NAACL 2025 Papers

Topic models are a popular approach for extracting semantic information from large document collections. However, recent studies suggest that the topics generated by these models often do not align well with human intentions. Although metadata such as labels and authorship information are available, it has not yet been effectively incorporated into neural topic models. To address this gap, we introduce FANToM, a novel method to align neural topic models with both labels and authorship information. FANToM allows for the inclusion of this metadata when available, producing interpretable topics and author distributions for each topic. Our approach demonstrates greater expressiveness than conventional topic models by learning the alignment between labels, topics, and authors. Experimental results show that FANToM improves existing models in terms of both topic quality and alignment. Additionally, it identifies author interests and similarities.

Existing studies in backdoor defense have predominantly focused on the training phase, overlooking the critical aspect of testing time defense. This gap becomes pronounced in the context of Large Language Models (LLMs) deployed as Web Services, which typically offer only black-box access, rendering training-time defenses impractical. To bridge this gap, this study critically examines the use of demonstrations as a defense mechanism against backdoor attacks in black-box LLMs. With an identified task, we retrieve task-relevant demonstrations from a clean data pool and integrate them with user queries during testing. Importantly, this approach does not necessitate modifications or tuning of the model, nor does it require insight into the model’s internal architecture. The alignment properties inherent in in-context learning play a pivotal role in mitigating the impact of backdoor triggers, effectively recalibrating the behavior of compromised models. Our experimental analysis demonstrates that this method robustly defends against both instance-level and instruction-level backdoor attacks, outperforming existing defense baselines across most evaluation scenarios.

Aspect Sentiment Triplet Extraction (ASTE) is a thriving research area with impressive outcomes being achieved on high-resource languages. However, the application of cross-lingual transfer to the ASTE task has been relatively unexplored, and current code-switching methods still suffer from term boundary detection issues and out-of-dictionary problems. In this study, we introduce a novel Test-Time Code-SWitching (TT-CSW) framework, which bridges the gap between the bilingual training phase and the monolingual test-time prediction. During training, a generative model is developed based on bilingual code-switched training data and can produce bilingual ASTE triplets for bilingual inputs. In the testing stage, we employ an alignment-based code-switching technique for test-time augmentation. Extensive experiments on cross-lingual ASTE datasets validate the effectiveness of our proposed method. We achieve an average improvement of 3.7% in terms of weighted-averaged F1 in four datasets with different languages. Additionally, we set a benchmark using ChatGPT and GPT-4, and demonstrate that even smaller generative models fine-tuned with our proposed TT-CSW framework surpass ChatGPT and GPT-4 by 14.2% and 5.0% respectively.

This paper presents Temporal-aware Soft Prompt Tuning (TASPT), a novel approach for automatic text dating. Unlike existing methods, which often overlook the evolution of word meanings in texts spanning long periods, TASPT incorporates the unique characteristics of historical texts. It introduces a temporal-aware text representation that dynamically captures both semantic variance and invariance. This representation is combined with a soft prompt, enabling efficient parameter tuning for automatic text dating. Experiments show that TASPT outperforms all existing methods on two diachronic datasets: the Twenty-Four Histories and the Royal Society Corpus.

Multimodal foundation models (MFMs) have demonstrated significant success in tasks such as visual captioning, question answering, and image-text retrieval. However, these models face inherent limitations due to their finite internal capacity, which restricts their ability to process extended temporal sequences—an essential requirement for comprehensive video and audio analysis. To overcome these challenges, we introduce a specialized cognitive module, temporal working memory (TWM), which aims to enhance the temporal modeling capabilities of MFMs. It selectively retains task-relevant information across temporal dimensions, ensuring that critical details are preserved throughout the processing of video and audio content. The TWM uses a query-guided attention approach to focus on the most informative multimodal segments within temporal sequences. By retaining only the most relevant content, TWM optimizes the use of the model’s limited capacity, enhancing its temporal modeling ability. This plug-and-play module can be easily integrated into existing MFMs. With our TWM, nine state-of-the-art models exhibit significant performance improvements across tasks such as video captioning, question answering, and video-text retrieval. By enhancing temporal modeling, TWM extends the capability of MFMs to handle complex, time-sensitive data effectively. Our code is available at https://github.com/xid32/NAACL_2025_TWM.

