Bridging the Gap between Training and Inference for Neural Machine Translation