Inferring User Political Preferences from Streaming Communications

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 186–196, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Inferring User Political Preferences from Streaming Communications Svitlana Volkova,1 Glen Coppersmith2 and Benjamin Van Durme1,2 1Center for Language and Speech Processing, 2Human Language Technology Center of Excellence, Johns Hopkins University, Baltimore, MD 21218 svitlana@jhu.edu, coppersmith@jhu.edu, vandurme@cs.jhu.edu Abstract Existing models for social media per- sonal analytics assume access to thou- sands of messages per user, even though most users author content only sporadi- cally over time. Given this sparsity, we: (i) leverage content from the local neigh- borhood of a user; (ii) evaluate batch mod- els as a function of size and the amount of messages in various types of neighbor- hoods; and (iii) estimate the amount of time and tweets required for a dynamic model to predict user preferences. We show that even when limited or no self- authored data is available, language from friend, retweet and user mention commu- nications provide sufficient evidence for prediction. When updating models over time based on Twitter, we find that polit- ical preference can be often be predicted using roughly 100 tweets, depending on the context of user selection, where this could mean hours, or weeks, based on the author’s tweeting frequency. 1 Introduction Inferring latent user attributes such as gender, age, and political preferences (Rao et al., 2011; Za- mal et al., 2012; Cohen and Ruths, 2013) auto- matically from personal communications and so- cial media including emails, blog posts or public discussions has become increasingly popular with the web getting more social and volume of data available. Resources like Twitter1 or Facebook2 become extremely valuable for studying the un- derlying properties of such informal communica- tions because of its volume, dynamic nature, and diverse population (Lunden, 2012; Smith, 2013). 1http://www.demographicspro.com/ 2http://www.wolframalpha.com/facebook/ The existing batch models for predicting latent user attributes rely on thousands of tweets per author (Rao et al., 2010; Conover et al., 2011; Pennacchiotti and Popescu, 2011a; Burger et al., 2011; Zamal et al., 2012; Nguyen et al., 2013). However, most Twitter users are less prolific than those examined in these works, and thus do not produce the thousands of tweets required to obtain their levels of accuracy e.g., the median number of tweets produced by a random Twitter user per day is 10. Moreover, recent changes to Twitter API querying rates further restrict the speed of access to this resource, effectively reducing the amount of data that can be collected in a given time period. In this paper we analyze and go beyond static models formulating personal analytics in social media as a streaming task. We first evaluate batch models that are cognizant of low-resource predic- tion setting described above, maximizing the effi- ciency of content in calculating personal analytics. To the best of our knowledge, this is the first work that makes explicit the tradeoff between accuracy and cost (manifest as calls to the Twitter API), and optimizes to a different tradeoff …

Content truncated to 500 words