Multimodal Sentiment Analysis (MSA) plays an important role in understanding social media content; however, existing methods often struggle with the heterogeneity and complex interactions between images and text. These challenges include inter-modal information asymmetry, insufficient feature fusion, and noise interference, which collectively limit robustness and accuracy. To address these issues, we propose a multimodal sentiment classification model termed Multi-Branch Cross-Modal Cross-Attention Gating (MB-CMCAG). The model first incorporates a Transformer-based image caption generation module to convert raw images into semantically rich auxiliary textual descriptions, which complement the original text and form paired textual inputs with enhanced visual semantics. To capture multi-source features, MB-CMCAG adopts a dual-branch feature extraction architecture: the visual branch encodes images using a Vision Transformer (ViT), while the textual branch encodes text with Bidirectional Encoder Representations from Transformers (BERT); a Contrastive Language-Image Pre-training (CLIP) model is also introduced for joint image–text feature extraction. To exploit cross-modal correlations and enable hierarchical fusion, we construct a cross-modal attention module that supports bidirectional information flow from image to text and from text to image. Building on this, a cross-modal gated mechanism is introduced to selectively regulate the transmission and aggregation of features from different sources, thereby improving noise suppression and sentiment sensitivity. Experimental results on the public MVSA-Single and MVSA-Multiple datasets show that MB-CMCAG achieves accuracies of 76.38% and 73.87%, respectively, outperforming existing baselines by a clear margin in image–text multimodal sentiment classification.
Huang et al. (Thu,) studied this question.