A Sociotechnical System Approach to Multimodal Emotion Propagation in Disaster Governance: Lessons from the 2021 Zhengzhou Flood

Abstract

Short-video platforms provide multimodal records of public affect during natural disasters and offer new opportunities for crisis-oriented social-media analysis. Using the 2021 Zhengzhou Flood as a single-event case study, we construct DMAC, which, to the best of our knowledge, is among the first Chinese-language real-world multimodal short-video datasets focused on a disaster event. The curated dataset contains 284 videos and 144,174 comments. We distinguish Video-Embedded Emotion (Evideo), sentiment expressed in collected comments (Eprimary), and a transparently constructed Composite Emotion Index (Ecomposite). We adopt and fine-tune the existing Multimodal End-to-End Sparse Model (MESM) and implement a governance-oriented proof-of-concept interface for timely analyst review. Under this challenging six-class setting, the models yield average accuracies of 57.18%, 63.42%, and 60.34% for Evideo, Eprimary, and Ecomposite, respectively, and the multimodal ablations show complementary value from text, audio, and visual signals. Cross-tier distributions exhibit an asymmetric, amplification-like concentration of negative affect in this case. These findings provide case-derived insights and a methodological reference for comparable Chinese short-video disaster contexts, while the system is intended to support rather than replace human interpretation and decision-making.