Speaker-Aware Emotion Recognition in Dialogues via SemGloVe- BERT and Graph Attention Networks

Emotion recognition in dialogues play a critical role in understanding human communications, with impactful applications in domains such as customer support, virtual assistants, and mental health monitoring. Traditional deep learning (DL) approaches like LSTM, Bi-GRU, and BiLSTM have shown promise in identifying emotions from text. However, these models often fail to capture long-range contextual dependencies and speaker-level interactions inherent in multi-turn conversations, leading to suboptimal performance. To address these limitations, this study proposes a novel framework that integrates DialogueGCN (Graph Convolutional Networks for Dialogue) with Graph Attention Networks (GAT) for improved emotion recognition. DialogueGCN is specifically designed to model both temporal dynamics and speaker-specific dependencies within a dialogue using graph structures. GAT enhances this representation by assigning varying levels of attention to different nodes, thereby emphasizing more relevant speaker interactions. The proposed model was implemented using Python and evaluated on the Daily Dialog dataset. The architecture outperforms conventional models significantly, achieving 93% accuracy. Compared to existing methods—LSTM (85%), Bi-GRU (87%), BiLSTM (89%), and BERT (75%)—the proposed DialogueGCN+GAT model also demonstrated superior accuracy(93%), precision (94%), recall (99%), and F1-score (97%). These findings validate the strength of graph-based approaches in emotion recognition tasks, particularly in handling complex dialogue structures. The results suggest that the proposed model offers a more context-aware and speaker-sensitive solution, making it highly effective for real-world dialogue systems. Future work aims to extend this model to other multimodal datasets to further evaluate its generalizability and performance.