Exploiting large pre-trained models for low-resource neural machine translation