Sentiment Analysis Based On Cascaded Conditional Random Fields

With the boarding popularized of Web 2.0 technology, the E-commerce platform, Micro-blog, forum and the third-party group purchase websites etc. have well developed, they are all well welcomed among the Internet users. In those platforms, the Internet users are willing to share comments to the products what they have bought. Always in these short review texts, which hold the sentiment orientation of the speakers, there are huge hidden values which are waiting for us to discover. With those values, for example, we can make use of them to support one company to make sale strategy or predict the development of an incident.Text Sentiment Analysis, also called Opinion Mining, is a natural language processing technology which was specialized used to analysis, manage, summarize and inference the text content. Its main research fields including the sentiment information classification, retrieving and extraction. In this paper, we focus on the fine-grained sentiment information extraction, to make it more details, in this paper we talk about the evaluation word(also called polar word), evaluation object extraction and the polar words’ sentiment orientation prediction. After reading the relevant literatures, there are some following problems:1) There are many unlisted words or new emerging attributes among evaluation objects, which make it hard to be discovered by the rule based or traditional statistic based method;2) The sentiment orientation ambiguity phenomenon of polar words is quite common, the sentiment orientation of a polar word varies in different context(especially depends on the evaluation object it modified), in the existence method, they extract the polar words and predict the sentiment orientation at the same time ignoring the inner dependence between polar word and evaluation object.In this paper we come up with a new method called Cascaded Conditional Random Fields Sentiment Analysis Model to extract polar words and evaluation objects, and make prediction to the polar word sentiment orientation. The advantages of this model are listed as follows:1) The cascaded model is composed by three well ordered, interdependent CRFs model which is design to extract the polar words, evaluation objects and predict the sentiment orientation of the polar words separately, the relative low degree of coupling can let each level CRFs model be focus on one job at a time which can ensure the accuracy;2) Each level CRFs model do job separately, so before the output of the lower level being inputted to the higher level, we can do some necessary pre-operation, which can stop the mistake spreading, therefore improve the accuracy;3) The single CRFs model is depend on other in the up-to-top order, so throw some well designed redundant labels, the higher level CRFs can make use of the lower level CRFs output as input feature, which can further improve the accuracy of the system.Through the experiment makes between the single CRFs and CCRFs, the result proves the advanced performance of CCRFs than the single-unified CRFs.