[ACL2019 Best Paper] Bridging the Gap between Training and Inference for Neural Machine Translation