ACL 2025 Papers

The few-shot relation extraction (FSRE) aims at enhancing the model’s generalization to new relations with very few labeled instances (support instances). Most existing studies use prototype networks (ProtoNets) for FSRE and assume that the support set, adapting the model to new relations, only contains accurately labeled instances. However, this assumption is usually unrealistic, as even carefully-annotated datasets often contain mislabeled instances. Thus, it is essential to enhance the robustness of FSRE models to noisy labels in support set, but this issue remains unexplored. In this paper, we first conduct a preliminary study, revealing the high sensitivity of ProtoNets to such noisy labels. Meanwhile, we discover that fully leveraging mislabeled support instances is crucial for enhancing the model’s robustness. To do this, we propose a self-denoising model for FSRE, which can automatically correct noisy labels of support instances. Specifically, our model comprises two core components: 1) a label correction module (LCM), used to correct mislabeled support instances based on the distances between them in the embedding space, and 2) a relation classification module (RCM), designed to achieve more robust relation prediction using the corrected labels generated by the LCM. Moreover, we propose a feedback-based training strategy, which focuses on training LCM and RCM to synergistically handle noisy labels in support set. Experimental results on two public datasets show the effectiveness and robustness of our model. Notably, even in scenarios without noisy labels, our model significantly outperforms all competitive baselines.

Disagreement in human labeling is ubiquitous, and can be captured in human judgment distributions (HJDs). Recent research has shown that explanations provide valuable information for understanding human label variation (HLV) and large language models (LLMs) can approximate HJD from a few human-provided label-explanation pairs. However, collecting explanations for every label is still time-consuming. This paper examines whether LLMs can be used to replace humans in generating explanations for approximating HJD. Specifically, we use LLMs as annotators to generate model explanations for a few given human labels. We test ways to obtain and combine these label-explanations with the goal to approximate human judgment distributions. We further compare the resulting human with model-generated explanations, and test automatic and human explanation selection. Our experiments show that LLM explanations are promising for NLI: to estimate HJDs, generated explanations yield comparable results to human’s when provided with human labels. Importantly, our results generalize from datasets with human explanations to i) datasets where they are not available and ii) challenging out-of-distribution test sets.

Understanding robustness is essential for building reliable NLP systems. Unfortunately, in the context of machine translation, previous work mainly focused on documenting robustness failures or improving robustness. In contrast, we study robustness from a model representation perspective by looking at internal model representations of ungrammatical inputs and how they evolve through model layers. For this purpose, we perform Grammatical Error Detection (GED) probing and representational similarity analysis. Our findings indicate that the encoder first detects the grammatical error, then corrects it by moving its representation toward the correct form. To understand what contributes to this process, we turn to the attention mechanism where we identify what we term *Robustness Heads*. We find that *Robustness Heads* attend to interpretable linguistic units when responding to grammatical errors, and that when we fine-tune models for robustness, they tend to rely more on *Robustness Heads* for updating the ungrammatical word representation.

Cross-lingual stance detection identifies users’ attitudes toward specific targets in texts by transferring knowledge from source languages to target languages. Previous studies have typically facilitated this transfer by translating and aligning labels or targets. However, these methods cannot effectively perform cross-lingual transfer of the complex reasoning processes in stance detection. To address this challenge, we propose a reinforcement learning framework using cross-lingual Chain-of-Thought (CoT) alignment, referred to as RCCA. Specifically, we adopt a cross-lingual CoT alignment strategy to obtain the high-quality CoTs generated from target language inputs. After that, we leverage reinforcement learning by sampling CoTs and assigning rewards according to predefined rules, aiming to enhance the model’s generalization capabilities in the target language. Experimental results on four multilingual datasets demonstrate that our approach outperforms competitive methods.

Retrieval-augmented generation (RAG) helps address the limitations of parametric knowledge embedded within a language model (LM). In real world settings, retrieved information can vary in complexity, yet most investigations of LM utilisation of context has been limited to synthetic text. We introduce DRUID (Dataset of Retrieved Unreliable, Insufficient and Difficult-to-understand contexts) with real-world queries and contexts manually annotated for stance. The dataset is based on the prototypical task of automated claim verification, for which automated retrieval of real-world evidence is crucial. We compare DRUID to synthetic datasets (CounterFact, ConflictQA) and find that artificial datasets often fail to represent the complexity and diversity of realistically retrieved context. We show that synthetic datasets exaggerate context characteristics rare in real retrieved data, which leads to inflated context utilisation results, as measured by our novel ACU score. Moreover, while previous work has mainly focused on singleton context characteristics to explain context utilisation, correlations between singleton context properties and ACU on DRUID are surprisingly small compared to other properties related to context source. Overall, our work underscores the need for real-world aligned context utilisation studies to represent and improve performance in real-world RAG settings.

