Research On Sentiment Analysis For Comments Of Online News

With the rapid development of Internet and the explosive growth of different types of social media, online news is an important type of information that attracts billions of users to read and express their opinions. Users often write comments in subjective emotion categories such as sadness, happiness and anger. Such emotions not only can help understand the preferences and perspectives of individual users, and therefore may facilitate online publishers to provide users with more relevant services, but also can help the government monitor public opinion and make public administration decision effectively. It is a challenging and practical research problem to automatically assign emotion tags to users’comments of online news.This paper tackles the task of predicting emotion for comments of online news by machine learning technology, including emotion tagging for comments of online news with single information sources, emotion tagging for comments of online news by Meta-classification with heterogeneous information sources and cross-domain cross-category emotion tagging for comments of online news. The main contributions and innovations include:Firstly, this paper proposes several approaches of predicting emotions for comments of online news with single information sources. In particular, this paper proposed to exploit single information sources including the content of the comments, the content of the news articles and the emotion votes of news articles generated by users, via content-based classifiers or statistical model to automatically assign emotion tags to users’ comments of online news. This paper further proposes an approach for emotion tagging for comments of online news relaxing the common assumption that emotions are independent with each other. Pearson Correlation Coefficient with user-generated emotion votes of news articles is utilized to measure the dependencies between emotions in the comments. Incorporating the emotion dependency into the discriminative model could facilitate emotion prediction for the comments.Empirical studies on two datasets from real online news services show the performance of all proposed methods, demonstrate the best effectiveness of the proposed model with the content of comments among all single information sources, and prove the benefits of emotion dependency for predicting emotion for comments of online news.Secondly, this paper proposes two approaches of emotion tagging for comments of online news with heterogeneous information sources. A novel research to exploit heterogeneous information sources such as the content of the comments, the content of the news articles and the emotion votes of news articles generated by users to bring their own insights on emotion tagging for the comments of online news is conducted. A basic Meta classification approach is proposed to merge and weight heterogeneous information sources in the context of emotion tagging for comments. Beyond just learning a fixed combination strategy for all comments, this paper further proposes a latent Meta classifier which automatically identifies different latent class variables underlying the observed comments and builds a regression model for each class of comments. In consequence, the combination weights of multiple information sources will be adaptive to specific comments according the characteristics of the comments. Empirical studies on two datasets from real online news services demonstrate the effectiveness of exploiting heterogeneous information sources and the proposed Meta classification approaches.Finally, this paper proposes a general solution of sentiment analysis for comments of online news in multiple domains to address the problem that labeling work of emotion tagging is labor-intensive and different domains of online news need different emotion classifiers and corresponding labeled data. This paper tackles the task of predicting emotions for comments of cross-domain online news by utilizing a relatively small amount of labeled data from a target domain while abundant labeled data available in a related but different source domain. In particular, two approaches have been proposed when the source domain and target domain share the same set of emotion categories or use different categories. A novel probabilistic framework is designed for modeling the relationship between different but related sets of emotion categories in source and target domains. The probabilistic framework transfers knowledge across different domains through the relationship to facilitate the emotion tagging in the target domain. An extensive set of experimental results on two datasets from popular online news services demonstrates the effectiveness of our proposed models in cross-domain emotion tagging for comments of online news in both the cases of using the same emotion categories and different categories in source and target domains.