Imbalanced datasets have uneven distribution of class labels, resulting in one class being significantly overrepresented compared to others. This imbalance creates a bias toward the majority class and is often neglecting the minority class. To address the class imbalance, this research introduces an oversampling approach T5-Corpus (T5C) method, using T5 transformer model and the WordNet lexical database. In this method, POS tagging is performed on the preprocessed dataset, and the sentences are passed into T5 model for paraphrasing. The performance of the paraphraser-based oversampling model, T5C, is evaluated with logistic regression (LR), linear SVC (L-SVC), multinomial NB (M-ND), and random forest (RF) on T5C with resampling and without resampling. The accuracy of T5C is compared with random oversampler and Synthetic Minority Oversampling Technique (SMOTE), and T5C outperformed with an accuracy of 85%. The findings demonstrate that the suggested oversampling technique enhances the text classification models. A streamlined oversampling process with robotic process automation (RPA) is incorporated to automate the generation of augmented training data and there by improves the accuracy and generalization of text classification algorithms.