π― What it does: This paper proposes the rhetorical parallel detection task, constructs two new datasets: the Latin sermons dataset ASP and the Chinese student essays collection PSE-I, designs evaluation metrics, and implements multiple baseline models.
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Evaluate the performance of larger PaLM language models on Inverse Scaling Prize tasks, finding that inverse scaling often turns into U-shaped scaling, and explore whether 1-shot demonstrations and chain-of-thought prompts can alleviate this issue.
KCTS: Knowledge-Constrained Tree Search Decoding with Token-Level Hallucination Detection
Sehyun Choi (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
CodeGenerationTransformerLarge Language ModelText
π― What it does: Propose a method that, without altering the weights of large language models, uses knowledge-constrained tree search (KCTS) and token-level hallucination detection (RIPA) to guide text generation, making it more aligned with given knowledge and reducing hallucinations.
π― What it does: This paper investigates the performance of retrieval-based interpolation language models (kNN-LM) in open-ended text generation, demonstrating that although they significantly reduce perplexity, they do not improve generation quality.
Knowledge Graph Compression Enhances Diverse Commonsense Generation
EunJeong Hwang (University of British Columbia), Tengfei Ma (Stony Brook University)
CodeGenerationCompressionGraph Neural NetworkTransformerLarge Language ModelMixture of ExpertsTextGraph
π― What it does: Utilize a differentiable graph compression algorithm to automatically select the most task-relevant concepts in the subgraph of the Commonsense Knowledge Graph (ConceptNet), and enhance the diversity and quality of generative models using the compressed subgraph.
Knowledge Rumination for Pre-trained Language Models
Yunzhi Yao (Zhejiang University), Ningyu Zhang (Zhejiang University)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose the Knowledge Rumination method, which leverages the hidden knowledge within pre-trained language models to improve the performance of knowledge-intensive tasks.
Jinheon Baek (Korea Advanced Institute Of Science And Technology), Sung Hwang
CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: Designed and evaluated a validator called KALMV, capable of detecting knowledge retrieval errors and generation errors, and enhancing the reliability of knowledge-augmented language models through multi-instruction integration and iterative error correction during reasoning.
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Propose the KRLS algorithm, combining offline reinforcement learning, next-word sampling, and fine-grained rewards to improve end-to-end response generation in task-oriented dialogue.
Language and Mental Health: Measures of Emotion Dynamics from Text as Linguistic Biosocial Markers
Daniela Teodorescu (University of Alberta), Saif Mohammad (National Research Council Canada)
CodeText
π― What it does: Investigated the relationship between the Utterance Emotion Dynamics (UED) metric in tweets and self-reported mental health diagnoses (e.g., ADHD, depression, bipolar disorder, PTSD), calculating indicators such as average emotion, emotional variability, rising rate, and recovery rate, and comparing them with a control group.
Language Model is Suitable for Correction of Handwritten Mathematical Expressions Recognition
Zui Chen (Shanghaitech University), Yi Zhou (Shanghaitech University)
CodeRecognitionConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelVision Language ModelImage
π― What it does: Proposed a Recognition and Language Fusion Network (RLFN) that jointly corrects errors in handwritten mathematical expression recognition by leveraging the MathBERTa language model and a visual recognition module.
Large Language Models and Multimodal Retrieval for Visual Word Sense Disambiguation
Anastasia Kritharoula (National Technical University of Athens), Giorgos Stamou (National Technical University of Athens)
CodeRetrievalTransformerLarge Language ModelVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Studied the Visual Word Sense Disambiguation (VWSD) task, proposing multiple methods such as vision/text retrieval, LLM knowledge enhancement, single-modal retrieval, learning to rank, and QA + chain-of-thought to improve retrieval accuracy.
