Topic Modeling Based Clustering of Disaster Tweets Using BERTopic
Julanta Leela Rachel J, A. Bhuvaneswari, M Kumudha
Abstract
The essence of catastrophe management is in the capacity to promptly and effectively addressing unforeseen circumstances, as well as to implement premeditated and flexible strategies in accordance with established protocols and norms. Twitter has become a prominent means of disaster communication, particularly in nations where it has a substantial presence. Twitter can provide valuable information regarding the effects of infrastructure on emergency management in the event of significant disasters. However, the process of sifting through numerous unrelated tweets can be laborious and time-consuming. Prior research has successfully recognized the various categories of communications present on social media platforms during times of disaster. However, there is a dearth of recommended methodologies aimed at effectively extracting valuable information from these messages. In this study, we propose a framework that may be efficiently utilized to promptly deliver disaster effect information derived from social media sources. The methodology involves the application of certain techniques to process the unprocessed Twitter data by employing keyword filtering, location-based filtering, and analysis of tweet properties. Subsequently, Bidirectional Encoder Representations from Transformers (BERT) technique is utilized to categorize the tweets originating from disaster-affected regions into distinct subjects that are relevant and valuable for emergency management. BERT analysis yielded a total of 376 themes within the dataset. Among these topics, a significant proportion of 196 were found to be closely associated with the impacts of disasters. These impacts encompassed many aspects such as outages, closures, flooded roadways, and damaged infrastructure.