ACL 2024 论文列表

In legal decisions, split votes (SV) occur when judges cannot reach a unanimous decision, posing a difficulty for lawyers who must navigate diverse legal arguments and opinions. In high-stakes domains, %as human-AI interaction systems become increasingly important, understanding the alignment of perceived difficulty between humans and AI systems is crucial to build trust. However, existing NLP calibration methods focus on a classifier’s awareness of predictive performance, measured against the human majority class, overlooking inherent human label variation (HLV). This paper explores split votes as naturally observable human disagreement and value pluralism. We collect judges’ vote distributions from the European Court of Human Rights (ECHR), and present SV-ECHR, a case outcome classification (COC) dataset with SV information. We build a taxonomy of disagreement with SV-specific subcategories. We further assess the alignment of perceived difficulty between models and humans, as well as confidence- and human-calibration of COC models. We observe limited alignment with the judge vote distribution. To our knowledge, this is the first systematic exploration of calibration to human judgements in legal NLP. Our study underscores the necessity for further research on measuring and enhancing model calibration considering HLV in legal decision tasks.

With the advent of large language models (LLM), the line between human-crafted and machine-generated texts has become increasingly blurred. This paper delves into the inquiry of identifying discernible and unique linguistic properties in texts that were written by humans, particularly uncovering the underlying discourse structures of texts beyond their surface structures. Introducing a novel methodology, we leverage hierarchical parse trees and recursive hypergraphs to unveil distinctive discourse patterns in texts produced by both LLMs and humans. Empirical findings demonstrate that, although both LLMs and humans generate distinct discourse patterns influenced by specific domains, human-written texts exhibit more structural variability, reflecting the nuanced nature of human writing in different domains. Notably, incorporating hierarchical discourse features enhances binary classifiers’ overall performance in distinguishing between human-written and machine-generated texts, even on out-of-distribution and paraphrased samples. This underscores the significance of incorporating hierarchical discourse features in the analysis of text patterns. The code and dataset will be available at [TBA].

There are two key challenges remaining for the document-level event argument extraction (D-EAE) tasks: key feature forgetting and cross-event argument confusion. The emergence capability of large language models (LLMs) holds promise for solving the above two challenges. In this paper, we propose a document-level event argument extraction method based on guided summarization and reasoning (EAESR), which leverages the emergence capabilities of LLMs to highlight key event information and to clarify the explicit and implicit association between multiple events. Specifically, we generate document summarization information that shorten the length of the event context while preserving the key event features. In addition, we generate inter-event reasoning information, which helps EAESR make sense of the correlations between events and reduces their dependence on the event context, especially to better cope with the few-shot D-EAE task. Then, we obtain named entity information to enable EAESR to learn argument boundary features to improve the sensitivity of its argument boundary recognition. Eventually, we fused the above features and sentence features to make EAESR have summarizing and reasoning capabilities simultaneously. Extensive experiments on WIKIEVENTS and RAMS have shown that EAESR achieves a new state-of-the-art that outperforms the baseline models by 1.3% F1 and 1.6% F1, respectively, and averages 11% F1 in few-shot settings.

Human interactions are deeply rooted in the interplay of thoughts, beliefs, and desires made possible by Theory of Mind (ToM): our cognitive ability to understand the mental states of ourselves and others. Although ToM may come naturally to us, emulating it presents a challenge to even the most advanced Large Language Models (LLMs). Recent improvements to LLMs’ reasoning capabilities from simple yet effective prompting techniques such as Chain-of-Thought (CoT) have seen limited applicability to ToM. In this paper, we turn to the prominent cognitive science theory “Simulation Theory” to bridge this gap. We introduce SimToM, a novel two-stage prompting framework inspired by Simulation Theory’s notion of perspective-taking. To implement this idea on current ToM benchmarks, SimToM first filters context based on what the character in question knows before answering a question about their mental state. Our approach, which requires no additional training and minimal prompt-tuning, shows substantial improvement over existing methods, and our analysis reveals the importance of perspective-taking to Theory-of-Mind capabilities. Our findings suggest perspective-taking as a promising direction for future research into improving LLMs’ ToM capabilities.

