Research On Sentiment Analysis Based On Web Text Mining

With the extremely rapid development of social networks, social network sites or community sites, such as Sina microblog and Douban, have already been important platform and pathway for releasing opinions and expressing emotions. In general a large number of messages published on social networks contains enrich subjective information, most of which are emotions or opinions that users expressed for substances or phenomena, in consequence mining users’ emotions from social networks data can better understand online users’ behaviors and predict market trends, in other words, mining users’ opinions have crucial value of academic research and practical application.Most of traditional researches on sentiment analysis pay close attention to text data, because subjective text usually consist of words that used to express sentiment, and analyzing sentiment words can apparently reveal users’ emotions. However, non-textual data(such as image, audio, video) can reveal plentiful emotions as well, although the way of expressing emotions differ from textual data. With an increasing number of different types of multimedia content on social networks, sentiment analysis on all kinds of user data have already attracted widely focus and research. Our paper focuses on sentiment analysis for multimedia content(textual and visual data) on social networks, aiming at analyzing users’ sentiment(positive or negative, for or against) by utilizing the relation between textual and visual sentiment. Subjective textual and visual data are both contain emotional information, and that share the same semantic space(namely they are correlative on semantic level), that is, textual and visual sentiment are complementary to each other.The main points of our paper are as follows:(1) Aimed at the problem of sentiment shortage of single text or image, a method based on convolutional neural networks(CNN) was proposed for multimedia sentiment analysis. Image features are combined with text features of different level(word-level, phrase-level and sentence-level) to construct CNN, and obtain performance of sentiment classification on three different semantic level. On the basis above, we consider taking advantage of ensemble classifier to merge three performances into a better one. The model not only mining the relation between textual and visual sentiment features, but also utilizing combined features of different level to enhance the performance of sentiment classification on multimedia content.(2) Aimed at the problem of lacking of training data and data contain plenty of redundancy and noise in the field of image sentiment analysis, a heterogeneous transfer learning method based on collective matrix factorization(CMF) was proposed. The method transfer the knowledge(model parameters) that learned from co-occurrence data in auxiliary domain into the domain of image sentiment analysis, thus images can be reconstructed with the help of model parameters, which result in effectively removing the noise of original images, consequently whose features can achieve a better performance on the task of image sentiment analysis.