NAACL 2025 论文列表

This paper presents the Korean National Educational Test Benchmark (KoNET), a new benchmark designed to evaluate Multimodal Generative AI Systems using Korean national educational tests. KoNET comprises four exams: the Korean Elementary General Educational Development Test (KoEGED), Middle (KoMGED), High (KoHGED), and College Scholastic Ability Test (KoCSAT). These exams are renowned for their rigorous standards and diverse questions, facilitating a comprehensive analysis of AI performance across different educational levels. By focusing on Korean, KoNET provides insights into model performance in less-explored languages. We assess a range of models—open-source, open-access, and closed APIs—by examining difficulties, subject diversity, and human error rates. The code and dataset builder will be made fully open-source.

Large language models (LLMs) have demonstrated significant progress in various natural language generation and understanding tasks. However, their linguistic generalization capabilities remain questionable, raising doubts about whether these models learn language similarly to humans. While humans exhibit compositional generalization and linguistic creativity in language use, the extent to which LLMs replicate these abilities, particularly in morphology, is under-explored. In this work, we systematically investigate the morphological generalization abilities of LLMs through the lens of compositionality. We define morphemes as compositional primitives and design a novel suite of generative and discriminative tasks to assess morphological productivity and systematicity. Focusing on agglutinative languages such as Turkish and Finnish, we evaluate several state-of-the-art instruction-finetuned multilingual models, including GPT-4 and Gemini. Our analysis shows that LLMs struggle with morphological compositional generalization particularly when applied to novel word roots, with performance declining sharply as morphological complexity increases. While models can identify individual morphological combinations better than chance, their performance lacks systematicity, leading to significant accuracy gaps compared to humans.

The advancement of large language models (LLMs) has led to a greater challenge of having a rigorous and systematic evaluation of complex tasks performed, especially in enterprise applications. Therefore, LLMs need to be benchmarked with enterprise datasets for a variety of NLP tasks. This work explores benchmarking strategies focused on LLM evaluation, with a specific emphasis on both English and Japanese. The proposed evaluation framework encompasses 25 publicly available domain-specific English benchmarks from diverse enterprise domains like financial services, legal, climate, cyber security, and 2 public Japanese finance benchmarks. The diverse performance of 8 models across different enterprise tasks highlights the importance of selecting the right model based on the specific requirements of each task. Code and prompts are available on GitHub.

Large Language Models (LLMs) have shown promising results in a variety of literary tasks, often using complex memorized details of narration and fictional characters. In this work, we evaluate the ability of Llama-3 at attributing utterances of direct-speech to their speaker in novels. The LLM shows impressive results on a corpus of 28 novels, surpassing published results with ChatGPT and encoder-based baselines by a large margin. We then validate these results by assessing the impact of book memorization and annotation contamination.We found that these types of memorization do not explain the large performance gain, making Llama-3 the new state-of-the-art for quotation attribution in English literature. We release publicly our code and data.

Explaining the decision-making process of machine learning models is crucial for ensuring their reliability and transparency for end users. One popular explanation form highlights key input features, such as i) tokens (e.g., Shapley Values and Integrated Gradients), ii) interactions between tokens (e.g., Bivariate Shapley and Attention-based methods), or iii) interactions between spans of the input (e.g., Louvain Span Interactions). However, these explanation types have only been studied in isolation, making it difficult to judge their respective applicability. To bridge this gap, we develop a unified framework that facilitates an automated and direct comparison between highlight and interactive explanations comprised of four diagnostic properties. We conduct an extensive analysis across these three types of input feature explanations – each utilizing three different explanation techniques–across two datasets and two models, and reveal that each explanation has distinct strengths across the different diagnostic properties. Nevertheless, interactive span explanations outperform other types of input feature explanations across most diagnostic properties. Despite being relatively understudied, our analysis underscores the need for further research to improve methods generating these explanation types. Additionally, integrating them with other explanation types that perform better in certain characteristics could further enhance their overall effectiveness.

