Sentiment and emotion analysis using pretrained deep learning models

Sentiment analysis has been pivotal in understanding emotional expressions and mental states. This research presents an innovative approach to sentiment analysis using text and image data using pretrained models. The study employs RoBERTa for textual sentiment prediction on Multiclass Emotion Model Dataset. DenseNet201 and MobileNetv2 was also used on Karolinska Directed Emotional Faces (KDEF) dataset as feature extractors and trained on several supervised learning classifiers. RoBERTa achieved an accuracy of 75%. On the image dataset, DenseNet201 and MobileNetv2 achieved 71% and 68% respectively using transfer learning. When used as feature extractor, DenseNet201 achieved 81% and MobileNetv2, 73% with Logistic Regression classifier. The findings of this research highlight the model’s potential in real-world applications, particularly in emotional well-being assessment. This study advances sentiment analysis by integrating both textual and visual data, providing a comprehensive approach to understanding and classifying emotional states. The innovations, contributions to the literature, and managerial significance of this study underscore its relevance and impact in fields such as education, mental health, and beyond.