Improving large language model assisted categorization and classification of highway construction accidents

The construction industry, particularly in highway work zones, grapples with persistently high rates of occupational injuries and fatalities, requiring a transformative approach to bolster safety measures. This study proceeds current research in safety by harnessing modern technologies, including natural language processing and machine learning. The primary objective presents a systematic approach for future understanding, preventing, and mitigating accidents, highlighting the need for enhanced safety protocols in targeted activities. The study conducts a comprehensive data-driven analysis based on text-rich narratives extracted from publicly available data, utilizing a subset of the Occupational Safety and Health Administration databases. The approach incorporates clustering and classification through machine learning techniques, with large language models (LLMs) playing a pivotal role in comprehension and dissemination. The research advances artificial intelligence (AI) in engineering by demonstrating innovative applications of LLMs in processing unstructured safety data, including the use of advanced prompting techniques like zero-shot and few-shot learning to automatically classify and cluster safety narratives. Anchored in OpenAI's Generative Pre-trained Transformer (GPT-3.5), the study positions LLMs as a foundation for further exploration, emphasizing their continuous evolution and ability to enhance safety analyses in construction or other industries with similar concerns.