Large Language Models are biased to overestimate profoundness
Eugenio Herrera-Berg (Centro Nacional de Inteligencia Artificial), Cristian Buc Calderon (Centro Nacional de Inteligencia Artificial)
CodeExplainability and InterpretabilityData-Centric LearningReinforcement Learning from Human FeedbackTransformerPrompt EngineeringTextChain-of-Thought
π― What it does: Evaluate the bias of GPT-4 and other LLMs in judging the depth of bland, motivational, and pseudo-profound sentences, and explore the impact of different prompting strategies (few-shot, chain-of-thought) on the results.
Larger Probes Tell a Different Story: Extending Psycholinguistic Datasets Via In-Context Learning
Namrata Shivagunde (University of Massachusetts Lowell), Anna Rumshisky (University of Massachusetts Lowell)
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: Extended the original psycholinguistic evaluation datasets NEG-136-SIMP and ROLE-88 by using templates and GPT-3 to generate larger-scale datasets, NEG-1500-SIMP (750 pairs each via two methods) and ROLE-1500 (750 pairs), and conducted zero-shot evaluations on 22 language models.
π― What it does: Researchers propose a multimodal model based on graph neural networks that uses voice and gesture resonance information for Aphasia type detection.
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
Danqing Wang (University of California Santa Barbara), Lei Li (Carnegie Mellon University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the SALAM framework, enabling large language models to interact with auxiliary learning assistants during the error collection phase, improving model performance through error memory retrieval and feedback guidance.
Learning Knowledge-Enhanced Contextual Language Representations for Domain Natural Language Understanding
Taolin Zhang (East China Normal University), Weining Qian (Alibaba Group)
CodeClassificationRecognitionRepresentation LearningTransformerLarge Language ModelContrastive LearningGraphBiomedical DataFinance Related
π― What it does: Proposes the KANGAROO framework, which enhances the knowledge representation of language models in closed-domain knowledge graphs using hyper-sphere embedding and multi-layer contrastive learning based on point-biconnected components.
CodeRecognitionData SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Using instruction-tuned large language models (Alpaca) to automatically generate synthetic context retrieval training sets, and training a BERT-based neural retriever to retrieve relevant contexts for named entity recognition (NER) in long texts.
Letβs Think Frame by Frame with VIP: A Video Infilling and Prediction Dataset for Evaluating Video Chain-of-Thought
Vaishnavi Himakunthala (University of California Santa Barbara), William Wang (University of California Santa Barbara)
CodeLarge Language ModelVideoTextBenchmarkChain-of-Thought
π― What it does: Propose the VIP dataset, which includes unstructured dense captions and structured FAMOUS descriptions for video keyframes, and define two tasks, video filling and prediction, to evaluate multi-frame reasoning capabilities.
CodeClassificationRecognitionTransformerLarge Language ModelTextBenchmark
π― What it does: Constructed the MCS-350 parallel children's story corpus with over 50K entries, and proposed the LIMIT hierarchical model for language identification and machine translation in low-resource languages.
LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers
Theo Olausson (MIT), Roger Levy (MIT)
CodeTransformerLarge Language ModelTextChain-of-Thought
π― What it does: Propose a neural-symbolic framework called LINC, which uses a large language model as a semantic parser to convert natural language premises into first-order logic expressions, then passes them to an external theorem prover Prover9 for symbolic reasoning, and obtains the final conclusion through majority voting.
Linear-Time Modeling of Linguistic Structure: An Order-Theoretic Perspective
Tianyu Liu (ETH Zurich), Ryan Cotterell (ETH Zurich)
CodeComputational EfficiencyText
π― What it does: This paper proposes a linear-time structural prediction framework based on partial order theory, modeling the relationship between word pairs in a sentence as partial orders, and recovering dependency trees and coreference networks by predicting the intersection of several total orders.
Ling-CL: Understanding NLP Models through Linguistic Curricula
Mohamed Elgaar (University of Massachusetts Lowell), Hadi Amiri (University of Massachusetts Lowell)
CodeOptimizationTransformerTextBenchmark
π― What it does: Propose a multi-perspective curriculum learning framework based on human-verified language complexity indices. By dynamically learning the importance weights of indices and aggregating them into difficulty scores, the framework designs time-varying sigmoid, negative sigmoid, and Gaussian weight curves for NLP model training. This enables gradual introduction of samples according to difficulty during training, revealing the core linguistic knowledge mastered by the model during learning.
