!Sentiment Classification: A Topic Sequence-Based Approach

With the development of Web 2.0, sentiment analysis has been widely used in many domains, such as information retrieval (IR), artificial intelligence and social networks. This paper focuses on the task of classifying a textual review as expressing a positive or negative sentiment, a core task of sentiment analysis called sentiment classification. To address this problem, we present a novel sentiment classification model based on topic sequence which refers to topics in descending order of their distribution probabilities. Topics' distribution probabilities are obtained after training the latent dirichlet allocation (LDA) model. To the best of our knowledge, previous work didn't consider the importance of the order relationships among topics. We work on exploiting the order relationships among topics and using this information for sentiment classification. Based on it, three steps are followed to tackle this task. First, we train the LDA model to get the topic distribution. Then, we sort these topics in descending order to get the topic sequence, which are used to construct topic co-occurrence matrices (positive and negative). Finally we use these two matrices to classify the test examples as positive or negative. The experiments show that our classification model obtains better results than many existing classifiers and the topic sequence plays an important role for sentiment classification.