A Study on Factors of Gender Bias Using the Attention Mechanism in BERT

Abstract

Corpora on the internet reflect various social biases inherent in human society, and large language models trained on such data inevitably inherit these biases. This study aims to investigate the factors contributing to gender bias by analyzing the attention mechanism of BERT. Using pairs of sentences that differ only in gender words and examining their attention weights, we conduct various analyses of their correlations to explore and discuss the gender bias present in BERT. As a result, we found that in certain layers and heads, there are cases where words of the same gender are assigned greater weight in contexts perceived as masculine/feminine, and cases where words of the opposite gender are assigned greater weight.