Online Active Learning for Cost Sensitive Domain Adaptation

Proceedings of the Seventeenth Conference on Computational Natural Language Learning, pages 1–9, Sofia, Bulgaria, August 8-9 2013. c⃝2013 Association for Computational Linguistics Online Active Learning for Cost Sensitive Domain Adaptation Min Xiao and Yuhong Guo Department of Computer and Information Sciences Temple University Philadelphia, PA 19122, USA {minxiao,yuhong}@temple.edu Abstract Active learning and domain adaptation are both important tools for reducing labeling effort to learn a good supervised model in a target domain. In this paper, we inves- tigate the problem of online active learn- ing within a new active domain adapta- tion setting: there are insufficient labeled data in both source and target domains, but it is cheaper to query labels in the source domain than in the target domain. Given a total budget, we develop two cost- sensitive online active learning methods, a multi-view uncertainty-based method and a multi-view disagreement-based method, to query the most informative instances from the two domains, aiming to learn a good prediction model in the target do- main. Empirical studies on the tasks of cross-domain sentiment classification of Amazon product reviews demonstrate the efficacy of the proposed methods on re- ducing labeling cost. 1 Introduction In many application domains, it is difficult or ex- pensive to obtain labeled data to train supervised models. It is critical to develop effective learning methods to reduce labeling effort or cost. Active learning and domain adaptation are both impor- tant tools for reducing labeling cost on learning good supervised prediction models. Active learn- ing reduces the cost of labeling by selecting the most informative instances to label, whereas do- main adaptation obtains auxiliary label informa- tion by exploiting labeled data in related domains. Combining the efforts from both areas to further reduce the labeling cost is an important research direction to explore. In this paper, we consider online active learn- ing with domain adaptations. Online learning has been widely studied (Borodin and El-Yaniv, 1998) due to its advantages of low memory requirement and fast computation speed. Dredze and Crammer (2008) applied online learning on domain adap- tation and proposed to combine multiple similar source domains to perform online learning for the target domain, which provides a new opportunity for conducting active learning with domain adap- tation. Online active learning with domain adap- tation, to our knowledge, has just gained atten- tion recently and has been addressed in (Rai et al., 2010; Saha et al., 2011). The active online do- main adaptation methods developed in (Rai et al., 2010; Saha et al., 2011) leverage information from the source domain by domain adaptation to intelli- gently query labels for instances only in the target domain in an online fashion with a given budget. They assumed a large amount of labeled data is readily available in the source domain. In this work, we however tackle online active learning with domain adaptation in a different set- ting, where source domains with a large amount of free labeled data are not available. Instead we as- sume there are very few labeled instances in …

Content truncated to 500 words