EMNLP 2023 Papers — Page 9
Conference on Empirical Methods in Natural Language Processing · 1047 papers
Retrofitting Light-weight Language Models for Emotions using Supervised Contrastive Learning
Sapan Shah (Tata Consultancy Services), Pushpak Bhattacharyya (Indian Institute of Technology Bombay)
RecognitionRepresentation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes a post-hoc method that utilizes supervised contrastive learning to reconstruct BERT and RoBERTa, generating sentiment-aware sentence representations;
Revisiting Automated Topic Model Evaluation with Large Language Models
Dominik Stammbach (ETH Zürich), Elliott Ash (ETH Zürich)
TransformerLarge Language ModelText
🎯 What it does: This paper evaluates the output quality of topic models using large language models (LLM) and explores the feasibility of LLM in automatically determining the number of topics;
Revisiting Block-based Quantisation: What is Important for Sub-8-bit LLM Inference?
Cheng Zhang (Imperial College London), Yiren Zhao (Imperial College London)
Computational EfficiencyHyperparameter SearchTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: To address low-precision inference for large language models, this paper proposes a block-level quantization scheme that utilizes shared exponents to reduce scale shift, achieving near-lossless inference at 6-bit and 4-bit precisions.
Revisiting De-Identification of Electronic Medical Records: Evaluation of Within- and Cross-Hospital Generalization
Yiyang Liu (Ningbo No.2 Hospital), Enwei Zhu
Domain AdaptationSafty and PrivacyConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelTextBiomedical DataElectronic Health RecordsBenchmark
🎯 What it does: Constructed a Chinese de-identified electronic medical record dataset from three hospitals, and evaluated model generalization in both same-hospital and cross-hospital scenarios.
Revisiting Instruction Fine-tuned Model Evaluation to Guide Industrial Applications
Manuel Faysse (Illuin Technology), Pierre Colombo (Illuin Technology)
Data SynthesisTransformerLarge Language ModelText
🎯 What it does: Investigated evaluation methods for Instruction Fine-Tuned (IFT) models and explored model specialization strategies for industrial application scenarios;
Revisiting Machine Translation for Cross-lingual Classification
Mikel Artetxe (Reka AI), Luke Zettlemoyer (Meta AI)
ClassificationDomain AdaptationTransformerSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Reevaluate the role of machine translation in cross-lingual classification, systematically compare translate-test, translate-train, and zero-shot methods, and propose MT adaptation and training data adaptation techniques.
Revisiting Source Context in Nearest Neighbor Machine Translation
Xuanhong Li (Central China Normal University), Po Hu (Central China Normal University)
GenerationDomain AdaptationTransformerTextRetrieval-Augmented Generation
🎯 What it does: This paper re-examines the role of source context in recent nearest neighbor machine translation (kNN-MT) and proposes a comprehensive method to enhance translation quality by introducing source context in three core steps: retrieval, calibration, and interpolation.
Revisiting Sparse Retrieval for Few-shot Entity Linking
Yulin Chen (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
RetrievalMeta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a few-shot entity linking framework based on sparse retrieval and keyword extraction, enhancing BM25 queries with keywords to improve retrieval effectiveness;
Revisiting the Knowledge Injection Frameworks
Peng Fu (Zhejiang University), Junbo Zhao (Zhejiang University)
Representation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextGraphReview/Survey PaperBenchmark
🎯 What it does: Systematically evaluate existing knowledge injection frameworks, finding that alignment injection performs almost equally to random or noise injection, with random even being superior; further analyze the reasons, propose using cleaned conceptual knowledge as the injection source, and verify that this method can significantly restore the advantages of alignment injection.
Revisiting the Optimality of Word Lengths
Tiago Pimentel (University of Cambridge), Ryan Cotterell (ETH Zürich)
OptimizationTransformerText
🎯 What it does: Explore the optimization of word length and compare the effectiveness of Zipf, CCH, and their lower bound prediction models.
Reward-Augmented Decoding: Efficient Controlled Text Generation With a Unidirectional Reward Model
Haikang Deng (UNC-Chapel Hill), Colin Raffel (University of Toronto)
GenerationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose Reward-Augmented Decoding (RAD), a method that utilizes a unidirectional reward model to guide language models in generating text with target attributes by adjusting sampling probabilities.
ROBBIE: Robust Bias Evaluation of Large Generative Language Models
David Esiobu (Meta), Eric Smith (Meta)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically evaluates five types of fundamental generative LLMs on 12 identity axes for bias and toxicity by constructing multi-dimensional metrics and a new dataset, while comparing three bias/toxicity mitigation techniques.
