arXivSub Start free trial

EMNLP 2023 Papers with Code

Conference on Empirical Methods in Natural Language Processing · 380 papers with a public code repository

‘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)

CodeGenerationData 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;

“Mistakes Help Us Grow”: Facilitating and Evaluating Growth Mindset Supportive Language in Classrooms

Kunal Handa (Brown University), Dorottya Demszky (Stanford University)

CodeTransformerLarge 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.

A Benchmark for Reasoning with Spatial Prepositions

Iulia Comsa, Srini Narayanan (Google DeepMind)

CodeTransformerLarge 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)

CodeGenerationTransformerLarge 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 Diachronic Analysis of Paradigm Shifts in NLP Research: When, How, and Why?

Aniket Pramanick (Ubiquitous Knowledge Processing Lab), Iryna Gurevych (Ubiquitous Knowledge Processing Lab)

CodeText

🎯 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 Diffusion Weighted Graph Framework for New Intent Discovery

Wenkai Shi, Ping Chen (Lenovo)

CodeClassificationGraph 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 Fair and In-Depth Evaluation of Existing End-to-End Entity Linking Systems

Hannah Bast (University of Freiburg), Natalie Prange (University of Freiburg)

CodeRetrievalTransformerSupervised 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 Generation-based Deductive Method for Math Word Problems

Yuxuan Hu (Renmin University of China), Hong Chen (Renmin University of China)

CodeGenerationRecurrent 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

CodeClassificationComputational 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 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)

CodeClassificationRecognitionTransformerLarge 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 Predictive Factor Analysis of Social Biases and Task-Performance in Pretrained Masked Language Models

Yi Zhou (Cardiff University), Danushka Bollegala (University of Liverpool)

CodeExplainability 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)

CodeGenerationRetrievalTransformerLarge 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 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)

CodeData-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)

CodeData-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 Simple Baseline for Knowledge-Based Visual Question Answering

Alexandros Xenos (Queen Mary University of London), Georgios Tzimiropoulos (Queen Mary University of London)

CodeKnowledge 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 Study on Accessing Linguistic Information in Pre-Trained Language Models by Using Prompts

Marion Di Marco (LMU Munich), Alexander Fraser (LMU Munich)

CodeRepresentation 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 Unified View of Evaluation Metrics for Structured Prediction

Yunmo Chen (Johns Hopkins University), Benjamin Van Durme (Johns Hopkins University)

CodeOptimizationText

🎯 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

CodeClassificationRetrievalTransformerLarge 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.

Accelerating Toeplitz Neural Network with Constant-time Inference Complexity

Zhen Qin (Shanghai Artificial Intelligence Laboratory), Yiran Zhong (Shanghai Artificial Intelligence Laboratory)

CodeComputational 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)

CodeRecognitionDomain 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.

AD-NLP: A Benchmark for Anomaly Detection in Natural Language Processing

Matei Bejan (University of Bucharest), Marius Popescu (University of Bucharest)

CodeAnomaly 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.

Adapting Language Models to Compress Contexts

Alexis Chevalier (Princeton University), Danqi Chen (Princeton University)

CodeCompressionComputational 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;

AdaSent: Efficient Domain-Adapted Sentence Embeddings for Few-Shot Classification

Yongxin Huang (Technical University of Darmstadt), Iryna Gurevych (University of Würzburg)

CodeClassificationDomain 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.

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)

CodeTransformerLarge Language ModelText

🎯 What it does: Studied Arabic grammar error detection (GED) and correction (GEC), and first achieved multi-class GED in Arabic;

ALCUNA: Large Language Models Meet New Knowledge

Xunjian Yin (Peking University), Xiaojun Wan (Peking University)

CodeData 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.

