Machine learning of motor vehicle accident categories from narrative data

Area of Science:

• Machine Learning
• Traffic Safety
• Data Science

Background:

• Accident narratives contain valuable data for understanding crash causes.
• Automated analysis of these narratives is challenging due to varied language and lack of keywords.
• Machine learning offers potential for extracting structured information from unstructured text.

Purpose of the Study:

• To evaluate Bayesian inferencing for identifying pre-crash activity and crash type.
• To assess the model's ability to learn and classify accident narratives lacking specific keywords.
• To compare Bayesian model performance against expert ratings and keyword search.

Main Methods:

• Applied Bayesian inferencing, a machine learning technique, to 3,686 motor vehicle crash narratives.
• Developed a keyword search for 63 accident-related terms to train the Bayesian model.
• Tested the model's classification accuracy on narratives with and without predefined keywords.

Main Results:

• When keywords were present, the Bayesian model closely matched expert ratings (P(detection) >= 0.9, P(false positive) <= 0.05).
• For narratives without keywords, detection rates varied from 67% to 12% as the Bayesian threshold changed.
• False positives decreased from 32% to 3% with increased Bayesian thresholds, indicating successful learning.

Conclusions:

• Bayesian inferencing is a viable machine learning technique for analyzing accident narratives.
• The model demonstrated learning capabilities, improving classification of narratives lacking initial keywords.
• This approach enhances the potential for automated traffic safety analysis from textual data.