EMNLP 2024 论文列表

Reinforcement learning from human feedback (RLHF) is a popular strategy for aligning large language models (LLMs) with desired behaviors. Reward modeling is a crucial step in RLHF. However, collecting paired preference data for training reward models is often costly and time-consuming, especially for domain-specific preferences requiring expert annotation. To address this challenge, we propose the **Do**main knowled**ge** merged **R**eward **M**odel (**DogeRM**), a novel framework that integrates domain-specific knowledge into a general reward model by model merging. The experiments demonstrate that DogeRM enhances performance across different benchmarks and provide a detailed analysis showcasing the effects of model merging, showing the great potential of facilitating model alignment.

Large vision-language models (LVLMs) have recently dramatically pushed the state of the art in image captioning and many image understanding tasks (e.g., visual question answering). LVLMs, however, often hallucinate and produce captions that mention concepts that cannot be found in the image. These hallucinations erode the trustworthiness of LVLMs and are arguably among the main obstacles to their ubiquitous adoption. Recent work suggests that addition of grounding objectives—those that explicitly align image regions or objects to text spans—reduces the amount of LVLM hallucination. Although intuitive, this claim is not empirically justified as the reduction effects have been established, we argue, with flawed evaluation protocols that (i) rely on data (i.e., MSCOCO) that has been extensively used in LVLM training and (ii) measure hallucination via question answering rather than open-ended caption generation.In this work, in contrast, we offer the first systematic analysis of the effect of fine-grained object grounding on LVLM hallucination under an evaluation protocol that more realistically captures LVLM hallucination in open generation. Our extensive experiments over three backbone LLMs reveal that grounding objectives have little to no effect on object hallucination in open caption generation.

Large language models (LLMs) have demonstrated remarkable capabilities in comprehensively handling various types of natural language processing (NLP) tasks. However, there are significant differences in the knowledge and abilities required for different tasks. Therefore, it is important to understand whether the same LLM processes different tasks in the same way. Are there specific neurons in a LLM for different tasks? Inspired by neuroscience, this paper pioneers the exploration of whether distinct neurons are activated when a LLM handles different tasks. Compared with current research exploring the neurons of language and knowledge, task-specific neurons present a greater challenge due to their abstractness, diversity, and complexity. To address these challenges, this paper proposes a method for task-specific neuron localization based on Causal Gradient Variation with Special Tokens (CGVST). CGVST identifies task-specific neurons by concentrating on the most significant tokens during task processing, thereby eliminating redundant tokens and minimizing interference from non-essential neurons. Compared to traditional neuron localization methods, our approach can more effectively identify task-specific neurons. We conduct experiments across eight different public tasks. Experiments involving the inhibition and amplification of identified neurons demonstrate that our method can accurately locate task-specific neurons.

When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training. It is often conjectured that this can teach the model the behavior of hallucinating factually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge. In this work, we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge. To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge. We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model’s knowledge. However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model’s tendency to hallucinate. Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently.

Neural Machine Translation models tend to perpetuate gender bias present in their training data distribution. Context-aware models have been previously suggested as a means to mitigate this type of bias. In this work, we examine this claim by analysing in detail the translation of stereotypical professions in English to German, and translation with non-informative context in Basque to Spanish. Our results show that, although context-aware models can significantly enhance translation accuracy for feminine terms, they can still maintain or even amplify gender bias. These results highlight the need for more fine-grained approaches to bias mitigation in Neural Machine Translation.

As an essential task in information extraction (IE), Event-Event Causal Relation Extraction (ECRE) aims to identify and classify the causal relationships between event mentions in natural language texts. However, existing research on ECRE has highlighted two critical challenges, including the lack of document-level modeling and causal hallucinations. In this paper, we propose a Knowledge-guided binary Question Answering (KnowQA) method with event structures for ECRE, consisting of two stages: Event Structure Construction and Binary Question Answering. We conduct extensive experiments under both zero-shot and fine-tuning settings with large language models (LLMs) on the MECI and MAVEN-ERE datasets. Experimental results demonstrate the usefulness of event structures on document-level ECRE and the effectiveness of KnowQA by achieving state-of-the-art on the MECI dataset. We observe not only the effectiveness but also the high generalizability and low inconsistency of our method, particularly when with complete event structures after fine-tuning the models.

