Broadly Improving User Classification via Communication-Based Name and Location Clustering on Twitter

Proceedings of NAACL-HLT 2013, pages 1010–1019, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Broadly Improving User Classification via Communication-Based Name and Location Clustering on Twitter Shane Bergsma, Mark Dredze, Benjamin Van Durme, Theresa Wilson, David Yarowsky Department of Computer Science and Human Language Technology Center of Excellence Johns Hopkins University Baltimore, MD 21218, USA shane.a.bergsma@gmail.com, mdredze@cs.jhu.edu, vandurme@cs.jhu.edu, taw@jhu.edu, yarowsky@cs.jhu.edu Abstract Hidden properties of social media users, such as their ethnicity, gender, and location, are of- ten reflected in their observed attributes, such as their first and last names. Furthermore, users who communicate with each other of- ten have similar hidden properties. We pro- pose an algorithm that exploits these insights to cluster the observed attributes of hundreds of millions of Twitter users. Attributes such as user names are grouped together if users with those names communicate with other similar users. We separately cluster millions of unique first names, last names, and user- provided locations. The efficacy of these clus- ters is then evaluated on a diverse set of clas- sification tasks that predict hidden users prop- erties such as ethnicity, geographic location, gender, language, and race, using only pro- file names and locations when appropriate. Our readily-replicable approach and publicly- released clusters are shown to be remarkably effective and versatile, substantially outper- forming state-of-the-art approaches and hu- man accuracy on each of the tasks studied. 1 Introduction There is growing interest in automatically classify- ing users in social media by various hidden prop- erties, such as their gender, location, and language (e.g. Rao et al. (2010), Cheng et al. (2010), Bergsma et al. (2012)). Predicting these and other proper- ties for users can enable better advertising and per- sonalization, as well as a finer-grained analysis of user opinions (O’Connor et al., 2010), health (Paul and Dredze, 2011), and sociolinguistic phenomena (Eisenstein et al., 2011). Classifiers for user prop- erties often rely on information from a user’s social network (Jernigan and Mistree, 2009; Sadilek et al., 2012) or the textual content they generate (Pennac- chiotti and Popescu, 2011; Burger et al., 2011). Here, we propose and evaluate classifiers that bet- ter exploit the attributes that users explicitly provide in their user profiles, such as names (e.g., first names like Mary, last names like Smith) and locations (e.g., Brasil). Such attributes have previously been used as “profile features” in supervised user classifiers (Pen- nacchiotti and Popescu, 2011; Burger et al., 2011; Bergsma et al., 2012). There are several motivations for exploiting these data. Often the only informa- tion available for a user is a name or location (e.g. for a new user account). Profiles also provide an orthogonal or complementary source of information to a user’s social network and textual content; gains based on profiles alone should therefore add to gains based on other data. The decisions of profile-based classifiers could also be used to bootstrap training data for other classifiers that use complementary fea- tures. Prior work has encoded profile attributes via lex- ical or character-based features (e.g. …

Content truncated to 500 words