NAACL 2025 Papers

Effective linguistic choices that attract potential customers play crucial roles in advertising success. This study aims to explore the linguistic features of ad texts that influence human preferences. Although the creation of attractive ad texts is an active area of research, progress in understanding the specific linguistic features that affect attractiveness is hindered by several obstacles. First, human preferences are complex and influenced by multiple factors, including their content, such as brand names, and their linguistic styles, making analysis challenging. Second, publicly available ad text datasets that include human preferences are lacking, such as ad performance metrics and human feedback, which reflect people’s interests. To address these problems, we present AdParaphrase, a paraphrase dataset that contains human preferences for pairs of ad texts that are semantically equivalent but differ in terms of wording and style. This dataset allows for preference analysis that focuses on the differences in linguistic features. Our analysis revealed that ad texts preferred by human judges have higher fluency, longer length, more nouns, and use of bracket symbols. Furthermore, we demonstrate that an ad text-generation model that considers these findings significantly improves the attractiveness of a given text. The dataset is publicly available at: https://github.com/CyberAgentAILab/AdParaphrase.

The rise of Large Language Models (LLMs) has boosted the use of Few-Shot Learning (FSL) methods in natural language processing, achieving acceptable performance even when working with limited training data. The goal of FSL is to effectively utilize a small number of annotated samples in the learning process. However, the performance of FSL suffers when unsuitable support samples are chosen. This problem arises due to the heavy reliance on a limited number of support samples, which hampers consistent performance improvement even when more support samples are added. To address this challenge, we propose an active learning-based instance selection mechanism that identifies effective support instances from the unlabeled pool and can work with different LLMs. Our experiments on five tasks show that our method frequently improves the performance of FSL. We make our implementation available on GitHub.

For centuries, writers have hidden messages as acrostics, in which initial letters of consecutive lines or paragraphs form meaningful words or phrases. Scholars searching for acrostics manually can only focus on a few authors at a time and often favor qualitative arguments about whether a given acrostic is accidental or intentional. Here we describe AcrosticSleuth, a first-of-its-kind approach to identify acrostics automatically and rank them by the probability that the corresponding sequence of characters does not occur by chance. Since acrostics are rare, we formalize the problem as a binary classification task in the presence of extreme class imbalance. To evaluate AcrosticSleuth, we present the Acrostic Identification Dataset (AcrostID), a collection of acrostics from the WikiSource online database. Despite the class imbalance, AcrosticSleuth achieves F1 scores of 0.39, 0.59, and 0.66 on the French, English, and Russian subdomains of WikiSource, respectively. We further demonstrate that AcrosticSleuth can identify previously unknown instances of wordplay in high-profile literary contexts, including the English philosopher Thomas Hobbes’ signature in the opening paragraphs of The Elements of Law.

Effective human-machine collaboration requires machine learning models to externalize uncertainty, so users can reflect and intervene when necessary. For language models, these representations of uncertainty may be impacted by sycophancy bias: proclivity to agree with users, even if they are wrong. For instance, models may be over-confident in (incorrect) problem solutions suggested by a user. We study the relationship between sycophancy and uncertainty estimation for the first time. We propose a generalization of the definition of sycophancy bias to measure downstream impacts on uncertainty estimation, and also propose a new algorithm (SyRoUP) to account for sycophancy in the uncertainty estimation process. Unlike previous works, we study a broad array of user behaviors, varying both correctness and confidence of user suggestions to see how model answers (and their certainty) change. Our experiments across conversation forecasting and question-answering tasks show that user confidence plays a critical role in modulating the effects of sycophancy, and that SyRoUP can better predict these effects. From these results, we argue that externalizing both model and user uncertainty can help to mitigate the impacts of sycophancy bias.

Many endangered languages are at risk of extinction due to barriers in communication and generational gaps that hinder their preservation. A cause for languages becoming endangered is the lack of language educational tools and artificial intelligence (AI) models for these low-resource languages. To address this, we propose the ATAIGI learning app designed with AI-powered models leveraging multimodal generative techniques. Our app offers users a comprehensive learning experience by providing translated phrases and definitions, example sentences, illustrative images, romanized pronunciation, and audio speech to accelerate language learning. ATAIGI is built on five AI models that are rigorously benchmarked individually, with our Transliteration Model achieving state-of-the-art results for Taiwanese Hokkien transliteration. ATAIGI is available for all to learn the endangered language of Taiwanese Hokkien, an endangered language spoken in Taiwan. A human evaluation conducted demonstrates the effectiveness of ATAIGI in improving language proficiency and cultural understanding, supporting its potential for the preservation and education of endangered languages like the Taiwanese Hokkien.

