A Text Augmentation Method Based on Dual Diversity Generation

In text augmentation in the field of natural language processing, it has been observed that most of the traditional methods often result in only minimal changes to the original text, which consequently limits the diversity of the augmented samples. In this paper, a novel text augmentation method based on Dual Diversity Generation is proposed, which effectively addresses the aforementioned limitation through two key components: keyword diversity and sample diversity. Initially, a sketch is constructed through keyphrase extraction, which is subsequently input into the text augmentation model for the generation of new samples. To mitigate the semantic overlap among the extracted keyphrases, keyphrase diversity is incorporated into the extraction process. Furthermore, sample diversity is emphasized during the generation phase to enhance the overall quality of the text augmentation. The performance of the proposed method is evaluated by text classification effect, and the experimental results indicate that the proposed method achieves significant improvements compared to other methods.