Detecting reports of mass emergency events on Twitter

The paper addresses the problem of detecting eyewitness reports of mass emergencies on Twitter. This is the first work to conduct a large-scale comparative evaluation of classification features extracted from Twitter posts, using different learning algorithms and datasets representing a broad range of mass emergencies including both natural and technological disasters. We investigate the relative importance of different feature types as well as on the effect of several feature selection methods applied to this problem. Because the task of detecting mass emergencies is characterized by high heterogeneity of the data, our primary focus is on identifying those features that are capable of separating mass emergency reports from other messages, irrespective of the type of the disaster.