Annual Meeting of the Association for Computational Linguistics · 356 papers
Superfiltering: Weak-to-Strong Data Filtering for Fast Instruction-Tuning
Ming Li (University of Maryland), Tianyi Zhou (University of Maryland)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the Superfiltering method, which uses weak models (e.g., GPT-2) to filter instruction tuning data, significantly accelerating and enhancing the training effectiveness of large LLMs.
Synergetic Event Understanding: A Collaborative Approach to Cross-Document Event Coreference Resolution with Large Language Models
Qingkai Min (Zhejiang University), Yue Zhang (Shanghai AI Laboratory)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a collaborative method that first uses a large language model to perform multi-step summarization of event abstracts, then inputs the summary information along with the original text into a small language model for fine-tuning, addressing cross-document event coreference resolution.
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: The InteR framework is proposed to achieve more accurate zero-shot retrieval by iteratively refining information through combining retrieval models (RM) with large language models (LLM).
Synthesizing Text-to-SQL Data from Weak and Strong LLMs
Jiaxi Yang (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences), Chang Zhou (Alibaba Group)
CodeData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextTabularBenchmark
🎯 What it does: Propose a text-to-SQL (text-to-SQL) model called SENSE based on open-source large language models (LLMs), leveraging 'strong data' generated by large models to enhance multi-domain generalization, and combining 'weak data' generated by small models with executor feedback for preference learning, enabling the model to learn from errors.
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTabularChain-of-Thought
🎯 What it does: Proposes the TAPERA framework for long-form table question answering (LFTQA), achieving modular reasoning and generation through QA-based content planning, table reasoning with executable Python programs, and answer generation.
🎯 What it does: Propose a continuous dialogue state tracking framework named TaSL based on task skill localization and integration, achieving continuous learning without memory replay;
Tell Me More! Towards Implicit User Intention Understanding of Language Model Driven Agents
Cheng Qian (Tsinghua University), Maosong Sun (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Proposed the IN3 benchmark to evaluate language model-driven agents' understanding of users' implicit intentions, and trained the Mistral-Interact model as a pre-interaction module to enhance agents' task understanding and execution efficiency.
Text Embedding Inversion Security for Multilingual Language Models
Yiyi Chen (Aalborg University), Johannes Bjerva (Aalborg University)
CodeSafty and PrivacyRepresentation LearningAdversarial AttackTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: Investigate multilingual embedding inversion attacks, explore the feasibility of black-box multilingual and cross-lingual text recovery, and propose a mask-based defense method.
Text-like Encoding of Collaborative Information in Large Language Models for Recommendation
Yang Zhang (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeRecommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a recommendation method called BinLLM, which converts collaborative information into a binary sequence (compressible as dot-decimal notation) and directly inputs it as text into LLMs.
🎯 What it does: Proposed and implemented the Text-to-Song task, using a two-stage Melodist model to first generate vocals and then accompaniment, ultimately synthesizing a complete song.
The Echoes of Multilinguality: Tracing Cultural Value Shifts during Language Model Fine-tuning
Rochelle Choenni (University of Amsterdam), Ekaterina Shutova (University of Amsterdam)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper investigates how the language and data sources during the fine-tuning of multilingual language models (MLMs) influence the cultural values encoded in the models, and evaluates the alignment of value changes with human values through Cloze probing using questions from the World Values Survey (WVS).
🎯 What it does: This paper first constructs a multimodal dataset named MERSA, which includes natural and scripted speech, text transcripts, physiological signals, and self-reported emotion questionnaires. It also proposes a Transformer framework based on wav2vec2.0, BERT, and LSTM, which performs early fusion of speech and text to predict emotion dimensions (arousal, valence, dominance).
Through the MUD: A Multi-Defendant Charge Prediction Benchmark with Linked Crime Elements
Xiao Wei (Shanghai University), Erik Cambria (Nanyang Technological University)
CodeClassificationExplainability and InterpretabilityTransformerContrastive LearningTextBenchmark
🎯 What it does: Proposed a benchmark for multi-defendant criminal cases called MUD, and designed the EJudge model to achieve explainable sentencing prediction based on criminal elements and legal rules.
