Gender Bias Mitigation of Persian Stereotypical Words Using FastText

Gender bias in word embeddings is a significant problem in Natural Language Processing tasks, which can reinforce harmful stereotypes. While a lot of effort has been devoted to debiasing word embeddings for high-resource languages like English, Persian remains a relatively under-studied area in this respect. This study addresses gender bias in Persian word embeddings through a step-by-step approach, beginning with utilizing pre-trained FastText as a comprehensive word embedding for Persian, followed by data cleaning to remove meaningless words while preserving inherently gendered words. We then created a unit gender vector to measure the gender bias of each word, followed by debiasing the words by removing the gender dimension in vector space without impacting other dimensions. Meanwhile, inherently gendered words remained unchanged. In addition, to ensure balanced representation, pairs of gendered words were equalized symmetrically in terms of gender through an equalization step. The findings revealed a substantial decrease in gender bias, as the projection analyses showed that stereotypical words no longer had gender features following debiasing, but words with inherent gender meanings preserved their semantic stability. A total of 1,249,168 out of 1,259,079 words were successfully debiased after using the debiasing method on stereotypical words, whereas 1,825 words with inherent gender meanings were not affected.