π― What it does: Proposes an open information extraction (OIE) fact linking task for large-scale knowledge graphs (KGs), and constructs a new multi-dimensional benchmark FaLB covering transductive, inductive, polysemous, and scenarios where OIE fragments are not in KG.
Shih-yang Liu (Hong Kong University of Science and Technology), Kwang-Ting Cheng (Hong Kong University of Science and Technology)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Proposes LLM-FP4, a quantization method that can compress large language models (LLMs) and their activations, weights to 4-bit floating point during the post-training phase.
LLM-powered Data Augmentation for Enhanced Cross-lingual Performance
Chenxi Whitehouse (City University of London), Alham Fikri Aji (MBZUAI)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Generate synthetic training samples for multilingual common-sense reasoning using large language models (Dolly-v2, StableVicuna, ChatGPT, GPT-4), and fine-tune small multilingual models (mBERT, XLM-R) with these samples to improve cross-lingual performance.
Localizing Active Objects from Egocentric Vision with Symbolic World Knowledge
Te-Lin Wu (University of California Los Angeles), Nanyun Peng (University of California Los Angeles)
CodeObject DetectionObject TrackingTransformerLarge Language ModelVision-Language-Action ModelContrastive LearningVideoText
π― What it does: Proposes an active object localization and tracking framework based on text instructions and symbolic knowledge, capable of identifying and tracking key objects (OUC) and tools (Tool) with state changes in the perspective camera view
Log-FGAER: Logic-Guided Fine-Grained Address Entity Recognition from Multi-Turn Spoken Dialogue
Xue Han (China Mobile Research Institute), Junlan Feng (China Mobile Research Institute)
CodeRecognitionData-Centric LearningTransformerLarge Language ModelText
π― What it does: This paper studies the recognition of fine-grained address entities in multi-turn speech dialogues, and proposes a logic-based PSL regularization method called Log-FGAER, as well as an ontology-driven data augmentation framework leveraging ChatGPT to generate low-cost annotated dialogue data.
Lost in Translation, Found in Spans: Identifying Claims in Multilingual Social Media
Shubham Mittal (Mohammed Bin Zayed University Of Artificial Intelligence), Preslav Nakov (Mohammed Bin Zayed University Of Artificial Intelligence)
CodeRecognitionTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper constructs a multilingual claim span identification (mCSI) dataset named X-CLAIM, proposes an automated annotation pipeline, and conducts experiments on six languages (English, Hindi, Punjabi, Tamil, Telugu, and Bengali).
MAGNIFICo: Evaluating the In-Context Learning Ability of Large Language Models to Generalize to Novel Interpretations
Arkil Patel (Mila and McGill University), Dzmitry Bahdanau (Mila and McGill University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper proposes the MAGNIFICO evaluation framework to systematically assess the ability of large language models to acquire and generalize new word meanings in context learning;
Dong-Ho Lee (University of Southern California), Sujay Jauhar
CodeData SynthesisDomain AdaptationData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Proposed a unified single-sample formatted example generation pipeline that leverages instruction-following large language models (LLMs) to generate diverse, structured annotated data under given JSON structure prompts, and continuously expands the training set through a self-reference strategy.
MeaeQ: Mount Model Extraction Attacks with Efficient Queries
Chengwei Dai (Chinese Academy of Sciences), Wei Zhou (Chinese Academy of Sciences)
CodeAdversarial AttackTransformerText
π― What it does: This paper studies NLP model extraction attacks, proposing the MeaeQ method, which achieves efficient querying through task-related filtering and clustering dimensionality reduction, thereby constructing a pirated model with performance comparable to the victim model.
π― What it does: Propose a domain-invariant representation learning method based on memory networks, leveraging meta-learning in multi-source domain training to learn memory gains and enhance text classification performance on unknown target domains.
