Detecting Disaster Related Tweets Using Hybrid Deep Neural Network Models

Abstract

People use various social media platforms like twitter to report the disaster and disaster related information during crisis situations such as natural and man-made disasters. Detecting these disaster events can improve situational awareness for general public, response agencies and aid agencies. Many works has done for disaster event detection which uses traditional machine learning techniques but in recent years deep neural network models have demonstrated better results for many problems over traditional machine learning models. It is advantageous of using deep learning models because the model has the capacity to capture more than one layers of information. Convolutional Neural Network (CNN), Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) are three such deep neural networks. CNN can recognize local features in a multidimensional field and LSTM and GRU network can learn sequential data as it has the ability of remembering previously read data. In this paper, we use two hybrid deep neural network models named CNN-LSTM and CNN-GRU, which is developed by combining both CNN, LSTM, and CNN, GRU networks for doing event detection during disaster situation on Twitter data. We provide the detailed explanation of both models and perform comparison against Support Vector Machine (SVM), CNN, and LSTM models. Our result shows that both models can successfully identify the presence of disaster related events accurately from twitter data. We found that CNN-LSTM model is the best model achieving significant improvement of 18.74% than regular SVM model.