MInD: Improving Multimodal Sentiment Analysis via Multimodal Information Disentanglement

Learning effective joint representations has been a central task in multimodal sentiment analysis. Previous methods focus on leveraging the correlations between different modalities and enhancing performance through sophisticated fusion techniques. However, challenges still exist due to the inherent heterogeneity of distinct modalities, which may lead to distributional gap, impeding the full exploitation of inter-modal information and resulting in redundancy and impurity in the information extracted from features. To address this problem, we introduce the Multimodal Information Disentanglement (MInD) approach. MInD decomposes the multimodal inputs into a modality-invariant component, a modality-specific component, and a remnant noise component for each modality through a shared encoder and multiple private encoders. The shared encoder aims to explore the shared information and commonality across modalities, while the private encoders are deployed to capture the distinctive information and characteristic features. These representations thus furnish a comprehensive perspective of the multimodal data, facilitating the fusion process instrumental for subsequent prediction tasks. Furthermore, MInD improves the learned representations by explicitly modeling the task-irrelevant noise in an adversarial manner. Experimental evaluations conducted on benchmark datasets, including CMU-MOSI, CMU-MOSEI, and UR-Funny, demonstrate MInD's superior performance over existing state-of-the-art methods in both multimodal emotion recognition and multimodal humor detection tasks.