Measuring and Mitigating Gender Bias in Contextualized Word Embeddings

Deep learning models like Transformers revolutionize the execution of automated tasks. However, these models are prone to many biases that can perpetuate societal inequalities. Sometimes disclosure of gender information by these models may lead to privacy related issues. These models are exploited most for Natural language processing (NLP) tasks and have attracted the attention of researchers to examine the inherent societal bias in them. In this paper we devise a post-processing mechanism that measure and mitigate the gender bias in the contextualized embeddings of the T5 model (Text-to-Text-Transfer-Transformer model).