Student Research Abstract: Unsupervised Key Term Extraction of Tornado Narratives from NOAA Storm Events Database

Disaster records are often composed of key metrics of the disaster event and its monetary cost. While the cost of the event is helpful for insurance adjustments, the monetary impact of a disaster is not conducive for disaster preparedness planners to build community resilience. Often, there are natural language narratives about the disaster event, like in the National Oceanic and Atmospheric Administration (NOAA) Storm Events Database, that contain essential information but are not easily retrievable because of their unstructured nature. These narratives need to be text mined in order to retrieve the impacts of the disasters in order to structure the data for further use. The method proposed in this abstract is a critical first step in the process. It is an unsupervised key term extraction method using sentence transformers to create embeddings that are then clustered, and assigned key terms by utilizing the highest term frequency-inverse document frequency (tf-idf) scores for the sentences in the narratives.