Transfer Learning for Entity Recognition of Novel Classes

Proceedings of the 27th International Conference on Computational Linguistics, pages 1974–1985 Santa Fe, New Mexico, USA, August 20-26, 2018. 1974 Transfer Learning for Entity Recognition of Novel Classes Juan Diego Rodriguez, Adam Caldwell and Alexander Liu Applied Research Laboratories The University of Texas at Austin Austin, Texas 78713 {juan.rodriguez, adam.caldwell, aliu}@arlut.utexas.edu Abstract In this reproduction paper, we replicate and extend several past studies on transfer learning for entity recognition. In particular, we are interested in entity recognition problems where the class labels in the source and target domains are different. Our work is the first direct comparison of these previously published approaches in this problem setting. In addition, we perform experi- ments on seven new source/target corpus pairs, nearly doubling the total number of corpus pairs that have been studied in all past work combined. Our results empirically demonstrate when each of the published approaches tends to do well. In particular, simpler approaches often work best when there is very little labeled target data, while neural transfer approaches tend to do better when there is more labeled target data. 1 Introduction Transfer learning methods can play an important role in improving the performance of machine learning algorithms in scenarios where in-domain data is scarce, such as entity extraction of protected health information in electronic health records. These methods make use of data in a source domain to improve performance in a target domain. Many transfer learning techniques have been developed for tasks such as named entity recognition (Daum´e III, 2007) and sentiment classification (Blitzer et al., 2007). A broad taxonomy of transfer learning scenarios in machine learning (only some of which are applied to NLP tasks) is given in (Zhang et al., 2017). In this paper we focus on the entity recognition task1 in which both the source and target domains have labeled data, but the class label sets are different. As is common in the literature, we frame entity recog- nition as a sequential classification problem, and restrict ourselves to local (sentence-level) supervised techniques. Most transfer learning methods have been developed under the assumption that the source and target domains have the same class labels (e.g., (Jiang and Zhai, 2007; Daum´e III, 2007)), and many NER corpora only annotate a small number of categories (e.g., 4 labels for the CoNLL 2003 corpus (Sang and Meulder, 2003) and 7 labels for the MUC 6 corpus (Grishman and Sundheim, 1996)). In many situations one would like to recognize categories that were not labeled in a given corpus, but that may be semantically related. This includes, for instance, recognizing companies or sports teams (Ritter et al., 2011), rather than organizations in general, or recognizing actors (Liu et al., 2013b), rather than people in general. In addition, techniques for transfer learning with different class labels may be useful even when the class label set is the same across domains. As Florian et al. (2004) point out, the same labels may be used in very different ways across corpora due to different annotation guidelines. …

Content truncated to 500 words