Mining help intent on twitter during disasters via transfer learning with sparse coding

Abstract

Citizens share a variety of information on social media during disasters, including messages with the intentional behavior of seeking or offering help. Timely identification of such help intent can operationally benefit disaster management by aiding the information collection and filtering for response planning. Prior research on intent identification has developed supervised learning methods specific to a disaster using labeled messages from that disaster. However, rapidly acquiring a large set of labeled messages is difficult during a new disaster in order to train a supervised learning classifier. In this paper, we propose a novel transfer learning method for help intent identification on Twitter during a new disaster. This method efficiently transfers the knowledge of intent behavior from the labeled messages of the past disasters using novel Sparse Coding feature representation. Our experiments using Twitter data from four disaster events show the performance gain up to 15% in both F-score and accuracy over the baseline of popular Bag-of-Words representation. The results demonstrate the applicability of our method to assist realtime help intent identification in future disasters.