Research of Models in the Field of Multimodal Sentiment Analysis

As social media is growing, more and more users tend to attach a picture with their posts to convey their tendency. So, it is necessary to find a way to realize the sentiment tendency to label the user, which can act as a reference when the system pushes personal information for the user. This article reviews the articles that make big contributions to Multimodal Sentiment Analysis and compares two types of multiply model sentiment analysis models, Entity-Sensitive Attention Fusion Network (ESAFN) and Multi-Interactive Memory Network (MIMN), from text process, picture process, the structure of the model, the generation of the dataset, and the result. The results indicate that ESAFN outperforms MIMN in both accuracy and MACRO-F1, suggesting that ESAFN demonstrates better stability and reliability across different datasets, potentially due to its more consistent ranking standards and data handling approach. This article also points out some limitations of these two models in model structure and dataset correctly.