All Things Considered: Detecting Partisan Events from News Media with Cross-Article Comparison

Yujian Liu (UC Santa Barbara), Lu Wang (University of Michigan)

CodeClassificationTransformerText

🎯 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 with Causal Hierarchical Attention and Pointers

Chao Lou (ShanghaiTech University), Kewei Tu (ShanghaiTech University)

CodeRepresentation 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)

CodeExplainability 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 Integrative Survey on Mental Health Conversational Agents to Bridge Computer Science and Medical Perspectives

Young Min Cho, Sharath Guntuku (University of Pennsylvania)

CodeTransformerLarge 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)

CodeExplainability 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.

Analysing State-Backed Propaganda Websites: a New Dataset and Linguistic Study

Freddy Heppell (University of Sheffield), Carolina Scarton (University of Sheffield)

CodeData-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)

CodeExplainability 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 Modular Approaches for Visual Question Decomposition

Apoorv Khandelwal (Brown University), Chen Sun (Brown University)

CodeExplainability 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)

CodeClassificationTransformerSupervised 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.

Anchoring Fine-tuning of Sentence Transformer with Semantic Label Information for Efficient Truly Few-shot Classification

Amalie Pauli, Ira Assent (Aarhus University)

CodeClassificationTransformerSupervised 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)

CodeExplainability 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)

CodeTransformerLarge 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-Bank: A Comprehensive Benchmark for Tool-Augmented LLMs

Minghao Li (Alibaba Group), Yongbin Li (Alibaba Group)

CodeTransformerLarge 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

Augmenting Zero-Shot Dense Retrievers with Plug-in Mixture-of-Memories

Suyu Ge (University of Illinois Urbana-Champaign), Paul Bennett (Spotify)

CodeRetrievalTransformerContrastive LearningTextBiomedical DataRetrieval-Augmented Generation

🎯 What it does: This paper proposes a retrieval-enhanced framework called MoMA based on multi-source memory to improve the generalization of zero-shot dense retrieval.

BasahaCorpus: An Expanded Linguistic Resource for Readability Assessment in Central Philippine Languages

Joseph Marvin Imperial (University of Bath), Ekaterina Kochmar (MBZUAI)

CodeClassificationText

🎯 What it does: Collected and released BASAHACORPUS, a corpus containing children's short stories in four Central Philippine languages (Hiligaynon, Minasbate, Karaya, and Rinconada) from Let’s Read Asia, and built a reading difficulty assessment model based on this corpus.

Beat LLMs at Their Own Game: Zero-Shot LLM-Generated Text Detection via Querying ChatGPT

Biru Zhu (Tsinghua University), Ming Gu (Tsinghua University)

CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringText

🎯 What it does: Propose a zero-shot black-box method that discriminates LLM-generated text by having ChatGPT revise the text to be examined and comparing the similarity between the original and revised versions.

Benchmarking and Improving Text-to-SQL Generation under Ambiguity

Adithya Bhaskar (Princeton University), Sunita Sarawagi (IIT Bombay)

CodeGenerationTransformerLarge Language ModelTextTabularBenchmark

🎯 What it does: This paper proposes a new evaluation benchmark, AmbiQT, specifically designed for ambiguous natural language queries in real databases, and develops a new decoding algorithm called LogicalBeam to improve the coverage of text-to-SQL generation models on ambiguous queries.

BERTie Bott’s Every Flavor Labels: A Tasty Introduction to Semantic Role Labeling for Galician

Micaella Bruton (Uppsala University), Meriem Beloucif (Uppsala University)

CodeData-Centric LearningTransformerSupervised Fine-TuningText

🎯 What it does: Constructed the first Galician semantic role labeling (SRL) dataset and proposed the Verbal Indexing preprocessing method;

Best of Both Worlds: Towards Improving Temporal Knowledge Base Question Answering via Targeted Fact Extraction

Nithish Kannen (Amazon Alexa AI, UK), L Subramaniam

CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation

🎯 What it does: Propose a temporal question answering system that integrates knowledge bases (KB) with textual resources, utilizing language models for targeted fact extraction to address failures caused by missing information in KBs;

Better Quality Pre-training Data and T5 Models for African Languages

Akintunde Oladipo (University Of Waterloo), Jimmy Lin (University Of Waterloo)

CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelText

🎯 What it does: This paper constructs a new high-quality pretraining corpus named WURA, covering 16 African languages and 4 high-resource languages, by auditing and improving the quality of existing multilingual pretraining corpora (especially mC4). Based on this, the T5 model AfriTeVa V2 was trained.