RoBoCoP: A Comprehensive ROmance BOrrowing COgnate Package and Benchmark for Multilingual Cognate Identification
Liviu Dinu, Laurentiu Zoicas (University of Bucharest)
ClassificationConvolutional Neural NetworkTransformerTextBenchmark
🎯 What it does: Built RoBoCoP—a complete, queryable dictionary database based on etymological information from five Romance language dictionaries, and designed and evaluated multiple machine learning and deep learning models for automatically detecting cross-lingual cognates (cognate detection)
Robust Prompt Optimization for Large Language Models Against Distribution Shifts
Moxin Li (National University of Singapore), Tat-Seng Chua (National University of Singapore)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied the robustness of LLM prompt optimization under distribution shift, and proposed a general prompt optimization framework GPO that leverages unlabeled target group data.
RobustGEC: Robust Grammatical Error Correction Against Subtle Context Perturbation
Yue Zhang (Soochow University), Shuming Shi (Soochow University)
GenerationLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Investigate the robustness of GEC systems when facing context perturbations unrelated to errors, and propose the RobustGEC benchmark and CPR post-training method.
Rumor Detection on Social Media with Crowd Intelligence and ChatGPT-Assisted Networks
Chang Yang (Shenyang University of Technology), Jiaming Zhao (Tianjin University)
ClassificationRecurrent Neural NetworkGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed the Crowd Intelligence and ChatGPT-Assisted Network (CICAN) model for social media rumor detection.
S2abEL: A Dataset for Entity Linking from Scientific Tables
Yuze Lou (University of Michigan), Doug Downey (Allen Institute for AI)
RetrievalTransformerLarge Language ModelTabularBenchmark
🎯 What it does: Created the first entity linking dataset for scientific tables (especially machine learning experiment result tables) called S2abEL, and designed a complete baseline model for table entity linking that supports detecting and annotating out-KB entities even when the knowledge base is incomplete.
SAMRank: Unsupervised Keyphrase Extraction using Self-Attention Map in BERT and GPT-2
Byungha Kang (Incheon National University), Youhyun Shin (Incheon National University)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Propose SAMRank, which performs unsupervised keyword extraction using only the self-attention maps of pre-trained language models.
Scalable-DSC: A Structural Template Prompt Approach to Scalable Dialogue State Correction
Haoxiang Su (Xinjiang University), Sijie Feng (Xinjiang Provincial Key Laboratory of Multi-lingual Information Technology)
TransformerPrompt EngineeringText
🎯 What it does: Propose a scalable dialogue state correction framework called Scalable-DSC, which utilizes Structural Template Prompt (STP) to convert dialogue states predicted by any DST model into natural language sequences, and generates corrected complete states based on this.
ScanDL: A Diffusion Model for Generating Synthetic Scanpaths on Texts
Lena Bolliger (University of Zurich), Lena Jäger (University of Zurich)
GenerationData SynthesisTransformerDiffusion modelTextSequential
🎯 What it does: Proposed and implemented SCANDL, a discrete sequence-to-sequence (seq2seq) generation framework based on diffusion models, for synthesizing human eye movement scan paths given text.
ScdNER: Span-Based Consistency-Aware Document-Level Named Entity Recognition
Ying Wei (Iowa State University), Qi Li (Iowa State University)
RecognitionTransformerTextBiomedical Data
🎯 What it does: Proposed a two-stage document-level named entity recognition model called ScdNER, which achieves more accurate and consistent entity predictions by leveraging span-level global feature fusion.
SCENE: Self-Labeled Counterfactuals for Extrapolating to Negative Examples
Deqing Fu (University of Southern California), Robin Jia (University of Southern California)
Data SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose the SCENE method, which perturbs positive samples using a masked filling model, and then generates high-quality negative samples through filtering and self-labeling, enabling the model to recognize unseen negative examples with only positive samples.
SciRepEval: A Multi-Format Benchmark for Scientific Document Representations
Amanpreet Singh (Allen Institute for Artificial Intelligence), Sergey Feldman (Allen Institute for Artificial Intelligence)
ClassificationRetrievalRepresentation LearningTransformerContrastive LearningTextBenchmark
🎯 What it does: Design and release the SciRepEval benchmark, which includes 24 multi-format tasks (classification, regression, proximity retrieval, retrieval), and use this benchmark to evaluate and improve scientific document representation models; based on this, propose the SPECTER2 multi-format model, which uses control codes and adapters to achieve task-format specific embeddings.
SCITAB: A Challenging Benchmark for Compositional Reasoning and Claim Verification on Scientific Tables
Xinyuan Lu (National University of Singapore), Min-Yen Kan (National University of Singapore)
TransformerLarge Language ModelTabularBenchmarkChain-of-Thought
🎯 What it does: Constructed a scientific table-based fact-checking benchmark, SCITAB, containing 1.2K expert-verified statements derived from real research papers, and generated forged and not enough information (NEI) statements through human-machine collaboration, forming a three-class (support, refute, not verifiable) dataset.
SEAHORSE: A Multilingual, Multifaceted Dataset for Summarization Evaluation
Elizabeth Clark (Google DeepMind), Ankur Parikh (Google DeepMind)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created the SEAHORSE multilingual and multidimensional summarization evaluation dataset and trained an evaluation metric based on mT5.
