Exploiting Text and Network Context for Geolocation of Social Media Users

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1362–1367, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Exploiting Text and Network Context for Geolocation of Social Media Users Afshin Rahimi,1 Duy Vu,2 Trevor Cohn,1 and Timothy Baldwin1 1Department of Computing and Information Systems 2Department of Mathematics and Statistics The University of Melbourne arahimi@student.unimelb.edu.au {duy.vu, t.cohn}@unimelb.edu.au tb@ldwin.net Abstract Research on automatically geolocating social media users has conventionally been based on the text content of posts from a given user or the social network of the user, with very lit- tle crossover between the two, and no bench- marking of the two approaches over compara- ble datasets. We bring the two threads of re- search together in first proposing a text-based method based on adaptive grids, followed by a hybrid network- and text-based method. Eval- uating over three Twitter datasets, we show that the empirical difference between text- and network-based methods is not great, and that hybridisation of the two is superior to the component methods, especially in contexts where the user graph is not well connected. We achieve state-of-the-art results on all three datasets. 1 Introduction There has recently been a spike in interest in the task of inferring the location of users of social me- dia services, due to its utility in applications in- cluding location-aware information retrieval (Ami- tay et al., 2004), recommender systems (Noulas et al., 2012) and rapid disaster response (Earle et al., 2010). Social media sites such as Twitter and Face- book provide two primary means for users to de- clare their location: (1) through text-based metadata fields in the user’s profile; and (2) through GPS- based geotagging of posts and check-ins. However, the text-based metadata is often missing, mislead- ing or imprecise, and only a tiny proportion of users geotag their posts (Cheng et al., 2010). Given the small number of users with reliable location infor- mation, there has been significant interest in the task of automatically geolocating (predicting lat/long co- ordinates) of users based on their publicly avail- able posts, metadata and social network information. These approaches are built on the premise that a user’s location is evident from their posts, or through location homophily in their social network. Our contributions in this paper are: a) the demon- stration that network-based methods are generally superior to text-based user geolocation methods due to their robustness; b) the proposal of a hybrid clas- sification method that backs-off from network- to text-based predictions for disconnected users, which we show to achieve state-of-the-art accuracy over all Twitter datasets we experiment with; and c) empir- ical evidence to suggest that text-based geolocation methods are largely competitive with network-based methods. 2 Related Work Past work on user geolocation falls broadly into two categories: text-based and network-based methods. Common to both methods is the manner of fram- ing the geolocation prediction problem. Geographic coordinates are real-valued, and accordingly this is most naturally modelled as (multiple) regression. However for …

Content truncated to 500 words