NAACL 2025 Papers

This study presents a novel approach to analyzing historical language change, focusing on the evolving semantics of the French term “indigène(s)” (“indigenous”) between 1825 and 1950. While existing approaches to measuring semantic change with contextual word embeddings (CWE) rely primarily on similarity measures or clustering, these methods may not be suitable for highly imbalanced datasets, and pose challenges for interpretation. For this reason, we propose an interpretable, feature-level approach to analyzing language change, which we use to trace the semantic evolution of “indigène(s)” over a 125-year period. Following recent work on sequence embeddings (O’Neill et al., 2024), we use k-sparse autoencoders (k-SAE) (Makhzani and Frey, 2013) to interpret over 210,000 CWEs generated using sentences sourced from the French National Library. We demonstrate that k-SAEs can learn interpretable features from CWEs, as well as how differences in feature activations across time periods reveal highly specific aspects of language change. In addition, we show that diachronic change in feature activation frequency reflects the evolution of French colonial legal structures during the 19th and 20th centuries.

While diffusion models excel at conditionally generating high-quality images, prior works in discrete diffusion models were not evaluated on conditional long-text generation. This work addresses the limitations of prior discrete diffusion models for conditional long-text generation, particularly in the long abstractive summarization task. Despite faster decoding speeds compared to autoregressive methods, previous discrete diffusion models failed on the abstractive summarization task due to the incompatibility between the backbone architectures and the random noising process. To overcome these challenges, we introduce a novel semantic-aware noising process that enables Transformer backbones to handle long sequences effectively. Additionally, we propose CrossMamba, an adaptation of the Mamba model to the encoder-decoder paradigm, which integrates seamlessly with the random absorbing noising process. Our approaches outperform existing discrete diffusion models on three benchmark summarization datasets: Gigaword, CNN/DailyMail, and Arxiv, while also achieving much faster inference speed compared to autoregressive models.

Natural language understanding over tabular data has played a significant role in data discovery tasks such as joinable and unionable table search. State-of-the-art approaches adopt large language models (LLMs) pre-trained over massive text corpora to learn and evaluate the table semantic relatedness. Existing methods typically follow a pretrain-and-finetune paradigm, namely fine-tuning an LLM using tabular data with table relatedness labels. To enhance model’s understanding of tabular data, recent studies include auxiliary tasks such as entity resolution and column type classification in the fine-tuning phase. In spite of achieving performance gain from these supervisions, there is a lack of study on how these supervisions complement or even contrast each other, leading to a subpar performance on the final data discovery tasks. In this paper, we propose a simple yet effective multi-task fine-tuning framework named DiscoverGPT that holistically discovers and leverages the intricate relationships among the supervisions to optimize the performance on the data discovery task. Moreover, DiscoverGPT is plug-and-play that allows a broad range of open-domain auxiliary tasks to be incorporated, by utilizing the generative power of LLMs. We demonstrate the usability and effectiveness of DiscoverGPT with baseline comparisons and ablation studies. DiscoverGPT outperforms the best performing baseline by up to 7% in F1 score.

Detecting factual inconsistency for long document summarization remains challenging, given the complex structure of the source article and long summary length. In this work, we study factual inconsistency errors and connect them with a line of discourse analysis. We find that errors are more common in complex sentences and are associated with several discourse features. We propose a framework that decomposes long texts into discourse-inspired chunks and utilizes discourse information to better aggregate sentence-level scores predicted by NLI models. Our approach shows improved performance on top of different model baselines over several evaluation benchmarks, covering rich domains of texts, focusing on long document summarization. This underscores the significance of incorporating discourse features in developing models for scoring summaries for long document factual inconsistency.

Summarizing movie screenplays presents a unique set of challenges compared to standard document summarization. Screenplays are not only lengthy, but also feature a complex interplay of characters, dialogues, and scenes, with numerous direct and subtle relationships and contextual nuances that are difficult for machine learning models to accurately capture and comprehend. Recent attempts at screenplay summarization focus on fine-tuning transformer-based pre-trained models, but these models often fall short in capturing long-term dependencies and latent relationships, and frequently encounter the “lost in the middle” issue. To address these challenges, we introduce DiscoGraMS, a novel resource that represents movie scripts as a movie character-aware discourse graph (CaD Graph). This approach is well-suited for various downstream tasks, such as summarization, question-answering, and salience detection. The model aims to preserve all salient information, offering a more comprehensive and faithful representation of the screenplay’s content. We further explore a baseline method that combines the CaD Graph with the corresponding movie script through a late fusion of graph and text modalities, and we present very initial promising results. We have made our code and dataset publicly available.

