Ordinal BERT-CNN: an intelligent semantic framework for automated risk extraction and severity classification from unstructured crash narratives

The widespread adoption of Intelligent Transportation Systems (ITS) generates massive volumes of unstructured data, particularly police crash narratives, which remain an underutilized resource for automated risk perception. Traditional traffic safety analysis relies heavily on manual feature engineering from structured databases, creating a bottleneck for real-time, data-driven decision support. To bridge this gap, this study proposes Ordinal BERT-CNN (O-BERT-CNN), an intelligent semantic fusion framework that automatically extracts latent behavioral risk factors from unstructured free-text and classifies crash severity. The framework integrates a pre-trained BERT encoder for global contextual understanding, a TextCNN module for local key-phrase extraction, and an ordinal regression layer that respects the hierarchical nature of injury severity (PDO, Injury, Fatality). Validated on 15,603 crash reports from the Michigan Traffic Crash Facts database, O-BERT-CNN achieves a classification accuracy of 0.8907 and a mean absolute error of 0.1137, significantly outperforming traditional and baseline deep learning models. Beyond classification, the framework employs BERTopic topic modeling and SHAP interpretation to translate black-box textual data into transparent, actionable intelligence. The results reveal distinct semantic patterns: low-speed operational errors dominate property-damage-only crashes, intersection conflicts and adverse weather link to injury crashes, and high-risk behaviors such as speeding and alcohol strongly predict fatal outcomes. By demonstrating how unstructured text can be automatically mined to uncover hidden risk dynamics, this study provides an offline analytical foundation for developing intelligent semantic modules in ITS, with a clear pathway toward real-time ITS architectures for proactive safety management and targeted enforcement.