Text Mining of Practical Disaster Reports: Case Study on Cascadia Earthquake Preparedness

Many practical disaster reports are published every day across the globe in various forms (e.g., after-action reports, response plans, impact assessments, and resiliency plans). These reports can enable the next generations to learn from past events and efforts to best mitigate and prepare for future disasters. However, the extensive practical literature unfortunately has limited impacts on research and practice in general because of the challenge in synthesizing and analyzing many relevant documents. In this study, we present a method to 1) prepare a corpus of practical reports about a topic of interest for text mining and 2)
extract insights from the corpus using select text mining tools. We validated the method through a case study that examines practical reports about the preparedness of the U.S. Pacific Northwest for a possible Cascadia Subduction Zone earthquake of magnitude 9, which can disrupt lifeline infrastructures for months. The case study illustrated the types of insights that the methodology can help extract from a corpus. For example, it identified potential differences in priorities between entities (Washington vs. Oregon state-level emergency management), highlighted latent sentiments expressed in the corpus, and recognized the inconsistent vocabulary across the field. We also discuss opportunities and challenges with text mining of practical disaster reports. For example, simple techniques can provide insights potentially only interpretable by those experienced in the field, whereas more complex techniques using large language models such as ChatGPT can provide more readily
accessible insights with known risks of artificial intelligence, such as made-up facts and misrepresentation.