Unequal Representations: Analyzing Intersectional Biases in Word Embeddings Using Representational Similarity Analysis

Abstract

We present a new approach for detecting human-like social biases in word embeddings using representational similarity analysis. Specifically, we probe contextualized and non-contextualized embeddings for evidence of intersectional biases against Black women. We show that these embeddings represent Black women as simultaneously less feminine than White women, and less Black than Black men. This finding aligns with intersectionality theory, which argues that multiple identity categories (such as race or sex) layer on top of each other in order to create unique modes of discrimination that are not shared by any individual category.

A Novel Approach to Fairness in Automated Decision-Making using Affective Normalization

“…From health inequities to refugee policies, word embedding maps have already proven to be an excellent decision-making tool by uncovering key connections between current system failures and potential solutions (Turan et al, 2019). To address the gaps in underreported bias types (Rozado, 2020), there exists a need to study underrepresented groups by deliberately seeking material written by members of the groups and encoding such research into public-facing word association libraries (Lepori, 2020;Nichols and Stahl, 2019). Efforts are underway to use word embeddings to estimate sentiments in EPA space from existing texts shared by members of the group of interest (Van Loon and Freese, 2021), and to estimate cultural structures more generally (Kozlowski et al, 2019).…”