ADVANCING SENTIMENT ANALYSIS OF USER REVIEWS: A HYBRID DEEP LEARNING APPROACH FOR MULTILINGUAL AND MULTIMODAL DATA

In the digitally interconnected age, social media talk and user reviews provide a goldmine of consumer opinion — if we can crack the code properly. Sentiment analysis (SA) is central to breaking this clatter of unstructured content into usable information for e-commerce sites and social applications. Based on a comprehensive survey of 19 recent papers, this work reviews a range of Machine Learning (ML) and Deep Learning (DL) models — from LSTM and CNN to the contextual abilities of BERT and new hybrid architectures. Grounded on this platform, we propose a new hybrid model, the BERT-BiLSTM-Attention (BBA) model, that combines BERT, Bi-LSTM, and self-attention mechanism, specifically designed to address the long-standing issues of multilingual data, textual noise, and class imbalance. The BBA model is also multimodal in sentiment analysis, encompassing text, audio, and visual features to capture the nature of user-generated content. With a mixed-methods methodology, we applied this model to multilingual e-commerce review and Twitter datasets with an impressive 97.5% accuracy. The results highlight the model's greater capacity for sensitivity to subtle semantics and bypassing data abnormalities, which bodes well for its potential in fine-tuning real-world sentiment analysis across platforms.