NAACL 2025 Papers

Natural Language Inference (NLI) is crucial for evaluating models’ Natural Language Understanding (NLU) and reasoning abilities. The development of NLI, in part, has been driven by the creation of large datasets, which require significant human effort. This has spurred interest in semi-supervised learning (SSL) that leverages both labeled and unlabeled data. However, the absence of hypotheses and class labels in NLI tasks complicates SSL. Prior work has used class-specific fine-tuned large language models (LLMs) to generate hypotheses and assign pseudo-labels but discarded many LLM-constructed samples during training to ensure the quality. In contrast, we propose to leverage all LLM-constructed samples by handling potentially noisy samples by injecting the moments of labeled samples during training to properly adjust the level of noise. Our method outperforms strong baselines on multiple NLI datasets in low-resource settings.

The vast amount of online information today poses challenges for non-English speakers, as much of it is concentrated in high-resource languages such as English and French. Wikipedia reflects this imbalance, with content in low-resource languages frequently outdated or incomplete. Recent research has sought to improve cross-language synchronization of Wikipedia tables using rule-based methods. These approaches can be effective, but they struggle with complexity and generalization. This paper explores large language models (LLMs) for multilingual information synchronization, using zero-shot prompting as a scalable solution. We introduce the Information Updation dataset, simulating the real-world process of updating outdated Wikipedia tables, and evaluate LLM performance. Our findings reveal that single-prompt approaches often produce suboptimal results, prompting us to introduce a task decomposition strategy that enhances coherence and accuracy. Our proposed method outperforms existing baselines, particularly in Information Updation (1.79%) and Information Addition (20.58%), highlighting the model’s strength in dynamically updating and enriching data across architectures.

Allophony refers to the variation in the phonetic realization of a phoneme based on its phonetic environment. Modeling allophones is crucial for atypical pronunciation assessment, which involves distinguishing atypical from typical pronunciations. However, recent phoneme classifier-based approaches often simplify this by treating various realizations as a single phoneme, bypassing the complexity of modeling allophonic variation. Motivated by the acoustic modeling capabilities of frozen self-supervised speech model (S3M) features, we propose MixGoP, a novel approach that leverages Gaussian mixture models to model phoneme distributions with multiple subclusters. Our experiments show that MixGoP achieves state-of-the-art performance across four out of five datasets, including dysarthric and non-native speech. Our analysis further suggests that S3M features capture allophonic variation more effectively than MFCCs and Mel spectrograms, highlighting the benefits of integrating MixGoP with S3M features.

Multimodal knowledge graph completion (MKGC) aims to predict missing triples in MKGs using multimodal information. Recent research typically either extracts information from each modality separately to predict, then ensembles the predictions at the decision stage, or projects multiple modalities into a unified feature space to learn multimodal representations for prediction. However, these methods usually overlook the intrinsic correlation between modalities in MKGs which should be leveraged in both unimodal knowledge extraction and multimodal knowledge fusion. Motivated by this, we propose a noval Modal-collaborative knowledge learning (Moodle) framework for MKGC, the key idea of which is to foster mutual guidance and collaboration during unimodal knowledge extraction, to let each modality acquire distinct and complementary knowledge that subsequently enhances the multimodal knowledge fusion. Specifically, Moodle preserves the representations of different modalities to learn unimodal knowledge while modeling the mutual guidance through multi-task learning. Furthermore, Moodle performs multimodal knowledge fusion and prediction guided by unimodal knowledge, capturing their synergistic relationships and acquire fine-grained semantic knowledge through contrastive learning. Extensive experiments on three real-world datasets demonstrate the advantages of Moodle over state-of-the-art methods.

Human evaluations lay at the heart of evaluations within the field of Natural Language Processing (NLP). Seen as the “golden standard” of evaluations, questions are being asked on whether these evaluations are both reproducible and repeatable. One overlooked aspect is the design choices made by researchers when designing user interfaces (UIs). In this paper, four UIs used in past NLP human evaluations are assessed by UX experts, based on standardized human-centered interaction principles. Building on these insights, we derive several recommendations that the NLP community should apply when designing UIs, to enable more consistent human evaluation responses.