Seeing through the mess: evolutionary dynamics of lexical polysemy
Andreas Baumann (University of Vienna), Benjamin Roth (University of Vienna)
TransformerTextOrdinary Differential Equation
🎯 What it does: This study systematically analyzes the evolutionary mechanisms of polysemy by combining a mathematical model based on adaptive dynamics with empirical testing, focusing on the impact of word frequency, non-conformist behavior, and semantic discriminability on semantic differentiation, and verifying the model's predictions on historical English data.
SEER : A Knapsack approach to Exemplar Selection for In-Context HybridQA
Jonathan Tonglet (TU Darmstadt), Bart Baesens (KU Leuven)
OptimizationLarge Language ModelPrompt EngineeringTextTabularFinance Related
🎯 What it does: This paper proposes the SEER method, which selects examples for In-Context HybridQA using integer linear programming (Knapsack);
Select, Prompt, Filter: Distilling Large Language Models for Summarizing Conversations
Minh-Quang Pham (Zoom Video Communications), Marco Turchi (Zoom Video Communications)
GenerationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Knowledge distillation for large language models (e.g., ChatGPT) using a three-step process (selection, prompting, filtering) to generate high-quality dialogue summary data, which is then used to fine-tune a small generative model (BART-large) for forum and email conversations.
Selective Labeling: How to Radically Lower Data-Labeling Costs for Document Extraction Models
Yichao Zhou (Google Research), Sandeep Tata (Google Research)
Data-Centric LearningTextFinance Related
🎯 What it does: Proposed a selective annotation method that simplifies the labeling task to 'yes/no' confirmation of model-predicted candidate boxes, combined with active learning to reduce annotation costs for document extraction models.
Selectively Answering Ambiguous Questions
Jeremy Cole, Jacob Eisenstein (Google DeepMind)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Studied self-calibration and selective answering of large language models when facing ambiguous questions, proposing a 'disambiguate first, then answer' paradigm, and estimating confidence through sampling repetition rate/diversity;
Self-Detoxifying Language Models via Toxification Reversal
Chak Tou Leong (Hong Kong Polytechnic University), Wenjie Li (Hong Kong Polytechnic University)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Achieve self-detoxification without fine-tuning or additional components by identifying the poisoning direction through the differences in toxicity vectors generated by negative prefixes via two forward passes, and reversing the information flow in the attention layer during the second inference.
Self-Ensemble of N-best Generation Hypotheses by Lexically Constrained Decoding
Ryota Miyano (Osaka University), Yuki Arase (Osaka University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposes a self-integration method that generates a candidate set using N-best, creates positive and negative word constraints through word-level quality estimation, and then generates higher quality text via word-constrained decoding.
Self-Evolution Learning for Mixup: Enhance Data Augmentation on Few-Shot Text Classification Tasks
Haoqi Zheng (National University of Defense Technology), Dacheng Tao (University of Sydney)
ClassificationMeta LearningTransformerText
🎯 What it does: For few-shot text classification tasks, a self-evolving learning Mixup method is proposed, which gradually generates pseudo-samples for data augmentation based on sample difficulty, progressing from easy to hard.
Self-ICL: Zero-Shot In-Context Learning with Self-Generated Demonstrations
Wei-Lin Chen (National Taiwan University), Hsin-Hsi Chen (National Taiwan University)
Representation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Propose the SELF-ICL framework, enabling zero-shot self-generated demonstrations and ICL based solely on test inputs and task descriptions.
Self-Improvement of Non-autoregressive Model via Sequence-Level Distillation
Yusheng Liao (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
Knowledge DistillationTransformerText
🎯 What it does: Generate self-distillation data using its own NAT model to eliminate multimodal issues and improve translation quality.
Self-Influence Guided Data Reweighting for Language Model Pre-training
Megh Thakkar (Mila Quebec AI Institute), Partha Talukdar (Google Research)
Data-Centric LearningTransformerTextBenchmark
🎯 What it does: Proposed the PRESENCE method, which utilizes self-influence scores to perform online reweighting and filtering of data during language model pretraining, aiming to improve pretraining quality.
SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
Potsawee Manakul (University of Cambridge), Mark Gales
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose SelfCheckGPT, a zero-resource, black-box method for detecting hallucinations in generative large language models, which determines the authenticity of sentences by measuring consistency across multiple sampled outputs.
Semantic matching for text classification with complex class descriptions
Brian De Silva (Amazon), Mingwei Shen (Amazon)
ClassificationMeta LearningTransformerLarge Language ModelText
🎯 What it does: Transform the text classification task into a matching problem, using semantic matching scores between example text and class descriptions for classification, supporting zero-shot and few-shot learning.
Semantic Similarity Models for Depression Severity Estimation
Anxo Pérez (Universidade da Coruña), Iryna Gurevych (Technical University of Darmstadt)
ClassificationExplainability and InterpretabilityTransformerTextRetrieval-Augmented Generation
🎯 What it does: Construct a sentence retrieval-based semantic pipeline to estimate the severity of depression symptoms by analyzing users' social media posts, with each symptom's response predicted as a multi-class label.
