[Paper Review] Debiasing Word Embeddings with Nonlinear Geometry

Cheng Lu, Nayoung Kim+1 more|arXiv (Cornell University)|Aug 29, 2022

Hate Speech and Cyberbullying Detection4 citations

TL;DR

This paper proposes JoSEC, a nonlinear geometry-based method to debias word embeddings for joint and intersectional social biases—such as those involving race and gender—by leveraging the intersection of individual bias subspaces without requiring additional human-annotated word sets. The approach effectively reduces intrinsic bias while preserving semantic utility, as validated on benchmark datasets.

ABSTRACT

Debiasing word embeddings has been largely limited to individual and independent social categories. However, real-world corpora typically present multiple social categories that possibly correlate or intersect with each other. For instance, "hair weaves" is stereotypically associated with African American females, but neither African American nor females alone. Therefore, this work studies biases associated with multiple social categories: joint biases induced by the union of different categories and intersectional biases that do not overlap with the biases of the constituent categories. We first empirically observe that individual biases intersect non-trivially (i.e., over a one-dimensional subspace). Drawing from the intersectional theory in social science and the linguistic theory, we then construct an intersectional subspace to debias for multiple social categories using the nonlinear geometry of individual biases. Empirical evaluations corroborate the efficacy of our approach. Data and implementation code can be downloaded at https://github.com/GitHubLuCheng/Implementation-of-JoSEC-COLING-22.

Motivation & Objective

• To address the limitation of existing debiasing methods that treat social categories in isolation, especially when multiple categories intersect or correlate.
• To study and mitigate two types of complex biases: joint biases (resulting from the union of categories) and intersectional biases (not overlapping with individual category biases).
• To develop a method that constructs bias subspaces for multiple categories without relying on predefined word lists for intersectional groups.
• To preserve the semantic utility of word embeddings while reducing intrinsic bias across multiple social categories.
• To provide a theoretically grounded, geometry-driven approach that aligns with intersectionality theory and linguistic theory.

Proposed method

• The method empirically observes that individual bias subspaces (e.g., for gender and race) intersect non-trivially, primarily along a one-dimensional subspace.
• It formulates the intersectional bias subspace as the nonlinear geometric intersection of individual bias subspaces, informed by intersectionality theory and Firth’s linguistic theory.
• JoSEC uses nonlinear manifold learning to model the complex, non-linear correlations between bias directions, avoiding linear combinations that may amplify bias.
• The approach leverages only the defining word sets for constituent categories (e.g., gendered or racial terms), eliminating the need for additional intersectional word lists.
• It applies a modified hard-debiasing procedure on the learned intersectional subspace to neutralize bias in word embeddings.
• The method is evaluated using intrinsic bias metrics and downstream NLP tasks to assess both bias reduction and semantic utility preservation.

Experimental results

Research questions

• RQ1 How do individual bias subspaces for multiple social categories (e.g., gender and race) intersect in the embedding space?
• RQ2 Can a nonlinear geometric approach better capture joint and intersectional biases than linear combinations of individual bias subspaces?
• RQ3 Does debiasing using the intersectional subspace preserve the semantic utility of word embeddings across NLP tasks?
• RQ4 Is there a reliable correlation between intrinsic bias reduction and extrinsic bias mitigation in downstream tasks like toxicity classification?
• RQ5 Can JoSEC reduce bias without requiring human-annotated word lists for intersectional identities?

Key findings

• The individual bias subspaces for different social categories (e.g., gender and race) intersect non-trivially, primarily along a one-dimensional subspace, challenging the assumption of linear independence.
• JoSEC successfully reduces intrinsic bias for both joint and intersectional biases without requiring additional human-labeled word sets for intersectional groups.
• Debiasing with JoSEC maintains semantic utility, as shown by statistically insignificant performance drops on NER, POS tagging, and POS chunking tasks.
• The method shows no significant correlation between intrinsic bias reduction and extrinsic bias in toxicity classification, suggesting intrinsic metrics may not reliably predict downstream fairness.
• Empirical results confirm that JoSEC outperforms linear combinations of bias subspaces, which can amplify bias due to nonlinear correlations.
• The approach is effective and generalizable, as it relies only on standard defining sets for constituent categories, making it practical for real-world deployment.