Retrieval-Augmented Generation (RAG) models have drawn considerable attention in modern open-domain question answering. The effectiveness of RAG depends on the quality of the top retrieved documents. However, conventional retrieval methods sometimes fail to rank the most relevant documents at the top. In this paper, we introduce ASRANK, a new re-ranking method based on scoring retrieved documents using zero-shot answer scent which relies on a pre-trained large language model to compute the likelihood of the document-derived answers aligning with the answer scent. Our approach demonstrates marked improvements across several datasets, including NQ, TriviaQA, WebQA, ArchivalQA, HotpotQA, and Entity Questions. Notably, ASRANK increases Top-1 retrieval accuracy on NQ from 19.2% to 46.5% for MSS and 22.1% to 47.3% for BM25. It also shows strong retrieval performance on several datasets compared to state-of-the-art methods (47.3 Top-1 by ASRANK vs 35.4 by UPR by BM25).

We propose ARISE, a framework that iteratively induces rules and generates synthetic data for text classification. We combine synthetic data generation and automatic rule induction, via bootstrapping, to iteratively filter the generated rules and data. We induce rules via inductive generalisation of syntactic-ngrams, enabling us to capture a complementary source of supervision. These rules alone lead to performance gains in both, in-context learning (ICL) and fine-tuning (FT) settings. Similarly, use of augmented data from ARISE alone improves the performance for a model, outperforming configurations that rely on complex methods like contrastive learning. Further, our extensive experiments on various datasets covering three full-shot, eight few-shot and seven multilingual variant settings demonstrate that the rules and data we generate lead to performance improvements across these diverse domains and languages.

Spontaneous or conversational multilingual speech presents many challenges for state-of-the-art automatic speech recognition (ASR) systems. In this work, we present a new technique AMPS, that augments a multilingual multimodal ASR system with paraphrase-based supervision for improved conversational ASR in multiple languages, including Hindi, Marathi, Malayalam, Kannada, and Nyanja. We use paraphrases of the reference transcriptions as additional supervision while training the multimodal ASR model and selectively invoke this paraphrase objective for utterances with poor ASR performance. Using AMPS with a state-of-the-art multimodal model SeamlessM4T, we obtain significant relative reductions in word error rates (WERs) of up to 5%. We present detailed analyses of our system using both objective and human evaluation metrics.

Large Language Models (LLMs) heavily rely on high-quality training data, making data valuation crucial for optimizing model performance, especially when working within a limited budget. In this work, we aim to offer a third-party data valuation approach that benefits both data providers and model developers. We introduce a linearized future influence kernel (LinFiK), which assesses the value of individual data samples in improving LLM performance during training. We further propose ALinFiK, a learning strategy to approximate LinFiK, enabling scalable data valuation. Our comprehensive evaluations demonstrate that this approach surpasses existing baselines in effectiveness and efficiency, demonstrating significant scalability advantages as LLM parameters increase.

Large Language Model (LLM) can enhance its credibility and verifiability by generating text with citations. However, existing research on citation generation is predominantly limited to sentence-level statements, neglecting the significance of positional fine-grained citations that can appear anywhere within sentences. To facilitate further exploration of the positional fine-grained citation generation, we propose ALiiCE, the first automatic evaluation framework for this task. Our method employs a dependency tree based approach to parse the sentence-level claim into atomic claims. Then ALiiCE evaluates citation quality using three metrics, including positional fine-grained citation recall, precision, and coefficient of variation of citation positions. We evaluate the positional fine-grained citation generation performance of several LLMs on long-form QA datasets. Our experiments and analyses demonstrate the effectiveness and reasonableness of ALiiCE. We offer our insights into the current advancements and future directions for the positional fine-grained citation generation task.

