Biographies or blenders: Which resource is best for cross-domain sentiment analysis?

Abstract

Domain adaptation is usually discussed from the point of view of new algorithms that minimise performance loss when applying a classifier trained on one domain to another. However, finding pertinent data similar to the test domain is equally important for achieving high accuracy in a cross-domain task. This study proposes an algorithm for automatic estimation of performance loss in the context of cross-domain sentiment classification. We present and validate several measures of domain similarity specially designed for the sentiment classification task. We also introduce a new characteristic, called domain complexity, as another independent factor influencing performance loss, and propose various functions for its approximation. Finally, a linear regression for modeling accuracy loss is built and tested in different evaluation settings. As a result, we are able to predict the accuracy loss with an average error of 1.5% and a maximum error of 3.4%. © 2012 Springer-Verlag.