CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Constructed the first open Chinese speech-to-text conversion dataset CS2W, covering four types of errors in spoken language and providing fine-grained annotations;
🎯 What it does: Propose a generative adversarial attack (CT-GAT) based on cross-task transferable features, which directly generates adversarial text by training a sequence-to-sequence generator, avoiding the need to construct an alternative model.
Cultural Concept Adaptation on Multimodal Reasoning
Zhi Li (Zhejiang University), Yin Zhang (Zhejiang University)
CodeData SynthesisDomain AdaptationSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: This paper proposes an unlabeled cultural concept adaptation method and designs a multimodal data augmentation technique called CultureMixup, aiming to improve the performance of cross-lingual and cross-cultural vision-text reasoning models.
DALE: Generative Data Augmentation for Low-Resource Legal NLP
Sreyan Ghosh (University of Maryland), Dinesh Manocha (University of Maryland)
CodeClassificationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In low-resource legal NLP tasks, the DALE framework is proposed, which uses BART pretraining and performs selective masked denoising sequence-to-sequence learning on legal texts, then generates diverse, coherent, and label-consistent synthetic data augmentation samples for downstream tasks.
DecoMT: Decomposed Prompting for Machine Translation Between Related Languages using Large Language Models
Ratish Puduppully (Institute for Infocomm Research, A STAR), Nancy Chen
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a machine translation method called DecoMT, which uses large language models for few-shot prompting by splitting sentences into chunks for decomposed prompting.
Democratizing Reasoning Ability: Tailored Learning from Large Language Model
Zhaoyang Wang (Sun Yat-sen University), Qi Zhang (Microsoft)
CodeKnowledge DistillationLarge Language ModelPrompt EngineeringContrastive LearningTextBenchmarkChain-of-Thought
🎯 What it does: Through multi-round interactive learning, a small LM is trained to possess chain-of-thought (CoT) capabilities, while self-reflective learning enhances its reasoning quality.
CodeSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the DEPN framework, which reduces the risk of models leaking private information by detecting and zeroing out privacy neurons in pre-trained language models.
DeSIQ: Towards an Unbiased, Challenging Benchmark for Social Intelligence Understanding
Xiao-Yu Guo (Monash University), Reza Haf
CodeKnowledge DistillationTransformerLarge Language ModelMultimodalityBenchmark
🎯 What it does: Identify biases in the Social-IQ dataset, construct a bias-free and more challenging DeSIQ benchmark, and provide new multimodal model baselines.
Detecting Spoilers in Movie Reviews with External Movie Knowledge and User Networks
Heng Wang (Xi'an Jiaotong University), Minnan Luo (Xi'an Jiaotong University)
CodeClassificationGraph Neural NetworkTransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
🎯 What it does: Proposed a multi-view movie review spoiler detection framework named MVSD, and constructed a large-scale LCS dataset and a UKM movie knowledge base based on IMDb.
🎯 What it does: This paper proposes to utilize two types of discourse structures, local and global, for fine-grained propaganda content identification at both the sentence-level and word-level.
🎯 What it does: This paper proposes a retrieval-augmented style transfer framework (RAST), which generates diverse question-answering questions by retrieving and combining context from an external question template library.
Diversity Enhanced Narrative Question Generation for Storybooks
Hokeun Yoon (Sungkyunkwan University), JinYeong Bak (Sungkyunkwan University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed and implemented a multi-question generation model called mQG, which can automatically generate diverse and answerable narrative questions given context, question type, and already generated questions.
🎯 What it does: Proposes a self-supervised framework DNA, which utilizes k-NN to retrieve neighbors and aggregate information, learning fine-grained category representations from coarse-grained labeled data.
Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models
Orevaoghene Ahia (University of Washington), Yulia Tsvetkov (University of Washington)
CodeComputational EfficiencyData-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Studied the usage cost and performance differences caused by tokenization imbalance in commercial language model APIs across different languages.
Document-Level Machine Translation with Large Language Models
Longyue Wang (Tencent AI Lab), Zhaopeng Tu (Tencent AI Lab)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper systematically evaluates the performance of large language models (ChatGPT GPT-3.5 and GPT-4) in document-level machine translation, conducting in-depth experiments from three aspects: prompt engineering, model comparison, and discourse modeling.
