Is linguistically-motivated data augmentation worth it?