Nowadays, more and more attentions and intentions concentrate on expressing personal opinions via the public blogs, forums, wiki, etc. Numerous available reviews and comments which cannot be collected manually contain a magnitude of valuable information. Automatical Analysis of these texts will be beneficial for both groups and individuals. Under this background, the paper will study intensively on two most essential issues in text orientation analysis, namely polarity classification and subjective classification. Our main contribution and focus are summarized as follows. (1) Polarity classification of short reviews Because they require feature-independency assumption which is not a fact on polarity classification of short reviews, traditional classification menthods perform poor on these kinds of tasks. Aiming at this problem, the paper proposes a combined-feature-based classification method, in which the object of the opinion is extracted at first, then the object is combined with the opinioned word, and finally the polarity of the short review is classified based on the combined features. The experiment on the polarity classification task for electronic product reviews shows that, the method can get better performance in comparison with the traditional method. (2) Polarity classification of long review The long reviews contain deeper structures, one of which is the sentiment/opinion structure. In order to model this structure, we present a new model named Roof-CRF, which can accommodate various kinds of features related to sentiment/opinion states and their transition. The model classifies sentiments of documents and sentences spontaneously, and the accuracies of the both are improved. (3) Sentiment grading of reviews Viewing the review grading tasks as multi-class classification will neglect the ordinal relationship between the labels. Therefore, we propose a novel Redundant-labeled-CRFs, which can deal with ranking or ordinal regression problems in more proper way. Besides, the model can integrate subjective/objective classification task and opinion grading task and then depress error by making safe decision over all of these subtasks. Experiment shows that the presented method outperforms basic CRF model in review grading. (4) Subjective/Objective classification Facing with the opinion search tasks in TREsC-Blog-07, we present a solution for subjective/objective classification for blog texts. A partial supervised learning method using only positive examples and unlabeled examples is adopted for subjective/objective classification, after selecting better examples for training with an approach of active learning. After that, we fuse relevance and subjectiveness with a Support Vector Regression method. Detailed experiments prove the effectiveness of this solution.
修改评论