Automatic Labeling of Tweets for Crisis Response Using Distant Supervision

Current tweet classification models aimed at enhancing crisis response are based on supervised deep learning. They rely on the quality and quantity of human-labeled training data. Still, the available training data is small in size and imbalanced in coverage of crisis types, which prevents the models from generalization, and as it is manually labeled, it is also expensive to produce. To overcome these problems, distant supervision can be applied to automatically generate large-scale labeled data for tweet classification for crisis response. Experimental results on different crisis events show that our work can produce good quality labeled data from past and recent events. Substituting automatically labeled training data for part of the manually labeled training data has a minimal impact on the model performance, indicating that automatically labeled data can be used when no hand-labeled data is available.