김석기  원정훈

Vol. 40,  No. 5, pp. 65-75, Oct.  2025
10.14341/JKOSOS.2025.40.5.65

small-scale construction BERTopic Large Language Model fatal accidents

PDF

This study analyzed the patterns and causes of fatal accidents at small-scale construction sites using an analysis technique that combines BERTopic and GPT-4o-based Large Language Models (LLM). To overcome the limitations of the c-TF-IDF method used in the existing BERTopic, a large-scale language model based on GPT-4o was used to significantly improve the contextual comprehension of accident history texts containing mixed technical terms. The patterns and causes of fatal accidents at 431 small-scale construction sites, which were excluded from the establishment of the safety management plan, were investigated. Eleven major accident clusters were identified, and unique accident patterns and causal characteristics for each cluster were determined. Among the clusters, the cluster with the highest accident frequency was ”worker negligence,” followed by ”inadequate safety measures” and ”inadequate safety management plan.” It was found that management factors accounted for a higher proportion of causes of accidental deaths than worker factors. In particular, upper-level management factors such as safety management plans, safety education systems, and safety investments accounted for a higher proportion. Therefore, addressing worker negligence, insufficient safety measures, and inadequate safety measures, and inadequate safety management plans at small-scale construction sites should be a top priority. Furthermore, the lack of a fundamental safety management system at small-scale construction sites is believed to be a key cause of accidents.