EMNLP 2023 Papers — Page 2
Conference on Empirical Methods in Natural Language Processing · 1047 papers
ATHENA: Mathematical Reasoning with Thought Expansion
Jb. Kim, Yo-Sub Han (Yonsei University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the ATHENA model, which solves math word problems through an attention-driven thinking expansion mechanism.
Augmenting Zero-Shot Dense Retrievers with Plug-in Mixture-of-Memories
Suyu Ge (University of Illinois Urbana-Champaign), Paul Bennett (Spotify)
RetrievalTransformerContrastive LearningTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper proposes a retrieval-enhanced framework called MoMA based on multi-source memory to improve the generalization of zero-shot dense retrieval.
Automated Fact-Checking in Dialogue: Are Specialized Models Needed?
Eric Chamoun (University of Cambridge), Andreas Vlachos (University of Cambridge)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: For fact-checking in dialogues, this paper proposes enhancing retrieval and sentence splitting, document retrieval, sentence retrieval, and claim detection techniques, enabling standard single-claim fact-checking models to simultaneously process dialogues and traditional claims.
Automatic Debate Evaluation with Argumentation Semantics and Natural Language Argument Graph Networks
Ramon Ruiz-Dolz (University of Dundee), Ana Garcia (Universitat Politècnica de València)
ClassificationGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose a hybrid method combining argumentation theory with Transformer + graph networks to automatically predict the winning stance in complete professional debates.
Automatic Prompt Optimization with “Gradient Descent” and Beam Search
Reid Pryzant (Microsoft), Michael Zeng (Microsoft)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a non-parametric automatic prompt optimization algorithm called ProTeGi, which iteratively improves prompts on LLM APIs by combining natural language 'gradients' with beam search;
Automatic Transcription of Handwritten Old Occitan Language
Esteban Garces Arias (LMU), Matthias Aßenmacher (LMU)
RecognitionData SynthesisTransformerLarge Language ModelVision Language ModelImageText
🎯 What it does: Built a Transformer-based Handwritten Text Recognition (HTR) model using a Swin visual encoder and a BERT language decoder, combined with large-scale synthetic data and image augmentation techniques, specifically for recognizing low-resource Old Occitan text;
AutoTrial: Prompting Language Models for Clinical Trial Design
Zifeng Wang (University of Illinois Urbana Champaign), Jimeng Sun (University of Illinois Urbana Champaign)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Using a large language model (AutoTrial) to assist in generating inclusion/exclusion criteria text for clinical trials.
Axiomatic Preference Modeling for Longform Question Answering
Corby Rosset (Microsoft Research), Paul Bennett (Microsoft Research)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This study proposes a preference modeling framework based on axiomatic principles, constructs diverse contrast training samples, and trains a preference model with approximately 220 million parameters that can uniformly score long answers generated by humans and LLMs.
Back Transcription as a Method for Evaluating Robustness of Natural Language Understanding Models to Speech Recognition Errors
Marek Kubis (Adam Mickiewicz University), Tomasz Ziętkiewicz (Samsung Research Poland)
ClassificationRecognitionData SynthesisTransformerTextAudio
🎯 What it does: This paper proposes an evaluation framework based on back-translation (Text-to-Speech + ASR) and fine-grained error classification to quantify the robustness of natural language understanding (NLU) models against speech recognition errors.
Background Summarization of Event Timelines
Adithya Pratapa (Carnegie Mellon University), Markus Dreyer (Amazon)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed and implemented the news background summarization task, generating concise background summaries for each update on a timeline to help readers quickly grasp the historical context.
Baize: An Open-Source Chat Model with Parameter-Efficient Tuning on Self-Chat Data
Canwen Xu (University of California San Diego), Julian McAuley (Microsoft Research Asia)
Computational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a pipeline based on ChatGPT's self-dialogue to automatically generate high-quality multi-turn chat data, followed by LoRA parameter-efficient fine-tuning of LLaMA, and further improved model performance by introducing Self-Distillation with Feedback.
Balance Act: Mitigating Hubness in Cross-Modal Retrieval with Query and Gallery Banks
Yimu Wang (University of Waterloo), Bo Xue (City University of Hong Kong)
RetrievalImageVideoTextMultimodalityAudio
🎯 What it does: Propose a post-processing method called Dual Bank Normalization (DBNORM) to address the hubness problem in cross-modal retrieval.
BanglaAbuseMeme: A Dataset for Bengali Abusive Meme Classification
Mithun Das (IIT Kharagpur), Animesh Mukherjee (IIT Kharagpur)
ClassificationConvolutional Neural NetworkTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Built and released an annotated dataset containing 4,043 Bengali malicious memes, and evaluated the performance of text, image, and multimodal models for malicious detection based on this dataset.
