Leveraging State-of-the-Art Large Language Models for Accident Analysis in the Highway Construction Industry

Abstract

Work zone hazards in the highway construction industry pose significant risks to worker safety,
necessitating effective accident prevention measures. According to the Occupational Safety and
Health Administration (OSHA)'s severe injuries database, the highway construction industry ranks
among the top 10% of contributors to all cases. However, existing research in construction safety
primarily focuses on work zone vehicle intrusion, building construction, and the mining industry,
resulting in a notable gap in understanding accidents specific to highway construction, including
rehabilitation and new construction projects.
To bridge this gap, this study proposes leveraging advanced large language models (LLMs), such
as GPT-3, to enhance data-driven analysis. While LLMs have demonstrated their effectiveness in
various scientific domains, their application in construction safety remains limited, with minimal
exploration of their potential for analyzing textual accident data. Previous industry research
predominantly relies on traditional machine learning approaches for natural language processing
(NLP) tasks, indicating untapped potential to gain deeper insights and improve accident analysis
in highway construction safety. This study aims to achieve efficient data processing, NLP
understanding, and contextualized insights by utilizing publicly available databases such as
OSHA's severe injuries database. The use of these databases underscores the novel and convenient
approach employed in this study, utilizing state-of-the-art LLMs to analyze extensive amounts of
textual accident data and facilitate identifying major causes of accidents. This analysis is
complemented by traditional statistical approaches, clustering algorithms, and classification
techniques. The research outcomes are expected to significantly contribute to improving accident
analysis and prevention in the highway construction industry.