Image And Text Fusion Sentiment Analysis For Social Media

As the main channels for people to transmit and exchange information,social platforms such as Weibo,Xiaohongshu and Twitter contain massive data resources.In order to better understand people’s true feelings about social hot topics,national policies,and product services,sentiment analysis has become an important research field in academia.However,most of the previous research on sentiment analysis was conducted on a single modality,which cannot be directly applied to the growing multimodal data.Therefore,this thesis conducts research on multimodal sentiment analysis of text and text.This thesis first proposes a graphic multi-modal sentiment analysis model AMNMD based on attention mechanism and momentum distillation technology.The model first uses ViT and RoBERTa to extract image and text features respectively,and then uses these features to interact with graphic and text information through the modal interaction module,and finally inputs the interacted features to the modal fusion module for graphic and text information fusion.to extract fused features for sentiment classification.In addition,in order to further improve the performance of the model,we introduce the momentum distillation technique to the model.Good experimental results obtained on three public datasets validate the effectiveness of the AMNMD model.For finer-grained multi-modal aspect-level sentiment analysis tasks of text and graphics,this thesis proposes an improved model AMNMDMABSA of the AMNMD model and a multi-interaction network model MINA based on the attention mechanism.Among them,the MINA model can fully carry out the information interaction between aspects and texts and between aspects and images when extracting features,and use multiple iterations to update aspect features to indirectly realize the information interaction between images and texts.Good experimental results obtained on two public datasets validate the effectiveness of these two models.Based on the AMNMD model,this thesis designs a social mediaoriented image and text fusion sentiment analysis system.This system can collect tweets containing specified search words from Twitter within a certain time range according to the user’s input on the webpage,perform sentiment analysis on them,and present the analysis results to the user in a visual interface.