Sentiment Classifiation Via Sequence Modeling

With the development of Internet, the value of network applications is always transfered. The user-center conception is coming true. In 2010, the number of visit of Facebook became larger than Google, and became the busiest website in America. Websites like Twitter, Movie Critic etc. are more and more popular. In this period of personal informations explosion, it's more and more important to know user's opinions, which became a very important problem recently.The task of Sentiment Analysis is to recognize the sentiment polarity of specific text. The earlier approaches consider only the frequency of words with polarity. But in more complex sentence, the performance become bad. So usually, people think parsed information like POS(position of speech), dependency etc. should be used in this problem.Original approaches option focus on text-level classification. However, in Twitter or Facebook, it contains only one sentence, without any contextual information. A kind of sentence-level sentiment classification method is necessary for this scenario.This paper proposed a sentence-level sentiment analysis framework, without POS or any other information. To attack this problem without the sentence parsing, we propose an approach whereby a given sentence is decomposed into a series of sub-sequences or sub-view representations. Sentence-level polarity is then determined by classifying within sub-views and fusing the obtained sub-view polarities. Two specific methods are instantiated: stacking-based maximum entropy model and hidden conditional random fields (HCRFs) based on contextual features.Extensive evaluations were carried out on two benchmark dataset, one is for sentence subjectivity classification and the other is for sentence polarity detection. Experimental results show that the performance of our proposed method is comparable to the state-of-the-art approaches.