Abstract Five machine learning models were trained using accident reports from the National Institute of Chemical Safety (NICS) that occurred in the Republic of Korea from 2014 to 2025 to classify the accidents’ causes. Even though some causes were difficult to distinguish because they were described using similar language in the accident reports, the models performed reasonably well with F1 scores of 0.7608, 0.8086, 0.7687, 0.8070, and 0.8215 and Area Under the Precision-Recall Curve (AUPRC) scores of 0.8284, 0.8354, 0.8548, 0.8434, and 0.8807, respectively, for Logistic Regression (LR), Random Forest (RF), Supporting Vector Machine (SVM), XGBoost, and Korean Bidirectional Encoder Representations from Transformers (KoBERT). A Local Interpretable Model-agnostic Explanations (LIME) XAI (explainable artificial intelligence) method was applied to analyze the decision-making process, which revealed symmetrical keyword reliance that reflects the inherent conceptual overlap between the two primary causes as defined by the NICS: “facility defect” and “noncompliance with safety standards”. Our results provide recommendations to overcome the inherent limitations of the NICS data set and highlight the necessity of a fine-grained classification framework using XAI approaches.
原文链接
分享
引用