Merging Experts into One: Improving Computational Efficiency of Mixture of Experts
Shwai He (University of Maryland), Dacheng Tao (University of Sydney)
CodeComputational EfficiencyTransformerMixture of ExpertsText
π― What it does: Propose an MEO (Merging Experts into One) method that merges multiple expert parameters into a single expert, preserving the representational diversity of multiple experts in the Mixture of Experts model while significantly reducing computational costs.
Merging Generated and Retrieved Knowledge for Open-Domain QA
Yunxiang Zhang (University of Michigan), Lu Wang (University of Illinois at Chicago)
CodeGenerationRetrievalRepresentation LearningTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: This paper proposes the COMBO framework, which improves the accuracy of open-domain question answering by pairing retrieved text with LLM-generated text through compatibility matching.
MILDSum: A Novel Benchmark Dataset for Multilingual Summarization of Indian Legal Case Judgments
Debtanu Datta (Indian Institute of Technology Kharagpur), Saptarshi Ghosh (Indian Institute of Technology Kharagpur)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Constructed the MILDSum dataset, containing 3,122 English legal judgments along with their English and Hindi summaries, and conducted benchmark evaluations of multiple summarization and translation methods
Mirror: A Universal Framework for Various Information Extraction Tasks
Tong Zhu (Soochow University), Min Zhang (Soochow University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed the Mirror framework, unifying various information extraction tasks into multi-slot tuples and mapping them to multi-span cyclic graphs, employing non-autoregressive graph decoding to achieve one-time extraction of all spans.
Mitigating Over-Generation for Unsupervised Keyphrase Extraction with Heterogeneous Centrality Detection
Mingyang Song (Beijing Jiaotong University), Liping Jing (Beijing Jiaotong University)
CodeRepresentation LearningGraph Neural NetworkLarge Language ModelTextBenchmark
π― What it does: This paper proposes an unsupervised keyword extraction method called CentralityRank, which evaluates the importance of candidate phrases by leveraging implicit and explicit centrality in heterogeneous graphs.
MMNMT: Modularizing Multilingual Neural Machine Translation with Flexibly Assembled MoE and Dense Blocks
Shangjie Li (Tianjin University), Deyi Xiong (Tianjin University)
CodeTransformerMixture of ExpertsText
π― What it does: Propose the MMNMT framework, flexibly assembling dense and Mixture-of-Experts (MoE) modules to achieve efficient capacity and low-resource robustness for multilingual translation;
π― What it does: Proposes using prompt tuning (MVP) as an alternative to traditional MLP head fine-tuning methods to achieve higher adversarial robustness in downstream classification tasks
More Than Spoken Words: Nonverbal Message Extraction and Generation
Dian Yu (Tencent AI Lab), Dong Yu (Tencent AI Lab)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study proposes and implements the task of extracting non-linguistic information (NM) from Chinese novel texts, as well as automatically generating NM for sentences in dialogue generation.
MoT: Memory-of-Thought Enables ChatGPT to Self-Improve
Xiaonan Li (Fudan University), Xipeng Qiu (Fudan University)
CodeRetrievalExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the MoT (Memory-of-Thought) framework, enabling large language models to enhance their reasoning capabilities by using pre-thought reasoning paths as external memory without labeled data or parameter updates, and retrieving relevant memories during testing.
mRedditSum: A Multimodal Abstractive Summarization Dataset of Reddit Threads with Images
Keighley Overbay (Seoul National University), Gunhee Kim (Seoul National University)
CodeTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
π― What it does: This paper constructs the MREDDITSUM dataset, which contains 3033 Reddit discussion threads with images and text, and provides human-generated summaries that cover image information.
Multi-level Contrastive Learning for Script-based Character Understanding
Dawei Li (University of California, San Diego), Shiping Yang (Independent Researcher)
CodeRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: To address role understanding in scripts, this paper proposes a multi-layer contrastive learning framework that leverages multi-perspective information from abstracts and dialogues to learn fine-grained and global role representations.