This paper investigates hybrid intelligence and collaboration between researchers of sexism and Large Language Models (LLMs), with afour-component pipeline. First, nine sexism researchers answer questions about their knowledge of sexism and of LLMs. They then participate in two interactive experiments involving an LLM (GPT3.5). The first experiment has experts assessing the model’s knowledgeabout sexism and suitability for use in research. The second experiment tasks them with creating three different definitions of sexism: anexpert-written definition, an LLM-written one, and a co-created definition. Lastly, zero-shot classification experiments use the three definitions from each expert in a prompt template for sexism detection, evaluating GPT4o on 2.500 texts sampled from five sexism benchmarks. We then analyze the resulting 67.500 classification decisions. The LLM interactions lead to longer and more complex definitions of sexism. Expert-written definitions on average perform poorly compared to LLM-generated definitions. However, some experts do improve classification performance with their co-created definitions of sexism, also experts who are inexperienced in using LLMs.

Large language models (LLMs) are susceptible to persuasion, which can pose risks when models are faced with an adversarial interlocutor. We take a first step towards defending models against persuasion while also arguing that defense against adversarial (i.e. *negative*) persuasion is only half of the equation: models should also be able to accept beneficial (i.e. *positive*) persuasion to improve their answers. We show that optimizing models for only one side results in poor performance on the other. In order to balance positive and negative persuasion, we introduce **P**ersuasion-**B**alanced **T**raining (or **PBT**), which leverages multi-agent recursive dialogue trees to create data and trains models via preference optimization to accept persuasion *when appropriate*. PBT allows us to use data generated from dialogues between smaller 7-8B models for training much larger 70B models. Moreover, PBT consistently improves resistance to misinformation and resilience to being challenged while also resulting in the best overall performance on holistic data containing both positive and negative persuasion. Crucially, we show that PBT models are better teammates in multi-agent debates across two domains (trivia and commonsense QA). We find that without PBT, pairs of stronger and weaker models have unstable performance, with the order in which the models present their answers determining whether the team obtains the stronger or weaker model’s performance. PBT leads to better and more stable results and less order dependence, with the stronger model consistently pulling the weaker one up.

Number-focused headline generation is a summarization task requiring both high textual quality and precise numerical accuracy, which poses a unique challenge for Large Language Models (LLMs). Existing studies in the literature focus only on either textual quality or numerical reasoning and thus are inadequate to address this challenge. In this paper, we propose a novel chain-of-thought framework for using rationales comprising key elements of the Topic, Entities, and Numerical reasoning (TEN) in news articles to enhance the capability for LLMs to generate topic-aligned high-quality texts with precise numerical accuracy. Specifically, a teacher LLM is employed to generate TEN rationales as supervision data, which are then used to teach and fine-tune a student LLM. Our approach teaches the student LLM automatic generation of rationales with enhanced capability for numerical reasoning and topic-aligned numerical headline generation. Experiments show that our approach achieves superior performance in both textual quality and numerical accuracy.

Extracting insights from text columns can bechallenging and time-intensive. Existing methods for topic modeling and feature extractionare based on syntactic features and often overlook the semantics. We introduce the semantictext column featurization problem, and presenta scalable approach for automatically solvingit. We extract a small sample smartly, use alarge language model (LLM) to label only thesample, and then lift the labeling to the wholecolumn using text embeddings. We evaluateour approach by turning existing text classification benchmarks into semantic categorization benchmarks. Our approach performs better than baselines and naive use of LLMs.

Although there has been a growing interest among industries in integrating generative LLMs into their services, limited experience and scarcity of resources act as a barrier in launching and servicing large-scale LLM-based services. In this paper, we share our experiences in developing and operating generative AI models within a national-scale search engine, with a specific focus on the sensitiveness of user queries. We propose a taxonomy for sensitive search queries, outline our approaches, and present a comprehensive analysis report on sensitive queries from actual users. We believe that our experiences in launching generative AI search systems can contribute to reducing the barrier in building generative LLM-based services.