How can we train models to perform well on hard test data when hard training data is by definition difficult to label correctly? This question has been termed the scalable oversight problem and has drawn increasing attention as language models have continually improved. In this paper, we present the surprising conclusion that current pretrained language models often generalize relatively well from easy to hard data, even performing as well as oracle models finetuned on hard data. We demonstrate this kind of easy-to-hard generalization using simple finetuning methods like in-context learning, linear classifier heads, and QLoRA for seven different measures of datapoint hardness, including six empirically diverse human hardness measures (like grade level) and one model-based measure (loss-based). Furthermore, we show that even if one cares most about model performance on hard data, it can be better to collect easy data rather than hard data for finetuning, since hard data is generally noisier and costlier to collect. Our experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. We conclude that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied.

Relation extraction (RE) extracts structured tuples of relationships (e.g. friend, enemy) between entities (e.g. Sherlock Holmes, John Watson) from text, with exciting potential applications. Hundreds of RE papers have been published in recent years; do their evaluation practices inform these goals? We review recent surveys and a sample of recent RE methods papers, compiling 38 datasets currently being used. Unfortunately, many have frequent label errors, and ones with known problems continue to be used. Many datasets focus on producing labels for a large number of relation types, often through error-prone annotation methods (e.g. distant supervision or crowdsourcing), and many recent papers rely exclusively on such datasets. We draw attention to a promising alternative: datasets with a small number of relations, often in specific domains like chemistry, finance, or biomedicine, where it is possible to obtain high quality expert annotations; such data can more realistically evaluate RE performance. The research community should consider more often using such resources.

Connecting text and visual modalities plays an essential role in generative intelligence. For this reason, inspired by the success of large language models, significant research efforts are being devoted to the development of Multimodal Large Language Models (MLLMs). These models can seamlessly integrate visual and textual modalities, while providing a dialogue-based interface and instruction-following capabilities. In this paper, we provide a comprehensive review of recent visual-based MLLMs, analyzing their architectural choices, multimodal alignment strategies, and training techniques. We also conduct a detailed analysis of these models across a wide range of tasks, including visual grounding, image generation and editing, visual understanding, and domain-specific applications. Additionally, we compile and describe training datasets and evaluation benchmarks, conducting comparisons among existing models in terms of performance and computational requirements. Overall, this survey offers a comprehensive overview of the current state of the art, laying the groundwork for future MLLMs.

In order to oversee advanced AI systems, it is important to understand their reasons for generating a given output. When prompted, large language models (LLMs) can provide natural language explanations or reasoning traces that sound plausible and receive high ratings from human annotators. However, it is unclear to what extent these explanations are truly capturing the factors responsible for the model’s predictions: the most “human-like” explanation may be different from the one that is most faithful to the model’s true decision making process. In this work, we introduce the correlational counterfactual test (CCT), a faithfulness metric based on counterfactual input edits that takes into account not just the binary label change, but the total shift in the model’s predicted label distribution. We evaluate the faithfulness of free-text explanations generated by few-shot-prompted LLMs from the Llama-2 family on three NLP tasks. We find that these explanations are indeed more likely to mention factors when they are impactful to the model’s prediction, with the degree of association increasing with model size but varying significantly by task.

Multi-document summarization (MDS) is a challenging task, often decomposed to subtasks of salience and redundancy detection, followed by text generation.In this context, alignment of corresponding sentences between a reference summary and its source documents has been leveraged to generate training data for some of the component tasks. Yet, this enabling alignment step has usually been applied heuristically on the sentence level on a limited number of subtasks.In this paper, we propose extending the summary-source alignment framework by (1) applying it at the more fine-grained proposition span level, (2) annotating alignment manually in a multi-document setup, and (3) revealing the great potential of summary-source alignments to yield several datasets for at least six different tasks. Specifically, for each of the tasks, we release a manually annotated test set that was derived automatically from the alignment annotation. We also release development and train sets in the same way, but from automatically derived alignments.Using the datasets, each task is demonstrated with baseline models and corresponding evaluation metrics to spur future research on this broad challenge.

Social networks have become ideal vehicles for news dissemination because posted content is easily able to reach users beyond a news outlet’s direct audience. Understanding how information is transmitted among communities of users is a critical step towards understanding the impact social networks have on real-world events. Two significant barriers in this vein of work are identifying user clusters and meaningfully characterizing these communities. Thus, we propose the PGNSC benchmark, which builds information pathways based on the audiences of influential news sources and uses their content to characterize the communities. We present methods of aggregating these news-source-centric communities and for constructing the community feature representations that are used sequentially to construct information pathway prediction pipelines. Lastly, we perform extensive experiments to demonstrate the performance of baseline pipeline constructions and to highlight the possibilities for future work.

