Ensemble-Based Medical Relation Classification

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1682–1693, Dublin, Ireland, August 23-29 2014. Ensemble-Based Medical Relation Classification Jennifer D’Souza and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {jld082000,vince}@hlt.utdallas.edu Abstract Despite the successes of distant supervision approaches to relation extraction in the news do- main, the lack of a comprehensive ontology of medical relations makes it difficult to apply such approaches to relation classification in the medical domain. In light of this difficulty, we propose an ensemble approach to this task where we exploit human-supplied knowledge to guide the de- sign of members of the ensemble. Results on the 2010 i2b2/VA Challenge corpus show that our ensemble approach yields a 19.8% relative error reduction over a state-of-the-art baseline. 1 Introduction Medical relation (MR) classification, an information extraction task in the clinical domain that was re- cently defined in the 2010 i2b2/VA Challenge (Uzuner et al., 2011), involves determining the relation between a pair of medical concepts (problems, treatments, or tests). The ability to classify MRs is indis- pensable to sound automatic analysis of patient health records. While MR classification is a relatively new task, there has been a lot of work on extracting semantic relations from news articles. Supervised approaches train classifiers on data annotated with the target relation types, typically using a rich feature set (Zhou et al., 2005; Surdeanu and Ciaramita, 2007; Zhou et al., 2007). Since obtaining annotated data is a time-consuming and labor-intensive process, researchers have considered unsupervised approaches (Shinyama and Sekine, 2006; Banko et al., 2007). While unsupervised approaches can use a large amount of unannotated data and extract a large number of relations, it may not be easy to map the resulting relations to those needed for a given knowledge base. One way to mitigate this problem is semi-supervised learning: starting from a given set of seed instances, a bootstrapping algorithm is used to iteratively learn extraction patterns and extract instances (Brin, 1999; Riloff and Jones, 1999; Agichtein and Gravano, 2000; Ravichandran and Hovy, 2002; Etzioni et al., 2005; Pantel and Pennacchiotti, 2006; Bunescu and Mooney, 2007; Rozenfeld and Feldman, 2008). However, the resulting patterns often suffer from semantic drift and low precision. Recent years have seen a surge of interest in distant supervision for relation extraction (Mintz et al., 2009; Nguyen and Moschitti, 2011; Krause et al., 2012; Min et al., 2013). The idea is to automatically create annotated relation instances by extracting their labels from relation instances in a knowledge base such as Freebase (Bollacker et al., 2008) and YAGO (Suchanek et al., 2007). Our goal in this paper is to advance the state of the art in MR classification. One of the major chal- lenges in MR classification is the scarcity of labeled data. At first glance, we can mitigate this problem using distant supervision approaches. However, there is difficulty in applying these approaches to MR classification: only one of the relation types defined in the 2010 …

Content truncated to 500 words