Data-driven machine learning methods have been proven highly successful in predicting glass properties, but hampered when dealing with small datasets, such as oxynitride glasses with excellent mechanical properties and chemical stability. Here, a data augmentation method based on the Wasserstein Generative Adversarial Network with Gradient Penalty (GP) and Content Constraint Penalty (CP) terms, a generative deep-learning model via the adversarial training of a generator and a discriminator, was established, in which the GP and CP terms ensure training stability and the physical rationality of the generated samples. The results indicate that the generated samples improve the performance of the oxynitride glass composition-property models trained with the XGBoost algorithm in terms of prediction accuracy and generalization capability. Furthermore, the augmented models outperform the general glass prediction model, GlassNet, over 101 experimental samples not included in the training datasets. Based on SHAP's single feature analysis and feature interaction analysis, the interpretability study further sheds light on the contributions of elements and the interactive effects of element pairs on the properties of oxynitride glasses. These achievements not only provide reliable models for the composition-property studies of oxynitride glasses but also offer a novel strategy for developing high-performance data-driven models under data scarcity scenarios.