NAACL 2024 论文列表

Most hate speech datasets neglect the cultural diversity within a single language, resulting in a critical shortcoming in hate speech detection. To address this, we introduce CREHate, a CRoss-cultural English Hate speech dataset. To construct CREHate, we follow a two-step procedure: 1) cultural post collection and 2) cross-cultural annotation. We sample posts from the SBIC dataset, which predominantly represents North America, and collect posts from four geographically diverse English-speaking countries (Australia, United Kingdom, Singapore, and South Africa) using culturally hateful keywords we retrieve from our survey. Annotations are collected from the four countries plus the United States to establish representative labels for each country. Our analysis highlights statistically significant disparities across countries in hate speech annotations. Only 56.2% of the posts in CREHate achieve consensus among all countries, with the highest pairwise label difference rate of 26%. Qualitative analysis shows that label disagreement occurs mostly due to different interpretations of sarcasm and the personal bias of annotators on divisive topics. Lastly, we evaluate large language models (LLMs) under a zero-shot setting and show that current LLMs tend to show higher accuracies on Anglosphere country labels in CREHate.Our dataset and codes are available at: https://github.com/nlee0212/CREHate

Multiple-choice questions (MCQs) are ubiquitous in almost all levels of education since they are easy to administer, grade, and are a reliable format in assessments and practices. One of the most important aspects of MCQs is the distractors, i.e., incorrect options that are designed to target common errors or misconceptions among real students. To date, the task of crafting high-quality distractors largely remains a labor and time-intensive process for teachers and learning content designers, which has limited scalability. In this work, we study the task of automated distractor generation in the domain of math MCQs and explore a wide variety of large language model (LLM)-based approaches, from in-context learning to fine-tuning. We conduct extensive experiments using a real-world math MCQ dataset and find that although LLMs can generate some mathematically valid distractors, they are less adept at anticipating common errors or misconceptions among real students.

Hierarchical classification frameworks have been widely used to process long sequences, especially in the legal domain for predictions from long legal documents. But being black-box models they are unable to explain their predictions making them less reliable for practical applications, more so in the legal domain. In this work, we develop an extractive explanation algorithm for hierarchical frameworks for long sequences based on the sensitivity of the trained model to its input perturbations. We perturb using occlusion and develop Ob-HEx; an Occlusion-based Hierarchical Explanation-extractor. We adapt Ob-HEx to Hierarchical Transformer models trained on long Indian legal texts. And use Ob-HEx to analyze them and extract their explanations for the ILDC-Expert dataset, achieving a minimum gain of 1 point over the previous benchmark on most of our performance evaluation metrics.

As Transformers have become state-of-the-art models for natural language processing (NLP) tasks, the need to understand and explain their predictions is increasingly apparent. Especially in unsupervised applications, such as information retrieval tasks, similarity models built on top of foundation model representations have been widely applied. However, their inner prediction mechanisms have mostly remained opaque. Recent advances in explainable AI have made it possible to mitigate these limitations by leveraging improved explanations for Transformers through layer-wise relevance propagation (LRP). Using BiLRP, an extension developed for computing second-order explanations in bilinear similarity models, we investigate which feature interactions drive similarity in NLP models. We validate the resulting explanations and demonstrate their utility in three corpus-level use cases, analyzing grammatical interactions, multilingual semantics, and biomedical text retrieval. Our findings contribute to a deeper understanding of different semantic similarity tasks and models, highlighting how novel explainable AI methods enable in-depth analyses and corpus-level insights.

As language models are adopted by a more sophisticated and diverse set of users, the importance of guaranteeing that they provide factually correct information supported by verifiable sources is critical across fields of study. This is especially the case for high-stakes fields, such as medicine and law, where the risk of propagating false information is high and can lead to undesirable societal consequences. Previous work studying attribution and factuality has not focused on analyzing these characteristics of language model outputs in domain-specific scenarios. In this work, we conduct human evaluation of responses from a few representative systems along various axes of attribution and factuality, by bringing domain experts in the loop. Specifically, we collect expert-curated questions from 484 participants across 32 fields of study, and then ask the same experts to evaluate generated responses to their own questions. In addition, we ask experts to improve upon responses from language models. The output of our analysis is ExpertQA, a high-quality long-form QA dataset with 2177 questions spanning 32 fields, along with verified answers and attributions for claims in the answers.

