A Transformer-Based Analysis of Tweets in Germany to Investigate the Appearance and Evolution of the 2021 Eifel Flood in Social Media

In July 2021 several European countries were hit by severe floods. Estimates by SwissRe indicate that a flood event caused by the low-pressure area “Bernd” caused 227 deaths and economic losses of 41 billion USD in Central and Western Europe, with hotspots in Germany, Belgium, and the Netherlands. An increasing number of studies focus on understanding and modelling the causes and evolution of this event, developing reliable estimates of the losses it caused, and recommending improved disaster management strategies. However, risk communication and flood-related citizens’ behaviors, attitudes, and perceptions before, during, and after the flood are currently understudied.

Here, we develop an analytical framework to extract information on these human-related elements based on social media data. We ultimately aim to understand how flood warnings, intensity and impact are reflected in social media topics. To this extent, we analyze differences between topics arising on social media for an event like the 2021 flood compared to less devastating floods that occurred in the past. This requires homogeneous automatic assessment of Twitter data over time. We analyse the content of 42,000 tweets containing selected keywords related to flooding posted in Germany since 2014. Keywords refer to both fluvial and flash floods. Bidirectional Encoder Representations from Transformers (BERT) in combination with unsupervised clustering techniques are implemented to classify the tweets in different topic groups (BERTopic). Further, we extract the temporal evolution of topic patterns for different flood types and phases of flooding. Our analysis contributes to understanding the patterns of key topics, reflecting behaviors before, during and after the flooding event - thus how these topics change over time. Using the new framework and understanding these dynamics supports (i) modelling risk communication, behavioral drivers, and social interactions in relation to different types of floods with different intensities, and (ii) identifying indirect flood impacts that are not reported in traditional flood documentation. Finally, our approach can be extended for analysis of other natural hazards as well as compound events.