Automated fact-checking systems often struggle with trustworthiness, as their generated explanations can include hallucinations. In this work, we explore evidence attribution for fact-checking explanation generation. We introduce a novel evaluation protocol, citation masking and recovery, to assess attribution quality in generated explanations. We implement our protocol using both human annotators and automatic annotators and found that LLM annotation correlates with human annotation, suggesting that attribution assessment can be automated. Finally, our experiments reveal that: (1) the best-performing LLMs still generate explanations that are not always accurate in their attribution; and (2) human-curated evidence is essential for generating better explanations.

As large language models (LLMs) expand into performing as agents for real-world applications beyond traditional NLP tasks, evaluating their robustness becomes increasingly important. However, existing benchmarks often overlook critical dimensions like cultural and social awareness. To address these, we introduce CASA, a benchmark designed to assess LLM agents’ sensitivity to cultural and social norms across two web-based tasks: online shopping and social discussion forums. Our approach evaluates LLM agents’ ability to detect and appropriately respond to norm-violating user queries and observations. Furthermore, we propose a comprehensive evaluation framework that measures awareness coverage, helpfulness in managing user queries, and the violation rate when facing misleading web content. Experiments show that current LLMs perform significantly better in non-agent than in web-based agent environments, with agents achieving less than 10% awareness coverage and over 40% violation rates. To improve performance, we explore two methods: prompting and fine-tuning, and find that combining both methods can offer complementary advantages – fine-tuning on culture-specific datasets significantly enhances the agents’ ability to generalize across different regions, while prompting boosts the agents’ ability to navigate complex tasks. These findings highlight the importance of constantly benchmarking LLM agents’ cultural and social awareness during the development cycle.

Lexical ambiguity—where a single wordform takes on distinct, context-dependent meanings–serves as a useful tool to compare across different language models’ (LMs’) ability to form distinct, contextualized representations of the same stimulus. Few studies have systematically compared LMs’ contextualized word embeddings for languages beyond English. Here, we evaluate semantic representations of Spanish ambiguous nouns in context in a suite of Spanish-language monolingual and multilingual BERT-based models. We develop a novel dataset of minimal-pair sentences evoking the same or different sense for a target ambiguous noun. In a pre-registered study, we collect contextualized human relatedness judgments for each sentence pair. We find that various BERT-based LMs’ contextualized semantic representations capture some variance in human judgments but fall short of the human benchmark. In exploratory work, we find that performance scales with model size. We also identify stereotyped trajectories of target noun disambiguation as a proportion of traversal through a given LM family’s architecture, which we partially replicate in English. We contribute (1) a dataset of controlled, Spanish sentence stimuli with human relatedness norms, and (2) to our evolving understanding of the impact that LM specification (architectures, training protocols) exerts on contextualized embeddings.

Large Language Models (LLMs) offer the potential to automate hiring by matching job descriptions with candidate resumes, streamlining recruitment processes, and reducing operational costs. However, biases inherent in these models may lead to unfair hiring practices, reinforcing societal prejudices and undermining workplace diversity. This study examines the performance and fairness of LLMs in job-resume matching tasks within the English language and U.S. context. It evaluates how factors such as gender, race, and educational background influence model decisions, providing critical insights into the fairness and reliability of LLMs in HR applications.Our findings indicate that while recent models have reduced biases related to explicit attributes like gender and race, implicit biases concerning educational background remain significant. These results highlight the need for ongoing evaluation and the development of advanced bias mitigation strategies to ensure equitable hiring practices when using LLMs in industry settings.

What ethical concerns, if any, do LLM researchers have? We introduce EthiCon, a corpus of 1,580 ethical concern statements extracted from scientific papers published in the ACL Anthology. We extract ethical concern keywords from the statements and show promising results in automating the concern identification process. Through a survey (N=200), we compare the ethical concerns of the corpus to the concerns listed by the general public and professionals in the field. Finally, we compare our retrieved ethical concerns with existing taxonomies and guidelines pointing to gaps and actionable insights.

