ScholarWorks@경상국립대학교

초록

In 2024, 25,027 occupational accidents were reported in the construction industry—the highest among all sectors—with 373 fatalities, accounting for 23.8% of all work-related deaths. A significant contributing factor to these accidents is non- compliance with safety regulations. Therefore, effective accident prevention requires accurately identifying relevant regulations and implementing systematic safety measures. However, retrieving appropriate regulations is often hindered by subjective interpretation and misunderstanding, which can lead to errors. Accident reports from construction sites contain unstructured text that reflects causal factors and offers valuable information for legal analysis. In contrast, legal documents are hierarchically structured, with cross-references between upper-level articles and sub-clauses. Ignoring this structure can result in the misapplication of laws. To address these challenges, this study proposes a legal clause retrieval method that accounts for the structural characteristics of legal texts. A dataset was created by dividing legal documents into upper-level articles and their corresponding sub-clauses. We developed a system that automatically retrieves the most relevant regulations from the Occupational Safety and Health Act based on accident narratives recorded in the construction safety information (CSI) system. Using Sentence-BERT, both accident descriptions and legal clauses were embedded as vectors, and semantic similarity was calculated. A two-stage retrieval process was implemented: first, identifying the most relevant article, then selecting the most appropriate sub-clause within it. This approach enables accurate and systematic legal retrieval, enhancing safety management and supporting regulatory compliance in the construction industry.

키워드