<p>User reviews are usually imbalanced data, which makes their classification challenging. Generative Adversarial Networks (GANs) have proven successful in synthetic data generation for addressing data scarcity in minority classes, but they come with the challenge of text being discrete. Despite their potentials, the use of generative methods in resolving the problem of imbalanced user reviews is rarely studied. In this research, we propose a framework to assess the value of GAN-based techniques in overcoming the challenge of imbalance in textual data considering different GAN architectures as well as traditional methods such as SMOTE and ADASYN. The framework also allows comparing the effectiveness of each model in targeted data augmentation to improve classification performance. We also investigate the quality and diversity of text produced by GANs, while looking at the impact on classification performance. The findings demonstrate enhanced text classification accuracy when utilizing GANs with an average of 2.48% improvement. Moreover, in most cases when balancing with GANs, higher quality and higher diversity in the generated text, leads to greater precision and recall of the minority class, respectively. This study not only highlights the benefits of GAN architectures for realistic synthetic text generation, but also provides new directions for studying the effect of GAN structure on the generated text properties.</p>