Robust Domain Adaptation Approach for Tweet Classification for Crisis Response

Information posted by people on Twitter during crises can significantly improve crisis response towards reducing human and financial loss. Deep learning algorithms can identify related tweets to reduce information overloaded which prevents humanitarian organizations from using Twitter posts. However, they heavily rely on labeled data which is unavailable for emerging crises. And because each crisis has its own features such as location, occurring time and social media response, current models are known to suffer from generalizing to unseen disaster events when pretrained on past ones. To solve this problem, we propose a domain adaptation approach that makes use of a distant supervision-based framework to label the unlabeled data from emerging crises. Then, pseudo-labeled target data, along with labeled-data from similar past disasters, are used to build the target model. Our results show that our approach can be seen as a general robust method to classify unseen tweets from current events.