# Black is to Criminal as Caucasian is to Police: Detecting and Removing Multiclass Bias in Word Embeddings

## TL;DR

This paper introduces a new method for detecting and removing biases related to race and religion in word embeddings, extending previous binary bias removal techniques, and provides a novel evaluation approach to ensure effectiveness across multiple classes.

## Contribution

It presents a novel multiclass debiasing method for word embeddings and a new evaluation framework to assess bias removal effectiveness in complex, real-world scenarios.

## Key findings

- Multiclass debiasing effectively reduces stereotypes in word embeddings.
- The proposed method maintains performance on standard NLP tasks.
- Debiasing is robust across different types of biases and datasets.

## Abstract

Online texts -- across genres, registers, domains, and styles -- are riddled with human stereotypes, expressed in overt or subtle ways. Word embeddings, trained on these texts, perpetuate and amplify these stereotypes, and propagate biases to machine learning models that use word embeddings as features. In this work, we propose a method to debias word embeddings in multiclass settings such as race and religion, extending the work of (Bolukbasi et al., 2016) from the binary setting, such as binary gender. Next, we propose a novel methodology for the evaluation of multiclass debiasing. We demonstrate that our multiclass debiasing is robust and maintains the efficacy in standard NLP tasks.