The Unreasonable Effectiveness of Random Target Embeddings for Continuous-Output Neural Machine Translation

Open in new window