Abstract: Much has been said about the value of social media messages for emergency services. The new uses related to these platforms bring users to share information, otherwise unknown in crisis events. Thus, many studies have been performed in order to identify tweets relating to a crisis event or to classify these tweets according to certain categories. However, determining the relevant information contained in the messages collected remains the responsibility of the emergency services. In this article, we introduce the issue of classifying the information contained in the messages. To do so, we use classes such as those used by the operators in the call centers. Particularly we show that this problem is related to named entities recognition on tweets. We then explain that a semi-supervised approach might be beneficial, as the volume of data to perform this task is low. In a second part, we present some of the challenges raised by this problematic and different ways to answer it. Finally, we explore one of them and its possible outcomes.