Automated Hazardous Action Classification Using Natural Language Processing and Machine-Learning Techniques

摘要

Information on hazardous actions of parties involved in a crash could support determination of crash responsibilities. However, information on hazardous actions that is explicitly recorded in a crash report may often be inconsistent with the narrative given in a crash report. Identification of such inconsistencies requires a large amount of manual effort. To address this issue, a new method is proposed in this paper to automatically classify hazardous actions described in a crash report based on the narrative. The proposed method leverages natural language processing (NLP) techniques to extract features from the narratives, as well as machine learning (ML) techniques to classify the hazardous action described in a narrative based on its values corresponding to selected features. The proposed method was preliminarily tested on a randomly selected set of crash reports from the State of Michigan. An accuracy of 92.77% and a Kappa statistic of 83.54% were achieved on the testing data, which shows that the proposed method is promising.