Document Image Translation (DIT), which aims at translating documents in images from source language to the target, plays an important role in Document Intelligence. It requires a comprehensive understanding of document multi-modalities and a focused concentration on relevant textual regions during translation. However, most existing methods usually rely on the vanilla encoder-decoder paradigm, severely losing concentration on key regions that are especially crucial for complex-layout document translation. To tackle this issue, in this paper, we propose a new Query-Response DIT framework (QRDIT). QRDIT reformulates the DIT task into a parallel response/translation process of the multiple queries (i.e., relevant source texts), explicitly centralizing its focus toward the most relevant textual regions to ensure translation accuracy. A novel dynamic aggregation mechanism is also designed to enhance the text semantics in query features toward translation. Extensive experiments in four translation directions on three benchmarks demonstrate its state-of-the-art performance, showing significant translation quality improvements toward whole-page complex-layout document images.

Multi-session personalized dialogue generation is one of the most important topics in open-domain dialogue. It aims to generate responses consistent with the dialogue history and personality information across multiple sessions to engage users’ interest in the dialogue. Recent approaches focusing on history modeling and persona modeling have advanced the development of this field. However, they overlook the importance of dialogue structure in helping large language models (LLMs) understand the dialogue context. Moreover, these methods do not efficiently expand and utilize personality information, reducing the responses’ consistency. In this paper, we propose a Persona-Aware LLM-enAnCEd(PALACE) framework for multi-session personalized dialogue generation. Specifically, the framework consists of three components: a topic-aware memory bank, a persona prompt learning module, and VAE-LoRA. The topic-aware memory bank works by retrieving historical information that possesses a certain dialogue structure and relevant topics. The persona prompt learning module enhances the LLM’s persona-aware capabilities by utilizing a persona commonsense knowledge graph and a query-driven graph neural network. Furthermore, to enhance the generative capabilities of the LLM and obtain more useful prior knowledge, we combine VAE with LoRA to propose VAE-LoRA. Experimental results on the MSC and DuLeMon dataset demonstrate that our framework outperforms the state-of-the-art methods in automatic and human evaluation metrics.

Current vision-language models (VLMs) understand complex vision-text tasks by extracting overall semantic information from large-scale cross-modal associations. However, extracting from large-scale cross-modal associations often smooths out semantic details and requires large computations, limiting multimodal fine-grained understanding performance and efficiency. To address this issue, this paper proposes a detail-oriented prompt learning (DoPL) method for vision-language models to implement fine-grained multi-modal semantic alignment with merely 0.25M trainable parameters. According to the low-entropy information concentration theory, DoPL explores shared interest tokens from text-vision correlations and transforms them into alignment weights to enhance text prompt and vision prompt via detail-oriented prompt generation. It effectively guides the current frozen layer to extract fine-grained text-vision alignment cues. Furthermore, DoPL constructs detail-oriented prompt generation for each frozen layer to implement layer-by-layer localization of fine-grained semantic alignment, achieving precise understanding in complex vision-text tasks. DoPL performs well in parameter-efficient fine-grained semantic alignment with only 0.12% tunable parameters for vision-language models. The state-of-the-art results over the previous parameter-efficient fine-tuning methods and full fine-tuning approaches on six benchmarks demonstrate the effectiveness and efficiency of DoPL in complex multi-modal tasks.

This paper proposes formulating Zipf’s meaning-frequency law, the power law between word frequency and the number of meanings, as a relationship between word frequency and contextual diversity. The proposed formulation quantifies meaning counts as contextual diversity, which is based on the directions of contextualized word vectors obtained from a Language Model (LM). This formulation gives a new interpretation to the law and also enables us to examine it for a wider variety of words and corpora than previous studies have explored. In addition, this paper shows that the law becomes unobservable when the size of the LM used is small and that autoregressive LMs require much more parameters than masked LMs to be able to observe the law.

