ACL 2023 Papers — Page 11
Annual Meeting of the Association for Computational Linguistics · 1074 papers
Unsupervised Extractive Summarization of Emotion Triggers
Tiberiu Sosea (University of Illinois Chicago), Cornelia Caragea (University of Illinois Chicago)
GenerationGraph Neural NetworkTransformerTextBenchmark
🎯 What it does: Proposed an unsupervised emotion trigger extraction and summarization method, and constructed the COVIDET-EXT dataset.
Unsupervised Graph-Text Mutual Conversion with a Unified Pretrained Language Model
Yi Xu (Shanghai Jiao Tong University), Chenghu Zhou (IGSNRR Chinese Academy Of Sciences)
GenerationData SynthesisTransformerLarge Language ModelTextGraph
🎯 What it does: Proposed an unsupervised image-text bidirectional translation framework named INFINITY, which uses a single pre-trained seq2seq model to translate between images and text, and automatically generates synthetic parallel data through back-translation for joint training.
Unsupervised Melody-to-Lyrics Generation
Yufei Tian (University of California, Los Angeles), Nanyun Peng (University of California, Los Angeles)
GenerationTransformerLarge Language ModelTextAudio
🎯 What it does: Propose an unsupervised melody-to-lyrics generation framework that employs a hierarchical Plan-and-Write model, leveraging melody segmentation and rhythm alignment constraints during inference to guide text generation;
Unsupervised Open-domain Keyphrase Generation
Lam Thanh Do (Hanoi University of Science and Technology), Kevin Chen-Chuan Chang (University of Illinois at Urbana-Champaign)
GenerationTransformerMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Built an unsupervised, cross-domain keyphrase generation model that utilizes a retrieval-enhanced seq2seq framework and automatically generates core concept phrases through information assessment.
Unsupervised Selective Rationalization with Noise Injection
Adam Storek (Columbia University), Kathleen McKeown (Columbia University)
Explainability and InterpretabilityTransformerLarge Language ModelContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an unsupervised selective reasoning method, using online noise injection technology to suppress the generation of infeasible reasoning justifications, thereby enhancing the feasibility of model explanations and prediction accuracy.
Unsupervised Subtitle Segmentation with Masked Language Models
David Ponce (Vicomtech Foundation, Basque Research and Technology Alliance), Victor Ruiz (University of Basque Country)
SegmentationTransformerLarge Language ModelText
🎯 What it does: This paper proposes an unsupervised subtitle segmentation method that uses a pre-trained masked language model (MLM) to predict line endings and subtitle separation points based on the probability of punctuation occurrence.
UPPAM: A Unified Pre-training Architecture for Political Actor Modeling based on Language
Xinyi Mou (Fudan University), Xuanjing Huang (Fudan University)
Representation LearningTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed a unified pre-training architecture called UPPAM, which uses linguistic representation to model political actors and infuses multi-dimensional political context knowledge through structure-aware contrastive learning and behavior-driven contrastive learning;
Using contradictions improves question answering systems
Etienne Fortier-Dubois, Domenic Rosati (Dalhousie University)
TransformerLarge Language ModelText
🎯 What it does: Introduce contradiction signals from natural language inference (NLI) models into QA systems. After restating the QA model's answer as a statement, use NLI to score the answer and context across three categories (entailment, contradiction, neutral), and combine the scores with QA confidence for answer re-ranking and selection.
Using counterfactual contrast to improve compositional generalization for multi-step quantitative reasoning
Armineh Nourbakhsh (Carnegie Mellon University), Carolyn Rosé (Carnegie Mellon University)
Representation LearningAI Code AssistantRecurrent Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: Improve the combinatorial generalization capability of multi-step quantitative reasoning models by constructing contrastive samples and introducing auxiliary triplet loss.
Using Domain Knowledge to Guide Dialog Structure Induction via Neural Probabilistic Soft Logic
Connor Pryor (University of California Santa Cruz), Lise Getoor (University of California Santa Cruz)
Explainability and InterpretabilityRepresentation LearningRecurrent Neural NetworkText
🎯 What it does: This paper proposes a neural symbolic method called NEUPSL DSI, which utilizes probabilistic soft logic to inject domain knowledge and guide unsupervised learning of dialogue structure.
