High-Performance Tagging on Medical Texts

High-Performance Tagging on Medical Texts Udo Hahn Joachim Wermter Computerlinguistik, Friedrich-Schiller-Universit¨at Jena F¨urstengraben 30, D-07743 Jena, Germany hahn@coling.uni-freiburg.de Abstract We ran both Brill’s rule-based tagger and TNT, a statistical tagger, with a default German newspaper-language model on a medical text corpus. Supplied with limited lexicon re- sources, TNT outperforms the Brill tagger with state-of-the-art performance figures (close to 97% accuracy). We then trained TNT on a large annotated medical text corpus, with a slightly extended tagset that captures certain medical language particularities, and achieved 98% tag- ging accuracy. Hence, statistical off-the-shelf POS taggers cannot only be immediately reused for medical NLP, but they also – when trained on medical corpora – achieve a higher perfor- mance level than for the newspaper genre. 1 Introduction The application of language technology in the med- ical field, dubbed as medical language processing (MLP), is gaining rapid recognition (for a survey, cf. Friedman and Hripcsak (1999)). It is both im- portant, because there is strong demand for all kinds of computer support for health care and clinical ser- vices, which aim at improving their quality and de- creasing their costs, and challenging — given the miracles of medical sublanguage, the various text genres one encounters and the enormous breadth of expertise surfacing as medical terminology. However, the development of human language technology for written language material has, up un- til now, almost exclusively focused on newswire or newspaper genres. This is most prominently ev- idenced by the PENN TREEBANK (Marcus et al., 1993). Its value as one of the most widely used language resources mainly derives from two fea- tures. First, it supplies everyday, non-specialist doc- ument sources, such as the Wall Street Journal, and, second, it contains value-added, viz. annotated, lin- guistic data. Since the understanding of newspaper material does not impose particular requirements on its reader, other than the mastery of general English and common-sense knowledge, it is easy for almost everybody to deal with. This is essential for the ac- complishment of the second task, viz. the annotation and reuse of part-of-speech (POS) tags and parse trees, as the result of linguistic analysis. With the help of such resources, whole generations of state- of-the-art taggers, chunkers, grammar and lexicon learners have evolved. The medical field poses new challenges. First, medical documents exhibit a large variety of struc- tural features not encountered in newspaper docu- ments (the genre problem), and, second, the under- standing of medical language requires an enormous amount of a priori medical expertise (the domain problem). Hence, the question arises, how portable results are from the newspaper domain to the medi- cal domain? We will deal with these issues, focusing on the portability of taggers, from two perspectives. We first pick up off-the-shelf technology, in our case the rule-based Brill tagger (Brill, 1995) and the statistically-based TNT tagger (Brants, 2000), both trained on newspaper data, and run it on medical text data. One may wonder how the taggers trained on newspaper language perform with medical lan- guage. Furthermore, one …

Content truncated to 500 words