Automated information mining in hazardous chemical accident reporting: An improved deep learning approach

The authors propose an automated accident-analysis framework for hazardous chemical accident reports using deep learning and topic mining. They construct a named‑entity dataset for accident entities, apply a RoBERTa–BiLSTM–Attention–CRF NER model to extract structured information from reports, and use topic analysis (e.g., LDA) to discover causation themes such as human error and rescue-related secondary accidents. For PSM incident investigation, this enables scalable extraction and aggregation of investigation findings, helping organizations analyze patterns and prevent recurrence.