[Paper Review] Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings

Tolga Bolukbasi, Kai-Wei Chang+3 more|arXiv (Cornell University)|Jul 21, 2016

Hate Speech and Cyberbullying Detection33 references1,357 citations

TL;DR

The paper shows that word embeddings exhibit strong gender bias aligned with stereotypes and presents methods to debias embeddings, reducing bias while preserving useful properties like clustering and analogy performance.

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.

Motivation & Objective

• Quantify direct and indirect gender bias in existing word embeddings.
• Identify a gender subspace that captures the majority of gender-related variance.
• Develop debiasing methods that reduce bias for gender-neutral words while preserving definitional gender associations and useful relationships.
• Evaluate whether debiasing preserves clustering and analogy capabilities and assess alignment with human stereotypes.

Proposed method

• Compute and align a gender subspace from pairs of gender-specific words (e.g., she-he, woman-man).
• Define direct bias as a function of cosine similarity to the gender direction for gender-neutral words.
• Decompose word vectors into gender and non-gender components to quantify indirect bias.
• Propose debiasing algorithms that reduce bias on gender-neutral words while preserving meaningful associations.
• Assess bias reduction using crowd-sourced evaluations and standard embedding tasks like clustering and analogy solving.

Experimental results

Research questions

• RQ1 How strongly do word embeddings encode gender bias, both directly and indirectly?
• RQ2 Can a gender subspace be robustly identified across embeddings, and how can it be used to measure bias?
• RQ3 Is it possible to debias embeddings to reduce gender bias while maintaining useful semantic structure and analogy performance?

Key findings

• Word embeddings from Google News (and other sources) exhibit gender stereotypes in occupations and analogies (e.g., associations aligned with female/male roles).
• A gender subspace can be identified, capturing the majority of variance in gender-related word differences.
• Direct and indirect gender biases can be quantified and targeted for debiasing without destroying key embedding utilities such as word clustering and analogy tasks.
• Debiasing approaches significantly reduce gender bias while preserving the embedding’s useful properties, enabling applications that are less likely to amplify bias.