Research on Multimodal Sentiment Analysis Technique Based on Global Attention Mechanism

Emotion analysis, as the most active research field at present, has gradually moved from single mode to multi-mode. The purpose is to identify and understand human emotion state more accurately and analyze and understand complex emotional information more comprehensively by comprehensively utilizing the information in multiple modes (such as language, vision, audio, etc.). It has been developed to develop complex feature fusion techniques. At present, researchers are committed to the development of complex feature fusion techniques. However, due to the differences in the dimensions of different modes, the fusion process faces many challenges, which becomes the bottleneck of achieving efficient sentiment analysis. Aiming at this challenge, this paper proposes a solution of sentiment analysis based on global perception ability to effectively improve the fusion effect between modes and improve the accuracy of sentiment analysis. Based on the MISA framework, we optimized the mode processing method, projected each mode into two identical intervals of complementing each other. By introducing the strategies of invariant mode and specific mode, we solved the problem of commonness and characteristics in multi-mode fusion, reduced the gap between modes, and retained the unique characteristics of each mode. In doing so, the Global View Attention Mechanism (GVT) is combined to better capture long-distance dependencies and global features, enabling more accurate prediction of sentiment analysis tasks. Experimental results show that the optimized model performs significantly better than the existing baseline methods on MOSI and MOSEI public datasets.