Legal documents including judgments and court orders require highly sophisticated legal knowledge for understanding. To disclose expert knowledge for non-experts, we explore the problem of visualizing legal texts with easy-to-understand diagrams and propose a novel dataset of LegalViz with 23 languages and 7,010 cases of legal document and visualization pairs, using the DOT graph description language of Graphviz. LegalViz provides a simple diagram from a complicated legal corpus identifying legal entities, transactions, legal sources, and statements at a glance, that are essential in each judgment. In addition, we provide new evaluation metrics for the legal diagram visualization by considering graph structures, textual similarities, and legal contents. We conducted empirical studies on few-shot and finetuning large language models for generating legal diagrams and evaluated them with these metrics, including legal content-based evaluation within 23 languages. Models trained with LegalViz outperform existing models including GPTs, confirming the effectiveness of our dataset.

In this paper, we address the task of semantic segmentation of legal documents through rhetorical role classification, with a focus on Indian legal judgments. We introduce **LegalSeg**, the largest annotated dataset for this task, comprising over 7,000 documents and 1.4 million sentences, labeled with 7 rhetorical roles. To benchmark performance, we evaluate multiple state-of-the-art models, including Hierarchical BiLSTM-CRF, TransformerOverInLegalBERT (ToInLegalBERT), Graph Neural Networks (GNNs), and Role-Aware Transformers, alongside an exploratory **RhetoricLLaMA**, an instruction-tuned large language model. Our results demonstrate that models incorporating broader context, structural relationships, and sequential sentence information outperform those relying solely on sentence-level features. Additionally, we conducted experiments using surrounding context and predicted or actual labels of neighboring sentences to assess their impact on classification accuracy. Despite these advancements, challenges persist in distinguishing between closely related roles and addressing class imbalance. Our work underscores the potential of advanced techniques for improving legal document understanding and sets a strong foundation for future research in legal NLP.

Legal judgment prediction (LJP) is an essential task for legal AI, aiming at predicting judgments based on the facts of a case. Legal judgments can involve multiple law articles and charges. Although recent methods in LJP have made notable progress, most are constrained to single-task settings (e.g., only predicting charges) or single-label settings (e.g., not accommodating cases with multiple charges), diverging from the complexities of real-world scenarios. In this paper, we address the challenge of predicting relevant law articles and charges within the framework of legal judgment prediction, treating it as a multi-task and multi-label text classification problem. We introduce a knowledge-enhanced approach, called K-LJP, that incorporates (I) ”label-level knowledge” (such as definitions and relationships among labels) to enhance the representation of case facts for each task, and (ii) ”task-level knowledge” (such as the alignment between law articles and corresponding charges) to improve task synergy. Comprehensive experiments demonstrate our method’s effectiveness in comparison to state-of-the-art (SOTA) baselines.

In real-world NLP applications, Large Language Models (LLMs) offer promising solutions due to their extensive training on vast datasets. However, the large size and high computation demands of LLMs limit their practicality in many applications, especially when further fine-tuning is required. To address these limitations, smaller models are typically preferred for deployment. However, their training is hindered by the scarcity of labeled data. In contrast, unlabeled data is often readily which can be leveraged by using LLMs to generate pseudo-labels for training smaller models. This enables the smaller models (student) to acquire knowledge from LLMs (teacher) while reducing computational costs. This process introduces challenges, such as potential noisy pseudo-labels. % and the high computational expense of processing large unlabeled datasets. Selecting high-quality and informative data is therefore critical to enhance model performance while improving the efficiency of data utilization. To address this, we propose LLKD that enables Learning with Less computational resources and less data for Knowledge Distillation from LLMs. LLKD is an adaptive sample selection method that incorporates signals from both the teacher and student. Specifically, it prioritizes samples where the teacher demonstrates high confidence in its labeling, indicating reliable labels, and where the student exhibits a high information need, identifying challenging samples that require further learning. Our comprehensive experiments show that LLKD achieves superior performance across various datasets with higher data efficiency.

Generative Large Language Models (LLMs) are capable of being in-context learners. However, the underlying mechanism of in-context learning (ICL) is still a major research question, and experimental research results about how models exploit ICL are not always consistent. In this work, we propose a framework for evaluating in-context learning mechanisms, which we claim are a combination of retrieving internal knowledge and learning from in-context examples by focusing on regression tasks. First, we show that LLMs can solve real-world regression problems and then design experiments to measure the extent to which the LLM retrieves its internal knowledge versus learning from in-context examples. We argue that this process lies on a spectrum between these two extremes. We provide an in-depth analysis of the degrees to which these mechanisms are triggered depending on various factors, such as prior knowledge about the tasks and the type and richness of the information provided by the in-context examples. We employ three LLMs and utilize multiple datasets to corroborate the robustness of our findings. Our results shed light on how to engineer prompts to leverage meta-learning from in-context examples and foster knowledge retrieval depending on the problem being addressed.

