Research And Implementation Of Multimodal Transformer Sentiment Analysis Technology

Sentiment is a complex and rich attribute.In reality,there are many ways to obtain human sentiment.Sentiment analysis is to analyze human sentiment tendency through data analysis,while multimodal sentiment analysis is to analyze and judge sentiment by using two or more singlemodal information.With the prosperity and development of multimedia era,the ways of human expressing sentiment are gradually diversified.Therefore,the study of multimodal sentiment analysis is of great significance not only to human life but also to social development.The current multimodal sentiment analysis has strict requirements on multimodal data,and is susceptible to problems such as lack of modal information,lack of multimodal data,and label noise.In addition,in terms of multi-modal information fusion,problems such as long-distance dependence,parallel computation,and more effective fusion methods have always been major bottlenecks to be studied and solved in the field of multi-modal sentiment analysis.Therefore,based on the existing research results of multimodal sentiment analysis,the following innovations are made:(1)Based on the model structure and design idea of Transformer,a multimodal sentiment analysis method is proposed.This method is used to extract and fuse features from text,image and audio modes,and is applied to specific classification tasks.This method can deal with the problems of unaligned modes,long-term dependence of modal elements and incomplete multi-modal data,and has good universality and robustness.(2)For multimodal emotional areas exist insufficient amount of data,cross modal data inconsistency,labels,noise and other issues,this article is through the label smoothing,data augmentation and application multi-task learning strategy,to improve the impact of the above problems,with an increase model generalization ability,alleviate model fitting effect.The proposed method is validated by experiments on CH-SIMS dataset.Compared with seven baseline models,the proposed method achieves the best performance on multiple classification and regression tasks.This paper designs and implements a multi-modal Transformer sentiment analysis system.The system provides a graphical interface for users to operate,and each processing unit is divided into modules according to their functions,including system front-end module,input preprocessing module,multi-modal sentiment analysis module and database storage module.In addition,the function and performance of the system are tested and analyzed.