BasahaCorpus: An Expanded Linguistic Resource for Readability Assessment in Central Philippine Languages
Joseph Marvin Imperial (University of Bath), Ekaterina Kochmar (MBZUAI)
ClassificationText
🎯 What it does: Collected and released BASAHACORPUS, a corpus containing children's short stories in four Central Philippine languages (Hiligaynon, Minasbate, Karaya, and Rinconada) from Let’s Read Asia, and built a reading difficulty assessment model based on this corpus.
Be Selfish, But Wisely: Investigating the Impact of Agent Personality in Mixed-Motive Human-Agent Interactions
Kushal Chawla (University of Southern California), Jonathan Gratch (University of Southern California)
Recurrent Neural NetworkReinforcement LearningAgentic AIText
🎯 What it does: Investigate the role of agent personality in mixed-motive negotiations by improving self-adversarial reinforcement learning training and adversarial roles, training and evaluating negotiation agents with multiple personalities.
Beat LLMs at Their Own Game: Zero-Shot LLM-Generated Text Detection via Querying ChatGPT
Biru Zhu (Tsinghua University), Ming Gu (Tsinghua University)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a zero-shot black-box method that discriminates LLM-generated text by having ChatGPT revise the text to be examined and comparing the similarity between the original and revised versions.
Benchmarking and Improving Text-to-SQL Generation under Ambiguity
Adithya Bhaskar (Princeton University), Sunita Sarawagi (IIT Bombay)
GenerationTransformerLarge Language ModelTextTabularBenchmark
🎯 What it does: This paper proposes a new evaluation benchmark, AmbiQT, specifically designed for ambiguous natural language queries in real databases, and develops a new decoding algorithm called LogicalBeam to improve the coverage of text-to-SQL generation models on ambiguous queries.
BERTie Bott’s Every Flavor Labels: A Tasty Introduction to Semantic Role Labeling for Galician
Micaella Bruton (Uppsala University), Meriem Beloucif (Uppsala University)
Data-Centric LearningTransformerSupervised Fine-TuningText
🎯 What it does: Constructed the first Galician semantic role labeling (SRL) dataset and proposed the Verbal Indexing preprocessing method;
Best of Both Worlds: Towards Improving Temporal Knowledge Base Question Answering via Targeted Fact Extraction
Nithish Kannen (Amazon Alexa AI, UK), L Subramaniam
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose a temporal question answering system that integrates knowledge bases (KB) with textual resources, utilizing language models for targeted fact extraction to address failures caused by missing information in KBs;
Better Quality Pre-training Data and T5 Models for African Languages
Akintunde Oladipo (University Of Waterloo), Jimmy Lin (University Of Waterloo)
Representation LearningData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper constructs a new high-quality pretraining corpus named WURA, covering 16 African languages and 4 high-resource languages, by auditing and improving the quality of existing multilingual pretraining corpora (especially mC4). Based on this, the T5 model AfriTeVa V2 was trained.
Beware of Model Collapse! Fast and Stable Test-time Adaptation for Robust Question Answering
Yi Su (Soochow University), Min Zhang (Soochow University)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a test-time adaptation method called Anti-CF, aiming to enhance model robustness under distribution drift by preventing model collapse and accelerating inference in QA tasks.
Beyond Detection: A Defend-and-Summarize Strategy for Robust and Interpretable Rumor Analysis on Social Media
Yi-Ting Chang (National Yang Ming Chiao Tung University), Hong-Han Shuai (National Yang Ming Chiao Tung University)
Anomaly DetectionExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelAuto EncoderText
🎯 What it does: Proposes the Defend-And-Summarize (DAS) framework and the BiTGN model, simultaneously enhancing robustness and explainability in rumor detection.
Beyond Factuality: A Comprehensive Evaluation of Large Language Models as Knowledge Generators
Liang Chen (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes a comprehensive evaluation framework named CONNER for automatically and reference-free assessing the intrinsic quality of knowledge generated by large language models (LLMs) and their extrinsic impact on downstream tasks;
Beyond Shared Vocabulary: Increasing Representational Word Similarities across Languages for Multilingual Machine Translation
Di Wu (University of Amsterdam), Christof Monz (University of Amsterdam)
Computational EfficiencyRepresentation LearningGraph Neural NetworkTransformerText
🎯 What it does: Reparameterize word vectors across languages using an equivalence graph based on word alignment, significantly enhancing cross-lingual word similarity on a shared vocabulary basis and achieving higher translation quality in multilingual neural machine translation.
Bias Neutralization in Non-Parallel Texts: A Cyclic Approach with Auxiliary Guidance
Karthic Madanagopal (Texas A&M University), James Caverlee (Texas A&M University)
GenerationDomain AdaptationTransformerGenerative Adversarial NetworkText
🎯 What it does: Propose a cyclic consistency adversarial network (FairBalance) that leverages non-parallel text to neutralize subjective bias while preserving content and improving fluency.