Using Neural Machine Translation for Generating Diverse Challenging Exercises for Language Learner
Frank Palma Gomez (CUNY), Alla Rozovskaya (CUNY)
GenerationTransformerLarge Language ModelText
🎯 What it does: Generate diverse and challenging fill-in-the-blank distractors for English learners automatically using back-translation neural machine translation technology.
USSA: A Unified Table Filling Scheme for Structured Sentiment Analysis
Zepeng Zhai (Beijing University of Posts and Telecommunications), Xiaojie Wang (Beijing University of Posts and Telecommunications)
ClassificationRecognitionGraph Neural NetworkTransformerTextTabular
🎯 What it does: This paper proposes a unified scheme for converting binary word pair dependency graphs into 2D table filling (USSA), and designs a fully end-to-end model based on this, achieving efficient processing of structured sentiment analysis (SSA).
UTC-IE: A Unified Token-pair Classification Architecture for Information Extraction
Hang Yan (Fudan University), Xipeng Qiu (Fudan University)
ClassificationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a unified token-pair classification architecture called UTC-IE, which converts information extraction tasks such as named entity recognition (NER), relation extraction (RE), and event extraction (EE) into a token-pair classification problem.
VendorLink: An NLP approach for Identifying & Linking Vendor Migrants & Potential Aliases on Darknet Markets
Vageesh Saxena (Maastricht University), Gerasimos Spanakis (Maastricht University)
ClassificationRecognitionDomain AdaptationKnowledge DistillationRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the VendorLink system, which identifies and links sellers migration and potential aliases in dark web markets through NLP writing style recognition, divided into three tasks: closed-set verification, open-set identification, and low-resource market adaptation;
Verify-and-Edit: A Knowledge-Enhanced Chain-of-Thought Framework
Ruochen Zhao (Nanyang Technological University), Lidong Bing (Alibaba Group)
Explainability and InterpretabilityTransformerTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose a Verify-and-Edit framework that performs post-editing on the generated reasoning chain during the Chain-of-Thought (CoT) inference process to improve the factual accuracy and correctness of answers.
Vision Language Pre-training by Contrastive Learning with Cross-Modal Similarity Regulation
Chaoya Jiang (Peking University), Shikun Zhang (Peking University)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Studies the problem of pseudo-negative samples in cross-modal contrastive learning, and proposes a cross-modal similarity-aware contrastive learning method based on mutual information theory.
Vision Meets Definitions: Unsupervised Visual Word Sense Disambiguation Incorporating Gloss Information
Sunjae Kwon (University of Massachusetts Amherst), Hong Yu (University of Massachusetts Lowell)
RetrievalRepresentation LearningLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: This paper proposes an unsupervised visual word sense disambiguation method that enhances image-text matching models through Bayesian inference using dictionary definitions.
VisText: A Benchmark for Semantically Rich Chart Captioning
Benny Tang (Massachusetts Institute of Technology), Arvind Satyanarayan (Massachusetts Institute of Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityGraphTabularBenchmark
🎯 What it does: Constructed the VisText dataset, providing 12,441 single-variable bar/line/area charts along with three representations (image, data table, scene graph), and generated L1 (structural information) and L2/L3 (statistical and cognitive features) descriptions for each chart; subsequently trained and evaluated three types of models (text translation, image guidance, and prefix tuning) on this dataset to automatically generate semantically rich chart descriptions.
Visually-augmented pretrained language models for NLP tasks without images
Hangyu Guo (Harbin Institute of Technology (Shenzhen)), Ji-Rong Wen (Renmin University of China)
Representation LearningTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelText
🎯 What it does: Propose a visual augmentation fine-tuning method called VAWI, which does not require image retrieval or generation. It first automatically identifies visual hunger words in text through three strategies (syntax, attention, and learnable), then generates visually aligned representations using the CLIP text encoder. These representations are refined through a position-aware reconstruction layer to obtain visually enhanced representations, which are finally injected into PLMs (supporting both full-parameter or parameter-efficient fine-tuning) to improve performance across various NLP tasks.
VLN-Trans: Translator for the Vision and Language Navigation Agent
Yue Zhang (Michigan State University), Parisa Kordjamshidi (Michigan State University)
Recurrent Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a translation module that converts complete instructions into recognizable and discriminative sub-instructions to enhance visual language navigation performance.
