| Title |
Automatic Classification of Unstructured Construction Accident Data and Risk Factor Identification Using Explainable AI |
| Authors |
고태규(Ko, Tae-Gyu) ; 한재영(Han, Jae-young) ; 김기남(Kim, Ki-nam) ; 이민재(Lee, Min-jae) |
| DOI |
https://doi.org/10.5659/JAIK.2026.42.8.397 |
| Keywords |
Construction Safety; Unstructured Data; Weak Supervision; BERT; SHAP |
| Abstract |
Construction accident data is an important resource for safety management, but its practical use is often limited by the high cost of labeling
unstructured text and the black-box nature of deep learning models. This study focuses on efficiently classifying unstructured accident data
and identifying risk factors by combining weak supervision and Explainable Artificial Intelligence. A total of 9,777 accident reports were
collected from the Construction Safety Management Integrated Information system. A weak supervision algorithm was applied to reduce the
need for extensive manual labeling. Using this labeling approach, the BERT classification model achieved a high accuracy of 0.845 and a
Macro-F1 score of 0.821. SHAP-based analysis was then used to visually explain how the model identified risk factors. The results showed
that the model recognized risks not only through explicit terms such as work processes and accident objects, but also through underlying
behavioral causes. In particular, local interpretation demonstrated that the model could accurately distinguish accident types based on the
subject or object associated with the same verb, indicating strong contextual understanding. Unlike conventional black-box AI models, this
approach clearly explains the basis of the model’s decision-making process. By improving labeling efficiency and providing transparent
reasoning, the proposed methodology can help safety practitioners proactively identify potential risks and develop effective preventive
measures. |