BiasX: “Thinking Slow” in Toxic Content Moderation with Explanations of Implied Social Biases
Yiming Zhang (Carnegie Mellon University), Maarten Sap (Carnegie Mellon University)
ClassificationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposed the BIASX framework, which leverages both human-generated and model-generated free-text explanations to assist content moderators in reflecting on and judging latent social biases and toxicity.
BioFEG: Generate Latent Features for Biomedical Entity Linking
Xuhui Sui (Nankai University), Wensheng Zhang (Nankai University)
RetrievalTransformerLarge Language ModelGenerative Adversarial NetworkBiomedical DataBenchmark
🎯 What it does: Propose the BioFEG framework, which enhances biomedical entity linking by generating potential features for unseen entities
Biomedical Named Entity Recognition via Dictionary-based Synonym Generalization
Zihao Fu (University of Cambridge), Nigel Collier (University of Cambridge)
RecognitionTransformerBiomedical Data
🎯 What it does: This paper proposes SynGen, a dictionary-based biomedical named entity recognition framework, which trains the model using synonyms present in the dictionary and achieves generalized recognition of out-of-dictionary synonyms through positive and negative sample learning.
BioPlanner: Automatic Evaluation of LLMs on Protocol Planning in Biology
Odhran O’Donoghue (Align to Innovate), Samuel Rodriques (Francis Crick Institute)
TransformerLarge Language ModelTextBiomedical DataChain-of-Thought
🎯 What it does: This paper proposes an automatic evaluation framework based on executable pseudocode to measure the ability of large language models in planning biological experiment protocols, and constructs and verifies the BIOPROT dataset on this basis.
BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations
Qizhi Pei (Renmin University of China), Rui Yan (Renmin University of China)
Drug DiscoveryTransformerLarge Language ModelPrompt EngineeringMultimodalityBiomedical Data
🎯 What it does: Propose BioT5, a cross-modal pre-training framework that leverages chemical knowledge and natural language associations for joint representation and generation of molecules, proteins, and text.
BLESS: Benchmarking Large Language Models on Sentence Simplification
Tannon Kew (University of Zurich), Matthew Shardlow (Manchester Metropolitan University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Construct the BLESS benchmark to evaluate the performance of 44 large language models on sentence simplification tasks using few-shot context learning, and conduct a systematic analysis of automatic evaluation metrics, edit operations, and manual quality assessment.
Boosting Summarization with Normalizing Flows and Aggressive Training
Yu Yang (University of Minnesota), Xiaotong Shen (University of Minnesota)
GenerationKnowledge DistillationTransformerFlow-based ModelText
🎯 What it does: Propose FlowSUM, a Transformer-based abstract framework based on variational encoding-decoding (VED), which enriches the latent space using normalizing flows and mitigates posterior collapse through controlled alternating aggressive training (CAAT) and an improved gating mechanism.
Bootstrapping Small & High Performance Language Models with Unmasking-Removal Training Policy
Yahan Yang (University of Pennsylvania), Dan Roth (University of Pennsylvania)
TransformerSupervised Fine-TuningText
🎯 What it does: This study explores the impact of the unmasking-removal training strategy, vocabulary size, and pretraining corpus on downstream tasks (semantic role labeling, QASRL, QAMR) using the small-scale language model BabyBERTa, and further improves model performance through continued pretraining.
BRAINTEASER: Lateral Thinking Puzzles for Large Language Models
Yifan Jiang (University of Southern California), Zhivar Sourati (University of Southern California)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed and released the BRAINTEASER benchmark, which converts lateral thinking puzzles into sentence-level and vocabulary-level multiple-choice questions, and tests the reasoning consistency of models through semantic reconstruction and context reconstruction.
Bridging Background Knowledge Gaps in Translation with Automatic Explicitation
HyoJung Han (University of Maryland), Marine Carpuat (University of Maryland)
GenerationData-Centric LearningLarge Language ModelText
🎯 What it does: Construct the WIKIEXPL dataset and propose a system for automatically detecting and generating explicitation in translation, helping translators fill the background knowledge gap of the target audience.
Bridging Continuous and Discrete Spaces: Interpretable Sentence Representation Learning via Compositional Operations
James Y. Huang (University of Southern California), Dong Yu (Tencent AI Lab)
Explainability and InterpretabilityRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Propose the INTERSENT framework to achieve interpretability and manipulability of sentence embeddings, supporting operations such as sentence fusion, difference, compression, and reconstruction.
