Structured Term Recognition in Medical Text

Proceedings of COLING 2012: Technical Papers, pages 943–958, COLING 2012, Mumbai, December 2012. Structured Term Recognition in Medical Text Michael Glass Al f io Gliozzo IBM T.J. Watson Research, Yorktown Heights ♠r❣❧❛ss❅✉s✳✐❜♠✳❝♦♠✱ ❣❧✐♦③③♦❅✉s✳✐❜♠✳❝♦♠ ABSTRACT In this work we present a novel approach to bootstrap domain specific terminology, namely Structured Term Recognition, and we apply it to the medical domain. In contrast to previous approaches, based on observing distributional properties of terminology with respect to their contexts, our method analyzes the “internal structure” of multi-word terms by learning patterns of word clusters. Evaluation shows that our method can be used to boost the performances of term recognition systems based on dictionary lookup while extending the coverage of large ontologies like UMLS. KEYWORDS: terminology recognition, bootstrapping, medical terminology, named entity recognition, joint inference. 943 1 Introduction Recognizing occurrences of domain specific terms and their types is important for many text processing applications. This problem is not easy, particularly in domains like medicine, where very rich terminology is generated by domain experts on a daily basis. In spite of the large interest in statistical term recognition in Natural Language Processing (NLP), state of the art approaches for term recognition in the medical domain are still based on dictionary lookup with some heuristics for partial mapping (Aronson, 2001). In fact, very large terminological resources, such as the Unified Medical Language System (UMLS) (Bodenreider, 2004), have been developed in the medical domain. The reason is that medical terminology can not be identified by looking at superficial features only, such us capitalization of words, prefixes and suffixes. In fact, diseases names, symptoms and most medical terms are not proper names, so they are not capitalized. In addition, they are usually characterized by a rather complex internal structure and composed by many words. In addition, distributional similarity metrics, i.e. recognition approaches based on the analysis of the local context where the term is located, work well when applied to single words or very frequent words, which is not the case for most of the medical terms we are interested in. In the context of the research on adapting a question answering system to the medical domain, we encountered the term recognition problem in many places. For example, recognizing names of diseases, symptoms and treatments is necessary to answer most of the Doctor’s DilemmaTMquestions (American College of Physicians, 2012), an evaluation benchmark we used to measure the ability of our system to answer medical questions. To assess the validity of the answer “HFE hereditary hemochromatosis” with respect to the question “Metabolic condition that can set off airport metal detector”, it is important to know that the answer is a type of metabolic condition. One way to address this problem is to use medical lexica where different terms are associated to semantic types, and then check whether the type of the candidate answer matches the type required by the question. On the other hand, UMLS is far from complete. Many disease names (especially multi-words) are not recognized by dictionary …

Content truncated to 500 words