The study investigates whether a classification model trained on accident narratives from one sector can identify structured information in narratives from other sectors and reporting environments. French accident narratives were segmented into factual units and annotated as work situation, unfavourable condition, accident event, or consequence. Models were trained on 42,244 units from 6,040 construction-sector narratives and tested without retraining on metallurgy, chemistry-plastics, and a company corpus.
The best-performing approach, task-adapted pretrained models, achieved an average balanced accuracy of about 85.7%, compared to 75.0% for a TF-IDF baseline and 76.9% for frozen pretrained representations. Performance was lower and more variable on the company corpus, indicating challenges when the reporting environment also changes. No single method consistently outperformed others across runs.
This work is interesting because it demonstrates that accident-process roles can be transferred across sectors, enabling systematic analysis of large, heterogeneous accident-report collections. The framework supports assisted coding, expert review, and cross-sector safety analysis, potentially helping to identify recurring patterns relevant to prevention.