Gender Bias in Multilingual Neural Machine Translation: The Architecture\n Matters

Multilingual Neural Machine Translation architectures mainly differ in the\namount of sharing modules and parameters among languages. In this paper, and\nfrom an algorithmic perspective, we explore if the chosen architecture, when\ntrained with the same data, influences the gender bias accuracy. Experiments in\nfour language pairs show that Language-Specific encoders-decoders exhibit less\nbias than the Shared encoder-decoder architecture. Further interpretability\nanalysis of source embeddings and the attention shows that, in the\nLanguage-Specific case, the embeddings encode more gender information, and its\nattention is more diverted. Both behaviors help in mitigating gender bias.\n