Multimodal sentiment analysis based on exchange fusion

To address the challenges faced by existing sentiment analysis models in handling redundant information, disparities in modality emotion representation, and low cross-modal interaction efficiency, this paper proposes a multimodal sentiment analysis model based on exchange fusion. The model first extracts features from each modality through a modality embedding layer and removes redundant information. Then, in the exchange fusion module, the language modality serves as the dominant modality, interacting layer by layer with the visual and audio modalities, gradually eliminating redundancies and conflicts between modalities while fully leveraging the complementary advantages of each modality. Experimental results show that the model achieves sentiment binary classification accuracy of 88.26%, 87.62%, and 81.18% on the public datasets CMU-MOSI, CMU-MOSEI, and CH-SIMS, respectively, demonstrating excellent performance.