Text mining on social media for disaster management

Abstract

Disaster management has played an increasingly important role in recent years due to the disruptive disasters affecting people’s daily routines and orders of governmental operation. To minimize the huge economic losses and potential life losses caused by disasters, governors and researchers are by all means for dig effective means to predict, prevent and respond to disasters. Compared to traditional information report systems and experience-based decision-making procedures, disaster management tools ensuring immediacy and automation are demanded. As social media platforms become a pivotal channel for information dissemination, especially when communicating with each other offline becomes difficult during disasters, the large-scale data in real-time is of great value for platforms to quickly recognize the risk of disasters and respond accordingly. The thesis is composed of three studies that provide insights into disaster management with regard to social media by designing a framework to assess the risks of disasters through social media data and discovering contextualized content features that imply the rules of disaster information dissemination. In the first work, by applying the design science paradigm, large language models-based sentiment analysis is combined with the expertise risk assessment method to form a novel disaster risk assessment model that outperforms benchmark models. In the second work, uncertainty-related and self-regulation-related textual features are summarized and they are empirically identified to have negative effects on the sharing of disaster-related social media content. In the last work, the effects of different types of personal concerns during different disaster stages on social media content sharing are discussed. The results suggest that social media user sharing preference changes during a life-threatening disaster. Overall, these three works supplement the literature on text mining on social media for public safety events. The novel features discovered and the designed framework provides further implications for disaster management. The research has the potential to significantly enhance disaster management efforts by providing real-time situational awareness, utilizing content features in disaster assessment, informing response planning and directing disaster information sharing and public engagement.