise and in one paragraph.
+
+# (d) Abstractive QA
+
+Here are a math question and the ground truth answer for a math problem task. Please concisely clarify the rationale behind the answer for this math question. Question: {input}. Answer: {label} Take a step-by-step approach in your response and give reasoning. You output should be concise and in one paragraph.
+
+# (e) Math
+
+Table 9: Teacher annotation prompts for all datasets. Note that for Math, Abstractive QA which has complex reasoning trajectories, we directly provide the ground-truth answer and ask the teacher model to explain the rationale behind the answer and for other datasets, we ask the teacher model to annotate the label and provide the reasoning steps.
+
+Input: ENERGY GIANT &lt;EGC&gt EXPANDS NATURAL GAS OPERATIONS. Energy Giant Corp (EGC) announced plans to expand its natural gas operations in the Midwest, following approval from the Federal Energy Regulatory Commission. The expansion will see EGC increase its natural gas production by  $15\%$ , aiming to meet the growing demand in the region. The company stated that the expansion will be funded by a mix of internal accruals and debt, with the project expected to be operational by the end of the next fiscal year. EGC also highlighted that the expansion aligns with its commitment to sustainable energy solutions, as natural gas is a cleaner-burning fuel compared to coal and oil. The company's shares rose  $2\%$  in early trading following the announcement.
+
+Rationale: The news story is about Energy Giant Corp (EGC) announcing plans to expand its natural gas operations in the Midwest. The company plans to increase its natural gas production by  $15\%$  to meet the growing demand in the region. The expansion will be funded by a mix of internal accrivals and debt. The company's shares rose  $2\%$  in early trading following the announcement. The story does not mention anything about copper, livestock, gold, money-fx, tea,ipi, trade,cocoa,iron-steel, reserves,zinc,nickel, ship,cotton,platinum, alum,strategic-metal,instal-debt,lead,housing,gnp, sugar,rubber,dlr,tin,interest,income,crude,coffee,jobs,meal-feed,lei,lumber,gas,veg-oil,orange, heat,wpi,cpi,earn,jet,potato,bop,money-supply,carcass,acq,pet-chem,grain,fuel,retail,cpu. However, the story is about the expansion of natural gas operations, which falls under the topic of 'nat-gas'. The answer is nat-gas.
+
+(a) The example of teacher synthetic data in R52 dataset for domain: nat-gas.
+
+Input: A rectangle has a length of 14 units and a width of 8 units. A diagonal is drawn from one corner of the rectangle to the opposite corner. What is the length of the diagonal in units?
+
+Rationale: The length of the diagonal of a rectangle can be found using the Pythagorean theorem, which relates the lengths of the sides of a right triangle. The theorem states that the square of the length of the hypotenuse (the side opposite the right angle, which is the diagonal in this case) is equal to the sum of the squares of the lengths of the other two sides (the length and width of the rectangle). For this rectangle, the length (l) is 14 units and the width (w) is 8 units. According to the Pythagorean theorem, the length of the diagonal (d) can be calculated as follows:  $d^2 = l^2 + w^2$ . Plugging in the values gives  $d^2 = 14^2 + 8^2 = 196 + 64 = 260$ . Taking the square root of both sides to solve for d gives  $d = \sqrt{260}$ . Factoring out the square root of 4 (which is a perfect square) from under the radical to simplify, we get  $d = 2\sqrt{65}$ . Thus, the length of the diagonal is  $2\sqrt{65}$  units. The answer is:  $2\sqrt{65}$ .
+
+(b) The example of teacher synthetic data in math dataset for domain: geometry.
+
+Table 10: Examples of the synthetic input and rationales from the teacher model for the tail domains.

2024/Multi-Stage Balanced Distillation_ Addressing Long-Tail Challenges in Sequence-Level Knowledge Distillation/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:89e9c49305957e6a28959741c7bbf162b3474da5fd4afc1426140454d8cbeece
+size 471133

2024/Multi-Target Cross-Lingual Summarization_ a novel task and a language-neutral approach/e5d21b5b-7d95-4d08-ad43-578c1b31719d_origin.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71b267a7710544340b89d0962f192b81fc5c4e781acd522b23ddea1e040d925b
+size 2188979

2024/Multi-Target Cross-Lingual Summarization_ a novel task and a language-neutral approach/full.md

@@ -0,0 +1,419 @@
+# Multi-Target Cross-Lingual Summarization: a novel task and a language-neutral approach
+
+Diogo Pernes Goncalo M. Correia Afonso Mendes
+
+Priberam Labs, Alameda D. Afonso Henriques, 41,  $2^{\circ}$ , 1000-123 Lisboa, Portugal
+
+{diogo.pernes, goncalo.correia, amm}@priberam.pt
+
+# Abstract
+
+Cross-lingual summarization aims to bridge language barriers by summarizing documents in different languages. However, ensuring semantic coherence across languages is an overlooked challenge and can be critical in several contexts. To fill this gap, we introduce multi-target cross-lingual summarization as the task of summarizing a document into multiple target languages while ensuring that the produced summaries are semantically similar. We propose a principled re-ranking approach to this problem and a multi-criteria evaluation protocol to assess semantic coherence across target languages, marking a first step that will hopefully stimulate further research on this problem.
+
+# 1 Introduction
+
+Cross-lingual summarization refers to the task of producing a summary in a different language than the original document and has the potential to break language barriers by helping people to effectively capture the essence of documents written in foreign languages (Wang et al., 2022). This is a very challenging task, as it combines the difficulties of monolingual summarization, such as factual inconsistencies with respect to the source document (Maynez et al., 2020), with those of machine translation, such as translation of idiomatic expressions and cultural references (Fadaee et al., 2018).
+
+The availability of large pre-trained multilingual transformers (Liu et al., 2020; Xue et al., 2021), followed by the widespread development and adoption of decoder-only language models (Radford et al., 2018; Touvron et al., 2023; Jiang et al., 2023; Team et al., 2024) has enabled a single model to perform cross-lingual summarization from multiple source languages to multiple target languages (many-to-many summarization, M2MS). Despite the increasing emphasis on this many-to-many paradigm, ensuring semantic coherence in summaries across different target languages has
+
+not been a primary focus of state-of-the-art methods, nor has it been systematically evaluated. Table 1 illustrates this issue by presenting an example where a state-of-the-art M2MS system based on mT5 (Xue et al., 2021) produces very different summaries, with one containing unfaithful content, depending on the chosen target language. Clearly, if information is not conveyed coherently across languages, the trustworthiness of the system is compromised. Users cannot rely on the summaries to be accurate and unbiased, regardless of the language in which they consume the content. In addition, in legal or regulatory contexts, ensuring that information is presented coherently across languages can be critical. This helps meet regulatory requirements and ensures that information is transmitted coherently across language boundaries.
+
+To fill this gap, we introduce a novel variant of cross-lingual summarization, which we call multi-target cross-lingual summarization (MTXLS), where we specifically address the challenge of promoting semantic coherence across target languages. This framework represents an important step towards more comprehensive cross-lingual summarization techniques and evaluation. Our main contributions in this work are summarized as follows: First, we introduce MTXLS formally as a novel task (Section 3), motivated by the need of producing summaries coherently for multiple target languages. Second, we present a re-ranking-based approach to address this problem (Section 4), where the re-ranking phase selects a set of summaries that exhibit superior semantic coherence across target languages compared to treating each cross-lingual summarization task independently. Notably, our approach circumvents the need for a pivot language. This language-neutral strategy ensures a more robust and unbiased multilingual summarization process. Finally, we propose and conduct a multicriteria evaluation protocol that goes beyond the simple evaluation of the similarity between gener
+
+<table><tr><td>Text (BBC)</td><td>en</td><td>Bitcoin uses more electricity annually than the whole of Argentina, analysis by Cambridge University suggests. &quot;Mining&quot; for the cryptocurrency is power-hungry, involving heavy computer calculations to verify transactions. Cambridge researchers say it consumes around 121.36 terawatt-hours (TWh) a year - and is unlikely to fall unless the value of the currency slumps. (...)</td></tr><tr><td rowspan="2">M2MS (mT5)</td><td>es</td><td>Bitcoin es la moneda más consumida del mundo, según un estudio de la Universidad de Cambridge.</td></tr><tr><td>fr</td><td>Bitcoinconsume plus d&#x27;électricité chaque année, selon une analyse de l&#x27;Université de Cambridge.</td></tr><tr><td rowspan="2">Mistral 7B</td><td>es</td><td>Los investigadores de la Universidad de Cambridge estiman que el Bitcoin consume alrededor de 121,36 TWh de energia electrica al año, lo que lo convierte en un proceso energetico intenso y poco probable que disminuya a menos que la valoración de la moneda bajo significativamente.</td></tr><tr><td>fr</td><td>Deschercheurs de l&#x27;Université de Cambridge esti-ment que le Bitcoin consomme environ 121,36 TWh d&#x27;électricité par an, ce qui en fait un processus énergi-vore qui est peu probable de décroître si la valeur de la monnaie chutes significativement.</td></tr><tr><td rowspan="2">NeutralRR (ours)</td><td>es</td><td>Bitcoin consume más electricidad que Argentina, según un estudio de la Universidad de Cambridge.</td></tr><tr><td>fr</td><td>Bitcoin consomme plus d&#x27;électricité que l&#x27;Argentine, selon une analyse de l&#x27;Université de Cambridge.</td></tr></table>
+
+Table 1: An example of an  $\mathrm{en}\rightarrow \{\mathrm{es,fr}\}$  summarization task solved by three different state-of-the-art systems, including ours. Text in red marks information that is present in a summary for one of the languages but not in the other summary.
+
+ated summaries and references (Section 5). Specifically, we incorporate the important aspect of evaluating the coherence of the entire set of generated summaries across all target languages using quality estimation methods for machine translation. The code and data used in our experiments are publicly available. $^{1}$
+
+# 2 Related Work
+
+# 2.1 Cross-Lingual Summarization
+
+Research in cross-lingual summarization has recently gained traction, in part due to the increased availability of large datasets for this task (Ladhak et al., 2020; Perez-Beltrachini and Lapata, 2021;Urlana et al., 2023). Among these, CrossSum (Bhattacharjee et al., 2023) stands out as the most resourceful. This news dataset contains document-summary pairs for 45 different languages and more than 1,500 language directions, and it was built by automatically pairing the data from the multilingual dataset XL-Sum (Hasan et al., 2021), which consists of news articles from BBC.
+
+Earlier cross-lingual summarization models operated on a per-language-pair basis (Zhu et al., 2019; Cao et al., 2020; Bai et al., 2021; Liang et al.,
+
+2022). However, with the emergence of large pretrained multilingual transformers like mBART (Liu et al., 2020) and mT5 (Xue et al., 2021), alongside extensive cross-lingual summarization datasets covering multiple language directions, a shift to many-to-many approaches occurred (Bhattacharjee et al., 2023; Chen et al., 2023b; Wang et al., 2023b). Evaluation expanded to include large decoder-only language models, including in a zero-shot setting, with only GPT-4 showing competitive performance compared to fine-tuned mBART-50 (Wang et al., 2023a; Tang et al., 2021). The approaches most akin to our setting in the cross-lingual summarization literature either involve first generating a summary in the source language and then using it to guide the generation of the target language summary (Bai et al., 2021), or employing a content plan generation step to condition the decoding of the target summary (Huot et al., 2024). However, they do not explicitly enforce or evaluate semantic similarity across summaries in different target languages.
