Hierarchical Discriminative Classification for Text-Based Geolocation

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 336–348, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Hierarchical Discriminative Classification for Text-Based Geolocation Benjamin Wing† Jason Baldridge† †Department of Linguistics, University of Texas at Austin ben@benwing.com, jbaldrid@utexas.edu Abstract Text-based document geolocation is com- monly rooted in language-based infor- mation retrieval techniques over geodesic grids. These methods ignore the natural hierarchy of cells in such grids and fall afoul of independence assumptions. We demonstrate the effectiveness of using lo- gistic regression models on a hierarchy of nodes in the grid, which improves upon the state of the art accuracy by several percent and reduces mean error distances by hundreds of kilometers on data from Twitter, Wikipedia, and Flickr. We also show that logistic regression performs fea- ture selection effectively, assigning high weights to geocentric terms. 1 Introduction Document geolocation is the identification of the location—a specific latitude and longitude—that forms the primary focus of a given document. This assumes that a document can be adequately associ- ated with a single location, which is only valid for certain documents, generally of fairly small size. Nonetheless, there are many natural situations in which such collections arise. For example, a great number of articles in Wikipedia have been man- ually geotagged; this allows those articles to ap- pear in their geographic locations in geobrowsers like Google Earth. Images in social networks such as Flickr may be geotagged by a camera and their textual tags can be treated as documents. Like- wise, tweets in Twitter are often geotagged; in this case, it is possible to view either an individual tweet or the collection of tweets for a given user as a document, respectively identifying the loca- tion as the place from which the tweet was sent or the home location of the user. Early work on document geolocation used heuristic algorithms, predicting locations based on toponyms in the text (named locations, determined with the aid of a gazetteer) (Ding et al., 2000; Smith and Crane, 2001). More recently, vari- ous researchers have used topic models for doc- ument geolocation (Ahmed et al., 2013; Hong et al., 2012; Eisenstein et al., 2011; Eisenstein et al., 2010) or other types of geographic document summarization (Mehrotra et al., 2013; Adams and Janowicz, 2012; Hao et al., 2010). A number of researchers have used metadata of various sorts for document or user geolocation, including doc- ument links and social network connections. This research has sometimes been applied to Wikipedia (Overell, 2009) or Facebook (Backstrom et al., 2010) but more commonly to Twitter, focusing variously on friends and followers (McGee et al., 2013; Sadilek et al., 2012), time zone (Mahmud et al., 2012), declared location (Hecht et al., 2011), or a combination of these (Schulz et al., 2013). We tackle document geolocation using super- vised methods based on the textual content of documents, ignoring their metadata. Metadata- based approaches can achieve great accuracy (e.g. Schulz et al. (2013) obtain 79% accuracy within …

Content truncated to 500 words