Extracting Humanitarian Information from Tweets

In this paper we describe the application of our methods to humanitarian information extraction from tweets and their performance in the scope of the SMERP 2017 Data Challenge task. Detecting and extracting the (scarce) relevant information from tweet collections as precisely, completely, and rapidly as possible is of the utmost importance during natural disasters and other emergency events. We applied a machine learning and a linguistically motivated approach. Both are designed to satisfy the information needs of an expert by allowing experts to define and find the target information. We found that the performance of this effort highly depends on the task definition and the ability to facilitate the feedback iteratively. The results of the current data challenge task demonstrate that it is realistic to expect a balanced performance across multiple metrics even under poor conditions.