Multi-Source Multi-Type Knowledge Exploration and Exploitation for Dialogue Generation
Xuanfan Ni (Nanjing University of Aeronautics and Astronautics), Piji Li (Shandong University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Designed and implemented a framework called KnowEE, which enhances the quality of open-domain multi-turn dialogue generation through multi-source, multi-type knowledge exploration and injection.
π― What it does: Propose a model combining multi-task knowledge distillation with cosine embedding constraints for key phrase boundary detection and classification in academic papers.
Multilingual estimation of political-party positioning: From label aggregation to long-input Transformers
Dmitry Nikolaev (University of Stuttgart), Sebastian PadΓ³ (University of Stuttgart)
CodeClassificationTransformerLarge Language ModelText
π― What it does: The paper positions multi-lingual party manifestos on the left-right spectrum using two methods (sentence-level label aggregation and direct prediction by long-text Transformers).
Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in Languages at Scale
Marta R. Costa-jussΓ (Meta), Carleigh Wood (Meta)
CodeExplainability and InterpretabilityData-Centric LearningTransformerTextBenchmark
π― What it does: Expand the original HolisticBias dataset into a multi-lingual version, generating a dataset with approximately 20,459 sentences across 50 languages and 13 demographic axes, and use this dataset to evaluate bias in machine translation and sentence embedding models related to gender and other demographic features.
Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer
Elizabeth Salesky (Johns Hopkins University), Matt Post (Human Language Technology Center of Excellence)
CodeRepresentation LearningTransformerVision Language ModelImageText
π― What it does: This paper proposes and verifies a method of using pixel-level visual representations (pixel representations) instead of traditional subword embeddings in multilingual machine translation, demonstrating its effectiveness during multilingual training.
π― What it does: This paper proposes a multimodal plan prediction model tailored for the TEACh dataset and designs an agenda-based synthetic dialogue generation framework to train plan prediction models in the absence of real human dialogue data.
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper constructs a linguistic marker framework expressing uncertainty and overconfidence, injects these markers into prompts for question-answering tasks, and systematically evaluates their impact on the performance of large language models.
π― What it does: Propose NeuSTIP, a neural symbolic model capable of simultaneously performing link prediction and time interval prediction on temporal knowledge graphs (TKG);
Non-autoregressive Streaming Transformer for Simultaneous Translation
Zhengrui Ma (Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences), Yang Feng (School of Future Science and Engineering Soochow University)
CodeTransformerText
π― What it does: Propose a non-autoregressive streaming Transformer (NAST) for real-time machine translation, addressing the non-monotonic alignment and source information leakage issues in traditional autoregressive models.
Non-autoregressive Text Editing with Copy-aware Latent Alignments
Yu Zhang (Soochow University), Guohong Fu (Soochow University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a non-autoregressive text editing method based on CTC, which completes the transformation from the source text to the target text by leveraging the potential alignment of three operations: copy (KEEP), delete (DELETE), and insert (ADD).
NormDial: A Comparable Bilingual Synthetic Dialog Dataset for Modeling Social Norm Adherence and Violation
Oliver Li (Columbia University), Smaranda Muresan (Columbia University)
CodeData SynthesisTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: A bilingual (Chinese-English) two-person dialogue dataset called NORMDIAL was constructed through a human-AI collaborative generation process, with each dialogue turn annotated for compliance and violation of social norms.
π― What it does: Constructed the video misleading title dataset VMH, analyzed its characteristics, and proposed a multimodal misinformation detection baseline.
On the Automatic Generation and Simplification of Childrenβs Stories
Maria Valentini (University of Colorado Boulder), Katharina von der Wense (Johannes Gutenberg University Mainz)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: This paper studies the automatic generation of children's stories appropriate for age groups by large language models (LLMs), as well as the application of lexical simplification models in this task;
On the Challenges of Using Black-Box APIs for Toxicity Evaluation in Research
Luiza Pozzobon, Sara Hooker
CodeTextBenchmark
π― What it does: This paper evaluates the experimental reproducibility issues caused by updates to commercial black-box toxicity detection APIs (e.g., Perspective API) over time, and demonstrates their impact by re-scoring the RTP dataset, HELM benchmark, and toxicity mitigation methods.