Beware of Model Collapse! Fast and Stable Test-time Adaptation for Robust Question Answering

Yi Su (Soochow University), Min Zhang (Soochow University)

CodeDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark

🎯 What it does: This paper proposes a test-time adaptation method called Anti-CF, aiming to enhance model robustness under distribution drift by preventing model collapse and accelerating inference in QA tasks.

Beyond Factuality: A Comprehensive Evaluation of Large Language Models as Knowledge Generators

Liang Chen (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)

CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: Proposes a comprehensive evaluation framework named CONNER for automatically and reference-free assessing the intrinsic quality of knowledge generated by large language models (LLMs) and their extrinsic impact on downstream tasks;

BioFEG: Generate Latent Features for Biomedical Entity Linking

Xuhui Sui (Nankai University), Wensheng Zhang (Nankai University)

CodeRetrievalTransformerLarge Language ModelGenerative Adversarial NetworkBiomedical DataBenchmark

🎯 What it does: Propose the BioFEG framework, which enhances biomedical entity linking by generating potential features for unseen entities

Biomedical Named Entity Recognition via Dictionary-based Synonym Generalization

Zihao Fu (University of Cambridge), Nigel Collier (University of Cambridge)

CodeRecognitionTransformerBiomedical Data

🎯 What it does: This paper proposes SynGen, a dictionary-based biomedical named entity recognition framework, which trains the model using synonyms present in the dictionary and achieves generalized recognition of out-of-dictionary synonyms through positive and negative sample learning.

BioPlanner: Automatic Evaluation of LLMs on Protocol Planning in Biology

Odhran O’Donoghue (Align to Innovate), Samuel Rodriques (Francis Crick Institute)

CodeTransformerLarge Language ModelTextBiomedical DataChain-of-Thought

🎯 What it does: This paper proposes an automatic evaluation framework based on executable pseudocode to measure the ability of large language models in planning biological experiment protocols, and constructs and verifies the BIOPROT dataset on this basis.

BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations

Qizhi Pei (Renmin University of China), Rui Yan (Renmin University of China)

CodeDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringMultimodalityBiomedical Data

🎯 What it does: Propose BioT5, a cross-modal pre-training framework that leverages chemical knowledge and natural language associations for joint representation and generation of molecules, proteins, and text.

BLESS: Benchmarking Large Language Models on Sentence Simplification

Tannon Kew (University of Zurich), Matthew Shardlow (Manchester Metropolitan University)

CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Construct the BLESS benchmark to evaluate the performance of 44 large language models on sentence simplification tasks using few-shot context learning, and conduct a systematic analysis of automatic evaluation metrics, edit operations, and manual quality assessment.

BRAINTEASER: Lateral Thinking Puzzles for Large Language Models

Yifan Jiang (University of Southern California), Zhivar Sourati (University of Southern California)

CodeTransformerLarge Language ModelTextBenchmark

🎯 What it does: Constructed and released the BRAINTEASER benchmark, which converts lateral thinking puzzles into sentence-level and vocabulary-level multiple-choice questions, and tests the reasoning consistency of models through semantic reconstruction and context reconstruction.

Byte Pair Encoding for Symbolic Music

Nathan Fradet (Sorbonne University), Jean-Pierre Briot (Sorbonne University)

CodeClassificationGenerationTransformerLarge Language ModelSequential

🎯 What it does: In symbolic music generation and classification tasks, this paper applies Byte Pair Encoding (BPE) to existing tokenization methods such as REMI and TSD, significantly shortening sequence lengths, expanding the vocabulary, and conducting experiments using Transformer models.

