Identifying and Mitigating Gender Bias in Hyperbolic Word Embeddings

TL;DR

This paper investigates gender bias in hyperbolic word embeddings, introduces a new bias measure called gyrocosine bias, and proposes a debiasing method called Poincaré Gender Debias that effectively reduces bias with minimal semantic distortion.

Contribution

It extends gender bias analysis to hyperbolic embeddings, introduces gyrocosine bias as a new measure, and presents PGD, a novel debiasing technique for hyperbolic word representations.

Findings

Gyrocosine bias reveals significant gender bias in hyperbolic embeddings.

PGD effectively reduces gender bias in hyperbolic embeddings.

Debiasing with PGD causes minimal semantic offset.

Abstract

Euclidean word embedding models such as GloVe and Word2Vec have been shown to reflect human-like gender biases. In this paper, we extend the study of gender bias to the recently popularized hyperbolic word embeddings. We propose gyrocosine bias, a novel measure for quantifying gender bias in hyperbolic word representations and observe a significant presence of gender bias. To address this problem, we propose Poincar\'e Gender Debias (PGD), a novel debiasing procedure for hyperbolic word representations. Experiments on a suit of evaluation tests show that PGD effectively reduces bias while adding a minimal semantic offset.