Bridging Information-Theoretic and Geometric Compression in Language Models
Emily Cheng (Universitat Pompeu Fabra), Marco Baroni (ICREA)
CompressionTransformerLarge Language ModelText
🎯 What it does: Investigate the geometric and information-theoretic compression of language data by pre-trained language models and demonstrate a strong correlation between the two;
Bridging the Digital Divide: Performance Variation across Socio-Economic Factors in Vision-Language Models
Joan Nwatu (University of Michigan), Rada Mihalcea (University of Michigan)
TransformerVision Language ModelImageMultimodality
🎯 What it does: Evaluate the performance differences of CLIP on images from families at different economic levels, and analyze the performance disparities caused by visual diversity.
Bridging the Gap between Synthetic and Authentic Images for Multimodal Machine Translation
Wenyu Guo (Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Beijing Language and Culture University)
Data SynthesisDomain AdaptationTransformerVision Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: This paper proposes a method in multimodal machine translation that bridges the distribution gap between synthetic and real images during training by using both synthetic and real images along with a consistency loss, enabling inference without real images.
Building Multi-domain Dialog State Trackers from Single-domain Dialogs
Qi Zhu (Tsinghua University), Minlie Huang (Tsinghua University)
Data SynthesisTransformerLarge Language ModelText
🎯 What it does: This paper proposes a divide-and-conquer paradigm for multi-domain dialogue state tracking and constructs a framework to synthesize multi-domain dialogues from single-domain dialogues.
Building Persona Consistent Dialogue Agents with Offline Reinforcement Learning
Ryan Shea (Columbia University), Zhou Yu (Columbia University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an offline reinforcement learning framework that utilizes human-annotated rewards to enhance the personality consistency of open-domain dialogue systems;
Byte Pair Encoding for Symbolic Music
Nathan Fradet (Sorbonne University), Jean-Pierre Briot (Sorbonne University)
ClassificationGenerationTransformerLarge Language ModelSequential
🎯 What it does: In symbolic music generation and classification tasks, this paper applies Byte Pair Encoding (BPE) to existing tokenization methods such as REMI and TSD, significantly shortening sequence lengths, expanding the vocabulary, and conducting experiments using Transformer models.
ByteSized32: A Corpus and Challenge Task for Generating Task-Specific World Models Expressed as Text Games
Ruoyao Wang (University of Arizona), Peter Jansen (University of Arizona)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelWorld ModelTextBenchmark
🎯 What it does: Proposed the BYTESIZED32 dataset and investigated the ability of LLMs to generate executable text games to express task-specific world models.
C-STS: Conditional Semantic Textual Similarity
Ameet Deshpande (Princeton University), Karthik Narasimhan (Princeton University)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageTextMultimodality
🎯 What it does: Proposed the Conditional Semantic Text Similarity (C-STC) task, which requires evaluating similarity between given sentence pairs under natural language conditions;
Cabbage Sweeter than Cake? Analysing the Potential of Large Language Models for Learning Conceptual Spaces
Usashi Chatterjee (Cardiff University), Steven Schockaert (Cardiff University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: The study utilizes large language models to learn concept spaces, focusing on taste and physical attributes, to investigate whether they can capture perceptual dimensions.
Calc-X and Calcformers: Empowering Arithmetical Chain-of-Thought through Interaction with Symbolic Systems
Marek Kadlčík (Masaryk University), Vlastimil Martinek (Masaryk University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Constructed a unified arithmetic chain-of-thought dataset called Calc-X, and trained a model named Calcformers capable of invoking external calculators via HTML-style gadget tags.
Can Language Models Laugh at YouTube Short-form Videos?
Dayoon Ko (Seoul National University), Gunhee Kim (Seoul National University)
Data SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Constructed the ExFunTube dataset, collecting and annotating 10,136 user-generated 30-second short videos, with annotated punchline timestamps and textual explanations; and designed a zero-shot video-to-text prompting method, converting visual, audio, and sound information into fine-grained text, which is then input into large language models (LLMs) for humorous explanations.
Can language models learn analogical reasoning? Investigating training objectives and comparisons to human performance
Molly Petersen, Lonneke van der Plas (University of Malta)
ClassificationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Investigating whether language models can learn analogical reasoning through specialized training objectives and comparing their performance with humans
Can Language Models Understand Physical Concepts?
Lei Li (University of Hong Kong), Qi Liu (University of Hong Kong)
Knowledge DistillationRepresentation LearningLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmarkPhysics Related
🎯 What it does: This paper proposes the VEC benchmark to evaluate language models' understanding of visual and embodied physical concepts, systematically comparing the performance of large-scale text models and vision-enhanced models; meanwhile, it introduces a method to distill embodied knowledge from VLMs into text models.
Can Large Language Models Capture Dissenting Human Voices?
Noah Lee (KAIST AI), James Thorne (KAIST AI)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Evaluate the reasoning ability of instruction-based large language models in natural language inference tasks and their consistency with human opinion distributions, conducting experiments using two distribution estimation methods (MCE and LPE).
