Text data augmentation based on semantic roles

The performance of deep learning models in natural language processing (NLP) is heavily dependent on the volume and quality of training data. However, acquiring large amounts of high-quality labeled data is often costly and challenging. Text data augmentation presents a low-cost and efficient solution to expand datasets by generating diversified new samples. A critical limitation of traditional augmentation methods, such as synonym replacement and random manipulation, is the "non-core word replacement" problem. These methods often replace functionally unimportant words (e.g., prepositions, conjunctions), resulting in limited semantic variation and ineffective augmented data. To address this issue, this paper innovatively introduces semantic roles into the data augmentation process, proposing a novel semantic role-based data augmentation (SRDA) method. Our approach consists of three key components: (1) we constructed an advanced MacBERT-CRF model for semantic role labeling (SRL) that outperforms traditional architectures; (2) we developed a tailored masked prediction model with semantic role-level masking, an increased masking ratio, dynamic N-gram masking, and a novel label embedding mechanism; and (3) we demonstrated the method's effectiveness through rigorous evaluation on four distinct Chinese datasets. Experimental results demonstrated that our method consistently outperforms various traditional augmentation techniques across different data volumes and domains, significantly enhancing downstream text classification performance.