Research On Sentiment Analysis In Product Reviews

With the rapid development of Internet, online shopping, the business modelhas entered the lives of ordinary people. More and more users begin to buy productsand make comments through the Internet. Product review analysis is to mineinformation in product reviews, so that potential consumers could understand thereputation of product and manufacturers can improve products quality and servicesin order to make the right marketing decisions.Compared with current sentence or chapter level emotion classificationresearch, users expect more on fine-grained attribute-level opinion mining results. Inthe existing sentiment analysis research, flexibility and scalability of extractionmethod based on the rules need to be improved, while the attribute extraction basedon HMM or CRFs can not handle long-distance emotional well-dependent elementsproblems. In addition, the current product reviews sentiment analysis often overlooktwo types of special emotional sentences which are compared emotional sentencesand negative emotional sentences, thus affecting the performance of sentimentanalysis.In this paper, the following aspects is contained. First, the product attributesextraction algorithm is improved. The extraction task is converted into a sequencelabeling problem. The algorithm introduces POS of word and Words dependenciesin syntactic dependency tree which contained sequential structure, conjunctionsstructure and syntactic structure of a word. In fine-grained sentiment analysis corpus(CUHK-HIT Opinmine) experimented to verify the the model. Second, we computesimilarity between instances of digital cameras ontology nodes and attribute wordand similarity between their associated opinion word， then use these values asfeatures to classify the instance to expand digital cameras ontology nodes. Third,comparative sentences were analyzed, including the identification of comparativesentences using a rule-based method and CSR combined with statistical algorithms.Then using the results of comparative sentences recognition extraced the element ofcomparative sentence and using the features of the comparative sentence analyzedthe sentiment of sentence. Fourth, this paper using method based on rule andnegative word dictionary analysised the clause opinion polarity, then caculated thesentiment of sentence.Contribution of this paper is as follows: improved fine-grained extractionalgorithm, making extraction precision and recall rate has increased4.8%and3.5%.The approach based on CSR combined statistical methods to identify CSRcomparative sentences obtained the result that recall rate is79.3%and precision is rate87.0%in COAE2012public data sets. In the sentiment analysis of negativesentence task, the approach based on rules is adopted, experiment results showthat the result is improved.