Text Orientation Analysis Based On Emotion Chunks Combining With Machine Learning

With the emergence of the virtual community as a new network group form, all kinds of forums ensue. These BBS are filled with plenty of evaluations and views about people's life and surroundings, which contain the speaker's subjective opinions, and express the commentators'emotional orientation. Identifying the emotional orientation of those reviews will help us better understand the attitude and the standpoint of the reviewers, and provide the support of information technology for many fields in the society, such as the feedbacks on products facing to the business, information filtering and public opinion analysis facing to the government administration. Therefore, the text emotional orientation analysis, which has an extensive application prospect, will be the key technology to solve those problems.Using the machine automatic processing means to analyze and estimate the emotional orientation of network comments, has currently become a hot point of the research of Internet intelligent information processing, having great practical value. But due to the particularity of network comments, the original text emotional classification methods cannot obtain the ideal effect. The main reasons are that:a. The text form of network comment is not standard, containing a number of network language, which cannot be correctly segmented; b. There are many statements of objective facts contained in the comments, and those statements which are useless information for the emotion classification, will impact the final results of the classification; c. Extracting the subjective sentences of comments simply will lost some emotional information, which not exists as the subjective sentence form, but contains the tendentiousness information. Hence, how to extract the emotion information of network comment text effectively is the key to improve the emotional classification results. In view of those problems, this paper proposes a method that combining the emotion chunk with machine learning algorithm, and makes an intensive study of the emotional of the network comment text. The SVM algorithm is added into the emotional orientation analysis system, and experiments on the military reviews under different sizes of sample sets are done. The main research of this thesis is as follows:Firstly, aiming at the problem that there are a lot of network language whose form lacking of standardization in the network comments, leading to the original participle software cannot identify the word correctly, this paper sets up a network language dictionary, recovering the nonstandard forms appearing in the text such as pinyin and abbreviation to the standard forms, so as to ensure the accuracy of the segmentation, and also keep the information including the author's emotion better but not be lost, in order to improve the accuracy of classification;Secondly, according to the characteristics of corpus, an emotion dictionary based on military field is established, which realizes the filtering of information with no emotion in the original comment text effectively, in order to reduce the irrelevant information for classification effect;Then, the concept of emotional chunk is proposed, which defines and marks those expression forms which is positive or negative as an emotional chunk, and it can be extracted as the emotional feature, so as to ensure the emotion information can be effectively reservation;And then, combining the emotional chunks with support vector machine, tests on different sizes of sample sets are done, and compare the results with those of classified using KNN classifier. It proves that in larger size of training sample sets, the classification results of SVM classifier. are better than those of KNN classifier. It shows that the method proposed in this paper can effectively improve the emotional classification accuracy of military reviews.Finally, a network text emotional orientation analysis system model for military reviews is designed and realized in this paper. Using this system model, we can judge the military review is positive or negative, check the results, and evaluate the function of the system. This system has passed the test, and it's feasible with a certain accuracy.