Researeh On Opinion Extraetion Of Chinese Produet Review

With the development of e-commerce and web 2.0, more and more consumers like publishing their own attitudes and views on e-ecommerce websites, forums and blogs after purchasing and using products, these reviews include attitudes about features, function and performance of the product. On one hand, consumers usually consult other people's suggestions before they purchase product to make more sensible decision; on the other hand, manufacturers can also improve their products according to these product reviews. Artificial reading massive reviews is time-consuming and inefficient, it also has hysteretic nature and one-sidedness. Recently, a variety of automatic opinion extraction approaches from unstructured web product reviews have built up a research hotspot.This paper makes deeply research on the construction of sentiment analysis resource and product feature dictionary, the matching approaches between feature words and sentiment words and the judgment criterion of the polarity of product feature words. The main research work is as follows.(1) Using open-source tools named Larbin and Xpath, aiming at mobile phone channel of shopping websites to crawl, and then according to webpage format to extract meta data, finally, we construct product review corpus about mobile phone.(2) In order to construct sentiment analysis resource, we proposed a method of building basic sentiment word dictionary based baidu baike, and then construct domain sentiment dictionary based on conjunction dictionary and dependency relations, network sentiment dictionary, sentiment modified dictionary.(3) In order to extract feature word, we proposed two methods, one is based rule and statistics, the other is base CRF machine learning, the precision and recall of the former reach 0.56 and 0.73 respectively, on the contrary, the precision of the latter is higher, but the recall is not well, which are 0.78 and 0.46 respectively, for the sake of comparison to other researchers, we use method of Hu and Liu on our experimental environment, the experiments show our both methods perform better then Hu and Liu's method.(4) In the area of matching between feature word and sentiment word, judging polarity of feature word, we proposed collocation recognition algorithm based on SVM machine learning, then make comparison with nearest matching algorithm and dependency relations algorithm, the experiment shows the SVM method perform best than the other methods, the precision, recall and F-measure reach 0.83, 0.62 and 0.71 respectively.