EMNLP 2023 Papers with AI Summaries
Conference on Empirical Methods in Natural Language Processing · 1047 papers
→ EMNLP 2023 papers with code (380)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 1047 EMNLP 2023 papers by keyword, author or institution —
free trial on arXivSub.
‘Don’t Get Too Technical with Me’: A Discourse Structure-Based Framework for Automatic Science Journalism
Ronald Cardenas (University of Edinburgh), Yufang Hou (IBM Research Europe)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This study proposes an automated scientific news writing framework and creates a new dataset called SCITECHNEWS, which includes scientific papers, corresponding news articles, and expert abstracts;
“Are Your Explanations Reliable?” Investigating the Stability of LIME in Explaining Text Classifiers by Marrying XAI and Adversarial Attack
Christopher Burger (University of Mississippi), Thai Le (University of Mississippi)
ClassificationExplainability and InterpretabilityAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper experimentally demonstrates the inherent instability of LIME in text classification tasks and proposes a new adversarial perturbation algorithm, XAIFOOLER, which can significantly alter the explanations generated by LIME while preserving the original prediction and semantics.
“Fifty Shades of Bias”: Normative Ratings of Gender Bias in GPT Generated English Text
Rishav Hada (Microsoft Research India), Kalika Bali (Microsoft Research India)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Created a collection of English sentences generated by GPT (1,000 sentences) and provided fine-grained gender bias scores using Best-Worst Scaling annotations;
“Mistakes Help Us Grow”: Facilitating and Evaluating Growth Mindset Supportive Language in Classrooms
Kunal Handa (Brown University), Dorottya Demszky (Stanford University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study creates Growth Mindset Supportive Language (GMSL) annotation guidelines and a parallel dataset, utilizes GPT-4 to generate growth mindset rewrites of teacher discourse, and evaluates the effectiveness of the rewrites through large-scale teacher-student questionnaires.
3DRP-Net: 3D Relative Position-aware Network for 3D Visual Grounding
Zehan Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
Object DetectionConvolutional Neural NetworkTransformerPoint Cloud
🎯 What it does: Propose a single-stage 3D visual localization network, 3DRP-Net, which achieves precise localization of target objects in 3D point clouds through 3D relative position multi-head attention (3DRP-MA) and a soft label strategy.
4 and 7-bit Labeling for Projective and Non-Projective Dependency Trees
Carlos Gómez-Rodríguez (Universidade da Coruña), David Vilares (Universidade da Coruña)
Representation LearningData-Centric LearningTransformerText
🎯 What it does: Propose two dependency tree sequence labeling encodings with finite label sets (4-bit and 7-bit), and provide linear-time encoding and decoding algorithms; evaluate their performance on multiple treebanks.
A Benchmark for Reasoning with Spatial Prepositions
Iulia Comsa, Srini Narayanan (Google DeepMind)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose and release a benchmark dataset for evaluating large models' ability to reason about spatial prepositions, containing 400 balanced samples in English and Romanian.
A Challenging Multimodal Video Summary: Simultaneously Extracting and Generating Keyframe-Caption Pairs from Video
Keito Kudo (Tohoku University), Nobuyuki Shimizu (LY Corporation)
GenerationTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
🎯 What it does: Proposed a multi-modal video summarization task named Multi-VidSum, which requires selecting keyframes and generating corresponding descriptions for each frame simultaneously under a given summary length, along with constructing the corresponding dataset and baseline models.
A Cheaper and Better Diffusion Language Model with Soft-Masked Noise
Jiaao Chen (Georgia Institute of Technology), Diyi Yang (Georgia Institute of Technology)
GenerationTransformerDiffusion modelText
🎯 What it does: Proposed a language-oriented diffusion model, Masked-Diffusion LM, which gradually destroys text during the forward diffusion process using soft masking noise based on word importance (TF-IDF + entropy), and directly predicts discrete vocabulary via cross-entropy during the reverse process, avoiding traditional high-dimensional continuous-to-discrete mapping.
A Comprehensive Evaluation of Biomedical Entity Linking Models
David Kartchner (Enveda Biosciences), Cassie Mitchell (Georgia Institute of Technology)
RetrievalTransformerLarge Language ModelBiomedical DataReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: This paper constructs a unified evaluation framework to systematically assess nine mainstream biomedical entity linking models across multiple datasets.
A Deeper (Autoregressive) Approach to Non-Convergent Discourse Parsing
Oren Tsur (Ben Gurion University of the Negev), Yoav Tulpan (Ben Gurion University of the Negev)
ClassificationTransformerLarge Language ModelText
🎯 What it does: Propose a unified autoregressive Transformer model named N-CoDiP for multi-label discourse parsing in non-convergent controversial dialogues;
A Diachronic Analysis of Paradigm Shifts in NLP Research: When, How, and Why?