Document hashing plays a crucial role in large-scale information retrieval. However, existing unsupervised document hashing methods merely consider flat semantics of documents, resulting in the inability of preserving hierarchical semantics in hash codes. In this paper, we propose a hierarchical generative model that can model and leverage the hierarchical structure of semantics. Specifically, we introduce hierarchical prototypes into the model to construct a hierarchical prior distribution, which is integrated into the variational auto-encoder (VAE) framework, enabling the model to produce hash codes preserving rough hierarchical semantics. To further promote the preservation of hierarchical structure, we force the hash code to preserve as much semantic information as possible via contrastive learning, which exploits the hierarchical pseudo labels produced during VAE training. Extensive experiments on three benchmarks outperform all baseline methods, demonstrating the superiority of our proposed model on both hierarchical datasets and flat datasets.

Visual document understanding (VDU) is a challenging task that involves understanding documents across various modalities (text and image) and layouts (forms, tables, etc.). This study aims to enhance generalizability of small VDU models by distilling knowledge from LLMs. We identify that directly prompting LLMs often fails to generate informative and useful data. In response, we present a new framework (called DocKD) that enriches the data generation process by integrating external document knowledge. Specifically, we provide an LLM with various document elements like key-value pairs, layouts, and descriptions, to elicit open-ended answers. Our experiments show that DocKD produces high-quality document annotations and surpasses the direct knowledge distillation approach that does not leverage external document knowledge. Moreover, student VDU models trained with solely DocKD-generated data is not only comparable to those trained with human-annotated data on in-domain tasks but also significantly excel them on out-of-domain tasks.

Parsing documents from pixels, such as pictures and scanned PDFs, into hierarchical structures is extensively demanded in the daily routines of data storage, retrieval and understanding. However, previously the research on this topic has been largely hindered since most existing datasets are small-scale, or contain documents of only a single type, which are characterized by a lack of document diversity. Moreover, there is a significant discrepancy in the annotation standards across datasets. In this paper, we introduce a large and diverse document hierarchy parsing (DHP) dataset to compensate for the data scarcity and inconsistency problem. We aim to set a new standard as a more practical, long-standing benchmark. Meanwhile, we present a new DHP framework designed to grasp both fine-grained text content and coarse-grained pattern at layout element level, enhancing the capacity of pre-trained text-layout models in handling the multi-page and multi-level challenges in DHP. Through exhaustive experiments, we validate the effectiveness of our proposed dataset and method.

Document structure editing involves manipulating localized textual, visual, and layout components in document images based on the user’s requests. Past works have shown that multimodal grounding of user requests in the document image and identifying the accurate structural components and their associated attributes remain key challenges for this task. To address these, we introduce the DocEditAgent, a novel framework that performs end-to-end document editing by leveraging Large Multimodal Models (LMMs). It consists of three novel components – (1) Doc2Command to simultaneously localize edit regions of interest (RoI) and disambiguate user edit requests into edit commands. (2) LLM-based Command Reformulation prompting to tailor edit commands originally intended for specialized software into edit instructions suitable for generalist LMMs. (3) Moreover, DocEditAgent processes these outputs via Large Multimodal Models like GPT-4V and Gemini, to parse the document layout, execute edits on grounded Region of Interest (RoI), and generate the edited document image. Extensive experiments on the DocEdit dataset show that DocEditAgent significantly outperforms strong baselines on edit command generation (2-33%), RoI bounding box detection (12-31%), and overall document editing (1-12%) tasks.

