An LLM-Based Approach for Stakeholder Classification in Disaster Contexts

Twitter is a prominent platform for social media users to disseminate information and articulate concerns during disasters, amassing critical disaster resilience knowledge. To enhance stakeholder-centered research through social media data mining, it is crucial to develop a user classification method based on stakeholder sectors involved in disaster resilience (e.g., public agencies, private industries). This study examined the feasibility of leveraging a large language model (LLM), generative pre-trained transformer (GPT) in particular, to detect stakeholder sectors to which users belong. A three-step stakeholder classification (TSSC) prompting framework was proposed by adapting various prompting methods. The proposed TSSC outperformed seven commonly used machine learning algorithms, demonstrating its effectiveness in categorizing users. By classifying users solely based on usernames and descriptions, this method could potentially reduce substantial computational costs typically associated with processing extensive user data, facilitating more efficient social media data mining for human-centered research in disaster contexts.