Benchmark plays a pivotal role in assessing the advancements of large language models (LLMs). While numerous benchmarks have been proposed to evaluate LLMs’ capabilities, there is a notable absence of a dedicated benchmark for assessing their musical abilities. To address this gap, we present ZIQI-Eval, a comprehensive and large-scale music benchmark specifically designed to evaluate the music-related capabilities of LLMs.ZIQI-Eval encompasses a wide range of questions, covering 10 major categories and 56 subcategories, resulting in over 14,000 meticulously curated data entries. By leveraging ZIQI-Eval, we conduct a comprehensive evaluation over 16 LLMs to evaluate and analyze LLMs’ performance in the domain of music.Results indicate that all LLMs perform poorly on the ZIQI-Eval benchmark, suggesting significant room for improvement in their musical capabilities.With ZIQI-Eval, we aim to provide a standardized and robust evaluation framework that facilitates a comprehensive assessment of LLMs’ music-related abilities. The dataset is available at GitHub and HuggingFace.

Research in the field of speech emotion recognition (SER) relies on the availability of comprehensive datasets to make it possible to design accurate emotion detection models. This study introduces the Multimodal Emotion Recognition and Sentiment Analysis (MERSA) dataset, which includes both natural and scripted speech recordings, transcribed text, physiological data, and self-reported emotional surveys from 150 participants collected over a two-week period. This work also presents a novel emotion recognition approach that uses a transformer-based model, integrating pre-trained wav2vec 2.0 and BERT for feature extractions and additional LSTM layers to learn hidden representations from fused representations from speech and text. Our model predicts emotions on dimensions of arousal, valence, and dominance. We trained and evaluated the model on the MSP-PODCAST dataset and achieved competitive results from the best-performing model regarding the concordance correlation coefficient (CCC). Further, this paper demonstrates the effectiveness of this model through cross-domain evaluations on both IEMOCAP and MERSA datasets.

As the influence of large language models (LLMs) spans across global communities, their safety challenges in multilingual settings become paramount for alignment research. This paper examines the variations in safety challenges faced by LLMs across different languages and discusses approaches to alleviating such concerns. By comparing how state-of-the-art LLMs respond to the same set of malicious prompts written in higher- vs. lower-resource languages,we observe that (1) LLMs tend to generate unsafe responses much more often when a malicious prompt is written in a lower-resource language, and (2) LLMs tend to generate more irrelevant responses to malicious prompts in lower-resource languages. To understand where the discrepancy can be attributed, we study the effect of instruction tuning with reinforcement learning from human feedback (RLHF) or supervised finetuning (SFT) on the HH-RLHF dataset. Surprisingly, while training with high-resource languages improves model alignment, training in lower-resource languages yields minimal improvement. This suggests that the bottleneck of cross-lingual alignment is rooted in the pretraining stage. Our findings highlight the challenges in cross-lingual LLM safety, and we hope they inform future research in this direction.

Large language models often necessitate grounding on external knowledge to generate faithful and reliable answers. Yet even with the correct groundings in the reference, they can ignore them and rely on wrong groundings or their inherent biases to hallucinate when users, being largely unaware of the specifics of the stored information, pose questions that might not directly correlate with the retrieved groundings. In this work, we formulate this knowledge alignment problem and introduce MixAlign, a framework that interacts with both the human user and the knowledge base to obtain and integrate clarifications on how the user question relates to the stored information. MixAlign employs a language model to achieve automatic knowledge alignment and, if necessary, further enhances this alignment through human user clarifications. Experimental results highlight the crucial role of knowledge alignment in boosting model performance and mitigating hallucination, with improvements noted up to 22.2% and 27.1% respectively. We also demonstrate the effectiveness of MixAlign in improving knowledge alignment by producing high-quality, user-centered clarifications.

