Fairness in Machine Learning: Detecting and Removing Gender Bias in Language Models

Introduction

This study addresses gender biases in language models by implementing and testing two approaches: debiasing the dataset and debiasing the corpus embeddings (hard debiasing), resulting in significant reductions of bias in the BERT model, with dataset debiasing reducing bias by 29.41% and hard debiasing by 11.76%, accompanied by minimal classification accuracy loss.