+
+# 2.2 Quality Estimation for Machine Translation
+
+In machine translation (MT), quality estimation methods aim to predict translation quality without access to gold standard outputs (Specia et al., 2013, 2018). Our focus is on using sentence-level MT quality estimation to evaluate semantic coherence in the generated summaries across target languages, by taking two system-generated summaries for different languages and evaluating how well one translates the other.
+
+Quality estimation methods for MT can be performed at various levels: word-level, where binary labels (OK or BAD) are assigned to each machine-translated word, and sentence- or document-level, where a score is generated as an estimate of the quality of the whole translated sentence or document. Many quality estimation methods produce both word-level and sentence-level scores (Wang et al., 2018; Kepler et al., 2019a,b; Lee, 2020). A sentence-level quality estimation method can arise from training multilingual sentence encoders like LASER (Artetxe and Schwenk, 2019) or SONAR (Duquenne et al., 2023). These models align representations of translated sentences, allowing embedding similarity metrics in the common space to serve as quality estimation metrics for MT. BLASER (Chen et al., 2023a), an automatic text-free metric for evaluating speech translation,
+
+refines this idea by using a regression model trained on the concatenation of the LASER embeddings of the source text and the reference and machine-generated translations. BLASER 2.0 (Communication et al., 2023) replaces LASER with SONAR embeddings, supports both speech and text modalities, and exists in both reference-dependent and reference-free (i.e., quality estimation) variants. Similarly, COMET (Rei et al., 2020) was initially introduced as a reference-dependent metric that cross-encodes the source text and the reference and machine-generated translations using an XLM-RoBERTa model (Conneau et al., 2020). Later, a similar idea was followed to build its reference-free version, called CometKiwi (Rei et al., 2022).
+
+# 3 Multi-Target Cross-Lingual Summarization
+
+# 3.1 Problem Formulation
+
+This section formalizes the task of MTXLS. Let  $\boldsymbol{x}_o \in \mathcal{X}$  represent a document in the source language  $o$ , and let  $\mathcal{T} = \{t_1, t_2, \ldots, t_N\}$  denote a set of  $N$  target languages. Without loss of generality, we assume that  $o \in \mathcal{T}$ . The primary goal of MTXLS is to generate a set of  $N$  summaries, denoted as  $S = \{\boldsymbol{y}_{t_1}, \boldsymbol{y}_{t_2}, \ldots, \boldsymbol{y}_{t_N}\}$ , where there is a summary  $\boldsymbol{y}_{t_i} \in \mathcal{Y}$  for each language in  $\mathcal{T}$ .
+
+It is evident that this task can be seen as a combination of a monolingual summarization task in language  $o$  and  $N - 1$  cross-lingual summarization tasks from  $o$  to each target language  $t \in \mathcal{T} \setminus \{o\}$ . While these tasks could be approached independently, we impose a constraint: all  $N$  summaries should convey identical information regardless of the language. This constraint ensures the alignment of information across different languages, thus promoting coherence in the resulting set of summaries.
+
+# 3.2 Summarize-and-Translate
+
+Consider a scenario where a summarization model is available for generating summaries from language  $o$  to a pivot language  $\pi$ . Additionally, there are models for translating from  $\pi$  to each language in  $\mathcal{T}$ . Common statistical approaches to these tasks involve modeling the summarization distribution  $p(\boldsymbol{y}_{\pi} \mid \boldsymbol{x}_{o}, \pi)$  and the translation distributions  $p(\boldsymbol{y}_t \mid \boldsymbol{y}_{\pi}, t)$  for each  $t \in \mathcal{T}$ .
+
+To enforce the desired coherence constraint across target languages, a simple strategy is to assume that the target summaries are conditionally independent of the source document given the pivot
+
+summary, expressed as  $(\pmb{y}_t \perp \pmb{x}_o) \mid \pmb{y}_{\pi}, \forall t \in \mathcal{T}$  and entailed by the Bayesian network in Figure 1a. This implies that, for each target language  $t$ , the information utilized to generate  $\pmb{y}_t$  from  $\pmb{x}_o$  comes solely from  $\pmb{y}_{\pi}$ . Notably, since translation is a more deterministic task than summarization, this assumption serves to mitigate the potential variability of  $\pmb{y}_t$  across different target languages.
+
+The previous assumption allows us to write the cross-lingual summarization distributions that use  $\pi$  as the pivot language as:
+
+$$
+\begin{array}{l} p \left(\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t, \pi\right) = \sum_ {\boldsymbol {y} _ {\pi}} p \left(\boldsymbol {y} _ {\pi} \mid \boldsymbol {x} _ {o}, \pi\right) p \left(\boldsymbol {y} _ {t} \mid \boldsymbol {y} _ {\pi}, t\right) \\ = \mathbb {E} _ {\boldsymbol {y} _ {\pi} \mid \boldsymbol {x} _ {o}, \pi} p (\boldsymbol {y} _ {t} \mid \boldsymbol {y} _ {\pi}, t), \tag {1} \\ \end{array}
+$$
+
+for each  $t \in \mathcal{T}$ . Approximating this expectation with a single sample and using the source language as the pivot language yields the conventional summarize-and-translate approach to cross-lingual summarization. While this baseline ensures coherence across multiple target languages by deriving summaries from the translation of the same pivot summary, it has inherent drawbacks. In particular, it involves two successive phases of decoding: first generating the pivot summary, and then generating summaries for each target language, thus potentially suffering from error accumulation from both decoding phases. Moreover, it is likely to degrade the similarity to the reference summaries in the target languages because it is biased towards the pivot language. Thus, all resulting summaries will reflect any biases introduced during the summarization from language  $o$  to language  $\pi$ .
+
+# 4 Methodology
+
+# 4.1 Beyond Summarize-and-Translate
+
+We now relax the conditional independence assumption made previously by explicitly conditioning  $\pmb{y}_t$  on  $\pmb{x}_o$ , as shown in Figure 1b. Notably, this approach does not involve decoding  $\pmb{y}_t$  after  $\pmb{y}_{\pi}$  but rather allows the two processes to run in parallel, and explicitly promotes semantic similarity between  $\pmb{y}_{\pi}$  and each  $\pmb{y}_t$ , as required to satisfy our constraint. We now have:
+
+$$
+p \left(\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t, \pi\right) = \mathbb {E} _ {\boldsymbol {y} _ {\pi} \mid \boldsymbol {x} _ {o}, \pi} p \left(\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, \boldsymbol {y} _ {\pi}, t\right). \tag {2}
+$$
+
+Let us impose that:
+
+$$
+p \left(\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, \boldsymbol {y} _ {\pi}, t\right) = \frac {1}{Z} \phi \left(\boldsymbol {y} _ {t}, \boldsymbol {y} _ {\pi}\right) q \left(\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t\right), \tag {3}
+$$
+
+![](images/f4749c9a709dcb73f3334096c501e12498ed68d54e261890e80f8a73ac51b496.jpg)  
+(a)
+
+![](images/986f8bf68ac3739b2cc909b0b054799ef387a8486af0040bde07d225c28d5b6c.jpg)  
+(b)  
+Figure 1: Graphical models representing summarize-and-translate (a), our method with a pivot language (b), and our language-neutral approach (c). Here,  $\pmb{x}_o$  denotes the document in the source language  $o$ ,  $\pmb{y}_{\pi}$  denotes the summary in the pivot language  $\pi$ , and  $\pmb{y}_{t_i}$  denotes the summary in the target language  $t_i$ ,  $i \in \{1,2,\dots,N\}$ .
+
+![](images/b3ca25d8b066f31959c8279f3791036b25a01fa51bc2e76f13e98f72b8a9cd06.jpg)  
+(c)
+
+where  $Z$  is a normalizing function independent of  $y_{t}, \phi : \mathcal{Y}^{2} \mapsto \mathbb{R}^{+}$  is a symmetric function measuring the semantic similarity between two texts in different languages and satisfies  $\sum_{\boldsymbol{y}_t} \phi(\boldsymbol{y}_t, \cdot) < \infty$ , and  $q(\boldsymbol{y}_t \mid \boldsymbol{x}_o, t)$  is modeled by a cross-lingual summarization system from language  $o$  to language  $t$ . This formulation explicitly addresses both of our goals: to produce a text  $y_{t}$  that serves as a good summary of  $x_o$  in language  $t$  and has a high similarity to the pivot  $y_{\pi}$ . Finally, we get:
+
+$$
+\begin{array}{l} p (\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t, \pi) = \mathbb {E} _ {\boldsymbol {y} _ {\pi} \mid \boldsymbol {x} _ {o}, \pi} \frac {1}{Z} \phi (\boldsymbol {y} _ {t}, \boldsymbol {y} _ {\pi}) q (\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t) \\ \approx \frac {1}{Z} \phi (\boldsymbol {y} _ {t}, \boldsymbol {y} _ {\pi}) q (\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t) \\ \propto \phi (\boldsymbol {y} _ {t}, \boldsymbol {y} _ {\pi}) q (\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t), \tag {4} \\ \end{array}
+$$
+
+where  $\pmb{y}_{\pi} \sim p(\pmb{y}_{\pi} \mid \pmb{x}_o, t)$ . This framework unveils diverse avenues for MTXLS. One is to directly train  $p(\pmb{y}_t \mid \pmb{x}_o, t, \pi)$  by jointly learning  $\phi$  and  $q$  from data, which requires cross-lingual document-summary pairs for all target languages and parallel data between the pivot and each target language. Alternatively,  $\phi$  could be used as a re-scoring function at each decoding step from  $q$ , but this would introduce a significant computational burden.
+
+In our work, we adopt a simpler re-ranking approach. We use  $q$  to generate  $k$  candidate summaries for each target language  $t$ , and then use  $\phi$  to select the optimal candidate. Notably, this allows simultaneous generation of candidate and pivot summaries, and enhances the semantic coherence of generated summaries while maintaining similarity to the reference cross-lingual distribution used to train the summarizer, which were not possible in the summarize-and-translate approach. As shown in Section 4.3, our approach has a deep connection with rejection sampling.
+
+# 4.2 A Language-Neutral Formulation
+
+Despite not using translation to obtain summaries for the target languages, the approach we have described in Section 4.1 still relies in a pivot language. However, following the same formulation, we can circumvent this issue by defining a joint distribution for the summaries in all the target languages:
+
+$$
+p (\mathcal {S} \mid \boldsymbol {x} _ {o}, \mathcal {T}) \propto \varphi (\mathcal {S}) \prod_ {i = 1} ^ {N} q \left(\boldsymbol {y} _ {t _ {i}} \mid \boldsymbol {x} _ {o}, t _ {i}\right), \tag {5}
+$$
+
+where
+
+$$
+\varphi (\mathcal {S}) = \frac {1}{\binom {N} {2}} \sum_ {i, j: j > i} \phi \left(\boldsymbol {y} _ {t _ {i}}, \boldsymbol {y} _ {t _ {j}}\right) \tag {6}
+$$
+
+measures the semantic similarity of the set of summaries  $\mathcal{S}$  by averaging all the pairwise similarities between each pair of summaries in  $\mathcal{S}$ . This model is represented graphically in Figure 1c. Note that the formulation in Section 4.1 is a particular case of this one where  $\mathcal{S} = \{\pmb{y}_t, \pmb{y}_\pi\}$  and  $p(\mathcal{S} \mid \pmb{x}_o, \mathcal{T}) = p(\pmb{y}_t \mid \pmb{x}_o, t, \pi) q(\pmb{y}_\pi \mid \pmb{x}_o, \pi)$ .
