Medical Relation Extraction with Manifold Models

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 828–838, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Medical Relation Extraction with Manifold Models Chang Wang IBM T. J. Watson Research Center Yorktown Heights, New York, 10598 changwangnk@gmail.com James Fan IBM T. J. Watson Research Center Yorktown Heights, New York, 10598 fanj@us.ibm.com Abstract In this paper, we present a manifold model for medical relation extraction. Our model is built upon a medical corpus containing 80M sentences (11 gigabyte text) and de- signed to accurately and efficiently detect the key medical relations that can facilitate clinical decision making. Our approach integrates domain specific parsing and typ- ing systems, and can utilize labeled as well as unlabeled examples. To provide users with more flexibility, we also take label weight into consideration. Effectiveness of our model is demonstrated both theo- retically with a proof to show that the so- lution is a closed-form solution and exper- imentally with positive results in experi- ments. 1 Introduction There exists a vast amount of knowledge sources and ontologies in the medical domain. Such in- formation is also growing and changing extremely quickly, making the information difficult for peo- ple to read, process and remember. The combi- nation of recent developments in information ex- traction and the availability of unparalleled medi- cal resources thus offers us the unique opportunity to develop new techniques to help healthcare pro- fessionals overcome the cognitive challenges they face in clinical decision making. Relation extraction plays a key role in informa- tion extraction. Using question answering as an example (Wang et al., 2012): in question analy- sis, the semantic relations between the question focus and each term in the clue can be used to identify the weight of each term so that better search queries can be generated. In candidate an- swer generation, relations enable the background knowledge base to be used for potential candidate answer generation. In candidate answer scoring, relation-based matching algorithms can go beyond explicit lexical and syntactic information to detect implicit semantic relations shared across the ques- tion and passages. To construct a medical relation extraction sys- tem, several challenges have to be addressed: • The first challenge is how to identify a set of relations that has sufficient coverage in the medical domain. To address this issue, we study a real-world diagnosis related question set and identify a set of relations that has a good coverage of the clinical questions. • The second challenge is how to efficiently de- tect relations in a large amount of medical text. The medical corpus underlying our re- lation extraction system contains 80M sen- tences (11 gigabytes pure text). To extract relations from a dataset at this scale, the re- lation detectors have to be fast. In this paper, we speed up relation detectors through pars- ing adaptation and replacing non-linear clas- sifiers with linear classifiers. • The third challenge is that the labeled rela- tion examples are often insufficient due to the high …

Content truncated to 500 words