Cabbage Sweeter than Cake? Analysing the Potential of Large Language Models for Learning Conceptual Spaces

Usashi Chatterjee (Cardiff University), Steven Schockaert (Cardiff University)

CodeRepresentation LearningTransformerLarge Language ModelText

🎯 What it does: The study utilizes large language models to learn concept spaces, focusing on taste and physical attributes, to investigate whether they can capture perceptual dimensions.

Calc-X and Calcformers: Empowering Arithmetical Chain-of-Thought through Interaction with Symbolic Systems

Marek Kadlčík (Masaryk University), Vlastimil Martinek (Masaryk University)

CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought

🎯 What it does: Constructed a unified arithmetic chain-of-thought dataset called Calc-X, and trained a model named Calcformers capable of invoking external calculators via HTML-style gadget tags.

Can Language Models Laugh at YouTube Short-form Videos?

Dayoon Ko (Seoul National University), Gunhee Kim (Seoul National University)

CodeData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityBenchmark

🎯 What it does: Constructed the ExFunTube dataset, collecting and annotating 10,136 user-generated 30-second short videos, with annotated punchline timestamps and textual explanations; and designed a zero-shot video-to-text prompting method, converting visual, audio, and sound information into fine-grained text, which is then input into large language models (LLMs) for humorous explanations.

Can language models learn analogical reasoning? Investigating training objectives and comparisons to human performance

Molly Petersen, Lonneke van der Plas (University of Malta)

CodeClassificationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningText

🎯 What it does: Investigating whether language models can learn analogical reasoning through specialized training objectives and comparing their performance with humans

Can Large Language Models Capture Dissenting Human Voices?

Noah Lee (KAIST AI), James Thorne (KAIST AI)

CodeExplainability and InterpretabilityTransformerLarge Language ModelText

🎯 What it does: Evaluate the reasoning ability of instruction-based large language models in natural language inference tasks and their consistency with human opinion distributions, conducting experiments using two distribution estimation methods (MCE and LPE).

Can LMs Generalize to Future Data? An Empirical Analysis on Text Summarization

Chi Seng Cheang (University of Macau), Lidia S. Chao (University of Macau)

CodeTransformerLarge Language ModelContrastive LearningTextBenchmark

🎯 What it does: Proposed the TEMPOSUM benchmark to evaluate the abstract summarization capability of pre-trained language models on news texts from future time periods, and refined hallucination types through human evaluation.

Can We Edit Factual Knowledge by In-Context Learning?

Ce Zheng (Peking University), Baobao Chang (Peking University)

CodeTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation

🎯 What it does: Proposes a knowledge editing method IKE based on context learning, which modifies the factual knowledge of large models through demonstrations without updating model parameters.

CAPSTONE: Curriculum Sampling for Dense Retrieval with Document Expansion

Xingwei He (University of Hong Kong), Nan Duan (Microsoft Research Asia)

CodeRetrievalTransformerContrastive LearningTextBenchmark

🎯 What it does: Proposes a document expansion strategy (CAPSTONE) using curriculum learning to improve query-aware document representations in dual-cross-encoders for dense retrieval;

ChatGPT to Replace Crowdsourcing of Paraphrases for Intent Classification: Higher Diversity and Comparable Model Robustness

Jan Cegin (Brno University of Technology), Peter Brusilovsky (University of Pittsburgh)

CodeClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark

🎯 What it does: Compare the effectiveness, diversity, and model robustness of paraphrases generated by ChatGPT and traditional crowdsourcing in intent classification data, using the same experimental process to verify whether ChatGPT can replace human workers.

