Handling Gender Bias in Neural Machine Translation: A Focus on English-Khasi Language Pair

Abstract

Machine Translation systems have progressed tremendously over the years and have enhanced cross-lingual communication. Although machine translation has advanced from a rule-based to a neural machine translation method, a new issue occurring in most machine translation systems, such as gender bias, cannot be ignored. Gender bias not only affects translation accuracy but also has an impact on societal prejudices and often promotes inequalities and stereotypes. This study attempts to address the challenges of gender bias in neural machine translation for English-Khasi, a low- resource language pair, where data scarcity increases the risk of biased translations. To the best of our knowledge, this is the first attempt to report a study on gender bias in the neural machine translation task of English-Khasi language pair. We use two different methods; a data augmentation technique and a transfer learning method tailored to the linguistic and socio-cultural characteristics of the target language. To implement the two methods, we manually build a sizeable gender-balanced English-Khasi parallel corpora to handle gender bias in English-Khasi neural machine translation systems. Through empirical evaluation of low-resource language pairs, English-Khasi, we demonstrate the effectiveness of the transfer learning approach in reducing gender bias while maintaining translation quality.