A Diversity-Enhanced and Constraints-Relaxed Augmentation for Low-Resource Classification

Previous studies on Data Augmentation (DA) mostly use a fine-tuned Language Model (LM) to strengthen the constraints but ignore fact that potential of diversity could improve effectiveness generated data. To address this dilemma, we propose Diversity-Enhanced and Constraints-Relaxed (DECRA) has two essential components top transformer-based backbone model, including \(\mathbf {k}\)-\(\varvec{\beta }\) augmentation masked language model loss. Extensive experiments demonstrate our DECRA outperforms state-of-the-art approaches by 3.8% in overall score.