Continuous-Output Neural Machine Translation with Random Target Embeddings Outperforms Pre-Trained Embeddings
Random target embeddings can outperform pre-trained embeddings, especially on larger datasets and for rare words, in continuous-output neural machine translation.