Toward explainable traffic crash liability analysis via neural-symbolic multimodal reasoning: Integrating visual perception with legal knowledge graphs.

Determining traffic crash liability relies on rigorous on-site evidence and legal provisions and involves substantial manual intervention and potential human error. However, intelligent traffic monitoring systems focus on detection and tracking and cannot provide structured, rule-grounded decision support for formal crash-liability analysis. Moreover, emerging multimodal large language models (MLLMs) suffer from object hallucination that distorts or omits physical evidence, and their opaque reasoning cannot produce explicit evidence-to-rule chains required for traceable and reviewable liability analysis. This paper proposes CrashLex, an explainable neural-symbolic decision-support framework that integrates visual perception with legal knowledge graph reasoning to align multimodal inference with predefined statutory rules and traceability requirements. Within CrashLex, a dual-query vision-language architecture employs parallel semantic and symbolic streams to mitigate object hallucination, utilizing Scene Queries for global context modeling and Perception Queries for fine-grained atomic feature extraction, respectively. The domain-specific legal knowledge graph maps these atomic facts from observable physical conditions to corresponding statutory elements, providing a structured basis for rule-grounded assessment. A logic-driven aggregation protocol then imposes explicit constraints derived from extracted facts and statutory provisions onto the inference process, thereby enabling traceable and interpretable preliminary liability assessment. On the held-out DeepAccident benchmark, CrashLex achieves a Legal Accuracy of 94.0% under the fixed legal-output evaluation protocol and an Object Hallucination Rate of 2.8%, corresponding to an 82.1% relative reduction from the 15.6% rate of the best-performing MLLM baseline on this metric. CrashLex grounds its assessment results in the deterministic execution of predefined legal rules rather than relying solely on unconstrained probabilistic generation, providing verifiable reasoning chains for rule-grounded human review.