Can LLMs Facilitate Interpretation of Pre-trained Language Models?
Basel Mousi (Hamad Bin Khalifa University), Fahim Dalvi (Hamad Bin Khalifa University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper uses ChatGPT to perform hierarchical clustering and automatically annotate the latent concepts of pre-trained language models, achieving large-scale, fine-grained model interpretation.
Can LMs Generalize to Future Data? An Empirical Analysis on Text Summarization
Chi Seng Cheang (University of Macau), Lidia S. Chao (University of Macau)
TransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Proposed the TEMPOSUM benchmark to evaluate the abstract summarization capability of pre-trained language models on news texts from future time periods, and refined hallucination types through human evaluation.
Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions?
Yang Chen (Georgia Institute of Technology), Ming-Wei Chang (Georgia Institute of Technology)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a new visual information retrieval question answering benchmark called INFOSEEK, and conducts a systematic evaluation of the performance of pre-trained vision-language models on this task;
Can We Edit Factual Knowledge by In-Context Learning?
Ce Zheng (Peking University), Baobao Chang (Peking University)
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposes a knowledge editing method IKE based on context learning, which modifies the factual knowledge of large models through demonstrations without updating model parameters.
Can We Edit Multimodal Large Language Models?
Siyuan Cheng (Zhejiang University), Ningyu Zhang (Zhejiang University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes research on editing multimodal large language models (MLLM), constructs the MMEdit benchmark dataset, and designs three evaluation metrics: reliability, locality, and generalization;
Can You Follow Me? Testing Situational Understanding for ChatGPT
Chenghao Yang (University of Chicago), Allyson Ettinger (Allen Institute for AI)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs a synthetic experimental environment based on a box/key game to test the tracking and reasoning capabilities of conversational large language models (e.g., ChatGPT) in situational understanding (SU) over time.
CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion
Xingwei He (University of Hong Kong), Nan Duan (Microsoft Research Asia)
RetrievalTransformerContrastive LearningTextBenchmark
🎯 What it does: Proposes a document expansion strategy (CAPSTONE) using curriculum learning to improve query-aware document representations in dual-cross-encoders for dense retrieval;
CaseEncoder: A Knowledge-enhanced Pre-trained Model for Legal Case Encoding
Yixiao Ma (Tsinghua University), Yiqun Liu (Tsinghua University)
RetrievalRepresentation LearningTransformerContrastive LearningText
🎯 What it does: Propose a pre-trained encoder named CaseEncoder, which enhances data sampling and pre-training through fine-grained legal knowledge to improve zero-shot legal case retrieval performance.
Causal Document-Grounded Dialogue Pre-training
Yingxiu Zhao (Hong Kong University of Science and Technology), Nevin Zhang (Hong Kong University of Science and Technology)
GenerationRetrievalData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a causal pre-training framework called CausalDD for document-driven dialogue (DocGD). First, construct a causally complete pre-training corpus (WikiDialog and Reddit), then introduce causal perturbation losses (NDE and TIE) during pre-training to explicitly capture causal relationships between documents, evidence, dialogue context, and responses.
Causal Reasoning through Two Cognition Layers for Improving Generalization in Visual Question Answering
Trang Nguyen (Tokyo Institute of Technology), Naoaki Okazaki (Tokyo Institute of Technology)
Explainability and InterpretabilityMixture of ExpertsVision Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose a visual question answering model named CopVQA based on a two-layer cognitive path, enhancing multimodal prediction through causal reasoning and improving out-of-distribution (OOD) generalization.
Centering the Margins: Outlier-Based Identification of Harmed Populations in Toxicity Detection
Vyoma Raman (University of California Berkeley), Dan Klein (University of California Berkeley)
ClassificationAnomaly DetectionExplainability and InterpretabilityTransformerTextTabularBenchmark
🎯 What it does: Using anomaly detection methods to identify potential marginalized groups in toxicity detection and evaluating model errors for these groups
CESAR: Automatic Induction of Compositional Instructions for Multi-turn Dialogs
Taha Aksu (National University of Singapore), Mahdi Namazifar (Amazon Alexa AI)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Proposed the CESAR framework, which can automatically generate composite instructions for multi-task dialogues, and extended the InstructDial++ benchmark based on this, covering 63 datasets, 86 atomic tasks, and 68 composite tasks.
Chain-of-Questions Training with Latent Answers for Robust Multistep Question Answering
Wang Zhu (University of Southern California), Robin Jia (University of Southern California)
TransformerTextChain-of-Thought
🎯 What it does: Designed and trained a Chain-of-Questions framework that solves multi-step reasoning QA tasks by generating and answering sub-questions, utilizing QDMR (Question Decomposition Semantic Representation) to provide supervision for sub-questions, with sub-answers treated as latent variables learned dynamically through Hard-EM and MAPO.