VSTAR: A Video-grounded Dialogue Dataset for Situated Semantic Understanding with Scene and Topic Transitions
Yuxuan Wang (Peking University), Dongyan Zhao (Peking University)
SegmentationGenerationConvolutional Neural NetworkTransformerLarge Language ModelVision Language ModelVideoTextMultimodalityBenchmark
🎯 What it does: Constructed the VSTAR dataset based on 395 TV series, annotated scene and topic boundaries in video-dialogue pairs, and designed three benchmark tasks: video scene segmentation, topic segmentation, and video dialog generation.
WACO: Word-Aligned Contrastive Learning for Speech Translation
Siqi Ouyang (University of California Santa Barbara), Lei Li (University of California Santa Barbara)
TransformerContrastive LearningTextMultimodalityAudio
🎯 What it does: Proposed a word-level alignment contrastive learning method called WACO, which directly brings speech and corresponding text word representations closer, enhancing speech translation performance under extremely low-resource conditions.
We Understand Elliptical Sentences, and Language Models should Too: A New Dataset for Studying Ellipsis and its Interaction with Thematic Fit
Davide Testa (University of Pisa), Alessandro Lenci (University of Pisa)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Studied the ability of language models (GPT-2 and BERT) to process elliptical sentences and explored the impact of typicality of event participants (topic adaptation) on the parsing of elliptical sentences.
Weaker Than You Think: A Critical Look at Weakly Supervised Learning
Dawei Zhu (Saarland University), Dietrich Klakow (Saarland University)
ClassificationMeta LearningTransformerSupervised Fine-TuningTextReview/Survey Paper
🎯 What it does: Systematically evaluate the practical effectiveness of weakly supervised learning (WSL) methods in low-resource scenarios, revealing that their benefits are overestimated and that direct fine-tuning on very few clean samples achieves better performance.
Weakly Supervised Vision-and-Language Pre-training with Relative Representations
Chi Chen (Tsinghua University), Yang Liu (Tsinghua University)
Representation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextRetrieval-Augmented Generation
🎯 What it does: Propose a weakly supervised vision-language pre-training framework named RELIT based on relative representations, which constructs weakly aligned data by using a small number of image-text pairs as anchors to complete two types of weakly aligned image-text pairs: retrieval and generation.
Weakly-Supervised Spoken Video Grounding via Semantic Interaction Learning
Ye Wang (Zhejiang University), Zhou Zhao (Zhejiang University)
RetrievalRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningVideoMultimodalityAudio
🎯 What it does: Studied the weakly supervised audio-visual localization task and proposed the SIL framework to achieve audio-visual localization without time labels.
WebCPM: Interactive Web Search for Chinese Long-form Question Answering
Yujia Qin (Tsinghua University), Jie Zhou (Tencent Inc)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: An interactive Web search interface based on Bing was constructed, and this interface was used to collect 5.5k Chinese long-form question-answer instances (including supporting facts and search behaviors), thereby creating the WebCPM dataset. Based on this dataset, the authors fine-tuned large-scale pre-trained language models (CPM, mT5, mBART, etc.) to learn four subtasks: action prediction, query generation, fact extraction, and answer synthesis, ultimately forming an end-to-end LFQA pipeline.
WebIE: Faithful and Robust Information Extraction on the Web
Chenxi Whitehouse (City, University of London), Andrea Pierleoni (Amazon Alexa AI)
GenerationTransformerPrompt EngineeringTextBenchmark
🎯 What it does: This paper introduces the WEBIE dataset, which automatically extracts 1.6M sentences from Common Crawl web text and incorporates negative samples, providing 21K triplets with multilingual translations based on human annotations, aiming to enhance the robustness and generalization capability of closed information extraction.
WeCheck: Strong Factual Consistency Checker via Weakly Supervised Learning
Wenhao Wu (Peking University), Yajuan Lyu (Baidu Inc)
Anomaly DetectionData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposed WeCheck, a fact consistency checking framework based on weakly supervised learning, which trains evaluation models directly on real generated text.
What about “em”? How Commercial Machine Translation Fails to Handle (Neo-)Pronouns
Anne Lauscher (University of Hamburg), Dirk Hovy
Text
🎯 What it does: Evaluate the quality of commercial machine translation systems (Google Translate, Bing, DeepL) in handling English third-person pronouns (including gendered, gender-neutral, traditional neopronouns, nounself, emoji-self, numberself, and other new pronouns), combining manual annotations of grammatical correctness, semantic consistency, and pronoun handling strategies, and conducting surveys to assess the preferences of affected groups regarding pronoun translations.
