Public Opinion Stage Segmentation in Disaster Events: A Study Based on Multimodal Sentiment Prediction Model

Existing approaches for predicting public sentiment and analyzing opinion evolution during disaster events using multimodal data (text, video, audio) suffer from several limitations: an inadequate dynamic fusion of heterogeneous multi-source data, and imprecise division of public opinion stages. To address these issues, this paper proposes an Enhanced Disentangled Cross Fusion (EDCF) model-based framework for analyzing the evolution of public opinion in disaster events. This framework integrates the E-Divisive with Medians (EDM) change point detection method with spatiotemporal sequence modeling techniques to achieve fine-grained stage segmentation. The EDCF model employs Transformers and positional encoding to process time-series signals (audio/video), effectively capturing long-range dependencies. It enhances modality-specific representation capabilities by introducing dedicated encoders for each modality, a shared encoder, and a reconstruction decoder for disentangled representation learning. Furthermore, the model utilizes a cross-modal language-guided attention mechanism for efficient and effective feature fusion. Experimental validation on the publicly available multimodal sentiment dataset CMU-MOSI demonstrates that the proposed EDCF framework significantly outperforms baseline methods on key sentiment prediction metrics.

@article{Yuan2025,
  title     = {Public Opinion Stage Segmentation in Disaster Events: A Study Based on Multimodal Sentiment Prediction Model},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {8},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Xiaogang Yuan and Jiaxi Chen and Dezhi An and Xiang Gong},
  doi       = {10.14569/IJACSA.2025.0160876},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160876}
}