Abstract

Gender stereotypes are manifest in most of the world's languages and areconsequently propagated or amplified by NLP systems. Although research hasfocused on mitigating gender stereotypes in English, the approaches that arecommonly employed produce ungrammatical sentences in morphologically richlanguages. We present a novel approach for converting betweenmasculine-inflected and feminine-inflected sentences in such languages. ForSpanish and Hebrew, our approach achieves F1 scores of 82% and 73% at the levelof tags and accuracies of 90% and 87% at the level of forms. By evaluating ourapproach using four different languages, we show that, on average, it reducesgender stereotyping by a factor of 2.5 without any sacrifice to grammaticality.

 

Quick Read (beta)