Emotion Recognition in Bangla Text: An Ensemble Approach with Data Augmentation Using BanglaBERT and MultiBERT

Recognizing emotions from text is challenging, but transformers have greatly improved Natural Language Processing (NLP), making emotion detection more accurate. In this study, we performed a multi-class classification to recognize human emotion from Bangla text by developing an ensemble model combining the strengths of two models, BanglaBERT and MultiBERT, thereby enhancing the classification performance. We addressed the class imbalance in the dataset by applying back-translation to the minority classes, doubling their size. We evaluated the ensemble model on original and augmented datasets, comparing its performance with individual models. Results show that the ensemble approach significantly improves performance. Without data augmentation, the ensemble model achieves an accuracy of 0.68, with macro precision, recall, and F1-scores of 0.65, 0.59, and 0.61, respectively. After applying back-translation, accuracy improves to 0.74, with macro precision, recall, and F1-scores rising to 0.74. Especially, the "Anger" and "Disgust" classes showed significant F1-score improvements, rising from 0.48 to 0.70 and 0.39 to 0.66.