Gender Bias in BERT -- Measuring and Analysing Biases through Sentiment Rating in a Realistic Downstream Classification Task

TL;DR

This paper introduces a new method to measure gender bias in BERT models based on sentiment differences and analyzes biases in a realistic movie classification task, revealing widespread biases across models and training conditions.

Contribution

It presents a novel bias measure based on sentiment valuation differences and provides a comprehensive analysis of gender biases in BERT across multiple models and training setups.

Findings

Most BERT models exhibit significant gender biases.

Biases mainly originate from public BERT models, not task-specific data.

Training conditions influence the extent of bias.

Abstract

Pretrained language models are publicly available and constantly finetuned for various real-life applications. As they become capable of grasping complex contextual information, harmful biases are likely increasingly intertwined with those models. This paper analyses gender bias in BERT models with two main contributions: First, a novel bias measure is introduced, defining biases as the difference in sentiment valuation of female and male sample versions. Second, we comprehensively analyse BERT's biases on the example of a realistic IMDB movie classifier. By systematically varying elements of the training pipeline, we can conclude regarding their impact on the final model bias. Seven different public BERT models in nine training conditions, i.e. 63 models in total, are compared. Almost all conditions yield significant gender biases. Results indicate that reflected biases stem from public BERT models rather than task-specific data, emphasising the weight of responsible usage.