Self-Explaining Counterfactual Data Augmentation for NLP

Counterfactual Data Augmentation (CDA) has proven effective in enhancing robustness and interpretability of Natural Language Processing (NLP) models. However, existing CDA methods often rely on post-hoc interpretability tools or manual annotations, which limit scalability and increase computational overhead. In this paper, we propose TCAD (Token-level Counterfactual Auto-augmentation with self-Explaining Decoder), a fully automated framework that integrates a self-explaining module into the T5 model to extract token-level saliency scores without external tools. These scores guide minimal token masking, after which the decoder generates label-consistent counterfactual texts. The generated samples are filtered for consistency and then used to iteratively enhance model training. Extensive experiments on Natural Language Inference (NLI) and Aspect-Based Sentiment Analysis (ATSA) demonstrate that TCAD achieves significant improvements in generalization, diversity, and label control. Moreover, compared to prior CDA methods, TCAD offers lower computational cost and stronger cross-task adaptability, making it a practical solution for scalable and interpretable data augmentation in NLP.