Gender Attribution: Tracing Stylometric Evidence Beyond Topic and Genre

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 78–86, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Gender Attribution: Tracing Stylometric Evidence Beyond Topic and Genre Ruchita Sarawgi, Kailash Gajulapalli, and Yejin Choi Department of Computer Science Stony Brook University NY 11794, USA {rsarawgi, kgajulapalli, ychoi}@cs.stonybrook.edu Abstract Sociolinguistic theories (e.g., Lakoff (1973)) postulate that women’s language styles differ from that of men. In this paper, we explore statistical techniques that can learn to iden- tify the gender of authors in modern English text, such as web blogs and scientific papers. Although recent work has shown the efficacy of statistical approaches to gender attribution, we conjecture that the reported performance might be overly optimistic due to non-stylistic factors such as topic bias in gender that can make the gender detection task easier. Our work is the first that consciously avoids gender bias in topics, thereby providing stronger evi- dence to gender-specific styles in language be- yond topic. In addition, our comparative study provides new insights into robustness of var- ious stylometric techniques across topic and genre. 1 Introduction Sociolinguistic theories (e.g., Lakoff (1973)) postu- late that women’s language styles differ from that of men with respect to various aspects of communi- cation, such as discourse behavior, body language, lexical choices, and linguistic cues (e.g., Crosby and Nyquist (1977), Tannen (1991), Argamon et al. (2003), Eckert and McConnell-Ginet (2003), Arga- mon et al. (2007)). In this paper, we explore statis- tical techniques that can learn to identify the gen- der of authors in modern English text, such as web blogs and scientific papers, motivated by sociolin- guistic theories for gender attribution. There is a broad range of potential applications across computational linguistics and social science where statistical techniques for gender attribution can be useful: e.g., they can help understanding de- mographic characteristics of user-created web text today, which can provide new insight to social sci- ence as well as intelligent marketing and opinion mining. Models for gender attribution can also help tracking changes to gender-specific styles in lan- guage over different domain and time. Gender de- tectors can be useful to guide the style of writing as well, if one needs to assume the style of a specific gender for imaginative writing. Although some recent work has shown the effi- cacy of machine learning techniques to gender at- tribution (e.g., Koppel et al. (2002), Mukherjee and Liu (2010)), we conjecture that the reported perfor- mance might be overly optimistic under scrutiny due to non-stylistic factors such as topic bias in gender that can make the gender detection task easier. In- deed, recent research on web blogs reports that there is substantial gender bias in topics (e.g., Janssen and Murachver (2004), Argamon et al. (2007)) as well as in genre (e.g., Herring and Paolillo (2006)). In order to address this concern, we perform the first comparative study of machine learning tech- niques for gender attribution after deliberately re- moving gender bias in topics and genre. Further- more, …

Content truncated to 500 words