Advancing construction safety analysis: A comparative study of SBERT-KMeans and LDA in extracting highway work-zone accident topics

Abstract

Highway work zones are widely recognized as some of the most hazardous occupational environments due to the simultaneous presence of construction activities and continuous traffic flow. The U.S. Occupational Safety and Health Administration (OSHA) accident reports are incredibly valuable for tracking safety risks and incident trends. Many studies have used latent Dirichlet allocation (LDA) to analyze these documents, but LDA does not usually capture the deeper semantic context and cannot handle the messy and diverse descriptions of accidents. This study explores an alternative approach to this issue by applying SBERT-KMeans topic modeling to analyze 1,051 OSHA reports from January 2015 to February 2022. The performance of SBERT-KMeans is first compared with a traditional Latent Dirichlet allocation (LDA) model. In addition to the classical measures of topic coherence and topic diversity, which are often optimized for single-word representations, the present study also evaluates the models' ability to preserve syntactic structure and recognize complex technical phrases. The results indicate that SBERT-KMeans provides a more phrase-aware representation of domain-specific concepts and technical expressions. Moreover, the extracted safety-risk-related topics are integrated with structured tabular data to evaluate severity correlations (hospitalization and amputation rates), while a topic semantic similarity network is employed to uncover highly interconnected topic hubs. The comparison suggests that SBERT-KMeans provides phrase-aware semantic representations that complement the unigram-based LDA baseline and can support data-driven safety analysis for highway construction management.