Emotion-Aware Multi-Modal Feature Extraction using CNN-Bi-LSTM with Dynamic Emotional Salience
S Indhumathi, F. Mary Harin Fernandez · 2025
Emotion-aware sentiment analysis has gained significant attention in recent years due to its applications in social media monitoring, mental health assessment, and customer feedback analysis. Traditional unimodal approaches often fail to capture the complex relationships between visual and textual data, leading to suboptimal sentiment classification. To address this limitation, a novel multi-modal feature extraction framework that integrates Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) networks for textual analysis and VGG16 for image feature extraction is proposed. This model incorporates a Dynamic Emotional Salience mechanism to assign adaptive weights to features based on their emotional significance, ensuring a more context-aware representation. The fusion of multi-modal features enhances sentiment classification accuracy by using deep contextual and spatial dependencies. Experimental results demonstrate that approach effectively captures fine-grained emotions, improving overall performance compared to conventional feature fusion techniques. The proposed framework provides a robust solution for emotion-aware sentiment analysis, making it suitable for real-world applications such as social media analytics, opinion mining, and psychological studies. Future enhancements include expanding the dataset diversity and optimizing feature fusion techniques for greater generalizability.