Construction occupational accident analysis performed by a Large Language Model using the Bow-Tie model

Preprint, 2024

This study explores the use of large language models (LLMs) to automate accident data analysis in a construction organization by mapping accident cases onto the Bow-Tie model template. Accident reports from a large Swedish contracting company were analyzed using the in-context learning (ICL) method. Results show that LLMs successfully performed accident analysis when provided with ICL demonstrations, reducing output hallucination and maintaining consistency with predefined analysis structures. However, challenges such as data quality, domain complexity, and minor output inconsistencies were observed. Additionally, the model's efficacy in proactive accident prevention remains inconclusive. Despite these limitations, the study demonstrates the potential of integrating LLMs with accident causation models (ACMs) to enhance learning from registered accident data in construction companies, though further research is needed.