Early detection of emergency events from social media: A new text clustering approach

Abstract Emergency events need early detection, quick response, and accuracy recover. In the era of big data, social media users can be seen as social sensors to monitor real time emergency events. This paper proposed an integrated approach to early detect all the four kinds of emergency events including natural disasters, man-made accidents, public health events and social security events. First, the BERT-Att-BiLSTM model is used to detect emergency related posts from the massive and irrelevant data. Then, the 3W attribute information (What, Where and When) of the emergency event is extracted. With the 3W attribute information, we create an unsupervised dynamical event clustering algorithm based on text-similarity and combine it with the supervised logistical regression model to cluster posts into different events. The experiments on Sina Weibo data demonstrate the superiority of the proposed framework. Case studies on some real emergency events show the proposed framework has good performance and high timeliness. Practical applications of the framework have also been discussed, following by some future directions for improvement.