EMNLP 2023 Papers — Page 6
Conference on Empirical Methods in Natural Language Processing · 1047 papers
Improving Image Captioning via Predicting Structured Concepts
Ting Wang (University of Science and Technology of China), Zhendong Mao (University of Washington)
GenerationRepresentation LearningGraph Neural NetworkTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose a Structured Concept Predictor (SCP) that simultaneously predicts semantic concepts and their structures in images, and feeds the structured concepts along with visual features into a Transformer decoder to achieve end-to-end image caption generation.
Improving Language Models’ Meaning Understanding and Consistency by Learning Conceptual Roles from Dictionary
Myeongjun Jang (University of Oxford), Thomas Lukasiewicz (Vienna University of Technology)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes an intermediate pre-training task, Concept Role Modeling (CRM), learned through dictionary definitions, to enhance the semantic understanding capability of pre-trained language models (PLMs). This significantly reduces model inconsistencies across multiple consistency types (semantic, negational, symmetric, transitive) and achieves efficient fine-tuning of large PLMs via parameter fusion techniques.
Improving Long Document Topic Segmentation Models With Enhanced Coherence Modeling
Hai Yu (Alibaba Group), Wen Wang (Alibaba Group)
SegmentationTransformerContrastive LearningText
🎯 What it does: This paper proposes two auxiliary tasks—Topic-Aware Sentence Structure Prediction (TSSP) and Contrastive Semantic Similarity Learning (CSSL)—and achieves performance improvements in long-text topic segmentation under the Longformer framework.
Improving Summarization with Human Edits
Zonghai Yao (University of Massachusetts), Sai Selvaraj (Abridge AI)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: This paper proposes a method to improve summary quality by leveraging human editor feedback, combining sequence alignment with forward/backward likelihood training, with a focus on clinical dialogue summaries in the healthcare domain.
Improving Transformer-based Program Repair Model through False Behavior Diagnosis
Youngkyoung Kim (Sungkyunkwan University), Eunseok Lee (Sungkyunkwan University)
AI Code AssistantTransformerLarge Language ModelText
🎯 What it does: By diagnosing the internal behavior of Transformer-based program repair models, we propose behavior vectors and a behavior discriminator (BeDisc), and introduce two handling approaches (early termination and masking skip) to eliminate erroneous behaviors and improve repair accuracy.
Improving Unsupervised Relation Extraction by Augmenting Diverse Sentence Pairs
Qing Wang (Iowa State University), Qi Li (Iowa State University)
Representation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the AugURE method, improving relation representation learning in unsupervised relation extraction through diverse positive sample augmentation and margin loss.
IMTLab: An Open-Source Platform for Building, Evaluating, and Diagnosing Interactive Machine Translation Systems
Xu Huang (Nanjing University), Shujian Huang (Nanjing University)
GenerationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed and implemented an open-source platform named IMTLAB for rapidly building, end-to-end evaluating, and diagnosing interactive machine translation (IMT) systems.
Incorporating Structured Representations into Pretrained Vision & Language Models Using Scene Graphs
Roei Herzig (Tel-Aviv University), Amir Globerson (Tel-Aviv University)
Representation LearningTransformerVision Language ModelContrastive LearningImageTextGraph
🎯 What it does: This paper proposes a method to enhance pre-trained vision-language models (VLM) using scene graphs (SG), constructing structured representations for both visual and textual modalities, and achieving prediction and learning of SG information through fine-grained positive and negative captions and visual layer 'Adaptive SG Tokens'.
Incorporating Worker Perspectives into MTurk Annotation Practices for NLP
Olivia Huang (University of California Berkeley), Dan Klein (University of California Berkeley)
Data-Centric LearningTextReview/Survey Paper
🎯 What it does: A critical review of common practices in natural language processing (NLP) data collection using Amazon Mechanical Turk (MTurk), combined with a survey launched on MTurk to systematically collect workers' real experiences and preferences regarding task clarity, compensation, privacy, response quality, and sensitive content, followed by specific recommendations for improving MTurk NLP data collection based on these results.
Increasing Coverage and Precision of Textual Information in Multilingual Knowledge Graphs
Simone Conia (Sapienza University of Rome), Yunyao Li (Apple)
Large Language ModelTextBenchmark
🎯 What it does: This paper proposes an automated knowledge graph embedding (KGE) method aimed at improving the coverage and accuracy of entity names and descriptions for non-English languages in Wikidata.
Increasing Probability Mass on Answer Choices Does Not Always Improve Accuracy
Sarah Wiegreffe (Allen Institute for AI), Ashish Sabharwal (Allen Institute for AI)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the problem of probability mass dispersion caused by surface form competition (SFC) in multiple-choice tasks and proposes a quantifiable measure of probability mass (PMA) for SFC along with its upper bound.
Indicative Summarization of Long Discussions
Shahbaz Syed (Leipzig University), Martin Potthast (Leipzig University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an unsupervised method for long discussion indicative summarization, leveraging LLMs to first cluster sentences, generate cluster labels, and then assign argument frameworks to form a hierarchical summary resembling a table of contents.
