A Hybrid Deep Learning Framework for Sentiment Analysis of Social Media’s Heterogeneous Data

Abstract

Sentiment analysis of social contents posted by people has become an essential task for understanding public opinion, customer feedback, and brand perception. However, the informal nature of Twitter (X) posts, short textual contents and the visual information make sentiment analysis a challenging task. This paper proposes a deep learning model for sentiment analysis of Twitter data using CNN. The textual sentiment analysis has been done using LSTM and CNN has been used for graphical data. The proposed framework uses inclusive preprocessing, tokenization, feature extraction, word embedding and deep feature learning to improve the results of while reducing noise and training complexity. Our experiment evaluation demonstrates that the LSTM model achieves 99% training accuracy and 94% validation accuracy for textual sentiment analysis. The CNN model improves the graphical sentiment analysis by improving overall accuracy from 92.48 to 94.92% compared to conventional CNN approach. The proposed framework also improves precision, recall, and F1-score across multiple sentiment classes. The results demonstrate that the combining deep learning architecture for textual and graphical modalities provides a robust and scalable solution for multimodal sentiment analysis using twitter data.