Aniket Pramanick (Ubiquitous Knowledge Processing Lab), Iryna Gurevych (Ubiquitous Knowledge Processing Lab)
Text
🎯 What it does: This study proposes a system framework based on causal discovery and inference to automatically analyze the evolutionary trends and causal relationships between natural language processing (NLP) research tasks and their related entities (tasks, methods, datasets, evaluation metrics) across different time periods.
A Diachronic Perspective on User Trust in AI under Uncertainty
Shehzaad Dhuliawala (ETH Zürich), Mrinmaya Sachan (ETH Zürich)
Explainability and InterpretabilityRecurrent Neural NetworkTextSequential
🎯 What it does: Design a betting game to measure user trust in AI by simulating AI confidence and predictions, and systematically evaluate the impact of miscalibration on trust and collaborative performance.
A Diffusion Weighted Graph Framework for New Intent Discovery
Wenkai Shi, Ping Chen (Lenovo)
ClassificationGraph Neural NetworkTransformerLarge Language ModelDiffusion modelContrastive LearningText
🎯 What it does: Propose a Diffusion Weighted Graph Framework (DWGF), which constructs structural relationship graphs through KNN diffusion in the new intent discovery task, and combines contrastive learning with global self-training; introduce a Graph Smoothing Filter (GSF) to smooth test features during inference.
A Digital Language Coherence Marker for Monitoring Dementia
Dimitris Gkoumas (Queen Mary University of London), Maria Liakata (Queen Mary University of London)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataAlzheimer's Disease
🎯 What it does: Propose a digital linguistic coherence marker based on logical topic consistency to monitor cognitive decline;
A Fair and In-Depth Evaluation of Existing End-to-End Entity Linking Systems
Hannah Bast (University of Freiburg), Natalie Prange (University of Freiburg)
RetrievalTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Conduct a fair and in-depth evaluation of existing end-to-end entity linking systems and propose two new bias-free benchmark datasets.
A Fine-Grained Taxonomy of Replies to Hate Speech
Xinchen Yu (University of Arizona TAMS, University of North Texas), Lingzi Hong (University of North Texas)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a fine-grained reply classification system based on argument theory (four main categories + eight subcategories), build and annotate 3,654 real replies to hate comments on Reddit, analyze language features of different reply types and the civility of subsequent dialogues, and train a model using pre-trained Transformers to predict reply categories.
A Framework for Vision-Language Warm-up Tasks in Multimodal Dialogue Models
Jaewook Lee (Konkuk University), Harksoo Kim (Konkuk University)
TransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the VLAW-MDM framework, which preheats the multimodal dialogue model on target task data through automatically generated image captions and four preheating tasks (caption generation, image replacement, occluded region modeling, occluded language modeling) to enhance the association between images and text;
A Frustratingly Easy Post-Training Quantization Scheme for LLMs
Yongkweon Jeon (Samsung Research), Ho-young Kim (Samsung Research)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a post-training quantization scheme named Z-FOLD, targeting the pre-layer normalization Transformer structure of large language models, achieving low-bit weight quantization (reducible to 2-bit) without additional parameters or computational costs.
A Generation-based Deductive Method for Math Word Problems
Yuxuan Hu (Renmin University of China), Hong Chen (Renmin University of China)
GenerationRecurrent Neural NetworkGraph Neural NetworkTransformerLarge Language ModelTextGraph
🎯 What it does: Propose a multivariate directed acyclic graph (mDAG) and a generative deductive method called GeDe, which automatically generates sequences of mathematical expressions containing advanced operators using a re-encoder and hierarchical beam search.
A linear time approximation of Wasserstein distance with word embedding selection
Sho Otao (Kyoto University), Makoto Yamada
ClassificationComputational EfficiencyRepresentation LearningText
🎯 What it does: Propose a TWD-GFS method that combines tree-structured Wasserstein distance with group feature selection for high-dimensional text distance computation.
A Mechanistic Interpretation of Arithmetic Reasoning in Language Models using Causal Mediation Analysis
Alessandro Stolfo (ETH Zürich), Mrinmaya Sachan (ETH Zürich)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Conduct causal mediation analysis on the internal mechanisms of Transformer language models in arithmetic reasoning tasks to identify which layers and modules are responsible for information transmission and result generation.
A Multi-Task Dataset for Assessing Discourse Coherence in Chinese Essays: Structure, Theme, and Logic Analysis
Hongyi Wu (East China Normal University), Yuanbin Wu (East China Normal University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created and annotated a multi-task discourse coherence dataset for middle school essays, CEDCC, covering three tasks: coherence scoring, topic sentences, and discourse relations, and built a baseline model on this dataset.
A Picture is Worth a Thousand Words: Language Models Plan from Pixels
Anthony Liu (LG AI Research), Honglak Lee (LG AI Research)
TransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes a method that encodes visual observations into learnable 'visual prompts' and directly inputs them into a pre-trained language model (PLM) for planning, achieving end-to-end planning from pixels to actions;
A Predictive Factor Analysis of Social Biases and Task-Performance in Pretrained Masked Language Models
Yi Zhou (Cardiff University), Danushka Bollegala (University of Liverpool)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper conducts a factor analysis on 39 pre-trained masked language models to investigate the impact of model factors on social bias and downstream task performance.
