Research On Opinion Mining Technology For Product Reviews On Websites

With the growing popularity of the Internet application, shopping website, review websites, as well as all kinds of BBS are gradually accumulating more and more product comments. In research about product comments, Aspect-based emotion analysis attracts much attention. From the consumer’s point of view, different consumer values different aspect, aspect-based emotion analysis results can help them to choose more suitable products. From the manufacturer’s point of view, aspect-based emotion analysis results can let them know what specific problems their products have, and help to improve the products.This paper uses the CRFs model to proceed aspect-based comment text sentiment analysis, mainly including three parts:aspects and evaluation term extraction, emotional polarity and strength analysis, emotional polarity and strength analysis result summarize based on theme-aspect.Regarding extraction of aspect and evaluation term, a semi-supervised method based on LDA and CRFs model is proposed in this paper, this method can extract aspect and evaluation term simultaneous. The initial aspect and evaluation term set is obtained from training results of the LDA topic model and sentiment word dictionary of the HowNet. Statistical feature (obtained from training results of the LDA topic model), semantic feature (word, part of speech, degree of modal particles, before and after relationship between words), as well as co-occurrence of aspect and evaluation term are all used as features in CRFs model. The F1-Measure of aspect and evaluate term extraction are 71.7% and 55.3% respectively. Compared with those supervised learning methods, method in this article can largely reduce the aspect and evaluation term marking workload. As is semi-supervised, the method also has certain degree of across domain character, a few of additional work is needed when applying to other areas.Regarding emotional polarity classification and strength analysis, a cascading CRFs model is used in this paper. Word, part of speech, modal particles, degree word, co-occurrence of aspect and evaluation term are all used as features in CRFs model.The F1-Measure of emotion sentence recognition, emotional polarity classification and emotional strength analysis are 86.3%,77.2% and 70.7% respectively.Regarding opinion summarization, a theme-aspect based method, which can summarize emotional polarity and strength analysis result, is proposed in this paper. In the end, a comment mining system is implemented to display the mining result visually.