On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation