Customizing Sentiment Classifiers to New Domains: a Case Study

摘要: Sentiment classification is a very domain specific problem; classifiers trained in one do not perform well others. Unfortunately, many domains are lacking large amounts of labeled data for fully-supervised learning approaches. At the same time, sentiment need to be customizable new order useful practice. We attempt address these difficulties and constraints this paper, where we survey four different approaches customizing system target absence data. base our experiments on from domains. After establishing that naive cross-domain results poor accuracy, compare obtained by using each discuss their advantages, disadvantages performance.