Investigating Gender Bias in Machine Translation. A Case Study between English and Italian

Neural machine translation systems have substantially improved the quality of translation output, yet many issues still need to be addressed: one major problem to be addressed concerns the presence of gender bias, the prejudice against one gender based on the perception that women and men are not equal. In this work, we will manually evaluate the translation of a sentence pattern previously employed for similar purposes by Escudé Font and Costa-jussà (2019) in the English-Italian language combination using two of the most popular MT systems, DeepL and Google Translate. The sets of sentences include 40 male- and female-dominated occupations and three adjectives, beautiful, wise and strong. The aim of this study is to evaluate gender bias, that becomes apparent when translating from a gender-neutral language to a gender-marked language, and to verify whether adjectives usually associated with female or male entities can affect the final MT output. Furthermore, we provide some relevant insights about gender bias in MT for post-editors and MT users, with a particular focus on the under-representation of women in the Italian language.