A Quality-based Syntactic Template Retriever for Syntactically-Controlled Paraphrase Generation
Xue Zhang (Beijing Jiaotong University), Jinan Xu (Beijing Jiaotong University)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a quality-based syntactic template retriever (QSTR) that retrieves more suitable syntactic templates by evaluating the impact of templates on the quality of generated rewrites, and designed a multi-template diversity search algorithm (DTS) to enhance the diversity and quality of multiple rewrites.
A Question Answering Framework for Decontextualizing User-facing Snippets from Scientific Documents
Benjamin Newman (University of Washington), Kyle Lo (Allen Institute for AI)
GenerationTransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This paper proposes a decontextualization method for scientific paper fragments based on a question-answering (QA) framework and implements a prompting strategy named QADECONTEXT.
A Rose by Any Other Name would not Smell as Sweet: Social Bias in Names Mistranslation
Sandra Sandoval (University of Maryland), Hal Daumé III (University of Maryland)
Data-Centric LearningText
🎯 What it does: By constructing a name-context dataset called DNIC and employing a back-translation evaluation method, systematically examine the mistranslation differences of machine translation on names associated with different races/genders.
A Scalable Framework for Table of Contents Extraction from Complex ESG Annual Reports
Xinyu Wang (University of Warwick), Yulan He (King's College London)
Data-Centric LearningRecurrent Neural NetworkGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityFinance Related
🎯 What it does: Propose a scalable three-step framework (CMM) for extracting directory structures from complex ESG annual reports.
A Self-enhancement Multitask Framework for Unsupervised Aspect Category Detection
Thi-Nhung Nguyen (VinAI Research), Tuan-Dung Cao (Hanoi University of Science and Technology)
ClassificationTransformerContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Designed the self-enhancing multi-task framework ASeM, leveraging seed word enhancement, retrieval-based data augmentation, and multi-task learning to achieve unsupervised Aspect Category Detection, Aspect Term Extraction, and Aspect Term Polarity.
A Self-training Framework for Automated Medical Report Generation
Siyuan Wang (University of Sydney), Bo Peng (Newcastle University)
GenerationConvolutional Neural NetworkTransformerVision Language ModelImageTextBiomedical Data
🎯 What it does: Propose an automatic medical report generation method based on a teacher-student self-training framework, leveraging unlabeled medical images to enhance model performance;
A Simple Baseline for Knowledge-Based Visual Question Answering
Alexandros Xenos (Queen Mary University of London), Georgios Tzimiropoulos (Queen Mary University of London)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: Proposed a training-free, simple baseline based on LLaMA-13B, solving knowledge-intensive visual question answering tasks through in-context learning using problem-informed caption generation.
A State-Vector Framework for Dataset Effects
Esmat Sahak (University of Toronto), Frank Rudzicz (University of Toronto)
ClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes a framework based on state vectors to quantify the individual and interactive effects of datasets on language models;
A Study on Accessing Linguistic Information in Pre-Trained Language Models by Using Prompts
Marion Di Marco (LMU Munich), Alexander Fraser (LMU Munich)
Representation LearningTransformerPrompt EngineeringText
🎯 What it does: This paper studies how to retrieve linguistic information from pre-trained multilingual language models through natural language prompts, covering morphological features (number, gender, case, tense) and more advanced syntactic tasks (subject-object distinction, verb particles, prepositional phrase attachment, passive voice).
A Suite of Generative Tasks for Multi-Level Multimodal Webpage Understanding
Andrea Burns (Boston University), Mandy Guo (Google)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Proposed the WikiWeb2M dataset, containing the complete text, images, and structural information of over 2 million English Wikipedia pages, and designed three generation tasks based on this: web page description generation, chapter summarization, and context-aware image caption generation.
A Systematic Study of Performance Disparities in Multilingual Task-Oriented Dialogue Systems
Songbo Hu (University of Cambridge), Ivan Vulić (University of Cambridge)
TransformerLarge Language ModelTextReview/Survey Paper
🎯 What it does: Systematically study performance differences in multilingual task-oriented dialogue systems, proposing absolute and relative equivalence metrics;
A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation
Giuseppe Attanasio (Bocconi University), Anne Lauscher
Explainability and InterpretabilityTransformerPrompt EngineeringText
🎯 What it does: This paper investigates gender bias in instruction-fine-tuned models for machine translation. It first analyzes errors caused by the model's omission of pronouns during translation using interpretability methods, and based on this analysis, proposes a debiasing method based on few-shot examples.
A Training-Free Debiasing Framework with Counterfactual Reasoning for Conversational Emotion Detection
Geng Tu (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
RecognitionAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a training-free bias-removal framework (TFD) that extracts and eliminates label, speaker, and neutral word biases in sentiment recognition models by generating adversarial (counterfactual) statements and contexts during the prediction phase.
