Pseudo-Label Training and Model Inertia in Neural Machine Translation