Research On Sentiment Classification And Opinion Mining Technique Of Online Reviews

With the rapid development of internet technology, internet has become a very important source from which people obtain information. More and more netizens publish opinions and comments on social events, public figures, products etc. through various channels. Analyzing and mining these massive comment resources, we can recognize emotional tendencies of the online reviews posted by users that make us better understanding of the consuming habits of various users, judge people's attitudes towards the hot topic and analyze the trend of public opinion. It is very significant to study the sentiment classification problem for government, enterprises and personage.This paper has conducted the research on sentiment classification and opinion mining technique of online reviews, including: research on comparation of word similarity algorithm, research on comparation of sentiment classification technique based on machine learning, research on two-stage sentiment classification based on semantic and machine learning, and research on opinion mining method based on semantic, the main research contents are listed as follows:(1) By researching traditional word similarity algorithm, we propose to use an improved word similarity algorithm based on Hownet which can recognize OOV easily to compute sentiment words' emotional tendencies so that it can improve the accuracy of words similarity calculation greatly. Finally, we can compute user sentiment words' emotional polarity and classify the attribute characteristics words.(2) In machine learning method, for NB, KNN and SVM classification algorithms, we do experiment which a single variable changed that include the method of changing feature selection, the method of changing weight calculation and the method of changing feature extraction number. Therefore we can get the best classification scheme of each classification algorithms.(3) In order to overcome weak completeness of semantic dictionary and the problem of needing huge training set of manual tagging, we propose a two-stage sentiment classification method based on semantic and machine learning. For testing set, firstly using the method based on semantic to mark and rank all the texts. By setting the document proportion, we regard the texts which have clear emotional polarity as the classified training set of the second stage and regard the texts which have unclear emotional polarity as the unclassified testing set of the second stage. Then using machine learning method, considering the best combination of feature selection method, weight calculation method and feature extraction number to experiment. The experimental results demonstrate that the two-stage classification method is effective.(4) Utilizing the method based on semantic to classify sentiment and extract opinion of reviews. Firstly, we establish various semantic dictionaries include basic sentiment dictionary, user dictionary, degree adverb dictionary, negative dictionary, dynamic sentiment dictionary and so on. Then using words similarity algorithm to build user sentiment dictionary and attribute characteristics dictionary, constructing general semantic rules and other special processing rules according to the relationship with the words and grammatical structure. For testing set, computing emotional scores of phrase and the scores are summed to judge the review's final emotional tendencies and each opinion attribute's emotional tendencies. Experimental results show that the method has higher classification accuracy and the results of opinion mining are corresponding to the actual situation, the method has great application value.