A Unified View of Evaluation Metrics for Structured Prediction
Yunmo Chen (Johns Hopkins University), Benjamin Van Durme (Johns Hopkins University)
OptimizationText
🎯 What it does: This paper proposes a unified evaluation framework to standardize the metrics for various structured prediction tasks (such as relation extraction, dependency parsing, event extraction, coreference resolution, template extraction, and AMR parsing), and demonstrates how to express existing metrics and design new ones using this framework.
A Video Is Worth 4096 Tokens: Verbalize Videos To Understand Them In Zero Shot
Aanisha Bhattacharyya, Changyou Chen
ClassificationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Feed long videos with multi-modal features (key frame descriptions, OCR, ASR, metadata) into an LLM to generate natural language stories, then perform video understanding tasks (emotion, theme, persuasion strategies) on the generated stories.
Absolute Position Embedding Learns Sinusoid-like Waves for Attention Based on Relative Position
Yuji Yamamoto (Tokyo University of Science), Takuya Matsuzaki (Tokyo University of Science)
Explainability and InterpretabilityRepresentation LearningTransformerText
🎯 What it does: By conducting frequency domain and linear dimensionality reduction analysis on the learnable absolute position embeddings in the RoBERTa model, this study reveals the fundamental mechanism behind its self-attention focusing on neighboring words: the sine wave components learned in the position embeddings cause phase shifts between queries and keys, leading to attention focusing on words with relative proximity.
Abstractive Open Information Extraction
Kevin Pei (University of Illinois at Urbana Champaign), Kevin Chang (University of Illinois at Urbana Champaign)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the abstract OpenIE task, constructing training sets, models, and semantic evaluation metrics;
Accelerating Toeplitz Neural Network with Constant-time Inference Complexity
Zhen Qin (Shanghai Artificial Intelligence Laboratory), Yiran Zhong (Shanghai Artificial Intelligence Laboratory)
Computational EfficiencyText
🎯 What it does: Proposed a closed-form algorithm ETSC to convert Toeplitz Neural Network (TNN) into State Space Model (SSM), achieving constant-time complexity for TNN inference.
Accented Speech Recognition With Accent-specific Codebooks
Darshan Prabhu (Indian Institute of Technology Bombay), Vinit Unni (Indian Institute of Technology Bombay)
RecognitionDomain AdaptationTransformerAudio
🎯 What it does: Proposes a cross-attention based codebook method to achieve accent adaptation for end-to-end ASR models, supporting known accents during training and zero-shot transfer for unknown accents at test time.
ACQUIRED: A Dataset for Answering Counterfactual Questions In Real-Life Videos
Te-Lin Wu (University of California, Los Angeles), Nanyun Peng (University of Southern California)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodality
🎯 What it does: Constructed the ACQUIRED dataset for counterfactual question answering in real-world videos, covering three common-sense dimensions (physical, social, temporal), and supporting first/third-person perspectives;
Active Instruction Tuning: Improving Cross-Task Generalization by Training on Prompt Sensitive Tasks
Po-Nien Kung (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
Data-Centric LearningMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the Active Instruction Tuning framework, which leverages Prompt Uncertainty to select new tasks most beneficial for the model, thereby enhancing cross-task zero-shot generalization capabilities.
Active Learning for Natural Language Generation
Yotam Perlitz (IBM Research AI), Liat Ein-Dor (IBM Research AI)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Systematic active learning evaluation across multiple natural language generation tasks (rewriting, style transfer, summarization, and question answering generation), comparing two categories of sampling strategies: representativeness and uncertainty.
Active Retrieval Augmented Generation
Zhengbao Jiang (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
GenerationRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose an active retrieval-augmented generation framework (FLARE), which dynamically determines when and what information to retrieve during long-text generation.
ACTOR: Active Learning with Annotator-specific Classification Heads to Embrace Human Label Variation
Xinpeng Wang (LMU Munich), Barbara Plank (LMU Munich)
ClassificationData-Centric LearningTransformerText
🎯 What it does: Propose an active learning framework that uses a multi-head model (each head corresponds to an annotator) to simultaneously select samples and annotators, achieving joint modeling of annotator diversity and uncertainty;
AD-NLP: A Benchmark for Anomaly Detection in Natural Language Processing
Matei Bejan (University of Bucharest), Marius Popescu (University of Bucharest)
Anomaly DetectionTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes a unified text anomaly detection benchmark, AD-NLP, covering various anomaly types such as syntax, semantics, pragmatics, and style, and introduces three new diverse datasets: SongGenres, GutenbergCategories, and GutenbergAuthors; on this benchmark, classical one-class SVM and Isolation Forest, as well as deep models CVDD and DATE, are systematically evaluated to explore their performance and interpretability; meanwhile, the complete data and code are publicly released.
Adapt in Contexts: Retrieval-Augmented Domain Adaptation via In-Context Learning
Quanyu Long (Nanyang Technological University), Sinno Pan
ClassificationRecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a retrieval-enhanced unsupervised domain adaptation framework (DAICL), which retrieves similar target domain texts as context and jointly trains task learning with target domain language modeling, applicable to both encoder and decoder large language models.