Document-level Relationship Extraction by Bidirectional Constraints of Beta Rules
Yichun Liu (Tianjin University), Yaxin Li (Tianjin University)
CodeGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: Propose a framework named BCBR that enhances document-level relation extraction models by utilizing a Beta rule miner and bidirectional logical constraints;
Don’t Trust ChatGPT when your Question is not in English: A Study of Multilingual Abilities and Types of LLMs
Xiang Zhang (University of Alberta), Grzegorz Kondrak (University of Alberta)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Systematically evaluate the multilingual capabilities of large language models using both quantitative and qualitative methods, and experimentally test GPT's performance on translation equivalence and variation tasks through prompt translation and response back-translation methods.
Doolittle: Benchmarks and Corpora for Academic Writing Formalization
Shizhe Diao (Hong Kong University of Science and Technology), Tong Zhang (Hong Kong University of Science and Technology)
CodeGenerationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Propose the academic writing formalization (AWF) task and the non-parallel dataset DOOLITTLE, using multi-objective reinforcement learning to enhance text quality.
DPP-TTS: Diversifying prosodic features of speech via determinantal point processes
Seongho Joo (Seoul National University), Kyomin Jung (Seoul National University)
CodeGenerationTransformerFlow-based ModelAudio
🎯 What it does: Proposed a text-to-speech model called DPP‑TTS based on Determinantal Point Processes (DPP), enhancing speech expressiveness diversity through diversified intonation and rhythm.
e-THERAPIST: I suggest you to cultivate a mindset of positivity and nurture uplifting thoughts
Kshitij Mishra (Indian Institute of Technology Patna), Asif Ekbal (Indian Institute of Technology Patna)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Developed a psychological therapy dialogue system called e-THERAPIST capable of generating polite and friendly conversations based on user gender, age, personality traits, and emotions.
EDIS: Entity-Driven Image Search over Multimodal Web Content
Siqi Liu (Cornell University), William Wang (University of California Santa Barbara)
CodeRetrievalTransformerSupervised Fine-TuningVision Language ModelContrastive LearningMultimodalityBenchmark
🎯 What it does: Propose a large-scale (millions of scale) news domain entity-driven image search dataset called EDIS and study how to retrieve multimodal candidates (image + title).
Anshita Gupta (University of Massachusetts Amherst), Niket Tandon (Allen Institute for AI)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed MEMITCSK, an improved parameter editing algorithm for directly correcting common-sense misjudgments in Transformer models, and constructed a new evaluation set called PROBE SET;
CodeClassificationRecognitionTransformerLarge Language ModelText
🎯 What it does: This paper introduces a co-sympathetic intent recognition task and jointly trains empathy detection, significantly improving the accuracy of empathy detection.
Empower Nested Boolean Logic via Self-Supervised Curriculum Learning
Hongqiu Wu (Shanghai Jiao Tong University), Min Zhang (Soochow University)
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Designed a self-supervised curriculum learning method called CLR to train language models to master multi-layer nested Boolean logic reasoning, and proposed the corresponding BoolKill benchmark dataset.
Juan Pablo Zuluaga-Gomez (Idiap Research Institute), Marcello Federico (AWS AI Labs)
CodeTransformerTextAudio
🎯 What it does: This paper proposes an end-to-end multi-speaker multi-turn dialogue speech translation system (STAC-ST), which simultaneously accomplishes ASR, speech translation, and speaker switch detection through multi-task training, achieving joint output of tasks using special marker sequences;
Enhancing Code-Switching for Cross-lingual SLU: A Unified View of Semantic and Grammatical Coherence
Zhihong Zhu (Peking University), Yuexian Zou (Peking University)
CodeDomain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a framework named SOGO for improving semantic and syntactic consistency in zero-shot cross-lingual SLU tasks through enhanced code-switching.
🎯 What it does: For low-resource fine-grained named entity recognition, a Fine-to-Coarse (F2C) mapping matrix is constructed using coarse-grained data, combined with inconsistency filtering to achieve joint training;
🎯 What it does: In open-domain fact verification, the SEE-ST method is proposed, improving table extraction and cell selection to enhance the efficiency of structured evidence retrieval.
