Augmentor or Filter? Reconsider the Role of Pre-trained Language Model in Text Classification Augmentation

Text augmentation is one of the most effective techniques to solve the critical problem of insufficient data in text classification. Existing text augmentation methods achieve hopeful performance in few-shot text data augmentation. However, these methods usually lead to performance degeneration on public datasets due to poor quality augmentation instances. Our study shows that even employing pre-trained language models, existing text augmentation methods generate numerous low-quality instances and lead to the feature space shift problem in augmentation instances. However, we note that the pre-trained language model is good at finding low-quality instances provided that it has been fine-tuned on the target dataset. To alleviate the feature space shift and performance degeneration in existing text augmentation methods, we propose BOOSTAUG, which reconsiders the role of the language model in text augmentation and emphasizes the augmentation instance filtering rather than generation. We evaluate BOOSTAUG on both sentence-level text classification and aspect-based sentiment classification. The experimental results on seven commonly used text classification datasets show that our augmentation method obtains state-of-the-art performance. Moreover, BOOSTAUG is a flexible framework; we release the code which can help improve existing augmentation methods.