EMNLP 2024 论文列表

Language models, characterized by their black-box nature, often hallucinate and display sensitivity to input perturbations, causing concerns about trust. To enhance trust, it is imperative to gain a comprehensive understanding of the model’s failure modes and develop effective strategies to improve their performance. In this study, we introduce a methodology designed to examine how input perturbations affect language models across various scales, including pre-trained models and large language models (LLMs). Utilizing fine-tuning, we enhance the model’s robustness to input perturbations. Additionally, we investigate whether exposure to one perturbation enhances or diminishes the model’s performance with respect to other perturbations. To address robustness against multiple perturbations, we present three distinct fine-tuning strategies. Furthermore, we broaden the scope of our methodology to encompass large language models (LLMs) by leveraging a chain of thought (CoT) prompting approach augmented with exemplars. We employ the Tabular-NLI task to showcase how our proposed strategies adeptly train a robust model, enabling it to address diverse perturbations while maintaining accuracy on the original dataset.

Large language models (LLMs) have demonstrated impressive performance and spurred numerous AI applications, in which role-playing agents (RPAs) are particularly popular, especially for fictional characters. The prerequisite for these RPAs lies in the capability of LLMs to understand characters from fictional works. Previous efforts have evaluated this capability via basic classification tasks or characteristic imitation, failing to capture the nuanced character understanding with LLMs. In this paper, we propose evaluating LLMs’ character understanding capability via the character profiling task, i.e., summarizing character profiles from corresponding materials, a widely adopted yet understudied practice for RPA development. Specifically, we construct the CROSS dataset from literature experts and assess the generated profiles by comparing them with ground truth references and evaluating their applicability in downstream tasks. Our experiments, which cover various summarization methods and LLMs, have yielded promising results. These results strongly validate the character understanding capability of LLMs. Resources are available at https://github.com/Joanna0123/character_profiling.

Chat-based large language models have the opportunity to empower individuals lacking high-quality healthcare access to receive personalized information across a variety of topics. However, users may ask underspecified questions that require additional context for a model to correctly answer. We study how large language model biases are exhibited through these contextual questions in the healthcare domain. To accomplish this, we curate a dataset of sexual and reproductive healthcare questions (ContextSRH) that are dependent on age, sex, and location attributes. We compare models’ outputs with and without demographic context to determine answer alignment among our contextual questions. Our experiments reveal biases in each of these attributes, where young adult female users are favored.

We introduce a dataset comprising commercial machine translations, gathered weekly over six years across 12 translation directions. Since human A/B testing is commonly used, we assume commercial systems improve over time, which enables us to evaluate machine translation (MT) metrics based on their preference for more recent translations. Our study not only confirms several prior findings, such as the advantage of neural metrics over non-neural ones, but also explores the debated issue of how MT quality affects metric reliability—an investigation that smaller datasets in previous research could not sufficiently explore. Overall, our research demonstrates the dataset’s value as a testbed for metric evaluation. We release our code.

Multilingual large language models (mLLMs) have significantly advanced machine translation, yet challenges remain for low-resource languages like Amharic. This study evaluates the performance of state-of-the-art mLLMs, specifically NLLB-200 (NLLB3.3, NLLB1.3 Distilled1.3, NLB600) and M2M (M2M1.2B, M2M418), in English-Amharic bidirectional translation using the Lesan AI dataset. We employed both automatic and human evaluation methods to analyze translation errors. Automatic evaluation used BLEU, METEOR, chrF, and TER metrics, while human evaluation assessed translation quality at both word and sentence levels. Sentence-level accuracy was rated by annotators on a scale from 0 to 5, and word-level quality was evaluated using Multidimensional Quality Metrics. Our findings indicate that the NLLB3.3B model consistently outperformed other mLLMs across all evaluation methods. Common error included mistranslation, omission, untranslated segments, and additions, with mistranslation being particularly common. Punctuation and spelling errors were rare in our experiment.

