Weakly Supervised User Profile Extraction from Twitter

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 165–174, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Weakly Supervised User Profile Extraction from Twitter Jiwei Li1, Alan Ritter2, Eduard Hovy1 1Language Technology Institute, 2Machine Learning Department Carnegie Mellon University, Pittsburgh, PA 15213, USA bdlijiwei@gmail.com, rittera@cs.cmu.edu, ehovy@andrew.cmu.edu Abstract While user attribute extraction on social media has received considerable attention, existing approaches, mostly supervised, encounter great difficulty in obtaining gold standard data and are therefore limited to predicting unary predicates (e.g., gen- der). In this paper, we present a weakly- supervised approach to user profile extrac- tion from Twitter. Users’ profiles from so- cial media websites such as Facebook or Google Plus are used as a distant source of supervision for extraction of their at- tributes from user-generated text. In addi- tion to traditional linguistic features used in distant supervision for information ex- traction, our approach also takes into ac- count network information, a unique op- portunity offered by social media. We test our algorithm on three attribute domains: spouse, education and job; experimental results demonstrate our approach is able to make accurate predictions for users’ at- tributes based on their tweets.1 1 Introduction The overwhelming popularity of online social me- dia creates an opportunity to display given as- pects of oneself. Users’ profile information in social networking websites such as Facebook2 or Google Plus3 provides a rich repository personal information in a structured data format, making it amenable to automatic processing. This includes, for example, users’ jobs and education, and pro- vides a useful source of information for applica- tions such as search4, friend recommendation, on- 1Both code and data are available at http://aclweb. org/aclwiki/index.php?title=Profile_data 2https://www.facebook.com/ 3https://plus.google.com/ 4https://www.facebook.com/about/ graphsearch I@[shanenicholson] has taken all the kids today so can go shopping-CHILD FREE! #iloveyoushano #iloveyoucreditcard Tamworth promo day with my handsome classy husband @[shanenicholson] Spouse: shanenicholson I got accepted to be part of the UofM engineering safety pilot program in [FSU] Here in class. (@ [Florida State University] - Williams Building) Don’t worry , guys ! Our beloved [FSU] will always con- tinue to rise ” to the top ! Education: Florida State University (FSU) first day of work at [HuffPo], a sports bar woo come visit me yo.. start to think we should just add a couple desks to the [HuffPo] newsroom for Business Insider writers just back from [HuffPo], what a hell ! Job: HuffPo Table 1: Examples of Twitter message clues for user profile inference. line advertising, computational social science and more. Although profiles exist in an easy-to-use, struc- tured data format, they are often sparsely popu- lated; users rarely fully complete their online pro- files. Additionally, some social networking ser- vices such as Twitter don’t support this type of structured profile data. It is therefore difficult to obtain a reasonably comprehensive profile of a user, or a reasonably complete facet of information (say, education level) for a class of users. While many users do not explicitly list all their personal information …

Content truncated to 500 words