Cross-domain Sentiment Orientation Research For Product Reviews

The Internet has provided such a lot of convenience for people that we can share information and communicate freely at any time. The number of Internet users is increasing and online shopping has become a kind of frequent behavior in people’s daily life. Users can buy products via the Internet, at the same time, they can make comments in order to share their shopping experience or their satisfaction of products and services. These product reviews contain rich emotional data and business value, not only can provide references for potential customers, helping them to make more rational purchase decisions, but also provide product feedback information for product producers and sellers, guiding them to improve the quality of products and services and develop more effective marketing strategy. The total number of product reviews on the Internet is huge and shows a tendency of surging. Mining and analyzing the sentiment information from product reviews by artificial alone is not possible. Sentiment tendency analysis technology arises at the historic moment. With the development of product variety, product reviews involve various domains, such as cars, electronic products, hotel, etc. Cross-domain sentiment tendency analysis technology uses the source domain data and their sentiment labels to judge the data label in the target domain.However, the data in different domains are usually subjected to different distribution, which is because that thay are acquired with various methods and approaches. This may cause certain challenges to traditional sentiment classification methods. Aiming at this problem, this paper proposes two kinds of cross-domain sentiment tendency analysis methods: one is the method of cross-domain opinion analysis based on the bootstrapping and propagation of trust-worthy Label, the other is the method of cross-domain sentiment tendency analysis for product reviews based on the combination framework model.The method of cross-domain opinion analysis based on the bootstrapping and propagation of trust-worthy Label is a half-supervised classification algorithm integrating the LPA algorithm and the concept of Bootstrapping. It aims at word level and be conducted as follows. Label propagation will be conducted between the seed words and the unlabeled words according to their similarities, during which, the words that are with trust-worthy labels will be chosen to extend the set of seed words iteratively. After that, the sentiment scores of the words from the target domain will be further improved by their prior scores.The method of cross-domain sentiment tendency analysis for product reviews based on the combination framework model aims to analyze the sentiment tendency of product reviews from different domains. Combining the different characteristics of lexicon rule classification sentiment method and machine learning sentiment classification, it builds a combination frame model to organically integrate the lexicon rule classifier and machine learning classifier. According to the label consistency principle, it choose the data with consistent label after classification by the two kinds of classifiers to join the training sets. Then it trains a new classifier to classify the other unlabeled data repeatedly, until the end of the iteration.To verify feasibility of these two kinds of cross-domain sentiment tendency analysis methods, this paper conducted experiments in some different product reviews domains. The results show that the first method plays an important role in word-level cross-domain sentiment tendency analysis and it can be applied for the expansion of domain sentiment lexicon. The experimental results also indicate that the proposed combination framework model, can improve the accuracy of cross-domain product reviews sentiment tendency analysis in a certain extent.