Beyond Weight Tying: Learning Joint Input-Output Embeddings for Neural Machine Translation | Read Paper on Bytez