Semantic Space Grounded Weighted Decoding for Multi-Attribute Controllable Dialogue Generation
Zhiling Zhang (Shanghai Jiao Tong University), Kenny Zhu (University of Texas at Arlington)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposed the DASC framework, which achieves multi-attribute controllable dialogue generation through semantic space weighted decoding
Semi-automatic Data Enhancement for Document-Level Relation Extraction with Distant Supervision from Large Language Models
Junpeng Li (National Key Laboratory of General Artificial Intelligence), Zilong Zheng (National Key Laboratory of General Artificial Intelligence)
GenerationData SynthesisData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Propose a semi-automated data augmentation framework that combines large language models (LLMs) with natural language inference (NLI) modules to automatically generate high-quality relation triplets for the document-level relation extraction (DocRE) task, and construct an improved test set DocGNRE and an expanded training set.
Semi-supervised multimodal coreference resolution in image narrations
Arushi Goel (University of Edinburgh), Hakan Bilen (A*STAR)
Representation LearningData-Centric LearningTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes a semi-supervised multimodal coreference resolution method that uses image-narrative pairs to simultaneously address coreference resolution and narrative localization tasks.
Sentiment Analysis on Streaming User Reviews via Dual-Channel Dynamic Graph Neural Network
Xin Zhang (Southeast University), Deyu Zhou (Southeast University)
ClassificationGraph Neural NetworkTransformerText
🎯 What it does: Proposed the DC-DGNN model, which uses a dual-channel dynamic graph neural network for sentiment analysis on streaming user reviews, capable of updating user and product representations over time and predicting ratings.
SentiStream: A Co-Training Framework for Adaptive Online Sentiment Analysis in Evolving Data Streams
Yuhao Wu (Singapore University of Technology and Design), Shuhao Zhang (Singapore University of Technology and Design)
ClassificationText
🎯 What it does: Proposed the SentiStream framework, achieving continuous adaptation and classification for online sentiment analysis in evolving data streams;
Seq2seq is All You Need for Coreference Resolution
Wenzheng Zhang (Rutgers University), Karl Stratos (Rutgers University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper redefines the coreference resolution task as a standard sequence-to-sequence (seq2seq) problem, directly fine-tuning pre-trained encoder-decoder models (such as T5, T0, FLAN-T5) on raw documents to output a serialized coreference annotation; no task-specific structures or hyperparameters are required.
SeqXGPT: Sentence-Level AI-Generated Text Detection
Pengyu Wang (Fudan University), Xipeng Qiu (Fudan University)
ClassificationAnomaly DetectionConvolutional Neural NetworkTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes SeqXGPT for sentence-level AI-generated text detection.
Set Learning for Generative Information Extraction
Jiangnan Li (Harbin Institute of Technology), Ruifeng Xu (Chinese University of Hong Kong)
GenerationText
🎯 What it does: To address sequence bias in information extraction tasks, this paper proposes a set learning method that reduces bias by sampling multiple permutations and using set probability to optimize the loss within a Seq2Seq generative framework.
Shall We Pretrain Autoregressive Language Models with Retrieval? A Comprehensive Study
Boxin Wang (University of Illinois Urbana-Champaign), Bryan Catanzaro (NVIDIA)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Reproduce and pretrain RETRO (including the RETRO++ variant) from scratch, compare with GPT on text generation, factuality, toxicity, and downstream tasks (e.g., open-domain QA), and validate the effectiveness of incorporating retrieval during pretraining.
SimCSE++: Improving Contrastive Learning for Sentence Embeddings from Two Perspectives
Jiahao Xu (Nanyang Technological University), Lemao Liu (Tencent AI Lab)
Representation LearningTransformerContrastive LearningText
🎯 What it does: Propose two improved methods for SimCSE: 1) off-dropout sampling, removing dropout noise in negative samples; 2) Dimension-wise Contrastive Learning (DCL), addressing the rank bottleneck caused by feature corruption.
Simple and Effective Input Reformulations for Translation
Brian Yu (University of California, Berkeley), Kurt Keutzer (University of California, Berkeley)
GenerationTransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Input reconstruction (POSE, ParSE, MiPS) for fine-tuning on translation tasks, enhancing mT5's performance in both monolingual and multilingual translation pairs.
Simple Temporal Adaptation to Changing Label Sets: Hashtag Prediction via Dense KNN
Niloofar Mireshghallah (University of California San Diego), Taylor Berg-Kirkpatrick (University of California San Diego)
ClassificationRetrievalDomain AdaptationTransformerText
🎯 What it does: Propose a non-parametric model based on dense KNN retrieval for temporal adaptation of long-term label sets (e.g., tweet label prediction).
Simplicity Level Estimate (SLE): A Learned Reference-Less Metric for Sentence Simplification
Liam Cripwell (Université de Lorraine), Claire Gardent (Université de Lorraine)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed a reference-free sentence simplification evaluation metric called SLE, which directly predicts the degree of sentence simplification and quantifies the simplification benefits.
