What Your Username Says About You

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2032–2037, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. What Your Username Says About You Aaron Jaech and Mari Ostendorf Dept. of Electrical Engineering University of Washington {ajaech,ostendor}@uw.edu Abstract Usernames are ubiquitous on the Internet, and they are often suggestive of user de- mographics. This work looks at the de- gree to which gender and language can be inferred from a username alone by making use of unsupervised morphology induction to decompose usernames into sub-units. Experimental results on the two tasks demonstrate the effectiveness of the proposed morphological features com- pared to a character n-gram baseline. 1 Introduction There is much interest in automatic recognition of demographic information of Internet users to improve the quality of online interactions. Re- searchers have looked into identifying a variety of factors about users, including age, gender, lan- guage, religious beliefs and political views. Most work leverages multiple sources of information, such as search query history, Twitter feeds, Face- book likes, social network links, and user profiles. However, in many situations, little of this infor- mation is available. Conversely, usernames are al- most always available. In this work, we look specifically at classify- ing gender and language based only on the user- name. Prior work by sociologists has established a link between usernames and gender (Cornetto and Nowak, 2006), and studies have linked user- names to other attributes, such as individual be- liefs (Crabill, 2007; Hassa, 2012) and shown how usernames shape perceptions of gender and eth- nicity in the absence of common nonverbal cues (Pelletier, 2014). The connections to ethnicity mo- tivate the exploration of language identification. Gender identification based on given names is very effective for English (Liu and Ruths, 2013), since many names are strongly associated with a particular gender, like “Emily” or “Mark”. Unfor- tunately, the requirement that each username be unique precludes use of given names alone. In- stead, usernames are typically a combination of component words, names and numbers. For ex- ample, the Twitter name @taylorswift13 might decompose into “taylor”, “swift” and “13”. The sub-units carry meaning and, importantly, they are shared with many other individuals. Thus, our ap- proach is to leverage automatic decomposition of usernames into sub-units for use in classification. We use the Morfessor algorithm (Creutz and Lagus, 2006; Virpioja et al., 2013) for unsuper- vised morphology induction to learn the decom- position of the usernames into sub-units. Morfes- sor has been used successfully in a variety of lan- guage modeling frameworks applied to a number of languages, particularly for learning concatena- tive morphological structure. The usernames that we analyze are a good match to the Morfessor framework, which allows us to push the boundary of how much can be done with only a username. The classifier design is described in the next section, followed by a description of experiments on gender and language recognition that demon- strate the utility of morph-based features com- pared to character n-gram features. The …

Content truncated to 500 words