Predicting Locations in Tweets

Abstract

Five hundred millions of tweets are posted daily, making Twitter a major social media from which topical information on events can be extracted. Events are represented by time, location and entity-related information. This paper focuses on location which is an important clue for both users and geo-spatial applications. We address the problem of predicting whether a tweet contains a location or not, as location prediction is a useful pre-processing step for location extraction, by defining a number of features to represent tweets and conducting intensive evaluation of machine learning parameters. We found that: (1) not only words appearing in a geography gazetteer are important but the occurrence of a preposition right before a proper noun also is. (2) it is possible to improve precision on location extraction if the occurrence of a location is predicted.