Adapting Language Models to Compress Contexts
Alexis Chevalier (Princeton University), Danqi Chen (Princeton University)
CompressionComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose AutoCompressors, which transform pre-trained language models into systems capable of compressing long texts into short summary vectors that can be used as soft prompts;
Adapting Offline Speech Translation Models for Streaming with Future-Aware Distillation and Inference
Biao Fu (Xiamen University), Xiaodong Shi (Xiamen University)
Knowledge DistillationTransformerContrastive LearningAudio
🎯 What it does: This paper proposes the FAST method, addressing the input mismatch issue of offline ST models during streaming inference. It designs Future-Aware Inference (FAI) and Future-Aware Distillation (FAD), enhancing the quality and latency balance of streaming speech translation by adding masked future context during inference and performing knowledge distillation during training.
Adaptive End-to-End Metric Learning for Zero-Shot Cross-Domain Slot Filling
Yuanjun Shi (Tianjin University), Minglai Shao (Tianjin University)
Domain AdaptationRepresentation LearningRecurrent Neural NetworkTransformerContrastive LearningText
🎯 What it does: Proposed an end-to-end adaptive metric learning framework for zero-shot cross-domain slot filling;
Adaptive Gating in Mixture-of-Experts based Language Models
Jiamin Li (City University of Hong Kong), Hong Xu (City University of Hong Kong)
Computational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: Propose an Adaptive Gating Mixture-of-Experts (MoE) language model that dynamically determines whether to use 1 or 2 experts based on the expert probability distribution of each token, thereby improving training efficiency while maintaining inference quality.
Adaptive Policy with Wait-k Model for Simultaneous Translation
Libo Zhao (South China University of Technology), Zhongqiang Huang (Alibaba DAMO Academy)
TransformerReinforcement LearningText
🎯 What it does: This paper proposes an adaptive read/write strategy (DaP) based on the differences in statistical distributions, coupling it with a pre-trained multi-path wait-k translation model to construct a real-time translation system with adjustable latency.
AdaSent: Efficient Domain-Adapted Sentence Embeddings for Few-Shot Classification
Yongxin Huang (Technical University of Darmstadt), Iryna Gurevych (University of Würzburg)
ClassificationDomain AdaptationComputational EfficiencyRepresentation LearningTransformerContrastive LearningTextFinance Related
🎯 What it does: This paper proposes the AdaSent method, combining domain adaptive pretraining (DAPT) with sentence embedding pretraining (SEPT) to achieve efficient few-shot sentence classification.
Addressing Linguistic Bias through a Contrastive Analysis of Academic Writing in the NLP Domain
Robert Ridley (Nanjing University), Xinyu Dai (Nanjing University)
Contrastive LearningText
🎯 What it does: By conducting a comparative analysis of academic paper abstracts in the NLP field, this study explores differences in linguistic dimensions such as vocabulary, morphology, syntax, and coherence among authors with different native language backgrounds, and proposes suggestions to improve publication fairness.
Addressing NER Annotation Noises with Uncertainty-Guided Tree-Structured CRFs
Jian Liu (Beijing Jiaotong University), Zhe Zhao (Tencent)
RecognitionTransformerText
🎯 What it does: Propose a unified framework that treats named entity recognition as a tree-structured CRF parsing problem with uncertain nodes, using MC-Dropout to assess uncertainty and enhancing model robustness to noisy labels through iterative co-learning.
Advancements in Arabic Grammatical Error Detection and Correction: An Empirical Investigation
Bashar Alhafni (New York University Abu Dhabi), Nizar Habash (New York University Abu Dhabi)
TransformerLarge Language ModelText
🎯 What it does: Studied Arabic grammar error detection (GED) and correction (GEC), and first achieved multi-class GED in Arabic;
AfriSenti: A Twitter Sentiment Analysis Benchmark for African Languages
Shamsuddeen Hassan Muhammad (University of Porto), Stephen Arthur
ClassificationTransformerSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a Twitter sentiment annotation dataset named AfriSenti covering 14 African languages with over 110k entries.
Air-Decoding: Attribute Distribution Reconstruction for Decoding-Time Controllable Text Generation
Tianqi Zhong (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
GenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes the Air-Decoding framework, which utilizes prefix tuning to generate an attribute-conditional language model and employs attribute distribution reconstruction to avoid the Attribute Collapse problem during decoding, thereby enhancing the fluency and attribute accuracy of controllable text generation.
ALCAP: Alignment-Augmented Music Captioner
Zihao He (USC Information Sciences Institute), Xuchen Song (TikTok Inc)
GenerationRetrievalTransformerContrastive LearningTextMultimodalityAudio
🎯 What it does: Propose a music caption generation model ALCAP that achieves cross-modal alignment between music and lyrics through contrastive learning;
ALCUNA: Large Language Models Meet New Knowledge
Xunjian Yin (Peking University), Xiaojun Wan (Peking University)
Data SynthesisTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Designed and implemented the KnowGen method for generating new knowledge, and built the ALCUNA artificial biological entity benchmark based on it, to evaluate the understanding, discrimination, and reasoning capabilities of large language models when facing new knowledge.