In this paper, we investigate the overlooked impact of instruction-tuning on memorization in large language models (LLMs), which has largely been studied in base, pre-trained models. We propose a black-box prompt optimization method where an attacker LLM agent uncovers higher levels of memorization in a victim agent, surpassing traditional approaches that prompt the model directly with training data. Using an iterative rejection-sampling process, we design instruction-based prompts that minimize overlap with training data to avoid providing direct solutions while maximizing overlap between the victim’s output and the training data to induce memorization. Our method shows 23.7% more overlap with training data compared to state-of-the-art baselines. We explore two attack settings: an analytical approach that determines the empirical upper bound of the attack, both with and without access to responses for prompt initialization, and a practical classifier-based method for assessing memorization without access to memorized data. Our findings reveal that instruction-tuned models can expose pre-training data as much as, or more than, base models; contexts beyond the original training data can lead to leakage; and instructions generated by other LLMs open new avenues for automated attacks, which we believe require further exploration.

The rise of Large Language Models (LLMs) revolutionizes information retrieval, allowing users to obtain required answers through complex instructions within conversations. However, publicly available services remain inadequate in addressing the needs of faculty and students to search campus-specific information. It is primarily due to the LLM’s lack of domain-specific knowledge and the limitation of search engines in supporting multilingual and timely scenarios. To tackle these challenges, we introduce ALOHA, a multilingual agent enhanced by hierarchical retrieval for university orientation. We also integrate external APIs into the front-end interface to provide interactive service. The human evaluation and case study show our proposed system has strong capabilities to yield correct, timely, and user-friendly responses to the queries in multiple languages, surpassing commercial chatbots and search engines. The system has been deployed and has provided service for more than 12,000 people.

Recommendation explanation systems have become increasingly vital with the widespread adoption of recommender systems. However, existing recommendation explanation evaluation benchmarks suffer from limited item diversity, impractical user profiling requirements, and unreliable and unscalable evaluation protocols. We present ALERT, a model-agnostic recommendation explanation evaluation benchmark. The benchmark comprises three main contributions: 1) a diverse dataset encompassing 15 Amazon e-commerce categories with 2,761 user-item interactions, incorporating implicit preferences through purchase histories;2) two novel LLM-powered automatic evaluators that enable scalable and human-preference aligned evaluation of explanations; and 3) a robust divide-and-aggregate approach that synthesizes multiple LLM judgments, achieving 70% concordance with expert human evaluation and substantially outperforming existing methods.ALERT facilitates comprehensive evaluation of recommendation explanations across diverse domains, advancing the development of more effective explanation systems.

As Large Language Models (LLMs) increasingly influence content generation across diverse platforms, there is a heightened urgency to regulate their outputs to ensure safe usage. However, defining safety is complex, given that entities across domains may interpret it through varied lenses and develop safety detectors—models trained to identify specific unsafe content based on predefined criteria. To address this complexity, we introduce the approach of Adaptive Integration of Detectors (AID) to orchestrate the strengths of multiple pretrained detectors to ensure comprehensive effectiveness in diverse scenarios. AID employs a Mixture-of-Experts (MoE) framework, wherein it dynamically assigns and learns data-adaptive weights for each detector using domain-specific annotated data and LLM-extracted features. We provide theoretical insights into why MoE can be effective by showing its optimality in a Neyman-Pearson setting. Our experimental studies using various detection tasks curated from benchmark datasets demonstrate AID’s ability to synergistically combine the unique capabilities of individual detectors. For example, it is observed that AID can improve the area under the curve (AUC) by an absolute value of 0.07 to 0.21, with a median of 0.12, compared to the best individual detectors developed for specific safety aspects. The improvement is particularly significant for complex detection tasks that mix different unsafe data sources.

Truthfulness (adherence to factual accuracy) and utility (satisfying human needs and instructions) are both fundamental aspects of Large Language Models, yet these goals often conflict (e.g., sell a car with known flaws), making it challenging to achieve both in real-world deployments. We propose AI-LieDar, a framework to study how LLM-based agents navigate these scenarios in an multi-turn interactive setting. We design a set of real-world scenarios where language agents are instructed to achieve goals that are in conflict with being truthful during a multi-turn conversation with simulated human agents. To evaluate the truthfulness at large scale, we develop a truthfulness detector inspired by psychological literature to assess the agents’ responses. Our experiment demonstrates that all models are truthful less than 50% of the time, although truthfulness and goal achievement (utility) rates vary across models. We further test the steerability of LLMs towards truthfulness, finding that models can be directed to be deceptive, and even truth-steered models still lie. These findings reveal the complex nature of truthfulness in LLMs and underscore the importance of further research to ensure the safe and reliable deployment of LLMs and AI agents.