π― What it does: This paper proposes a Generalized Relation Discovery task that simultaneously handles known and unknown relations in open-world relation extraction, while considering negative samples and long-tail distributions.
π― What it does: Research and implemented a dynamic token filtering method (Token Filtering) in the FiD generative question-answering model, combining it with CALM's layer skipping technique to reduce decoder cross-attention computation and improve the speed of long text generation.
William Rudman (Brown University), Carsten Eickhoff (Brown University)
CodeClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Investigate the persistence of outlier dimensions in large language models and whether they retain task-specific knowledge after fine-tuning, exploring whether a single dimension can accomplish downstream binary classification tasks.
π― What it does: Proposed the P5 method, which achieves personalized response selection under zero-shot and fine-tuning scenarios by inputting the persona sentence with the highest similarity to the response as a prompt along with the dialogue context;
PAC-tuning: Fine-tuning Pre-trained Language Models with PAC-driven Perturbed Gradient Descent
Guangliang Liu (Michigan State University), Rongrong Wang (Michigan State University)
CodeClassificationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose a two-stage fine-tuning method called PAC-tuning, which learns noise variance through PAC-Bayes training and subsequently perturbs gradient descent with this noise, enhancing the generalization ability of pre-trained language models on few-shot text classification tasks.
CodeExplainability and InterpretabilityData-Centric LearningTransformerSupervised Fine-TuningPrompt EngineeringText
π― What it does: Propose a debiasing method called PEFTDebias based on parameter-efficient fine-tuning (PEFT), which first trains the PEFT module on unlabeled axis-related corpora using CDA, and then freezes this module during downstream tasks for fine-tuning to maintain debiasing effects;
People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection
Indira Sen (RWTH Aachen University), Claudia Wagner (RWTH Aachen University)
CodeClassificationData SynthesisAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Compare the effectiveness of manually generated versus automatically generated Counterfactually Augmented Data (CAD) in gender bias and hate speech detection tasks.
Physician Detection of Clinical Harm in Machine Translation: Quality Estimation Aids in Reliance and Backtranslation Identifies Critical Errors
Nikita Mehandru (University of California, Berkeley), Niloufar Salehi (University of California, Berkeley)
CodeLarge Language ModelTextBiomedical DataElectronic Health Records
π― What it does: This paper evaluates the effectiveness of Quality Estimation (QE) and Back-Translation (BT) as feedback mechanisms in helping physicians determine whether machine translation (MT) outputs are suitable for patients to read, designing and conducting a randomized controlled experiment.
Pointwise Mutual Information Based Metric and Decoding Strategy for Faithful Generation in Document Grounded Dialogs
Yatin Nandwani (IBM Research, AI), Luis Lastras (IBM Research, AI)
CodeGenerationTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed a novel conditional point-wise mutual information-based metric PMI-FAITH and the corresponding decoding strategy PMI-DECODE for evaluating and generating dialogue responses consistent with documents.
Post-hoc Utterance Refining Method by Entity Mining for Faithful Knowledge Grounded Conversations
Yoonna Jang (Korea University), Heuiseok Lim (Korea University)
CodeGenerationRetrievalTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: Propose a post-refinement method called REM, which improves the authenticity and consistency of generated statements in knowledge-driven dialogues by mining entities and implicitly learning entity information from given knowledge.
π― What it does: Investigate the computational power of linear Transformer (LT) and its Fast Weight Programmer (FWP), and conduct experimental validation on formal language recognition tasks.
Pragmatic Reasoning Unlocks Quantifier Semantics for Foundation Models
Yiyuan Li (UNC Chapel Hill), Shashank Srivastava (UNC Chapel Hill)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper proposes the PRESQUE framework, which combines natural language reasoning with pragmatic reasoning (RSA) to infer the percentage range of quantifiers using pre-trained language models.