Existing approaches to relation extraction obtain relation embeddings by concatenating embeddings of the head and tail entities. Despite the popularity of this approach, we find that such representations mostly capture the types of the entities involved, leading to false positives and confusion between relations that involve entities of the same type. Another possibility is to use a prompt with a [MASK] token to directly learn relation embeddings, but this approach tends to perform poorly. We show that this underperformance comes from the fact that information about entity types is insufficiently captured by the [MASK] embeddings. We therefore propose a simple model, which combines such [MASK] embeddings with entity embeddings. Despite its simplicity, our model consistently outperforms the state-of-the-art across several benchmarks, even when the entity embeddings are obtained from a pre-trained entity typing model. We also experiment with a self-supervised pre-training strategy which further improves the results.

Links are a fundamental part of information networks, turning isolated pieces of knowledge into a network of information that is much richer than the sum of its parts. However, adding a new link to the network is not trivial: it requires not only the identification of a suitable pair of source and target entities but also the understanding of the content of the source to locate a suitable position for the link in the text. The latter problem has not been addressed effectively, particularly in the absence of text spans in the source that could serve as anchors to insert a link to the target entity. To bridge this gap, we introduce and operationalize the task of entity insertion in information networks. Focusing on the case of Wikipedia, we empirically show that this problem is, both, relevant and challenging for editors. We compile a benchmark dataset in 105 languages and develop a framework for entity insertion called LocEI (Localized Entity Insertion) and its multilingual variant XLocEI. We show that XLocEI outperforms all baseline models (including state-of-the-art prompt-based ranking with LLMs such as GPT-4) and that it can be applied in a zero-shot manner on languages not seen during training with minimal performance drop. These findings are important for applying entity insertion models in practice, e.g., to support editors in adding links across the more than 300 language versions of Wikipedia.

The black-box nature of large language models (LLMs) poses challenges in interpreting results, impacting issues such as data intellectual property protection and hallucination tracing. Training data attribution (TDA) methods are considered effective solutions to address these challenges.Most recent TDA methods rely on influence functions, assuming the model achieves minimized empirical risk. However, achieving this criterion is difficult, and sourcing accuracy can be compromised by fitting errors during model training. In this paper, we introduce a novel TDA method called Debias and Denoise Attribution (DDA), which enhances influence functions by addressing fitting errors. Specifically, the debias strategy seeks to improve the performance of influence functions by eliminating the knowledge bias present in the base model before fine-tuning, while the denoise strategy aims to reduce discrepancies in influence scores arising from varying degrees of fitting during the training process through smoothing techniques.Experimental results demonstrate that our method significantly outperforms existing approaches, achieving an averaged AUC of 91.64%. Moreover, DDA exhibits strong generality and scalability across various sources and different-scale models like LLaMA2, QWEN2, and Mistral.

Tool learning aims to enhance and expand large language models’ (LLMs) capabilities with external tools, which has gained significant attention recently. Current methods have shown that LLMs can effectively handle a certain amount of tools through in-context learning or fine-tuning. However, in real-world scenarios, the number of tools is typically extensive and irregularly updated, emphasizing the necessity for a dedicated tool retrieval component. Tool retrieval is nontrivial due to the following challenges: 1) complex user instructions and tool descriptions; 2) misalignment between tool retrieval and tool usage models. To address the above issues, we propose to enhance tool retrieval with iterative feedback from the large language model. Specifically, we prompt the tool usage model, i.e., the LLM, to provide feedback for the tool retriever model in multi-round, which could progressively improve the tool retriever’s understanding of instructions and tools and reduce the gap between the two standalone components. We build a unified and comprehensive benchmark to evaluate tool retrieval models. The extensive experiments indicate that our proposed approach achieves advanced performance in both in-domain evaluation and out-of-domain evaluation.

Time-Sensitive Question Answering (TSQA) demands the effective utilization of specific temporal contexts, encompassing multiple time-evolving facts, to address time-sensitive questions. This necessitates not only the parsing of temporal information within questions but also the identification and understanding of time-evolving facts to generate accurate answers. However, current large language models still have limited sensitivity to temporal information and their inadequate temporal reasoning capabilities. In this paper, we propose a novel framework that enhances temporal awareness and reasoning through Temporal Information-Aware Embedding and Granular Contrastive Reinforcement Learning. Experimental results on four TSQA datasets demonstrate that our framework significantly outperforms existing LLMs in TSQA tasks, marking a step forward in bridging the performance gap between machine and human temporal understanding and reasoning.