As Large Language Models (LLMs) and generative AI become increasingly widespread, concerns about content safety have grown in parallel. Currently, there is a clear lack of high-quality, human-annotated datasets that address the full spectrum of LLM-related safety risks and are usable for commercial applications. To bridge this gap, we propose a comprehensive and adaptable taxonomy for categorizing safety risks, structured into 12 top-level hazard categories with an extension to 9 fine-grained subcategories. This taxonomy is designed to meet the diverse requirements of downstream users, offering more granular and flexible tools for managing various risk types. Using a hybrid data generation pipeline that combines human annotations with a multi-LLM “jury” system to assess the safety of responses we obtain Aegis2.0, a carefully curated collection of 34,248 samples of human-LLM interactions, annotated according to our proposed taxonomy. To validate its effectiveness, we demonstrate that several lightweight models, trained using parameter-efficient techniques on Aegis2.0, achieve performance competitive with leading safety models fully fine-tuned on much larger, non-commercial datasets generated leveraging GPT-4. Additionally, we introduce a novel training blend that combines topic following data with safety data. This approach enhances the adaptability of guard models, enabling them to generalize to new risk categories defined during inference. We plan to open-source Aegis2.0 data and models to the research community to aid in safety guardrailing of LLMs.

We present ACCORD, a framework and benchmark suite for disentangling the commonsense grounding and reasoning abilities of large language models (LLMs) through controlled, multi-hop counterfactuals. ACCORD introduces formal elements to commonsense reasoning to explicitly control and quantify reasoning complexity beyond the typical 1 or 2 hops. Uniquely, ACCORD can automatically generate benchmarks of arbitrary reasoning complexity, so it scales with future LLM improvements. Indeed, our experiments on state-of-the-art LLMs show performance degrading to below random chance with only moderate scaling, leaving substantial headroom for improvement. We release a leaderboard of the benchmark suite tested in this work, as well as code for automatically generating more complex benchmarks.

Dialogue State Tracking (DST) is crucial for understanding user needs and executing appropriate system actions in task-oriented dialogues. Majority of existing DST methods are designed to work within predefined ontologies and assume the availability of gold domain labels, struggling with adapting to new slots values. While Large Language Models (LLMs)-based systems show promising zero-shot DST performance, they either require extensive computational resources or they underperform existing fully-trained systems, limiting their practicality. To address these limitations, we propose a zero-shot, open-vocabulary system that integrates domain classification and DST in a single pipeline. Our approach includes reformulating DST as a question-answering task for less capable models and employing self-refining prompts for more adaptable ones. Our system does not rely on fixed slot values defined in the ontology allowing the system to adapt dynamically. We compare our approach with existing SOTA, and show that it provides up to 20% better Joint Goal Accuracy (JGA) over previous methods on datasets like MultiWOZ 2.1, with up to 90% fewer requests to the LLM API.

The Dialogue Topic Segmentation task aims to divide a dialogue into different topic paragraphs in order to better understand the structure and content of the dialogue. Due to the short sentences, serious references and non-standard language in the dialogue, it is difficult to determine the boundaries of the topic. Although the unsupervised approaches based on LLMs performs well, it is still difficult to surpass the supervised methods based on classical models in specific domains. To this end, this paper proposes UPS (Utterance Pair Segment), a dialogue topic segmentation method based on utterance pair relationship modeling, unifying the supervised and unsupervised network architectures. For supervised pre-training, the model predicts the adjacency and topic affiliation of utterances in dialogues. For unsupervised pre-training, the dialogue-level and utterance-level relationship prediction tasks are used to train the model. The pre-training and fine-tuning strategies are carried out in different scenarios, such as supervised, few-shot, and unsupervised data. By adding a domain adapter and a task adapter to the Transformer, the model learns in the pre-training and fine-tuning stages, respectively, which significantly improves the segmentation effect. As the result, the proposed method has achieved the best results on multiple benchmark datasets across various scenarios.