Enhancing the Ranking Context of Dense Retrieval through Reciprocal Nearest Neighbors
George Zerveas (Brown University), Carsten Eickhoff (University of Tübingen)
CodeRetrievalText
🎯 What it does: Propose an evidence label smoothing method based on reciprocal nearest neighbors (rNN) to improve the training of dense retrieval models under sparse annotations, and perform efficient re-ranking using rNN during the inference phase.
Sagnik Ray Choudhury (University of Michigan), Isabelle Augenstein (University of Copenhagen)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Create and use the multi-annotator SpanEx dataset, manually annotate cross-sentence span interactions in natural language inference (NLI) and fact checking (FC) tasks, and evaluate the decision-making process of large language models (LLMs) based on this dataset; propose an unsupervised method based on community detection to automatically extract interpretable text span interactions.
Explaining with Contrastive Phrasal Highlighting: A Case Study in Assisting Humans to Detect Translation Differences
Eleftheria Briakou (Google), Marine Carpuat (University of Maryland)
CodeExplainability and InterpretabilityTransformerContrastive LearningText
🎯 What it does: Proposed a post-hoc interpretable method that uses contrastive phrase highlighting to explain predictions of NLP models comparing two texts, focusing on identifying translation differences;
Explanation Selection Using Unlabeled Data for Chain-of-Thought Prompting
Xi Ye (University of Texas at Austin), Greg Durrett (University of Texas at Austin)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose a black-box search framework based on unlabeled data to automatically generate and optimize chain-of-thought (explanation) prompts, thereby improving the performance of large language models on text reasoning tasks.
Explicit Planning Helps Language Models in Logical Reasoning
Hongyu Zhao (University of Chicago), Hongyuan Mei (Toyota Technological Institute at Chicago)
CodeTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed the LEAP system, which utilizes language models for multi-step logical reasoning and incorporates explicit planning during the reasoning process.
Exploiting Asymmetry for Synthetic Training Data Generation: SynthIE and the Case of Information Extraction
Martin Josifoski (EPFL), Robert West (EPFL)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a process for generating synthetic training data through reverse tasks, utilizing a large language model (LLM) to first sample structured outputs (entity–relation–object triplets) and then prompt the LLM to generate corresponding natural language text, thereby obtaining high-quality, balanced closed information extraction (cIE) training samples.
Explore-Instruct: Enhancing Domain-Specific Instruction Coverage through Active Exploration
Fanqi Wan (Sun Yat Sen University), Shuming Shi (Tencent Ai Lab)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Construct domain-specific instruction sets and perform instruction tuning by combining active exploration (lookahead, backtracking) with LLMs to enhance model performance in specific domains.
🎯 What it does: Explore the impact of discourse structure on translation quality in document-level machine translation, and propose a multi-granularity attention mechanism based on RST.
Exploring Jiu-Jitsu Argumentation for Writing Peer Review Rebuttals
Sukannya Purkayastha (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
CodeClassificationGenerationDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Constructed the JITSUPEER dataset and proposed a rebuttal generation task for peer review based on the 'Jiu-Jitsu' argumentation framework, focusing on attitude origins and theme-driven rebuttals; designed three subtasks (feasibility scoring, description generation, terminal generation) and conducted baseline experiments.
Exploring Linguistic Probes for Morphological Generalization
Jordan Kodner (Stony Brook University), Sarah Payne (Stony Brook University)
CodeExplainability and InterpretabilityRepresentation LearningRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: Designed and evaluated language-specific morphological probes to examine the morphological generalization capabilities of different models on English, Spanish, and Swahili.
FactKB: Generalizable Factuality Evaluation using Language Models Enhanced with Factual Knowledge
Shangbin Feng (University of Washington), Yulia Tsvetkov (University of Washington)
CodeClassificationRepresentation LearningData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a knowledge-based pre-training method that first performs fact-based pre-training on language models and then conducts fact consistency evaluation.
Fair Text Classification with Wasserstein Independence
Thibaud Leteno (University of Saint Etienne), Christophe Gravier (University of Saint Etienne)
CodeClassificationTransformerLarge Language ModelGenerative Adversarial NetworkText
🎯 What it does: Propose an unsupervised fair text classification method (WFC) based on Wasserstein mutual information minimization, which enforces independence between the target and sensitive attributes in the latent space through a pre-trained sensitive attribute model and Wasserstein regularization, without requiring access to sensitive labels during training or inference.
