Abstract Manager

Many Canadian agencies maintain large collision datasets, but the most detailed information is buried in unstructured text such as claims descriptions, police notes, and free‑text reports. This information is prohibitively time‑consuming to manually extract and structure. At the same time, privacy constraints often prevent sending narratives to third‑party AI services, limiting their use in safety analytics and planning.

This presentation describes how Parsons used ParsonsGPT, a privately hosted large language model, to transform over 41,000 incident records into structured, analysis‑ready data. The goal was to break down narrative collision records into standardized categories for entities, movements, conflict types, and contributing factors, in a form that can be used directly in statistical and spatial analysis. This structured dataset is intended to reveal systemic collision patterns and support concrete decisions about where and how the City of Vancouver should act next on road safety.

The project involved preprocessing collision records, integrating supplementary geospatial datasets (City intersections, streets, signals, active transportation facilities, amenities), and defining a detailed data dictionary and attribute schema. For each incident, ParsonsGPT received standardized fields (datetime, severity, mode flags), multi‑report narratives, and selected spatial context derived from existing geocoding (latitude/longitude, nearest street and intersection, basic signal, weather, and lighting attributes). AI interpreted narratives using already‑geocoded and GIS‑linked data to infer road users, direction of travel, movement preceding collision, primary contributing factors, likely at‑fault parties, and high‑level collision type.

Quality control combined manually auditing a sample of records with monitoring “unknown” and ambiguous classifications to identify systematic issues and refine prompts and schema. The resulting dataset now powers further analysis of conflict types and contributing factors by mode, location type, and infrastructure context. The presentation will share practical lessons from this privacy‑preserving AI workflow that can help other jurisdictions safely unlock similar narrative collision data for transportation safety analytics.