What are the Desired Characteristics of Calibration Sets? Identifying Correlates on Long Form Scientific Summarization
Griffin Adams (Columbia University), Noémie Elhadad (Columbia University)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTextBiomedical DataElectronic Health Records
🎯 What it does: This paper investigates effective features of calibration sets in long scientific abstracts and systematically evaluates the impact of different candidate set construction strategies on relevance and faithfulness.
What Are You Token About? Dense Retrieval as Distributions Over the Vocabulary
Ori Ram (Tel Aviv University), Amir Globerson (Tel Aviv University)
RetrievalExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningText
🎯 What it does: By projecting the query and passage vectors of dense retrieval models into a lexical space, the internal representations of the model are explained and the 'token amnesia' problem is revealed;
What Do NLP Researchers Believe? Results of the NLP Community Metasurvey
Julian Michael (New York University), Samuel R. Bowman (New York University)
TextReview/Survey Paper
🎯 What it does: Conducted a large-scale community meta-survey targeting NLP researchers, systematically collecting views and sociological beliefs on controversial issues such as scaling, AGI, and ethics.
What does a Text Classifier Learn about Morality? An Explainable Method for Cross-Domain Comparison of Moral Rhetoric
Enrico Liscio (Tudelft), Pradeep K. Murukannaiah (Tudelft)
ClassificationDomain AdaptationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Proposes the Tomea method, which uses the SHAP interpreter to generate domain-specific moral lexicons, and compares how text classifiers represent moral discourse across different domains.
What does the Failure to Reason with “Respectively” in Zero/Few-Shot Settings Tell Us about Language Models?
Ruixiang Cui (University of Copenhagen), Anders Søgaard (University of Copenhagen)
ClassificationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper systematically investigates the reasoning ability of language models in zero/one-shot scenarios for the 'respectively' structure, constructs synthetic dataset WikiResNLI and natural dataset NatResNLI, and evaluates the performance of multiple Transformer language models on explicit vs. implicit, synthetic vs. natural corpora.
What Is Overlap Knowledge in Event Argument Extraction? APE: A Cross-datasets Transfer Learning Model for EAE
Kaihang Zhang (Beijing University of Posts and Telecommunications), Jinyu Guo (Beijing University of Posts and Telecommunications)
TransformerSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes an event argument extraction model APE based on cross-dataset knowledge transfer. It first learns overlapping knowledge across datasets through a pseudo entity recognition (PER) task, then uses a dedicated Adapter to learn dataset-specific knowledge, and activates overlapping knowledge in both stages of training using prompts of the same style.
What is the best recipe for character-level encoder-only modelling?
Kris Cao (DeepMind)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: This study conducts a unified evaluation of various character-level encoder models to identify the optimal training strategy and compare it with subword-level BERT.
What is the Real Intention behind this Question? Dataset Collection and Intention Classification
Maryam Sadat Mirzaei (RIKEN Center for Advanced Intelligence Project (AIP)), Satoshi Sekine (RIKEN Center for Advanced Intelligence Project (AIP))
ClassificationTransformerLarge Language ModelTextBenchmark
🎯 What it does: Investigated the positive and negative intentions behind questions, constructed the 'Question Intention Dataset,' and performed classification and evaluation of question intentions.
What Makes Pre-trained Language Models Better Zero-shot Learners?
Jinghui Lu (SenseTime Research), Fei Tan (SenseTime Research)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose Perplection, which uses perplexity to filter prompt templates for zero-shot text classification, avoiding the use of labeled development sets;
What social attitudes about gender does BERT encode? Leveraging insights from psycholinguistics
Julia Watson (University of Toronto), Suzanne Stevenson (University of Toronto)
Representation LearningTransformerLarge Language ModelText
🎯 What it does: Evaluate BERT's social attitudes in gender-related language choices by directly linking language selections from human psycholinguistic experiments and social attitude questionnaire results with model predictions.
