TomBERT-based local-global multimodal sentiment classification for social platforms
With the rise of the self-media era,the emotions and attitudes expressed by users on social media have become im-portant sources of information reflecting the public's emotions.However,existing multimodal sentiment classification methods al-ways overlook scene factors outside the target when dealing with text and image fusions,which affects the accuracy of sentiment classification.Aiming at this problem,a local-global multimodal sentiment classification model for social platform based on TomBERT is proposed.This model uses the TomBERT model as the basic architecture,divides the input information into two parts:the main body(local)and the scene(global)for image and text matching,and obtains the final multimodal hidden representation through a multimodal encoder for classification,fully considering the correlation between subject information and scene informa-tion,using scene factors to enhance subject features and assist in sentiment classification.Experimental results have shown that local-global multimodal sentiment classification model for social platform based on TomBERT,compared to traditional methods,not only captures the relationships between modalities but also comprehensively considers the influence of subjects and scenes,im-proving the accuracy of sentiment classification.
万方数据
维普