While broad-coverage multilingual natural language processing tools have been developed, a significant portion of the world’s over 7000 languages are still neglected. One reason is the lack of evaluation datasets that cover a diverse range of languages, particularly those that are low-resource or endangered. To address this gap, we present a large-scale text classification dataset encompassing 1504 languages many of which have otherwise limited or no annotated data. This dataset is constructed using parallel translations of the Bible. We develop relevant topics, annotate the English data through crowdsourcing and project these annotations onto other languages via aligned verses. We benchmark a range of existing multilingual models on this dataset. We make our dataset and code available to the public.

Continual learning is vital for task-oriented dialogue systems (ToDs), and AdapterCL, equipped with residual adapters, has proven effectiveness in this domain. However, its performance is limited by training separate adapters for each task, preventing global knowledge sharing. To address this, we propose **Task-wrapped Continual Learning (TCL)**, a novel framework that employs **Task-Wrapped Adapters (TWAs)**, to simultaneously learn both global and task-specific information through parameter sharing. TCL leverages task-conditioned hypernetworks to transfer global knowledge across tasks, enabling TWAs to start from more informed initialization, efficiently learning task-specific details while reducing model parameters. Additionally, the simple, linear structure of both hypernetworks and TWAs ensure stable training, with task-free inference supported through effective loss utilization. Across 37 ToD domains, TCL consistently outperforms AdapterCL, significantly reducing forgetting. Remarkably, by setting the task embedding dimension to 1, TCL achieves a 4.76% improvement over AdapterCL while using only 46% of the parameters. These findings position TWA as a lightweight, powerful alternative to traditional adapters, offering a promising solution for continual learning in ToDs. The code is availableat https://github.com/cloversjtu/TCL.

Modern language models (LMs) have significantly advanced generative modeling in natural language processing (NLP). Despite their success, LMs often struggle with adaptation to new contexts in real-time applications. A promising approach to task adaptation is activation intervention, which steers the LMs’ generation process by identifying and manipulating the activations. However, existing interventions rely heavily on heuristic rules or require many prompt inputs to determine effective interventions. In this paper, we propose a layer-wise additive activation intervention framework that optimizes the intervention process, thereby enhancing sample efficiency. We evaluate our framework on various datasets, demonstrating improvements in the accuracy of pretrained LMs and competing intervention baselines.

Sarcasm is a linguistic phenomenon that intends to ridicule a target (e.g., entity, event, or person) in an inherent way. Multimodal Sarcasm Explanation (MuSE) aims at revealing the intended irony in a sarcastic post using a natural language explanation. Though important, existing systems overlooked the significance of the target of sarcasm in generating explanations. In this paper, we propose a Target-aUgmented shaRed fusion-Based sarcasm explanatiOn model, aka. TURBO. We design a novel shared-fusion mechanism to leverage the inter-modality relationships between an image and its caption. TURBO assumes the target of the sarcasm and guides the multimodal shared fusion mechanism in learning intricacies of the intended irony for explanations. We evaluate our proposed TURBO model on the MORE+ dataset. Comparison against multiple baselines and state-of-the-art models signifies the performance improvement of TURBO by an average margin of +3.3%. Moreover, we explore LLMs in zero and one-shot settings for our task and observe that LLM-generated explanation, though remarkable, often fails to capture the critical nuances of the sarcasm. Furthermore, we supplement our study with extensive human mevaluation on TURBO’s generated explanations and find them out to be comparatively better than other systems.