SKD-NER: Continual Named Entity Recognition via Span-based Knowledge Distillation with Reinforcement Learning
Yi Chen (Xinjiang University), Liang He (Xinjiang University)
RecognitionKnowledge DistillationTransformerReinforcement LearningText
🎯 What it does: Proposed the SKD-NER model, achieving continual learning for named entity recognition (NER) that balances learning of both new and existing entity categories.
Skill-Based Few-Shot Selection for In-Context Learning
Shengnan An (Xi'an Jiaotong University), Jian-Guang Lou (Microsoft Corporation)
RetrievalMeta LearningLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose SKILL-KNN, a training-free few-shot selection method that generates skill descriptions through prompting and performs KNN retrieval based on these descriptions, aiming to enhance the contextual learning effectiveness of large language models.
SLOG: A Structural Generalization Benchmark for Semantic Parsing
Bingzhi Li (Universite Paris Cite), Najoung Kim (Boston University)
Data SynthesisTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposed the SLOG dataset, specifically designed to evaluate the performance of semantic parsing models on structural generalization (new syntactic structures), expanding the existing COGS benchmark;
Small Language Models Fine-tuned to Coordinate Larger Language Models improve Complex Reasoning
Gurusha Juneja (IIT Delhi), Tanmoy Chakraborty (IIT Delhi)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Proposes the DaSLaM framework, which delegates the decomposition of complex problems and their solution to specialized smaller decomposers and arbitrary solvers, respectively. The decomposer generates effective subproblems through reinforcement learning by observing feedback from the solver, thereby enhancing overall reasoning performance.
SMoP: Towards Efficient and Effective Prompt Tuning with Sparse Mixture-of-Prompts
Joon-Young Choi (Korea University), SangKeun Lee (Korea University)
Computational EfficiencyTransformerPrompt EngineeringMixture of ExpertsTextBenchmark
🎯 What it does: Propose the SMoP (Sparse Mixture-of-Prompts) method, which achieves efficient prompt tuning by using extremely short soft prompts through sparse mixing of prompts.
Sociocultural Norm Similarities and Differences via Situational Alignment and Explainable Textual Entailment
Sky CH-Wang (Columbia University), Smaranda Muresan (Columbia University)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper proposes a method for cross-cultural social norm discovery and comparison, utilizing the Chinese Zhihu platform and the American SOCIALCHEMISTRY dataset. By employing cross-lingual contextual alignment, few-shot prompt-based rule extraction, and chain-of-thought reasoning, it generates interpretable text entailment pairs, constructing a dataset of 3,069 cross-cultural social norm comparisons between China and the U.S.
SODA: Million-scale Dialogue Distillation with Social Commonsense Contextualization
Hyunwoo Kim (Allen Institute for Artificial Intelligence), Yejin Choi (Allen Institute for Artificial Intelligence)
GenerationKnowledge DistillationTransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Construct a large-scale social dialogue dataset SODA with millions of dialogues, and distill dialogues from large models using the CO3 framework.
Solving Hard Analogy Questions with Relation Embedding Chains
Nitesh Kumar (Cardiff University), Steven Schockaert (Cardiff University)
Representation LearningTransformerText
🎯 What it does: This paper proposes a hybrid method that combines knowledge graph paths with relation embedding chains to capture fine-grained relationships between concepts and solve challenging analogy problems.
Somali Information Retrieval Corpus: Bridging the Gap between Query Translation and Dedicated Language Resources
Abdisalam Badel, Fan Zhou (University Of Electronic Science And Technology Of China)
RetrievalTextBenchmark
🎯 What it does: Constructed the first public Somali language information retrieval evaluation corpus and implemented a retrieval system based on pseudo-relevance feedback.
SOUL: Towards Sentiment and Opinion Understanding of Language
Yue Deng (DAMO Academy, Alibaba Group), Lidong Bing (DAMO Academy, Alibaba Group)
GenerationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the SOUL task, which includes two subtasks: Review Comprehension (RC) and Justification Generation (JG), to evaluate models' capabilities in sentiment understanding and reasoning.
Sparse Low-rank Adaptation of Pre-trained Language Models
Ning Ding (Tsinghua University), Maosong Sun (Tsinghua University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose the sparse low-rank adaptation (SoRA) method, achieving dynamic low-rank adjustment of LoRA through gate vectors and proximal gradient, enabling parameter sparsification and post-training pruning.
Sparse Universal Transformer
Shawn Tan (Mila, University of Montreal), Chuang Gan (MIT-IBM Watson AI Lab)
Computational EfficiencyRepresentation LearningTransformerMixture of ExpertsContrastive LearningText
🎯 What it does: Propose Sparse Universal Transformer (SUT), integrating sparse expert mixture and dynamic stopping mechanism based on stick-breaking into Universal Transformer;
Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4
Kent K. Chang (University of California, Berkeley), David Bamman (University of California, Berkeley)
Explainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically evaluates the memory capacity of ChatGPT and GPT-4 for 571 novels using 'name-filling' membership inference queries, analyzes the impact of memory bias on downstream task performance, conducts benchmark comparisons with models like BERT, and proposes the use of models with publicly available training data to improve interpretability and reproducibility.
