LLM-driven causal chain extraction: An interpretable framework for autonomous vehicle crash narrative analysis.

Area of Science:

• Artificial Intelligence
• Autonomous Systems
• Transportation Safety

Background:

• Existing autonomous vehicle (AV) crash analysis lacks interpretable causal attribution.
• Limited utilization of unstructured textual data hinders mechanistic insights into AV accidents.
• Fragmented causal attribution in AV crashes necessitates advanced analytical frameworks.

Purpose of the Study:

• To establish an interpretable framework for analyzing root causes of autonomous vehicle (AV) crashes.
• To leverage unstructured crash narratives for mechanistic insights into AV accident causation.
• To address gaps in fragmented causal attribution in current AV safety research.

Main Methods:

• Integrated framework combining Large Language Models (LLMs) and Chain-of-Thought (CoT) reasoning.
• Sentence-level resampling for data augmentation and an instruction-tuned LLM for extracting Crash Causality Frames (CCFs).
• System-theoretic taxonomy mapping CCFs to causal indicators and CoT for generating natural-language explanations.

Main Results:

• Optimized LLaMA-70B+LoRA model achieved 97.93% accuracy in CCF extraction after data resampling.
• Identified five dominant causation patterns, with AV-CV interaction failures (51.9%) being most prevalent.
• CoT module generated auditable causal chains with 91.04% accuracy, enhancing interpretability.

Conclusions:

• The framework transforms unstructured narratives into interpretable causal models for AV crashes.
• Systemic interactions, especially AV-CV behavioral mismatches, are primary crash catalysts.
• Practical implications include enhanced intention prediction, context-aware sensor fusion, and improved takeover training protocols.