Does Context Help Mitigate Gender Bias in Neural Machine Translation?

First, we evaluated the performance of contextaware models in the translation of stereotypical Neural machine translation (NMT) models tend to professions from English into German and French, exhibit gender bias, originating from their training measuring translation accuracy on gender-based data (Stanovsky et al., 2019; Saunders and subsets of the data. Our results in this case indicate Byrne, 2020). A typical example is the translation that, although context-aware models lead to significantly of gender-neutral professions in a language like increasing the use of feminine forms, this English, into languages with differentiated feminine was achieved mainly for professions that are stereotypically and masculine forms. In this case, NMT systems viewed as feminine, thus with limited bias often produce translations that reflect genderstereotypical mitigation.