CHEF in the Language Kitchen: A Generative Data Augmentation Leveraging Korean Morpheme Ingredients

Jaehyung Seo (Korea University), Heuiseok Lim (Korea University)

CodeClassificationData SynthesisTransformerContrastive LearningTextBenchmark

🎯 What it does: Propose CHEF, a generative data augmentation method based on Korean morphemes, which synthesizes new sentences using a morpheme mixer and a label discriminator while maintaining label consistency.

Chinese Lexical Substitution: Dataset and Method

Jipeng Qiang (Yangzhou University), Xiaoye Ouyang (China Academy of Electronic and Information Technology)

CodeData SynthesisRepresentation LearningTransformerLarge Language ModelMixture of ExpertsTextBenchmark

🎯 What it does: Constructed the Chinese Lexical Substitution (CHNLS) dataset using a human-machine collaborative annotation approach.

CiteBench: A Benchmark for Scientific Citation Text Generation

Martin Funkquist (Linköping University), Iryna Gurevych (Technical University of Darmstadt)

CodeGenerationTransformerSupervised Fine-TuningTextBenchmark

🎯 What it does: Constructed the CITEBENCH benchmark, unifying four distinct reference text generation tasks and providing baselines and evaluation tools.

CleanCoNLL: A Nearly Noise-Free Named Entity Recognition Dataset

Susanna Rücker (Humboldt University of Berlin), Alan Akbik (Humboldt University of Berlin)

CodeRecognitionTextBenchmark

🎯 What it does: By conducting a comprehensive re-annotation of CoNLL-03 with the addition of an entity linking layer and consistency checks, the CLEANCONLL dataset was generated, nearly free of noise.

Clustering Pseudo Language Family in Multilingual Translation Models with Fisher Information Matrix

Xinyu Ma (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)

CodeTransformerText

🎯 What it does: Propose a pseudo-language family clustering method based on the Fisher Information Matrix (FIM) to select auxiliary language pairs in multilingual neural machine translation;

Coarse-to-Fine Contrastive Learning in Image-Text-Graph Space for Improved Vision-Language Compositionality

Harman Singh (Meta AI), Yu Chen (Anytime.AI)

CodeRetrievalRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningImageTextMultimodalityGraph

🎯 What it does: This paper designs a coarse-to-fine contrastive learning framework based on scene graphs, which significantly enhances the compositional reasoning ability of vision-language models by decomposing text-parsed scene graphs into multi-level subgraphs and jointly training images with multi-text contrast using hard negative subgraphs generated through graph augmentation.

CoCo: Coherence-Enhanced Machine-Generated Text Detection Under Low Resource With Contrastive Learning

Xiaoming Liu (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)

CodeClassificationGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText

🎯 What it does: Proposed the COCO model, which detects differences between machine-generated text and human-generated text using a consistency-enhanced contrastive learning approach.

CodeT5+: Open Code Large Language Models for Code Understanding and Generation

Yue Wang (Salesforce AI Research), Steven Hoi

CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextRetrieval-Augmented Generation

🎯 What it does: Proposed CodeT5+, which can dynamically switch between encoder, decoder, or encoder-decoder modes, and improved code understanding and generation performance through mixed pre-training objectives.

CoF-CoT: Enhancing Large Language Models with Coarse-to-Fine Chain-of-Thought Prompting for Multi-domain NLU Tasks

Hoang Nguyen (University of Illinois at Chicago), Philip Yu (University of Illinois at Chicago)

CodeTransformerLarge Language ModelTextChain-of-Thought

🎯 What it does: Propose a coarse-to-fine hierarchical chained reasoning (CoF-CoT) framework that decomposes NLU tasks into a process from coarse to fine using multi-step reasoning sequences;

Cognate Transformer for Automated Phonological Reconstruction and Cognate Reflex Prediction

V.S.D.S.Mahesh Akavarapu (Indian Institute of Technology Kanpur), Arnab Bhattacharya (Indian Institute of Technology Kanpur)

CodeGenerationTransformerText

🎯 What it does: Proposed and implemented the Cognate Transformer model for speech reconstruction, which can automatically generate phoneme sequences in two tasks: parent language reconstruction and cognitive reflection prediction.

