Reseach On Extended Topic Model For Fine-grained Opinion Mining Of Online Reviews

With the development of the Internet and E-commerce technology,consumers' shopping online becomes a boom.People can buy everything they want without leaving home.However,buyers disable to check the products in person,neither the quality nor the function.So,comments made by pre-buyers seem more important.It acts as reference for the consumers.In addition,merchants get the feedback from users' reviews to improve the quality of their products and services.At present,comments in web site involve the great amount of data.It is difficult to obtain useful information quickly by manual-reading method,and people want to obtain evaluation information about all aspects of products or services not overall evaluation of them.So fine-grained opinion mining becomes a hot research topic.Considering the short shortcoming of the current fine-grained opinion methods such as based on artificial definition,frequency,supervised learning method,LDA as an unsupervised method not only doesn't need manual labeled training data,but also overcomes the some shortcomings of above-methods.It is widely used by researchers.However,the LDA can't extract the aspect from reviews and the three-level model of LDA can't achieve sentiment analysis.It needs to be improved and expanded to achieve opinion mining.To solve this problem,we propose Fine-grained Topic Sentiment Unification Model(FG-TSU Model).First,the model introduces the sliding window to extract semantic aspects;besides,an indicator variables is introduced distinguish aspect words and opinion word based on lexicon-opinion;Finally,in order to achieve sentiment analysis,we incorporate an additional sentiment layer between a document and a topic,it not only obtain the entire sentiment polarity of the whole review,but also entire sentiment polarity of aspects.In order to verify the domain adaptability of this model,we select two datasets of hotel and mobile phone to carry out experiments,and design three experiments.The experimental results show the feasibility of this model in the realization of fine-grained opinion mining.