Knowledge graph embedding techniques have emerged as a critical approach for addressing the issue of missing relations in knowledge graphs. However, existing methods often suffer from limitations, including high intra-group similarity, loss of semantic information, and insufficient inference capability, particularly in complex relation patterns such as 1-N and N-1 relations. To address these challenges, we introduce a novel KGE framework that leverages mutual information maximization to improve the semantic representation of entities and relations. By maximizing the mutual information between different components of triples, such as (h, r) and t, or (r, t) and h, the proposed method improves the model’s ability to preserve semantic dependencies while maintaining the relational structure of the knowledge graph. Extensive experiments on benchmark datasets demonstrate the effectiveness of our approach, with consistent performance improvements across various baseline models. Additionally, visualization analyses and case studies demonstrate the improved ability of the MI framework to capture complex relation patterns.

Argument quality assessment faces inherent challenges due to its subjective nature, where different evaluators may assign varying quality scores for an argument based on personal perspectives. Although existing datasets collect opinions from multiple annotators to model subjectivity, most existing computational methods fail to consider multi-perspective evaluation. To address this issue, we propose MPAQ, a multi-persona framework for argument quality assessment that simulates diverse evaluator perspectives through large language models. It first dynamically generates targeted personas tailored to an input argument, then simulates each persona’s reasoning process to evaluate the argument quality from multiple perspectives. To effectively generate fine-grained quality scores, we develop a coarse-to-fine scoring strategy that first generates a coarse-grained integer score and then refines it into a fine-grained decimal score. Experiments on IBM-Rank-30k and IBM-ArgQ-5.3kArgs datasets demonstrate that MPAQ consistently outperforms strong baselines while providing comprehensive multi-perspective rationales.

Online discourse is increasingly trapped in a vicious cycle where polarizing language fuelstoxicity and vice versa. Identity, one of the most divisive issues in modern politics, oftenincreases polarization. Yet, prior NLP research has mostly treated toxicity and polarization asseparate problems. In Indonesia, the world’s third-largest democracy, this dynamic threatens democratic discourse, particularly in online spaces. We argue that polarization and toxicity must be studied in relation to each other. To this end, we present a novel multi-label Indonesian dataset annotated for toxicity, polarization, and annotator demographic information. Benchmarking with BERT-base models and large language models (LLMs) reveals that polarization cues improve toxicity classification and vice versa. Including demographic context further enhances polarization classification performance.

Temporal knowledge graph reasoning aims to predict future events with knowledge of existing facts and plays a key role in various downstream tasks. Previous methods focused on either graph structure learning or semantic reasoning, failing to integrate dual reasoning perspectives to handle different prediction scenarios. Moreover, they lack the capability to capture the inherent differences between historical and non-historical events, which limits their generalization across different temporal contexts. To this end, we propose a **M**ulti-**E**xpert **S**tructural-**S**emantic **H**ybrid (MESH) framework that employs three kinds of expert modules to integrate both structural and semantic information, guiding the reasoning process for different events. Extensive experiments on three datasets demonstrate the effectiveness of our approach.

Privacy policies inform users about data collection and usage, yet their complexity limits accessibility for diverse populations. Existing Privacy Policy Question Answering (QA) systems exhibit performance disparities across English dialects, disadvantaging speakers of non-standard varieties. We propose a novel multi-agent framework inspired by human-centered design principles to mitigate dialectal biases. Our approach integrates a Dialect Agent, which translates queries into Standard American English (SAE) while preserving dialectal intent, and a Privacy Policy Agent, which refines predictions using domain expertise. Unlike prior approaches, our method does not require retraining or dialect-specific fine-tuning, making it broadly applicable across models and domains. Evaluated on PrivacyQA and PolicyQA, our framework improves GPT-4o-mini’s zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data. These results highlight the effectiveness of structured agent collaboration in mitigating dialect biases and underscore the importance of designing NLP systems that account for linguistic diversity to ensure equitable access to privacy information.