Inference-Time Policy Adapters (IPA): Tailoring Extreme-Scale LMs without Fine-tuning
Ximing Lu (Allen Institute for Artificial Intelligence), Yejin Choi (Allen Institute for Artificial Intelligence)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose Inference-time Policy Adapters (IPA), which optimize the outputs of large language models (e.g., GPT-3) during inference using a lightweight adapter, avoiding the need for model fine-tuning.
Influence Scores at Scale for Efficient Language Data Sampling
Nikhil Anand, Maria Minakova (Amazon Alexa AI)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: This paper evaluates and applies various influence scores for efficient sampling of language data. First, it benchmarks metrics such as VoG, EL2N, TracIn, and PVI on the SNLI dataset, then migrates VoG to the NLU stack of a commercial voice assistant, demonstrating that model performance can be maintained or slightly improved even with approximately halved training data.
INFORM : Information eNtropy based multi-step reasoning FOR large language Models
Chuyue Zhou (Soochow University), Min Zhang (Soochow University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the INFORM framework, which uses information entropy to select problems, automatically generates chained reasoning steps, and enhances LLM reasoning performance through information entropy-based self-consistent reasoning;
Information Value: Measuring Utterance Predictability as Distance from Plausible Alternatives
Mario Giulianelli (University of Amsterdam), Raquel Fernández (University of Amsterdam)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes an 'information value' metric based on a complete sentence alternative set, which quantifies the predictability of a sentence relative to feasible alternative sentences that can be generated in a given context.
Instruct and Extract: Instruction Tuning for On-Demand Information Extraction
Yizhu Jiao (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBenchmarkChain-of-Thought
🎯 What it does: Proposed the 'on-demand information extraction' task and constructed the corresponding benchmark dataset INSTRUCTIE; based on this, trained an instruction-tuned model ODIE to extract structured tables from text according to user instructions.
Instructed Language Models with Retrievers Are Powerful Entity Linkers
Zilin Xiao (Rice University), Daxin Jiang (Microsoft STCA)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes and evaluates the generative entity linking model INSGENEL, along with its retrieval-augmented version INSGENEL-R and ICL version, demonstrating how decoder-only language models can achieve high-precision generation in entity linking tasks.
Instructive Dialogue Summarization with Query Aggregations
Bin Wang (Institute for Infocomm Research, A*STAR), Nancy Chen
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the InstructDS model, which supports query-based instruction dialogue summarization.
INSTRUCTSCORE: Towards Explainable Text Generation Evaluation with Automatic Feedback
Wenda Xu (University of California, Santa Barbara), Lei Li (Carnegie Mellon University)
GenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Developed INSTRUCTSCORE, an interpretable text generation evaluation metric that provides scores and diagnostic reports.
Integrating Language Models into Direct Speech Translation: An Inference-Time Solution to Control Gender Inflection
Dennis Fucci (University of Trento), Luisa Bentivogli (Fondazione Bruno Kessler)
TransformerLarge Language ModelTextAudio
🎯 What it does: Control gender morphology in translation by injecting gender-specific external language models during direct speech translation.
Interactive Text Generation
Felix Faltings (MIT), Bill Dolan (Microsoft)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose an interactive text generation framework that can be trained without requiring real users. The model alternately edits text with users/user simulators during generation, aiming to rapidly approach a predefined target text.
Interactive Text-to-SQL Generation via Editable Step-by-Step Explanations
Yuan Tian (Purdue University), Tianyi Zhang (University Of Sydney)
Explainability and InterpretabilityAI Code AssistantLarge Language ModelTextTabularChain-of-Thought
🎯 What it does: Developed an interactive text-to-SQL system called STEPS, which utilizes editable step-by-step natural language explanations to enable users to directly correct erroneous SQL clauses.
InterFair: Debiasing with Natural Language Feedback for Fair Interpretable Predictions
Bodhisattwa Majumder (Allen Institute for AI), Julian McAuley (University of San Diego)
ClassificationExplainability and InterpretabilityRecurrent Neural NetworkLarge Language ModelBiomedical Data
🎯 What it does: Propose INTERFAIR, an interactive framework that enables users to adjust the model's explanation (bias rationale) during testing through natural language feedback to balance task performance and bias mitigation.
Interpreting and Exploiting Functional Specialization in Multi-Head Attention under Multi-task Learning
Chong Li (Chinese Academy of Sciences), Chengqing Zong (University of Chinese Academy of Sciences)
Explainability and InterpretabilityComputational EfficiencyTransformerSupervised Fine-TuningText
🎯 What it does: This paper quantifies the functional specialization of multi-head attention in multi-task learning through the Importance Attention Head Pruning (IAP) method, and proposes Importance Attention Head Training (IAT) to enhance multi-task and transfer learning performance.