Event extraction aims to identify events and then extract the arguments involved in those events. In recent years, there has been a gradual shift from sentence-level event extraction to document-level event extraction research. Despite the significant success achieved in English domain event extraction research, event extraction in Chinese still remains largely unexplored. However, a major obstacle to promoting Chinese document-level event extraction is the lack of fine-grained, wide domain coverage datasets for model training and evaluation. In this paper, we propose DocEE-zh, a new Chinese document-level event extraction dataset comprising over 36,000 events and more than 210,000 arguments. DocEE-zh is an extension of the DocEE dataset, utilizing the same event schema, and all data has been meticulously annotated by human experts. We highlight two features: focus on high-interest event types and fine-grained argument types. Experimental results indicate that state-of-the-art models still fail to achieve satisfactory performance, with an F1 score of 45.88% on the event argument extraction task, revealing that Chinese document-level event extraction (DocEE) remains an unresolved challenge. DocEE-zh is now available at https://github.com/tongmeihan1995/DocEE.git.

Recent developments show that Large Language Models (LLMs) produce state-of-the-art performance on natural language (NL) to code generation for resource-rich general-purpose languages like C++, Java, and Python. However, their practical usage for structured domain-specific languages (DSLs) such as YAML, JSON is limited due to domain-specific schema, grammar, and customizations generally unseen by LLMs during pre-training. Efforts have been made to mitigate this challenge via in-context learning through relevant examples or by fine-tuning. However, it suffers from problems, such as limited DSL samples and prompt sensitivity but enterprises maintain good documentation of the DSLs. Therefore, we propose DocCGen, a framework that can leverage such rich knowledge by breaking the NL-to-Code generation task for structured code languages into a two-step process. First, it detects the correct libraries using the library documentation that best matches the NL query. Then, it utilizes schema rules extracted from the documentation of these libraries to constrain the decoding. We evaluate our framework for two complex structured languages, Ansible YAML and Bash command, consisting of two settings: Out-of-domain (OOD) and In domain (ID). Our extensive experiments show that DocCGen consistently improves different sized language models across all six evaluation metrics, reducing syntactic and semantic errors in structured code.

Recent advancements of large language models (LLMs)have led to claims of AI surpassing humansin natural language processing NLP tasks such as textual understanding and reasoning.%This work investigates these assertions by introducingCAIMIRA, a novel framework rooted in item response theory IRTthat enables quantitative assessment and comparison of problem-solving abilities inquestion-answering QA agents.%Through analysis of over 300,000 responses from ~ 70 AI systemsand 155 humans across thousands of quiz questions, CAIMIRA uncovers distinctproficiency patterns in knowledge domains and reasoning skills. %Humans outperform AI systems in knowledge-grounded abductive and conceptual reasoning,while state-of-the-art LLMs like GPT-4 Turbo and Llama-3-70B demonstrate superior performance ontargeted information retrieval and fact-based reasoning, particularly when information gapsare well-defined and addressable through pattern matching or data retrieval.%These findings identify key areas for future QA tasks and model development,highlighting the critical need for questions that not only challengehigher-order reasoning and scientific thinking, but also demand nuanced linguisticand cross-contextual application.

Language models (LMs) are known to suffer from hallucinations and misinformation. Retrieval augmented generation (RAG) that retrieves verifiable information from an external knowledge corpus to complement the parametric knowledge in LMs provides a tangible solution to these problems. However, the generation quality of RAG is highly dependent on the relevance between a user’s query and the retrieved documents. Inaccurate responses may be generated when the query is outside of the scope of knowledge represented in the external knowledge corpus or if the information in the corpus is out-of-date. In this work, we establish a statistical framework that assesses how well a query can be answered by an RAG system by capturing the relevance of knowledge. We introduce an online testing procedure that employs goodness-of-fit (GoF) tests to inspect the relevance of each user query to detect out-of-knowledge queries with low knowledge relevance. Additionally, we develop an offline testing framework that examines a collection of user queries, aiming to detect significant shifts in the query distribution which indicates the knowledge corpus is no longer sufficiently capable of supporting the interests of the users. We demonstrate the capabilities of these strategies through a systematic evaluation on eight question-answering (QA) datasets, the results of which indicate that the new testing framework is an efficient solution to enhance the reliability of existing RAG systems.

