Domain adaptation with active learning for named entity recognition

One of the dominant problems facing Named Entity Recognition is that when a system trained on one domain is applied to a different domain, a substantial drop in performance is frequently observed. In this paper, we apply active learning strategies to domain adaptation for named entity recognition systems and show that adaptive learning combining the source and target domains is more effective than nonadaptive learning directly from the target domain. Active learning aims to minimize labeling effort by selecting the most informative instances to label. We investigate several sample selection techniques such as Maximum Entropy and Smallest Margin and apply them to the ACE corpus. Our results show that the labeling cost can be reduced by over 92% without degrading the performance.