COHESENTIA: A Novel Benchmark of Incremental versus Holistic Assessment of Coherence in Generated Texts

Aviya Maimon (Bar Ilan University), Reut Tsarfaty (Bar Ilan University)

CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark

🎯 What it does: Constructed the COHESENTIA benchmark and proposed two annotation protocols (global and incremental) for evaluating the coherence of stories generated by GPT-3.

CoLT5: Faster Long-Range Transformers with Conditional Computation

Joshua Ainslie (Google), Sumit Sanghai (Google)

CodeComputational EfficiencyRepresentation LearningTransformerMixture of ExpertsTextBenchmark

🎯 What it does: Propose COLT5, a long-text Transformer accelerated by conditional computation, capable of efficient inference and training on long sequences.

Combining Denoising Autoencoders with Contrastive Learning to fine-tune Transformer Models

Alejo Lopez-Avila, Víctor Suárez-Paniagua (Huawei Ireland Research Center)

CodeRepresentation LearningTransformerSupervised Fine-TuningAuto EncoderContrastive LearningText

🎯 What it does: Propose a three-stage fine-tuning method: first adapt data distribution using a denoising autoencoder, then cluster representations and correct class imbalance via supervised contrastive learning, and finally add a classification head for final training during the fine-tuning stage.

Composable Text Controls in Latent Space with ODEs

Guangyi Liu (CUHK-Shenzhen), Zhiting Hu (UC San Diego)

CodeGenerationTransformerGenerative Adversarial NetworkTextOrdinary Differential Equation

🎯 What it does: Proposes LATENTOPS, which enables composable text control operations in the text latent space and achieves efficient text generation or editing through ODE sampling.

CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations

Myra Cheng (Stanford University), Diyi Yang (Stanford University)

CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText

🎯 What it does: Proposed the CoMPosT framework to systematically describe and evaluate four dimensions of LLM simulations (Context, Model, Persona, Topic), and designed a characterization method based on persona and topic semantic axes to detect 'caricaturization' tendencies in LLM simulations.

Content- and Topology-Aware Representation Learning for Scientific Multi-Literature

Kai Zhang (Worcester Polytechnic Institute), Xiaozhong Liu (Worcester Polytechnic Institute)

CodeRepresentation LearningGraph Neural NetworkTransformerAuto EncoderContrastive LearningTextMultimodalityGraphBiomedical DataBenchmark

🎯 What it does: In a multi-document learning scenario, the SMRC 2 model is proposed, which generates document-level and global multi-document representations by simultaneously learning the semantics (entity graph and abstract text) and topological structure (citation network) of the documents.

Context Compression for Auto-regressive Transformers with Sentinel Tokens

Siyu Ren (Shanghai Jiao Tong University), Kenny Zhu (University of Texas at Arlington)

CodeCompressionComputational EfficiencyTransformerText

🎯 What it does: Proposed a context compression method based on sentinel tokens, which inserts <CL> and <CR> tokens into autoregressive Transformers and modifies the attention mask to compress contiguous token ranges into compact representations, significantly reducing key-value (KV) cache usage and computational complexity.

Contextual Interaction for Argument Post Quality Assessment

Yiran Wang (University of Chinese Academy of Sciences), Le Sun (Chinese Academy of Sciences)

CodeClassificationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningText

🎯 What it does: Proposed and compared two argument quality assessment methods—supervised contrastive learning and example-based LLM prompting, emphasizing the role of contextual interaction between arguments in distinguishing subtle differences.

Continual Event Extraction with Semantic Confusion Rectification

Zitao Wang (Nanjing University), Wei Hu (Nanjing University)

CodeKnowledge DistillationRepresentation LearningTransformerTextBenchmark

🎯 What it does: Proposed a novel continual event extraction model that can avoid forgetting and correct semantic confusion while new event types continuously emerge.

