Social networks like Twitter are good means for people to express themselves and ask for help in times of crisis. However, to provide help, authorities need to identify informative posts on the network from the vast amount of non-informative ones to better know what is actually happening. Traditional methods for identifying informative posts put emphasis on the presence or absence of certain words which has limitations for classifying these posts. In contrast, in this paper, we propose to consider the (overall) distribution of words in the post. To do this, based on the distributional hypothesis in linguistics, we assume that each tweet is a distribution from which we have drawn a sample of words. Building on recent developments in learning methods, namely learning on distributions, we propose an approach which identifies informative tweets by using distributional assumption. Extensive experiments have been performed on Twitter data from more than 20 crisis incidents of nearly all types of incidents. These experiments show the superiority of the proposed approach in a number of real crisis incidents. This implies that better modelling of the content of a tweet based on recent advances in estimating distributions and using domain-specific knowledge for various types of crisis incidents such as floods or earthquakes, may help to achieve higher accuracy in the task.

Keywords