Location Mention Recognition from Japanese Disaster-Related Tweets

Abstract

In order to minimize the damage inflicted by large-scale disasters, it is essential to collect and disseminate information quickly and accurately. In recent years, various national agencies and local municipalities have used Twitter and other highly immediate social media to help focus their disaster relief efforts. Because the volume of information circulating on social media increases rapidly during a disaster, the ability to quickly sort out valuable posts from the massive volume of posts that appear is essential. In the case of Twitter, it is vital for early responders to identify the location of relevant tweets in order to facilitate decision making and focus their response. To help in this task, attempts have been made to use machine learning to classify genres, extract useful information, and identify locations and points of interest for groups of tweets posted during a disaster. However, since preparing training data and building a model during the early stages of a disaster are extremely challenging, using a model built on past disaster tweet data offers a promising possibility. In this study, we focus on three heavy rain disasters that occurred in Japan and examine the extraction of the location mentions in tweets using models learned from tweets posted during prior disasters.