+
+# 4.3 Summary Sampling
+
+Our primary goal is now to conceive a method that allows us to sample summaries from:
+
+$$
+p (\mathcal {S} \mid \boldsymbol {x} _ {o}, \mathcal {T}) = \frac {\varphi (\mathcal {S})}{Z ^ {\prime}} \prod_ {i = 1} ^ {N} q \left(\boldsymbol {y} _ {t _ {i}} \mid \boldsymbol {x} _ {o}, t _ {i}\right). \tag {7}
+$$
+
+We demonstrate we can achieve this goal through rejection sampling, which works as follows. Given a distribution  $f(\pmb{x})$  from which we aim to sample and a proposal distribution  $g(\pmb{x})$  satisfying  $\sup_{\pmb{x}} \frac{f(\pmb{x})}{g(\pmb{x})} \leq M$ , we start by generating a sample  $\pmb{x}$  from  $g$  and a sample  $u$  uniformly in  $[0,1]$ . Subsequently, we accept  $\pmb{x}$  if  $\frac{f(\pmb{x})}{Mg(\pmb{x})} \geq u$  and reject it otherwise.
+
+In our context, we may use  $\prod_{i=1}^{N} q(\pmb{y}_{t_i} \mid \pmb{x}_o, t)$  as the proposal distribution and assume without loss of generality that  $\phi$  is bounded in  $[0,1]$ , so  $\varphi$  is also bounded in  $[0,1]$  and therefore:
+
+$$
+\sup  _ {\mathcal {S}} \frac {p (\mathcal {S} \mid \boldsymbol {x} _ {o} , \mathcal {T})}{\prod_ {i = 1} ^ {N} q \left(\boldsymbol {y} _ {t _ {i}} \mid \boldsymbol {x} _ {o} , t _ {i}\right)} = \sup  _ {\mathcal {S}} \frac {\varphi (\mathcal {S})}{Z ^ {\prime}} \leq \frac {1}{Z ^ {\prime}}. \tag {8}
+$$
+
+Thus,  $M = \frac{1}{Z'}$  satisfies the condition above. The rejection sampling procedure for sampling from  $p(S \mid x_o, \mathcal{T})$  is then:
+
+1. Sample  $\mathcal{S}$  by sampling  $\boldsymbol{y}_t \sim q(\boldsymbol{y}_t \mid \boldsymbol{x}_o, t)$  independently for each  $t \in \mathcal{T}$ .  
+2. Sample  $u\sim U(0,1)$  
+3. Accept  $S$  if  $\varphi(S) \geq u$ ; otherwise, reject it.
+
+In step 1, summaries can be sampled independently and in parallel for each target language because of the factorized form of the proposal distribution.
+
+# 4.4 A Mode-Seeking Heuristic
+
+The procedure presented in Section 4.3 offers a systematic means to sample sets of summaries from the distribution  $p(S \mid x_o, \mathcal{T})$ . However, in many practical scenarios, the objective is to obtain a single set of high-quality summaries, i.e., a set with high probability under this distribution. This goal motivates the approach we present here.
+
+Let us assume we can generate  $k$  candidate summaries for each target language using diverse beam search (Vijayakumar et al., 2018) or a sampling algorithm. In this setup, there are  $k^N$  different sets of summaries resulting from the different combinations of selecting a candidate from each target language. Among these sets, we wish to choose the set  $S^*$  that maximizes  $\varphi(S)$ , in order to achieve our goal of having a maximally semantically coherent set of summaries. Interestingly, this criterion corresponds to choosing the set  $S^*$  with maximum probability of being accepted in the rejection sampling procedure described in Section 4.3.
+
+However, finding  $S^*$  among the  $k^N$  candidate sets is an instance of the generalized maximum clique problem, which is NP-hard (Feremans et al., 2003), and therefore we must resort to a heuristic search. For this purpose, we introduce a random permutation  $\sigma$  of the target languages  $\mathcal{T}$ , e.g.  $\sigma(\mathcal{T}) = (t_N, t_{N-1}, \ldots, t_1)$ , and define the proxy similarity function as follows:
+
+$$
+\hat {\varphi} (\mathcal {S}; \sigma) = \frac {1}{N - 1} \sum_ {i = 1} ^ {N - 1} \phi \left(\boldsymbol {y} _ {\sigma (\mathcal {T}) _ {i}}, \boldsymbol {y} _ {\sigma (\mathcal {T}) _ {i + 1}}\right). \tag {9}
+$$
+
+Algorithm 1 Language-neutral multi-target cross-lingual summarization
+
+Require: Input document  $(\pmb{x}_o)$ ; Set of target languages  $(\mathcal{T},$  with size  $N$ ); Number of candidates per language  $(k)$ ; Number of random permutations  $(m)$ .
+
+for each  $t\in \mathcal{T}$  do
+
+$$
+\begin{array}{l} \text {f o r} i \leftarrow 1 \text {t o} k \text {d o} \\ \text {S a m p l e} \boldsymbol {y} _ {t} ^ {(i)} \sim q (\boldsymbol {y} _ {t} \mid \boldsymbol {x} _ {o}, t). \\ \end{array}
+$$
+
+$$
+e n d \mathrm {f o r}
+$$
+
+end for
+
+for  $i\gets 1$  to  $m$  do  $\triangleright$  Find set with high similarity Build a weighted directed graph  $\mathcal{G} = (\mathcal{V},\mathcal{E})$  , where  $\mathcal{V}$  has  $Nk + 2$  nodes, one for each candidate summary plus a source and a sink node, and  $\mathcal{E}\gets \emptyset$
+
+$$
+\begin{array}{l} \begin{array}{c c c c c} \text {S a m p l e} & \text {a r a n d o m} & \text {p e r m u t a t i o n} & \sigma (\mathcal {T}) \\ (t _ {1} ^ {\prime}, t _ {2} ^ {\prime}, \ldots , t _ {N} ^ {\prime}). \end{array} = \\ \mathcal {E} \leftarrow \mathcal {E} \cup \left\{\left(\text {s o u r c e} \rightarrow \boldsymbol {y} _ {t _ {1} ^ {\prime}} ^ {(i)}, 0\right)\right\} _ {i = 1} ^ {k} \\ \end{array}
+$$
+
+$$
+\mathcal {E} \leftarrow \mathcal {E} \cup \left\{\left(\boldsymbol {y} _ {t ^ {\prime}} ^ {(i)} \rightarrow \operatorname {s i n k}, 0\right)\right\} _ {i = 1} ^ {k}
+$$
+
+$$
+\begin{array}{l} \text {f o r} l \leftarrow 1 \text {t o} N - 1 \mathbf {d o} \\ \mathcal {E} \leftarrow \mathcal {E} \cup \left\{\left(\boldsymbol {y} _ {t ^ {\prime} l} ^ {(i)} \rightarrow \boldsymbol {y} _ {t ^ {\prime} l + 1} ^ {(j)}, 1 - \phi \left(\boldsymbol {y} _ {t ^ {\prime} l} ^ {(i)}, \boldsymbol {y} _ {t ^ {\prime} l + 1} ^ {(j)}\right)\right)\right\} _ {i, j = 1} ^ {k} \\ \end{array}
+$$
+
+end for
+
+$$
+\hat {S} _ {i} ^ {*} \leftarrow \text {s h o r t e s t p a t h} (\mathcal {G}, \text {s o u r c e}, \text {s i n k})
+$$
+
+end for
+
+$$
+\operatorname {r e t u r n} \hat {S} ^ {*} \leftarrow \arg \max  _ {\mathcal {S} \in \{\hat {S} _ {1} ^ {*}, \dots , \hat {S} _ {m} ^ {*} \}} \varphi (\mathcal {S}) \quad \triangleright \text {e q .} \tag {6}
+$$
+
+This proxy represents a sparsification of the clique in the graphical model shown in Figure 1c, where only the edges connecting adjacent target summaries according to the permutation  $\sigma$  are retained. This sparsification embodies the assumption of transitivity in semantic similarity: For any three languages  $a$ ,  $b$ , and  $c$ , if the summary  $\mathbf{y}_a$  is similar to  $\mathbf{y}_b$ , and  $\mathbf{y}_b$  is similar to  $\mathbf{y}_c$ , then it follows that  $\mathbf{y}_a$  should also share a significant degree of similarity with  $\mathbf{y}_c$ . Notably, the set that maximizes  $\hat{\varphi}(\mathcal{S}; \sigma)$  can be found in  $O(Nk^2 \log(Nk))$  time using dynamic programming. This observation motivates Algorithm 1, where we consider  $k$  candidate summaries per target language and  $m \ll N!$  random permutations of the target languages. Then, for each permutation, we find the candidate set  $\hat{S}_i^*$  that maximizes  $\hat{\varphi}(\mathcal{S}; \sigma_i)$  using dynamic programming. Finally, we choose the set among  $\hat{S}_1^*, \hat{S}_2^*, \dots, \hat{S}_m^*$  that has the highest score according to  $\varphi$ .
+
+# 4.5 Choice of  $\phi$
+
+So far, we have presented our methodology in a formal manner, but have not yet provided specifics on implementing a function  $\phi$  capable of measuring the semantic similarity between two summaries in different languages. In practice, any quality estimation model for MT (Section 2.2) could be used. In our experiments, we leverage the cosine similarity of SONAR embeddings (Duquenne et al., 2023) as the similarity metric, reserving BLASER 2.0 (Chen
+
+et al., 2023a) and CometKiwi (Rei et al., 2022) for evaluation. Our selection of the cosine similarity of SONAR embeddings is motivated by its symmetry, unlike the remaining options, and the fact that the SONAR encoder is relatively lightweight. Specifically, we define the similarity function as:
+
+$$
+\phi \left(\boldsymbol {y} _ {a}, \boldsymbol {y} _ {b}\right) = \frac {1 + \boldsymbol {s} _ {a} ^ {\top} \boldsymbol {s} _ {b}}{2}, \tag {10}
+$$
+
+where  $s_a$  and  $s_b$  represent the  $L_2$ -normalized SONAR embeddings of summaries  $y_a$  and  $y_b$ .
+
+# 5 Experiments
+
+# 5.1 Dataset
+
+We use a subset of the CrossSum dataset, including data in the following languages: Arabic, Chinese (simplified), English, French, Portuguese, Russian, and Spanish. CrossSum pairs documents in one language with summaries from documents in another language, using automatic similarity metrics. However, mispairings are frequent due to this automated process. Additionally, the dataset is designed for single-target cross-lingual summarization and does not perfectly fit our multi-target setting. To adapt the dataset to our needs, we restructured the dataset into clusters. This process is explained in Appendix A. Each resulting cluster consists of up to seven multilingual document-summary pairs, with one such pair for each language. This allows us to select any document within the cluster as a source for summarization, with all summaries within the cluster serving as references for each of the languages. Statistics about the clustered data and an analysis of the semantic coherence of the dataset summaries are also provided in Appendix A.
