Construction accident narrative classification: An evaluation of text mining techniques

Area of Science:

• Construction Safety
• Data Science
• Machine Learning

Background:

• Accident and near-miss reporting are crucial for preventing future incidents.
• Classifying large volumes of safety data manually is time-consuming and resource-intensive.

Purpose of the Study:

• To evaluate the effectiveness of various text mining classification techniques for construction accident narratives.
• To identify the optimal machine learning algorithm for automating the classification of accident causes.

Main Methods:

• Utilized six machine learning algorithms: Support Vector Machine (SVM), Linear Regression (LR), Random Forest (RF), K-Nearest Neighbor (KNN), Decision Tree (DT), and Naive Bayes (NB).
• Applied techniques including tokenization and hyperparameter tuning (grid search) to optimize SVM models.
• Evaluated classifier performance on 1000 publicly available construction accident narratives from the US OSHA website.

Main Results:

• Support Vector Machine (SVM) algorithms, specifically linear SVM and Radial Basis Function (RBF) SVM with unigram tokenization, exhibited the best classification performance.
• Linear SVM achieved precision ranging from 0.5 to 1, recall from 0.36 to 0.9, and F1 scores between 0.45 and 0.92 across 11 accident cause labels.
• The study identified reasons for misclassification and proposed methods for performance enhancement.

Conclusions:

• Text mining, particularly using linear SVM, offers a viable and efficient solution for classifying construction accident narratives.
• Automated classification can significantly reduce the time and effort required for safety data analysis.
• Further research can refine models to improve accuracy and address complex misclassifications.