Bingyou Dai, Xuesong Wang, Fengchun Yang, Yuxiang Feng, Yinhai Wang, Mohammed Quddus
Abstract: Crash causation analysis is essential for designing effective safety countermeasures. However, distilling lengthy in-depth crash investigation records into structured and interpretable causation representations remains time-consuming, labor-intensive, and highly dependent on expert judgment, limiting its scalability. To address this limitation, this study develops an LLM-based framework for case-specific crash causation-chain inference and constructs the Road-Crash-Causation-Chain dataset. Specifically, a semi-automatic annotation pipeline was designed based on the Driving Reliability and Error Analysis Method (DREAM), and the original graph-based DREAM causation structure was reformulated into a linearized representation suitable for LLM generation. After expert manual review, a total of 4000 crash cases were used to construct the Road-Crash-Causation-Chain dataset. To better evaluate structured DREAM outputs, two task-specific metrics, vehicle-node F1 and chain-validity rate, were further introduced to measure node-level correctness and structural validity, respectively. Based on this framework, multiple open-source LLM backbones were fine-tuned and evaluated, and Qwen3-14B was adopted to build the final model, CrashCausation-Qwen14B. Experimental results show that CrashCausation-Qwen14B achieves strong performance on the DREAM crash causation-chain inference task, with a vehicle-node F1 of 0.824, a ROUGE-L of 0.899, and a chain-validity rate of 0.973. Additional evaluations, including cross-dataset testing, human evaluation, time-cost comparison, ablation analysis, and qualitative comparison with existing traffic-safety frameworks, further demonstrate its generalizability, practical usefulness, and task-specific advantages in case-specific crash causation-chain inference. Overall, the proposed framework provides a scalable approach for case-specific crash causation-chain inference and establishes a foundation for integrating in-depth crash investigation, the DREAM framework, and large language models in traffic safety research.
引用：Bingyou Dai, Xuesong Wang, Fengchun Yang, Yuxiang Feng, Yinhai Wang, Mohammed Quddus. Leveraging Large Language Models for Crash Causation Chain Inference with In-Depth Accident Investigation Data. Transportation Research Part C: Emerging Technologies, 2026, 191: 105820.