Research On Web-based Opinion Analysis For Stock Reviews

In the securities trading, the swapper acts according to the information which one has to make the investment decision, and the information is one of the key factors which decides the income. At present in China's stock market the individual investors account for above 90%. However, because of their limited capability of searching and analyzing information, they obviously are in the information inferiority position in the stock market. The purpose of this paper is to provide the service of stock reviews'opinion analysis by improving the algorithms'recall and precision which will help him to obtain prompt, comprehensive and accurate securities information and provide the information service support for their decision-making.The main task of this paper is:(1) An information extraction method without template based on CRF is applied. This paper introduces several popular Web information extraction technology, and carries on the analysis and contrast to their technical characteristic and the applicable scope. The extraction goal of this system is mainly the half structured data in the network, and the extraction technology is requested to have the effectiveness and the easy extension, so the information extraction method without template based on CRF and an extraction method based on DOM are selected to implement the extraction module of the negotiable securities information service system.(2) The positive and negative judgment of stock reviews is studied. This paper proposes an improved approach with stock reviews'discourse structure and SVM algorithms to classify the sentiment of the stock reviews which decides whether a review is positive, neutral or negative. After the analysis of the discourse structure of stock reviews, it extracts the title and the forecasting sentences and constructs Title Classifier and Body Classifier using SVM algorithm to classify reviews into three categories (positive, neutral, and negative). To compare with PMI algorithm and SVM algorithm not based on the discourse structure, recall and precision of this method increase nearly 10%, reaching 88.0% and 86.8% respectively. (3) An approach based on pattern for opinion classification of stock recommendations is proposed in this paper. This approach takes advantage of the discourse structure, syntactical features and lexical features to identify and extract forecasting sentences. In this paper, we adopt a semi-supervised statistical approach based on features to construct a sentiment lexicon. Further, rule modes are obtained by the sequence of tokens concatenated by the ^ operator. Then use the patterns to determine all the forecasting sentences'polarities in a text. Finally, the text's polarity is calculated by forecasting sentences'polarities and weights. Experimental results on large-scale data show that the method is effective and the precision reaches 91.64%.(4) A stock reviews'opinion analysis system is designed. On the base of system's overall demand analysis, this paper sets up a business model which integrates the collection, processing, service functions. This system can extract specific information from the vastness of the Internet securities information and provide users stock reviews'opinion analysis, trend analysis, presentation production services and so on.Features and innovations of this paper are to apply the method of opinion analysis to the field of stock reviews. With the analysis of discourse structure, improve the SVM algorithm in the field of stock reviews and design stock reviews'opinion analysis algorithm based on patterns. Finally combining the advantages of both methods, propose a combined approach that recall and precision of this approach reaches 90.2% and 84.8% respectively. This paper provides a solid data base for the future analysis and study.