Tweets Classification with BERT in the Field of Disaster Management

Crisis informatics focus on the contribution of user generated content (UGC) to disaster management. To leverage the social media data effectively, it is crucial to filter out noisy information from the large volume of data flow so that we could better estimate disaster damage with these data. Not satisfied with basic keyword-based filtration, many researchers turn to machine learning for solution. In this project, I apply deep learning techniques to address Tweets classification problem in disaster management field. The labels of Tweets reflect different types of disaster-related information, which have different potential usage in emergency response. In particular, BERT is used for transfer learning. The standard BERT architecture for classification and several other customized BERT architectures are trained to compare with the baseline bidirectional LSTM with pretrained Glove Twitter embeddings. Results show that BERT and BERT-based LSTM attain the best results, outperforming the baseline model by 3.29% on average in terms of F-1 score respectively. Ambiguity and subjectivity affect the performance of these models considerably. In some examples the models can surpass human performance.