Data selection for fine-tuning large language models (LLMs) aims to choose a high-quality subset from existing datasets, allowing the trained model to outperform baselines trained on the full dataset. However, the expanding body of research lacks a clear, unified framework, and the variability in experimental settings complicates systematic comparisons.While existing surveys comprehensively overview the stages and methods of data selection, they often overlook an in-depth exploration of the fine-tuning phase. In this paper, we conduct a focused review of recent data selection techniques for fine-tuning LLMs, analyzing a dozen key studies. We introduce a novel three-stage scheme—comprising feature extraction, criteria design, and selector evaluation—to systematically categorize and evaluate these methods. Additionally, we propose a unified comparison approach that incorporates ratio-based efficiency and ranking-based feasibility metrics to address inconsistencies across experiments. Our findings reveal that methods emphasizing more targeted quality measurement achieve higher efficiency but at the cost of feasibility. Finally, we discuss trends and highlight four key challenges in fine-tuning data selection, offering potential directions for future research.

Toxicity text detectors can be vulnerable to adversarial examples - small perturbations to input text that fool the systems into wrong detection. Existing attack algorithms are time-consuming and often produce invalid or ambiguous adversarial examples, making them less useful for evaluating or improving real-world toxicity content moderators. This paper proposes an annotation pipeline for quality control of generated toxic adversarial examples (TAE). We design model-based automated annotation and human-based quality verification to assess the quality requirements of . Successful should fool a target toxicity model into making benign predictions, be grammatically reasonable, appear natural like human-generated text, and exhibit semantic toxicity. When applying these requirements to more than 20 state-of-the-art (SOTA) TAE attack recipes, we find many invalid samples from a total of 940k raw TAE attack generations. We then utilize the proposed pipeline to filter and curate a high-quality TAE dataset we call TaeBench (of size 264k). Empirically, we demonstrate that TaeBench can effectively transfer-attack SOTA toxicity content moderation models and services. Our experiments also show that TaeBench with adversarial training achieve significant improvements of the robustness of two toxicity detectors.

Eradicating poverty is the first goal in the U.N. Sustainable Development Goals. However, aporophobia – the societal bias against people living in poverty – constitutes a major obstacle to designing, approving and implementing poverty-mitigation policies. This work presents an initial step towards operationalizing the concept of aporophobia to identify and track harmful beliefs and discriminative actions against poor people on social media. In close collaboration with non-profits and governmental organizations, we conduct data collection and exploration. Then we manually annotate a corpus of English tweets from five world regions for the presence of (1) direct expressions of aporophobia, and (2) statements referring to or criticizing aporophobic views or actions of others, to comprehensively characterize the social media discourse related to bias and discrimination against the poor. Based on the annotated data, we devise a taxonomy of categories of aporophobic attitudes and actions expressed through speech on social media. Finally, we train several classifiers and identify the main challenges for automatic detection of aporophobia in social networks. This work paves the way towards identifying, tracking, and mitigating aporophobic views on social media at scale.

Reaching a human-level understanding of real-world documents necessitates effective machine reading comprehension, yet recent developments in this area often struggle with table images. In response, we introduce the Visual Table Reading Comprehension (TabComp) dataset, which includes table images, questions, and generative answers designed to evaluate OCR-free models. Unlike general Visual Question Answering (VQA) datasets, TabComp uniquely focuses on table images, fostering the development of systems which obviate the use of optical character recognition (OCR) technology, which often struggles with complex table layouts. Our findings reveal that current OCR-free models perform poorly on TabComp, highlighting the need for robust, specialized models for accurate table reading comprehension. We propose TabComp as a benchmark for evaluating OCR-free models in table reading comprehension and encourage the research community to collaborate on developing more effective solutions. The code and data are available at - https://github.com/dialabiitj/TabComp/

Ensuring the trustworthiness of Generative Foundation Models (GenFMs) is a pressing challenge as they gain widespread use. Existing evaluation toolkits are often limited in scope, dynamism, and flexibility. This paper introduces TRUSTEVAL, a dynamic and comprehensive toolkit designed for evaluating GenFMs across various dimensions. TRUSTEVAL supports both dynamic dataset generation and evaluation, offering advanced features including comprehensiveness, usability, and flexibility. TRUSTEVAL integrates diverse generative models, datasets, evaluation methods, metrics, inference efficiency enhancement, and evaluation report generation. Through case studies, we demonstrate TRUSTEVAL’s potential to advance the trustworthiness evaluation of GenFMs.