Semi-Supervised Maximum Entropy Based Approach to Acronym and Abbreviation Normalization in Medical Texts

Semi-Supervised Maximum Entropy Based Approach to Acronym and Abbreviation Normalization in Medical Texts Serguei Pakhomov, Ph.D. Mayo Foundation, Rochester, MN pakhomov.sergey@mayo.edu Abstract Text normalization is an important aspect of successful information retrieval from medical documents such as clinical notes, radiology reports and discharge summaries. In the medical domain, a significant part of the general problem of text normalization is abbreviation and acronym disambiguation. Numerous abbreviations are used routinely throughout such texts and knowing their meaning is critical to data retrieval from the document. In this paper I will demonstrate a method of automatically generating training data for Maximum Entropy (ME) modeling of abbreviations and acronyms and will show that using ME modeling is a promising technique for abbreviation and acronym normalization. I report on the results of an experiment involving training a number of ME models used to normalize abbreviations and acronyms on a sample of 10,000 rheumatology notes with ~89% accuracy. 1 Introduction and Background Text normalization is an important aspect of successful information retrieval from medical documents such as clinical notes, radiology reports and discharge summaries, to name a few. In the medical domain, a significant part of the general problem of text normalization is abbreviation and acronym1 disambiguation. Numerous abbreviations are used routinely throughout such texts and identifying their meaning is critical to understanding of the document. The problem is that abbreviations are highly ambiguous with respect to their meaning. For example, according to UMLS2 (2001), RA may stand for “rheumatoid arthritis”, “renal artery”, “right atrium”, “right atrial”, “refractory anemia”, “radioactive”, “right arm”, “rheumatic arthritis,” etc. Liu et al. (2001) show that 33% of abbreviations listed in UMLS are ambiguous. In addition to problems with text interpretation, Friedman, et al. (2001) also point out that abbreviations constitute a major source of errors in a system that automatically generates lexicons for medical NLP applications. Ideally, when looking for documents containing “rheumatoid arthritis”, we want to retrieve everything that has a mention of RA in the sense of “rheumatoid arthritis” but not those documents where RA means “right atrial.” In a way, abbreviation normalization problem is a special case of the word sense disambiguation (WSD) problem. Modern approaches to WSD include supervised machine learning techniques, where some amount of training 1 To save space and for ease of presentation, I will use the word “abbreviation” to mean both “abbreviation” and “acronym” since the two could be used interchangeably for the purposes described in this paper. 2 Unified Medical Language System, a database containing biomedical information and a tools repository developed at the National Library of Medicine to help helath professionals as well as medical informatics researchers. Computational Linguistics (ACL), Philadelphia, July 2002, pp. 160-167. Proceedings of the 40th Annual Meeting of the Association for data is marked up by hand and is used to train a classifier. One such technique involves using a decision tree classifier (Black 1988). On the other side of the spectrum, the fully unsupervised learning methods such as clustering have also been successfully used (Shutze …

Content truncated to 500 words