Comparative Analysis of BERT and RoBERTa for Construction Site Incident Report Classification

Most studies that investigated the automated classification of construction incident reports relied on traditional machine learning approaches, line support vector machines, decision trees, and logistic regression. Those approaches, however, have limited capacity to capture rich contextual text information to inform predictions. This study, therefore, investigates the effectiveness of two transformer-based models, BERT and RoBERTa, in classifying construction site incident reports. The purpose of this research is to evaluate the performance of these models in capturing contextual and semantic nuances in incident reports and to determine which model is better suited for supporting safety management tasks in the construction industry. A comparative framework was employed, where both models were fine-tuned on a dataset of construction site incident reports obtained from the United States Occupational Safety and Health Administration (OSHA). The results show that while both models achieved high accuracy, precision, and recall, RoBERTa demonstrated superiority in capturing more relevant context to inform its predictions. Specifically, RoBERTa outperformed BERT in classifying incidents related to "Caught in/between objects" and showed marginal improvements in other categories. The findings of this research have implications for construction companies seeking to automate site incident report analysis for timely decision-making and provide insights for researchers deciding between BERT and RoBERTa for classification tasks. This study contributes to developing more accurate and reliable safety management systems in the construction industry by adopting transformer-based models to automate the analysis of incident reports.