Developing effective text summarizers remains a challenge due to issues like hallucinations, key information omissions, and verbosity in LLM-generated summaries. This work explores using LLM-generated feedback to improve summary quality by aligning the summaries with human preferences for faithfulness, completeness, and conciseness. We introduce FeedSum, a large-scale dataset containing multi-dimensional LLM feedback on summaries of varying quality across diverse domains. Our experiments show how feedback quality, dimensionality, and granularity influence preference learning, revealing that high-quality, multi-dimensional, fine-grained feedback significantly improves summary generation. We also compare two methods for using this feedback: supervised fine-tuning and direct preference optimization. Finally, we introduce SummLlama3-8b, a model that outperforms the nearly 10x larger Llama3-70b-instruct in generating human-preferred summaries, demonstrating that smaller models can achieve superior performance with appropriate training. The full dataset and SummLlama3-8B model are available at https://huggingface.co/datasets/DISLab/FeedSum and https://huggingface.co/DISLab/SummLlama3-8B.

Smart word substitution aims to enhance sentence quality by improving word choices, however current benchmarks rely on human-labeled data , which suffers from subjectivity and lacks diversity due to limitations in the number of annotators. Since word choices are inherently subjective, ground-truth word substitutions generated by a small group of annotators are often incomplete and likely not generalizable. To circumvent this issue, we instead employ a model-based scoring (BARTScore) to quantify sentence quality, thus forgoing the need for human annotations. Specifically, we use this score to define a distribution for each word substitution, allowing one to test whether a substitution is statistically superior relative to others. Further, we propose a loss function that directly optimizes the alignment between model predictions and sentence scores, while also enhancing the overall quality score of a substitution. Crucially, model learning no longer requires human labels, thus avoiding the cost of annotation while maintaining the quality of the text modified with substitutions. Experimental results show that the proposed approach outperforms both masked language models (BERT, BART) and large language models (GPT-4, LLaMA).

Domain Large Language Models (LLMs) are developed for domain-specific tasks based on general LLMs. But it still requires professional knowledge to facilitate the expertise for some domain-specific tasks. In this paper, we investigate into knowledge-intensive calculation problems. We find that the math problems to be challenging for LLMs, when involving complex domain-specific rules and knowledge documents, rather than simple formulations of terminologies. Therefore, we propose a pipeline to solve the domain-specific calculation problems with Knowledge-Intensive Programs Generator more effectively, named as KIPG. It generates knowledge-intensive programs according to the domain-specific documents. For each query, key variables are extracted, then outcomes which are dependent on domain knowledge are calculated with the programs. By iterative preference alignment, the code generator learns to improve the logic consistency with the domain knowledge. Taking legal domain as an example, we have conducted experiments to prove the effectiveness of our pipeline, and extensive analysis on the modules. We also find that the code generator is also adaptable to other domains, without training on the new knowledge.

Large language models (LLMs) demonstrate impressive few-shot learning capabilities, but their performance varies widely based on the sequence of in-context examples. Key factors influencing this include the sequence’s length, composition, and arrangement, as well as its relation to the specific query. Existing methods often tackle these factors in isolation, overlooking their interdependencies. Moreover, the extensive search space for selecting optimal sequences complicates the development of a holistic approach. In this work, we introduce Beam Search-based Example Sequence Constructor (BESC), a novel method for learning to construct optimal example sequences. addresses all key factors involved in sequence selection by considering them jointly during inference, while incrementally building the sequence. This design enables the use of beam search to significantly reduce the complexity of the search space. Experiments across various datasets and language models show notable improvements in performance.

Interactions with large language models (LLMs) often yield long and detailed responses, leveraging both parametric knowledge and retrieval-augmented generation (RAG). While these responses can provide rich insights, they often include redundant or less engaging content not aligned with user interests. This issue becomes apparent when users specify particular subtopics to include or exclude – termed **coverage-conditioned (C2)** queries – as LLMs often struggle to provide tailored responses. To address this challenge, we investigate the role of query outlines, sequences of subqueries designed to guide LLMs in generating responses that meet specific user requirements. To systematically create and evaluate these outlines, we introduce **QTree**, a dataset of 10K hierarchical sets of information-seeking subqueries that define structured boundaries for outline creation and evaluation in C2 scenarios. Additionally, we develop **QPlanner**, a 7B language model trained to generate customized outlines within boundaries of QTree. We evaluate the effectiveness of the generated outlines through automatic and human judgements, focusing on their impact within retrieval-augmented generation (RAG) systems. Experimental results demonstrate that QPlanner, especially when trained with alignment techniques like DPO, generates higher-quality outlines that better fulfill diverse user needs.

