Deep Learning and Clustering-Based Analysis of Text Narratives for Identification of Traffic Crash Severity Contributors

Crash narratives provide valuable information to understand traffic crashes and develop roadway safety countermeasures. However, manually reading long text narratives is time-consuming and error-prone. This study presents a deep-learning and clustering-based approach to identifying contributors to traffic crash severity in text narratives. We evaluate the approach using a dataset of narratives from Massachusetts and compare different deep-learning models for semantic similarity. The approach clusters semantically similar phrases in the narratives and provides an overview of frequent topics related to severe crashes, offering a valuable tool for roadway safety analysis and countermeasure development.