Time is Encoded in the Weights of Finetuned Language Models
Kai Nylund (Paul G. Allen School of Computer Science & Engineering, University of Washington), Noah Smith (Paul G. Allen School of Computer Science & Engineering, University of Washington)
🎯 What it does: Studied the impact of time on language model performance, introduced the concept of 'time vectors' derived by fine-tuning on a single time period and calculating the difference, and used these vectors for interpolation and analogy in the weight space to enhance the model's generalization ability for new time periods without additional training.
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequential
🎯 What it does: This paper proposes a time-axis-based sentence decomposition method, combining it with the context learning of large language models (LLMs), and achieves high-quality extraction of temporal facts from complex sentences using a fine-tuned small pre-trained language model (PLM).
To be Continuous, or to be Discrete, Those are Bits of Questions
Yiran Wang (National Institute of Information and Communications Technology), Masao Utiyama (National Institute of Information and Communications Technology)
CodeTransformerContrastive LearningText
🎯 What it does: In traditional deep learning models, inputs/outputs often use continuous vectors, while natural language is inherently discrete. This paper introduces binary representations (between continuous and discrete) into the model's output, proposing a structured contrastive hashing method, and directly represents syntactic trees or entity hierarchical structures with binary labels via bit-level CKY parsing.
To Generate or to Retrieve? On the Effectiveness of Artificial Contexts for Medical Open-Domain Question Answering
Giacomo Frisoni (University of Bologna), Zaiqiao Meng (University of Glasgow)
CodeTransformerLarge Language ModelPrompt EngineeringTextBiomedical DataBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose MEDGENIE, a fully generative open-domain medical question-answering framework that first generates synthetic context using a medical large language model and then reads and answers.
Toward In-Context Teaching: Adapting Examples to Students’ Misconceptions
Alexis Ross (MIT), Jacob Andreas (MIT)
CodeTransformerLarge Language ModelText
🎯 What it does: Proposed the ADAPT evaluation framework and the ATOM adaptive teaching method, studying how large language models (e.g., GPT-4) and probabilistic models can teach students to address misconceptions;
🎯 What it does: This paper proposes the i-NID task, addressing unknown intent discovery under long-tailed distributions by designing the ImbaNID framework and creating the ImbaNIDBench benchmark.
Alexander Spangher (University of Southern California), Jonathan May (University of California, Los Angeles)
CodeClassificationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: The study establishes a framework that links local government policy documents with news reports and predicts which policies will be covered.
🎯 What it does: Proposed and implemented a fine-tuning framework called TRANSLICO based on transliteration contrastive learning to eliminate representation barriers between different writing systems in multilingual pre-trained models.
Transparent and Scrutable Recommendations Using Natural Language User Profiles
Jerome Ramos (University College London), Aldo Lipani (University College London)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Leverage instruction-tuned LLMs to generate natural language user profiles, then fine-tune the LLM based on these profiles for rating prediction, thereby achieving a transparent and auditable recommendation system.
Tree-of-Traversals: A Zero-Shot Reasoning Algorithm for Augmenting Black-box Language Models with Knowledge Graphs
Elan Markowitz (Amazon AGI), Aram Galstyan (Amazon AGI)
CodeTransformerLarge Language ModelTextGraphBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a Tree-of-Traversals zero-shot reasoning algorithm, enabling black-box large language models (LLMs) to perform tree search reasoning through knowledge graph (KG) interfaces, thereby augmenting model knowledge and answering question-answering tasks.
Trial and Error: Exploration-Based Trajectory Optimization of LLM Agents
Yifan Song (Peking University), Bill Yuchen Lin (Allen Institute for AI)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIContrastive LearningTextSequentialBenchmark
🎯 What it does: This paper proposes an Exploratory Trajectory Optimization (ETO) method, which enhances the performance of open-ended large language model (LLM) agents by allowing LLMs to learn through trial-and-error in environments, collecting failure trajectories, and comparing them with expert trajectories.
TruthX: Alleviating Hallucinations by Editing Large Language Models in Truthful Space
Shaolei Zhang (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Yang Feng (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelAuto EncoderContrastive LearningText
🎯 What it does: Propose the TruthX method, which edits the internal representations of LLMs during inference to enhance the truthfulness of their answers.
