Natural disasters often trigger a surge of social media posts, unfortunately including misinformation that can disrupt response efforts, necessitating rapid fact-checking. To facilitate fact-checking, accurately identifying the specific disaster type mentioned in a social media post is crucial for effective evidence collection. However, a comprehensive multimodal social media dataset for disaster-type identification in fact-checking has been lacking. We address this gap by proposing the task of disaster-type classification from multimodal social media posts and introducing DTCD-AFC (Disaster-Type Classification Dataset designed for Automated Fact-Checking). This dataset, derived from CrisisMMD, features annotations for seven disaster types based on post claims, enabling evaluation of classification performance in automated fact-checking systems and supporting social media content analysis during disasters. Using GPT-4o as a zero-shot baseline, we found that evaluations grounded in DTCD-AFC labels better reflect post content than evaluations using event-derived CrisisMMD labels. Furthermore, we demonstrated that DTCD-AFC encompasses challenging tasks, such as identifying closely related disaster types (e.g., floods, hurricanes, and landslides), as well as tasks in which keywords within text exert significant influence.

PDF