Research On Sentiment Analysis Of Product Reviews Based On Domain Ontology And Conditional Random Fields

With the rapid expansion of internet resources such as forums, web sites for reviewing product, people can freely express their opinion or experience on the web. Consequently, interent become to an important source to obtain product reviews. Product reviews expressed by customers contain much important message such as the quality of products and so on. But the huge amount of product reviews is impsible for customers to totally browse; as a result, sentiment analysis of product reviews becomes a hot focus in opinion analysis area. How to obtain the valuable information in product reviews is waiting to be resolved. Sentiment analysis for product reviews contains two tasks:the first task is how to extract the comment-target in product reviews; the second task points to two problems, opinion sentences recognition and how to judge the polarity of opinion sentences. In this paper, the two tasks were mainly discussed on product reveiw in Chinese form Interenet.In comment-target extraction, the method to build Product Reviews in Chinese Ontology was proposed, and was realized on prouduct reviews about mobile phone. Then the senmantic relations between classes were discussed, which were intergrated to define SWRL rules for implicity comment-target extraction. In the same time, CRFs model was used to extract explicity comment-target, and the domian ontology feature and sentiment feature of comment-target were added to the feature set. Experiments results on the corpus of digital area from the task 3 from COAE2011 shows remarkable performance of the method, and the recall approachs to 83%.In the second task, sentiment words from HowNet were used as the seed opinion words, then semi-supervised CRFs model were used to iteratively construct the opinion words; then opinion sentences were recognized whether they contains opinion words; at last, the polarity of opinion sentences are judged according to the senmatic rules based on syntax of chinese language. The proposed method was attended to the evalation of task2 in COAE2011. The results in digital area proves the effectiveness of the method, the precision is 0.5839, the recall is 0.5559, and F-measure is about 0.576245 which are better than the avarage level in the evaluation.