Overview
Injury records usually include a short free-text description of what happened, and these narratives are rarely analyzed at scale. In this proceedings paper with David Passmore, Yulia Kustikova, Rose Baker, and Jeong-Ha Yim, we explored whether unsupervised machine learning can summarize such narratives in a way that is useful for assessing risk.
The data were narrative reports of 77,215 injuries that occurred in coal mines in the United States between 2000 and 2015. We estimated a topic model with Latent Dirichlet Allocation and then related topic emphasis to characteristics of the injuries.
Key points
- The modeling process identified six topics in the free-text reports.
- One topic described mainly strains and sprains of the musculoskeletal system.
- Emphasis on that topic differed by where on the mine property the injury occurred, by the degree of injury, and by year.
- Narratives close to this topic referred more often to surface or other locations than to underground locations.
- The exploratory result suggests that further topic mining of injury narratives is worth pursuing.
Event
IPICSE-2018, published in MATEC Web of Conferences, volume 251 (2018).
Citation
Passmore, D., Chae, C., Kustikova, Y., Baker, R., & Yim, J.-H. (2018). An exploration of text mining of narrative reports of injury incidents to assess risk. MATEC Web of Conferences, 251, Article 06020. https://doi.org/10.1051/matecconf/201825106020