Chufan Gao (University of Illinois Urbana-Champaign), Jimeng Sun (University of Illinois Urbana-Champaign)
CodeClassificationTransformerSupervised Fine-TuningTextBiomedical Data
🎯 What it does: Proposed a document-level relation extraction model TTM-RE that integrates a built-in learnable memory module (Token Turing Machine) with a noise-robust loss.
Two Issues with Chinese Spelling Correction and A Refinement Solution
Changxuan Sun (East China Normal University), Xuesong Lu (East China Normal University)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper addresses two major issues in the Chinese Spelling Correction (CSC) task: first, the SIGHAN13/14/15 datasets contain a large number of errors, leading to inaccurate model evaluation; second, existing models have limited room for improvement on the SIGHAN test set. To address these issues, the authors manually corrected three SIGHAN datasets and retrained and evaluated four representative CSC models (PLOME, REALISE, LEAD, SCOPE) using the corrected versions. Based on this, the authors proposed a post-processing scheme based on ChineseBERT, which significantly improves the precision, recall, and F1 scores of all models in detection and correction by masking model outputs, using BERT inference, and combining pinyin edit distance thresholds to determine whether to accept the inferred results.
🎯 What it does: Constructed an open-source multilingual supervised fine-tuning dataset called UltraLink, balancing language-specific knowledge and language-agnostic capabilities;
Uncovering the Full Potential of Visual Grounding Methods in VQA
Daniel Reich (University of Bremen), Tanja Schultz (University of Bremen)
CodeVision Language ModelImageTextMultimodality
🎯 What it does: The study identifies flaws in the training and testing practices of visual grounding (VG) methods in VQA evaluation and proposes a 'True Visual Grounding (TVG)' framework to address these issues.
UniCoder: Scaling Code Large Language Model via Universal Code
Tao Sun (Beihang University), Zhoujun Li (Beihang University)
CodeGenerationData SynthesisAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed a code generation framework called UNICODER, which uses UniCode as an intermediate representation, constructs a large-scale instruction dataset named UNICODER-INSTRUCT, and fine-tunes a general-purpose code LLM under a multi-task learning objective.
CodeAnomaly DetectionLarge Language ModelVision Language ModelMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a unified multi-modal hallucination detection framework called UNIHD and constructed the MHaluBench unified evaluation benchmark for fine-grained detection of hallucinations in image-text and text-image tasks.
Unlocking the Power of Large Language Models for Entity Alignment
Xuhui Jiang (CAS Key Laboratory of AI Safety Institute of Computing Technology), Yuanzhuo Wang (CAS Key Laboratory of AI Safety Institute of Computing Technology)
CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: This paper proposes the ChatEA framework, combining large language models (LLM) with traditional knowledge representation learning (KRL) methods to enhance entity alignment (EA) performance by leveraging the background knowledge and reasoning capabilities of LLM.
ValueBench: Towards Comprehensively Evaluating Value Orientations and Understanding of Large Language Models
Yuanyi Ren (Peking University), Guojie Song (Peking University)
CodeLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose ValueBench, a comprehensive benchmark integrating 44 psychological assessment tools and 453 value dimensions, designed to evaluate the value orientation and value understanding of large language models (LLMs);
VariErr NLI: Separating Annotation Error from Human Label Variation
Leon Weber-Genzel (LMU Munich), Barbara Plank (UCLouvain)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Construct the VARIERR dataset by designing a two-round annotation process (first providing labels and reasons, then conducting self-evaluation and peer evaluation on label-reason pairs) to distinguish between annotation errors and human label variations.
Virtual Compiler Is All You Need For Assembly Code Search
Zeyu Gao (Tsinghua University), Chao Zhang (Beijing University of Posts and Telecommunications)
CodeData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Train a large language model as a virtual compiler to generate assembly code and expand the assembly code retrieval dataset, thereby improving natural language retrieval performance in reverse engineering.
