This thesis presents the development of an emotion classification pipeline designed to improve multi-label text classification, with a particular focus on addressing the underrepresentation of minority classes. In multi-label tasks, class imbalance often causes models to perform poorly on infrequent labels compared to majority ones. To mitigate this issue, the proposed approach generates synthetic sentences to enrich minority class samples and enhance overall classification performance. The study begins by establishing a baseline model based on the BERT-base-uncased architecture, trained on the original dataset. A data-driven analysis is then conducted to identify the most representative examples of underperforming labels. These examples are used as input for two data augmentation methods: a traditional synonym replacement technique and a large language model based generation approach.