Recent work has proposed explicitly inducing language-wise modularity in multilingual LMs via sparse fine-tuning (SFT) on per-language subnetworks as a means of better guiding cross-lingual sharing. In this paper, we investigate (1) the degree to which language-wise modularity *naturally* arises within models with no special modularity interventions, and (2) how cross-lingual sharing and interference differ between such models and those with explicit SFT-guided subnetwork modularity. In order to do so, we use XLM-R as our multilingual LM. Moreover, to quantify language specialization and cross-lingual interaction, we use a Training Data Attribution method that estimates the degree to which a model’s predictions are influenced by in-language or cross-language training examples. Our results show that language-specialized subnetworks do naturally arise, and that SFT, rather than always increasing modularity, can decrease language specialization of subnetworks in favor of more cross-lingual sharing.

The proliferation of online misinformation has posed significant threats to public interest. While numerous online users actively participate in the combat against misinformation, many of such responses can be characterized by the lack of politeness and supporting facts. As a solution, text generation approaches are proposed to automatically produce counter-misinformation responses. Nevertheless, existing methods are often trained end-to-end without leveraging external knowledge, resulting in subpar text quality and excessively repetitive responses. In this paper, we propose retrieval augmented response generation for online misinformation (RARG), which collects supporting evidence from scientific sources and generates counter-misinformation responses based on the evidences. In particular, our RARG consists of two stages: (1) evidence collection, where we design a retrieval pipeline to retrieve and rerank evidence documents using a database comprising over 1M academic articles; (2) response generation, in which we align large language models (LLMs) to generate evidence-based responses via reinforcement learning from human feedback (RLHF). We propose a reward function to maximize the utilization of the retrieved evidence while maintaining the quality of the generated text, which yields polite and factual responses that clearly refutes misinformation. To demonstrate the effectiveness of our method, we study the case of COVID-19 and perform extensive experiments with both in- and cross-domain datasets, where RARG consistently outperforms baselines by generating high-quality counter-misinformation responses.

Understanding complex events from different modalities, associating to external knowledge and generating response in a clear point of view are still unexplored in today’s multi-modal dialogue research. The great challenges include 1) lack of event-based multi-modal dialogue dataset; 2) understanding of complex events and 3) heterogeneity gap between different modalities. To overcome these challenges, we firstly introduce a novel event-oriented video-dialogue dataset called SportsVD (Sports-domain Video-dialogue Dataset). To our best knowledge, SportsVD is the first dataset that consists of complex events videos and opinion-based conversations with regards to contents in these events. Meanwhile, we present multi-modal dialogue generation method VCD (Video Commentary Dialogue) to generate human-like response according to event contents in the video and related external knowledge. In contrast to previous video-based dialogue generation, we focus on opinion-based response and the understanding of longer and more complex event contents. We evaluate VCD’s performance on SportsVD and other baselines under several automatic metrics. Experiments demonstrate VCD can outperform among other state-of-the-art baselines. Our work is available at https://github.com/Cheng-Fenghua/SportsVD.

Social media is an easy-to-access platform providing timely updates about societal trends and events. Discussions regarding epidemic-related events such as infections, symptoms, and social interactions can be crucial for informing policymaking during epidemic outbreaks. In our work, we pioneer exploiting Event Detection (ED) for better preparedness and early warnings of any upcoming epidemic by developing a framework to extract and analyze epidemic-related events from social media posts. To this end, we curate an epidemic event ontology comprising seven disease-agnostic event types and construct a Twitter dataset SPEED with human-annotated events focused on the COVID-19 pandemic. Experimentation reveals how ED models trained on COVID-based SPEED can effectively detect epidemic events for three unseen epidemics of Monkeypox, Zika, and Dengue; while models trained on existing ED datasets fail miserably. Furthermore, we show that reporting sharp increases in the extracted events by our framework can provide warnings 4-9 weeks earlier than the WHO epidemic declaration for Monkeypox. This utility of our framework lays the foundations for better preparedness against emerging epidemics.

Cognitive science and symbolic AI research suggest that event causality provides vital information for story understanding. However, machine learning systems for story understanding rarely employ event causality, partially due to the lack of methods that reliably identify open-world causal event relations. Leveraging recent progress in large language models, we present the first method for event causality identification that leads to material improvements in computational story understanding. Our technique sets a new state of the art on the COPES dataset (Wang et al., 2023c) for causal event relation identification. Further, in the downstream story quality evaluation task, the identified causal relations lead to 3.6-16.6% relative improvement on correlation with human ratings. In the multimodal story video-text alignment task, we attain 4.1-10.9% increase on Clip Accuracy and 4.2-13.5% increase on Sentence IoU. The findings indicate substantial untapped potential for event causality in computational story understanding. The codebase is at https://github.com/insundaycathy/Event-Causality-Extraction.

