Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

TL;DR

This paper identifies gender biases in word embeddings trained on news data and proposes a method to effectively reduce these biases while preserving the embeddings' utility for NLP tasks.

Contribution

It introduces a geometric approach to detect and remove gender bias in word embeddings, along with algorithms and metrics for debiasing.

Findings

Gender bias is captured by a specific direction in the embedding space.

Debiasing reduces gender stereotypes while maintaining embedding utility.

The method effectively diminishes bias in practical NLP applications.

Abstract

The blind application of machine learning runs the risk of amplifying biases present in data. Such a danger is facing us with word embedding, a popular framework to represent text data as vectors which has been used in many machine learning and natural language processing tasks. We show that even word embeddings trained on Google News articles exhibit female/male gender stereotypes to a disturbing extent. This raises concerns because their widespread use, as we describe, often tends to amplify these biases. Geometrically, gender bias is first shown to be captured by a direction in the word embedding. Second, gender neutral words are shown to be linearly separable from gender definition words in the word embedding. Using these properties, we provide a methodology for modifying an embedding to remove gender stereotypes, such as the association between between the words receptionist and female, while maintaining desired associations such as between the words queen and female. We define metrics to quantify both direct and indirect gender biases in embeddings, and develop algorithms to "debias" the embedding. Using crowd-worker evaluation as well as standard benchmarks, we empirically demonstrate that our algorithms significantly reduce gender bias in embeddings while preserving the its useful properties such as the ability to cluster related concepts and to solve analogy tasks. The resulting embeddings can be used in applications without amplifying gender bias.