Video Large Language Models (Video LLMs) have achieved impressive performance on video-and-language tasks, such as video question answering. However, most existing Video LLMs neglect temporal information in video data, leading to struggles with temporal-aware video understanding. To address this gap, we propose a Time Gating Video LLM (TG-Vid) designed to enhance temporal modeling through a novel Time Gating module (TG). The TG module employs a time gating mechanism on its sub-modules, comprising gating spatial attention, gating temporal attention, and gating MLP. This architecture enables our model to achieve a robust understanding of temporal information within videos. Extensive evaluation of temporal-sensitive video benchmarks (i.e., MVBench, TempCompass, and NExT-QA) demonstrates that our TG-Vid model significantly outperforms the existing Video LLMs. Further, comprehensive ablation studies validate that the performance gains are attributed to the designs of our TG module. Our code is available at https://github.com/LaVi-Lab/TG-Vid.

Recent language models enable new opportunities for structured reasoning with text, such as the construction of intuitive, proof-like textual entailment trees without relying on brittle formal logic. However, progress in this direction has been hampered by a long-standing lack of a clear protocol for determining what _valid decompositional entailment_ is. This absence causes noisy datasets and limited performance gains by modern neuro-symbolic entailment engines. To address these problems, we formulate a consistent and theoretically grounded approach to annotating decompositional entailment and evaluate its impact on LLM-based textual inference. We find that our new dataset, RDTE (Recognizing Decompositional Textual Entailment), has a substantially higher internal consistency than prior decompositional entailment datasets, suggesting that RDTE is a significant step forward in the long-standing problem of forming a clear protocol for discerning entailment. We also find that training an RDTE-oriented entailment classifier via knowledge distillation and employing it in an entailment tree reasoning engine significantly improves both accuracy and proof quality, illustrating the practical benefit of this advance for textual inference.

Topic modeling is a powerful technique for uncovering hidden themes within a collection of documents. However, the effectiveness of traditional topic models often relies on sufficient word co-occurrence, which is lacking in short texts. Therefore, existing approaches, whether probabilistic or neural, frequently struggle to extract meaningful patterns from such data, resulting in incoherent topics. To address this challenge, we propose a novel approach that leverages large language models (LLMs) to extend short texts into more detailed sequences before applying topic modeling. To further improve the efficiency and solve the problem of semantic inconsistency from LLM-generated texts, we propose to use prefix tuning to train a smaller language model coupled with a variational autoencoder for short-text topic modeling. Our method significantly improves short-text topic modeling performance, as demonstrated by extensive experiments on real-world datasets with extreme data sparsity, outperforming current state-of-the-art topic models.

Reinforcement learning (RL) can align language models with non-differentiable reward signals, such as human preferences. However, a major challenge arises from the sparsity of these reward signals - typically, there is only a single reward for an entire output. This sparsity of rewards can lead to inefficient and unstable learning. To address this challenge, our paper introduces an novel framework that utilizes the critique capability of Large Language Models (LLMs) to produce intermediate-step rewards during RL training. Our method involves coupling a policy model with a critic language model, which is responsible for providing comprehensive feedback of each part of the output. This feedback is then translated into token or span-level rewards that can be used to guide the RL training process. We investigate this approach under two different settings: one where the policy model is smaller and is paired with a more powerful critic model, and another where a single language model fulfills both roles. We assess our approach on three text generation tasks: sentiment control, language model detoxification, and summarization. Experimental results show that incorporating artificial intrinsic rewards significantly improve both sample efficiency and the overall performance of the policy model, supported by both automatic and human evaluation.

