Research On Analysis Method Of Social Media Sentiment Based On Text And Image Feature Fusion

Sentiment analysis has always been an important research field,which plays an important role in e-commerce platform,public opinion analysis,personalized recommendation,network monitoring and other aspects.However,the more mature research at present is only based on a single mode of text.With the development of Internet technology and the promotion of social media,the expression of users' sentiment and the evaluation of things in social media are not limited to text.People increasingly like to express their personal views,comment on an event and express their sentiment through text,images and other data forms on social media.The sentimental information of images in social media is more abundant than that of texts.However,if only the sentimental information of images is considered,the background information of the context will be ignored,leading to inaccurate sentiment recognition.Therefore,image and text features are fused in this paper to achieve combin sentiment analysis.In the process of text sentimental analysis,in order to retain the colloquial and daily features of social network text expression,the collected data sets of network hot words,popular words,popular words in Sogou input method and Weibo data after Jieba word segmentation are combined with Chinese Wikipedia thesaurus to input into word2vec for word vector training,and then establish the Self-attention+BiLSTM sentiment classification model,keep social media texts classification of sentimental expression characteristics,and then improve the classification accuracy of the model.In the process of image sentimental analysis,the color features and LBP texture features of the image are extracted,and the depth semantic features are extracted by VGG19 network model.The Lp-norm multi-core learning is used to fuse multiple features of the image for image sentiment recognition.Compared with the model which only uses neural network for image sentiment recognition,this method can alleviate the semantic gap between image bottom features and deep semantic features,and effectively improve the classification accuracy.For the fusion of image and text features,the text and image feature layer and feature fusion decision makers respectively,and the fused features are input into the classifier to realize sentiment classification.The experimental comparison with the sentiment classification of text and image shows that the sentiment classification accuracy of the fusion of the two features is higher,and the accuracy rate reaches 91.47%.Compared with the single use of text classification results improved by 3.30%,single use of image classification results improved by 4.94%.