Real-time monitoring of public affective risk perception during extreme rainstorm events: A deep learning framework with disaster knowledge-enhanced attention mechanism

Real-time monitoring of changes in public affective risk perception under extreme rainstorm events is crucial for improving disaster management effectiveness. To this end, this study develops a fine-grained sentiment analysis framework by integrating a disaster knowledge-enhanced attention mechanism with deep learning models. This framework enables the dynamic tracking of public affective risk perception during extreme rainstorm events by analyzing large-scale social media data. Specifically, we collected 157,372 Weibo posts related to rainstorm events and applied the feature selection method to identify key seed words that reflect public risk-related emotions. Dynamic word embedding techniques were then used to expand these seed words, leading to the creation of a specialized sentiment lexicon for the rainstorm disaster domain. Next, an emotion-word-based attention mechanism was incorporated into the training of the BERT-TextCNN model, enhancing its ability to recognize key emotional features. Compared to the optimal benchmark model, our approach improved accuracy by 6.1 %, precision by 5.0 %, recall by 5.5 %, and F1 score by 5.6 %. Furthermore, an empirical analysis of the 2023 extreme rainstorm disaster in Hebei Province validated the framework's efficacy. Findings reveal that our method can effectively achieve comprehensive situational monitoring and spatiotemporal evolution analysis of public affective risk perception during extreme rainstorm events. Additionally, it provides early warnings for severe rainstorm disasters by detecting anomalies in the affective perception index. Our recommendations provide actionable insights for disaster management agencies in formulating more effective emergency responses to extreme rainfall events. • Build a rainstorm-specific sentiment lexicon to capture public affective risk perception. • Develop an emotion-enhanced attention deep-learning model for dynamic monitoring of affective risk perception. • Derive extreme-rainstorm response strategies informed by the monitored dynamics of public affective risk perception.