Abstract:

To address the limitations of relying on manual extraction of knowledge from complaint texts in automotive safety risk management, this study employed LLMs for automated risk discrimination. First, over 50 000 complaint texts covering eight major subsystems such as the engine were collected. A demonstration sampling method based on Bilingual and Crosslingual Embedding(BCEmbedding) model and community detection algorithm was proposed to construct a diverse and high-quality example knowledge base. Secondly, prompts were designed from perspectives such as scenarios, skills, and examples to develop agents capable of performing risk keyword extraction and expansion, as well as subsystem risk categorization. Finally, by analyzing reasoning knowledge from retrieved texts and utilizing chain of thought techniques, a risk level knowledge base for retrieved texts was established. A consensus-seeking multi-agent(MA) system based on LLMs was designed, resulting in a discriminative model for automotive safety risk levels. The results show that the model not only reduces labor costs but also achieves high accuracy and efficiency. It effectively supports risk term extraction, risk categorization, and risk level assessment in complaint incidents, thereby enhancing safety risk management.

Key words: large language models (LLMs), agent, automobile safety risk, risk discrimination, complaint text, prompt