Chain of Thought (CoT) is significant in improving the reasoning abilities of large language models (LLMs). However, the correlation between the effectiveness of CoT and the length of reasoning steps in prompts remains largely unknown. To shed light on this, we have conducted several empirical experiments to explore the relations. Specifically, we design experiments that expand and compress the rationale reasoning steps within CoT demonstrations, while keeping all other factors constant. We have the following key findings. First, the results indicate that lengthening the reasoning steps in prompts, even without adding new information into the prompt, considerably enhances LLMs’ reasoning abilities across multiple datasets. Alternatively, shortening the reasoning steps, even while preserving the key information, significantly diminishes the reasoning abilities of models. This finding highlights the importance of the number of steps in CoT prompts and provides practical guidance to make better use of LLMs’ potential in complex problem-solving scenarios. Second, we also investigated the relationship between the performance of CoT and the rationales used in demonstrations. Surprisingly, the result shows that even incorrect rationales can yield favorable outcomes if they maintain the requisite length of inference. Third, we observed that the advantages of increasing reasoning steps are task-dependent: simpler tasks require fewer steps, whereas complex tasks gain significantly from longer inference sequences.

In-context learning is a popular inference strategy where large language models solve a task using only a few labeled demonstrations without needing any parameter updates. Although there have been extensive studies on English in-context learning, multilingual in-context learning remains under-explored, and we lack an in-depth understanding of the role of demonstrations in this context. To address this gap, we conduct a multidimensional analysis of multilingual in-context learning, experimenting with 5 models from different model families, 9 datasets covering classification and generation tasks, and 56 typologically diverse languages. Our results reveal that the effectiveness of demonstrations varies significantly across models, tasks, and languages. We also find that strong instruction-following models including Llama 2-Chat, GPT-3.5, and GPT-4 are largely insensitive to the quality of demonstrations. Instead, a carefully crafted template often eliminates the benefits of demonstrations for some tasks and languages altogether. These findings show that the importance of demonstrations might be overestimated. Our work highlights the need for granular evaluation across multiple axes towards a better understanding of in-context learning.

We analyze the operation of transformer language adapters, which are small modules trained on top of a frozen language model to adapt its predictions to new target languages. We show that adapted predictions mostly evolve in the source language the model was trained on, while the target language becomes pronounced only in the very last layers of the model. Moreover, the adaptation process is gradual and distributed across layers, where it is possible to skip small groups of adapters without decreasing adaptation performance. Last, we show that adapters operate on top of the model’s frozen representation space while largely preserving its structure, rather than on an isolated subspace. Our findings provide a deeper view into the adaptation process of language models to new languages, showcasing the constraints imposed on it by the underlying model and introduces practical implications to enhance its efficiency.

Prior work has found that pretrained language models (LMs) fine-tuned with different random seeds can achieve similar in-domain performance but generalize differently on tests of syntactic generalization. In this work, we show that, even within a single model, we can find multiple subnetworks that perform similarly in-domain, but generalize vastly differently. To better understand these phenomena, we investigate if they can be understood in terms of “competing subnetworks”: the model initially represents a variety of distinct algorithms, corresponding to different subnetworks, and generalization occurs when it ultimately converges to one. This explanation has been used to account for generalization in simple algorithmic tasks (“grokking”). Instead of finding competing subnetworks, we find that all subnetworks—whether they generalize or not—share a set of attention heads, which we refer to as the _heuristic core_. Further analysis suggests that these attention heads emerge early in training and compute shallow, non-generalizing features. The model learns to generalize by incorporating additional attention heads, which depend on the outputs of the “heuristic” heads to compute higher-level features. Overall, our results offer a more detailed picture of the mechanisms for syntactic generalization in pre-trained LMs.

Retrieval-augmented generation (RAG) is a powerful technique to facilitate language model generation with proprietary and private data, where data privacy is a pivotal concern. Whereas extensive research has demonstrated the privacy risks of large language models (LLMs), the RAG technique could potentially reshape the inherent behaviors of LLM generation, posing new privacy issues that are currently under-explored. To this end, we conduct extensive empirical studies with novel attack methods, which demonstrate the vulnerability of RAG systems on leaking the private retrieval database. Despite the new risks brought by RAG on the retrieval data, we further discover that RAG can be used to mitigate the old risks, i.e., the leakage of the LLMs’ training data. In general, we reveal many new insights in this paper for privacy protection of retrieval-augmented LLMs, which could benefit both LLMs and RAG systems builders.