Screening Gender Transfer in Neural Machine Translation

Abstract

This paper aims at identifying the information flow in state-of-the-art machine translation systems, taking as example the transfer of gender when translating from French into English.Using a controlled set of examples, we experiment several ways to investigate how gender information circulates in a encoder-decoder architecture considering both probing techniques as well as interventions on the internal representations used in the MT system. Our results show that gender information can be found in all token representations built by the encoder and the decoder and lead us to conclude that there are multiple pathways for gender transfer.

Evaluer, diagnostiquer et analyser la traduction automatique neuronale

“…Comme le note toutefois [Vanmassenhove et al, 2017], qui étendent les travaux de [Shi et al, 2016] à l'étude de l'aspect du verbe principal, ce n'est pas parce qu' une information utile pour la traduction est correctement extraite qu'elle sera correctement utilisée. Cette observation est corroborrée par [Wisniewski et al, 2021b]…”

“Le professeur est très intelligent / La prof est très attirante”. Recognizing and Reducing Gender Bias in Neural Machine Translation

“…Hence, focusing on these issues is extremely relevant, in order to act on the products of Neural Machine Translation systems which are not only non-transparent for the user, but quite often also for their creators. As Wisniewski, Zhu, Ballier and Yvon and Zhu argue, "contrary to previous generations of MT engines where transfer rules were quite transparent, understanding this information flow [from the encoder to the decoder] within state-of-the-art neural MT systems is a challenging task, and a key step for their interpretability" (Wisniewski et al 2021b).…”