联系我们

论文查询结果	返回搜索

论文编号：	15512
作者编号：	1120211212
上传时间：	2025/6/14 0:00:25
中文题目：	基于用户意图识别与多轮交互机制的法律自动问答研究
英文题目：	Research on Legal Question Answering Based on User Intent Recognition and Multi-turn Interaction Mechanisms
指导老师：	王芳教授
中文关键字：	信息服务；大语言模型；大小模型协同；法律自动问答
英文关键字：	Information Services; Large Language Models; Large-Small Model Collaboration; Legal Question Answering
中文摘要：	在数字化浪潮以及“依法治国”战略的双重推动下，公众对法律信息服务的需求呈现爆发式增长，传统的法律服务模式在资源分配、服务效率等方面的局限性日益凸显。大语言模型的快速迭代与发展为法律信息服务带来了新的机遇，其能够通过对话的方式为用户提问提供即时的回复。尽管大语言模型在各种通用领域的应用中都展现出了强大的能力，但由于法律领域的高度专业性与复杂性，大语言模型在法律领域的应用研究仍存在较大的挑战。 本文聚焦于用户意图识别、多轮交互信息补全、法律信息抽取、答案生成四个法律自动问答的关键环节，以大语言模型与领域小模型协同的方式对模型进行构建与优化，提升法律自动问答系统回复的专业性、准确性与针对性。具体而言，本文主要围绕以下四个核心研究问题展开：第一，法律咨询过程中，用户往往采用口语化、非专业化的表达方式，且存在表述模糊、歧义等问题，给意图识别任务带来显著挑战。如何在此情况下准确识别用户意图，是本研究的首要问题。第二，传统自动问答系统在用户提问中存在关键信息缺失情况时往往表现欠佳。为此，本文探讨如何在用户提问存在信息缺失情况下，设计有效的机制以补全关键信息。第三，用户提问中法律实体的识别影响着自动问答系统的准确性，为此，本文对如何从用户提问中准确抽取其中包含的法律关键实体进行了探索。第四，检索增强生成方法是一种缓解大语言模型幻觉的有效方式，但现有方法主要关注语义层面的相似性，忽视了法律领域中逻辑结构与推理路径的重要性。如何设计模型使其生成专业性强、准确性高的法律回复，降低大语言模型因幻觉等问题给用户带来的风险也是本文的重要研究问题。 针对上述研究问题，本文的研究工作主要包括以下方面： 第一，提出了一种多模型协同的意图识别算法。该算法利用大语言模型强大的语义理解能力对领域小模型的意图识别能力进行补充，并为了缓解标签噪声问题，提出了一种多粒度融合的标签优化模块，将Label Smoothing方法和Bootstrapping方法进行了融合，利用元学习的思想对融合系数进行求解。第二，提出了一种基于强化学习的多轮交互机制。该机制首先利用微调大模型的方式对用户问题的完整性进行判别，对于缺失信息的问题，利用强化学习方法预测问题中缺少的关键要素，并根据关键要素生成针对性的反问问题及对应的选项，引导用户对相关内容进行澄清和补充，直至用户所提供的信息完整时结束交互。第三，构建了基于词嵌入向量增强与样本学习难度的法律信息抽取算法。该算法利用大语言模型对用户提问中的法律命名实体进行预识别，将大语言模型预识别得到的法律实体的词嵌入向量与领域小模型的词嵌入向量结合，对词嵌入表示进行增强。此外，利用大语言模型在训练初期评估样本的学习难度，利用自步学习策略指导训练中后期样本的选择，进一步提升模型的性能。第四，提出了一种思维链与RAG协同的答案生成算法。该算法利用可学习的思维链策略构建法律咨询文本的逻辑结构，利用监督式DSSM检索模型从数据库中检索相似案例，利用法律上下文学习生成专业、准确的法律建议。第五，基于上述提出的算法，以中国和美国的法律体系为例，构建了法律自动问答系统。通过在中文和英文数据集上的实验验证了上述算法的有效性。本研究共八章内容，包括图22幅，表45个，参考文献346篇。
英文摘要：	Under the dual impetus of the digital wave and the "rule of law" strategy, public demand for legal information services has shown explosive growth. The limitations of traditional legal service models in terms of resource allocation and service efficiency have become increasingly evident. The rapid iteration and development of large language models (LLMs) present new opportunities for legal information services, enabling immediate responses to user inquiries through conversational interfaces. Despite the powerful capabilities demonstrated by LLMs in applications across various general domains, due to the high level of specialization and complexity in the legal field, significant challenges remain in the application research of LLMs in the legal domain. The paper focuses on four key steps in legal question answering: user intent recognition, multi-turn interactive information completion, legal information extraction, and answer generation. Models are constructed and optimized through a collaborative approach between LLMs and domain-specific small models, with the aim of enhancing the professionalism, accuracy, and relevance of legal QA system responses. Specifically, the paper addresses the following four core problems: (1) In legal consultations, users often express themselves in colloquial and non-specialized language, with vagueness and ambiguity posing significant challenges to intent recognition. Accurately identifying user intent under such circumstances is the first key issue addressed in this study. (2) Traditional QA systems often underperform when key information is missing in user queries. To address the issue, this study explores mechanisms to effectively complete missing key information when user queries are incomplete. (3) The recognition of legal entities within user queries directly affects the accuracy of QA systems. This study investigates how to accurately extract key legal entities from user queries. (4) Retrieval-augmented generation (RAG) is an effective method to mitigate hallucination in LLMs. However, existing RAG methods mainly focus on semantic similarity, overlooking the importance of logical structure and reasoning paths in legal contexts. Designing models that generate highly professional and accurate legal responses, while minimizing the risks posed by hallucination, is key challenge tackled in this work. To address the above research questions, the main research contents of the paper are as follows. (1) A multi-model collaborative intent recognition algorithm is proposed. This algorithm leverages the strong semantic understanding capabilities of LLMs to complement the intent recognition performance of domain-specific small models. To mitigate label noise, a multi-granularity label optimization module is introduced, which integrates Label Smoothing and Bootstrapping methods. The fusion coefficient is determined through meta-learning. (2) A reinforcement learning-based multi-turn interaction mechanism is proposed. The method uses a fine-tuned LLM to assess the completeness of user queries. For queries lacking key information, reinforcement learning is used to predict the missing elements. Targeted clarifying questions and options are then generated based on the missing elements, guiding users to clarify and supplement their input until the information is deemed complete. (3) A legal information extraction algorithm is proposed based on word embedding enhancement and sample learning difficulty. The algorithm leverages the LLM to pre-identify legal named entities in user queries, and then combines the word embedding vectors of the identified legal entities from the LLM with the word embedding vectors from the domain-specific small model to enhance representations. Additionally, the LLM is used to assess the learning difficulty of samples during the early stages of training, and the self-paced learning strategy is employed to guide sample selection in the later stages, further enhancing the model's performance. (4) A Chain-of-Thought (CoT) and RAG collaborative answer generation algorithm is proposed. This algorithm constructs the logical structure of legal queries using a learnable CoT strategy, retrieves relevant cases from a legal database using a supervised DSSM retrieval model, and generates professional and accurate legal responses by legal in-context learning. (5) Based on the algorithms proposed above, legal QA systems are constructed using the Chinese and U.S. legal systems as case studies. Experiments conducted on both Chinese and English datasets validate the effectiveness of the proposed algorithms. The paper comprises eight chapters and includes 22 figures, 45 tables, and 346 references.
查看全文：	预览  下载(下载需要进行登录)