A Systematic Literature Review on Multimodal Sentiment Analysis Trends, Challenges and Opportunities

Multimodal Sentiment Analysis (MSA) has gained significant attention in artificial intelligence for its ability to interpret human emotions and opinions through the integration of text, images, and video data. Traditional sentiment analysis techniques, which rely on unimodal data (text only), often fall short in capturing the complexities of human expression. This study presents a Systematic Literature Review (SLR) of recent advancements in multimodal sentiment analysis, with a particular focus on the use of visual content, such as images and videos. The review includes 10 Scopus-indexed journal articles from a pool of approximately 400 publications, identified through specific keywords related to sentiment analysis and visual content. The analysis employs the PICOC (Population, Intervention, Comparison, Outcomes and Context) framework to categorize and assess key methods and algorithms, including fusion models and transformer-based architectures, which have been employed to enhance sentiment classification. Findings indicate that multimodal models, such as Contrastive Learning and Multi-Layer Fusion (CLMLF), IMCN, and TBJE, significantly outperform traditional unimodal approaches, achieving accuracies up to 96.98%. However, challenges remain, particularly in terms of processing complexity, computational costs, and the lack of standardized datasets for comparison. This review highlights the latest trends, evaluates the effectiveness of various multimodal approaches, and discusses future research directions, such as the development of lightweight models and more effective crossmodal fusion strategies. The findings provide valuable insights for researchers seeking to improve the accuracy, efficiency, and interpretability of multimodal sentiment analysis techniques.