Chain-of-Thought Tuning: Masked Language Models can also Think Step By Step in Natural Language Understanding
Caoyun Fan (Shanghai Jiao Tong University), Yaohui Jin (Shanghai Jiao Tong University)
ClassificationRecognitionTransformerPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: Proposes Chain-of-Thought Tuning (CoTT), a two-step reasoning framework based on prompt tuning, enabling small-scale masked language models (MLM) to perform step-by-step reasoning in natural language understanding tasks.
Challenges in Context-Aware Neural Machine Translation
Linghao Jin (University of Southern California), Xuezhe Ma (University of Southern California)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigated the core bottleneck of context-aware neural machine translation (NMT), systematically evaluated multiple languages, model architectures, and context concatenation methods, proposed a more realistic paragraph-to-paragraph (PARA2PARA) translation setup, and subsequently released a Chinese-English literary translation dataset consisting of 10,545 paragraphs.
Character-LLM: A Trainable Agent for Role-Playing
Yunfan Shao (Fudan University), Xipeng Qiu (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Train a tunable agent Character-LLM through supervised fine-tuning using scenarios generated from historical figure profiles, enabling the model to role-play as specific historical figures.
Characterizing and Verifying Scientific Claims: Qualitative Causal Structure is All You Need
Jinxuan Wu (Beihang University), Zhunchen Luo (PLA Academy of Military Science)
Explainability and InterpretabilityGraph Neural NetworkTransformerTextBiomedical Data
🎯 What it does: Propose a scientific claim verification framework VerQCS based on qualitative causal structures, abstracting causal relationships between claims and evidence into heterogeneous graphs, and using attention-based graph neural networks for causal reasoning
Characterizing Mechanisms for Factual Recall in Language Models
Qinan Yu (Brown University), Ellie Pavlick (Brown University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper investigates the fact recall mechanisms of language models when pre-training memory conflicts with contextual prompts, using the world capital question-answering task to explore whether models tend to retain memory or overwrite new information; it also analyzes the role of attention heads and proposes a technique to dynamically adjust model behavior during runtime.
ChatEdit: Towards Multi-turn Interactive Facial Image Editing via Dialogue
Xing Cui, Zhaofeng He
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: Propose a multi-round interactive facial image editing task, construct the CHATEDIT dataset, and provide a baseline framework.
ChatGPT to Replace Crowdsourcing of Paraphrases for Intent Classification: Higher Diversity and Comparable Model Robustness
Jan Cegin (Brno University of Technology), Peter Brusilovsky (University of Pittsburgh)
ClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Compare the effectiveness, diversity, and model robustness of paraphrases generated by ChatGPT and traditional crowdsourcing in intent classification data, using the same experimental process to verify whether ChatGPT can replace human workers.
CHEF in the Language Kitchen: A Generative Data Augmentation Leveraging Korean Morpheme Ingredients
Jaehyung Seo (Korea University), Heuiseok Lim (Korea University)
ClassificationData SynthesisTransformerContrastive LearningTextBenchmark
🎯 What it does: Propose CHEF, a generative data augmentation method based on Korean morphemes, which synthesizes new sentences using a morpheme mixer and a label discriminator while maintaining label consistency.
Chinese Lexical Substitution: Dataset and Method
Jipeng Qiang (Yangzhou University), Xiaoye Ouyang (China Academy of Electronic and Information Technology)
Data SynthesisRepresentation LearningTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: Constructed the Chinese Lexical Substitution (CHNLS) dataset using a human-machine collaborative annotation approach.
CiteBench: A Benchmark for Scientific Citation Text Generation
Martin Funkquist (Linköping University), Iryna Gurevych (Technical University of Darmstadt)
GenerationTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed the CITEBENCH benchmark, unifying four distinct reference text generation tasks and providing baselines and evaluation tools.
CLAD-ST: Contrastive Learning with Adversarial Data for Robust Speech Translation
Sathish Indurthi, Marco Turchi (Zoom Video Communications)
TransformerContrastive LearningTextAudio
🎯 What it does: By incorporating adversarial ASR outputs and human text at the sentence level through contrastive learning in Transformer NMT models, the model's robustness to ASR noise is enhanced in a cascading architecture for speech translation.
CLAIR: Evaluating Image Captions with Large Language Models
David M. Chan (University of California, Berkeley), John Canny (University of California, Berkeley)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmark
🎯 What it does: Utilizing large language models (LLMs) for zero-shot evaluation of image captions, generating numerical scores and providing explanatory justifications.