+
+# 5.2 Methods
+
+Our pivot-free re-ranking method (NeutralRR) proposed in Algorithm 1 was tested using  $k = 8$  candidates per target language for re-ranking and  $m = 6$  language permutations, unless otherwise specified. We study the effects of varying  $k$  and  $m$  in Section 5.5 and Appendix D.2, respectively. We compare our method with four other approaches, namely: a many-to-many summarizer with beam search decoding (M2MS) with a beam size of 8; the summarize-and-translate approach (S&T), where summaries are obtained in the source language and then translated to each of the target languages using beam search with a beam size of 8 in both decoding steps; a Mistral 7B (Jiang et al., 2023)
+
+large language model (LLM) used in a zero-shot setting and instructed to write summaries with identical information for all the target languages (see Appendix C); our pivot-dependent re-ranking approach (PivotRR) as described in Section 4.1, where we use the source language as the pivot.
+
+All summaries except those of Mistral 7B were decoded from the same mT5 base model (Xue et al., 2021) fine-tuned in CrossSum. In the S&T approach, translations were performed using the NLLB 1.3B model (Costa-jussa et al., 2022). NeutralRR and PivotRR used beam search multinomial sampling using with 5 beams and a temperature of 1.0 for candidate generation. The pivot summary in PivotRR was decoded using beam search with 8 beams. For Mistral 7B, we used multinomial sampling with a temperature of 0.1. Further implementation details are provided in Appendix B.
+
+# 5.3 Evaluation Metrics
+
+Throughout this work, we emphasize the importance of evaluating MTXLS not only by comparing the generated summaries for each target language with their respective references, but also by evaluating the semantic coherence across different target languages. To evaluate the former, we present the ROUGE-2 F1 scores (Lin, 2004) for each generated summary against its corresponding reference in the same target language. In addition, we calculate the BLASER 2.0 score (Communication et al., 2023) by treating the generated summary as the translation and the reference summary for the source language as the source text. This evaluation metric is justified due to mismatched articles in CrossSum, as explained in Section 5.1, which reduces the reliability of reference summaries in languages other than the source.
+
+To assess semantic coherence across various target languages, we evaluate how well each generated summary translates the generated summaries for the remaining target languages. For this purpose, we use two quality estimation models for MT, namely CometKiwi (Rei et al., 2022) and BLASER 2.0. Here, for each target language, we use the generated summary as the translation and the summaries generated for all the other target languages as the source texts and then report the average across those languages.
+
+<table><tr><td rowspan="2">Source</td><td rowspan="2">Method</td><td colspan="3">ROUGE-2 (R)</td><td colspan="3">BLASER 2.0 (R)</td><td colspan="3">CometKiwi (C)</td><td colspan="3">BLASER 2.0 (C)</td><td rowspan="2">T</td><td rowspan="2">#P</td></tr><tr><td>en</td><td>zh</td><td>rest</td><td>en</td><td>zh</td><td>rest</td><td>en</td><td>zh</td><td>rest</td><td>en</td><td>zh</td><td>rest</td></tr><tr><td rowspan="5">en</td><td>M2MS</td><td>17.88</td><td>18.63</td><td>13.20</td><td>3.52</td><td>3.04</td><td>3.25</td><td>59.28</td><td>61.00</td><td>60.31</td><td>3.48</td><td>3.26</td><td>3.61</td><td>0.52</td><td>582</td></tr><tr><td>S&amp;T</td><td>17.88</td><td>7.51</td><td>11.77</td><td>3.52</td><td>2.73</td><td>3.26</td><td>85.00</td><td>79.24</td><td>86.40</td><td>4.67</td><td>3.87</td><td>4.79</td><td>0.53</td><td>1,953</td></tr><tr><td>Mistral 7B</td><td>6.52</td><td>3.18</td><td>4.57</td><td>2.45</td><td>2.13</td><td>2.31</td><td>69.77</td><td>65.95</td><td>71.09</td><td>3.09</td><td>3.24</td><td>3.16</td><td>4.64</td><td>7,241</td></tr><tr><td>PivotRR (ours)</td><td>17.88</td><td>17.54</td><td>13.12</td><td>3.52</td><td>3.09</td><td>3.28</td><td>63.72</td><td>64.42</td><td>63.35</td><td>3.71</td><td>3.45</td><td>3.81</td><td>0.96</td><td>1,348</td></tr><tr><td>NeutralRR (ours)</td><td>17.59</td><td>17.87</td><td>12.90</td><td>3.53</td><td>3.08</td><td>3.29</td><td>64.34</td><td>65.43</td><td>64.76</td><td>3.76</td><td>3.49</td><td>3.89</td><td>0.99</td><td>1,348</td></tr><tr><td rowspan="5">zh</td><td>M2MS</td><td>17.95</td><td>24.13</td><td>16.32</td><td>3.58</td><td>3.14</td><td>3.31</td><td>61.95</td><td>60.23</td><td>60.56</td><td>3.40</td><td>3.20</td><td>3.39</td><td>0.40</td><td>582</td></tr><tr><td>S&amp;T</td><td>13.51</td><td>24.13</td><td>12.11</td><td>3.48</td><td>3.14</td><td>3.25</td><td>83.61</td><td>82.50</td><td>82.09</td><td>4.26</td><td>4.10</td><td>4.29</td><td>0.52</td><td>1,953</td></tr><tr><td>Mistral 7B</td><td>4.58</td><td>3.93</td><td>3.68</td><td>2.47</td><td>2.02</td><td>2.39</td><td>67.28</td><td>66.40</td><td>66.98</td><td>3.19</td><td>2.98</td><td>3.15</td><td>11.48</td><td>7,241</td></tr><tr><td>PivotRR (ours)</td><td>18.32</td><td>24.13</td><td>16.36</td><td>3.60</td><td>3.14</td><td>3.36</td><td>64.73</td><td>62.99</td><td>61.90</td><td>3.54</td><td>3.37</td><td>3.54</td><td>0.89</td><td>1,348</td></tr><tr><td>NeutralRR (ours)</td><td>18.34</td><td>23.72</td><td>16.37</td><td>3.61</td><td>3.18</td><td>3.35</td><td>66.94</td><td>63.74</td><td>63.23</td><td>3.63</td><td>3.43</td><td>3.62</td><td>0.90</td><td>1,348</td></tr><tr><td rowspan="4">rest</td><td>M2MS</td><td>16.73</td><td>23.83</td><td>13.83</td><td>3.48</td><td>3.07</td><td>3.23</td><td>60.50</td><td>60.33</td><td>61.13</td><td>3.55</td><td>3.15</td><td>3.54</td><td>0.56</td><td>582</td></tr><tr><td>S&amp;T</td><td>11.88</td><td>7.63</td><td>11.41</td><td>3.38</td><td>2.72</td><td>3.20</td><td>85.63</td><td>80.38</td><td>85.67</td><td>4.71</td><td>3.88</td><td>4.75</td><td>0.59</td><td>1,953</td></tr><tr><td>PivotRR (ours)</td><td>16.32</td><td>23.56</td><td>13.66</td><td>3.50</td><td>3.12</td><td>3.25</td><td>63.63</td><td>62.04</td><td>63.28</td><td>3.72</td><td>3.33</td><td>3.73</td><td>0.98</td><td>1,348</td></tr><tr><td>NeutralRR (ours)</td><td>16.48</td><td>23.01</td><td>13.75</td><td>3.51</td><td>3.12</td><td>3.27</td><td>65.37</td><td>63.30</td><td>64.62</td><td>3.83</td><td>3.39</td><td>3.82</td><td>1.02</td><td>1,348</td></tr></table>
+
+Table 2: Results of evaluated methods in CrossSum for multi-target cross-lingual summarization using different languages as the source language. The language in each column is the target, with "rest" indicating the average for the remaining target languages. Metrics with (R) evaluate similarity to reference summaries, while those with (C) evaluate semantic coherence across languages. ROUGE-2 and CometKiwi range from 0 to 100, while BLASER 2.0 ranges from 1 to 5 (higher values are better). Best results are bold, second best results are underlined. Columns T and #P indicate the average computation time per generated summary in seconds and the number of model parameters in millions, respectively.
+
+# 5.4 Main Results
+
+In this section, we present results on MTXLS considering all the seven languages mentioned in Section 5.1 as targets. To perform this task, we took each of the seven languages as the source in turn and discarded the clusters that lacked a document in the source language. Then, we iterated through the remaining clusters taking the document in the source language as the input for summarization and we generated summaries for all the languages in the cluster, including the source language, using each of the methods mentioned in Section 5.2.
+
+The results are in Table 2 and are presented per language pair. Due to space limitations, we present detailed results only for English (en) and Chinese (zh), and show the averages for the remaining source and target languages (rest). An extended version of this table, including detailed results for more languages, confidence intervals, and the accuracy of each approach on following the target language is shown in Appendix D.1. When the source and target languages are the same, S&T and PivotRR reduce to M2MS because we use the source language as the pivot. Consequently, the results of these three methods for ROUGE-2 and BLASER 2.0 (R) coincide for en→en and zh→zh.
+
+We begin by discussing the results of Mistral 7B, as these deserve special attention. Interestingly, the model always performs worst in terms of simi
+
+larity to the reference summaries (ROUGE-2 and BLASER 2.0 (R)), even though it was instructed that the articles were obtained from the BBC and that the summaries should follow the BBC style (see Appendix C). Regarding the coherence across target languages, we observe that the model has a very decent performance, as illustrated in Table 1, ranking second in CometKiwi scores, only surpassed by S&T. However, the model often failed to produce the output in the requested format, in which case we had to repeat the request, or did not produce text in the specified target language (see Appendix D.1). For these reasons, we did not extend its evaluation to other source languages beyond English and Chinese.
+
+The method M2MS conducts cross-lingual summarization for each target language independently, disregarding semantic coherence across languages. Consequently, it consistently achieves the highest ROUGE-2 scores but ranks lowest in coherence metrics (CometKiwi and BLASER 2.0 (C)). Conversely, S&T ensures the best semantic coherence across target languages by directly translating the source language summary for each target language. However, this often results in significant degradation in similarity with the references for each target language, as measured by ROUGE-2, and, in many cases, even diminishes similarity to the reference summary for the source language, as measured by
+
+![](images/0cbaefc2f129d676c361b3a2c2a9e75b00757ff19918e2d44cc79b2700a8a8ed.jpg)
+
+![](images/caf1bcc8efa2d520ec290ff3ce47d4db1542f1ea6e5f9fbabf87b1ca46acc3de.jpg)
+
+![](images/26b3b082615a555e8633edf2de712b0836d78697f8a51ffd71f815842f63e6a0.jpg)  
+Figure 2: Results of PivotRR and NeutralRR as a function of the number of candidates per target language for re-ranking ( $k$  in Algorithm 1). Error bars indicate the standard deviations across the target languages.
+
+![](images/1cae351b7543e59f9d4d10b17f1f6c3ce71ff355bceea15afd31d0061565799f.jpg)
+
+BLASER 2.0 (R). This indicates that the MT model introduced errors compromising summary quality.
+
+Our approaches (PivotRR and NeutralRR) do not significantly degrade ROUGE-2 scores compared to M2MS and notably achieve the highest similarity to the reference summary for the source language. As expected, our methods also significantly improve semantic coherence across different target languages compared to M2MS. NeutralRR performs comparably to PivotRR in terms of similarity to the reference summaries, and consistently outperforms it in terms of semantic coherence across target languages. This was expected because NeutralRR treats all languages equally and aims for a set of summaries with high similarity. Conversely, PivotRR utilizes a fixed pivot summary and seeks candidates in each target language that closely resemble the pivot.
