Cross-sector generalization of accident-process role classification in occupational accident narratives

TL;DR

The study evaluates how well automated role classification models trained on construction-sector accident narratives generalize to metallurgy, chemistry-plastics, and company-specific corpora. Task-specific adaptation significantly improves cross-sector performance over frozen pretrained models, achieving consistent balanced accuracy above 85% across domains.

Why It Matters

Automated analysis of occupational accident narratives can support large-scale risk prevention across industries. Reliable cross-sector generalization enables the deployment of unified coding systems without retraining, reducing barriers to adopting AI-assisted safety analysis in diverse workplaces.

Key Findings

• Task-specific adaptation of pretrained models outperforms frozen representations in cross-sector transfer for accident narrative role classification.
• The best-performing models maintained balanced accuracy between 85.6% and 85.8% across three unseen industrial sectors without any domain-specific tuning.
• A single model trained on construction data can effectively classify roles in heterogeneous accident narratives from other sectors, supporting scalable safety analysis.