Active Learning with Transfer Learning

In sentiment classification, unlabeled user reviews are often free to collect for new products, while sentiment labels are rare. In this case, active learning is often applied to build a high-quality classifier with as small amount of labeled instances as possible. However, when the labeled instances are insufficient, the performance of active learning is limited. In this paper, we aim at enhancing active learning by employing the labeled reviews from a different but related (source) domain. We propose a framework Active Vector Rotation (AVR), which adaptively utilizes the source domain data in the active learning procedure. Thus, AVR gets benefits from source domain when it is helpful, and avoids the negative affects when it is harmful. Extensive experiments on toy data and review texts show our success, compared with other state-of-theart active learning approaches, as well as approaches with domain adaptation.