Construction and Application of Traffic Accident Knowledge Graph Based on LLM

Records of traffic accidents contain a wealth of information regarding accident causes and consequences. It provides a valuable data foundation for accident analysis. The diversity and complexity of textual data pose significant challenges in knowledge extracting. Previous research primarily relies on Natural Language Processing (NLP) to extract knowledge from texts and uses knowledge graphs (KGs) to store information in a structured way. However, the process based on NLP typically necessitates extensive annotated datasets for model training, which is complex and time-consuming. Moreover, the application of traffic accident knowledge graphs by direct information querying within the graph requiring complex commands, which leads to poor interaction capabilities. In this study, we adapt an innovative approach integrates Large Language Models (LLMs) for the construction and application of a traffic accident knowledge graph. Based on the defined schema layer of the traffic accident knowledge graph, we employ LLMs to extract knowledge from accident records and refine the extraction process by using prompts and few-shot learning mechanism. To ensure the accuracy of the extracted result, we employ a dual verification method combines self-verification of LLMs with manual inspection. Then we visualize the knowledge by using Neo4j. Finally, we explore the application of KGs within the framework of Retrieval-Augmented Generation (RAG) and construct an intelligent question-answering system. The combination of LLMs and KGs facilitates a framework of semi-automated knowledge extraction and analysis. The Knowledge Graph-Based Retrieval-Augmented Generation Question Answering System for Traffic Accidents enables complex query and answering tasks such as causation analysis and scenario generation for autonomous driving tests. The integration of KGs and LLMs not only expands the application scenarios of KGs but also reduces the risk of hallucination in responses generated by LLMs. This method efficiently Extracting information from unstructured textual data, advances the digitalization and intelligence of traffic accident management.