Interpreting Embedding Spaces by Conceptualization
Adi Simhi (Technion Israel Institute of Technology), Shaul Markovitch (Technion Israel Institute of Technology)
ClassificationExplainability and InterpretabilityRepresentation LearningLarge Language ModelTextGraph
🎯 What it does: Proposes an algorithm called Concept Embedding Space (CES) that maps non-interpretable embedding spaces generated by LLMs to an interpretable concept space, enabling explanation and analysis of embeddings through the concept space.
Interventional Rationalization
Linan Yue (University of Science and Technology of China), Zhenya Huang (University of Science and Technology of China)
ClassificationExplainability and InterpretabilityRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper proposes an Inter-RAT method based on causal intervention, which constructs a causal graph to identify and eliminate spurious correlations (shortcuts) between inputs, rationalized subsequences, and labels, achieving more reliable reasoning extraction and prediction.
Interview Evaluation: A Novel Approach for Automatic Evaluation of Conversational Question Answering Models
Xibo Li (Soochow University), Yu Hong (Soochow University)
Large Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Proposed a new evaluation framework called Interview Evaluation, which uses a question-answering agent (Q agent) and a response agent (A agent) to simulate an interview process, dynamically generating questions and conducting round-by-round evaluations based on model responses.
Introducing Rhetorical Parallelism Detection: A New Task with Datasets, Metrics, and Baselines
Stephen Bothwell (University of Notre Dame), David Chiang (University of Notre Dame)
RecognitionRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper proposes the rhetorical parallel detection task, constructs two new datasets: the Latin sermons dataset ASP and the Chinese student essays collection PSE-I, designs evaluation metrics, and implements multiple baseline models.
Inverse Scaling Can Become U-Shaped
Jason Wei (OpenAI), Quoc Le
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Evaluate the performance of larger PaLM language models on Inverse Scaling Prize tasks, finding that inverse scaling often turns into U-shaped scaling, and explore whether 1-shot demonstrations and chain-of-thought prompts can alleviate this issue.
Investigating Bias in Multilingual Language Models: Cross-Lingual Transfer of Debiasing Techniques
Manon Reusens (KU Leuven), Bart Baesens (KU Leuven)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated biases in multilingual BERT models regarding gender, race, and religion, and evaluated the migration effectiveness of various debiasing techniques across English, French, German, and Dutch.
Investigating Efficiently Extending Transformers for Long Input Summarization
Jason Phang (New York University), Peter Liu (Google Research)
GenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This study proposes an efficient block-local + global token (Global-Local) Transformer through systematic experiments on the Transformer architecture and pre-training methods, combined with a short-long pre-training strategy, successfully extending the PEGASUS model to handle summarization tasks for documents with up to 16K tokens;
Is ChatGPT a General-Purpose Natural Language Processing Task Solver?
Chengwei Qin (Nanyang Technological University), Diyi Yang (Shanghai Jiao Tong University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Systematically evaluates ChatGPT's performance across seven task categories through zero-shot assessment on 20 NLP datasets.
Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking Agents
Weiwei Sun (Shandong University), Zhaochun Ren (Leiden University)
RetrievalKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Investigated the ability of large language models (ChatGPT, GPT-4) to re-rank retrieval results in information retrieval, and proposed an instructional method for directly generating sorted permutations;
It Ain’t Over: A Multi-aspect Diverse Math Word Problem Dataset
Jiwoo Kim (Sungkyunkwan University, Republic of Korea), Jongwuk Lee (Sungkyunkwan University, Republic of Korea)
Data SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: A multi-faceted and diverse mathematical word problem dataset called DMath is proposed, covering different question types, vocabulary diversity, bilingual support (English and Korean), and intermediate reasoning formats (expression trees and Python code).
JASMINE: Arabic GPT Models for Few-Shot Learning
El Moatez Billah Nagoudi (University of British Columbia), Md Tawkat Islam Khondaker (University of British Columbia)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Built four autoregressive Transformer language models (JASMINE) with parameter scales ranging from 300 million to 6.7 billion, and provided a multilingual benchmark for Arabic;
Joint Entity and Relation Extraction with Span Pruning and Hypergraph Neural Networks
Zhaohui Yan (ShanghaiTech University), Kewei Tu (ShanghaiTech University)
Graph Neural NetworkLarge Language ModelText
🎯 What it does: Propose a joint entity and relation extraction model HGERE, combining a high-recall span pruner with a hypergraph neural network to achieve high-order interactive reasoning
Joint Geometrical and Statistical Domain Adaptation for Cross-domain Code Vulnerability Detection
Qianjin Du, Jidong Zhai (Tsinghua University)
Domain AdaptationTransformerContrastive LearningTextSequential
🎯 What it does: This paper proposes the MNCRI framework to address domain discrepancy issues in cross-domain code vulnerability detection, integrating two techniques: geometric domain alignment and statistical domain alignment;
JointMatch: A Unified Approach for Diverse and Collaborative Pseudo-Labeling to Semi-Supervised Text Classification
Henry Peng Zou (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
ClassificationTextBenchmark
🎯 What it does: Proposes JointMatch, a unified framework for semi-supervised text classification, which generates high-quality pseudo-labels and enhances model performance by leveraging adaptive thresholds, cross-annotation, and weighted consistency updates.