Shahar Jacob (Hebrew University of Jerusalem), Dafna Shahaf (Hebrew University of Jerusalem)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes FAME, a system that performs analogy mapping using only entity names, automatically extracts common-sense relationships, and generates explainable mappings.
Fast and Accurate Factual Inconsistency Detection Over Long Documents
Barrett Lattimer (ASAPP), Yi Yang (ASAPP)
CodeAnomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposed the SCALE method based on NLI, which detects factual inconsistencies in generated text by splitting long texts into chunks and reasoning between each chunk and generated sentences.
Faster Minimum Bayes Risk Decoding with Confidence-based Pruning
Julius Cheng (University of Cambridge), Andreas Vlachos (University of Cambridge)
CodeComputational EfficiencyText
🎯 What it does: Proposed an iterative confidence pruning algorithm to accelerate Minimum Bayes Risk (MBR) decoding and reduce the number of utility function calls.
Federated Learning of Large Language Models with Parameter-Efficient Prompt Tuning and Adaptive Optimization
Tianshi Che (Auburn University), Dejing Dou (Baidu Inc)
CodeFederated LearningSafty and PrivacyLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the FedPepTAO framework in a federated learning environment for parameter-efficient prompt tuning of large language models, enhancing performance through adaptive optimization.
FedID: Federated Interactive Distillation for Large-Scale Pretraining Language Models
Xinge Ma (Yunnan University), Xuejie Zhang (Yunnan University)
CodeFederated LearningKnowledge DistillationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the Federated Interactive Distillation (FedID) framework, which uses a small amount of labeled data on the server to correct confirmation bias in traditional federated distillation, achieving decentralized training of large-scale pre-trained language models
FedTherapist: Mental Health Monitoring with User-Generated Linguistic Expressions on Smartphones via Federated Learning
Jaemin Shin (KAIST), Sung-Ju Lee (KAIST)
CodeClassificationFederated LearningSafty and PrivacyComputational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: Proposes FedTherapist, a mobile mental health monitoring system based on federated learning, utilizing user-generated speech transcriptions and keyboard input text data;
Fidelity-Enriched Contrastive Search: Reconciling the Faithfulness-Diversity Trade-Off in Text Generation
Wei-Lin Chen (National Taiwan University), Chung-Chi Chen (Artificial Intelligence Research Center AIST)
CodeGenerationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a new decoding strategy called Fidelity-Enriched Contrastive Search (FECS), which enhances the faithfulness of text generation and suppresses repetition by incorporating a semantic similarity reward into Contrastive Search.
Fighting Fire with Fire: The Dual Role of LLMs in Crafting and Detecting Elusive Disinformation
Jason Lucas (Pennsylvania State University), Dongwon Lee (Pennsylvania State University)
CodeGenerationAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose the F3 framework, leveraging LLMs for generating, purifying, and zero-shot detection of fake information, constructing an end-to-end 'fighting fire with fire' process.
Find-2-Find: Multitask Learning for Anaphora Resolution and Object Localization
Cennet Oguz (German Research Center for Artificial Intelligence), Josef van Genabith (German Research Center for Artificial Intelligence)
CodeObject DetectionTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: Propose the Find2Find dataset and design a joint multi-task learning model to address coreference resolution and object localization under multimodal visual-language ambiguity.
🎯 What it does: A corpus consisting of 250 hate tweets and 54,816 paragraphs was constructed, with real counter-arguments against hate annotated; a method to retrieve online articles for real counter-arguments was proposed.
🎯 What it does: This paper proposes an embedding initialization method called FOCUS, which rapidly adapts the target language's dedicated tokenizer while keeping the structure of multilingual pre-trained models unchanged.
From Values to Opinions: Predicting Human Behaviors and Stances Using Value-Injected Large Language Models
Dongjun Kang (Sungkyunkwan University), JinYeong Bak (Sungkyunkwan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In this study, the authors propose a Value Injection Method (VIM), which injects fine-grained value distributions into LLAMA through two tasks: argument generation and question answering. Subsequently, the model with injected values is used to predict group opinions and behaviors.