Large language models (LLMs) exhibit exceptional performance across a wide range of natural language processing tasks, often relying on lengthy prompts to harness their full capabilities. However, extended prompts can lead to substantial computational overhead and increased hardware demands, limiting the scalability and efficiency of such models. In this paper, we propose DisComp, a two-stage prompt compression framework based on knowledge distillation that combines task-agnostic and task-aware strategies, designed to efficiently compress prompt length without compromising performance.In the first stage, task-agnostic compression is achieved through knowledge distillation, transferring the summarization capabilities of a LLM to a smaller, more efficient model. The distillation process combines cross-entropy loss and keyword matching loss to ensure the smaller model generates concise and informative summaries. In the second stage, sentence-level pruning is applied, where sentences are ranked by relevance to the query, and irrelevant sentences are pruned to retain only task-critical information. We evaluate our method on three benchmark datasets, LongBench , ZeroSCROLLS and NaturalQuestions. The results show that DisComp significantly outperforms previous task-agnostic and task-specific compression approaches, and it is up to 6.56× faster at inference compared to the best token-level compression method.

Ambiguity is a linguistic tool for encoding information efficiently, yet it also causes misunderstandings and disagreements. It is particularly relevant to the domain of misinformation, as fact-checking ambiguous claims is difficult even for experts. In this paper we argue that instead of predicting a veracity label for which there is genuine disagreement, it would be more beneficial to explain the ambiguity. Thus, this work introduces claim disambiguation, a constrained generation task, for explaining ambiguous claims in fact-checking. This involves editing them to spell out an interpretation that can then be unequivocally supported by the given evidence. We collect a dataset of 1501 such claim revisions and conduct experiments with sequence-to-sequence models. The performance is compared to a simple copy baseline and a Large Language Model baseline. The best results are achieved by employing Minimum Bayes Decoding, with a BertScore F1 of 92.22. According to human evaluation, the model successfully disambiguates the claims 72% of the time.

Preference modeling techniques, such as direct preference optimization (DPO), has shown effective in enhancing the generalization abilities of large language model (LLM). However, in tasks involving video instruction-following, providing informative feedback, especially for open-ended conversations, remains a significant challenge. While previous studies have explored using large multimodal models (LMMs) as reward models for guiding preference modeling, their ability to accurately assess the quality of generated responses and their alignment with video content has not been conclusively demonstrated. This paper introduces a novel framework that utilizes detailed video captions as a proxy of video content, enabling language models to incorporate this information as supporting evidence for scoring video Question Answering (QA) predictions. Our approach demonstrates robust alignment with OpenAI GPT-4V model’s reward mechanism, which directly takes video frames as input. Furthermore, we show that applying our reward mechanism to DPO algorithm significantly improves model performance on open-ended video QA tasks.

Differentially private SGD (DPSGD) enables privacy-preserving training of language models, but often reduces utility, diversity, and linguistic quality. We introduce DPRefine, a three-phase method that initializes a model using data synthesis from a small pre-trained LM with rigorous filtering, applies DP finetuning on private data, and performs self-distillation to refine outputs. This approach significantly outperforms vanilla DPSGD, with AlpacaEval preferring DPRefine’s generations in 78.38% of cases across all datasets and metrics, while also demonstrating substantial improvements in lexical diversity, achieving 85.31% in MSTTR and 86.82% in Jaccard similarity. Our fine-grained analysis reveals that DPRefine reduces linguistic errors in generated text by 84%, mitigating grammar errors, spelling mistakes, and missing punctuation commonly associated with DPSGD. It also reduces inconsistencies present in non-private models, such as fabricated details and misattributed quotes. We find that small models like GPT-2 and T5 are effective for initialization and distillation, highlighting their potential in enabling scalable and efficient deployment of high-performing, privacy-preserving language models with improved linguistic quality and consistency.

Current text-to-image (T2I) synthesis diffusion models raise misuse concerns, particularly in creating prohibited or not-safe-for-work (NSFW) images. To address this, various safety mechanisms and red teaming attack methods are proposed to enhance or expose the T2I model’s capability to generate unsuitable content. However, many red teaming attack methods assume knowledge of the text encoders, limiting their practical usage. In this work, we rethink the case of purely black-box attacks without prior knowledge of the T2l model. To overcome the unavailability of gradients and the inability to optimize attacks within a discrete prompt space, we propose DiffZOO which applies Zeroth Order Optimization to procure gradient approximations and harnesses both C-PRV and D-PRV to enhance attack prompts within the discrete prompt domain. We evaluated our method across multiple safety mechanisms of the T2I diffusion model and online servers. Experiments on multiple state-of-the-art safety mechanisms show that DiffZOO attains an 8.5% higher average attack success rate than previous works, hence its promise as a practical red teaming tool for T2l models.