+
+# 5.5 Effect of Varying the Number of Candidates
+
+In this experiment, we investigate how the performance of our methods changes as we vary the number of candidates for re-ranking, using English as the source language. To vary the number of candidates generated by beam search multinomial sampling, we kept the number of beams per output sequence constant and equal to 5 and varied the number of output sequences. The results are in Figure 2, where we show the averages and standard deviations across the seven target languages.
+
+Interestingly, increasing the number of candidates does not affect the similarity between the selected summaries and their respective references,
+
+as evaluated by ROUGE-2. In addition, it has a positive effect on the similarity between the selected summaries and the reference in the source language, as measured by BLASER 2.0 (R). We justify this observation by the hypothesis that a set of summaries with high similarity can serve as a reliable indicator of summary quality, since it is unlikely that the model generates the same false information in multiple languages. This was illustrated in the example in Table 1 for NeutralRR. Finally, as more candidates are considered, computation time increases, yet so does the similarity of selected summaries, as evaluated by CometKiwi. Notably, this similarity increase is more significant for NeutralRR, which is not limited by maximizing similarity to a fixed pivot summary.
+
+# 6 Conclusion
+
+This work introduces multi-target cross-lingual summarization to address the challenge of achieving coherent summaries across multiple target languages. We propose two re-ranking approaches tailored to this task, which improve semantic coherence across languages compared to conventional beam search decoding, while still preserving similarity to the reference summaries. In particular, one of these methods eliminates the need for a pivot language, thus treating all languages equally and eliminating potential biases arising from pivot language selection. Furthermore, we extended the evaluation framework for cross-lingual summarization by including the assessment of semantic coherence across different target languages.
+
+# Limitations
+
+While we believe that our approach has merit, it is equally important to recognize its inherent limitations. First, we anticipate that as large language models continue to improve and become fluent in more languages, instructing the model to produce summaries with identical information for all target languages will eventually be sufficient to satisfy our semantic coherence constraint. However, due to the autoregressive nature of language models, the order in which summaries are decoded in different languages may introduce bias, which is not present in our approach. Second, the success of our re-ranking approaches depends on the quality of the sampled candidates. If all candidates are of low quality, or if they have poor semantic coherence across target languages, our approaches will inevitably fail. Investigating computationally efficient ways to incorporate the semantic coherence constraint directly at decoding time is an interesting research direction. Finally, our method introduces increased computational complexity compared to the usual beam search decoding.
+
+In our experimental evaluation, we acknowledge it would be valuable to conduct further validation through more datasets and human evaluation. Using more datasets would be difficult as most cross-lingual summarization datasets are presented as pairs of source-language document and target-language summary, limiting applicability to our multi-target setting. Human evaluation was constrained by the need for a pool of polyglot speakers and a considerable sample size, which were not feasible. Nevertheless, we believe our contributions are valuable as they identify a gap in cross-lingual summarization and propose a principled method to address it, irrespective of the particular performance of the models implementing it.
+
+# Acknowledgements
+
+The authors thank André F. T. Martins (Unbabel, Instituto Superior Tecnico, and Instituto de Telecomunicações) and João Gante (Hugging Face) for their insightful comments and suggestions, which were very helpful in shaping and implementing the ideas presented in this work. This research was supported by the Portuguese Recovery and Resilience Plan through project C645008882-00000055 (i.e., the Center For Responsible AI).
+
+# References
+
+Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the Association for Computational Linguistics, 7:597-610.  
+Yu Bai, Yang Gao, and Heyan Huang. 2021. Crosslingual abstractive summarization with limited parallel resources. 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 6910-6924, Online. Association for Computational Linguistics.  
+Abhik Bhattacharjee, Tahmid Hasan, Wasi Uddin Ahmad, Yuan-Fang Li, Yong-Bin Kang, and Rifat Shahriyar. 2023. *CrossSum: Beyond English-centric cross-lingual summarization for 1,500+ language pairs*. In *Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics* (Volume 1: Long Papers), pages 2541–2564, Toronto, Canada. Association for Computational Linguistics.  
+Yue Cao, Hui Liu, and Xiaojun Wan. 2020. Jointly learning to align and summarize for neural cross-lingual summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6220–6231, Online. Association for Computational Linguistics.  
+Mingda Chen, Paul-Ambroise Duquenne, Pierre Andrews, Justine Kao, Alexandre Mourachko, Holger Schwenk, and Marta R. Costa-jussa. 2023a. BLASER: A text-free speech-to-speech translation evaluation metric. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9064-9079, Toronto, Canada. Association for Computational Linguistics.  
+Yulong Chen, Huajian Zhang, Yijie Zhou, Xuefeng Bai, Yueguan Wang, Ming Zhong, Jianhao Yan, Yafu Li, Judy Li, Xianchao Zhu, and Yue Zhang. 2023b. Revisiting cross-lingual summarization: A corpus-based study and a new benchmark with improved annotation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9332–9351, Toronto, Canada. Association for Computational Linguistics.  
+Seamless Communication, Loic Barrault, Yu-An Chung, Mariano Cora Meglioli, David Dale, Ning Dong, Paul-Ambroise Duquenne, Hady Elsahar, Hongyu Gong, Kevin Heffernan, John Hoffman, Christopher Klaiber, Pengwei Li, Daniel Licht, Jean Maillard, Alice Rakotoarison, Kaushik Ram Sadagopan, Guillaume Wenzek, Ethan Ye, Bapi Akula, Peng-Jen Chen, Naji El Hachem, Brian Ellis, Gabriel Mejia Gonzalez, Justin Haaheim, Prangthip Hansanti, Russ Howes, Bernie Huang, Min-Jae Hwang, Hirofumi Inaguma, Somya Jain, Elahe Kalbassi, Amanda Kallet,
+
+Ilia Kulikov, Janice Lam, Daniel Li, Xutai Ma, Ruslan Mavlyutov, Benjamin Peloquin, Mohamed Ramadan, Abinesh Ramakrishnan, Anna Sun, Kevin Tran, Tuan Tran, Igor Tufanov, Vish Vogeti, Carleigh Wood, Yilin Yang, Bokai Yu, Pierre Andrews, Can Balioglu, Marta R. Costa-jussa, Onur Celebi, Maha Elbayad, Cynthia Gao, Francisco Guzmán, Justine Kao, Ann Lee, Alexandre Mourachko, Juan Pino, Sravya Popuri, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, Paden Tomasello, Changhan Wang, Jeff Wang, and Skyler Wang. 2023. SeamlessM4T: Massively multilingual & multimodal machine translation. arXiv preprint arXiv:2308.11596.  
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishray Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle-moyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440-8451, Online. Association for Computational Linguistics.  
+Marta R Costa-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, et al. 2022. No language left behind: Scaling human-centered machine translation. arXiv preprint arXiv:2207.04672.  
+Paul-Ambroise Duquenne, Holger Schwenk, and Benoit Sagot. 2023. SONAR: Sentence-level multimodal and language-agnostic representations. arXiv preprint arXiv:2308.11466.  
+Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2018. Examining the tip of the iceberg: A data set for idiom translation. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA).  
+Corinne Feremans, Martine Labbe, and Gilbert Laporte. 2003. Generalized network design problems. European Journal of Operational Research, 148(1):1-13.  
+Tahmid Hasan, Abhik Bhattacharjee, Md. Saiful Islam, Kazi Mubasshir, Yuan-Fang Li, Yong-Bin Kang, M. Sohel Rahman, and Rifat Shahriyar. 2021. XLsum: Large-scale multilingual abstractive summarization for 44 languages. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4693-4703, Online. Association for Computational Linguistics.  
+Fantine Huot, Joshua Maynez, Chris Alberti, Reinald Kim Amplayo, Priyanka Agrawal, Stanza Fierro, Shashi Narayan, and Mirella Lapata. 2024.  $\mu$ PLAN: Summarizing using a content plan as cross-lingual bridge. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2146–2163, St. Julian's, Malta. Association for Computational Linguistics.
+
+Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825.  
+Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Herve Jégou, and Tomas Mikolov. 2016a. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651.  
+Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2016b. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759.  
+Fabio Kepler, Jonay Trénous, Marcos Treviso, Miguel Vera, Antonio Góis, M. Amin Farajian, Antonio V. Lopes, and André F. T. Martins. 2019a. Unbabel's participation in the WMT19 translation quality estimation shared task. In Proceedings of the Fourth Conference on Machine Translation (Volume 3: Shared Task Papers, Day 2), pages 78-84, Florence, Italy. Association for Computational Linguistics.  
+Fabio Kepler, Jonay Trénous, Marcos Treviso, Miguel Vera, and André F. T. Martins. 2019b. OpenKiwi: An open source framework for quality estimation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 117-122, Florence, Italy. Association for Computational Linguistics.  
+Faisal Ladhak, Esin Durmus, Claire Cardie, and Kathleen McKeown. 2020. WikiLingua: A new benchmark dataset for cross-lingual abstractive summarization. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4034-4048, Online. Association for Computational Linguistics.  
+Dongjun Lee. 2020. Two-phase cross-lingual language model fine-tuning for machine translation quality estimation. In Proceedings of the Fifth Conference on Machine Translation, pages 1024–1028, Online. Association for Computational Linguistics.  
+Yunlong Liang, Fandong Meng, Chulun Zhou, Jinan Xu, Yufeng Chen, Jinsong Su, and Jie Zhou. 2022. A variational hierarchical model for neural cross-lingual summarization. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2088–2099, Dublin, Ireland. Association for Computational Linguistics.  
+Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain. Association for Computational Linguistics.  
+Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and
+
+Luke Zettlemoyer. 2020. Multilingual denoising pretraining for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742.  
+Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020. On faithfulness and factuality in abstractive summarization. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1906-1919, Online. Association for Computational Linguistics.  
+Laura Perez-Beltrachini and Mirella Lapata. 2021. Models and datasets for cross-lingual summarisation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9408–9423, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.  
+Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Improving language understanding by generative pre-training. Technical report, OpenAI.  
+Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685-2702, Online. Association for Computational Linguistics.  
+Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and André F. T. Martins. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. In Proceedings of the Seventh Conference on Machine Translation (WMT), pages 634-645, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.  
+Lucia Specia, Carolina Scarton, and Gustavo Henrique Paetzold. 2018. Quality estimation for machine translation, volume 11. Springer.  
+Lucia Specia, Kashif Shah, Jose G.C. de Souza, and Trevor Cohn. 2013. QuEst - a translation quality estimation framework. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 79-84, Sofia, Bulgaria. Association for Computational Linguistics.  
+Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, and Angela Fan. 2021. Multilingual translation from denoising pre-training. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3450-3466, Online. Association for Computational Linguistics.  
+Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Riviere, Mihir Sanjay
+
+Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Antonia Paterson, Beth Tsai, Bobak Shahriari, Charline Le Lan, Christopher A. Choquette-Choo, Clément Crepy, Daniel Cer, Daphne Ippolito, David Reid, Elena Buchatskaya, Eric Ni, Eric Noland, Geng Yan, George Tucker, George-Christian Muraru, Grigory Rozhdestvenskiy, Henryk Michalewski, Ian Tenney, Ivan Grishchenko, Jacob Austin, James Keeling, Jane Labanowski, Jean-Baptiste Lespiau, Jeff Stanway, Jenny Brennan, Jeremy Chen, Johan Ferret, Justin Chiu, Justin Mao-Jones, Katherine Lee, Kathy Yu, Katie Millican, Lars Lowe Sjoesund, Lisa Lee, Lucas Dixon, Machel Reid, Maciej Mikuła, Mateo Wirth, Michael Sharman, Nikolai Chinaev, Nithum Thain, Olivier Bachem, Oscar Chang, Oscar Wahltinez, Paige Bailey, Paul Michel, Petko Yotov, Pier Giuseppe Sessa, Rahma Chaabouni, Ramona Comanescu, Reena Jana, Rohan Anil, Ross McIlroy, Ruibo Liu, Ryan Mullins, Samuel L Smith, Sebastian Borgeaud, Sertan Girgin, Sholto Douglas, Shree Pandya, Siamak Shakeri, Soham De, Ted Klimenko, Tom Hennigan, Vlad Feinberg, Wojciech Stokowiec, Yu hui Chen, Zafarali Ahmed, Zhitao Gong, Tris Warkentin, Ludovic Peran, Minh Giang, Clément Farabet, Oriol Vinyals, Jeff Dean, Koray Kavukcuoglu, Demis Hassabis, Zoubin Ghahramani, Douglas Eck, Joelle Barral, Fernando Pereira, Eli Collins, Armand Joulin, Noah Fiedel, Evan Senter, Alek Andreev, and Kathleen Kenealy. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295.  
+Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. 2023. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288.  
+Ashok Urlana, Pinzhen Chen, Zheng Zhao, Shay Cohen, Manish Shrivastava, and Barry Haddow. 2023. PMIndiaSum: Multilingual and cross-lingual headline summarization for languages in India. In Findings of the Association for Computational Linguistics: EMNLP
+
+2023, pages 11606-11628, Singapore. Association for Computational Linguistics.  
+Ashwin Vijayakumar, Michael Cogswell, Ramprasaath Selvaraju, Qing Sun, Stefan Lee, David Crandall, and Dhruv Batra. 2018. Diverse beam search for improved description of complex scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32.  
+Jiaan Wang, Yunlong Liang, Fandong Meng, Beiqi Zou, Zhixu Li, Jianfeng Qu, and Jie Zhou. 2023a. Zero-shot cross-lingual summarization via large language models. In Proceedings of the 4th New Frontiers in Summarization Workshop, pages 12-23, Singapore. Association for Computational Linguistics.  
+Jiaan Wang, Fandong Meng, Duo Zheng, Yunlong Liang, Zhixu Li, Jianfeng Qu, and Jie Zhou. 2022. A survey on cross-lingual summarization. Transactions of the Association for Computational Linguistics, 10:1304-1323.  
+Jiaan Wang, Fandong Meng, Duo Zheng, Yunlong Liang, Zhixu Li, Jianfeng Qu, and Jie Zhou. 2023b. Towards unifying multi-lingual and cross-lingual summarization. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15127–15143, Toronto, Canada. Association for Computational Linguistics.  
+Jiayi Wang, Kai Fan, Bo Li, Fengming Zhou, Boxing Chen, Yangbin Shi, and Luo Si. 2018. Alibaba submission for WMT18 quality estimation task. In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 809-815, Belgium, Brussels. Association for Computational Linguistics.  
+Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483-498, Online. Association for Computational Linguistics.  
+Junnan Zhu, Qian Wang, Yining Wang, Yu Zhou, Jiajun Zhang, Shaonan Wang, and Chengqing Zong. 2019. NCLS: Neural cross-lingual summarization. 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 3054-3064, Hong Kong, China. Association for Computational Linguistics.  
+A Dataset Clustering and Analysis  
+As mentioned in Section 5.1, the original CrossSum dataset presents documents in one language paired with summaries in another language, a format that does not serve our multi-target setting. Therefore,
+
+we clustered the dataset to obtain clusters of multilingual document-summary pairs about the same story. To achieve this, we aggregated all documents across the mentioned languages and constructed an undirected graph representing their pairwise connections. In this graph, two documents in different languages are connected if they are paired in CrossSum. We then built clusters by extracting all maximal cliques from this graph and we discarded all singleton cliques. Consequently, each maximal clique is a cluster of up to seven multilingual documents pertaining to the same story, where each document is accompanied by a summary in its respective language.
+
+This clustering procedure was applied separately to the CrossSum validation and test splits. The resulting validation set consisted of 4,525 clusters and 10,479 documents, while the test set consisted of 4,560 clusters and 10,535 documents. Table 3 provides a breakdown of cluster sizes in the test set, as well as the distribution of documents for each language and cluster size. Notably, none of the clusters in the test set are complete, indicating that no cluster includes a document for all seven languages considered. In addition, we conducted an analysis of the co-occurrence of different language pairs within the clusters to verify whether a robust evaluation of cross-lingual summarization was possible across all language directions. Figure 3 illustrates the distribution of clusters containing examples of each language pair. While certain language pairs have higher representation than others, it is noteworthy that even the least represented pair (fr, zh) is found in 35 clusters, indicating a diverse linguistic coverage across the dataset.
+
+Since one of our goals is to assess the semantic coherence of the generated summaries in different target languages, it is crucial to evaluate the coherence of reference summary clusters in this regard. This evaluation helps to determine the level of coherence that can be achieved in the generated summaries without degrading similarity to the reference summaries. To achieve this, we computed BLASER 2.0 and CometKiwi scores between reference summaries within the same cluster for each language pair. The results are shown in Figure 4. It is important to note that the matrices are nonsymmetric due to the nature of BLASER 2.0 and CometKiwi metrics. Firstly, we note a significant agreement between the two metrics, as anticipated. Additionally, coherence tends to be higher among
+
+<table><tr><td rowspan="2">Language</td><td colspan="7">Cluster Size</td></tr><tr><td>2</td><td>3</td><td>4</td><td>5</td><td>6</td><td>7</td><td>All</td></tr><tr><td>ar</td><td>1,022</td><td>455</td><td>153</td><td>34</td><td>6</td><td>0</td><td>1,670</td></tr><tr><td>en</td><td>1,780</td><td>598</td><td>176</td><td>37</td><td>7</td><td>0</td><td>2,598</td></tr><tr><td>es</td><td>1,271</td><td>367</td><td>130</td><td>31</td><td>7</td><td>0</td><td>1,806</td></tr><tr><td>fr</td><td>224</td><td>84</td><td>46</td><td>14</td><td>5</td><td>0</td><td>373</td></tr><tr><td>pt</td><td>1,027</td><td>280</td><td>118</td><td>33</td><td>6</td><td>0</td><td>1,464</td></tr><tr><td>ru</td><td>1,077</td><td>482</td><td>140</td><td>38</td><td>5</td><td>0</td><td>1,742</td></tr><tr><td>zh</td><td>531</td><td>224</td><td>93</td><td>28</td><td>6</td><td>0</td><td>882</td></tr><tr><td>All</td><td>3,466</td><td>830</td><td>214</td><td>43</td><td>7</td><td>0</td><td>4,560</td></tr></table>
+
+Table 3: Number of clusters in the test set containing a document of each language, organized by cluster size.
+
+![](images/96ff2d1290ac350e3e782e6e0a79640275d62d4737e351ac489725d1d6ae850b.jpg)  
+Figure 3: Number of clusters in the test set containing documents of each language pair.
+
+languages using the Latin script. However, for most language pairs, coherence remains above 3.40 BLASER 2.0 points and 70.0 CometKiwi points. This suggests room for improvement compared to the results outlined in Table 2.
+
+# B Implementation Details
+
+To represent the cross-lingual summarization distribution  $q(\boldsymbol{y}_t \mid \boldsymbol{x}_o, t)$ , we use an mT5 model (Xue et al., 2021) for all the methods except Mistral 7B. mT5 allows us to perform summarization across all language directions by conditioning the decoder on a unique start-of-sequence token that specifies the intended target language.
+
+We used the publicly available SONAR checkpoint text_sonar_basicEncoder to implement  $\phi$ , the mT5 checkpoint csebuetnp/mT5_m2m.CrossSum_enhanced, which was fine-tuned in the CrossSum dataset, and the Mistral 7B checkpoint mistralai/Mistral-7B-Instruct-v0.2. All of these checkpoints are available at the Hugging
+
+Face model hub.3
+
+The optimal beam size and sampling temperature for beam search multinomial sampling were determined through a grid search. We explored beam sizes of 1, 3, and 5, and temperatures of 0.1, 0.3, 0.5, 1.0, 1.5, and 2.0 in order to maximize the ROUGE-2 score on the validation set of English-to-all summarization. We also tried with other decoding strategies, namely (single-beam) multinomial sampling and diverse beam search (Vijayakumar et al., 2018), but these degraded ROUGE scores considerably. The number of random language permutations ( $m$  in Algorithm 1) used by NeutralRR was set to 6 when the number of target languages was at least three and was set to 2 if there were only two target languages, since there are only two possible permutations of two languages.
+
+Regarding the evaluation metrics, we used the multilingual implementation of ROUGE by Hasan et al. (2021). For CometKiwi and BLASER 2.0, we used the Unbabel/wmt22-cometkiwi-da and blaser_2_0_qe checkpoints, respectively.
+
+All experiments were run on an 80-core Intel Xeon Gold 5218R CPU @ 2.10GHz with 800GB of RAM and an NVIDIA A100 GPU with 80GB of memory.
+
+# C LLM Prompt
+
+The following prompt was used on the experiments with Mistral 7B:
+
+For the <source-lang> news article from BBC written below, provide a summary in <target-lang_1>, a summary in <target-lang_2>, ... and a summary in <target-lang_N>. All summaries should be one or two sentences long and follow the style of BBC. All summaries must contain the same information. Present the answer in the format of a JSON object where the keys are the language codes and the values are the summaries.
+
+Text:
+
+<source_document>
+
+![](images/030c617ce35b9842eb43905a71240917b7543d118cc81464ffd5fb7e7614a67f.jpg)  
+(a)
+
+![](images/8891a957c2aa4118553c3a9183b8a3e137d366df0b1609a7120056b321170147.jpg)  
+(b)
+
+![](images/00872c69f26696e3801d8b7edff8060b56bd566765c8d1eba06e5e10fc36842e.jpg)  
+Figure 4: Average BLASER 2.0 (a) and CometKiwi (b) scores between reference summaries within the same cluster for each language pair in the test set.
+
+![](images/540113e1b051fd794bc15c73c3ba17ba9caabb4277c4859fd3e76f30d800c71b.jpg)
+
+![](images/459fcb0e4dbb9d9abab600f02541a644635a7d01feec19528fcf5599e3350aba.jpg)  
+Figure 5: Effect of varying the number of language permutations ( $m$  in Algorithm 1) on the results of NeutralRR. The results of doing an exhaustive search for the most coherent set are also shown for comparison. The error bars and the shaded area indicate the standard deviations across the target languages.
+
+![](images/f928763c892029e9bd1869b7e74c8ea5c49e11bb1a9a8e8b87324dcfe5e451ca.jpg)
+
+# D Further Experimental Results
+
+# D.1 Main Results Extended
+
+An extended version of the results presented in Table 2 is shown in Tables 4 and 5. In addition to English and Chinese, we also show results for Spanish and French. Spanish is the second most represented language in the dataset, surpassed only by English, while French is the least represented (see Table 3). All the results are accompanied by  $95\%$  bootstrap confidence intervals with 1,000 resamples. Apart from the metrics mentioned in Section 5.3, we also include the target language accuracy in Table 4. This metric corresponds to the percentage of times a method generated text in
+
+the specified target language, and is calculated by comparing the specified language with the dominant language identified in the generated text by the fastText model (Joulin et al., 2016a,b). We observe that the mT5-based methods generate text in the correct target language in the vast majority (if not all) of the cases. Mistral 7B sometimes struggles to generate text in the correct target language, especially in Arabic.