Joyful: Joint Modality Fusion and Graph Contrastive Learning for Multimoda Emotion Recognition
Dongyuan Li (Tokyo Institute of Technology), Manabu Okumura (Tokyo Institute of Technology)
RecognitionGraph Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: Proposed a framework named JOYFUL that combines joint multimodal fusion with graph contrastive learning for dialogue emotion recognition.
Just Adjust One Prompt: Enhancing In-Context Dialogue Scoring via Constructing the Optimal Subgraph of Demonstrations and Prompts
Jiashu Pu, Rongsheng Zhang (Fuxi AI Lab, NetEase Inc.)
TransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: Propose the ADOROR method, leveraging in-context learning of large language models for dimension-agnostic dialogue evaluation, and enhance scoring quality through automatic prompt generation, personalized selection of demonstrations and prompts, and constructing optimal subgraphs.
Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback
Katherine Tian (Harvard University), Christopher Manning (Stanford University)
Reinforcement Learning from Human FeedbackLarge Language ModelPrompt EngineeringText
🎯 What it does: Studies the calibration of RLHF fine-tuned LMs, systematically evaluates multiple methods for extracting confidence from models, and proposes obtaining better calibrated confidence through model verbalized probabilities.
KCTS: Knowledge-Constrained Tree Search Decoding with Token-Level Hallucination Detection
Sehyun Choi (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
GenerationTransformerLarge Language ModelText
🎯 What it does: Propose a method that, without altering the weights of large language models, uses knowledge-constrained tree search (KCTS) and token-level hallucination detection (RIPA) to guide text generation, making it more aligned with given knowledge and reducing hallucinations.
KEBAP: Korean Error Explainable Benchmark Dataset for ASR and Post-processing
Seonmin Koo (Korea University), Heuiseok Lim (Korea University)
Explainability and InterpretabilityLarge Language ModelTextBenchmarkAudio
🎯 What it does: This paper proposes and implements an explainable benchmark dataset for Korean speech recognition systems—KEBAP—and performs fine-grained classification and evaluation of speech-level and text-level errors on this dataset. Additionally, the benchmark is used to diagnose the performance of Google Cloud Speech-to-Text and CLOVA Speech, with attempts to verify error type classification using ChatGPT.
KEPL: Knowledge Enhanced Prompt Learning for Chinese Hypernym-Hyponym Extraction
Ningchen Ma (China University of Mining & Technology), Suncong Zheng (Chinese Academy of Sciences)
RecognitionRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Knowledge-Enhanced Prompt Learning (KEPL) model, which utilizes Hearst-like patterns as prompts to simultaneously embed patterns and text, achieving Chinese hierarchical relation extraction.
kNN-LM Does Not Improve Open-ended Text Generation
Shufan Wang (University of Massachusetts Amherst), Mohit Iyyer (University of Massachusetts Amherst)
GenerationRetrievalTextRetrieval-Augmented Generation
🎯 What it does: This paper investigates the performance of retrieval-based interpolation language models (kNN-LM) in open-ended text generation, demonstrating that although they significantly reduce perplexity, they do not improve generation quality.
Knowledge Distillation \approx Label Smoothing: Fact or Fallacy?
Md Sultan
ClassificationKnowledge DistillationTransformerText
🎯 What it does: This paper theoretically analyzes and empirically verifies the similarity between knowledge distillation (KD) and label smoothing (LS) in NLP text classification tasks, focusing on comparing the impact of the two methods on model confidence and generalization ability.
Knowledge Graph Compression Enhances Diverse Commonsense Generation
EunJeong Hwang (University of British Columbia), Tengfei Ma (Stony Brook University)
GenerationCompressionGraph Neural NetworkTransformerLarge Language ModelMixture of ExpertsTextGraph
🎯 What it does: Utilize a differentiable graph compression algorithm to automatically select the most task-relevant concepts in the subgraph of the Commonsense Knowledge Graph (ConceptNet), and enhance the diversity and quality of generative models using the compressed subgraph.
Knowledge Rumination for Pre-trained Language Models
Yunzhi Yao (Zhejiang University), Ningyu Zhang (Zhejiang University)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the Knowledge Rumination method, which leverages the hidden knowledge within pre-trained language models to improve the performance of knowledge-intensive tasks.
Knowledge-Augmented Language Model Verification
Jinheon Baek (Korea Advanced Institute Of Science And Technology), Sung Hwang
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphRetrieval-Augmented Generation
🎯 What it does: Designed and evaluated a validator called KALMV, capable of detecting knowledge retrieval errors and generation errors, and enhancing the reliability of knowledge-augmented language models through multi-instruction integration and iterative error correction during reasoning.