This study finds that existing information retrieval (IR) models show significant biases based on the linguistic complexity of input queries, performing well on linguistically simpler (or more complex) queries while underperforming on linguistically more complex (or simpler) queries.To address this issue, we propose EqualizeIR, a framework to mitigate linguistic biases in IR models. EqualizeIR uses a linguistically biased weak learner to capture linguistic biases in IR datasets and then trains a robust model by regularizing and refining its predictions using the biased weak learner. This approach effectively prevents the robust model from overfitting to specific linguistic patterns in data. We propose four approaches for developing linguistically-biased models. Extensive experiments on several datasets show that our method reduces performance disparities across linguistically simple and complex queries, while improving overall retrieval performance.

Augmenting Large Language Models (LLMs) with retrieved external knowledge has proven effective in improving the factual accuracy of generated responses. Despite their success, retrieval-augmented LLMs still face the distractibility issue, where the generated responses are negatively influenced by noise from both external and internal knowledge sources. In this paper, we introduce a novel, training-free decoding method guided by entropy considerations to mitigate this issue. Our approach utilizes entropy-based document-parallel ensemble decoding to prioritize low-entropy distributions from retrieved documents, thereby enhancing the extraction of relevant information of context. Additionally, it incorporates a contrastive decoding mechanism that contrasts the obtained low-entropy ensemble distribution with the high-entropy distribution derived from the model’s internal knowledge across layers, which ensures a greater emphasis on reliable external information. Extensive experiments on open-domain question answering datasets demonstrate the superiority of our method.

Document-level relation extraction (DocRE) aims to extract relations between entities in a document. While previous research has primarily focused on traditional small models, recent studies have extended the scope to large language models (LLMs). Current LLM-based methods typically focus on filtering all potential relations (candidate relations) within a document at one time and then performing triplet fact extraction. However, most approaches for candidate relation filtering are based on the document level, which results in insufficient correlation between candidate relations and entity pairs. In addition, the data imbalance problem caused by a large amount of no-relation data (NA problem) is another important reason for the suboptimal performance of LLM-based methods. To address these issues, we propose an entity pair-guided relation summarization and retrieval model (EP-RSR) for DocRE, which introduces an innovative LLM-based document-level relation extraction paradigm, EPRF (Entity Pair-Relation-Fact), along with an entity pair-level candidate relation filtering method. Our approach first selects entity pairs that potentially contain relations and uses them to guide relation summarization and retrieval for extracting relation facts. This enhances the relevance between candidate relations and entity pairs while alleviating the issue of imbalanced NA data. Benchmark testing on three datasets demonstrates that our approach achieves state-of-the-art (SOTA) performance for LLM-based models. Our code is available at https://github.com/LookingYu/EP-RSR.

Clinical named entity recognition (NER) aims to retrieve important entities within clinical narratives. Recent works have demonstrated that large language models (LLMs) can achieve strong performance in this task. While previous works focus on proprietary LLMs, we investigate how open NER LLMs, trained specifically for entity recognition, perform in clinical NER. Our initial experiment reveals significant contrast in performance for some clinical entities and how a simple exploitment on entity types can alleviate this issue. In this paper, we introduce a novel framework, entity decomposition with filtering, or EDF. Our key idea is to decompose the entity recognition task into several retrievals of entity sub-types and then filter them. Our experimental results demonstrate the efficacies of our framework and the improvements across all metrics, models, datasets, and entity types. Our analysis also reveals substantial improvement in recognizing previously missed entities using entity decomposition. We further provide a comprehensive evaluation of our framework and an in-depth error analysis to pave future works.

Recent research efforts have explored the potential of leveraging natural language inference (NLI) techniques to enhance relation extraction (RE). In this vein, we introduce MetaEntail-RE, a novel adaptation method that harnesses NLI principles to enhance RE performance. Our approach follows past works by verbalizing relation classes into class-indicative hypotheses, aligning a traditionally multi-class classification task to one of textual entailment. We introduce three key enhancements: (1) Meta-class analysis which, instead of labeling non-entailed premise-hypothesis pairs with the less informative “neutral” entailment label, provides additional context by analyzing overarching meta-relationships between classes; (2) Feasible hypothesis filtering, which removes unlikely hypotheses from consideration based on domain knowledge derived from data; and (3) Group-based prediction selection, which further improves performance by selecting highly confident predictions. MetaEntail-RE is conceptually simple and empirically powerful, yielding significant improvements over conventional relation extraction techniques and other NLI formulations. We observe surprisingly large F1 gains of 17.6 points on BioRED and 13.4 points on ReTACRED compared to conventional methods, underscoring the versatility of MetaEntail-RE across both biomedical and general domains.

