Extract and Attend: Improving Entity Translation in Neural Machine Translation