G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment
Yang Liu (Microsoft Azure AI), Chenguang Zhu (Microsoft Azure AI)
CodeGenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Propose the G-EVAL framework, which combines large language models (GPT-4) with chain-of-thought (CoT) reasoning, further integrated with a form-filling paradigm, to perform reference-free automatic evaluation of natural language generation (NLG) text.
GATITOS: Using a New Multilingual Lexicon for Low-resource Machine Translation
Alexander Jones (Google Research), Ishank Saxena (Google Research)
CodeData-Centric LearningTransformerText
🎯 What it does: This paper significantly improves translation quality by incorporating bilingual dictionaries for data augmentation in low-resource and unsupervised machine translation training.
GazeVQA: A Video Question Answering Dataset for Multiview Eye-Gaze Task-Oriented Collaborations
Muhammet Ilaslan, Mike Shou
CodeObject DetectionSegmentationTransformerVision Language ModelContrastive LearningImageVideoTextMultimodalityBenchmark
🎯 What it does: Constructed a collaborative task VQA dataset named GazeVQA containing multi-perspective videos and eye movement data, and proposed the AssistGaze model capable of answering questions with text, images, and videos.
GEM: Gestalt Enhanced Markup Language Model for Web Understanding via Render Tree
Zirui Shao (Zhejiang University), Xiaozhong Liu (Worcester Polytechnic Institute)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes GEM (Gestalt Enhanced Markup Language Model), which enhances web understanding capabilities without requiring visual input by leveraging visual information from web rendering trees (style and position) during pre-training, and introducing two pre-training tasks based on Gestalt principles (Same Textual Style Prediction and Proximate Nodes Prediction).
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a framework for generating causal counterfactual data (CCG) that complies with common-sense constraints, aiming to enhance the stability of relation extraction models in low-resource, cross-domain, and adversarial scenarios.
Generating Data for Symbolic Language with Large Language Models
Jiacheng Ye (University of Hong Kong), Tao Yu (University of Hong Kong)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Leveraging large language models (e.g., Codex) as data generators, combined with prompt engineering and self-consistent verification to generate symbolic language data with high annotation costs, then training with a small-scale task model (e.g., T5-Large) to significantly reduce deployment and inference costs;
GLEN: General-Purpose Event Detection for Thousands of Types
Sha Li (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Constructed a large-scale general event detection dataset GLEN (3,465 event types, 208,000 sentences) and proposed a multi-stage event detection model CEDAR, which performs trigger word identification, sentence-level type ranking, and trigger word-level type classification separately.
GLEN: Generative Retrieval via Lexical Index Learning
Sunkyung Lee (Sungkyunkwan University), Jongwuk Lee (Sungkyunkwan University)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed a generative retrieval framework called GLEN, which utilizes dynamic vocabulary-based indexing learning to directly generate document identifiers corresponding to queries, and enhances retrieval performance through two-phase training (keyword assignment and ranking-based ID fine-tuning) as well as collision-free inference.
GlobalBench: A Benchmark for Global Progress in Natural Language Processing
Yueqi Song (Carnegie Mellon University), Graham Neubig (George Mason University)
CodeTextBenchmark
🎯 What it does: Propose GlobalBench, a continuously expanding multilingual multitask benchmark and leaderboard, tracking the performance, utility, and fairness of NLP systems across all languages, and incentivizing improvements for under-served languages through a reward mechanism.
Goal-Driven Explainable Clustering via Language Descriptions
Zihan Wang (University of California San Diego), Ruiqi Zhong (University of California Berkeley)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposed a new text clustering task called GOALEX, which requires clustering results to align with user-specified goals and provide natural language explanations for the meaning of each cluster.
🎯 What it does: Studied a language grouping method based on gradient similarity to help select sets of languages that mutually positively influence each other during multilingual training.
Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning
Saibo Geng (EPFL), Robert West (EPFL)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the Grammar-Constrained Decoding (GCD) framework, which uses formal grammar to constrain the generation process of large language models (LLMs), enabling them to perform various structured NLP tasks without fine-tuning.
Hallucination Detection for Generative Large Language Models by Bayesian Sequential Estimation
Xiaohua Wang (Fudan University), Xuanjing Huang (Fudan University)
CodeRetrievalAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: Propose a hallucination detection framework based on Bayesian sequential analysis, dynamically determining the number of external evidence retrievals and performing hierarchical judgment.
HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
Junyi Li (Gaoling School of Artificial Intelligence, Renmin University of China), Ji-Rong Wen (Gaoling School of Artificial Intelligence, Renmin University of China)
CodeTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Constructed HaluEval, an evaluation benchmark for large language models containing 35,000 hallucination samples generated both by human annotation and automatic generation, used to test models' ability to detect hallucinations.
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Proposed JEEBENCH, a challenging question bank consisting of 515 high-difficulty mathematics, physics, and chemistry problems from 8 years of the Indian IIT JEE-Advanced exams, to evaluate LLMs' long-range reasoning and professional knowledge application capabilities.
CodeRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a hierarchical argumentation graph Hi-ArG and built an automated construction process, followed by further pretraining of language models using the GreaseArG multimodal model and custom pretraining tasks
Hidding the Ghostwriters: An Adversarial Evaluation of AI-Generated Student Essay Detection
Xinlin Peng (University of Chinese Academy of Sciences), Yingfei Sun (University of Chinese Academy of Sciences)
CodeClassificationAdversarial AttackTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the AIG-ASAP dataset to evaluate adversarial perturbations on AI-generated student essays and tested the robustness of existing AI-generated content detectors on this dataset.
Hierarchical Pretraining on Multimodal Electronic Health Records
Xiaochen Wang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
CodeClassificationRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningTextMultimodalityTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Proposes a hierarchical multimodal pretraining framework called MEDHMP for pretraining on electronic health records (EHR) and downstream tasks;
🎯 What it does: Designed a multi-round conversational recommendation framework called HutCRS based on a hierarchical interest tree, which operates in two phases: first identifying the aspects of user interest, then asking about attributes related to these aspects or directly recommending items.
Identification of Multimodal Stance Towards Frames of Communication
Maxwell Weinzierl, Sanda Harabagiu
CodeClassificationData SynthesisGraph Neural NetworkTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposed the first multimodal stance annotation dataset MMVAX-STANCE for 113 COVID-19 vaccine communication frameworks, generated 46,000 synthetic multimodal instances by inferring text-image relationships, and constructed a stance detection model.
Identifying Informational Sources in News Articles
Alexander Spangher (University of Southern California), Jonathan May (University of Southern California)
CodeClassificationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the largest annotated news source dataset and proposed a source prediction task to study the combination patterns of sources in news.
CodeRecognitionRetrievalExplainability and InterpretabilityTransformerContrastive LearningTextBenchmark
🎯 What it does: Constructed and released a trafficker author attribution dataset named IDTraffickers, trained and evaluated author identification and verification models using this dataset to help law enforcement link potential traffickers in online prostitution ads.
Image Manipulation via Multi-Hop Instructions - A New Dataset and Weakly-Supervised Neuro-Symbolic Approach
Harman Singh (Indian Institute Of Technology Delhi), Parag Singla (Indian Institute Of Technology Delhi)
CodeGenerationExplainability and InterpretabilityGraph Neural NetworkVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: Propose a weakly supervised neuro-symbolic image editing framework called NEUROSIM, which can perform add, delete, and modify operations on multi-object images based on multi-step natural language instructions;
CodeData SynthesisTransformerLarge Language ModelText
🎯 What it does: Utilize constrained Beam Search (CBSQE) during NMT decoding to retain high-probability words from the reference, generating more accurate pseudo QE data;
Improved Unsupervised Chinese Word Segmentation Using Pre-trained Knowledge and Pseudo-labeling Transfer
Hsiu-Wen Li (National Cheng Kung University), Hung-Yu Kao (National Cheng Kung University)
CodeSegmentationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a Chinese word segmentation framework that utilizes an unsupervised segmentation model to generate pseudo labels and fine-tunes a pre-trained BERT model based on these pseudo labels, improving segmentation performance while significantly reducing training time.
Improving Chinese Pop Song and Hokkien Gezi Opera Singing Voice Synthesis by Enhancing Local Modeling
Peng Bai (Xiamen University), Xiaodong Shi (Xiamen University)
CodeGenerationData SynthesisTransformerAudio
🎯 What it does: Study the singing voice synthesis of Mandarin pop songs and Minnan opera, proposing two techniques to enhance local modeling, significantly improving the quality of synthesized audio.
