Investigating the Roots of Gender Bias in Machine Translation: Observations on Gender Transfer between French and English

Abstract

This paper aims at identifying the inner mechanisms that make a translation model choose a masculine rather than a feminine form, an essential step to mitigate gender bias in MT. We conduct two series of experiments using probing and comparing the predictions of translation and a language models to show that i) gender information is encoded in all decoder's and encoder's representations and ii) the translation model does not need to use information from the source to predict his.

Tasks

Reproductions