Leveraging Transformers for Workplace Safety: Insights from OSHA Accident Data

Abstract

Accidents at work present serious hazards to employee safety and a loss of productivity for organizations, which makes a proactive analysis of accident data of paramount importance. The Occupational Safety and Health Administration (OSHA) has provided a rich dataset of accidents, which has a variety of structured (organized) fields and unstructured (narrative) text that describe circumstances, hazards, and consequences. While the structured data will accommodate some basic statistical analysis, the narrative text has not been analyzed, simply because it was not usable. Thus, this study demonstrates how to employ transformer-based natural language processing (NLP) models (BERT and RoBERTa) to analyze OSHA accident reports and process their information. We fine-tune each of the transformer models to automatically classify accident types, identify contributory factors, and discover recurring patterns in hazard types across different industries. In addition, while attempting to examine the narrative data collectively, we employ clustering techniques and semantic similarity to create clusters of related incidents. We also analyzed a text summarization approach to condense the length of reports into concise safety notes, facilitating a quicker review process. Our analysis practices have uncovered high-risk industries, frequent factors that could lead to an accident, and common scenarios that could lead to an accident occurring, which we view as valuable input for safety managers and policy makers. Our study speaks well to the ability of transformer models when examining and transforming raw narrative data into knowledge in a structured manner, and identifying entangled evidence-based interventions and prioritizing hazards to mitigate risk. This study provides evidence that deep learning has enhanced our analysis capacity through a framework with the potential to support change that reduces the incidence of workplace accidents and sustainable change to safer workplaces.