Removing gender bias from word embeddings is essential for numerous downstream applications. Unfortunately, the entirety of the relevant literature and proposed methodologies is aimed at English, a high-resource language which possesses a linguistic structure fundamentally different from that of other world languages. In this study, we evaluate the relevance of gender-debiasing methods to low-resource, non-English languages, taking Greek as a case study. Our results show that these methodologies are successful both in removing gender bias and in preserving useful semantic information. This suggests that other non-English languages may also benefit from these techniques for debiasing word embeddings.
Tsimenidis et al. (Thu,) studied this question.