Automatically Inferring Gender Associations from Language

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 5746–5752, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 5746 Automatically Inferring Gender Associations from Language Serina Chang Department of Computer Science Columbia University∗ serinac@stanford.edu Kathleen McKeown Department of Computer Science Columbia University kathy@cs.columbia.edu Abstract In this paper, we pose the question: do people talk about women and men in different ways? We introduce two datasets and a novel integra- tion of approaches for automatically inferring gender associations from language, discover- ing coherent word clusters, and labeling the clusters for the semantic concepts they repre- sent. The datasets allow us to compare how people write about women and men in two dif- ferent settings - one set draws from celebrity news and the other from student reviews of computer science professors. We demonstrate that there are large-scale differences in the ways that people talk about women and men and that these differences vary across domains. Human evaluations show that our methods sig- nificantly outperform strong baselines. 1 Introduction It is well-established that gender bias exists in language – for example, we see evidence of this given the prevalence of sexism in abusive lan- guage datasets (Waseem and Hovy, 2016; Jha and Mamidi, 2017). However, these are extreme cases of gender norms in language, and only encompass a small proportion of speakers or texts. Less studied in NLP is how gender norms man- ifest in everyday language – do people talk about women and men in different ways? These types of differences are far subtler than abusive language, but they can provide valuable insight into the roots of more extreme acts of discrimination. Subtle dif- ferences are difficult to observe because each case on its own could be attributed to circumstance, a passing comment or an accidental word. How- ever, at the level of hundreds of thousands of data points, these patterns, if they do exist, become un- deniable. Thus, in this work, we introduce new datasets and methods so that we can study subtle gender associations in language at the large-scale. ∗ Since writing this paper, Serina Chang has moved to the Department of Computer Science at Stanford University. Our contributions include: • Two datasets for studying language and gen- der, each consisting of over 300K sentences. • Methods to infer gender-associated words and labeled clusters in any domain. • Novel findings that demonstrate in both do- mains that people do talk about women and men in different ways. Each contribution brings us closer to modeling how gender associations appear in everyday lan- guage. In the remainder of the paper, we present related work, our data collection, methods and findings, and human evaluations of our system.1 2 Related Work The study of gender and language has a rich his- tory in social science. Its roots are often attributed to Robin Lakoff, who argued that language is fun- damental to gender inequality, “reflected in both …

Content truncated to 500 words