Analysis On User Review Information In E-commerce Website Based On POS Pattern Matching

With the development of e-commerce and the growth of the amount of site users, user reviews is explosively growing. User reviews in the e-commerce sites can provide important references for both the potential users who need to make purchase decisions, and the producers who want to improve their products. It is difficult to obtain the really useful information from the huge number of reviews just by manually reading. Thus, realizing the automatic processing of user reviews and then generating effective results, has very important meaning in both research and application fields.Analysis on user reviews includes mining feature words and opinion words form user reviews, and analyzing the sentiment polarity of the reviews. At present, researchers have proposed a lot of methods on analyzing user reviews. However, there are some problems on feature-opinion pair recognition including incorrect collocation relation between feature words and opinion words and not accounting for potential feature words prediction for implicit feature-opinion pairs. In addition, it is necessary to make sentiment lexicon and polarity judgment rules better. This thesis includes three aspects as follows:(1)Extracting feature-opinion pairs from review text based on POS rules matching. Firstly, effective POS patterns are selected from a large number of training review texts by using a method of extending and matching. And then, feature-opinion pairs are extracted from test review texts by using effective POS patterns. Besides, three pruning methods are designed to remove useless feature-opinion pairs.(2)Predicting potential feature words for implicit feature-opinion pairs based on TF-IDF formula. Opinion sentences without a feature word are widely found in user reviews. But most of existing research pays more attention to the extraction of the explicit feature-opinion pairs, ignoring the implicit feature-opinion pairs. This thesis attempts to use a TF-IDF schema to predict a possible feature word for implicit feature-opinion pairs based on a set of explicit feature-opinion pairs.(3)Introducing a method of sentiment polarity analysis on evaluation units based on the extended sentiment lexicons. It is found that a part of adverbs and feature words with polarity will affect the polarity of relative sentiment words except negation words. Some polarity judgment rules are made in consideration of the above factors. What's more, network words and field words are added into sentiment lexicons. Sentiment analysis of evaluation units is based on these sentiment lexicons and judgment rules.Phone reviews are used as a test dataset, and a set of the effective POS patterns are extracted. The experiments on extracting feature-opinion pairs and analyzing sentiment polarity are conducted by the methods proposed. The experiment has indicated the method proposed is effective and feasible.