A study on extracting disaster information from tweets

Abstract

In Japan, where natural disasters occurs frequently, obtaining and delivering accurate information promptly when a disaster occurs is essential to minimize damage. Information from traditional mass media contain a number of general information unrelated to disaster, so there are limitations in delivering necessary information to the resident in affected area. On the other hand, Twitter, one of the popular social media, is expected to play an important role during disaster because of its simplicity, promptness and wide propagation. However, because of its huge size of users, there are too many tweets which hinders timely extraction of relevant information. Disaster information is also useful for business travellers and tourists. They are less informed about the area and the challenge is to provide them with accurate information promptly. Our study proposes to establish a system to assist real time understanding of disaster by extracting relevant information efficiently from messages tweeted during two typhoons. First, binary classification is applied to classify and extract disaster tweets from tweets group. By using BNS method, the improvement in accuracy is confirmed. Then clustering is applied to the disaster tweets. The tweets are classified by 15 clusters generated. The result yields F measure of 0.59.