VISTA: Visualized Text Embedding For Universal Multi-Modal Retrieval
Junjie Zhou (Beijing University of Posts and Telecommunications), Yongping Xiong (Beijing Academy of Artificial Intelligence)
CodeRetrievalRepresentation LearningTransformerVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a unified embedding model VISTA, which is built upon a frozen strong text encoder and uses Vision Transformer as an image tokenizer, capable of handling text, images, and their combinations. The model achieves multimodal retrieval capability through a two-stage training process.
Visualization Recommendation with Prompt-based Reprogramming of Large Language Models
Xinhang Li (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeRecommendation SystemTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: This paper proposes an HTP framework based on hierarchical table prompts for LLM reprogramming, used for automatically recommending visualization charts.
WatME: Towards Lossless Watermarking Through Lexical Redundancy
Liang Chen (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelText
🎯 What it does: This paper proposes WatME, a watermarking method based on lexical redundancy, which can embed detectable watermarks into large language model texts without significantly compromising generation quality;
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose WaveCoder, which utilizes multi-task, scalable high-quality instruction data (CodeSeaXDataset) for instruction tuning of Code LLMs, and constructs an LLM-based generation-discrimination framework to achieve automatic generation and quality control of instruction data.
WebCiteS: Attributed Query-Focused Summarization on Chinese Web Search Results with Citations
Haolin Deng (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the WebCiteS dataset and proposed and experimented with the Chinese Web Search Result Query-Focused Summarization (AQFS) task.
WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models
Hongliang He (Zhejiang University), Dong Yu (Tencent AI Lab)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringVision-Language-Action ModelMultimodality
🎯 What it does: Designed and implemented WebVoyager, a large multimodal model (LMM)-based system capable of end-to-end completion of interactive tasks on real websites, using screenshots and webpage element text jointly to guide decision-making and execute actions such as clicking, typing, and scrolling.
What Does the Bot Say? Opportunities and Risks of Large Language Models in Social Media Bot Detection
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextMultimodalityGraphBenchmark
🎯 What it does: Explore opportunities and risks of applying large language models (LLM) to social media bot detection, propose a hybrid heterogeneous expert framework to enhance detection performance, and investigate LLM-driven text and structural information tampering strategies to evade detection.
What is the Best Way for ChatGPT to Translate Poetry?
Shanshan Wang (University of Macau), Lidia Chao (University of Macau)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigates the capability of ChatGPT in translating modern English-Chinese poetry and proposes a two-step translation approach (EAPMT) based on poetic interpretation assistance.
When Benchmarks are Targets: Revealing the Sensitivity of Large Language Model Leaderboards
Norah Alzahrani (National Center for AI), Haidar Khan (Saudi Data and AI Authority)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigate how small perturbations in multiple-choice question (MCQ) benchmarks affect large language model (LLM) leaderboards, systematically evaluating sensitivity to answer order, symbols, scoring methods, and prompts;
When Good and Reproducible Results are a Giant with Feet of Clay: The Importance of Software Quality in NLP
Sara Papi (Fondazione Bruno Kessler), Matteo Negri (Fondazione Bruno Kessler)
CodeTransformerTextAudio
🎯 What it does: Analyzed and fixed three categories of bugs in the widely used Conformer implementation, and through experiments demonstrated the potential misleading effects of these bugs on ASR and ST results; proposed the pangoliNN library for neural network unit testing, and provided a code quality checklist to enhance the verifiability of research software.
🎯 What it does: Research and improve the correlation between fairness evaluation metrics based on prompts, proposing the CAIRO method that maximizes the Pearson correlation coefficient between metrics through multi-model prompt enhancement and combination selection.
🎯 What it does: Proposed the Weight Recover Prune (WRP) method, which recovers a small number of critical weights on the basis of 2:4 structured sparsity to improve the accuracy of large language models (LLMs) while maintaining compression effects
Zero-Shot Cross-Lingual Reranking with Large Language Models for Low-Resource Languages
Mofetoluwa Adeyemi (University of Waterloo), Jimmy Lin (University of Waterloo)
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the zero-shot listwise re-ranking effectiveness of large language models (LLMs) in low-resource African languages (Hausa, Somali, Swahili, Yoruba), and systematically evaluated across three scenarios: cross-lingual, monolingual (original language), and self-translation.