Measuring Bias in Contextualized Word Representations

TL;DR

This paper introduces a template-based method to measure social biases in contextualized word embeddings like BERT, demonstrating improved consistency over traditional methods and applying it to gender bias in pronoun resolution, with potential for broader bias detection.

Contribution

The paper presents a novel, more reliable technique for quantifying social biases in contextual word embeddings, surpassing traditional cosine similarity methods.

Findings

Template-based bias measurement is more consistent than cosine methods.

Gender bias in BERT affects downstream pronoun resolution tasks.

Method generalizes to other social biases beyond gender.

Abstract

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on and amplify social stereotypes present in the data as well. In this study, we (1)~propose a template-based method to quantify bias in BERT; (2)~show that this method obtains more consistent results in capturing social biases than the traditional cosine based method; and (3)~conduct a case study, evaluating gender bias in a downstream task of Gender Pronoun Resolution. Although our case study focuses on gender bias, the proposed technique is generalizable to unveiling other biases, including in multiclass settings, such as racial and religious biases.