Why Gender and Age Prediction from Tweets is Hard: Lessons from a Crowdsourcing Experiment

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1950–1961, Dublin, Ireland, August 23-29 2014. Why Gender and Age Prediction from Tweets is Hard: Lessons from a Crowdsourcing Experiment Dong Nguyen14∗ Dolf Trieschnigg14 A. Seza Do˘gru¨oz23 Rilana Gravel4 Mari¨et Theune1 Theo Meder4 Franciska de Jong1 (1) Human Media Interaction, University of Twente, Enschede, The Netherlands (2) Netherlands Institute for Advanced Studies, Wassenaar, NL (3) Tilburg School of Humanities, Tilburg University, Tilburg, NL (4) Meertens Institute, Amsterdam, The Netherlands ∗Corresponding author: d.nguyen@utwente.nl Abstract There is a growing interest in automatically predicting the gender and age of authors from texts. However, most research so far ignores that language use is related to the social identity of speak- ers, which may be different from their biological identity. In this paper, we combine insights from sociolinguistics with data collected through an online game, to underline the importance of approaching age and gender as social variables rather than static biological variables. In our game, thousands of players guessed the gender and age of Twitter users based on tweets alone. We show that more than 10% of the Twitter users do not employ language that the crowd as- sociates with their biological sex. It is also shown that older Twitter users are often perceived to be younger. Our findings highlight the limitations of current approaches to gender and age prediction from texts. 1 Introduction A major thrust of research in sociolinguistics aims to uncover the relationship between social variables such as age and gender, and language use (Holmes and Meyerhoff, 2003; Eckert and McConnell-Ginet, 2013; Eckert, 1997; Wagner, 2012). In line with scholars from a variety of disciplines, including the so- cial sciences and philosophy, sociolinguists consider age and gender as social and fluid variables (Eckert, 2012). Gender and age are shaped depending on the societal context, the culture of the speakers involved in a conversation, the individual experiences and the multitude of social roles: a female teenager might also be a high school student, a piano player, a swimmer, etc. (Eckert, 2008). Speakers use language as a resource to construct their identity (Bucholtz and Hall, 2005). For example, a person’s gender identity is constructed through language by using linguistic features associated with male or female speech. These features gain social meaning in a cultural and societal context. On Twitter, users construct their identity through interacting with other users (Marwick and boyd, 2011). Depending on the context, they may emphasize specific aspects of their identity, which leads to linguistic variation both within and between speakers. We illustrate this with the following three tweets: Tweet 1: I’m walking on sunshine <3 #and don’t you feel good Tweet 2: lalaloveya <3 Tweet 3: @USER loveyou ;D In these tweets, we find linguistic markers usually associated with females (e.g. a heart represented as <3). Indeed, 77% of the 181 players guessed that a female wrote these tweets in our online game. However, this is a 16-year old biological male, whose Twitter account reveals that …

Content truncated to 500 words