Investigating Gender Bias in BERT

In this work, we analyze the gender bias induced by BERT in downstream tasks. We also propose solutions to reduce bias. Contextual language models (CLMs) have pushed NLP benchmarks a new height. It has become norm utilize CLM-provided word embeddings tasks such as text classification. However, unless addressed, CLMs are prone learn intrinsic dataset. As result, predictions of can vary noticeably varying words, replacing “he” “she”, or even gender-neutral words. paper, focus our analysis on popular CLM, i.e., $$\text {BERT}$$ . it induces five related emotion and sentiment intensity prediction. For each task, train simple regressor utilizing ’s embeddings. then evaluate regressors using an equity evaluation corpus. Ideally from specific design, should discard informative features input. results show significant dependence system’s gender-particular words phrases. claim that biases be reduced removing gender-specific embedding. Hence, for layer BERT, identify directions primarily encode information. The space formed is referred subspace semantic algorithm finds fine-grained directions, one primary direction layer. This obviates need realizing multiple dimensions prevents other crucial information being omitted. Experiments embedding components achieves great success reducing BERT-induced investigation reveals contextualized model ( ) proposed solution seems promising biases.