The Exacerbation of (Grammatical) Gender Stereotypes in English–Spanish Machine Translation

Abstract

The information required to select grammatical gender in machine translation of isolated sentences for gender-marking languages is frequently missing or difficult to extract. Our text-centric, black-box study demonstrates how the gender distribution of the training set is distorted at the output. Human evaluation reveals that gender clues are frequently absent from the source, resulting in stereotyped translations.