Opinion Mining Research Based On Topic And Sentiment Unification Maximum Entropy Model

With the development of Electronic Commerce, tens of thousands of users have began to purchase various products and services through the network and publish the related online reviews. Analysis of these reviews can not only help potential customers make an intelligent decision, but also guide enterprises to timely improve the quality of their products and services. However, the number of online reviews is enormous, which is impractical, if not impossible, to use traditional manual-methods for fast access. Therefore, it has been an important research topic for researchers to develop opinion mining of online reviews through automatic analyzing and extracting methods.The rule-based method and language-based method, statistical machine learning method are mainly used in opinion mining. Many domain experts is needed to define characteristics of words and rules in different areas in rule-based method, which can not meet the continual emergence of new words., and the rule can not be used in different domain. The language-based method is based on the use of linguistic rules of the language syntax characteristic to recognize aspect word. However grammar of different languages is very different, which cause this method is not easily language adaptable. Both of the mentioned methods have poor portability, can not automatically cluster words which have similar significance to a topic. As an unsupervised statistical topic model, LDA topic model not only can well overcome the disadvantage of the above methods, but also do not require a lot of manually labeled training set which is needed by other supervised and semi-supervised statistical models, so it has been widely used by researchers. Standard LDA is a bag-of-word model, which assumes that a document is a set of independent words. Locations and semantic information of words are not considered in the model. Thus, it is not appropriate for word extraction in fine-grained methods.To solve this problem, a topic and sentiment unification maximum entropy LDA model (TSU MaxEnt-LDA) is proposed for fine-grained opinion mining. Topics and sentiments are simultaneously considered on word or phrase level to get more specific sentiment polarity analysis. First, maximum entropy component is added in TSU MaxEnt-LDA to distinguish background words, aspect words and opinion words and further realize both the local and global extraction of these words. Then, an indicator variables is introduced to distinguish local and global aspect words and opinion words.Finally, sentiment layer is inserted between the topic layer and the word layer to get sentiment polarity of each topic and the entire sentiment polarity of the whole review, automatically produce fine-grained topics emotional summary chart.In order to validate this model is domain adaptable, two different corpus is used in the experiment. One is a collection of restaurant review from Citysearch New York, and the other is electronic reviews from Amazon.experimental results show that the proposed theoretical research perform better than the past proving the correctness of the model.The research background and research significance, current research status have been introduced in the first chapter. The second chapter describe the fine-grained opinion mining tasks, such as sentiment classification, opinion extraction and opinion analysis. At the same time, it introduces the basic mathematical knowledge and models which is involved in this paper. The third Chapter first describe the TSU MaxEnt-LDA model in detail; then introduce the whole generation process of a document and the inference of the proposed model. In the forth chapter, we present some experimental results and verify the correctness of TSU MaxEnt-LDA model theory thorough analysis of the experimental results.