Research On Feature-granularity Tree For Mining Product Reviews

With the explosive growth of the network information, how to find useful informationfrom it comes to a hot research focus. Mining product reviews is to extract the productfeatures, users’ attitudes and judge the emotional polarity, in order to offer referenceinformation for potential users and merchants. However, after analyzing the extracted productfeatures, we find that the granularities of product features which users concern are different.So the paper studies this problem, and the main work as follows:Using the method of label path basin on index, this study finds the path of data area, andextracts the product manual as well as original product reviews. Then define the label rulesand mark reviews artificially, preparing the adequate data for follow-up.This paper proposes a method about how to get the granularity distribution of featuresbased on feature-granularity tree. Firstly, because the category of the feature-groups fromsingle specification file is indistinct, we judge the similarity of feature-records from fromdifferent specification files by using an improved formula of similarity calculation whichimproves the precision of judging the similarity of feature-records. Secondly, restructure thefeature-groups based on similar feature-records. After these, a feature-granularity tree is builtaccording to the new feature-groups. Secondly, restructure the feature-groups based on similarfeature-records. After these, a feature-granularity tree is built according to the newfeature-groups. The feature-records come from the same product, so the coverage of thefeatures of product is not complete. In order to solve this problem, we select specification filesof several types of products to enrich the feature-granularity tree. At the end, according tosimilarity calculation and tongyicicilin, judge the similarity between the mined features andthe nodes in the feature-granularity tree to locate the mined features in the feature-granularitytree, in order to describe the granularity relabions among the features.Experimental results show the improved formula of similarity calculation raises theprecision of judging the similarity of feature-records, the approach about how to get thegranularity distribution of features based on feature-granularity tree is effective and thefeature-granularity tree built by this paper is practical.