Cross-domain informativeness classification for disaster situations

Social Media services gain increasing importance as a new data source for achieving Situation Awareness in disaster management. One crucial prerequisite is to automatically filter social media messages towards informativeness commonly realized by supervised machine learning. Since disaster situations are different, most classification approaches focus on informativeness classification of similar disasters. Thus their use is limited to particular disaster types, for instance earthquakes or floods, lacking general applicability. At the same time, how to get accurate informativeness classification for new disaster events is not yet totally understood due to variations in training data, features, classification algorithms and their settings. To address these issues, our contribution is threefold: First, a systematic and in-depth analysis of an existing twitter crisis data set is provided along four different dimensions in order to gain a comprehensive understanding of those characteristics indicating informative Tweets in disaster situations. On basis of these insights, a cross domain classifier is engineered, which is applicable not only across different disaster events but also across disaster events of different types. Finally, systematic classification experiments are conducted, demonstrating that our classification approach is more accurate than other disaster type specific ones.