How Can Natural Language Processing Support Emergency Management? NLP for Classification of Tweets During Crisis-Events

Social media has become an increasingly important source of information for emergency services during disasters. In this project, we analyze and compare the effectiveness of three state of the art deep learning models for detecting informativeness of disaster-related tweets in real time. We used the CrisisLexT26 dataset which comprises of 250,000 tweets regarding 12 disaster types from a total of 26 different crisis-events which occurred in 2012 and 2013. Our findings have shown that generalized detection models work better when being trained on all type of crisis compared to the specialized models. The best architecture (XLNet) achieved a AUC between 0.86 and 0.94, showcasing its potential as a useful tool for future emergency response efforts.