A Data-Driven Multimodal Mining Framework for Emergency Information: Quantitative Visual Feature and Satisfaction Modeling

The losses caused by frequent natural disasters are increasing day by day, and the short-video platform has become the core digital space for the public to pay attention to disasters and express their demands. In the face of massive multimodal data, how to automatically extract features and quantify their impact on public behavior is a key technical challenge facing information systems and computational social sciences. To address this issue, this study proposes an automated multimodal data mining and modeling framework that integrates YOLOv11-based computer vision with BERT-based natural language processing for disaster short-video analysis. Real-world short-video and interaction data were automatically collected using web crawling. YOLOv11 was employed to identify and quantify two types of visual information—relief information and suffering information—while BERT was used to extract a text-based rescue satisfaction index. Then, the partial least squares structural equation model was used to explore the driving mechanism of information characteristics on public engagement. It was found that the content of relief information in videos has a significant positive impact on satisfaction but a significant negative impact on engagement. The content of suffering information has a significant negative impact on satisfaction but a significant positive impact on engagement. In addition, satisfaction has a significant negative impact on engagement and plays an intermediary role between the two types of information content and engagement. By integrating YOLOv11 and BERT into a unified short-video analytics framework, this study extends automated multimodal disaster information analysis and provides practical support for optimizing emergency communication and disaster information-release strategies.