Accurately attributing answer text to its source document is crucial for developing a reliable question-answering system. However, attribution for long documents remains largely unexplored. Post-hoc attribution systems are designed to map answer text back to the source document, yet the granularity of this mapping has not been addressed. Furthermore, a critical question arises: What exactly should be attributed? This involves identifying the specific information units within an answer that require grounding. In this paper, we propose and investigate a novel approach to the factual decomposition of generated answers for attribution, employing template-based in-context learning. To accomplish this, we utilize the question and integrate negative sampling during few-shot in-context learning for decomposition. This approach enhances the semantic understanding of both abstractive and extractive answers. We examine the impact of answer decomposition by providing a thorough examination of various attribution approaches, ranging from retrieval-based techniques to LLM-based attributors.

The creation of artificial polyglot voices remains a challenging task, despite considerable progress in recent years. This paper investigates self-supervised learning for voice conversion to create native-sounding polyglot voices. We introduce a novel cross-lingual any-to-one voice conversion system that is able to preserve the source accent without the need for multilingual data from the target speaker. In addition, we show a novel cross-lingual fine-tuning strategy that further improves the accent and reduces the training data requirements. Objective and subjective evaluations with English, Spanish, French and Mandarin Chinese confirm that our approach improves on state-of-the-art methods, enhancing the speech intelligibility and overall quality of the converted speech, especially in cross-lingual scenarios. Audio samples are available at: https://giuseppe-ruggiero.github.io/a2o-vc-demo/

Recent advancements in noisy multi-label text classification have primarily relied on the class-conditional noise (CCN) assumption, which treats each label independently undergoing label flipping to generate noisy labels. However, in real-world scenarios, noisy labels often exhibit dependencies with true labels. In this study, we validate through hypothesis testing that real-world datasets are unlikely to adhere to the CCN assumption, indicating that label noise is dependent on the labels. To address this, we introduce a label-specific denoising framework designed to counteract label-dependent noise. The framework initially presents a holistic selection metric that evaluates noisy labels by concurrently considering loss information, ranking information, and feature centroid. Subsequently, it identifies and corrects noisy labels individually for each label category in a fine-grained manner. Extensive experiments on benchmark datasets demonstrate the effectiveness of our method under both synthetic and real-world noise conditions, significantly improving performance over existing state-of-the-art models.

Legal case retrieval (LCR) aims to provide similar cases as references for a given fact description. This task is crucial for promoting consistent judgments in similar cases, effectively enhancing judicial fairness and improving work efficiency for judges. However, existing works face two main challenges for real-world applications: existing works mainly focus on case-to-case retrieval using lengthy queries, which does not match real-world scenarios; and the limited data scale, with current datasets containing only hundreds of queries, is insufficient to satisfy the training requirements of existing data-hungry neural models. To address these issues, we introduce an automated method to construct synthetic query-candidate pairs and build the largest LCR dataset to date, LEAD, which is hundreds of times larger than existing datasets. This data construction method can provide ample training signals for LCR models. Experimental results demonstrate that model training with our constructed data can achieve state-of-the-art results on two widely-used LCR benchmarks. Besides, the construction method can also be applied to civil cases and achieve promising results. The data and codes can be found in https://github.com/thunlp/LEAD.

Compared to identifying binary versions of the same function under different compilation options, existing Learning-Based Binary Code Similarity Detection (LB-BCSD) methods exhibit lower accuracy in recognizing functions with the same functionality but different implementations. To address this issue, we introduces an adversarial attack method called FuncFooler, which focuses on perturbing critical code to generate multiple variants of the same function. These variants are then used to retrain the model to enhance its robustness. Current adversarial attacks against LB-BCSD mainly draw inspiration from the FGSM (Fast Gradient Sign Method) method in the image domain, which involves generating adversarial bytes and appending them to the end of the executable file. However, this approach has a significant drawback: the appended bytes do not affect the actual code of the executable file, thus failing to create diverse code variants. To overcome this limitation, we proposes a gradient-guided adversarial attack method based on critical code—FuncFooler. This method designs a series of strategies to perturb the code while preserving the program’s semantics. Specifically, we first utilizes gradient information to locate critical nodes in the control flow graph. Then, fine-grained perturbations are applied to these nodes, including control flow, data flow, and internal node perturbations, to obtain adversarial samples. The experimental results show that the application of the FuncFooler method can increase the accuracy of the latest LB-BCSD model by 5%-7%.