Text mining of practical disaster reports: Case study on Cascadia earthquake preparedness

Many practical disaster reports are published daily worldwide in various forms, including after-action reports, response plans, impact assessments, and resiliency plans. These reports serve as vital resources, allowing future generations to learn from past events and better mitigate and prepare for future disasters. However, this extensive practical literature often has limited impact on research and practice due to challenges in synthesizing and analyzing the reports. In this study, we 1) present a corpus of practical reports for text mining and 2) introduce an approach to extract insights from the corpus using select text mining tools. We validate the approach through a case study examining practical reports on the preparedness of the U.S. Pacific Northwest for a magnitude 9 Cascadia Subduction Zone earthquake, which has the potential to disrupt lifeline infrastructures for months. To explore opportunities and challenges associated with text mining of practical disaster reports, we conducted a brief survey of potential user groups. The case study illustrates the types of insights that our approach can extract from a corpus. Notably, it reveals potential differences in priorities between Washington and Oregon state-level emergency management, uncovers latent sentiments expressed within the reports, and identifies inconsistent vocabulary across the field. Survey results highlight that while simple tools may yield insights that are primarily interpretable by experienced professionals, more advanced tools utilizing large language models, such as Generative Pre-trained Transformer (GPT), offer more accessible insights, albeit with known risk associated with current artificial intelligence technologies. To ensure reproducibility, all supporting data and code are made publicly available (DOI: 10.17603/ds2-9s7w-9694).