Despite recent advances in large language models, open-source models often struggle to consistently perform well on complex reasoning tasks. Existing ensemble methods, whether applied at the token or output levels, fail to address these challenges. In response, we present Language model Ensemble with Monte Carlo Tree Search (LE-MCTS), a novel framework for process-level ensembling of language models. LE-MCTS formulates step-by-step reasoning with an ensemble of language models as a Markov decision process. In this framework, states represent intermediate reasoning paths, while actions consist of generating the next reasoning step using one of the language models selected from a predefined pool. Guided by a process-based reward model, LE-MCTS performs a tree search over the reasoning steps generated by different language models, identifying the most accurate reasoning chain. Experimental results on five mathematical reasoning benchmarks demonstrate that our approach outperforms both single language model decoding algorithms and language model ensemble methods. Notably, LE-MCTS improves performance by 3.6% and 4.3% on the MATH and MQA datasets, respectively, highlighting its effectiveness in solving complex reasoning problems.

Few-Shot Relation Extraction (FSRE) aims to achieve high classification performance by training relation classification models with a small amount of labeled data. Prototypical networks serve as a straightforward and efficient method for optimizing model performance by combining similarity evaluation and contrastive learning. However, directly integrating these methods can introduce unpredictable noise, such as information redundancy, which hinders classification performance and negatively affects embedding space learning. The technique presented in this paper applies Local-To-Global optimization to enhance prototypical networks in few-shot relation extraction. Specifically, this paper develops a local optimization strategy that indirectly optimizes the prototypes by optimizing the other information contained within the prototypes. It considers relation prototypes as global anchors and incorporates the techniques introduced in this paper, such as information alignment, local contrastive learning, and a local adaptive focal loss function, to address the issues of information redundancy. This approach enables the model to learn a unified and effective embedding space. We conduct extensive experiments on the FewRel 1.0 and FewRel 2.0 datasets to validate the effectiveness of the proposed model.

Large vision-language models (LVLMs) have achieved impressive results in visual question-answering and reasoning tasks through vision instruction tuning on specific datasets. However, there remains significant room for improvement in aligning visual and language modalities. Existing methods often depend on external models or data, leading to uncontrollable and unstable alignment results. In this paper, we propose SIMA, a self-improvement framework that enhances visual and language modality alignment without external dependencies. SIMA leverages existing vision instruction tuning datasets to self-generate responses, incorporating an in-context self-critic mechanism that constructs preference pairs for tuning. Crucially, our approach allows LVLMs to act as critics by designing effective critic prompts, eliminating the need for additional fine-tuning with external instruction data. We introduce three novel visual metrics within the self-critic process to guide judgement, significantly improving the accuracy of self-critic. Through extensive experiments across 14 hallucination and comprehensive benchmarks, we demonstrate that SIMA significantly improves LVLM’s performance and outperforms previous approaches, achieving superior modality alignment.

Large Language Models (LLMs) have recently demonstrated remarkable performance in Text-to-SQL tasks. However, existing research primarily focuses on the optimization of prompts and improvements in workflow, with few studies delving into the exploration of the questions. In this paper, we propose a Text-to-SQL framework based on question classification and multi-agent collaboration (QCMA-SQL). Specifically, we first employ multiple cross-attention mechanisms to train a schema selector to classify questions and select the most suitable database schema. Subsequently, we employ the appropriate agents based on the varying difficulty levels of the questions to generate preliminary SQL queries. Moreover, we implement syntax validation and execution optimization steps to generate final SQL queries. Experimental results on the Spider dataset show that the QCMA-SQL framework achieves an execution accuracy of 87.4%, outperforming state-of-the-art methods. Through ablation studies, we find that classifying the questions ultimately leads to a 2.8% increase in execution accuracy.