Dynamic Associations Between Public Risk Perception and Behavioral Responses on Weibo During the 2023 Beijing–Tianjin–Hebei Extreme Rainfall Event

Social media has become an important platform for the public to obtain information, express emotions, seek help, and participate in collaborative responses during major natural disasters. However, existing studies have mainly focused on the intensity of disaster-related discussions, emotional evolution, or patterns of information dissemination, while systematic examination remains limited regarding how public risk perception is associated with specific behavioral responses and whether these associations vary across disaster stages, degrees of regional impact, and user types. Using the stimulus–organism–response (SOR) framework as an organizing heuristic, this study examines the 2023 Beijing–Tianjin–Hebei extreme rainfall event. A total of 31,102 Sina Weibo posts were collected. By integrating LDA topic modeling, risk-perception dictionary matching, a BERT-based semantic robustness check, and chi-square tests, this study analyzes descriptive associations between public risk perception and behavioral responses. The results show that Weibo discussions evolved across disaster stages, with topics shifting from meteorological warnings and disaster reports to mutual aid, rescue operations, material support, and post-disaster reflection. Public risk perception shifted from information-oriented to relationship-oriented cognition, while behavioral responses transitioned from information attention to emergency mutual aid, sustained support, and reflective expression. Further analysis indicates that uncertainty perception was mainly associated with information attention; susceptibility perception co-occurred with help-seeking and emotional support; positive trust perception was strongly associated with emotional support; and negative trust perception co-occurred with questioning, accountability, and reflective suggestions. Because the negative-trust classifier performed substantially less well than the other dimensions, findings involving negative trust are treated as exploratory. Heterogeneity analyses show that observed perception–behavior associations varied across disaster stage, regional impact, and user type. These findings deepen understanding of social-media discourse during disasters while supporting cautious, stage-, region-, and group-sensitive risk communication.