Random target embeddings can outperform pre-trained embeddings, especially on larger datasets and for rare words, in continuous-output neural machine translation.
