Domain adaptive bootstrapping for named entity recognition

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1523–1532, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Domain adaptive bootstrapping for named entity recognition Dan Wu1, Wee Sun Lee2, Nan Ye2 1Singapore MIT Alliance 2Department of Computer Science National University of Singapore {dwu@,leews@comp,g0701171@}nus.edu.sg Hai Leong Chieu DSO National Laboratories chaileon@dso.org.sg Abstract Bootstrapping is the process of improving the performance of a trained classifier by iteratively adding data that is labeled by the classifier itself to the training set, and retraining the classifier. It is often used in situations where labeled training data is scarce but unlabeled data is abundant. In this paper, we consider the problem of do- main adaptation: the situation where train- ing data may not be scarce, but belongs to a different domain from the target appli- cation domain. As the distribution of un- labeled data is different from the training data, standard bootstrapping often has dif- ficulty selecting informative data to add to the training set. We propose an effective domain adaptive bootstrapping algorithm that selects unlabeled target domain data that are informative about the target do- main and easy to automatically label cor- rectly. We call these instances bridges, as they are used to bridge the source domain to the target domain. We show that the method outperforms supervised, transduc- tive and bootstrapping algorithms on the named entity recognition task. 1 Introduction Most recent researches on natural language pro- cessing (NLP) problems are based on machine learning algorithms. High performance can often be achieved if the system is trained and tested on data from the same domain. However, the perfor- mance of NLP systems often degrades badly when the test data is drawn from a source that is differ- ent from the labeled data used to train the system. For named entity recognition (NER), for example, Ciaramita and Altun (2005) reported that a system trained on a labeled Reuters corpus achieved an F-measure of 91% on a Reuters test set, but only 64% on a Wall Street Journal test set. The task of adapting a system trained on one do- main (called the source domain) to a new domain (called the target domain) is called domain adap- tation. In domain adaptation, it is generally as- sumed that we have labeled data in the source do- main while labeled data may or may not be avail- able in the target domain. Previous work in do- main adaptation can be classified into two cate- gories: [S+T+], where a small, labeled target do- main data is available, e.g. (Blitzer et al., 2006; Jiang and Zhai, 2007; Daum´e III, 2007; Finkel and Manning, 2009), or [S+T-], where no labeled tar- get domain data is available, e.g. (Blitzer et al., 2006; Jiang and Zhai, 2007). In both cases, and es- pecially for [S+T-], domain adaptation can lever- age on large amounts of unlabeled data in the tar- get domain. In practice, it is often unreasonable to expect labeled data for every new domain that we …

Content truncated to 500 words