KRLS: Improving End-to-End Response Generation in Task Oriented Dialog with Reinforced Keywords Learning
Xiao Yu (Columbia University), Zhou Yu (Columbia University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose the KRLS algorithm, combining offline reinforcement learning, next-word sampling, and fine-grained rewards to improve end-to-end response generation in task-oriented dialogue.
Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning
Lean Wang (Peking University), Xu Sun (Peking University)
ClassificationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Investigated the information flow mechanism in in-context learning of large language models, proposing that label words serve as anchor points for information aggregation and extraction, thereby improving inference speed and accuracy.
LACMA: Language-Aligning Contrastive Learning with Meta-Actions for Embodied Instruction Following
Cheng-Fu Yang (Ucla), Kai-Wei Chang (Ucla)
Robotic IntelligenceConvolutional Neural NetworkTransformerVision-Language-Action ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the LACMA method, which significantly enhances instruction following performance in unseen environments through contrastive learning that aligns hidden states with instructions and introduces meta-actions.
Language and Mental Health: Measures of Emotion Dynamics from Text as Linguistic Biosocial Markers
Daniela Teodorescu (University of Alberta), Saif Mohammad (National Research Council Canada)
Text
🎯 What it does: Investigated the relationship between the Utterance Emotion Dynamics (UED) metric in tweets and self-reported mental health diagnoses (e.g., ADHD, depression, bipolar disorder, PTSD), calculating indicators such as average emotion, emotional variability, rising rate, and recovery rate, and comparing them with a control group.
Language Model is Suitable for Correction of Handwritten Mathematical Expressions Recognition
Zui Chen (Shanghaitech University), Yi Zhou (Shanghaitech University)
RecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: Proposed a Recognition and Language Fusion Network (RLFN) that jointly corrects errors in handwritten mathematical expression recognition by leveraging the MathBERTa language model and a visual recognition module.
Language Model Quality Correlates with Psychometric Predictive Power in Multiple Languages
Ethan Wilcox (ETH Zürich), Tiago Pimentel (ETH Zürich)
TransformerLarge Language ModelTextMultimodality
🎯 What it does: This paper systematically examines the correlation between language model quality and its predictive ability for human reading time (QP hypothesis) through cross-lingual evaluation across 13 languages.
Language Models with Rationality
Nora Kassner (Allen Institute for AI), Peter Clark (Allen Institute for AI)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Building upon large language models, a self-reflective 'rational layer' is introduced, constructing a belief graph and modifying inconsistencies through constraint solving to achieve interpretable and consistent answers.
Language Representation Projection: Can We Transfer Factual Knowledge across Languages in Multilingual Language Models?
Shaoyang Xu (Tianjin University), Deyi Xiong (Tianjin University)
RetrievalDomain AdaptationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes a parameter-agnostic language representation projection framework called LRP2, which can map representations of non-English languages into the English representation space and then back, significantly improving the accuracy of multilingual language models in fact knowledge retrieval tasks.
Large Language Models and Multimodal Retrieval for Visual Word Sense Disambiguation
Anastasia Kritharoula (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
RetrievalTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Studied the Visual Word Sense Disambiguation (VWSD) task, proposing multiple methods such as vision/text retrieval, LLM knowledge enhancement, single-modal retrieval, learning to rank, and QA + chain-of-thought to improve retrieval accuracy.
Large Language Models are biased to overestimate profoundness
Eugenio Herrera-Berg (Centro Nacional de Inteligencia Artificial), Cristian Buc Calderon (Centro Nacional de Inteligencia Artificial)
Explainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: Evaluate the bias of GPT-4 and other LLMs in judging the depth of bland, motivational, and pseudo-profound sentences, and explore the impact of different prompting strategies (few-shot, chain-of-thought) on the results.
Large Language Models are Complex Table Parsers
Bowen Zhao (Fudan University), Xiaobo Zhang (Fudan University)
TransformerLarge Language ModelPrompt EngineeringTabularChain-of-Thought
🎯 What it does: Using GPT-3.5 as a parser for complex tables, first reconstruct JSON format tables into tuples containing hierarchy, row/column positions, and content; subsequently, complete table Q&A tasks through carefully designed single-round and multi-round prompts (including chain-of-thought and code assistance).
Large Language Models are Temporal and Causal Reasoners for Video Question Answering
Dohwan Ko (Korea University), Hyunwoo Kim (Korea University)
TransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper studies the temporal and causal reasoning capabilities of large language models (LLMs) in video question answering (VideoQA), and proposes the Flipped-VQA framework. By flipping the input and output of three tasks (VQ→A, VA→Q, QA→V), the framework enables LLMs to better leverage their pre-trained reasoning knowledge for VideoQA.
Large Language Models Can Self-Improve
Jiaxin Huang (University of Illinois at Urbana-Champaign), Jiawei Han (University of Illinois at Urbana-Champaign)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Leverage pre-trained large language models on unannotated question-answering data, generating high-confidence reasoning paths through Chain-of-Thought (CoT) and self-consistency, then fine-tuning the model using these self-generated reasoning answers to achieve self-improvement.
