Wei, Bangyan (Sun Yat-sen University), Xu, Yonggao (Affairs Center of Road Transport of Guangdong Provice), He, Zhaocheng (Sun Yat-Sen University), Yi, Zhijun (Affairs Center of Road Transport of Guangdong Provice), Wu, Xiaoyun (Affairs Center of Road Transport of Guangdong Province), Zhu, Yiting (Sun Yat-Sen University)

Knowledge Graph and Large Language Model-Driven Intelligent Traffic Assistant for Question Answering and Risk Analysis of Commercial Vehicles

Scheduled for presentation during the Invited Session "S22b-Emerging Trends in AV Research" (TH-EA-T22), Thursday, November 20, 2025, 13:50−14:10, Coolangata 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

Reducing the accident risk of commercial vehicles that transport passengers, heavy goods, and dangerous goods is significant to society. By installing cameras and positioning devices on these vehicles, transportation authorities can effectively identify risky driving behaviors (e.g., overspeed, using mobile phones, etc.) and obtain spatiotemporal information about when and where such events occur. Leveraging traffic incident data, including risky driving behaviors and traffic accidents, for risk analysis of commercial vehicles and providing an intelligent assistant for question-answering (Q&A) for transportation departments is highly valuable research. This study first extracts traffic object entities and relationships from traffic incident data to construct Risk Knowledge Graphs (RKG) of commercial vehicles. Next, taking the RKG triplets as input, this study designs an intelligent traffic assistant based on Large Language Models (LLMs) for knowledge retrieval and interactive Q&A on commercial vehicle data. Additionally, this study designs a Self-Verification Chain-of-Thought(SV-CoT) incorporating prompt engineering and self-feedback mechanisms, which enhances the reasoning accuracy and reliability of the LLM. In the experiments, this study analyzes the distribution characteristics of commercial vehicle risk values and designs 13 representative questions across five categories to test the ability of different LLMs to answer these questions.