Towards a Better Understanding of Label Smoothing in Neural Machine Translation