Kim J, Trueblood AB, Kum HC, Shipp EM. Crash narrative classification: Identifying agricultural crashes using machine learning with curated keywords.
TRAFFIC INJURY PREVENTION 2020;
22:74-78. [PMID:
33206551 DOI:
10.1080/15389588.2020.1836365]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [MESH Headings] [Track Full Text] [Subscribe] [Scholar Register] [Received: 04/07/2020] [Revised: 09/21/2020] [Accepted: 10/08/2020] [Indexed: 06/11/2023]
Abstract
OBJECTIVE
Traditionally, structured or coded data fields from a crash report are the basis for identifying crashes involving different types of vehicles, such as farm equipment. However, using only the structured data can lead to misclassification of vehicle or crash type. The objective of the current article is to examine the use of machine learning methods for identifying agricultural crashes based on the crash narrative and to transfer the application of models to different settings (e.g., future years of data, other states).
METHODS
Different data representations (e.g., bag-of-words [BoW], bag-of-keywords [BoK]) and document classification algorithms (e.g., support vector machine [SVM], multinomial naïve Bayes classifier [MNB]) were explored using Texas and Louisiana crash narratives across different time periods.
RESULTS
The BoK-support vector classifier (SVC), BoK-MNB, and BoW-SVC models trained with Texas data were better predictive models than the baseline rule-based algorithm on the future year test data, with F1 scores of 0.88, 0.89, 0.85 vs. 0.84. The BoK-MNB trained with Louisiana data performed the closest to the baseline rule-based algorithm on the future year test data (F1 scores, 0.91 baseline rule-based algorithm vs. 0.89 BoK-MNB). The BoK-SVC and BoK-MNB models trained with Texas and Louisiana data were better productive models for Texas future year test data with F1 scores 0.89 and 0.90 vs. 0.84. The BoK-MNB model trained with both states' data was a better predictive model for the Louisiana future year test data, F1 score 0.94 vs. 0.91.
CONCLUSIONS
The findings of this study support that machine learning methodologies can potentially reduce the amount of human power required to develop key word lists and manually review narratives.
Collapse