A Causal Inference Method for Reducing Gender Bias in Word Embedding Relations

TL;DR

This paper introduces a causal method to reduce gender bias in word embeddings by leveraging statistical dependencies, achieving state-of-the-art results across multiple bias mitigation and NLP tasks.

Contribution

A novel causal approach that targets gender bias in word embedding relations, addressing limitations of previous methods focused only on gender direction bias.

Findings

Achieves state-of-the-art results on gender-debiasing tasks

Improves performance on lexical and sentence-level evaluations

Enhances downstream coreference resolution accuracy

Abstract

Word embedding has become essential for natural language processing as it boosts empirical performances of various tasks. However, recent research discovers that gender bias is incorporated in neural word embeddings, and downstream tasks that rely on these biased word vectors also produce gender-biased results. While some word-embedding gender-debiasing methods have been developed, these methods mainly focus on reducing gender bias associated with gender direction and fail to reduce the gender bias presented in word embedding relations. In this paper, we design a causal and simple approach for mitigating gender bias in word vector relation by utilizing the statistical dependency between gender-definition word embeddings and gender-biased word embeddings. Our method attains state-of-the-art results on gender-debiasing tasks, lexical- and sentence-level evaluation tasks, and downstream coreference resolution tasks.