Gender in Danger? Evaluating Speech Translation Technology on the MuST-SHE Corpus

摘要: Translating from languages without productive grammatical gender like English into gender-marked is a well-known difficulty for machines. This also due to the fact that training data on which models are built typically reflect asymmetries of natural languages, bias included. Exclusively fed with textual data, machine translation intrinsically constrained by input sentence does not always contain clues about identity referred human entities. But what happens speech translation, where an audio signal? Can provide additional information reduce bias? We present first thorough investigation in contributing with: i) release benchmark useful future studies, and ii) comparison different technologies (cascade end-to-end) two language directions (English-Italian/French).