Sentiment classification systems are vital for applications such as targeted and contextual advertising, market analysis, opinion mining, and so on. Sentiments are expressed on sites across the web, from online shopping to movie reviews. Using classifiers to analyze these sentiments--positive or negative--in various domains, service providers can present users with the most relevant data for advertisements, recommendations, and relevant search results. But training a classifier to serve as a generic model for all types of content is difficult. For example, “horrific” can be used to describe a good horror movie (a positive review), but it would not appear in a positive review of an electronic item, so cross-domain classification is a challenging problem. In this paper, the authors propose a new cross-domain sentiment classification system.

The authors first discuss the two main challenges in creating a better classification system: identifying the source domain (parent domain), and building “a learning framework to incorporate the information ... of source and target domains.” The authors overcame these challenges by expanding associated features. For example, a car can simply be classified as a car or an automobile. By adding more related features to the feature vectors, the authors were able to reduce mismatches between the two domains. They used both labeled and unlabeled data from source and target domains to distribute the features. With lexical and sentiment elements, the authors created a sentiment-sensitive thesaurus (a list of related words and metadata for the contextual data). One of the unique contributions of the paper is that the proposed method creates sentiment labels as a part of the context feature expansion.

The authors conducted extensive experiments to support their proposed method for cross-domain sentiment classification. For these experiments, they used the benchmark experimental dataset containing Amazon.com product reviews for books, DVDs, electronics, and kitchen appliances. The dataset was labeled using a zero- to five-star system, where more than three stars is considered a positive review and fewer than three stars is considered a negative review; the authors did not consider the case where the review gives exactly three stars. The authors compare the accuracy of their method with that of other previously proposed methods: structural correspondence learning (SCL), spectral feature alignment (SFA), and SentiWordNet. The new cross-domain sentiment classification system considerably outperforms all three of the other systems in terms of accuracy and relatedness.

The authors have developed a robust and generic cross-domain sentiment classification method, and have shown that their new classifier is significantly better than the existing classifiers. They should have considered the case of a review with exactly three stars in their experimental analysis. The authors claim that feature expansion has not previously been applied to any cross-domain classification systems, but they do not provide any references to their claim, even in the related work section.

Aside from these two issues, the paper is well written and provides extensive theoretical analysis and experimental results to support the new sentiment classifier.