Continually Improving Extractive QA via Human Feedback

Ge Gao (Cornell University), Eunsol Choi (University of Texas at Austin)

CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: This paper studies leveraging user feedback during continuous interaction to continuously improve the performance of extractive question answering systems, proposing a framework of iterative deployment and offline contextual reinforcement learning;

Contrastive Learning for Inference in Dialogue

Etsuko Ishii (Hong Kong University of Science and Technology), Pascale Fung (Hong Kong University of Science and Technology)

CodeTransformerLarge Language ModelContrastive LearningText

🎯 What it does: This paper studies the information gap issue in dialogue reasoning and improves the model's performance in reasoning tasks through contrastive learning methods.

Contrastive Learning of Sentence Embeddings from Scratch

Junlei Zhang (Zhejiang University), Junxian He (Hong Kong University of Science and Technology)

CodeData SynthesisRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText

🎯 What it does: Leverage large language models (e.g., ChatGPT) to generate synthetic positive and negative sentence pairs, then train sentence embeddings under a contrastive learning framework.

Controllable Contrastive Generation for Multilingual Biomedical Entity Linking

Tiantian Zhu (Harbin Institute Of Technology), Yang Xiang (Peng Cheng Laboratory)

CodeTransformerPrompt EngineeringContrastive LearningBiomedical Data

🎯 What it does: Propose the Con2GEN framework, which employs a prompt-controlled contrastive generation method to address ambiguity and information missing issues in multilingual biomedical entity linking (MBEL).

Controlling Pre-trained Language Models for Grade-Specific Text Simplification

Sweta Agrawal (University of Maryland), Marine Carpuat (University of Maryland)

CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper experimentally evaluates the impact of using low-level control markers on output simplification degree and quality in text simplification tasks, and proposes an instance-level control prediction method that predicts control markers based on input text and target reading level to improve text simplification effectiveness.

Conversational Semantic Parsing using Dynamic Context Graphs

Parag Jain (University of Edinburgh), Mirella Lapata (University of Edinburgh)

CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelTextGraph

🎯 What it does: Propose a semantic parsing model for conversational knowledge graph question answering, which can map user utterances into executable SPARQL queries within the context of conversation history.

CORE: A Few-Shot Company Relation Classification Dataset for Robust Domain Adaptation.

Philipp Borchert (Lille Economie Management), Marie-Francine Moens (University of Lille)

CodeClassificationDomain AdaptationMeta LearningTransformerPrompt EngineeringTextBenchmark

🎯 What it does: Constructed and evaluated the CORE dataset for few-shot company relationship classification and cross-domain adaptation.

CoRec: An Easy Approach for Coordination Recognition

Qing Wang (Iowa State University), Qi Li (Iowa State University)

CodeRecognitionTransformerLarge Language ModelText

🎯 What it does: Proposes CoRec—a pipeline model for coordinating structure recognition that does not rely on a syntactic parser, divided into two steps: identifying coordinating words and detecting parallel clause boundaries;

CorefPrompt: Prompt-based Event Coreference Resolution by Measuring Event Type and Argument Compatibilities

Sheng Xu (Soochow University), Qiaoming Zhu (Soochow University)

CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText

🎯 What it does: In this paper, the authors propose a Prompt-based event coreference resolution framework called CorefPrompt, which transforms event coreference determination into a masked language model (MLM) task, completing event modeling and coreference judgment within the same template; simultaneously, two auxiliary Prompt tasks, event type compatibility and argument compatibility, are introduced to explicitly demonstrate the reasoning process; finally, a mask token update mechanism is utilized to enhance the model's interactive expression.

CoSyn: Detecting Implicit Hate Speech in Online Conversations Using a Context Synergized Hyperbolic Network

Sreyan Ghosh (University of Maryland College Park), Dinesh Manocha (University of Maryland College Park)

CodeClassificationGraph Neural NetworkText

🎯 What it does: This paper proposes the CoSyn framework, aiming to detect implicit hate speech in online conversations by combining user history, social context, and conversation context within a hyperbolic space.