ALDi: Quantifying the Arabic Level of Dialectness of Text
Amr Keleg (University of Edinburgh), Walid Magdy (University of Edinburgh)
RecognitionTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigate and propose a quantification metric ALDi for the 'dialectality' of Arabic dialect sentences, and construct the corresponding annotated dataset AOC-ALDi.
Aligning Large Language Models through Synthetic Feedback
Sungdong Kim (NAVER Cloud), Minjoon Seo (KAIST AI)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a novel framework that aligns large language models using synthetic feedback, completely without relying on human examples or proprietary LLMs.
All Things Considered: Detecting Partisan Events from News Media with Cross-Article Comparison
Yujian Liu (UC Santa Barbara), Lu Wang (University of Michigan)
ClassificationTransformerText
🎯 What it does: Developed a framework based on event selection, utilizing cross-article event comparisons to detect partisan events in news and simultaneously predict the ideology of articles.
AMR Parsing is Far from Solved: GrAPES, the Granular AMR Parsing Evaluation Suite
Jonas Groschwitz (University of Amsterdam), Meaghan Fowlie (Utrecht University)
TextBenchmark
🎯 What it does: Proposed the Granular AMR Parsing Evaluation Suite (GrAPES), providing 36 fine-grained evaluation categories and corresponding new metrics that cover diverse semantic phenomena such as lexical ambiguity, rare words, coreference, and structural generalization, enabling systematic assessment of specific capabilities of AMR parsers.
AMR Parsing with Causal Hierarchical Attention and Pointers
Chao Lou (ShanghaiTech University), Kewei Tu (ShanghaiTech University)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningGraph
🎯 What it does: Propose the CHAP AMR parser, which uses multi-layer target forms and pointers combined with causal hierarchical attention to explicitly model graph structures in the Transformer decoder.
An Attribution Method for Siamese Encoders
Lucas Moeller, Sebastian Padó (University of Stuttgart)
Explainability and InterpretabilityTransformerSupervised Fine-TuningText
🎯 What it does: Proposes a local attribution method for Siamese encoders (e.g., sentence Transformers), generalizing integrated gradients to dual-input models to obtain a feature pair attribution matrix, which can be reduced to a word pair attribution map.
An Empirical Study of Translation Hypothesis Ensembling with Large Language Models
António Farinhas (Instituto Superior Técnico), André F. T. Martins (Instituto Superior Técnico)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study explores how to improve translation quality and reduce hallucinations in machine translation based on large language models (LLMs) by integrating multiple translation hypotheses.
An Exploration of Left-Corner Transformations
Andreas Opedal (ETH Zürich), Tim Vieira (ETH Zürich)
OptimizationComputational EfficiencyText
🎯 What it does: This paper proposes a generic weighted left-corner transformation (GLCT), unifying traditional left-corner transformation with speculative transformation, providing fine-grained control over non-terminals and rules, and proving its equivalence to weighted context-free grammars (WCFG) and left-recursion elimination.
An Expression Tree Decoding Strategy for Mathematical Equation Generation
Wenqi Zhang (Zhejiang University), Weiming Lu (University of Shanghai for Science and Technology)
GenerationTransformerText
🎯 What it does: Proposed an expression tree decoding strategy that uses hierarchical parallel decoding to simultaneously generate multiple mathematical expressions and construct expression trees;
An Integrative Survey on Mental Health Conversational Agents to Bridge Computer Science and Medical Perspectives
Young Min Cho, Sharath Guntuku (University of Pennsylvania)
TransformerLarge Language ModelTextTabularBiomedical DataReview/Survey PaperRetrieval-Augmented Generation
🎯 What it does: Reviewed 534 papers on mental health chatbots from computer science and medicine, screened through the PRISMA framework to obtain 136 core papers, systematically organized their technical implementations, experimental designs, evaluation methods, and ethical considerations, and compared research differences between the two fields.
An Investigation of LLMs’ Inefficacy in Understanding Converse Relations
Chengwen Qi (Beihang University), Yuanjun Laili (Beihang University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkChain-of-Thought
🎯 What it does: To investigate whether large language models (LLMs) truly understand the semantics of formal languages, the authors constructed a new benchmark, ConvRe, focusing on dual (converse) relationships in binary relations. This benchmark includes 17 relations and 1,240 triplets, and proposes two multiple-choice question-answering tasks (Re2Text and Text2Re). The authors also detect whether models rely on 'shortcut learning' by varying test text and few-shot example text variants. Subsequently, zero-shot and few-shot reasoning experiments were conducted on three categories of LLMs (GPT-3/4, Claude, Flan-T5), evaluating the impact of different prompting methods, prompt lengths, and whether prompts include hints or chain-of-thought (CoT) reasoning.
