Research On Topic Model For Online Review Texts

With the rapid development of e-commerce and the popularity of online shopping, the Web storages huge number of product and service reviews comment by customers. Product reviews contain subjective feelings of customers who have used some products. These subjective texts reflect people’s opinions, attitudes and positions. So information of subjective texts provides important commercial value. However, it is not impossible to read all the information while much information floods on the Internet. Therefore, a valid text mining method which applies to online reviews texts is urgent.With review texts being widely used in people’s life, mining the topic information only could not satisfy people’s requirement of the information about sentiment analysis in reviews. For a given set of online review texts, through analysis and mining, we can not only get topics information but also sentiment information about the review texts, which is the goal of detects topics and sentiment simultaneously from review texts.In this paper, in order to mining topic information and opposite sentiment information in each topic from online review texts, we propose a topic model called Topic-Opposite Sentiment Mining Model (TOSM) based on Latent Dirichlet Allocation (LDA). We extend LDA to TOSM with adding the sentiment layer, which makes TOSM has four layers like "document-topic-sentiment-word" than LDA has three layers like "document-topic-word". Moreover, we import sentiment prior information into TOSM to make our opposite sentiment represented clearly. It contains two methods to import sentiment prior information. First, we use sentiment lexicon. Second, we use Polarity Shifting Detection method to detect the negation in review texts.TOSM considers the correlation between topics and sentiments, and TOSM model topic information and sentiment information simultaneously. One important advantage of TOSM is that each topic detected by TOSM contains two opposite sentiments.Experimental results show that the topic-sentiment found by TOSM matches evaluative details of the reviews. Experimental results of sentiment classification show that TOSM outperforms ASUM and JST model.