Maintaining persona consistency is paramount in the application of open-domain dialogue systems, as exemplified by models like ChatGPT. Despite significant advancements, the limited scale and diversity of current persona dialogue datasets remain challenges to achieving robust persona-consistent dialogue models. In this study, drawing inspiration from the success of large-scale pre-training, we introduce PPDS, an open-domain persona dialogue system that employs extensive generative pre-training on a persona dialogue dataset to enhance persona consistency. Specifically, we present a persona extraction model designed to autonomously and precisely generate vast persona dialogue datasets. Additionally, we unveil a pioneering persona augmentation technique to address the invalid persona bias inherent in the constructed dataset. Both quantitative and human evaluations consistently highlight the superior response quality and persona consistency of our proposed model, underscoring its effectiveness.

Text-to-image (T2I) generation models have significantly advanced in recent years. However, effective interaction with these models is challenging for average users due to the need for specialized prompt engineering knowledge and the inability to perform multi-turn image generation, hindering a dynamic and iterative creation process. Recent attempts have tried to equip Multi-modal Large Language Models (MLLMs) with T2I models to bring the user’s natural language instructions into reality. Hence, the output modality of MLLMs is extended, and the multi-turn generation quality of T2I models is enhanced thanks to the strong multi-modal comprehension ability of MLLMs. However, many of these works face challenges in identifying correct output modalities and generating coherent images accordingly as the number of output modalities increases and the conversations go deeper. Therefore, we propose DialogGen, an effective pipeline to align off-the-shelf MLLMs and T2I models to build a Multi-modal Interactive Dialogue System (MIDS) for multi-turn Text-to-Image generation. It is composed of drawing prompt alignment, careful training data curation, and error correction. Moreover, as the field of MIDS flourishes, comprehensive benchmarks are urgently needed to evaluate MIDS fairly in terms of output modality correctness and multi-modal output coherence. To address this issue, we introduce the Multi-modal Dialogue Benchmark (DialogBen), a comprehensive bilingual benchmark designed to assess the ability of MLLMs to generate accurate and coherent multi-modal content that supports image editing. It contains two evaluation metrics to measure the model’s ability to switch modalities and the coherence of the output images. Our extensive experiments on DialogBen and user study demonstrate the effectiveness of DialogGen in producing correct output modalities and coherent multi-modal outputs compared with other State-of-the-Art models. We hope that DialogBen can contribute to the community for building more powerful MIDS.

There is increasing interest in looking at dialects in NLP. However, most work to date still treats dialects as discrete categories. For instance, evaluative work in variation-oriented NLP for English often works with Indian English or African-American Venacular English as homogeneous categories, yet even within one variety there is substantial variation. We examine within-dialect variation and show that performance critically varies within categories. We measure speech-to-text performance on Italian dialects, and empirically observe a geographical performance disparity. This disparity correlates substantially (-0.5) with linguistic similarity to the highest performing dialect variety. We cross-examine our results against dialectometry methods, and interpret the performance disparity to be due to a bias towards dialects that are more similar to the standard variety in the speech-to-text model examined. We additionally leverage geostatistical methods to predict zero-shot performance at unseen sites, and find the incorporation of geographical information to substantially improve prediction performance, indicating there to be geographical structure in the performance distribution.

Video-to-speech (V2S) synthesis, the task of generating speech directly from silent video input, is inherently more challenging than other speech synthesis tasks due to the need to accurately reconstruct both speech content and speaker characteristics from visual cues alone. Recently, audio-visual pretraining has eliminated the need for additional acoustic hints in V2S, which previous methods often relied on to ensure training convergence. However, even with pretraining, existing methods continue to face challenges in achieving a balance between acoustic intelligibility and the preservation of speaker-specific characteristics. We analyzed this limitation and were motivated to introduce DiVISe (Direct Vsual-Input Speech Synthesis), an end-to-end V2S model that predicts Mel-spectrograms directly from video frames alone. Despite not taking any acoustic hints, DiVISe effectively preserves speaker characteristics in the generated audio, and achieves superior performance on both objective and subjective metrics across the LRS2 and LRS3 datasets. Our results demonstrate that DiVISe not only outperforms existing V2S models in acoustic intelligibility but also scales more effectively with increased data and model parameters. Code and weights will be made publicly available after acceptance of this paper.