An Iteratively Parallel Generation Method with the Pre-Filling Strategy for Document-level Event Extraction
Guanhua Huang (University of Science and Technology of China), Weinan E (Peking University)
TransformerTextFinance Related
🎯 What it does: Proposed an iterative parallel generation with pre-filling strategy for document-level event extraction method called IPGPF, avoiding trigger words and eliminating dependency on the order of role generation.
Analysing State-Backed Propaganda Websites: a New Dataset and Linguistic Study
Freddy Heppell (University of Sheffield), Carolina Scarton (University of Sheffield)
Data-Centric LearningTransformerTextMultimodalityBenchmark
🎯 What it does: Collected and analyzed the content of two Russian-backed multilingual fake news websites (RRN and WarOnFakes), completing article retrieval, topic clustering, language and time analysis, and article backtracking detection.
Analyzing Cognitive Plausibility of Subword Tokenization
Lisa Beinborn (Vrije Universiteit Amsterdam), Yuval Pinter (Ben Gurion University of Negev)
Explainability and InterpretabilityText
🎯 What it does: Propose a subword tokenization evaluation paradigm based on cognitive explainability, assessing tokenizers by leveraging the correlation between reaction times and accuracy in lexical decision tasks and the chunkability of subword splits;
Analyzing Film Adaptation through Narrative Alignment
Tanzir Pial (Stony Brook University), Steven Skiena (Stony Brook University)
Representation LearningTransformerContrastive LearningText
🎯 What it does: Proposes a book-to-movie script alignment method combining the Smith-Waterman local alignment algorithm with SBERT semantic embeddings, applying it to automatically analyze 40 book-to-film adaptations to reveal patterns in fidelity, dialogue importance, narrative order, and gender representation.
Analyzing Modular Approaches for Visual Question Decomposition
Apoorv Khandelwal (Brown University), Chen Sun (Brown University)
Explainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark
🎯 What it does: Analyzes and dissects the modular structure of ViperGPT, evaluates the sources of its performance, and compares problem-solving approaches between program generation and natural language prompting.
Analyzing Norm Violations in Live-Stream Chat
Jihyung Moon (SoftlyAI Research), Sungjoon Park (SoftlyAI Research)
ClassificationTransformerSupervised Fine-TuningText
🎯 What it does: Conducted toxicity and norm violation detection on Twitch live chat, constructing the first live chat toxicity detection dataset NormVio-RT and performing in-depth analysis.
Anaphor Assisted Document-Level Relation Extraction
Chonggang Lu (Beihang University), Yongyi Mao (University of Ottawa)
Graph Neural NetworkTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Propose the Anaphor-Assisted (AA) framework, which constructs a document graph using pronouns and coreference information, and achieves cross-sentence entity interaction through dynamic graph convolution;
Anchoring Fine-tuning of Sentence Transformer with Semantic Label Information for Efficient Truly Few-shot Classification
Amalie Pauli, Ira Assent (Aarhus University)
ClassificationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed a few-shot text classification method called AncSetFit, which utilizes anchor sentences from a sentence embedding model to semantically guide classes, achieving efficient training and inference with only 2-8 samples per class.
Answering Questions by Meta-Reasoning over Multiple Chains of Thought
Ori Yoran (Tel Aviv University), Jonathan Berant (Tel Aviv University)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the Multi-Chain Reasoning (MCR) method, enabling large language models to perform meta-reasoning across multiple chains of thought (CoT), thereby achieving higher accuracy simultaneously in both individual answers and explanations.
AnyTOD: A Programmable Task-Oriented Dialog System
Jeffrey Zhao (Google Research), Yonghui Wu (Google Research)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the ANYTOD system, an end-to-end task-oriented dialogue system that can achieve zero-shot task adaptation through programming.
API-Assisted Code Generation for Question Answering on Varied Table Structures
Yihan Cao (Carnegie Mellon University), Daniel Fried (Carnegie Mellon University)
AI Code AssistantTransformerLarge Language ModelPrompt EngineeringTabular
🎯 What it does: Propose a unified TableQA framework that converts various table structures into Pandas multi-index dataframes, utilizes Python code generation models (CODEX, STARCODER) to generate executable query programs under few-shot prompting, and supports custom API calls to expand functionality.
API-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs
Minghao Li (Alibaba Group), Yongbin Li (Alibaba Group)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextBenchmark
🎯 What it does: Proposed the API-Bank benchmark for tool-enhanced LLMs, and built an executable evaluation system, annotated data, and automatically generated training sets, training the Lynx model
APoLLo : Unified Adapter and Prompt Learning for Vision Language Models
Sanjoy Chowdhury (University of Maryland), Dinesh Manocha (University of Maryland)
ClassificationDomain AdaptationRepresentation LearningLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a unified adapter and prompt learning framework, APoLLo, to enhance the generalization performance of vision-language models in few-shot scenarios.
