Abstraction and Gender Bias in Dissertation Writing

• In Event: Text-Based Machine Learning, Big Data, and the Social Study of Science - I

Abstract

Do gendered expectations in the use of concepts impact the structure of semantic networks? Recent studies show that one dimension of how academic language gets gendered is that men tend to use more abstract terms. Scholars use abstraction not only to solve problems but to lay claim over them. We test how shifts of gender balance in a discipline’s student body affect this social function of abstraction. To this end, we built a series of networks that represent hierarchies of abstraction between nouns used in Education dissertations defended between 1980 and 2010. These networks proxy abstraction as relations of semantic entailment, whereby “a term v is distributionally more general than u if u appears in a subset of the contexts in which v is found.” Then we modelled the evolution of this hierarchy as a function of gendered-biased use of concepts. This bias is measured as a log odds ratio between the chances that a term is key in dissertations written by female students against the term being distinctive for male students. Preliminary results show that higher levels of abstraction correspond to less gendered concept use. This relation holds over time, as the distribution of gendered concept use fails to change over time.

Authors

• Sebastian Munoz-Najar Galvez, Stanford University

• Bas Hofstra, Stanford University