What the DAAM: Interpreting Stable Diffusion Using Cross Attention
Raphael Tang (Comcast Applied AI), Ferhan Ture (Comcast Applied AI)
Explainability and InterpretabilityTransformerDiffusion modelImageText
🎯 What it does: This paper proposes the DAAM method to generate word-pixel attribution maps by aggregating cross-attention weights from Stable Diffusion and performing bilinear interpolation and thresholding.
What’s the Meaning of Superhuman Performance in Today’s NLU?
Simone Tedeschi (Babelscape), Roberto Navigli (Sapienza University of Rome)
Explainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelContrastive LearningTextReview/Survey PaperBenchmark
🎯 What it does: This paper critically evaluates claims of 'superhuman performance' in the current NLP field, systematically organizing issues such as evaluation design flaws, data bias, improper measurement methods, and insufficient quality and diversity of human annotations in mainstream NLU benchmarks (e.g., SuperGLUE, SQuAD, RACE, XTREME, etc.).
When and how to paraphrase for named entity recognition?
Saket Sharma (JPMorgan Chase & Co.), Sashank Santhanam (Apple)
RecognitionData SynthesisTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper investigates the effectiveness of using different paraphrasing methods for data augmentation in named entity recognition tasks, systematically evaluates six paraphrasers, and proposes a simple span-level annotation strategy, exploring the relationship between paraphrasing intensity, data scarcity, and model performance;
When Does Aggregating Multiple Skills with Multi-Task Learning Work? A Case Study in Financial NLP
Jingwei Ni (ETH Zürich), Markus Leippold (University of Zürich)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
🎯 What it does: This paper uses financial NLP as a case study to systematically evaluate the effectiveness of multi-task learning (MTL) in aggregating diverse skills, and proposes a parameter-efficient SPAL-FinBERT architecture;
When Does Translation Require Context? A Data-driven, Multilingual Exploration
Patrick Fernandes (Carnegie Mellon University), Graham Neubig (Carnegie Mellon University)
Data-Centric LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose the MuDA benchmark, which uses P-CXMI to identify context-dependent words in translation and automatically annotate five discourse phenomena, including lexical coherence, politeness forms, pronoun selection, verb forms, and ellipsis, constructing evaluation data for 14 language pairs.
When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
Alex Mallen (University of Washington), Hannaneh Hajishirzi (University of Washington)
RetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Investigate the capabilities and limitations of large language models in entity knowledge memory, construct a long-tail entity QA dataset POPQA, evaluate the performance of multiple large models in zero/few-shot QA, explore the complement of retrieval-enhanced methods to model memory, and propose an adaptive retrieval method.
When to Use Efficient Self Attention? Profiling Text, Speech and Image Transformer Variants
Anuj Diwan (University of Texas at Austin), David Harwath (University of Texas at Austin)
Computational EfficiencyTransformerImageTextMultimodalityBenchmarkAudio
🎯 What it does: Conduct unified experiments on Transformers and their efficient variants of self-attention across text, speech, and visual modalities, evaluating multiple efficiency metrics and determining when to use them.
When to Use What: An In-Depth Comparative Empirical Analysis of OpenIE Systems for Downstream Applications
Kevin Pei (University of Illinois at Urbana-Champaign), Yunyao Li (Apple)
Recurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper provides an application-oriented empirical comparison for the OpenIE task, systematically evaluates different neural models, training sets, and benchmarks, and offers model selection recommendations for various downstream tasks.
Where’s the Point? Self-Supervised Multilingual Punctuation-Agnostic Sentence Segmentation
Benjamin Minixhofer (Cohere for AI), Ivan Vulić (University of Cambridge)
SegmentationTransformerSupervised Fine-TuningContrastive LearningText
🎯 What it does: Proposed a multi-lingual, punctuation-free self-supervised sentence segmentation method (WtP), which uses paragraph line breaks as implicit markers and trains a character-level language model to predict line break positions, thereby achieving sentence segmentation.
White-Box Multi-Objective Adversarial Attack on Dialogue Generation
Yufei Li (University of California, Riverside), Cong Liu (University of California, Riverside)
GenerationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Propose a white-box multi-objective adversarial attack method called DGSlow for dialogue generation models, which can force the generation models to output longer and more irrelevant responses by replacing a few words while maintaining semantic similarity.