Appraising the Potential Uses and Harms of LLMs for Medical Systematic Reviews
Hye Yun, Byron Wallace
TransformerLarge Language ModelTextBiomedical Data
🎯 What it does: In this study, the authors explore the potential uses and risks of large language models (LLM) in writing medical systematic reviews by collecting interviews with 16 medical systematic review experts, and distill evaluation criteria from an expert perspective.
APrompt: Attention Prompt Tuning for Efficient Adaptation of Pre-trained Language Models
Qifan Wang (Meta AI), Dongfang Liu (Rochester Institute of Technology)
Domain AdaptationComputational EfficiencyTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes APROMPT, a parameter-efficient fine-tuning method that introduces query, key, and value prompts into the self-attention layer of Transformers.
Architectural Sweet Spots for Modeling Human Label Variation by the Example of Argument Quality: It’s Best to Relate Perspectives!
Philipp Heinisch (Bielefeld University), Philipp Cimiano (Bielefeld University)
ClassificationRecommendation SystemTransformerLarge Language ModelText
🎯 What it does: This paper studies how to directly utilize non-aggregated multi-annotator labels in subjective annotation tasks (especially argument quality assessment), proposes and compares a series of model architectures ranging from fully shared to fully personalized, and significantly improves single-annotator prediction performance through a recommendation system-inspired model.
Are All Steps Equally Important? Benchmarking Essentiality Detection in Event Processes
Haoyu Wang (University of Pennsylvania), Dan Roth (University of Southern California)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposed and implemented a benchmark task to evaluate a machine's understanding of the importance of internal steps in event processes—Essential Step Detection (ESD).
Are Compressed Language Models Less Subgroup Robust?
Leonidas Gee (University of Sussex), Novi Quadrianto (University of Sussex)
ClassificationCompressionKnowledge DistillationTransformerText
🎯 What it does: This paper systematically evaluates the impact of 18 model compression methods on the robustness of BERT subgroups, exploring performance changes of compressed models on minority subgroups;
Are Embedded Potatoes Still Vegetables? On the Limitations of WordNet Embeddings for Lexical Semantics
Xuyou Cheng (University of Cambridge), Guy Emerson (University of Cambridge)
Representation LearningGraph Neural NetworkTextBenchmark
🎯 What it does: This paper systematically evaluates the semantic capabilities of knowledge graph embedding (KBE) models trained on WordNet beyond link prediction, comparing the performance of various traditional KBE models (e.g., TransE, DistMult, MuRP, FuncE) and graph neural network (GNN)-based encoders (KBGAT, rGAT) on semantic tasks such as word similarity, word analogy, POS tagging, and NER, revealing that high link prediction performance does not equate to superior semantic representation.
Argue with Me Tersely: Towards Sentence-Level Counter-Argument Generation
Jiayu Lin (Fudan University), Zhongyu Wei (Fudan University)
GenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Constructed the ArgTersely benchmark and dataset for sentence-level refutation generation, and proposed the Arg-LlaMA framework and Arg-Judge evaluator based on instruction tuning and multi-step reasoning
Argument-based Detection and Classification of Fallacies in Political Debates
Pierpaolo Goffredo (Université Côte d'Azur), Elena Cabrio (Université Côte d'Azur)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Built and expanded the Fallacy corpus of U.S. presidential debates from 1960-2020, annotating argument components, relations, and six categories of fallacies.
ART: rule bAsed futuRe-inference deducTion
Mengze Li (Zhejiang University), Fei Wu (Shanghai Institute for Advanced Study of Zhejiang University)
Explainability and InterpretabilityGraph Neural NetworkTransformerVision-Language-Action ModelMultimodalityChain-of-Thought
🎯 What it does: Propose a rule-driven future reasoning inference task (ART), construct a large-scale multimodal dataset Video-ART, and design a baseline model ARTNet.
Assessing Step-by-Step Reasoning against Lexical Negation: A Case Study on Syllogism
Mengyu Ye (Tohoku University), Hiroaki Funayama (Tohoku University)
TransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper evaluates the performance of large language models (LLMs) in chain-of-thought (CoT) reasoning involving lexical negation through a series of controlled experiments (BASE, FIC, FICNEG, FICNEG-O).
Assessing the influence of attractor-verb distance on grammatical agreement in humans and language models
Christos Zacharopoulos, Mathias Sablé-Meyer (Cognitive Neuroimaging Unit, NeuroSpin center, France)
TransformerLarge Language ModelText
🎯 What it does: This paper investigates the impact of the distance between attractor words and verbs on syntax consistency judgments in humans and language models.
ATFormer: A Learned Performance Model with Transfer Learning Across Devices for Deep Learning Tensor Programs
Yang Bai (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
OptimizationComputational EfficiencyHyperparameter SearchTransformerSupervised Fine-TuningBenchmark
🎯 What it does: Propose ATFormer, an attention-based cost model that rapidly and accurately predicts tensor operation performance on different hardware, and accelerates the search process through transfer learning.