CleanCoNLL: A Nearly Noise-Free Named Entity Recognition Dataset
Susanna Rücker (Humboldt University of Berlin), Alan Akbik (Humboldt University of Berlin)
RecognitionTextBenchmark
🎯 What it does: By conducting a comprehensive re-annotation of CoNLL-03 with the addition of an entity linking layer and consistency checks, the CLEANCONLL dataset was generated, nearly free of noise.
clembench: Using Game Play to Evaluate Chat-Optimized Language Models as Conversational Agents
Kranti Chalamalasetti, David Schlangen (University Of Potsdam)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a game-based evaluation framework called clembench, designed to systematically test the performance of chat-optimized large language models (cLLMs) in constrained, rule-based interactive language games; by implementing five different dialogue games (such as Taboo, Wordle, drawing instructions, reference games, and private/shared information games), and having various open-source and closed-source models play these games in self-dialogue, their instruction-following and task-completion capabilities are measured.
CLEME: Debiasing Multi-reference Evaluation for Grammatical Error Correction
Jingheng Ye (Tsinghua Shenzhen International Graduate School, Tsinghua University), Ying Shen (Tsinghua Shenzhen International Graduate School, Tsinghua University)
Text
🎯 What it does: Propose and implement CLEME, a syntactic error correction evaluation metric that eliminates multi-reference evaluation bias by utilizing consistent block boundaries and the error correction independence assumption.
CLEVR-Implicit: A Diagnostic Dataset for Implicit Reasoning in Referring Expression Comprehension
Jingwei Zhang (South China University of Technology), Yi Cai (South China University of Technology)
RecognitionTransformerPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: For the task of pointing expression understanding, this paper constructs the CLEVR-Implicit dataset and proposes the TIE method to convert implicit expressions into explicit expressions.
ClimateBERT-NetZero: Detecting and Assessing Net Zero and Reduction Targets
Tobias Schimanski (University of Zurich), Markus Leippold (University of Oxford)
ClassificationTransformerSupervised Fine-TuningTextFinance Related
🎯 What it does: Developed the ClimateBERT-NetZero model to automatically detect and evaluate net-zero and emission reduction targets of enterprises, countries, or regions, demonstrating its application in two real-world use cases: first, combining a QA model to extract target years, base years, and emission reduction magnitudes; second, conducting large-scale analysis of U.S. listed companies' earnings call texts from 2003–2022.
Clinical Contradiction Detection
Dave Makhervaks (Technion Israel Institute of Technology), Kira Radinsky (Technion Israel Institute of Technology)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records
🎯 What it does: Using SNOMED clinical ontology for remote supervision to construct naturally occurring medical sentence pairs and detect contradictions within the sentence pairs.
Clustering Pseudo Language Family in Multilingual Translation Models with Fisher Information Matrix
Xinyu Ma (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerText
🎯 What it does: Propose a pseudo-language family clustering method based on the Fisher Information Matrix (FIM) to select auxiliary language pairs in multilingual neural machine translation;
ClusterLLM: Large Language Models as a Guide for Text Clustering
Yuwei Zhang (University of California, San Diego), Jingbo Shang (University of California, San Diego)
Representation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Propose the CLUSTERLLM framework, which leverages API LLM for guided text clustering, combining triplet and pairwise tasks to optimize embeddings and determine clustering granularity.
CoAnnotating: Uncertainty-Guided Work Allocation between Human and Large Language Models for Data Annotation
Minzhi Li (National University of Singapore), Diyi Yang (Stanford University)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the CoAnnotating framework, which dynamically assigns text annotation tasks to humans or ChatGPT by leveraging uncertainty information from LLMs, aiming to reduce costs while maintaining quality.
Coarse-to-Fine Contrastive Learning in Image-Text-Graph Space for Improved Vision-Language Compositionality
Harman Singh (Meta AI), Yu Chen (Anytime.AI)
RetrievalRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImageTextMultimodalityGraph
🎯 What it does: This paper designs a coarse-to-fine contrastive learning framework based on scene graphs, which significantly enhances the compositional reasoning ability of vision-language models by decomposing text-parsed scene graphs into multi-level subgraphs and jointly training images with multi-text contrast using hard negative subgraphs generated through graph augmentation.
CoCo: Coherence-Enhanced Machine-Generated Text Detection Under Low Resource With Contrastive Learning
Xiaoming Liu (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
ClassificationGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed the COCO model, which detects differences between machine-generated text and human-generated text using a consistency-enhanced contrastive learning approach.
Code-Switching Metrics Using Intonation Units
Rebecca Pattichis (University of California Los Angeles), Rena Torres Cacoullos (Pennsylvania State University)
Text
🎯 What it does: This paper migrates traditional word-level code-switching (CS) measurement methods (multilingual measure M-Index and CS probability I-Index) to phonological units—intonation units (Intonation Units, IUs)—and categorizes word-level mixing into two types: word-level lone items and multi-word strings, investigating CS patterns at the IU level.