Precedent-Enhanced Legal Judgment Prediction with LLM and Domain-Model Collaboration
Yiquan Wu (Zhejiang University), Kun Kuang (Zhejiang University)
CodeClassificationConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
π― What it does: This paper proposes a precedent-enhanced legal judgment prediction framework called PLJP, which leverages a domain model to provide candidate labels and precedent retrieval, followed by an LLM synthesizing precedents in context to make the final judgment.
π― What it does: Propose the TextReact method, which integrates retrieved natural language text with chemical reaction data to enhance the performance of chemical prediction models.
Preserving Knowledge Invariance: Rethinking Robustness Evaluation of Open Information Extraction
Ji Qi (Tsinghua University), Xu Bin (Tsinghua University)
CodeTextBenchmark
π― What it does: Proposed the ROBUST benchmark to evaluate the robustness of open information extraction models under syntactic and expression distribution drift;
π― What it does: Propose a BART-based multi-document summarization model called FABRIC, which enhances cross-document coherence and factual consistency by leveraging theme-assisted document segmentation, degenerate composite layers, and layered graph attention.
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose GDP-ZERO, an open MCTS decision framework that requires no model training, leveraging large language models to simulate user-system interactions and evaluate task progress in dialogue tree search, directly planning goal-oriented dialogue strategies;
Prompting Large Language Models with Chain-of-Thought for Few-Shot Knowledge Base Question Generation
Yuanyuan Liang (East China Normal University), Yunshi Lan (East China Normal University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: Propose the KQG-CoT framework, which selects representative logical forms by structural encoding and clustering from an unlabeled data pool, then constructs prompts using chain-of-thought (CoT) to enable large language models to generate natural language questions consistent with given logical forms under few-shot settings.
Mayank Mishra (IBM Research AI), Srikanth Tamilselvam (IBM Research AI)
CodeClassificationGenerationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose and evaluate the use of pseudo-code as a prompt method, investigating its impact on large language models (LLMs) across multi-task scenarios (classification, question answering, and generation);
π― What it does: This paper introduces soft prompts into end-to-end speech translation models to enhance the representation capability of high-level encoders.
PROSE: A Pronoun Omission Solution for Chinese-English Spoken Language Translation
Ke Wang (Huawei IT Innovation and Research Center), Wei Peng (Huawei IT Innovation and Research Center)
CodeTransformerContrastive LearningTextBenchmark
π― What it does: Built PROSE, a Chinese-English spoken translation document-level dataset covering four spoken genres (conversations, TV series, movies, vlogs), and analyzed and addressed translation errors caused by omitted pronouns in Chinese.
π― What it does: Propose Prototype-based HyperAdapter (PHA), which generates task-specific adapters in multi-task and few-shot transfer scenarios through an instance-dense retriever and prototypical hypernetwork.
Pushdown Layers: Encoding Recursive Structure in Transformer Language Models
Shikhar Murty (Stanford University), Christopher Manning (Stanford University)
CodeRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Propose Pushdown Layers that integrate stack memory with Transformer self-attention to automatically infer and maintain recursive syntactic structures.
π― What it does: Built a self-supervised vision transformer model (HOMOGLYPH) to measure the visual similarity between homoglyphs in OCR text, using this similarity as the character substitution cost in the Levenshtein edit distance, thereby improving record linkage accuracy across different OCR engines, scripts, and languages (CJK and ancient scripts).
Query Rewriting in Retrieval-Augmented Large Language Models
Xinbei Ma (Shanghai Jiao Tong University), Nan Duan (Microsoft Research Asia)
CodeRetrievalComputational EfficiencyTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
π― What it does: Proposed the Rewrite-Retrieve-Read framework, introducing a query rewriting step into retrieval-augmented LLMs and training a tunable lightweight rewriter to enhance retrieval quality and final answer accuracy.
