Deep Active Learning Toward Crisis-related Tweets Classification
DescriptionWhen a crisis occurs people use social medias to share re- lated information. Social medias like Twitter can be considered as a platform for distributing online information such as asking help, reporting death or injured people, offering advise or donation [7, 8, 11, 5, 3, 4]. In addition, identify- ing the crisis-relevant information and classifying them into different informative categories can help the humanitarian response organizations to take quick and efficient actions. However, detecting and extracting useful information due to overwhelming amount of brief and informal messages during crisis situations are still challenging tasks. In this paper, first we develop Deep Neural Networks (DNNs) to perform a binary classification task. We evaluate our model on CrisisNLP dataset released by Imran el al  and the results of using Bi-LSTM on top of pretrained Word2vec word em- bedding shows the 91% accuracy. Even though DNN models yielded state-of-the-art performance on tweet classification task, it needs large amount of labeled data which is expensive and time consuming. To solve this issue, we apply an active learning approach to select the most informative tweets based on three different strategies. Combining active learn- ing with deep learning, we achieve acceptable classification accuracy using less annotated tweets.