Large Language Models Meet Open-World Intent Discovery and Recognition: An Evaluation of ChatGPT
Xiaoshuai Song (Beijing University of Posts and Telecommunications), Weiran Xu (Meituan)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringTextBenchmarkFinance Related
🎯 What it does: Evaluate ChatGPT's performance in open-domain (OOD) intent discovery and general intent discovery (GID) tasks, conducting experiments through three prompt designs: zero-shot, few-shot, and no-shot.
Large Language Models Only Pass Primary School Exams in Indonesia: A Comprehensive Test on IndoMMLU
Fajri Koto (MBZUAI), Timothy Baldwin (University of Melbourne)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper creates and publishes IndoMMLU, a multi-task, multi-language evaluation benchmark covering Indonesian primary to university entrance levels, assessing LLMs' reasoning and knowledge capabilities in Indonesian language and culture using 14,981 multiple-choice questions collected by teachers.
Large Language Models: The Need for Nuance in Current Debates and a Pragmatic Perspective on Understanding
Bram van Dijk (Leiden Institute of Advanced Computer Science), Max Johannes van Duijn (Leiden Institute of Advanced Computer Science)
TransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: This paper conducts a literature review and theoretical analysis, thoroughly examining three major criticisms regarding the capabilities of large language models (LLMs) (only performing next-word prediction, only mastering formal language, and being unable to provide information for human language acquisition), and proposes more detailed, empirically supported interpretations of these arguments. It then introduces a pragmatic philosophy perspective, discussing the concepts and roles of 'true understanding' and 'intentionality' in human-LLM interaction, and explores the rationality and limitations of attributing mental states to LLMs in human-computer interaction practices.
Large-scale similarity search with Optimal Transport
Cléa Laouar (Okinawa Institute of Science and Technology), Makoto Yamada (Okinawa Institute of Science and Technology)
RetrievalOptimizationComputational EfficiencyText
🎯 What it does: Propose an efficient k-NN search method based on tree-structured approximation of Wasserstein distance, and implement large-scale text retrieval.
Larger Probes Tell a Different Story: Extending Psycholinguistic Datasets Via In-Context Learning
Namrata Shivagunde (University of Massachusetts Lowell), Anna Rumshisky (University of Massachusetts Lowell)
Data SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Extended the original psycholinguistic evaluation datasets NEG-136-SIMP and ROLE-88 by using templates and GPT-3 to generate larger-scale datasets, NEG-1500-SIMP (750 pairs each via two methods) and ROLE-1500 (750 pairs), and conducted zero-shot evaluations on 22 language models.
Lazy-k Decoding: Constrained Decoding for Information Extraction
Arthur Hemmer (Shift Technology), Jean-marc Ogier
ClassificationTransformerText
🎯 What it does: Proposed a new constrained decoding method called Lazyk, which can perform global hard constraint search on probability-based token-classification models without the need for explicit linearization of constraints;
Leap-of-Thought: Accelerating Transformers via Dynamic Token Routing
Yeachan Kim (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyTransformerText
🎯 What it does: Introduce a dynamic token router in Transformer, allowing each token to decide at each layer whether to be processed by the current layer or skipped to the next layer, merging into pseudo tokens when skipped; train the router with gradient guidance to retain important information while significantly reducing the number of tokens per layer.
Learn and Consolidate: Continual Adaptation for Zero-Shot and Multilingual Neural Machine Translation
Kaiyu Huang (Tsinghua University), Yang Liu (Dalian University of Technology)
Domain AdaptationKnowledge DistillationRepresentation LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: This paper proposes a two-stage continual adaptation framework called LCCA to improve zero-shot translation and multilingual translation performance in neural machine translation under newly added parallel corpora.
Learning Co-Speech Gesture for Multimodal Aphasia Type Detection
Daeun Lee (Sungkyunkwan University), Jinyoung Han (University of South Florida)
ClassificationGraph Neural NetworkTransformerTextMultimodalityAudio
🎯 What it does: Researchers propose a multimodal model based on graph neural networks that uses voice and gesture resonance information for Aphasia type detection.
Learning From Free-Text Human Feedback – Collect New Datasets Or Extend Existing Ones?
Dominic Petrak (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
GenerationData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: This paper studies how to evaluate the feasibility of using existing dialogue data, including system errors and types and frequencies of user free-text feedback, to learn free-text human feedback; proposes modified integrated error classification and user response type classification methods, and validates their effectiveness through manual annotation; experiments show that incorporating error information and user feedback as additional inputs can improve dialogue generation quality.
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
Danqing Wang (University of California Santa Barbara), Lei Li (Carnegie Mellon University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the SALAM framework, enabling large language models to interact with auxiliary learning assistants during the error collection phase, improving model performance through error memory retrieval and feedback guidance.