+
+# D.2 Effect of the Heuristic Search
+
+In this experiment, we investigate the effect of the number of language permutations (m in Algorithm 1) on the performance of NeutralRR. In this experiment, we always use English as the source language and only consider clusters of documents with 4 languages, allowing up to 24 language permutations. The number of candidate summaries
+
+per language is kept fixed at 8. For this cluster size and number of candidates, maximizing  $\varphi$  (equation (6)) directly with an exhaustive search is feasible since there are only  $8^4 = 4,096$  possible sets of summaries. Therefore, we also compare the results of our approach with the exhaustive search. The results are shown in Figure 5.
+
+The first observation is that changing  $m$  or performing an exhaustive search does not significantly affect the similarity to the reference summaries. Changing  $m$  also has no significant effect on the computation time, which is natural since the time required by the dynamic programming optimization is much smaller than the decoding time of the summarization model. However, an exhaustive search obviously increases the computation time, and the difference would only become larger for larger cluster sizes or more candidate summaries per language. Regarding the semantic coherence of the resulting set of summaries, an exhaustive search yields the best results as expected, but they are only slightly better than our heuristic search with a sufficiently large number of language permutations.
+
+<table><tr><td rowspan="2">Source</td><td rowspan="2">Method</td><td colspan="5">ROUGE-2 (R)</td><td colspan="5">BLASER 2.0 (R)</td><td colspan="5">Target Lang. Acc.</td></tr><tr><td>en</td><td>es</td><td>fr</td><td>zh</td><td>rest</td><td>en</td><td>es</td><td>fr</td><td>zh</td><td>rest</td><td>en</td><td>es</td><td>fr</td><td>zh</td><td>rest</td></tr><tr><td rowspan="5">en</td><td>M2MS</td><td>17.88 ± 0.59</td><td>13.91 ± 0.97</td><td>21.55 ± 3.14</td><td>18.63 ± 1.82</td><td>10.19 ± 0.83</td><td>3.52 ± 0.02</td><td>3.03 ± 0.03</td><td>3.31 ± 0.08</td><td>3.04 ± 0.05</td><td>3.31 ± 0.03</td><td>100.0</td><td>99.7</td><td>100.0</td><td>97.8</td><td>99.6</td></tr><tr><td>S&amp;T</td><td>17.88 ± 0.59</td><td>11.91 ± 0.84</td><td>18.98 ± 2.52</td><td>7.51 ± 0.78</td><td>9.33 ± 0.77</td><td>3.52 ± 0.02</td><td>3.07 ± 0.04</td><td>3.30 ± 0.07</td><td>2.73 ± 0.04</td><td>3.31 ± 0.04</td><td>100.0</td><td>99.9</td><td>100.0</td><td>99.8</td><td>99.9</td></tr><tr><td>Mistral 7B</td><td>6.52 ± 0.28</td><td>6.68 ± 0.58</td><td>7.86 ± 1.57</td><td>3.18 ± 0.37</td><td>2.78 ± 0.29</td><td>2.45 ± 0.02</td><td>2.17 ± 0.03</td><td>2.46 ± 0.08</td><td>2.13 ± 0.03</td><td>2.31 ± 0.03</td><td>99.7</td><td>99.8</td><td>100.0</td><td>99.2</td><td>82.2</td></tr><tr><td>PivotRR</td><td>17.88 ± 0.59</td><td>13.58 ± 0.96</td><td>20.77 ± 3.08</td><td>17.54 ± 1.69</td><td>10.42 ± 0.85</td><td>3.52 ± 0.02</td><td>3.06 ± 0.03</td><td>3.32 ± 0.08</td><td>3.09 ± 0.04</td><td>3.34 ± 0.03</td><td>100.0</td><td>99.4</td><td>100.0</td><td>98.8</td><td>99.6</td></tr><tr><td>NeutralRR</td><td>17.59 ± 0.58</td><td>13.70 ± 0.94</td><td>20.08 ± 3.24</td><td>17.87 ± 1.75</td><td>10.23 ± 0.86</td><td>3.53 ± 0.02</td><td>3.06 ± 0.03</td><td>3.34 ± 0.08</td><td>3.08 ± 0.04</td><td>3.34 ± 0.03</td><td>100.0</td><td>99.7</td><td>99.5</td><td>99.4</td><td>99.5</td></tr><tr><td rowspan="4">es</td><td>M2MS</td><td>15.88 ± 1.26</td><td>14.37 ± 0.66</td><td>20.48 ± 4.77</td><td>20.01 ± 3.38</td><td>9.12 ± 1.04</td><td>3.40 ± 0.04</td><td>2.99 ± 0.02</td><td>3.26 ± 0.15</td><td>2.92 ± 0.10</td><td>3.20 ± 0.04</td><td>99.7</td><td>99.8</td><td>100.0</td><td>97.9</td><td>99.9</td></tr><tr><td>S&amp;T</td><td>10.71 ± 0.80</td><td>14.37 ± 0.66</td><td>14.28 ± 2.91</td><td>7.99 ± 1.31</td><td>8.01 ± 0.85</td><td>3.29 ± 0.04</td><td>2.99 ± 0.02</td><td>3.16 ± 0.12</td><td>2.60 ± 0.07</td><td>3.15 ± 0.04</td><td>100.0</td><td>99.8</td><td>100.0</td><td>99.5</td><td>99.8</td></tr><tr><td>PivotRR</td><td>15.35 ± 1.20</td><td>14.37 ± 0.66</td><td>18.70 ± 4.55</td><td>19.72 ± 3.03</td><td>9.37 ± 1.02</td><td>3.42 ± 0.04</td><td>2.99 ± 0.02</td><td>3.27 ± 0.14</td><td>2.95 ± 0.09</td><td>3.23 ± 0.04</td><td>99.7</td><td>99.8</td><td>99.0</td><td>99.5</td><td>99.7</td></tr><tr><td>NeutralRR</td><td>15.59 ± 1.25</td><td>14.68 ± 0.65</td><td>20.68 ± 4.61</td><td>19.69 ± 3.07</td><td>9.54 ± 1.07</td><td>3.43 ± 0.04</td><td>3.05 ± 0.02</td><td>3.33 ± 0.14</td><td>2.98 ± 0.08</td><td>3.23 ± 0.04</td><td>99.7</td><td>99.9</td><td>99.0</td><td>100.0</td><td>99.7</td></tr><tr><td rowspan="4">fr</td><td>M2MS</td><td>21.25 ± 3.19</td><td>15.49 ± 3.49</td><td>23.78 ± 2.46</td><td>38.04 ± 10.21</td><td>13.38 ± 3.65</td><td>3.57 ± 0.08</td><td>3.05 ± 0.10</td><td>3.33 ± 0.07</td><td>3.24 ± 0.24</td><td>3.29 ± 0.11</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td></tr><tr><td>S&amp;T</td><td>16.14 ± 2.08</td><td>13.19 ± 2.56</td><td>23.78 ± 2.46</td><td>9.87 ± 3.11</td><td>11.01 ± 2.54</td><td>3.43 ± 0.08</td><td>3.11 ± 0.11</td><td>3.33 ± 0.07</td><td>2.78 ± 0.18</td><td>3.28 ± 0.10</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td></tr><tr><td>PivotRR</td><td>20.11 ± 2.98</td><td>15.31 ± 3.00</td><td>23.78 ± 2.46</td><td>38.21 ± 9.44</td><td>13.16 ± 3.33</td><td>3.54 ± 0.08</td><td>3.10 ± 0.11</td><td>3.33 ± 0.07</td><td>3.35 ± 0.21</td><td>3.34 ± 0.11</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td></tr><tr><td>NeutralRR</td><td>20.95 ± 2.79</td><td>14.42 ± 3.13</td><td>22.76 ± 2.41</td><td>36.28 ± 10.27</td><td>13.18 ± 3.29</td><td>3.59 ± 0.08</td><td>3.08 ± 0.11</td><td>3.33 ± 0.06</td><td>3.26 ± 0.21</td><td>3.36 ± 0.11</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td><td>100.0</td></tr><tr><td rowspan="5">zh</td><td>M2MS</td><td>17.95 ± 1.52</td><td>15.54 ± 2.24</td><td>30.91 ± 10.19</td><td>24.13 ± 1.54</td><td>11.72 ± 1.77</td><td>3.58 ± 0.05</td><td>3.04 ± 0.07</td><td>3.50 ± 0.27</td><td>3.14 ± 0.04</td><td>3.34 ± 0.06</td><td>98.0</td><td>100.0</td><td>100.0</td><td>99.8</td><td>99.6</td></tr><tr><td>S&amp;T</td><td>13.51 ± 1.17</td><td>11.36 ± 1.42</td><td>21.79 ± 5.13</td><td>24.13 ± 1.54</td><td>9.13 ± 1.41</td><td>3.48 ± 0.04</td><td>3.00 ± 0.08</td><td>3.32 ± 0.16</td><td>3.14 ± 0.04</td><td>3.31 ± 0.07</td><td>100.0</td><td>99.5</td><td>100.0</td><td>99.8</td><td>99.3</td></tr><tr><td>Mistral 7B</td><td>4.58 ± 0.43</td><td>4.39 ± 0.65</td><td>8.22 ± 3.00</td><td>3.93 ± 0.30</td><td>1.92 ± 0.45</td><td>2.47 ± 0.04</td><td>2.26 ± 0.05</td><td>2.44 ± 0.15</td><td>2.02 ± 0.03</td><td>2.43 ± 0.05</td><td>99.8</td><td>91.6</td><td>100.0</td><td>97.4</td><td>63.1</td></tr><tr><td>PivotRR</td><td>18.32 ± 1.46</td><td>15.73 ± 2.09</td><td>30.67 ± 8.23</td><td>24.13 ± 1.54</td><td>11.80 ± 1.65</td><td>3.60 ± 0.05</td><td>3.09 ± 0.08</td><td>3.51 ± 0.18</td><td>3.14 ± 0.04</td><td>3.39 ± 0.06</td><td>98.4</td><td>100.0</td><td>100.0</td><td>99.8</td><td>99.9</td></tr><tr><td>NeutralRR</td><td>18.34 ± 1.46</td><td>15.63 ± 2.11</td><td>29.25 ± 7.83</td><td>23.72 ± 1.47</td><td>12.32 ± 1.73</td><td>3.61 ± 0.05</td><td>3.09 ± 0.07</td><td>3.50 ± 0.20</td><td>3.18 ± 0.04</td><td>3.39 ± 0.06</td><td>98.6</td><td>100.0</td><td>100.0</td><td>99.9</td><td>99.9</td></tr><tr><td rowspan="4">rest</td><td>M2MS</td><td>15.50 ± 1.15</td><td>12.82 ± 1.10</td><td>22.16 ± 5.59</td><td>20.36 ± 3.16</td><td>11.02 ± 0.99</td><td>3.48 ± 0.04</td><td>2.99 ± 0.04</td><td>3.29 ± 0.14</td><td>3.06 ± 0.08</td><td>3.30 ± 0.04</td><td>99.9</td><td>99.4</td><td>99.7</td><td>99.6</td><td>99.9</td></tr><tr><td>S&amp;T</td><td>10.84 ± 0.81</td><td>10.95 ± 0.90</td><td>14.69 ± 2.68</td><td>6.76 ± 1.15</td><td>9.52 ± 0.80</td><td>3.38 ± 0.03</td><td>3.01 ± 0.04</td><td>3.19 ± 0.10</td><td>2.74 ± 0.06</td><td>3.28 ± 0.04</td><td>100.0</td><td>99.8</td><td>100.0</td><td>100.0</td><td>99.8</td></tr><tr><td>PivotRR</td><td>15.38 ± 1.09</td><td>12.93 ± 1.08</td><td>20.91 ± 4.98</td><td>19.96 ± 3.08</td><td>11.13 ± 0.98</td><td>3.51 ± 0.04</td><td>3.03 ± 0.04</td><td>3.32 ± 0.13</td><td>3.10 ± 0.08</td><td>3.32 ± 0.04</td><td>99.9</td><td>99.4</td><td>99.6</td><td>99.8</td><td>99.9</td></tr><tr><td>NeutralRR</td><td>15.29 ± 1.13</td><td>13.04 ± 1.09</td><td>21.14 ± 4.78</td><td>19.70 ± 2.94</td><td>11.18 ± 0.99</td><td>3.50 ± 0.04</td><td>3.03 ± 0.04</td><td>3.34 ± 0.13</td><td>3.11 ± 0.08</td><td>3.34 ± 0.04</td><td>99.8</td><td>99.5</td><td>100.0</td><td>99.6</td><td>100.0</td></tr></table>
+
+Table 4: Extended results of multi-target cross-lingual summarization in CrossSum for the metrics evaluating similarity to the reference summaries. The target language accuracy is also shown.