CodeBERTScore: Evaluating Code Generation with Pretrained Models of Code
Shuyan Zhou (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes CodeBERTScore, a code generation evaluation metric based on BERTScore, which considers both natural language context and code semantic consistency.
CodeFusion: A Pre-trained Diffusion Model for Code Generation
Mukul Singh (Microsoft), Gust Verbruggen (Microsoft)
Data SynthesisAI Code AssistantTransformerDiffusion modelText
🎯 What it does: Developed a conditional diffusion-based NL-to-code generation model called CODEFUSION, which progressively generates diverse and syntactically correct code using diffusion denoising and a Transformer decoder.
CodeT5+: Open Code Large Language Models for Code Understanding and Generation
Yue Wang (Salesforce AI Research), Steven Hoi
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposed CodeT5+, which can dynamically switch between encoder, decoder, or encoder-decoder modes, and improved code understanding and generation performance through mixed pre-training objectives.
CoF-CoT: Enhancing Large Language Models with Coarse-to-Fine Chain-of-Thought Prompting for Multi-domain NLU Tasks
Hoang Nguyen (University of Illinois at Chicago), Philip Yu (University of Illinois at Chicago)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose a coarse-to-fine hierarchical chained reasoning (CoF-CoT) framework that decomposes NLU tasks into a process from coarse to fine using multi-step reasoning sequences;
COFFEE: Counterfactual Fairness for Personalized Text Generation in Explainable Recommendation
Nan Wang (Netflix Inc.), Shaoliang Nie (Meta AI)
Recommendation SystemExplainability and InterpretabilityTransformerReinforcement LearningGenerative Adversarial NetworkText
🎯 What it does: Propose the COFFEE framework, achieving personalized text generation with individual counterfactual fairness by decoupling attribute representation and policy learning.
Cognate Transformer for Automated Phonological Reconstruction and Cognate Reflex Prediction
V.S.D.S.Mahesh Akavarapu (Indian Institute of Technology Kanpur), Arnab Bhattacharya (Indian Institute of Technology Kanpur)
GenerationTransformerText
🎯 What it does: Proposed and implemented the Cognate Transformer model for speech reconstruction, which can automatically generate phoneme sequences in two tasks: parent language reconstruction and cognitive reflection prediction.
Cognitive Dissonance: Why Do Language Model Outputs Disagree with Internal Representations of Truthfulness?
Kevin Liu (MIT), Jacob Andreas (MIT)
Explainability and InterpretabilityRepresentation LearningLarge Language ModelText
🎯 What it does: This paper investigates the inconsistency between output probabilities and internal hidden representations (probes) in language models when answering factual questions, and proposes three types of errors;
COHESENTIA: A Novel Benchmark of Incremental versus Holistic Assessment of Coherence in Generated Texts
Aviya Maimon (Bar Ilan University), Reut Tsarfaty (Bar Ilan University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
🎯 What it does: Constructed the COHESENTIA benchmark and proposed two annotation protocols (global and incremental) for evaluating the coherence of stories generated by GPT-3.
Collaborative Generative AI: Integrating GPT-k for Efficient Editing in Text-to-Image Generation
Wanrong Zhu (University Of California Santa Barbara), William Wang
GenerationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: Use large language models (GPT-2, GPT-3, GPT-3.5) to perform one-time edits on StableDiffusion's text prompts, evaluating their effectiveness in image generation.
CoLT5: Faster Long-Range Transformers with Conditional Computation
Joshua Ainslie (Google), Sumit Sanghai (Google)
Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsTextBenchmark
🎯 What it does: Propose COLT5, a long-text Transformer accelerated by conditional computation, capable of efficient inference and training on long sequences.
Combining Denoising Autoencoders with Contrastive Learning to fine-tune Transformer Models
Alejo Lopez-Avila, Víctor Suárez-Paniagua (Huawei Ireland Research Center)
Representation LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningText
🎯 What it does: Propose a three-stage fine-tuning method: first adapt data distribution using a denoising autoencoder, then cluster representations and correct class imbalance via supervised contrastive learning, and finally add a classification head for final training during the fine-tuning stage.
CombLM: Adapting Black-Box Language Models through Small Fine-Tuned Models
Aitor Ormazabal (HiTZ Center, University Basque Country), Eneko Agirre (HiTZ Center, University Basque Country)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a method for domain adaptation and downstream task transfer on black-box large language models. The core idea is to first perform domain fine-tuning using a white-box model, and then fuse the outputs of the small and large models at the probabilistic level through a small network to obtain an adapted model.
Comparing Biases and the Impact of Multilingual Training across Multiple Languages
Sharon Levy (University of California Santa Barbara), Dan Roth (AWS AI Labs)
ClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: Evaluate bias in sentiment analysis tasks across multiple languages (English, Italian, Chinese, Hebrew, Spanish) using four attributes: ethnicity, religion, nationality, and gender.