Learning Knowledge-Enhanced Contextual Language Representations for Domain Natural Language Understanding
Taolin Zhang (East China Normal University), Weining Qian (Alibaba Group)
ClassificationRecognitionRepresentation LearningTransformerLarge Language ModelContrastive LearningGraphBiomedical DataFinance Related
🎯 What it does: Proposes the KANGAROO framework, which enhances the knowledge representation of language models in closed-domain knowledge graphs using hyper-sphere embedding and multi-layer contrastive learning based on point-biconnected components.
Learning Language-guided Adaptive Hyper-modality Representation for Multimodal Sentiment Analysis
Haoyu Zhang (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
ClassificationTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes an Adaptive Language-Guided Multimodal Transformer (ALMT), which enhances the robustness and accuracy of Multimodal Sentiment Analysis (MSA) by suppressing irrelevant and conflicting information in visual and audio modalities.
Learning Preference Model for LLMs via Automatic Preference Data Generation
Shijia Huang (Chinese University of Hong Kong), Liwei Wang (Chinese University of Hong Kong)
Data-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a framework for automatically generating preference data (AutoPM) to train preference models of large language models, thereby enhancing their alignment with human values.
Learning Retrieval Augmentation for Personalized Dialogue Generation
Qiushi Huang (University of Surrey), Lilian Tang (University of Surrey)
GenerationRetrievalTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose the LAPDOG framework, leveraging external story knowledge to enhance personalized dialogue generation.
Learning the Visualness of Text Using Large Vision-Language Models
Gaurav Verma (Georgia Institute of Technology), Ani Nenkova (Adobe Research)
Explainability and InterpretabilityRepresentation LearningData-Centric LearningSupervised Fine-TuningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Studied sentence-level visuality (visuality) prediction, constructed the TIMED dataset with 3,620 manually annotated English sentences, and proposed the TIP-CLIP fine-tuning method by matching NULL images through CLIP.
Learning to Describe for Predicting Zero-shot Drug-Drug Interactions
Fangqi Zhu (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: This paper proposes a zero-shot drug-drug interaction (DDI) prediction method called TextDDI for new drugs, which predicts interactions between unknown drugs by utilizing text descriptions of drugs from online databases.
Learning to Predict Task Transferability via Soft Prompt
Lingyun Feng (China Mobile Information Technology Center)
Representation LearningData-Centric LearningMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A soft prompt-based task affinity score function was studied and implemented on 50 diverse NLP tasks to predict transfer benefits from source tasks to target tasks, enabling efficient selection of the most helpful intermediate tasks.
Learning to Rank Context for Named Entity Recognition Using a Synthetic Dataset
Arthur Amalvy (Laboratoire Informatique d'Avignon), Richard Dufour (Laboratoire des Sciences du Numérique de Nantes)
RecognitionData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Using instruction-tuned large language models (Alpaca) to automatically generate synthetic context retrieval training sets, and training a BERT-based neural retriever to retrieve relevant contexts for named entity recognition (NER) in long texts.
Learning to Rank Generation with Pairwise Partial Rewards
Youngwon Lee (Seoul National University), Seung-won Hwang (Seoul National University)
GenerationTransformerReinforcement LearningContrastive LearningText
🎯 What it does: This paper proposes a Pairwise Partial Reward (PPR) method based on a prefix tree, which uses reinforcement learning to provide partial rewards for actions generated at each step, thereby learning to rank different actions and significantly improving the quality of conditional text generation.
Length Does Matter: Summary Length can Bias Summarization Metrics
Xiaobo Guo (Dartmouth), Soroush Vosoughi (Dartmouth)
GenerationTransformerText
🎯 What it does: Investigated the bias of generated summary length on 14 automatic evaluation metrics and proposed a Bayesian network-based length normalization method to mitigate this bias.
Length is a Curse and a Blessing for Document-level Semantics
Chenghao Xiao (Durham University), Noura Al Moubayed (Durham University)
RetrievalRepresentation LearningTransformerContrastive LearningText
🎯 What it does: The study investigates the generalization ability of contrastive learning models under different text lengths, finding that they are vulnerable to 'length attacks,' and proposes a framework named LA(SER)3 for unsupervised contrastive learning using length-invariant semantics.
Let GPT be a Math Tutor: Teaching Math Word Problem Solvers with Customized Exercise Generation
Zhenwen Liang, Ashwin Kalyan
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose the CEMAL method, which utilizes LLMs to generate practice problems targeting students' weaknesses, thereby enhancing the learning effectiveness of a small-scale math word problem solver.
Let’s Sample Step by Step: Adaptive-Consistency for Efficient Reasoning and Coding with LLMs
Pranjal Aggarwal (Indian Institute of Technology), Mausam (Indian Institute of Technology)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Propose the Adaptive-Consistency method, which dynamically adjusts the sampling quantity in large language model inference and decides whether to stop sampling based on sample consistency.
Let’s Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought
Vaishnavi Himakunthala (University of California Santa Barbara), William Wang (University of California Santa Barbara)
Large Language ModelVideoTextBenchmarkChain-of-Thought
🎯 What it does: Propose the VIP dataset, which includes unstructured dense captions and structured FAMOUS descriptions for video keyframes, and define two tasks, video filling and prediction, to evaluate multi-frame reasoning capabilities.
