Research On Sentiment Classification For Microblogging With Text And Image

With the development of Internet,micro-blogging has been acknowledged as one of the most important social network application since 2006.Users can post mini messages using various mobile devices,such as laptops,mobile phones and tablets.Mining valuable information from micro-blogging has drawn many researchers' atten-tion in recent years and sentiment analysis on micro-blogging has been applied in more applications.However,most contributions about sentiment analysis on micro-blogging focused on the features mining from the texts.Few contributions combine texts with images to classify sentiment.The role of images in conveying emotional messages and opinions is recognized as more important than the text.Meanwhile,emotional mining for only images will ignore the context and background,and only considering texts for sentiment classification probably generates poor results.It is not good to take only one factor to do sentiment classification.In the thesis,we study the sentiment classification combining texts with images.The main work is as follows:Firstly,based on the kernel method of multi-view learning technique SVM-2K,correlation between image and text is identified.We take advantage of the correlation to reduce the semantic expression of diversity.The results of experiments on Sina weibo data has shown the effectiveness of our proposed method based on multi-view method SVM-2K,which combines text features and image features to classify sentiment.Secondly,we propose a new feature fusion method based on latent semantic anal-ysis(LSA),which fuses texts feature with image feature in parallel strategy.Seman-tic features of texts and images are generated by single value decomposition(SVD).Experiments results have shown that our method improves the presion of sentiment classification,which validates the effectiveness of our feature fusion method.Thirdly,we propose a new neighborhood classifier based on the characteristic of words limitation and quantities of patterns of micro-blogging.It maps the text and images into the plane of two coordinates,and classifies the sentiment by coordinate distance.Experiments prove that our proposed method performs better than existing methods(NaiveBayes,SVM).