RtGender: A Corpus for Studying Differential Responses to Gender

RtGender: A Corpus for Studying Differential Responses to Gender Rob Voigta, David Jurgensb, Vinodkumar Prabhakarana, Dan Jurafskya, Yulia Tsvetkovc aStanford University, bUniversity of Michigan, cCarnegie Mellon University robvoigt@stanford.edu, jurgens@umich.edu, vinodkpg@stanford.edu, jurafsky@stanford.edu, ytsvetko@cs.cmu.edu Abstract Like many social variables, gender pervasively influences how people communicate with one another. However, prior computational work has largely focused on linguistic gender difference and communications about gender, rather than communications directed to people of that gender, in part due to lack of data. Here, we fill a critical need by introducing a multi-genre corpus of more than 25M com- ments from five socially and topically diverse sources tagged for the gender of the addressee. Using these data, we describe pilot studies on how differential responses to gender can be measured and analyzed and present 30k annotations for the sentiment and relevance of these responses, showing that across our datasets responses to women are more likely to be emotive and about the speaker as an individual (rather than about the content being responded to). Our dataset enables studying socially important questions like gender bias, and has potential uses for downstream applications such as dialogue systems, gender detection or obfuscation, and debiasing language generation. Keywords: gender-annotated corpora, gender difference, gender bias, discourse, computational social science 1. Introduction Language is a means for the construction of identity and social categories like gender; social issues such as gen- der bias, in turn, often take form in language. Linguis- tic datasets have been used both to debunk gender-biased myths — for example, contrary to stereotype women are not actually more talkative than men (Mehl et al., 2007) — and to identify social issues. For instance, women1 journal- ists reach a smaller audience in terms of social media im- pressions (Matias and Wallach, 2012), and traditional gen- der stereotypes and unbalanced gender representation occur even in contemporary stories and movies (Fast et al., 2016; Sap et al., 2017). Large datasets are particularly of use in this context due to the complex nature of differential responses to gender. However, previous computational work on language and gender has focused mainly on language about or portraying persons of a particular gender (Wagner et al., 2015; Flekova et al., 2016; Agarwal et al., 2015). We thus present a large multi-genre dataset of online com- munication to enable research in a category of gender dif- ference understudied in computational work: responses to gender in language. These include posts and talks labeled for the gender of the source,2 along with comments given in response to the source texts. We collect such data from a variety of contexts, including: • Facebook (Politicians): Responses to Facebook posts from members of the U.S. House and Senate • Facebook (Public Figures): Responses to Facebook posts from other public figures, e.g., television hosts, journalists, and athletes • TED: Responses to presentations from TED speakers 1Throughout this paper we use the terms “woman” and “man” as labels for gender in preference to “female” and “male” since the latter terms are more commonly used as …

Content truncated to 500 words