Lifelong Sequence Generation with Dynamic Module Expansion and Adaptation
Chengwei Qin (Nanyang Technological University), Shafiq Joty (Nanyang Technological University)
GenerationMeta LearningNeural Architecture SearchRecurrent Neural NetworkTextSequential
🎯 What it does: Propose a Dynamic Module Expansion and Adaptation (DMEA) framework for lifelong sequence generation, combining pseudo-sample replay and dynamic gradient scaling to mitigate catastrophic forgetting and promote forward knowledge transfer.
LIMIT: Language Identification, Misidentification, and Translation using Hierarchical Models in 350+ Languages
Milind Agarwal (George Mason University), Antonios Anastasopoulos (George Mason University)
ClassificationRecognitionTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the MCS-350 parallel children's story corpus with over 50K entries, and proposed the LIMIT hierarchical model for language identification and machine translation in low-resource languages.
LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers
Theo Olausson (MIT), Roger Levy (MIT)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose a neural-symbolic framework called LINC, which uses a large language model as a semantic parser to convert natural language premises into first-order logic expressions, then passes them to an external theorem prover Prover9 for symbolic reasoning, and obtains the final conclusion through majority voting.
Linear-Time Modeling of Linguistic Structure: An Order-Theoretic Perspective
Tianyu Liu (ETH Zurich), Ryan Cotterell (ETH Zurich)
Computational EfficiencyText
🎯 What it does: This paper proposes a linear-time structural prediction framework based on partial order theory, modeling the relationship between word pairs in a sentence as partial orders, and recovering dependency trees and coreference networks by predicting the intersection of several total orders.
Ling-CL: Understanding NLP Models through Linguistic Curricula
Mohamed Elgaar (University of Massachusetts Lowell), Hadi Amiri (University of Massachusetts Lowell)
OptimizationTransformerTextBenchmark
🎯 What it does: Propose a multi-perspective curriculum learning framework based on human-verified language complexity indices. By dynamically learning the importance weights of indices and aggregating them into difficulty scores, the framework designs time-varying sigmoid, negative sigmoid, and Gaussian weight curves for NLP model training. This enables gradual introduction of samples according to difficulty during training, revealing the core linguistic knowledge mastered by the model during learning.
Linking Surface Facts to Large-Scale Knowledge Graphs
Gorjan Radevski (NEC Laboratories Europe), Goran Glavaš (NEC Laboratories Europe)
Representation LearningData-Centric LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: Proposes an open information extraction (OIE) fact linking task for large-scale knowledge graphs (KGs), and constructs a new multi-dimensional benchmark FaLB covering transductive, inductive, polysemous, and scenarios where OIE fragments are not in KG.
Lion: Adversarial Distillation of Proprietary Large Language Models
Yuxin Jiang (Hong Kong University of Science and Technology), Wei Wang (Hong Kong University of Science and Technology)
Knowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes a framework based on adversarial knowledge distillation, iteratively transferring knowledge from proprietary large language models (e.g., ChatGPT) to small open-source models (Lion), leveraging three roles—teacher, judge, and generator—to form a simulation-discrimination-generation cycle.
LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models
Zhiqiang Hu (Singapore University Of Technology And Design), Roy Lee
Computational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Developed the LLM-Adapters framework, integrating multiple adapters for parameter-efficient fine-tuning of large language models (LLMs), and conducted systematic empirical studies on arithmetic reasoning and common sense reasoning tasks.
LLM-enhanced Self-training for Cross-domain Constituency Parsing
Jianling Li (Tianjin University), Yue Zhang (Westlake University)
Domain AdaptationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Generate raw corpus conforming to target domain grammar rules using large language models (LLM) during self-training, iteratively train and adapt for cross-domain syntactic parsing;
LLM-FP4: 4-Bit Floating-Point Quantized Transformers
Shih-yang Liu (Hong Kong University of Science and Technology), Kwang-Ting Cheng (Hong Kong University of Science and Technology)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes LLM-FP4, a quantization method that can compress large language models (LLMs) and their activations, weights to 4-bit floating point during the post-training phase.
LLM-powered Data Augmentation for Enhanced Cross-lingual Performance
Chenxi Whitehouse (City University of London), Alham Fikri Aji (MBZUAI)
Data SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Generate synthetic training samples for multilingual common-sense reasoning using large language models (Dolly-v2, StableVicuna, ChatGPT, GPT-4), and fine-tune small multilingual models (mBERT, XLM-R) with these samples to improve cross-lingual performance.
LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models
Huiqiang Jiang (Microsoft Corporation), Lili Qiu (Microsoft Corporation)
CompressionComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose LLMLingua, a coarse-to-fine hierarchical prompt compression framework that achieves large-scale language model inference acceleration while preserving semantic integrity through a budget controller, iterative token-level compression, and distribution alignment.