This paper proposes a hybrid multimodal sentiment analysis model designed for analyzing customer reviews in small and medium-sized enterprises (SMEs). Online reviews on platforms such as e-commerce sites often include both textual descriptions and user-uploaded images, providing rich multimodal signals for understanding customer sentiment. The proposed framework integrates contextual textual representations derived from BERT with sequential modeling using BiLSTM and attention mechanisms. Visual features are extracted from review images using a ResNet-50 convolutional neural network. A late-fusion architecture with cross-modal attention is employed to effectively combine textual and visual representations. Experiments conducted on a curated dataset of 10,000 SME customer reviews demonstrate that the proposed model significantly outperforms unimodal approaches and baseline multimodal architectures. The system achieves an accuracy of 92.3% and a macro F1-score of 0.91, highlighting the importance of multimodal alignment for sentiment classification tasks. The results indicate that incorporating visual cues alongside textual signals can substantially improve sentiment analysis performance in real-world SME review scenarios. The framework offers a scalable approach for businesses to better understand customer feedback and enhance product and service quality.
Lin He (Wed,) studied this question.