Thematic analysis of urban highway crash narratives using machine learning classifiers

This study presents a novel and robust framework to consequently classify fataland injury crash narratives obtained from urban highway crash reports of Tehran,and extract accident causal factors from unstructured raw texts. With an evergrowingamount of textual information stored in crash narratives, automaticinformation retrieval is highly regarded. Five main accident causal categories,namely, driver behavior, inadequate road characteristics, inadequate trafficdesign, roadway intrusion, and vehicle malfunction have been chosen as classlabels, to classify crash narratives in this research. The approach towards solvingthis multiclass classification problem has been multiple binary classifications.Machine learning classifier SVM (support vector machines) with linear kerneland Decision Tree Networks have proved to outperform other ML classifieralgorithms. Classification reports reveal the sensitivity of drivers’ perilousbehavior, among all possible accidental causes, in injury and fatal crashes ofTehran urban highways. This research proves Natural language processing (NLP)algorithms implementation in accident analysis to be promising particularlywhen trained by a large corpus of narratives.

Keywords:

Text analytics , machine learning , classification , Accident analysis