Enhanced Hybrid Emotion Recognition System: Addressing Class Imbalance and Emotion Hierarchies through Feature-Rich Deep Learning

Abstract

The task of detecting emotions at a detailed level in written text continues to be difficult because of three main issues which include the large difference between the number of examples in each category and insufficient features and intricate connections between different emotions. The current transformer models which analyze text data achieve unsatisfactory results by producing F1 scores which are less than 0.54 when evaluated on the GoEmotions benchmark. The research solves these problems by using a three-stage emotion recognition system which identifies 28 different emotions. The research introduces an EmotionAwareRebalancer system which uses transformer-based text hybridization to create realistic synthetic data that solves class imbalance problems by reducing the 81:1 ratio to 4:1. The research presents three main contributions which include a 172-dimensional feature extraction system that combines contextual embeddings with linguistic markers and lexicon-based signals and stylistic patterns and a deep learning system that uses multi-level attention mechanisms and hierarchical loss functions to model psychological emotion relationships. The research method uses DistilRoBERTa embeddings with cosine similarity thresholds (0.85) to preserve semantic meaning in synthetic data generation and trains the model with focal loss (γ=2) and hierarchy-constrained regularization. The experimental results show that the proposed method reaches 0.624 F1 score on test data which surpasses RoBERTa-large (0.503) by 24.1% and EmoBERT (0.523) by 19.3%. The results show that minority class performance reaches significant improvement because grief increases from 0.21 to 0.541 F1 and relief increases from 0.18 to 0.540 F1. The ablation study demonstrates that model rebalancing leads to a 0.066 F1 score improvement and the hierarchical loss function adds another 0.081 F1 score improvement. The framework shows how it can be used in real-world applications to monitor mental health and analyze customer feelings and develop human-computer interfaces which need advanced emotional processing capabilities.