Morphological Analysis of Biomedical Terminology with Analogy-Based Alignment

Proceedings of Recent Advances in Natural Language Processing, pages 347–354, Hissar, Bulgaria, 12-14 September 2011. Morphological Analysis of Biomedical Terminology with Analogy-Based Alignment Vincent Claveau IRISA-CNRS Vincent.Claveau@irisa.fr Ewa Kijak IRISA-Univ. Rennes 1 Ewa.Kijak@irisa.fr Abstract In the biomedical domain, many terms are neoclassical compounds (composed of several Greek or Latin roots). The study of their morphology is important for numer- ous applications since it makes it possible to structure, translate, retrieve them effi- ciently... In this paper, we propose an original yet fruitful approach to carry out this morpho- logical analysis by relying on Japanese, more precisely on terms written in kanjis, as a pivot language. In order to do so, we have developed a specially crafted align- ment algorithm relying on analogy learn- ing. Aligning terms with their kanji-based counterparts provides at the same time a decomposition of the term into morphs, and a kanji label for each morph. Evaluated on a dataset of French terms, our approach yields a precision greater than 70% and shows its relevance com- pared with existing techniques. We also illustrate the interest of this approach through two direct applications of the pro- duced alignments: translating unknown terms and discovering relationships be- tween morphs for terminological structur- ing. 1 Introduction In many domains, accessing the information in documents or collections of documents is guided by the use of well-defined terms, which form a terminology of the domain. This is particularly true in the biomedical domain where there is a long tradition of terminologies development for structuring the knowledge as well as accessing it. An example is the MeSH (Medical Subject Head- ings) www.nlm.nih.gov/mesh terminology which is used to index the very popular PubMED database (www.pubmed.gov). Knowing how to handle these terms, understanding them, translat- ing them or building semantic relationships be- tween them are thus essential operations for ap- plications like enrichment of bilingual lexicons, or more generally machine translation, information retrieval... In this framework, the work presented here is interested in the morphology of simple terms from the biomedical domain as a basis for the termi- nological analysis. More precisely, we present a technique aiming at breaking up a term into its morphological components, namely morphs, and associating in the same time semantic knowl- edge to these morphs. Note that in this pa- per, we distinguish morphs, elementary linguistic signs (segments), from morphemes, equivalence classes with identical signified and close signifi- cants (Mel’ˇcuk, 2006). We therefore tackle the same issue already raised in some studies (Deléger et al., 2008; Markó et al., 2005, for example), but we try here to suppress the costly human opera- tions required by these studies. The original idea at the heart of our approach is to use the multilingualism of existing termino- logical databases. We exploit Japanese as a pivot language, or more precisely terms written in kan- jis, to help decomposing the terms of other lan- guages into morphs and associate them with the corresponding kanjis, in a fully automatic way. Thus, kanjis play the role of a …

Content truncated to 500 words