CP-BCS: Binary Code Summarization Guided by Control Flow Graph and Pseudo Code

Tong Ye (Zhejiang University), Wenhai Wang (Zhejiang University)

CodeGenerationAI Code AssistantGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityGraph

🎯 What it does: The study focuses on generating readable summaries for de-symbolized binary functions, addressing the high reverse engineering difficulty caused by the lack of symbolic information in binary functions.

CQE: A Comprehensive Quantity Extractor

Satya Almasian (Heidelberg University), Michael Gertz (Heidelberg University)

CodeTransformerLarge Language ModelTextBenchmark

🎯 What it does: Developed a comprehensive quantity extraction framework (CQE) capable of identifying and standardizing numerical values, units, change trends, and associated concepts, providing a unified normalized representation.

CRAB: Assessing the Strength of Causal Relationships Between Real-world Events

Angelika Romanou (École Polytechnique Fédérale de Lausanne), Antoine Bosselut (École Polytechnique Fédérale de Lausanne)

CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark

🎯 What it does: Constructed the CRAB benchmark, containing approximately 2.7K fine-grained causal relationship annotations of real-world events, and used it to evaluate the causal reasoning ability of large language models.

CRaSh: Clustering, Removing, and Sharing Enhance Fine-tuning without Full Large Language Model

Kaiyan Zhang (Tsinghua University), Bowen Zhou (Tsinghua University)

CodeComputational EfficiencyKnowledge DistillationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: Explored a no-training-step method called CRaSh, which extracts efficient submodels (emulators) from large language models (LLMs) by leveraging layer clustering, elimination, and sharing techniques, and achieves privacy-friendly fine-tuning of LLMs through Offsite-Tuning (OFT).

Cross-Lingual Cross-Target Stance Detection with Dual Knowledge Distillation Framework

Ruike Zhang (Institute of Automation, Chinese Academy of Sciences), Wenji Mao (Institute of Automation, Chinese Academy of Sciences)

CodeClassificationKnowledge DistillationGraph Neural NetworkTransformerPrompt EngineeringContrastive LearningText

🎯 What it does: Propose a cross-lingual and cross-target stance detection task, and develop a dual-teacher distillation framework (CCSD)

Cross-Modal Conceptualization in Bottleneck Models

Danis Alukaev (Innopolis University), Ivan Titov (University of Edinburgh)

CodeExplainability and InterpretabilityRepresentation LearningConvolutional Neural NetworkVision Language ModelContrastive LearningImageTextMultimodalityBiomedical Data

🎯 What it does: This study proposes a cross-modal conceptual bottleneck model (XCB) that automatically induces interpretable concepts by leveraging text descriptions corresponding to training images, thereby eliminating the need for manually defining and annotating concepts.

CRoW: Benchmarking Commonsense Reasoning in Real-World Tasks

Mete Ismayilzada (EPFL), Antoine Bosselut

CodePrompt EngineeringTextBenchmarkChain-of-Thought

🎯 What it does: Constructed a multi-task Commonsense Reasoning Benchmark (CROW), generating Winograd-style commonsense violation samples across six real-world NLP tasks through a manually designed multi-stage data collection pipeline, and evaluated models' reasoning capabilities on these tasks.

CRT-QA: A Dataset of Complex Reasoning Question Answering over Tabular Data

Zhehao Zhang (Darmouth College), Jian-Guang Lou (Microsoft Research Asia)

CodeData SynthesisLarge Language ModelAgentic AIPrompt EngineeringTabularBenchmarkChain-of-Thought

🎯 What it does: Constructed the CRT-QA dataset, focusing on multi-step reasoning and informal reasoning with table data, and provided fine-grained annotations (question directness, sub-question combination types, human reasoning paths) as well as unanswerable/uncertain questions.