Research on sarcasm recognition in text data is becoming more and more important since sarcastic content is so common in internet communication. This study uses multi-objective genetic algorithm-based augmentation and thorough data pre-processing on a variety of datasets to overcome issues with the small dataset and class imbalance in the sarcasm detections. This work presents a new pipeline for multi-objective genetic algorithm augmentation of sarcastic data. During training, the suggested approach inverts the labels on the synthetic and original data. The generator receives input from this label inversion, enabling it to produce high-quality of data that close resembles original distributions. Interestingly, suggested system performs similarly to a normal multi-objective genetic algorithm, demonstrating its strong effectiveness in enhancing textual data. The examination of multiple datasets reveals the complex effects of augmentation on model performance, along with helpful advice for preserving a fine balance between artificial and real data. The methodological framework includes both based augmentation and thorough data pre-processing. Overall, our suggested technique’s F1-score performs better than the synonym substitution augmentation technique’s. The experiments showed that the F1-score increased by 0.067–1.053%, and that the adoption of standard led to a 2.89% increasing in the F1-score. The suggested method performed better than traditional multi-objective genetic algorithm and showed comparable performance, highlighting its effectiveness in text data augmentation.