Research On Sentiment Orientation Technology For Review Texts

The behavior of the user on the Internet is just "receiving" information in the past and becomes now involved in "creating" information that people can publish their own comments on merchandise,business,and service.In order to make more informed decisions,more and more consumers began to choose to understand the word of mouth of the product or service before buying.These comments can also be used as a kind of feedback to help producers understand the strengths and weaknesses of their products and then gain the competitive advantage.Base on the relevant research at home and abroad,this paper uses the theories and methods of natural language processing and machine learning to analyze the tendency of Chinese comment text.The main goal of this paper is to minimize the manual annotation work.The main contents include three parts:the evaluation words and objects extraction,evaluation unit extraction and emotion computing.First,the evaluation words and objects extraction part,In order to reduce the workload of manual annotation,this paper joints unsupervised and supervised method to extract evaluation words and objects.The Apriori algorithm is used in the paper as the unsupervised method to extract the frequent items,and making the result as the seed words of Conditional Random Field Models(CRFs)to iteratively extract the final evaluation information.The seed words also include the words in the emotional vocabulary ontology library in the process of extracting the evaluation words.The F1-measure of the evaluation words and evaluation objects extraction are 69.33%,60.13%respectively.Compared with those supervised methods,this method can greatly reduce manual tagging work and also has certain degree of across domain ability.Second,the evaluation unit is extracted in front of the evaluation words and evaluation objects extracted before,Using the method of fixed the evaluation words and taken evaluation object as the linear chain of CRFs model,combined with feature templates(words,POS,location and modifier features).Finally,the comment text is expressed in the form of several evaluation units(<evaluation object,evaluation word,negative multiplicity,whether containing degree adverb,text sentence>).Third,the polarity value of the context is calculated by calculating the polarity value of each evaluation unit.Considering that some neutral words will show different emotional polarity in different contexts,the evaluation objects are divided into four categories,evaluation words are divided into five categories.The calculation fonnula of polarity value in different categories of evaluation words and different categories of evaluation objects is given with the impact of negative word,degree word,and text document patterns to the polarity.Finally,the F1 value obtained by the text emotion calculation method is 73.20%,and the experiment results also show that the method has certain cross-domain.