Sentiment Analysis for Product Reviews Using a Transformer-Based Approach

Abstrak

Developing multimodal sentiment analysis systems for low-resource languages remains challenging because constructing manually annotated multimodal datasets is both expensive and time-consuming. To address this limitation, this study proposes a lightweight Multimodal Transformer framework that integrates textual, audio, and visual information using pseudo-labeling and cross-modal attention for Indonesian product review analysis. A multimodal dataset was constructed from product review videos through clip-level segmentation, while sentiment labels were generated using an ensemble-based pseudo-labeling strategy to minimize manual annotation efforts. Two cross-modal interaction strategies were investigated, namely a text-centered interaction scheme (CM2) and a full three-modal interaction scheme (CM3). Experimental results demonstrate that CM3 generally achieves superior classification performance, reaching a macro F1-score of 0.789 and an accuracy of 0.876, whereas CM2 provides more stable optimization under moderate training conditions. External evaluation further indicates that both architectures remain affected by distribution shifts, highlighting the importance of dataset diversity for improving model robustness. Overall, the findings show that effective multimodal sentiment analysis depends not only on richer cross-modal interactions but also on selecting an appropriate model complexity for the available training data, providing a practical solution for multimodal sentiment analysis in low-resource settings.

Identifier

• OpenAlex: https://openalex.org/W7169838159
• DOI: https://doi.org/10.67561/jaids.v2i1.3