Examining hurricane–related social media topics longitudinally and at scale: A transformer-based approach

Abstract

Instead of turning to emergency phone systems, social media platforms, such as Twitter, have emerged as alternative and sometimes preferred venues for members of the public in the US to communicate during hurricanes and other natural disasters. However, relevant posts are likely to be missed by responders given the volume of content on platforms. Previous work successfully identified relevant posts through machine-learned methods, but depended on human annotators. Our study indicates that a GPU-accelerated version of BERTopic, a transformer-based topic model, can be used without human training to successfully discern topics during multiple hurricanes. We use 1.7 million tweets from four US hurricanes over seven years and categorize identified topics as temporal constructs. Some of the more prominent topics related to disaster relief, user concerns, and weather conditions. Disaster managers can use our model, data, and constructs to be aware of the types of themes social media users are producing and consuming during hurricanes.