LangLearn is an open-source framework designed to facilitate autonomous learning of low-resource languages (LRL). By combining a language-agnostic approach with AI-enhanced flashcards, LangLearn empowers users to generate custom flashcards for their vocabulary, while offering structured learning through both pre-curated and self-curated decks. The framework integrates six key components: the word definition, corresponding Hanji characters, romanization with numeric tones, audio pronunciation, a sample sentence, as well as a contextual AI-generated image. LangLearn currently supports English and Taiwanese Hokkien (a variety of Southern Min), with plans to extend support for other dialects. Our preliminary study demonstrates that LangLearn positively empowers users to engage with LRLs using their vocabulary preferences, with a comprehensive user study currently underway. LangLearn’s modular structure enables future expansion, including ASR-based pronunciation practice. The code is available at https://github.com/HokkienTranslation/HokkienTranslation.

Large language models (LLMs) have demonstrated remarkable capabilities in handling complex dialogue tasks without requiring use case-specific fine-tuning. However, analyzing live dialogues in real-time necessitates low-latency processing systems, making it impractical to deploy models with billions of parameters due to latency constraints. As a result, practitioners often prefer smaller models with millions of parameters, trained on high-quality, human-annotated datasets. Yet, curating such datasets is both time-consuming and costly. Consequently, there is a growing need to combine the scalability of LLM-generated labels with the precision of human annotations, enabling fine-tuned smaller models to achieve both higher speed and accuracy comparable to larger models. In this paper, we introduce a simple yet effective framework to address this challenge. Our approach is specifically designed for per-utterance classification problems, which encompass tasks such as intent detection, dialogue state tracking, and more. To mitigate the impact of labeling errors from LLMs – the primary source of inaccuracies in student models – we propose a noise-reduced preference learning loss. Experimental results demonstrate that our method significantly improves accuracy across utterance-level dialogue tasks, including sentiment detection (over 2%), dialogue act classification (over 1.5%), etc.

Prior case retrieval (PCR) is crucial for legal practitioners to find relevant precedent cases given the facts of a query case. Existing approaches often overlook the underlying semantic intent in determining relevance with respect to the query case. In this work, we propose LeCoPCR, a novel approach that explicitly generate intents in the form of legal concepts from a given query case facts and then augments the query with these concepts to enhance models understanding of semantic intent that dictates relavance. To overcome the unavailability of annotated legal concepts, We employ a weak supervision approach to extract key legal concepts from the reasoning section using Determinantal Point Process (DPP) to balance quality and diversity. Experimental results on the ECtHR-PCR dataset demonstrate the effectiveness of leveraging legal concepts and DPP-based key concept extraction.

As large language models (LLMs) are increasingly deployed in diverse applications, including chatbot assistants and code generation, aligning their behavior with safety and ethical standards has become paramount. However, jailbreak attacks, which exploit vulnerabilities to elicit unintended or harmful outputs, threaten LLMs safety significantly. In this paper, we introduce Layer-AdvPatcher, a novel methodology designed to defend against jailbreak attacks by utilizing an unlearning strategy to patch specific layers within LLMs through self-augmented datasets. Our insight is that certain layer(s), tend to produce affirmative tokens when faced with harmful prompts. By identifying these layers and adversarially exposing them to generate more harmful data, one can understand their inherent and diverse vulnerabilities to attacks. With these exposures, we then “unlearn” these issues, reducing the impact of affirmative tokens and hence minimizing jailbreak risks while keeping the model’s responses to safe queries intact.We conduct extensive experiments on two models, four benchmark datasets, and multiple state-of-the-art jailbreak attacks to demonstrate the efficacy of our approach. Results indicate that our framework reduces the harmfulness and attack success rate of jailbreak attacks without compromising utility for benign queries compared to recent defense methods. Our code is publicly available at: https://github.com/oyy2000/LayerAdvPatcher