Characterizing accident narratives with word embeddings: Improving accuracy, richness, and generalizability

Area of Science:

• Occupational Health and Safety
• Data Science
• Machine Learning

Background:

• Workplace accidents pose significant risks to employee well-being and organizational finances.
• Analyzing textual accident narratives is challenging but crucial for identifying safety risks.
• Existing methods for analyzing accident data are often limited in scope and efficiency.

Purpose of the Study:

• To develop and apply machine learning models for automated analysis of occupational accident narratives.
• To convert unstructured textual data into structured, analyzable fields for improved safety insights.
• To enhance organizational capabilities in mitigating future workplace accidents through data-driven analysis.

Main Methods:

• Collected a large dataset of worker injury narratives from the U.S. Occupational Safety and Health Administration (OSHA).
• Utilized word embeddings-based text mining for advanced natural language processing of accident reports.
• Developed machine learning models to classify injury details across five dimensions: body part, injury source, event type, hospitalization, and amputation.

Main Results:

• Achieved excellent performance in analyzing accident narratives compared to previous methodologies.
• Successfully extracted and structured critical information from unstructured text data.
• Demonstrated the generalizability of the models through successful deployment on construction and mining industry datasets (transfer learning).

Conclusions:

• The developed machine learning approach significantly enhances the ability to rapidly analyze textual accident narratives.
• This methodology provides valuable, structured insights for improving occupational health and safety protocols.
• The findings support the broader application of AI in workplace safety management and accident prevention.