Beyond Single Emotions: A Dual-Label Classification System for Enhanced Emotion Understanding in NLP Using Transformer-Based Models

Emotion detection in text is a significant subfield of natural language processing (NLP), with applications ranging from sentiment analysis to tracking mental health. Although single-label emotion classification has been extensively researched, multi-label emotion classification, especially dual-label classification, is not well explored. This paper explores dual-label emotion classification theoretically grounded in Plutchik's emotion dyads, suggesting that dyads of emotions are combined to produce more complex emotions. The transformer models Bidirectional Encoder Representations from Transformers (BERT), XLNet, and Robustly Optimized BERT Pre-training Approach (RoBERTa) are used to predict dual emotions on the GoEmotions dataset containing 28 classes of emotions. For better performance additional sentiment features Text Blob and Vader are included. Some emotions, like gratitude and admiration, are relatively easy to classify, but others, like grief and relief, are much more complicated. BERT achieved the highest accuracy (63.81%), while XLNet and RoBERTa performed better for subtle emotions. Overall, this approach gives a balance between emotional complexity and model practicality and a better understanding of the emotional nuance of the text without complicating the model excessively.