Representation learning of crash narrative using natural language processing models

Understanding crash scenarios is critical for traffic safety research and automated driving system (ADS) evaluation. This paper presents a framework that leverages large language models and statistical learning techniques to construct a continuous crash space from unstructured narrative text.

Crash narratives are embedded into a high-dimensional semantic space using Sentence-BERT. Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP) are applied for dimensionality reduction to support quantitative analysis and visualization. K-means clustering combined with term frequency-inverse document frequency (TF-IDF) is used to extract interpretable, population-level crash characteristics. Distance-based measures, including Gaussian density estimation, enable identification of similar and dissimilar crashes and detection of within-category corner cases.

The framework is applied to 12,741 crashes from the Crash Investigation Sampling System (CISS), a nationally representative sample of U.S. motor-vehicle crashes. The resulting crash space shows clear structural organization: similar crashes cluster together, distinct crash types form separate regions, and rare or atypical events appear as low-density outliers. Three-dimensional visualization further supports interpretation of crash relationships.

The proposed approach effectively represents a wide variety of crash scenarios and demonstrates that language models substantially enhance feature extraction from narrative text beyond what is possible using structured variables alone.

The crash space supports intuitive visualization of crash patterns, grouping of similar scenarios, and systematic identification of rare and safety-critical corner cases, providing a useful tool for traffic safety analysis, policy development, crash mitigation strategies, and ADS evaluation.