🎯 What it does: The study proposes a multi-task framework that jointly learns dialogue discourse parsing and addressee recognition to leverage partially overlapping structural information between the two tasks for improved parsing performance.
Improving Image Captioning via Predicting Structured Concepts
Ting Wang (University of Science and Technology of China), Zhendong Mao (University of Washington)
CodeGenerationRepresentation LearningGraph Neural NetworkTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Propose a Structured Concept Predictor (SCP) that simultaneously predicts semantic concepts and their structures in images, and feeds the structured concepts along with visual features into a Transformer decoder to achieve end-to-end image caption generation.
Improving Language Models’ Meaning Understanding and Consistency by Learning Conceptual Roles from Dictionary
Myeongjun Jang (University of Oxford), Thomas Lukasiewicz (Vienna University of Technology)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes an intermediate pre-training task, Concept Role Modeling (CRM), learned through dictionary definitions, to enhance the semantic understanding capability of pre-trained language models (PLMs). This significantly reduces model inconsistencies across multiple consistency types (semantic, negational, symmetric, transitive) and achieves efficient fine-tuning of large PLMs via parameter fusion techniques.
Zonghai Yao (University of Massachusetts), Sai Selvaraj (Abridge AI)
CodeGenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: This paper proposes a method to improve summary quality by leveraging human editor feedback, combining sequence alignment with forward/backward likelihood training, with a focus on clinical dialogue summaries in the healthcare domain.
Improving Unsupervised Relation Extraction by Augmenting Diverse Sentence Pairs
Qing Wang (Iowa State University), Qi Li (Iowa State University)
CodeRepresentation LearningData-Centric LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Propose the AugURE method, improving relation representation learning in unsupervised relation extraction through diverse positive sample augmentation and margin loss.
Increasing Coverage and Precision of Textual Information in Multilingual Knowledge Graphs
Simone Conia (Sapienza University of Rome), Yunyao Li (Apple)
CodeLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes an automated knowledge graph embedding (KGE) method aimed at improving the coverage and accuracy of entity names and descriptions for non-English languages in Wikidata.
Shahbaz Syed (Leipzig University), Martin Potthast (Leipzig University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Propose an unsupervised method for long discussion indicative summarization, leveraging LLMs to first cluster sentences, generate cluster labels, and then assign argument frameworks to form a hierarchical summary resembling a table of contents.
Inference-Time Policy Adapters (IPA): Tailoring Extreme-Scale LMs without Fine-tuning
Ximing Lu (Allen Institute for Artificial Intelligence), Yejin Choi (Allen Institute for Artificial Intelligence)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelReinforcement LearningMixture of ExpertsText
🎯 What it does: Propose Inference-time Policy Adapters (IPA), which optimize the outputs of large language models (e.g., GPT-3) during inference using a lightweight adapter, avoiding the need for model fine-tuning.
INFORM : Information eNtropy based multi-step reasoning FOR large language Models
Chuyue Zhou (Soochow University), Min Zhang (Soochow University)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: Propose the INFORM framework, which uses information entropy to select problems, automatically generates chained reasoning steps, and enhances LLM reasoning performance through information entropy-based self-consistent reasoning;
Information Value: Measuring Utterance Predictability as Distance from Plausible Alternatives
Mario Giulianelli (University of Amsterdam), Raquel Fernández (University of Amsterdam)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes an 'information value' metric based on a complete sentence alternative set, which quantifies the predictability of a sentence relative to feasible alternative sentences that can be generated in a given context.
Instruct and Extract: Instruction Tuning for On-Demand Information Extraction
Yizhu Jiao (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextTabularBenchmarkChain-of-Thought
🎯 What it does: Proposed the 'on-demand information extraction' task and constructed the corresponding benchmark dataset INSTRUCTIE; based on this, trained an instruction-tuned model ODIE to extract structured tables from text according to user instructions.
Interpreting Embedding Spaces by Conceptualization
Adi Simhi (Technion Israel Institute of Technology), Shaul Markovitch (Technion Israel Institute of Technology)
CodeClassificationExplainability and InterpretabilityRepresentation LearningLarge Language ModelTextGraph
🎯 What it does: Proposes an algorithm called Concept Embedding Space (CES) that maps non-interpretable embedding spaces generated by LLMs to an interpretable concept space, enabling explanation and analysis of embeddings through the concept space.