R2H: Building Multimodal Navigation Helpers that Respond to Help Requests
Yue Fan (University of California, Santa Cruz), Xin Wang
CodeAutonomous DrivingReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideoTextMultimodalityBenchmark
π― What it does: Proposed the Respond to Help Requests (R2H) benchmark to evaluate the response capability of multimodal navigation assistant agents in dialogue history and interactive processes, and developed two assistant models based on this benchmark: SeeRee (self-supervised vision-language Transformer combined with COS Attention Mask and Parse by Step preprocessing) and zero-shot multimodal LLM (mPLUG-Owl).
π― What it does: This paper proposes a black-box OOD detection framework for text generation models called RAINPROOF, and constructs the LOFTER benchmark tailored to language, domain, and dialogue shift, to evaluate model robustness in open-world scenarios.
Rather a Nurse than a Physician - Contrastive Explanations under Investigation
Oliver Eberle (Technische UniversitΓ€t Berlin), Stephanie Brandl (University of Copenhagen)
CodeExplainability and InterpretabilityTransformerTextBiomedical Data
π― What it does: Compared controlled and uncontrolled explanations between humans and models in text classification tasks to examine whether controlled explanations better align with human cognition.
CodeData SynthesisTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
π― What it does: Automatically construct a billion-level long-turn dialogue corpus by retrieving, recombining, and remarking short-turn dialogues to concatenate them into long turns, thereby enhancing the utilization of long contexts;
Reading Order Matters: Information Extraction from Visually-rich Documents by Token Path Prediction
Chong Zhang (Fudan University), Tao Gui (Fudan University)
CodeClassificationTransformerVision Language ModelImageTextBenchmark
π― What it does: This paper proposes the Token Path Prediction (TPP) framework, modeling the named entity recognition (NER) problem in visually rich documents as predicting token paths in a complete directed graph, thereby addressing the issue of BIO label failure caused by the uncertainty of text reading order generated by OCR.
CodeTransformerLarge Language ModelTextBenchmarkChain-of-Thought
π― What it does: Proposes the RECEVAL framework, which utilizes reference-free NLI and V-Information metrics to conduct fine-grained evaluation of the correctness and informativeness of multi-step reasoning chains.
Reducing Sequence Length by Predicting Edit Spans with Large Language Models
Masahiro Kaneko (MBZUAI), Naoaki Okazaki (Tokyo Institute of Technology)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose using edit spans to predict modified parts in the source text, thereby reducing the target sequence length and lowering inference costs in local sequence transduction tasks.
Referring Image Segmentation via Joint Mask Contextual Embedding Learning and Progressive Alignment Network
Ziling Huang (University of Tokyo), Shinβichi Satoh
CodeSegmentationConvolutional Neural NetworkVision Language ModelMultimodality
π― What it does: Propose a joint mask and context embedding learning network (JMCELN), achieving multi-stage segmentation reasoning through dynamically learned context embeddings and a progressive alignment network;
π― What it does: This paper proposes the RegNLP (Regulation and NLP) research direction, aiming to systematically integrate regulatory studies with natural language processing to promote risk assessment and governance of large language models (LLMs).
Huy Dao, Yuxiang Nie (Hong Kong University of Science and Technology)
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: This paper proposes an RTCP, a goal-driven conversation promotion framework that integrates short-term and long-term planning along with prefix tuning to generate high-quality dialogues for targeted products;
Rethinking Word-Level Auto-Completion in Computer-Aided Translation
Xingyu Chen (Shanghai Jiao Tong University), Rui Wang (Tencent AI Lab)
CodeTransformerText
π― What it does: Redefine the evaluation criteria for Word-Level Auto Completion (WLAC), introducing the 'consistency' criterion, and propose a joint training method based on machine translation, significantly improving WLAC performance.
Retrieval-Generation Alignment for End-to-End Task-Oriented Dialogue System
Weizhou Shen (Sun Yat-sen University), Wei Bi (Tencent)
CodeGenerationRetrievalTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextRetrieval-Augmented Generation
π― What it does: This paper proposes an end-to-end task-oriented dialogue system, MK-TOD, which jointly trains a retriever and a generator using maximum marginal likelihood, and introduces retrieval-related meta-knowledge to alleviate the retrieval-generation imbalance problem.