Large Reasoning Models (LRMs) have significantly advanced beyond traditional Large Language Models (LLMs) with their exceptional logical reasoning capabilities, yet these improvements introduce heightened safety risks. When subjected to jailbreak attacks, their ability to generate more targeted and organized content can lead to greater harm. Although some studies claim that reasoning enables safer LRMs against existing LLM attacks, they overlook the inherent flaws within the reasoning process itself. To address this gap, we propose the first jailbreak attack targeting LRMs, exploiting their unique vulnerabilities stemming from the advanced reasoning capabilities. Specifically, we introduce a Chaos Machine, a novel component to transform attack prompts with diverse one-to-one mappings. The chaos mappings iteratively generated by the machine are embedded into the reasoning chain, which strengthens the variability and complexity and also promotes a more robust attack. Based on this, we construct the Mousetrap framework, which makes attacks projected into nonlinear-like low sample spaces with mismatched generalization enhanced. Also, due to the more competing objectives, LRMs gradually maintain the inertia of unpredictable iterative reasoning and fall into our trap. Success rates of the Mousetrap attacking o1-mini, Claude-Sonnet and Gemini-Thinking are as high as 96%, 86% and 98% respectively on our toxic dataset Trotter. On benchmarks such as AdvBench, StrongREJECT, and HarmBench, attacking Claude-Sonnet, well-known for its safety, Mousetrap can astonishingly achieve success rates of 87.5%, 86.58% and 93.13% respectively. Attention: This paper contains inappropriate, offensive and harmful content.

High computation costs and latency of large language models such as GPT-4 have limited their deployment in clinical settings. Small language models (SLMs) offer a cost-effective alternative, but their limited capacity requires biomedical domain adaptation, which remains challenging. An additional bottleneck is the unavailability and high sensitivity of clinical data. To address these challenges, we propose a novel framework for adapting SLMs into high-performing clinical models. We introduce the MediPhi collection of 3.8B-parameter SLMs developed with our novel framework: pre-instruction tuning of experts on relevant medical and clinical corpora (PMC, Medical Guideline, MedWiki, etc.), model merging, and clinical-tasks alignment. To cover most clinical tasks, we extended the CLUE benchmark to CLUE+, doubling its size. Our expert models deliver relative improvements on this benchmark over the base model without any task-specific fine-tuning: 64.3% on medical entities, 49.5% on radiology reports, and 44% on ICD-10 coding (outperforming GPT-4-0125 by 14%). We unify the expert models into MediPhi via model merging, preserving gains across benchmarks. Furthermore, we built the MediFlow collection, a synthetic dataset of 2.5 million high-quality instructions on 14 medical NLP tasks, 98 fine-grained document types, and JSON format support. Alignment of MediPhi using supervised fine-tuning and direct preference optimization achieves further gains of 18.9% on average.

Automated metrics for Machine Translation have made significant progress, with the goal of replacing expensive and time-consuming human evaluations. These metrics are typically assessed by their correlation with human judgments, which captures the monotonic relationship between human and metric scores. However, we argue that it is equally important to ensure that metrics treat all systems fairly and consistently. In this paper, we introduce a method to evaluate this aspect.

Scientific Natural Language Inference (NLI) is the task of predicting the semantic relation between a pair of sentences extracted from research articles. Existing datasets for this task are derived from various computer science (CS) domains, whereas non-CS domains are completely ignored. In this paper, we introduce a novel evaluation benchmark for scientific NLI, called MisMatched. The new MisMatched benchmark covers three non-CS domains–Psychology, Engineering, and Public Health, and contains 2,700 human annotated sentence pairs. We establish strong baselines on MisMatched using both Pre-trained Small Language Models (SLMs) and Large Language Models (LLMs). Our best performing baseline shows a Macro F1 of only 78.17% illustrating the substantial headroom for future improvements. In addition to introducing the MisMatched benchmark, we show that incorporating sentence pairs having an implicit scientific NLI relation between them in model training improves their performance on scientific NLI. We make our dataset and code publicly available on GitHub.

Faced with an expensive human annotation process, creators of NLP systems increasingly turn to synthetic data generation. While this method shows promise, the extent to which synthetic data can replace human annotation is poorly understood. We investigate the use of synthetic data in Fact Verification (FV) and Evidence-based Question Answering (QA) by incrementally replacing human-generated data with synthetic points on eight diverse datasets. Strikingly, replacing up to 90% of the training data only marginally decreases performance, but replacing the final 10% leads to severe declines. We find that models trained on purely synthetic data can be improved by including as few as 125 human-generated data points. We show that matching the performance gain of a little human data requires an order of magnitude more synthetic data, and then estimate price ratios at which human annotation would be a more cost-effective solution. Our results suggest that even when human annotation at scale is infeasible, there is great value to having a small proportion of the dataset being human-generated.