Graph-Based Emotion Sense Fusion Model for Contextual Sentiment Interpretation

Abstract

The task of sentiment perception becomes challenging because words obtain several meanings according to contextual meaning and affective tone. Traditional sentiment models fail to take WSD into consideration at the cost of frequent mislabelling of polarities and misunderstanding of emotions. The proposed work is to bridge this gap by presenting an ESF Net - a Graph Based Emotion-Sense Fusion Model that fuses contextual semantics, word sense, and emotional data for holistic sentiment interpretation. To prepare an integrated setting, the authors propose an ESA Dataset containing word sense labels along with emotion classes and sentiment polarity. This architecture creates an Emotion-Sense Graph where each node is a triplet-word, sense, and emotion - edges reflect semantic and affective similarities. A Graph Convolutional Network combines contextualized embeddings from BERT and sense representation from Word-Net into joint sense emotion embeddings. Experimental evaluation shows significant improvements in the classification of emotion and sentiment prediction tasks compared to standard text baselines. Confusion matrices specify better discrimination between emotionally confusable classes like fear-sadness and joy-trust, whereas UMAP projections are shown to reveal clearly separated sentiment clusters. Results confirm the proposition that the integration of sense-level emotional dependencies leads to more interpretable human-aligned sentiment representation. Present research defines a new direction in sentiment analysis by treating emotion and sense disambiguation as interdependent in its place of separate sub-tasks. It gives the basis for explainable affective computing and multimodal emotion research.