The development of highly fluent large language models (LLMs) has prompted increased interest in assessing their reasoning and problem-solving capabilities. We investigate whether several LLMs can solve a classic type of deductive reasoning problem from the cognitive science literature. The tested LLMs have limited abilities to solve these problems in their conventional form. We performed follow up experiments to investigate if changes to the presentation format and content improve model performance. We do find performance differences between conditions; however, they do not improve overall performance. Moreover, we find that performance interacts with presentation format and content in unexpected ways that differ from human performance. Overall, our results suggest that LLMs have unique reasoning biases that are only partially predicted from human reasoning performance and the human-generated language corpora that informs them.

Pre-trained language models (LMs) have shown remarkable reasoning performance using explanations or chain-of-thoughts (CoT)) for in-context learning. On the other hand, these reasoning tasks are usually presumed to be more approachable for symbolic programming. To understand the mechanism of reasoning of LMs, we curate synthetic datasets containing equivalent (natural, symbolic) data pairs, where symbolic examples contain first-order logic rules and predicates from non-parametric knowledge bases (KBs), supporting automated verification of intermediate reasoning results. Then we revisit neuro-symbolic approaches and propose to learn from demonstrations containing logic rules and corresponding examples to iteratively reason over KBs, recovering Prolog’s backward chaining algorithm and supporting automated verification of LMs’ outputs. Comprehensive experiments are included to systematically compare LMLP with CoT in deductive reasoning settings, showing that LMLP enjoys more than 25% higher accuracy than CoT on length generalization benchmarks even with smaller model sizes.

Synthetic users are cost-effective proxies for real users in the evaluation of conversational recommender systems. Large language models show promise in simulating human-like behavior, raising the question of their ability to represent a diverse population of users. We introduce a new protocol to measure the degree to which language models can accurately emulate human behavior in conversational recommendation. This protocol is comprised of five tasks, each designed to evaluate a key property that a synthetic user should exhibit: choosing which items to talk about, expressing binary preferences, expressing open-ended preferences, requesting recommendations, and giving feedback. Through evaluation of baseline simulators, we demonstrate these tasks effectively reveal deviations of language models from human behavior, and offer insights on how to reduce the deviations with model selection and prompting strategies.

Contemporary Large Language Models (LLMs) exhibit a high degree of code generation and comprehension capability. A particularly promising area is their ability to interpret code modules from unfamiliar libraries for solving user-instructed tasks. Recent work has shown that large proprietary LLMs can learn novel library usage in-context from demonstrations. These results raise several open questions: whether demonstrations of library usage is required, whether smaller (and more open) models also possess such capabilities, etc. In this work, we take a broader approach by systematically evaluating a diverse array of LLMs across three scenarios reflecting varying levels of domain specialization to understand their abilities and limitations in generating code based on libraries defined in-context. Our results show that even smaller open-source LLMs like Llama-2 and StarCoder demonstrate an adept understanding of novel code libraries based on specification presented in-context. Our findings further reveal that LLMs exhibit a surprisingly high proficiency in learning novel library modules even when provided with just natural language descriptions or raw code implementations of the functions, which are often cheaper to obtain than demonstrations. Overall, our results pave the way for harnessing LLMs in more adaptable and dynamic coding environments.

Language models (LMs) have greatly propelled the research on natural language processing. However, LMs also raise concerns regarding the generation of biased or toxic content and the potential disclosure of private information from the training dataset. In this work, we present a new efficient approach, Ethos, that rectifies LMs to mitigate toxicity and bias in outputs and avoid privacy leakage. Ethos is built on task arithmetic. However, unlike current task arithmetic algorithms, Ethos distinguishes general beneficial and undesired knowledge when reconstructing task vectors. Specifically, Ethos first obtains a set of principal components from the pre-trained models using singular value decomposition. Then, by projecting the task vector onto principal components, Ethos separates the principal components that encode general from those associated with undesired knowledge. Ethos performs forgetting or unlearning by only negating the task vector with undesired knowledge, thereby minimizing collateral damage on general model utility. We demonstrate the efficacy of our approach on three different tasks: bias, toxicity, and memorization unlearning. Evaluations show Ethos is more effective in removing undesired knowledge while maintaining the overall model performance compared to current task arithmetic methods.

Entity disambiguation (ED), which links the mentions of ambiguous entities to their referent entities in a knowledge base, serves as a core component in entity linking (EL). Existing generative approaches demonstrate improved accuracy compared to classification approaches under the standardized ZELDA benchmark. Nevertheless, generative approaches suffer from the need for large-scale pre-training and inefficient generation. Most importantly, entity descriptions, which could contain crucial information to distinguish similar entities from each other, are often overlooked.We propose an encoder-decoder model to disambiguate entities with more detailed entity descriptions. Given text and candidate entities, the encoder learns interactions between the text and each candidate entity, producing representations for each entity candidate. The decoder then fuses the representations of entity candidates together and selects the correct entity.Our experiments, conducted on various entity disambiguation benchmarks, demonstrate the strong and robust performance of this model, particularly +1.5% in the ZELDA benchmark compared with GENRE. Furthermore, we integrate this approach into the retrieval/reader framework and observe +1.5% improvements in end-to-end entity linking in the GERBIL benchmark compared with EntQA.

Large Language Models (LLMs) exhibit impressive capabilities but also present risks such as biased content generation and privacy issues. One of the current alignment techniques includes principle-driven integration, but it faces challenges arising from the imprecision of manually crafted rules and inadequate risk perception in models without safety training. To address these, we introduce Guide-Align, a two-stage approach. Initially, a safety-trained model identifies potential risks and formulates specific guidelines for various inputs, establishing a comprehensive library of guidelines and a model for input-guidelines retrieval. Subsequently, the retrieval model correlates new inputs with relevant guidelines, which guide LLMs in response generation to ensure safe and high-quality outputs, thereby aligning with human values. An additional optional stage involves fine-tuning a model with well-aligned datasets generated through the process implemented in the second stage.Our method customizes guidelines to accommodate diverse inputs, thereby enhancing the fine-grainedness and comprehensiveness of the guideline library. Furthermore, it incorporates safety expertise from a safety-trained LLM through a lightweight retrieval model.We evaluate our approach on three benchmarks, demonstrating significant improvements in LLM security and quality. Notably, our fine-tuned model, Labrador, even at 13 billion parameters, outperforms GPT-3.5-turbo and surpasses GPT-4 in alignment capabilities.

Open-source pre-trained Large Language Models (LLMs) exhibit strong language understanding and generation capabilities, making them highly successful in a variety of tasks. However, when used as agents for dealing with complex problems in the real world, their performance is far inferior to large commercial models such as ChatGPT and GPT-4. As intelligent agents, LLMs need to have the capabilities of task planning, long-term memory, and the ability to leverage external tools to achieve satisfactory performance. Various methods have been proposed to enhance the agent capabilities of LLMs. On the one hand, methods involve constructing agent-specific data and fine-tuning the models. On the other hand, some methods focus on designing prompts that effectively activate the reasoning abilities of the LLMs. We explore both strategies on the 7B and 13B models. We propose a comprehensive method for constructing agent-specific data using GPT-4. Through supervised fine-tuning with constructed data, we find that for these models with a relatively small number of parameters, supervised fine-tuning can significantly reduce hallucination outputs and formatting errors in agent tasks. Furthermore, techniques such as multi-path reasoning and task decomposition can effectively decrease problem complexity and enhance the performance of LLMs as agents. We evaluate our method on five agent tasks of AgentBench and achieve satisfactory results.

We introduce Enhancing Perception, a framework for Large Language Models (LLMs) designed to streamline the time-intensive task typically undertaken by professional fact-checkers of crafting explanations for fake news. This study investigates the effectiveness of enhancing LLM explanations through conversational refinement. We compare various questioner agents, including state-of-the-art LLMs like GPT-4, Claude 2, PaLM 2, and 193 American participants acting as human questioners. Based on the histories of these refinement conversations, we further generate comprehensive summary explanations. We evaluated the effectiveness of these initial, refined, and summary explanations across 40 news claims by involving 2,797 American participants, measuring their self-reported belief change regarding both real and fake claims after receiving the explanations. Our findings reveal that, in the context of fake news, explanations that have undergone conversational refinement—whether by GPT-4 or human questioners, who ask more diverse and detail-oriented questions—were significantly more effective than both the initial unrefined explanations and the summary explanations. Moreover, these refined explanations achieved a level of effectiveness comparable to that of expert-written explanations. The results highlight the potential of automatic explanation refinement by LLMs in debunking fake news claims.