Research On Aspect-Level Opinion Mining Technology With Product Reviews

With the growing popularity of the Internet and online shopping,product reviews are usually used as an important foundation for purchasing.Hence,it has gradually become a research hotspots that how to obtain information efficiently from numerous product reviews that is more valuable to potential consumers and mershants.In this thesis,the research on aspect-level opinion mining technology of product reviews can obtain the evaluation of product attributes quickly and accurately,and is of great significance for potential consumers to choose and mershants to improve their products.The main contents in this thesis are as follows:1.In view of the problems in most methods of evaluation collocation extraction,such as high workload of manual annotation,dependence on external emotion dictionary and less consideration on the part of speech of verbs,etc,a rule-based evaluation collocation extraction method is proposed.This approach formulates rules by the result of part-ofspeech tagging,dependency parsing and semantic dependency parsing.Firstly,the core collocation is extracted,and then the algorithm for identifying the parallel evaluation targets and the improved algorithm for identifying the modified part are combined with the different part of speech of the core collocation component,which further develop the rule to identify complete evaluation targets and phrases.Experiments are conducted on Chinese mobile phone and hotel reviews dataset,for the former,the precision is 71.95%,the recall is 66.74%,and the F1 value is 69.25%.And for the latter,the precision is 60.42%,the recall is 62.24%,and the F1 value is 61.31%.The results show that the method is effective in evaluating collocation extraction.2.In view of the problems in the existing research on the classification of aspect-level sentiment,for instance,most methods usually need to build a perfect emotional dictionary and judgment rules,the classification performance of the basic machine learning methods need to be improved and so on,an aspect-level sentiment classification method based on emotional lexicon and machine learning is proposed.It selects some positive,neutral and negative emotional words to form emotional lexicon,and takes the proportion of each category in the mutual information as the weight of the classification probability in the basic machine learning method,and selects the category with the highest weighted probability as the tendency of sentiment.The macro average value of the experiments on the Chinese mobile phone reviews dataset reaches 84.46%,which is 4.53% higher than the basic machine learning method,and the micro average value reaches 84.49%,which is 4.55% higher.Both value of the experiment on the data set of Chinese hotel reviews reaches 83.62%,an increase of 4.11%.The results show that the method can effectively improve the performance of aspect-level sentiment classification performance.3.An aspect-based opinion mining system with product reviews is designed and implemented.This system automatically performs reviews data crawling,pre-processing,evaluation collocation extraction and emotion classification through the product ID input by the user,and displays the results of the opinion mining with the graphical interface.