Existing work on improving language model reasoning typically explores a single solution path, which can be prone to errors. Inspired by perspective-taking in social studies, this paper introduces DiPT, a novel approach that complements current reasoning methods by explicitly incorporating diversified viewpoints. This approach allows the model to gain a deeper understanding of the problem’s context and identify the most effective solution path during the inference stage. Additionally, it provides a general data-centric AI recipe for augmenting existing data to improve their quality for fine-tuning. Our empirical results demonstrate that DiPT can be flexibly integrated into existing methods that focus on a single reasoning approach, enhancing their reasoning performance and stability when presented with paraphrased problems. Furthermore, we illustrate improved context understanding by maintaining the model’s safe outputs against “jailbreaking” prompts intentionally designed to bypass safeguards built into deployed models. Lastly, we show that fine-tuning with data enriched with diverse perspectives can boost the reasoning capabilities of the model compared to fine-tuning with raw data alone.

We present lightweight flow matching multilingual text-to-speech (TTS) systems for Ojibwe, Mi’kmaq, and Maliseet, three Indigenous languages in North America. Our results show that training a multilingual TTS model on three typologically similar languages can improve the performance over monolingual models, especially when data are scarce. Attention-free architectures are highly competitive with self-attention architecture with higher memory efficiency. Our research provides technical development to language revitalization for low-resource languages but also highlights the cultural gap in human evaluation protocols, calling for a more community-centered approach to human evaluation.

Hallucination, a phenomenon where large language models (LLMs) produce output that is factually incorrect or unrelated to the input, is a major challenge for LLM applications that require accuracy and dependability. In this paper, we introduce a reliable and high-speed production system aimed at detecting and rectifying the hallucination issue within LLMs. Our system encompasses named entity recognition (NER), natural language inference (NLI), span-based detection (SBD), and an intricate decision tree-based process to reliably detect a wide range of hallucinations in LLM responses. Furthermore, we have crafted a rewriting mechanism that maintains an optimal mix of precision, response time, and cost-effectiveness. We detail the core elements of our framework and underscore the paramount challenges tied to response time, availability, and performance metrics, which are crucial for real-world deployment of these technologies. Our extensive evaluation, utilizing offline data and live production traffic, confirms the efficacy of our proposed framework and service.

Multimodal machine translation (MMT) aims to leverage additional modalities to assist in language translation. With limited parallel data, current MMT systems rely heavily on monolingual English captioning data. These systems face three key issues: they often overlook that visual signals are unnecessary in many cases, they lack transparency in how visual information is used for disambiguation when needed, and they have yet to fully explore the potential of large-scale vision-language models (LVLMs) for MMT tasks. To address these issues, we propose the Detect, Disambiguate, and Translate (DeDiT) framework, the first reasoning-based framework for MMT leveraging LVLMs. DeDiT detects ambiguity in the input sentence, performs visual reasoning only when ambiguity is found, and generates the final translation.We implemented two versions of DeDiT: a prompting method for large proprietary LVLMs and a fine-tuning method for smaller LVLMs using synthetic data. Experiments on the Multi30K and CoMMuTE benchmarks show that DeDiT outperforms state-of-the-art models in disambiguation accuracy and translation quality. We also introduce an improved evaluation metric for disambiguation accuracy that enhances performance assessment and can be applied to proprietary models accessed via APIs.

Generative AI has made rapid advancements in recent years, achieving unprecedented capabilities in multimodal understanding and code generation. This can enable a new paradigm of front-end development in which multimodal large language models (MLLMs) directly convert visual designs into code implementations. In this work, we construct Design2Code – the first real-world benchmark for this task. Specifically, we manually curate 484 diverse real-world webpages as test cases and develop a set of automatic evaluation metrics to assess how well current multimodal LLMs can generate the code implementations that directly render into the given reference webpages, given the screenshots as input. We also complement automatic metrics with comprehensive human evaluations to validate the performance ranking. To rigorously benchmark MLLMs, we test various multimodal prompting methods on frontier models such as GPT-4o, GPT-4V, Gemini, and Claude. Our fine-grained break-down metrics indicate that models mostly lag in recalling visual elements from the input webpages and generating correct layout designs.