On the Impact of Data Augmentation on Downstream Performance in Natural Language Processing