GenderQuant: Quantifying Mention-Level Genderedness

Proceedings of NAACL-HLT 2019, pages 2959–2969 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2959 GenderQuant: Quantifying Mention-Level Genderedness Ananya University of California Irvine, CA aananya@uci.edu Nitya Parthasarthi Massachusetts Institute of Technology Cambridge, MA Sameer Singh University of California Irvine, CA sameer@uci.edu Abstract Language is gendered if the context surround- ing a mention is suggestive of a particular bi- nary gender for that mention. Detecting the different ways in which language is gendered is an important task since gendered language can bias NLP models (such as for coreference resolution). This task is challenging since gen- deredness is often expressed in subtle ways. Existing approaches need considerable anno- tation efforts for each language, domain, and author, and often require handcrafted lexicons and features. Additionally, these approaches do not provide a quantifiable measure of how gendered the text is, nor are they applicable at the fine-grained mention level. In this paper, we use existing NLP pipelines to automatically annotate gender of mentions in the text. On corpora labeled using this method, we train a supervised classifier to predict the gender of any mention from its context and evaluate it on unseen text. The model con- fidence for a mention’s gender can be used as a proxy to indicate the level of gendered- ness of the context. We test this gendered lan- guage detector on movie summaries, movie re- views, news articles, and fiction novels, achiev- ing an AUC-ROC of up to 0.71, and observe that the model predictions agree with human judgments collected for this task. We also pro- vide examples of detected gendered sentences from aforementioned domains. 1 Introduction Language can be extraordinarily gendered (Moul- ton et al., 1978). Genderedness in language is when we use words or phrases that are stereotypical or indicative of a particular gender (we only consider male vs female in this work) (Prior, 2017). It is im- portant to detect this bias in language since not only is this bias propagated to the readers (Menegatti and Rubini, 2017), but also machine learning algo- rithms trained on gendered corpora tend to become biased (Zhao et al., 2018a; Rudinger et al., 2018), often aggravating the disparity (Zhao et al., 2017). Bias in language and machine learning sys- tems can lead to unfair treatment, e.g., early work by Moulton et al. (1978) shows that males have an advantage in contexts where they are referred to by a putative neutral term. Recent work on corefer- ence resolution systems (Zhao et al., 2018a) shows that bias in machine learning systems originates from training on existing corpora, resulting in male- stereotyped professions like surgeon and president incorrectly resolved to males instead of females. Such biases in machine learning systems can lead to unintentional biases in downstream tasks pro- ducing effects like preferential treatment to male candidates over female candidates when selecting resumes (Dastin, 2018). Detecting these biases is the first step in finding a solution. Most of the current works for related problems tend to be domain-specific …

Content truncated to 500 words