Integrated Content-Graph Analysis to Characterize Social Media Conversations During Disaster Evacuations

Social media content posted by users during an evacuation phase of a disaster can provide a valuable source of information for situational awareness of emergency responders for decision making. However, understanding different aspects of the situation from unstructured data of social media and how they can be categorized to help emergency responders are the challenges that must be addressed. In this work, we design and conduct an integrated analytical framework using content and graph analysis methods to address these challenges by first detecting communities on the graph of user-entity interactions, followed by understanding different aspects of the users’ interests by topic modeling. Experimenting on a dataset of Twitter/X posts from hurricane Irma in 2017, our results reveal that the users mainly discussed about different aspects of the evacuation situation including locations, time of action, empathy and mental health issues, family constraints, pets and sheltering, event descriptors, transportation and supplies, and accommodation. Additionally, our experiments show the significant alignment between topic representation based on, simple yet effective, frequent set of entities and topic prediction by BERTopic as a popular topic modeling method.