Social media can provide insights into natural hazard events and people's emergency responses. In this study, we present a natural language processing analytic framework to extract and categorize information from 43 287 textual Twitter (X) posts in German since 2014. We implement bidirectional encoder representations from transformers in combination with unsupervised clustering techniques (BERTopic) to automatically extract social media content, addressing transferability issues that arise from commonly used bag-of-words representations. We analyze the temporal evolution of topic patterns, reflecting behaviors and perceptions of citizens before, during, and after flood events. Topics related to low-impact riverine flooding contain descriptive hazard-related content, while the focus shifts to catastrophic impacts and responsibilities during high-impact events. Our analytical framework enables the analysis of temporal dynamics of citizens’ behaviors and perceptions, which can facilitate lessons-learned analyses and improve risk communication and management.