Understanding workplace accidents is crucial for improving safety and preventing future incidents. However, accident reports are often written differently depending on the industry or company, making it hard for automated systems to analyze them at scale. A newly published research paper explores how well AI can classify key parts of accident narratives across different industrial sectors, potentially paving the way for better tools to support workplace safety experts.
Key Takeaways
- The study developed AI classifiers that categorize sentences in accident reports into four roles: work situation, unfavorable conditions, accident events, and consequences.
- These classifiers were trained only on construction sector reports but tested on reports from metallurgy, chemistry-plastics, and an independent company dataset without additional retraining.
- Task-specific fine-tuning of AI models significantly improved their ability to generalize across sectors compared to using fixed, pretrained models.
- The best AI models achieved balanced accuracy scores above 85% when classifying accident narratives in unseen industries.
The researchers began with a large collection of French workplace accident reports from the construction industry. Experts annotated over 42,000 factual units—essentially sentences or phrases—labeling them according to their role in describing the accident process. These roles included the work situation (what was happening at work), unfavorable conditions (hazards or risks explicitly mentioned), the accident event itself, and the consequences of the accident.
Using this annotated data, the team trained AI models to automatically assign these role labels to new accident report texts. The real test was whether these models, trained only on construction reports, could accurately classify similar information in reports from very different sectors like metallurgy and chemical manufacturing, which often use distinct terminology and writing styles.
To tackle this challenge, the researchers compared different AI approaches. One approach used “frozen” pretrained language models, which means the AI’s understanding of language was fixed and not adjusted during training for this specific task. The other approaches involved “task-specific fine-tuning” or supervised learning, where the AI models were further trained to better recognize the particular patterns and vocabulary related to accident narratives.
Results showed that models adapted through task-specific fine-tuning consistently outperformed the frozen models. This means that by teaching the AI to focus on the accident classification task itself, it became better at handling the variations in language and style found in different industries. Across multiple test datasets from sectors the AI had never seen before, the fine-tuned models achieved balanced accuracies around 85.6% to 85.8%, indicating strong and reliable performance.
These findings have important implications for occupational safety. Automated systems that can accurately structure and classify accident reports from multiple industries could help safety experts identify common risk factors and accident patterns more efficiently. This cross-sector generalization means that companies and regulators may one day use AI tools to analyze diverse accident data without needing to retrain models for each specific context.
Looking ahead, further research could explore applying these methods to accident reports in other languages or sectors, as well as integrating such AI tools into real-world safety management systems. By improving how workplace accidents are documented and understood at scale, this technology could contribute to safer work environments across a wide range of industries.
Based on research published on arXiv by Aho Yapi, Pierre Latouche, Arnaud Guillin et al..