Specialist or Generalist? Instruction Tuning for Specific NLP Tasks
Chufan Shi (Tsinghua University), Deng Cai (Tencent)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically evaluates the impact of general instruction data on building specialized models by conducting two-stage instruction tuning on large language models, first using general instruction data (such as GPT4-Instruct, LIMA) and then task-specific data.
Speech Recognition and Meaning Interpretation: Towards Disambiguation of Structurally Ambiguous Spoken Utterances in Indonesian
Ruhiyah Widiaputri, Sakriani Sakti (Japan Advanced Institute of Science and Technology)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: This paper constructs a speech recognition and meaning interpretation system for Indonesian structural ambiguous sentences. First, 400 ambiguous sentences (totaling 4,800 audio recordings) were collected and annotated. Subsequently, through two frameworks—Cascade (ASR+TD) and Direct (SD)—utilizing acoustic features such as Mel-spectrogram, F0, and energy, along with meaning labels, the system automatically disambiguates ambiguous sentences into unambiguous text.
Speech-enriched Memory for Inference-time Adaptation of ASR Models to Word Dictionaries
Ashish Mittal (IBM Research), Gakuto Kurata (IIT Bombay)
RecognitionData SynthesisDomain AdaptationTextAudio
🎯 What it does: Construct a dictionary memory using synthetic speech, and during inference, combine KNN matching with a trie structure to perform dictionary-level zero-shot adaptation on existing Transducer and Whisper ASR models, thereby improving the recognition accuracy of rare words.
SpEL: Structured Prediction for Entity Linking
Hassan Shavarani, Anoop Sarkar (Simon Fraser University)
RetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Developed a structured prediction-based entity linking system called SPEL, which accurately maps text fragments to Wikipedia entities through subword-level classification and aggregation strategies.
Spoiler Detection as Semantic Text Matching
Ryan Tran (University of California, San Diego), Julian McAuley (University of California, San Diego)
ClassificationRetrievalData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a spoiler matching task that aligns TV show comments with TV show summaries through semantic matching, addressing the issue that traditional spoiler detection methods are insensitive to viewers' watching progress.
SPT: Learning to Selectively Insert Prompts for Better Prompt Tuning
Wei Zhu (East China Normal University), Ming Tan (Southern University of Science and Technology)
ClassificationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Learning to automatically select which intermediate layers in a pre-trained model (PTM) to insert instance-aware prompts, thereby improving prompt tuning performance.
STAIR: Learning Sparse Text and Image Representation in Grounded Tokens
Chen Chen (Apple), Yinfei Yang (Apple)
RetrievalExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the STAIR (Sparse Text And Image Representation) model, which maps images and text into a sparse and interpretable dictionary space, achieving efficient and interpretable cross-modal retrieval;
Stance Detection on Social Media with Background Knowledge
Ang Li (Harbin Institute of Technology), Ruifeng Xu (SIAT Chinese Academy of Sciences)
ClassificationRetrievalTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Explored how to leverage background knowledge (narrative knowledge and discourse knowledge) to enhance social media stance detection, using ChatGPT for retrieval and filtering, and proposed a knowledge-enhanced framework called KASD.
Standardizing Distress Analysis: Emotion-Driven Distress Identification and Cause Extraction (DICE) in Multimodal Online Posts
Gopendra Singh, Asif Ekbal (IIT Patna)
ClassificationRecognitionConvolutional Neural NetworkGraph Neural NetworkTransformerVision Language ModelAuto EncoderMultimodality
🎯 What it does: This paper proposes a unified multi-task deep framework called DICE for simultaneously identifying distress content (Distress Identification) and extracting text fragments causing distress (Cause Extraction) from social media posts containing both text and images.
Statistical Depth for Ranking and Characterizing Transformer-Based Text Embeddings
Parker Seegmiller (Dartmouth College), Sarah Masud Preum (Dartmouth College)
ClassificationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a Transformer-based Text Embedding Statistical Depth (TTE Depth) method to perform center-outlier ranking on text collections, and designs a Wilcoxon rank-sum test based on this depth to determine whether the distributions of two groups of text embeddings are significantly different.
StereoMap: Quantifying the Awareness of Human-like Stereotypes in Large Language Models
Sullam Jeoung (University of Illinois at Urbana-Champaign), Jana Diesner (University of Illinois at Urbana-Champaign)
Explainability and InterpretabilityLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Built a framework called STEREOMAP based on the psychology Stereotype Content Model (SCM) to quantify the stereotypes of large language models (LLMs) regarding the warmth and competence dimensions of social groups.
