Enhancing emotion classification on the ISEAR dataset using fine-tuning and data augmentation with hybrid transformer models

Emotion detection is a natural language processing task used in many applications, including customer support, human mental disorder identification, and analysis of social platforms. This study examines how data augmentation methods can be combined with convolutional neural networks (CNNs) and transformer models to improve emotion detection on the International Survey on Emotion Antecedents and Reactions (ISEAR) dataset affected by scarce labeled data and class imbalances. The study evaluates the performance outputs between five transformer models: Electra and XLNet with RoBERTa and T5 along with DeBERTa, while processing anger, joy, sadness, fear, disgust, guilt, and shame emotions. Combining DeBERTa-v3-large with CNN achieves an accuracy rate of 94.94%, the top result for identifying Shame and Disgust classification. The XLNet-base and Electra-base-discriminator produce results at 93% accuracy, but the T5-base displays notably lower results at 69%. Implementing CNN layers with transformer models like Electra and RoBERTa improves model performance, particularly in precision and recall measurements. The performance metrics of Electra Base + CNN surpass those measured in the baseline Electra Base model through better precision and recall values. Such models become stronger at identifying local dependencies when CNN techniques are integrated, thus boosting their capacity to detect emotional nuances. Synonym replacement methods extended the training data to overcome class imbalance problems while minimizing the occurrence of overfitting. This technique improved the model’s generalization, producing higher classification accuracy. The research shows that DeBERTa-v3-large + CNN performs optimally as an emotion classifier since it demonstrated superior precision and steadfast emotional identification among alternative models. The research adds knowledge to developing resistant emotion recognition algorithms, particularly in cases where training data availability is restricted. It also creates new pathways for improved augmentation practices involving back-translation and generative adversarial networks (GANs).