Research On Textual Opinion Summarization

Opinion summarization aims to concentrate and refine the text data so as to generate asummary of the text regarding the expressed opinion. It makes the users easier to read andunderstand the content of the opinion text. This study focuses on multi-document opinionsummarization. The key issues of our research are summarized as follows:1) We collect and annotate a Chinese and English multi-document corpus on productreviews, which provides a basic resource for the research on textual opinionsummarization.2) We adopt opinion-based unsupervised models to generate opinion summarization inthis paper. Specifically, we incorporate the opinion informantion into the PageRank modeland Cluster_HITS model to consider the impacts of both topic and opinion relations amongthe reviews. Empirical studies on the both Chinese and English corpus demonstrate that theproposed method apparently outperforms existing approaches in terms of ROUGEmeasurement.3) We construct a quality-based unsupervised opinion summarization method byincorporating the review quality informantion into PageRank model and Cluster_HITSmodel to fully consider the impact of the reviews quality on opinion summarization.Empirical studies on the English corpus demonstrate that quality information is effectivefor multi-document opinion summarization.4) We propose a supervised learning method that explores the maximum entropymethod to generate the opinion summarization of product reviews. In this method, theopinion summarization is treated as a binary classification problem. The internal feature oftext, topic feature, opinion feature and quality feature are explored in the maximumentropy model to get the opinion summarization of reviews. Experiment on both Chinesecorpus and English corpus show that the supervised learning method can be more effective in integrating the theme information of the text, the opinion information of the text and thequality information of the text.