Explainable Social Sensing of Public Sentiment During the 2025 Pakistan Floods: Implications for Disaster Governance and Community Resilience

Flood disasters disrupt infrastructure and shape public perceptions of warnings and emergency institutions. This study examined 26,010 posts from X during the 2025 Pakistan floods using BERTopic, transformer-based sentiment classification, observational event analysis, SHapley Additive exPlanations, exploratory forecasting, and a responsible-AI checklist. Five discussion domains concerned relief operations, politics, media and warnings, government response, and transboundary water politics and infrastructure failure. Transboundary-water discussion had the lowest mean sentiment (−0.2094), whereas government-response discussion was more positive (0.1802). Adjusted event-associated contrasts were −0.184 after a reported dam-release event and +0.118 after a prime-ministerial visit, conditional on observational assumptions. The classifier agreed with human consensus for 169 of 200 audited posts (84.5%); language-specific reliability remains unresolved. Model explanations associated selected institutional and infrastructure terms with sentiment predictions, without establishing communication effects. Forecasting within one flood episode remains exploratory because improvement over persistence and independent-event performance have not been established. The contribution is an applied integration of established methods that separates descriptive discourse, observational associations, model explanations, and governance requirements. These outputs may inform human-reviewed disaster communication, but do not directly measure institutional trust or community resilience.