基于大语言模型的刑事案件智能判决研究  

Research on Intelligent Judgment of Criminal Cases Based on Large Language Models

作　　者：丛颖男 韩林睿 马佳羽 朱金清 CONG Yingnan;HAN Linrui;MA Jiayu;ZHU Jinqing(Business School,China University of Political Science and Law,Beijing 100088,China;Ministry of Education Laboratory of Philosophy and Social Sciences-The Data Law Lab,China University of Political Science and Law,Beijing 100088,China;The Institute for Data Law,China University of Political Science and Law,Beijing 100088,China;School of Law,Tsinghua University,Beijing 100084,China;Beijing Bytedance Network Technology Co.Ltd.,Beijing 100043,China)

机构地区：[1]中国政法大学商学院,北京100088 [2]教育部哲学社会科学实验室——中国政法大学数据法治实验室,北京100088 [3]中国政法大学数据法治研究院,北京100088 [4]清华大学法学院,北京100084 [5]北京字节跳动网络技术有限公司,北京100043

出　　处：《计算机科学》2025年第5期248-259,共12页Computer Science

基　　金：2025年中国政法大学青年教师学术创新团队支持计划(25CXTD04);2022年国家重点研发计划“社会治理与智慧社会科技支撑”重点专项(2022YFC3303000);教育部人文社会科学研究一般项目(22YJC190003)。

摘　　要：刑事案件判决的智能化一直是数字法院建设中的研究热点。传统方法基于自然语言处理技术,由模型依据案件事实直接预测判决结果,但应对复杂刑事案件案情时,模型难以发现法律要件之间的逻辑依赖关系,也难以清晰表达法律推理过程。文中提出一种基于大语言模型的刑事案件智能判决方法,该方法以“标记案件语料-预训练大模型-强化判决逻辑”为思路,首先通过自动化标注与人工校正相结合的方式,标注案情中的主体、客体、主观要件和客观要件等法律要素,构建结构化的推理数据集;其次基于GLM预训练框架,选取ChatGLM3-6b-32k作为基座大语言模型进行增量预训练;最后采用LoRA参数高效微调策略与大模型检索增强技术对模型进行参数调优与法律知识扩展,实现判决逻辑的强化。实验结果表明,与Qwen-7B-Chat和Baichuan2-7B-Chat相比,ChatGLM3-6b-32k模型在指令监督微调后性能更优。引入司法三段论显著增强了判决文本的逻辑性,使其更贴近人类法官的裁判说理。在罪名预测和刑期预测任务中,所提模型准确率相较于MTL-Fusion,Lawformer和BERT模型均有显著提升。此外,与基于欧美法律文本训练的Legal-BERT和CaseLawBERT相比,所提模型更适应中国刑事案件的判决逻辑,在处理长文本任务上展现出更强的能力。该研究不仅探索了大语言模型在刑事案件智能判决中的应用,还为司法领域大模型研究的范式提供了有益参考。The intelligentization of criminal case trials has been a hot research topic in the development of digital courts.In the conventional method based on natural language processing,the model directly predicts the final judgment based on the facts of the case.However,when dealing with complex criminal cases,the model may fail to identify the logical dependencies between legal elements and to clearly present the legal reasoning process.The intelligent criminal case trial method based on large language models proposed in this paper follows the approach of“annotating case corpus-pre-training large language model-reinforcing trial logic”.The first step is to annotate the legal elements of the case such as subjects,objects,subjective elements,and objective elements by combining automated annotat ing with manual correction and create a structured reasoning dataset.The second step is to use ChatGLM3-6b-32k as the foundation al large language model for incremental pre-training based on the GLM pre-training framework.The last step is to fine-tune the parameter and increase legal knowledge using the LoRA parameter-efficient fine-tu ning strategy and large language model retrieval enhancement technology,thereby reinforcing the trial logic.Experimental results indicate that,compared to Qwen-7B-Chat and Baichuan2-7B-Chat,the ChatGLM3-6b-32k model exhibits superior performance after supervised fine-tuning.The introduction of judicial syllogism significantly enhances the logicality of the judgment texts,ma-king them closer to the reasoning of human judges.In the tasks of charge prediction and sentencing prediction,the model created using this method shows a significant improvement in accuracy compared to the MTL-Fusion,Lawformer,and BERT models.In addition,compared to Legal-BERT and CaseLawBERT,which are trained on European and American legal texts,the ChatGLM3-6b-32k model better suits the trial logic of Chinese criminal cases and demonstrates stronger capabilities in handling long texts.This paper not only explores the applic

关 键 词：数字法院 法律判决预测 司法三段论 大语言模型 参数高效微调 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]