Integrating pavement condition records with LLM-based crash narrative analysis for pavement safety assessment

This paper proposes an integrated framework that couples Large Language Models (LLMs)-based crash narrative analysis with quantile regression to identify and quantify pavement-related crash risk. The LLM component converts unstructured police narratives into structured, mechanism-specific labels (e.g., hydroplaning, curve-related loss of control), which enables outcomes that are directly linked to pavement and roadway-surface conditions that are often missing from conventional structured crash fields. These LLM-derived mechanism labels are then matched to segment-level pavement condition, friction, texture, traffic exposure, and geometric characteristics and modeled using quantile regression to characterize how covariate effects vary across the full distribution of crash risk rather than only at the mean. A case study using over 24,000 police crash narratives linked to a pavement management dataset of approximately 180,000 data records demonstrates strong associations between friction/texture measures and wet-pavement crash mechanisms. These results can help transportation agencies select candidate pavement-safety projects by identifying pavement conditions associated with elevated crash risk and prioritizing targeted, cost-effective countermeasures.