Wasserstein Distance Regularized Sequence Representation for Text Matching in Asymmetrical Domains

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 314–331, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 314 Multi-Dimensional Gender Bias Classification Emily Dinan∗, Angela Fan∗†, Ledell Wu, Jason Weston, Douwe Kiela, Adina Williams Facebook AI Research †Laboratoire Lorrain d’Informatique et Applications (LORIA) Abstract Machine learning models are trained to find patterns in data. NLP models can inadvertently learn socially undesirable patterns when train- ing on gender biased text. In this work, we pro- pose a novel, general framework that decom- poses gender bias in text along several prag- matic and semantic dimensions: bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker. Using this fine-grained framework, we automatically annotate eight large scale datasets with gen- der information. In addition, we collect a new, crowdsourced evaluation benchmark. Distin- guishing between gender bias along multiple dimensions enables us to train better and more fine-grained gender bias classifiers. We show our classifiers are valuable for a variety of ap- plications, like controlling for gender bias in generative models, detecting gender bias in ar- bitrary text, and classifying text as offensive based on its genderedness. 1 Introduction Language is a primary means by which people com- municate, express their identities, and categorize themselves and others both explicitly and implic- itly. Such social information is present in the words we write and, consequently, in the text we use to train our NLP models. In particular, models often can unwittingly learn negative associations about protected groups present in their training data and propagate them. In particular, NLP models often learn to replicate unwanted gender biases present in society (Bolukbasi et al., 2016; Hovy and Spruit, 2016; Caliskan et al., 2017; Rudinger et al., 2017; Garg et al., 2018; Gonen and Goldberg, 2019; Di- nan et al., 2020). Since unwanted gender biases can affect downstream applications—sometimes even ∗Joint first authors. Figure 1: Framework for Gender Bias in Dialogue. We propose a framework separating gendered language based on who you are speaking ABOUT, speaking TO, and speaking AS. leading to poor user experiences—understanding and mitigating gender bias is an important step towards making NLP tools and models safer and more equitable. In this paper, we provide a fine- grained framework for that purpose, analyze the presence of gender bias in models and data, and em- power others by releasing tools that can be further applied to numerous text-based use-cases. While many works have explored methods for re- moving gender bias from text (Emami et al., 2019; Hall Maudslay et al., 2019; Ravfogel et al., 2020), no extant work on classifying gender or removing gender bias has foregrounded the fact that we use language, at least in part, to collaboratively and so- cially construct our gender identities. We propose a pragmatic and semantic framework for measur- ing bias along three dimensions that is sensitive to conversational and performative aspects of gender, as illustrated in Figure …

Content truncated to 500 words