STINMatch: Semi-Supervised Semantic-Topological Iteration Network for Financial Risk Detection via News Label Diffusion
Xurui Li, Xiaozhong Liu (Worcester Polytechnic Institute)
Data-Centric LearningGraph Neural NetworkTransformerTextGraphBenchmarkFinance Related
🎯 What it does: Propose a semi-supervised semantic-topology iterative network, STINMatch, integrated with the news-enterprise knowledge graph (NEKG), to achieve financial risk detection.
Stop Uploading Test Data in Plain Text: Practical Strategies for Mitigating Data Contamination by Evaluation Benchmarks
Alon Jacovi (Bar Ilan University), Yoav Goldberg (Bar Ilan University)
Safty and PrivacyTextBenchmark
🎯 What it does: This paper addresses the issue of large language models inadvertently leaking evaluation data during training by proposing three actionable prevention strategies: encrypting test data with public-key encryption and using a No Derivatives license; refusing to send evaluation data to closed-source APIs without training exclusion controls; and avoiding the use of data with publicly available answers on the internet while simultaneously publishing the original context.
StoryAnalogy: Deriving Story-level Analogies from Large Language Models to Unlock Analogical Understanding
Cheng Jiayang (Hong Kong University of Science and Technology), Zheng Zhang (Westlake University)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBenchmark
🎯 What it does: Built a large-scale story-level analogy corpus named STORYANALOGY, and designed annotation and evaluation methods for entity similarity and relation similarity based on an extended structural mapping theory. Subsequently, systematically evaluated the model's performance on story analogy recognition and generation tasks.
StrAE: Autoencoding for Pre-Trained Embeddings using Explicit Structure
Mattia Opper (University of Edinburgh), N. Siddharth (University of Edinburgh)
Representation LearningAuto EncoderContrastive LearningText
🎯 What it does: A structured autoencoder StrAE is studied, which encodes pre-trained corpora using a given tree structure and contrastive learning objectives, exploring the role of explicit hierarchical structures in multi-layer semantic representation learning, and proposes a self-structured version Self-StrAE, demonstrating that strong representations can be achieved solely through prior hierarchical merging.
Struct-XLM: A Structure Discovery Multilingual Language Model for Enhancing Cross-lingual Transfer through Reinforcement Learning
Linjuan Wu (Zhejiang University), Weiming Lu (Zhejiang University)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose Struct-XLM, which automatically discovers general syntactic structures through reinforcement learning and integrates them into multilingual pre-trained models to enhance cross-lingual representation alignment and transfer learning effectiveness.
StructGPT: A General Framework for Large Language Model to Reason over Structured Data
Jinhao Jiang (Renmin University of China), Ji-Rong Wen (Renmin University of China)
TransformerLarge Language ModelPrompt EngineeringGraphTabular
🎯 What it does: Designed a unified framework called StructGPT, which constructs specialized interfaces (KG, tables, databases) to enable large language models (LLMs) to first read relevant structured data and then perform reasoning, thereby enhancing the zero-shot/few-shot performance of LLMs on structured data reasoning tasks.
Structural generalization in COGS: Supertagging is (almost) all you need
Alban Petit (Université Paris-Saclay), François Yvon (Sorbonne Université)
OptimizationRecurrent Neural NetworkTextBenchmark
🎯 What it does: Introduce a supertagging step in semantic parsing and ensure structural consistency through integer linear programming to enhance compositional generalization.
Structural Priming Demonstrates Abstract Grammatical Representations in Multilingual Language Models
James A. Michaelov (University of California San Diego), Benjamin K. Bergen (University of California San Diego)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This paper uses structural priming experiments to evaluate whether multilingual Transformer models (XGLM, PolyLM) possess cross-linguistic abstract syntactic representations.
Structure-aware Knowledge Graph-to-text Generation with Planning Selection and Similarity Distinction
Feng Zhao (Huazhong University of Science and Technology), Cheng Yan (Huazhong University of Science and Technology)
GenerationTransformerLarge Language ModelContrastive LearningGraph
🎯 What it does: This paper proposes a two-stage knowledge graph to text (KG-to-Text) generation framework: the first stage uses a planning scorer to select the optimal sequence among all possible triplet linearization sequences; the second stage enhances the BART encoder with a graph-aware module incorporating relative distance encoding and distinguishes similar graphs and entities through contrastive learning, thereby generating accurate and coherent text.
SummEdits: Measuring LLM Ability at Factual Reasoning Through The Lens of Summarization
Philippe Laban (Salesforce AI), Chien-Sheng Wu (Salesforce AI)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the SUMMEDITS benchmark and an editing-based evaluation protocol to test the fact consistency detection capability of large language models (LLMs) in text summarization.
SuperDialseg: A Large-scale Dataset for Supervised Dialogue Segmentation
Junfeng Jiang (University of Tokyo), Akiko Aizawa (National Institute of Informatics)
SegmentationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a large-scale supervised dialogue segmentation dataset called SuperDialseg and conducted benchmark experiments on multiple models.
