Multiomdal-MLP: Enhanced multimodal sentiment analysis using feature vectors and binding learning mechanisms

The key challenge of multimodal emotion analysis is to deal with the heterogeneity of different modes, such as text, audio and image, which requires a flexible and efficient fusion method. This paper presents a Multimodal-MLP model. The core of this model lies in its unique MLP-Communicate module, which can flexibly mix features in length, mode and time, and effectively promote the interaction and fusion of in-depth information among multi-modal data. In addition, the feature vector CLS is introduced to enhance the precision of emotion modeling and alleviate the problem of data sparsity. Furthermore, in order to improve the model's ability to integrate multi-modal information, bimodal binding strategies are designed. Extensive experiments on benchmark datasets CMU-MOSI and CMU-MOSEI show that our model outperforms the most advanced methods.