+
+<table><tr><td rowspan="2">Source</td><td rowspan="2">Method</td><td colspan="5">CometKiwi (C)</td><td colspan="5">BLASER 2.0 (C)</td></tr><tr><td>en</td><td>es</td><td>fr</td><td>zh</td><td>rest</td><td>en</td><td>es</td><td>fr</td><td>zh</td><td>rest</td></tr><tr><td rowspan="5">en</td><td>M2MS</td><td>59.28 ± 0.56</td><td>61.79 ± 1.16</td><td>58.21 ± 2.39</td><td>61.00 ± 1.40</td><td>60.52 ± 1.06</td><td>3.48 ± 0.02</td><td>3.52 ± 0.04</td><td>3.63 ± 0.08</td><td>3.26 ± 0.01</td><td>3.6 ± 0.01</td></tr><tr><td>S&amp;T</td><td>85.00 ± 0.20</td><td>87.33 ± 0.33</td><td>87.93 ± 0.39</td><td>79.24 ± 1.15</td><td>85.58 ± 0.39</td><td>4.67 ± 0.01</td><td>4.80 ± 0.02</td><td>4.81 ± 0.03</td><td>3.87 ± 0.05</td><td>4.78 ± 0.03</td></tr><tr><td>Mistral 7B</td><td>69.77 ± 0.66</td><td>79.40 ± 0.75</td><td>77.97 ± 2.01</td><td>65.95 ± 1.04</td><td>66.04 ± 0.96</td><td>3.09 ± 0.04</td><td>3.44 ± 0.08</td><td>3.19 ± 0.15</td><td>3.24 ± 0.07</td><td>3.06 ± 0.06</td></tr><tr><td>PivotRR</td><td>63.72 ± 0.52</td><td>64.79 ± 1.15</td><td>61.51 ± 2.48</td><td>64.42 ± 1.28</td><td>63.49 ± 1.03</td><td>3.71 ± 0.02</td><td>3.73 ± 0.03</td><td>3.83 ± 0.06</td><td>3.45 ± 0.04</td><td>3.83 ± 0.03</td></tr><tr><td>NeutralRR</td><td>64.34 ± 0.54</td><td>66.12 ± 1.21</td><td>63.20 ± 2.33</td><td>65.43 ± 1.25</td><td>64.83 ± 1.04</td><td>3.76 ± 0.02</td><td>3.81 ± 0.03</td><td>3.91 ± 0.05</td><td>3.49 ± 0.04</td><td>3.91 ± 0.03</td></tr><tr><td rowspan="4">es</td><td>M2MS</td><td>61.25 ± 1.06</td><td>61.61 ± 0.72</td><td>60.40 ± 3.15</td><td>59.23 ± 2.23</td><td>61.65 ± 1.39</td><td>3.52 ± 0.04</td><td>3.44 ± 0.03</td><td>3.54 ± 0.11</td><td>3.07 ± 0.10</td><td>3.5 ± 0.01</td></tr><tr><td>S&amp;T</td><td>86.31 ± 0.33</td><td>86.36 ± 0.29</td><td>86.81 ± 1.05</td><td>79.53 ± 1.72</td><td>85.61 ± 0.59</td><td>4.69 ± 0.02</td><td>4.72 ± 0.02</td><td>4.78 ± 0.05</td><td>3.82 ± 0.07</td><td>4.76 ± 0.03</td></tr><tr><td>PivotRR</td><td>64.42 ± 1.08</td><td>65.18 ± 0.74</td><td>61.44 ± 3.10</td><td>60.56 ± 2.12</td><td>64.26 ± 1.39</td><td>3.70 ± 0.04</td><td>3.68 ± 0.03</td><td>3.65 ± 0.10</td><td>3.24 ± 0.07</td><td>3.76 ± 0.04</td></tr><tr><td>NeutralRR</td><td>65.82 ± 1.07</td><td>65.59 ± 0.75</td><td>63.79 ± 3.25</td><td>61.90 ± 2.23</td><td>64.92 ± 1.39</td><td>3.80 ± 0.03</td><td>3.78 ± 0.02</td><td>3.78 ± 0.09</td><td>3.36 ± 0.06</td><td>3.84 ± 0.04</td></tr><tr><td rowspan="4">fr</td><td>M2MS</td><td>60.63 ± 2.11</td><td>63.81 ± 2.92</td><td>61.62 ± 1.63</td><td>59.83 ± 5.47</td><td>59.98 ± 3.05</td><td>3.57 ± 0.07</td><td>3.47 ± 0.09</td><td>3.60 ± 0.05</td><td>3.04 ± 0.20</td><td>3.5 ± 0.10</td></tr><tr><td>S&amp;T</td><td>86.70 ± 0.38</td><td>87.27 ± 0.81</td><td>86.51 ± 0.57</td><td>81.17 ± 3.12</td><td>85.42 ± 1.22</td><td>4.68 ± 0.03</td><td>4.70 ± 0.05</td><td>4.76 ± 0.03</td><td>3.72 ± 0.18</td><td>4.66 ± 0.06</td></tr><tr><td>PivotRR</td><td>64.67 ± 2.02</td><td>66.29 ± 2.93</td><td>64.62 ± 1.53</td><td>62.68 ± 4.89</td><td>62.06 ± 3.14</td><td>3.74 ± 0.06</td><td>3.67 ± 0.09</td><td>3.81 ± 0.05</td><td>3.27 ± 0.14</td><td>3.66 ± 0.09</td></tr><tr><td>NeutralRR</td><td>66.34 ± 1.99</td><td>66.42 ± 3.29</td><td>65.32 ± 1.60</td><td>63.04 ± 5.42</td><td>64.02 ± 2.98</td><td>3.86 ± 0.05</td><td>3.76 ± 0.08</td><td>3.89 ± 0.05</td><td>3.25 ± 0.14</td><td>3.75 ± 0.07</td></tr><tr><td rowspan="5">zh</td><td>M2MS</td><td>61.95 ± 1.16</td><td>59.45 ± 2.13</td><td>61.08 ± 5.42</td><td>60.23 ± 1.05</td><td>60.76 ± 2.04</td><td>3.40 ± 0.04</td><td>3.21 ± 0.07</td><td>3.41 ± 0.16</td><td>3.20 ± 0.01</td><td>3.4 ± 0.10</td></tr><tr><td>S&amp;T</td><td>83.61 ± 0.53</td><td>82.72 ± 1.54</td><td>83.48 ± 3.32</td><td>82.50 ± 0.57</td><td>81.42 ± 1.40</td><td>4.26 ± 0.04</td><td>4.22 ± 0.06</td><td>4.30 ± 0.13</td><td>4.10 ± 0.02</td><td>4.31 ± 0.06</td></tr><tr><td>Mistral 7B</td><td>67.28 ± 1.27</td><td>68.27 ± 1.66</td><td>70.39 ± 4.04</td><td>66.40 ± 0.86</td><td>65.42 ± 1.67</td><td>3.19 ± 0.06</td><td>3.07 ± 0.09</td><td>3.44 ± 0.23</td><td>2.98 ± 0.05</td><td>3.08 ± 0.09</td></tr><tr><td>PivotRR</td><td>64.73 ± 1.13</td><td>60.92 ± 2.15</td><td>61.59 ± 5.66</td><td>62.99 ± 1.01</td><td>62.33 ± 2.04</td><td>3.54 ± 0.04</td><td>3.37 ± 0.06</td><td>3.56 ± 0.15</td><td>3.37 ± 0.03</td><td>3.60 ± 0.05</td></tr><tr><td>NeutralRR</td><td>66.94 ± 0.96</td><td>62.77 ± 2.20</td><td>62.83 ± 6.14</td><td>63.74 ± 1.03</td><td>63.51 ± 1.88</td><td>3.63 ± 0.03</td><td>3.47 ± 0.06</td><td>3.64 ± 0.14</td><td>3.43 ± 0.03</td><td>3.66 ± 0.05</td></tr><tr><td rowspan="4">rest</td><td>M2MS</td><td>60.21 ± 0.98</td><td>62.63 ± 1.33</td><td>60.04 ± 3.37</td><td>60.86 ± 1.94</td><td>60.87 ± 1.22</td><td>3.56 ± 0.03</td><td>3.45 ± 0.04</td><td>3.51 ± 0.10</td><td>3.22 ± 0.06</td><td>3.59 ± 0.04</td></tr><tr><td>S&amp;T</td><td>85.05 ± 0.26</td><td>85.82 ± 0.54</td><td>85.72 ± 1.17</td><td>80.39 ± 1.35</td><td>85.22 ± 0.53</td><td>4.72 ± 0.02</td><td>4.75 ± 0.03</td><td>4.72 ± 0.05</td><td>3.96 ± 0.06</td><td>4.78 ± 0.03</td></tr><tr><td>PivotRR</td><td>63.02 ± 0.98</td><td>64.63 ± 1.34</td><td>61.47 ± 3.15</td><td>62.33 ± 2.03</td><td>63.02 ± 1.20</td><td>3.73 ± 0.03</td><td>3.69 ± 0.04</td><td>3.68 ± 0.09</td><td>3.37 ± 0.06</td><td>3.77 ± 0.04</td></tr><tr><td>NeutralRR</td><td>64.90 ± 0.94</td><td>66.30 ± 1.27</td><td>62.98 ± 3.15</td><td>63.86 ± 1.98</td><td>64.42 ± 1.21</td><td>3.83 ± 0.03</td><td>3.78 ± 0.04</td><td>3.78 ± 0.08</td><td>3.45 ± 0.05</td><td>3.86 ± 0.04</td></tr></table>
+
+Table 5: Extended results of multi-target cross-lingual summarization in CrossSum for the metrics evaluating semantic coherence across target languages.

2024/Multi-Target Cross-Lingual Summarization_ a novel task and a language-neutral approach/images.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1ec70a61a359cc0c96d7a25f4212e415fab2fb8d4a40c8737bfcd79a099cf5d
+size 970185