Article Preview
Top1. Introduction
With the explosive growth of Chinese users of the Internet, there are astronomical websites that provide the Chinese news and Chinese articles. The enormous information will be overwhelming for the Chinese readers. However, news always contain different opinion (e.g., positive opinion or negative opinion) (Godbole et al., 2007), that is, news articles not only convey the information people would like to know but also the sentimental components which may influence the emotions of the readers (Russell 2009) and may also affect the productivity of the readers (Wong and Law 2002). Filtering out the news with bad emotions (negative) become an essential task for the readers who only want to know the global trends without affecting performance of their daily works.
However, Chinese news contain a variety of different topics ranging from the business trends to the global warming issues. Investigating the sentiment of topic-independent Chinese online news is much difficult from the sentiment analysis within the same topic in many aspects. First, language-specific characteristic of the Chinese is that there is no delimiter within the Chinese words. When processing Chinese, the information need to be segmented at the beginning that will be affected by the accuracy of the segmenter. Second, Chinese sentence may not comply with the restrict grammar which makes the articles containing more complicated sentence difficult for analyzing the word usage. Finally, the online news covering a variety of topics containing different word usage makes the sentiment analysis of the topic-independent documents (the online news) more difficult.
In this paper, we develop a framework to deal with the sentiment detection of online topic-independent Chinese news. The framework consists of two components: feature selection and sentiment classification. We propose new features to represent the sentimental documents, and employ the support vector machine (SVM) to predict the document sentiments.
The rest of the paper is organized as follows. Section 2 gives the related literatures to our work. In Section 3, we detail each feature we used, and Section 4 provides experimental results to show the effectiveness of our features. Section 5 gives a conclusion and future works.