A multimodal sentiment analysis model based on semantic enrichment

At present, more and more multimodal pre-training models are applied to multimodal tasks, especially the sentiment analysis task of graphic modalities. However, people have not considered the problem of information richness when using multimodal pre-trained models for image semantic representation, So we propose a simple semantic enrichment method with flexible adjustment mechanism, and build a sentiment analysis model of graphic modalities to address above problem. It has been verified that the accuracy and F1 score of MVSA-multi and MVSA-single datasets are improved by 1 to 3 percentage points compared with the existing mainstream models.