Exploiting Twitter for Informativeness Classification in Disaster Situations

Disaster management urgently requires mechanisms for achieving situation awareness (SA) in a timely manner, allowing authorities to react in an appropriate way to reduce the impact on affected people and infrastructure. In such situations, no matter if they are human-induced like shootings or natural ones like earthquakes or floods, social media such as Twitter are frequently used communication channels, making them a highly valuable additional data source for enhancing SA. The challenge is, however, to identify out of the tremendous mass of irrelevant and non informative social media data those messages being really “informative”, i.e., contributing to SA in a certain disaster situation. Existing approaches on machine-learning driven informativeness classification most often focus on specific disaster types, such as shootings or floods, thus lacking general applicability and falling short in classification of new disaster events. Therefore, this article puts forward the following three contributions: First, in order to better understand the underlying social media data source, an in-depth analysis of existing Twitter data on 26 different disaster events is provided along temporal, spatial, linguistic, and source dimensions. Second, based thereupon, a cross-domain informativeness classifier is proposed being not focused on specific disaster types but rather allowing for classifications across different types. Third, the applicability of this cross-domain classifier is demonstrated, showing its accuracy compared to other disaster type specific approaches.