Unsupervised Discovery of Gendered Language through Latent-Variable Modeling

Abstract

Studying the ways in which language is gendered has long been an area of interest in sociolinguistics. Studies have explored, for example, the speech of male and female characters in film and the language used to describe male and female politicians. In this paper, we aim not to merely study this phenomenon qualitatively, but instead to quantify the degree to which the language used to describe men and women is different and, moreover, different in a positive or negative way. To that end, we introduce a generative latent-variable model that jointly represents adjective (or verb) choice, with its sentiment, given the natural gender of a head (or dependent) noun. We find that there are significant differences between descriptions of male and female nouns and that these differences align with common gender stereotypes: Positive adjectives used to describe women are more often related to their bodies than adjectives used to describe men.

Content Analysis of Textbooks via Natural Language Processing: Findings on Gender, Race, and Ethnicity in Texas U.S. History Textbooks

“…1 Some early efforts apply machine counting of words to scanned textbooks, such as Lachmann and Mitchell (2014)'s study on depictions of war. A number of recent studies outside education have used NLP methods to study the reflection of gender and other social variables in text: Fast et al (2016) look at gender stereotypes in online fiction; Hoyle et al (2019) measured the association of adjectives and verbs with different genders in a million digitized books; Garg et al (2018) quantified a century of gender and ethnic stereotypes using word representations learned from books, newspapers, and other texts; and Ash et al (2020) examine the role of gender slant in judicial behavior using text written by judges. We build on this line of work examining depictions of social groups in texts (see also Field et al, 2019;Joseph et al, 2017;Ornaghi et al, 2019), extending NLP methods to textbooks.…”

“…To extract verbs and adjectives associated with people, we used a part-of-speech tagger and dependency parser, a tool that annotates dependency relations between words (we used a parser by Dozat et al, 2017). This approach is similar to those used by previous work for gathering descriptive attributes of entities in movie plot summaries, books, and news (Bamman et al, 2013;Card et al, 2016;Hoyle et al, 2019). We perform dependency parsing to extract verbs and adjectives associated with people-related terms.…”

Multi-Dimensional Gender Bias Classification

“…It's well known that we change how we speak about others depending on who they are (Hymes, 1974;Rickford and McNair-Knox, 1994), and what their gender identity is (Lakoff, 1973;Eckert and McConnell-Ginet, 1992). For example, adjectives which describe women have been shown to differ from those used to describe men in numerous situations (Trix and Psenka, 2003;Gaucher et al, 2011;Moon, 2014;Hoyle et al, 2019), as do verbs that take nouns referring to men as opposed to women (Guerin, 1994;Hoyle et al, 2019).…”

“…Previous work on gender bias classification has been predominantly single-task-often supervised on the task of analogy-and relied mainly on word lists, that are binarily gendered (Bolukbasi et al, 2016;Zhao et al, 2018b)-sometimes also explicitly (Caliskan et al, 2017;Hoyle et al, 2019 based approaches provided a solid start, they ultimately prove insufficient. First, they conflate different conversational dimensions of gender bias, and are therefore unable to detect the subtle, but very well-described, pragmatic differences of interest here.…”

“…We also compare to a baseline for which the control tokens are determined by a word list: if an utterance contains more masculine-gendered words than feminine-gendered words from the word list it is labeled as masculine (and vice versa for feminine); if it contains no gendered words or an equal number of masculine and feminine gendered words, it is labeled as neutral. Following Dinan et al (2020), we combine several existing word lists (Zhao et al, 2018bHoyle et al, 2019).…”