交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (3): 83-92.DOI: 10.16097/j.cnki.1009-6744.2026.03.008

• 青年基金项目成果 • 上一篇    下一篇

融合检索增强生成与思维链的交通事故责任认定大模型构建

王正礼* ，唐子墨，郑振杰   

  1. 南京大学，工程管理学院，南京210093
  • 收稿日期:2026-02-28 修回日期:2026-04-23 接受日期:2026-05-06 出版日期:2026-06-25 发布日期:2026-06-22
  • 作者简介:王正礼（1991—），男，河南商丘人，副教授。
  • 基金资助:
    国家自然科学基金青年科学基金及面上项目(72101012,72571126)。

Traffic Accident Liability Determination Using Multimodal Large Language Model Enhanced by Retrieval-Augmented Generation and Chain-of-Thought Reasoning

WANG Zhengli*, TANG Zimo, ZHENG Zhenjie   

  1. School of Management and Engineering, Nanjing University, Nanjing 210093, China
  • Received:2026-02-28 Revised:2026-04-23 Accepted:2026-05-06 Online:2026-06-25 Published:2026-06-22
  • Supported by:
    National Natural Science Foundation of China(72101012,72571126)。

PDF

摘要： 交通事故责任认定是提升道路交通安全治理能力的关键环节。针对我国当前交通事故责任认定存在人工依赖度高、主观性强及处理效率低等问题，本文提出一种融合检索增强生成与思维链的多模态大语言模型交通事故责任认定方法。首先，以交通事故监控视频为输入，通过采样提取事故关键帧序列；其次，构建交通法规知识库，利用BM25与向量检索的混合检索策略及倒数排序融合算法召回相关法律条款，以抑制大模型在法律引用上的“幻觉”现象；最后，设计思维链提示策略驱动模型进行多步责任推理，自动生成包含事故事实描述、法律条款依据及责任认定结论的结构化《道路交通事故认定书》。实验结果表明：本文方法在真实事故数据集上的责任认定准确率达80.00%，较未引入检索增强生成与思维链机制的基线模型提升6.67%；法条引用准确率由6.67%提升至56.67%，显著缓解条款错引与内容编造现象；生成文书在关键信息覆盖、语义一致性和格式规范性等方面达到可用水平。

关键词: 智能交通, 交通事故责任认定, 多模态大语言模型, 道路交通事故, 检索增强生成, 思维链

Abstract: Traffic accident liability determination is a key aspect in improving the governance capacity of traffic safety. In view of the high manual dependence, strong subjectivity, and low efficiency in current traffic accident liability determination in China, this paper proposes a Multimodal Large Language Model that incorporates Retrieval-Augmented Generation (RAG) and Chain-of Thought (CoT). First, accident keyframe sequences are sampled and extracted from surveillance videos. Then, a traffic law knowledge base is constructed, and relevant legal clauses are retrieved using a hybrid strategy that combines BM25 and vector retrieval with Reciprocal Rank Fusion (RRF), to suppress "hallucinations" in legal citations. A CoT prompting strategy is then designed to drive the model to perform multi-step liability reasoning, automatically generating a structured Road Traffic Accident Determination Report, including accident facts, legal bases, and the final liability conclusion. Experimental results on a real-world accident dataset show that the proposed method attains an accident liability determination accuracy of 80.00%, improving by 6.67% over the baseline without RAG and CoT. Meanwhile, the legal provision citation accuracy increases from 6.67% to 56.67%, significantly alleviating erroneous citations and fabricated legal content. The generated reports demonstrate practical usability in terms of key information coverage, semantic consistency, and format standardization.

Key words: intelligent transportation, traffic accident liability determination, multimodal large language model, road traffic accident, retrieval-augmented generation, chain-of-thought

中图分类号: