文章编号： 1672-5328（2025）02-0001-12

李健1, 2，朱国军1，王奥1，夏强1, 2，周胥君1，李毅喆1
(1. 同济大学道路与交通工程教育部重点实验室，上海201804；2. 同济大学城市交通研究院，上海201804)

摘要： 大语言模型凭借其强大的语义理解和生成能力成为街头巷尾热议的话题。虽然大语言模型在 处理通识性问答方面表现出色，但是在涉及复杂决策的行业领域仍普遍存在“幻觉”现象，且在可 解释性、可信度等方面问题突出。在梳理国内外研究现状的基础上，从知识图谱与大语言模型融合 的思路出发，提出了城市交通知识增强大语言模型系统架构，探索了提示词工程、检索增强生成、 模型融合及智能体构建技术，研发了城市交通知识增强大语言模型(TransKG-LLM)，并从数据增 强、知识增强、模型增强及任务增强等4 个维度进行了实践探索。研究结果表明，所提出的模型可 以缓解通用大模型的“幻觉”现象，有助于提升城市交通治理能力的科学化、精细化和智能化 水平。

关键词： 城市交通；生成式人工智能；知识增强生成；知识图谱；大语言模型

中图分类号： U491

文献标识码：A

Knowledge-Enhanced Large Language Models for Urban Transportation: Modeling and Applications

LI Jian1, 2, ZHU Guojun1,WANG Ao1, XIA Qiang1, 2, ZHOU Xujun1, LI Yizhe1
(1. Key Laboratory of Road and Traffic Engineering of Ministry of Education, Tongji University, Shanghai 201804, China; 2. Urban Mobility Institute, Tongji University, Shanghai 201804, China)

Abstract： Large language models (LLMs) have become a hot topic of discussion because of their powerful semantic understanding and generation ability. Although the large language models perform well in dealing with general knowledge-based questions and answers, there is still a widespread phenomenon of“hallucinations” in industries involving complex decision-making, and problems such as interpretability and credibility are prominent. On the basis of combining current domestic and aboard research, this paper puts forward a knowledge-enhanced system architecture for large language models in urban transportation starting from the perspective of integrating knowledge graphs with large language models. Furthermore, the paper explores the technologies of prompt word engineering, retrieval enhancement generation, model integration, and agent construction. A knowledge-enhanced LLM for urban transportation (TransKG-LLM) is developed. Practical explorations are conducted from four dimensions: data enhancement, knowledge enhancement, model enhancement, and task enhancement. The results indicate that the proposed model can alleviate the“hallucinations”phenomenon of the general large language model, and help to improve the scientific, refined, and intelligent level of urban transportation management ability.

Keywords： urban transportation; generative artificial intelligence; knowledge- augmented generation; knowledge graph; large language model