WhitenedCSE: Whitening-based Contrastive Learning of Sentence Embeddings
Wenjie Zhuo (Zhejiang University), Yi Yang (Zhejiang University)
Representation LearningTransformerContrastive LearningText
🎯 What it does: Propose WhitenedCSE, a sentence embedding framework that integrates whitening techniques into contrastive learning;
Why Aren’t We NER Yet? Artifacts of ASR Errors in Named Entity Recognition in Spontaneous Speech Transcripts
Piotr Szymański (Wroclaw University of Science and Technology), Piotr Żelasko (Meaning.Team Inc)
ClassificationTransformerLarge Language ModelTextFinance RelatedAudio
🎯 What it does: Systematic evaluation of named entity recognition (NER) in spontaneous speech transcription, analyzing and summarizing the interactive effects between ASR errors and NER errors, and proposing a comprehensive error classification system.
Why Did the Chicken Cross the Road? Rephrasing and Analyzing Ambiguous Questions in VQA
Elias Stengel-Eskin (Johns Hopkins University), Benjamin Van Durme (Johns Hopkins University)
GenerationTransformerVision Language ModelMultimodality
🎯 What it does: Created and annotated a specialized corpus for studying linguistic ambiguity in visual question answering (VQA), and developed a visual question generation model to rewrite and eliminate ambiguity in original questions.
WikiBio: a Semantic Resource for the Intersectional Analysis of Biographical Events
Marco Antonio Stranisci (Università degli Studi di Torino), Tommaso Caselli (University of Groningen)
ClassificationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Construct the WikiBio corpus for biomedical event detection and use it to assess gender and ethnic biases in Wikipedia.
WikiHowQA: A Comprehensive Benchmark for Multi-Document Non-Factoid Question Answering
Valeriia Bolotova-Baranova (Rmit University), Mark Sanderson (Rmit University)
GenerationRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Constructed the WikiHowQA multi-document non-factual question answering (MD-NFQA) benchmark, collecting 11,746 'How to' questions, corresponding human answers, and 74,527 supporting documents, focusing on the generation and evaluation of multi-document non-factual question answering;
WinoQueer: A Community-in-the-Loop Benchmark for Anti-LGBTQ+ Bias in Large Language Models
Virginia Felkner, Jonathan May (Information Sciences Institute University of Southern California)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Construct the WinoQueer benchmark through community surveys to evaluate negative biases of large language models toward the LGBTQ+ community;
With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness
Julius Steen (Heidelberg University), Katja Markert (Heidelberg University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerTextBenchmark
🎯 What it does: To address the issue of NLI models misjudging the authenticity of generated text, this study significantly improves faithfulness prediction on the TRUE benchmark through task-adaptive data augmentation, leveraging contradiction scores, and employing Monte Carlo Dropout.
Won’t Get Fooled Again: Answering Questions with False Premises
Shengding Hu (Tsinghua University), Maosong Sun (Tsinghua University)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the robustness of pre-trained language models in answering questions with false premises (FPQ) and proposes a method to activate the model's rebuttal capability;
Word sense extension
Lei Yu, Yang Xu (University of Toronto)
Representation LearningMeta LearningTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: Proposed the Word Sense Extension (WSE) paradigm, simulating how language users create new meanings for words in novel contexts, and constructed a training framework based on semantic partitioning;
World-to-Words: Grounded Open Vocabulary Acquisition through Fast Mapping in Vision-Language Models
Ziqiao Ma (University of Michigan), Joyce Chai (University of Michigan)
Object DetectionTransformerVision Language ModelMultimodality
🎯 What it does: By introducing a word-region alignment objective into visual language models, the study explores learning word semantics from image-text pairs and achieves fast mapping of new words in few-shot scenarios.
WSPAlign: Word Alignment Pre-training via Large-Scale Weakly Supervised Span Prediction
Qiyu Wu (University of Tokyo), Yoshimasa Tsuruoka (University of Tokyo)
Representation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose a weakly supervised span prediction-based word alignment pre-training method called WSPAlign.
Wukong-Reader: Multi-modal Pre-training for Fine-grained Visual Document Understanding
Haoli Bai (Huawei Technologies), Qun Liu (Huawei Technologies)
ClassificationRecognitionTransformerVision Language ModelContrastive LearningMultimodality
🎯 What it does: Proposes WUKONG-READER, a multimodal pre-training model for fine-grained visual document understanding.
XDailyDialog: A Multilingual Parallel Dialogue Corpus
Zeming Liu (Harbin Institute of Technology), Kaiping Peng (Tsinghua University)
GenerationData SynthesisRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper constructs the first publicly available multilingual parallel open-domain dialogue corpus, XDailyDialog, and proposes three dialogue tasks: multilingual, cross-lingual, and multilingual mixed.
XL-LEXEME: WiC Pretrained Model for Cross-Lingual LEXical sEMantic changE
Pierluigi Cassotti (University of Bari Aldo Moro), Pierpaolo Basile (University of Bari Aldo Moro)
TransformerLarge Language ModelContrastive LearningText
🎯 What it does: Proposed and implemented the XL-LEXEME model for cross-lingual lexical change detection;
XSemPLR: Cross-Lingual Semantic Parsing in Multiple Natural Languages and Meaning Representations
Yusen Zhang (Penn State University), Rui Zhang (Penn State University)
Representation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose XSEMPLR, a unified cross-lingual semantic parsing benchmark covering 22 natural languages and 8 meaning representations, systematically evaluating multilingual model performance.
xSIM++: An Improved Proxy to Bitext Mining Performance for Low-Resource Languages
Mingda Chen (Meta AI Research), Holger Schwenk (Meta AI Research)
Representation LearningData-Centric LearningTransformerContrastive LearningText
🎯 What it does: Proposed and implemented an improved xsim++ proxy metric for evaluating the quality of bitext phrase pair mining in low-resource languages, validated through extensive mining experiments and NMT training.
Your spouse needs professional help: Determining the Contextual Appropriateness of Messages through Modeling Social Relationships
David Jurgens (University of Michigan), Michael Geraci (University of Buffalo)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: By constructing a relational context appropriateness dataset containing 12,236 annotated examples, and training large language models to determine whether messages are appropriate in different social relationship contexts; subsequently evaluating the model's applicability in real conversations and exploring its predictive capability for subtle offensive language.
Z-Code++: A Pre-trained Language Model Optimized for Abstractive Summarization
Pengcheng He (Microsoft Azure AI), Xuedong Huang (Microsoft Azure AI)
GenerationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Develop and release Z-Code++, a pre-trained language model for abstractive text summarization.
Z-ICL: Zero-Shot In-Context Learning with Pseudo-Demonstrations
Xinxi Lyu (University of Washington), Hannaneh Hajishirzi (University of Washington)
ClassificationTransformerPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
🎯 What it does: Propose Z-ICL, a zero-shot context learning method that retrieves pseudo demonstrations from the raw text corpus using synonym labels to enhance text classification performance.
Zero- and Few-Shot Event Detection via Prompt-Based Meta Learning
Zhenrui Yue (University of Illinois Urbana Champaign), Dong Wang (University of Illinois Urbana Champaign)
ClassificationMeta LearningTransformerPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes the MetaEvent framework for zero-shot and few-shot event detection, enabling the model to quickly adapt to unknown event types through meta-learning.
Zero-Shot and Few-Shot Stance Detection on Varied Topics via Conditional Generation
Haoyang Wen (Carnegie Mellon University), Alexander Hauptmann
ClassificationGenerationTransformerTextRetrieval-Augmented Generation
🎯 What it does: Propose a zero-shot and few-shot stance detection framework based on conditional generation, which uses templates to denoise and generate complete sentences for stance determination;
Zero-shot Approach to Overcome Perturbation Sensitivity of Prompts
Mohna Chakraborty (Iowa State University), Qi Li (Iowa State University)
ClassificationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a prompt enhancement and unsupervised ranking method under zero-shot settings to address the sensitivity of prompts to language model predictions.
Zero-shot Cross-lingual Transfer With Learned Projections Using Unlabeled Target-Language Data
Ujan Deb (IIT Bhilai), Preethi Jyothi (IIT Bombay)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: In zero-shot cross-lingual transfer with a known target language, construct a language subspace using unlabeled target language text, and during task fine-tuning, project source language representations into the target subspace to improve cross-lingual performance.
Zero-shot Faithful Factual Error Correction
Kung-Hsiang Huang (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
Domain AdaptationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataRetrieval-Augmented Generation
🎯 What it does: Proposes a zero-shot fact error correction framework ZEROFEC, which corrects factual errors under given evidence using question answering and document-level entailment assessment, while maintaining faithfulness to the original sentence.