Gender bias in a neural machine translation system: a study of cross-language transfer mechanisms

This paper describes a study on gender bias in French/English neural machine trans-lation (MT) systems. We introduce a controlled corpus to measure the intensity of such biases in the two translation directions (from and into English). This corpus also allows us to investigate the information flow in a encoder-decoder architecture and to identify how gender information can be transfered between languages. Considering both probing as well as interventions on the internal representations of the MT system, we show that gender information is encoded in all token representations built by the encoder and the decoder and that there are multiple paths to transfer gender.