Superlim: A Swedish Language Understanding Evaluation Benchmark
Aleksandrs Berdicevskis (University of Gothenburg), Nina Tahmasebi (University of Gothenburg)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes Superlim, a multitask natural language understanding benchmark and evaluation platform for Swedish, covering 14 datasets, 15 tasks, and equipped with an online leaderboard and documentation.
Support or Refute: Analyzing the Stance of Evidence to Detect Out-of-Context Mis- and Disinformation
Xin Yuan (Shanghai Jiao Tong University), Shujun Li (University of Kent)
ClassificationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningMultimodalityRetrieval-Augmented Generation
🎯 What it does: Propose a unified framework that leverages multimodal evidence to detect out-of-context information in image-text pairs
SUT: Active Defects Probing for Transcompiler Models
Mengnan Qi (Microsoft Cloud and AI), Neel Sundaresan (Microsoft Cloud and AI)
Explainability and InterpretabilityData-Centric LearningAI Code AssistantLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Developed an evaluation framework based on syntax unit testing called SUT, which includes interpretable SUT Acc and SETS metrics, capable of accurately diagnosing syntax errors in cross-language translation models.
Syllogistic Reasoning for Legal Judgment Analysis
Wentao Deng (Shandong University), Pengjie Ren (Shandong University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper constructs a syllogistic legal judgment analysis (SLJA) framework based on syllogistic reasoning, generates the first Chinese SLJA dataset (SLJA-COR) containing syllogistic reasoning processes from Chinese criminal cases, and systematically evaluates the performance of four advanced large language models on four tasks (document retrieval, crime element generation, document interpretation, and judgment prediction).
Symbol tuning improves in-context learning in language models
Jerry Wei (Google), Quoc Le
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: During the training phase, natural language labels are replaced with arbitrary symbols, and input-label pairs across various NLP classification tasks are fine-tuned, resulting in a method called Symbol Tuning.
Symbolic Planning and Code Generation for Grounded Dialogue
Justin Chiu (Cornell Tech), Daniel Fried (Carnegie Mellon University)
Explainability and InterpretabilityAI Code AssistantLarge Language ModelPoint Cloud
🎯 What it does: Designed a dialogue system based on symbolic planning and code generation, which can achieve image point localization and interaction by converting language into executable Python code in the ONECOMMON task.
Syntactic Substitutability as Unsupervised Dependency Syntax
Jasper Jian (Stanford University), Siva Reddy (Mila Quebec AI Institute and McGill University)
TransformerLarge Language ModelText
🎯 What it does: Propose an unsupervised dependency syntactic parsing method called SSUD based on sentence syntactic substitutability, leveraging the averaging of BERT's self-attention distributions to extract syntactic information;
Synthetic Data Generation with Large Language Models for Text Classification: Potential and Limitations
Zhuoyan Li (Purdue University), Ming Yin (Purdue University)
ClassificationData SynthesisTransformerLarge Language ModelText
🎯 What it does: Explored the effectiveness of large language models generating synthetic data in text classification tasks, and investigated the moderating effect of task and instance subjectivity on model performance.
System Combination via Quality Estimation for Grammatical Error Correction
Muhammad Reza Qorib (National University of Singapore), Hwee Tou Ng (National University of Singapore)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a new grammar error correction quality estimation model, GRECO, and applies it to a multi-model system combination; by re-ranking all edited beam search results with quality estimation scores, higher quality correction outcomes are achieved.
Systematic word meta-sense extension
Lei Yu (University of Toronto)
Representation LearningTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: This paper proposes and implements the Systematic Word Sense Expansion (SWORME) task, constructing an 880k-sentence dataset covering 50 pairs of systematic meta-words, and evaluates the performance of language models in word sense expansion and metaphor understanding by comparing analogy chains and association chain-based sequential learning methods.
TacoPrompt: A Collaborative Multi-Task Prompt Learning Method for Self-Supervised Taxonomy Completion
Hongyuan Xu (Nankai University), Xiaojie Yuan (Nankai University)
ClassificationRepresentation LearningTransformerPrompt EngineeringText
🎯 What it does: Propose a collaborative multi-task framework called TacoPrompt based on prompt learning, which uses a self-supervised classifier to accomplish tasks, capable of simultaneously identifying appropriate parent-child node pairs for new concepts.
Tagging-Assisted Generation Model with Encoder and Decoder Supervision for Aspect Sentiment Triplet Extraction
Luo Xianlong (Sun Yat-Sen University), Yihao Wang (Sun Yat-Sen University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the TAGS model, which combines sequence labeling with generative models to improve the extraction quality of Aspect Sentiment Triplet Extraction (ASTE) tasks through multi-perspective label assistance and label semantic alignment.
Target-Agnostic Gender-Aware Contrastive Learning for Mitigating Bias in Multilingual Machine Translation
Minwoo Lee (Seoul National University), Kyomin Jung (Seoul National University)
Representation LearningTransformerContrastive LearningText
🎯 What it does: To address gender bias in multilingual machine translation, we propose a target-language-agnostic gender-aware contrastive learning method called GACL, which corrects gender bias by injecting contextual gender information into encoder representations.