This paper investigates the potential benefits of language-specific fact-checking models, focusing on the case of Chinese using CHEF dataset. To better reflect real-world fact-checking, we first develop a novel Chinese document-level evidence retriever, achieving state-of-the-art performance. We then demonstrate the limitations of translation-based methods and multilingual language models, highlighting the need for language-specific systems. To better analyze token-level biases in different systems, we construct an adversarial dataset based on the CHEF dataset, where each instance has a large word overlap with the original one but holds the opposite veracity label. Experimental results on the CHEF dataset and our adversarial dataset show that our proposed method outperforms translation-based methods and multilingual language models and is more robust toward biases, emphasizing the importance of language-specific fact-checking systems.

Large language models are able to generate code for visualisations in response to simple user requests.This is a useful application and an appealing one for NLP research because plots of data provide grounding for language.However, there are relatively few benchmarks, and those that exist may not be representative of what users do in practice.This paper investigates whether benchmarks reflect real-world use through an empirical study comparing benchmark datasets with code from public repositories.Our findings reveal a substantial gap, with evaluations not testing the same distribution of chart types, attributes, and actions as real-world examples.One dataset is representative, but requires extensive modification to become a practical end-to-end benchmark. This shows that new benchmarks are needed to support the development of systems that truly address users’ visualisation needs.These observations will guide future data creation, highlighting which features hold genuine significance for users.

Large Language Models (LLMs) have emerged as highly capable systems and are increasingly being integrated into various uses. Nevertheless, the rapid advancement in their deployment trails a comprehensive understanding of their internal mechanisms, as well as a delineation of their capabilities and limitations. A desired characteristic of an intelligent system is its ability to recognize the scope of its own knowledge. To investigate whether LLMs embody this attribute, we develop a benchmark that challenges these models to enumerate all information they possess on specific topics. This benchmark assesses whether the models recall excessive, insufficient, or the precise amount of required information, thereby indicating their awareness of how much they know about the given topic. Our findings reveal that the emergence of this property varies across different architectures and manifests at diverse rates. However, with sufficient scaling, all tested models are ultimately capable of performing this task. The insights gained from this research advance our understanding of LLMs, shedding light on their operational capabilities and contributing to the ongoing exploration of their intricate dynamics.

Assessing the performance of systems to classify Multi-Party Conversations (MPC) is challenging due to the interconnection between linguistic and structural characteristics of conversations. Conventional evaluation methods often overlook variances in model behavior across different levels of structural complexity on interaction graphs. In this work, we propose a methodological pipeline to investigate model performance across specific structural attributes of conversations. As a proof of concept we focus on Response Selection and Addressee Recognition tasks, to diagnose model weaknesses. To this end, we extract representative diagnostic subdatasets with a fixed number of users and a good structural variety from a large and open corpus of online MPCs. We further frame our work in terms of data minimization, avoiding the use of original usernames to preserve privacy, and propose alternatives to using original text messages. Results show that response selection relies more on the textual content of conversations, while addressee recognition requires capturing their structural dimension. Using an LLM in a zero-shot setting, we further highlight how sensitivity to prompt variations is task-dependent.

Do large multilingual language models learn language universals? We consider a candidate universal much-discussed in the linguistics literature, the Final-over-Final Condition (Sheehan et al., 2017b). This Condition is syntactic in the sense that it can only be stated by reference to abstract sentence properties such as nested phrases and head direction. A study of typologically diverse “mixed head direction” languages confirms that the Condition holds in corpora. But in a targeted syntactic evaluation, Gemini Pro only seems to respect the Condition in German, Russian, Hungarian and Serbian. These relatively high-resource languages contrast with Basque, where Gemini Pro does not seem to have learned the Condition at all. This result suggests that modern language models may need additional sources of bias in order to become truly human-like, within a developmentally-realistic budget of training data.