Improving Neural Machine Translation by Transferring Knowledge from Syntactic Constituent Alignment Learning