Unsupervised and Semi-Supervised Morphological Analysis for Information Retrieval in the Biomedical Domain

Proceedings of COLING 2012: Technical Papers, pages 629–646, COLING 2012, Mumbai, December 2012. Unsupervised and semi-supervised morphological analysis for Information Retrieval in the biomedical domain Vincent CLAVEAU IRISA-CNRS, Campus de Beaulieu F35042 Rennes, France Vincent.Claveau@irisa.fr Abstract In the biomedical field, the key to access information is the use of specialized terms. However, in most of Indo-European languages, these terms are complex morphological structures. The aim of the presented work is to identify the various meaningful components of these terms and use this analysis to improve biomedical Information Retrieval. We present an approach combining an automatic alignment using a pivot language, and an analogical learning that allows an accurate morphological analysis of terms. These morphological analysis are used to improve the indexing of medical documents. The experiments reported in this paper show the validity of this approach with a 10% improvement in MAP over a standard IR system. Keywords: Morphology, biomedical terminology, alignment, analogical learning, morpho- semantic indexing, biomedical information retrieval. 629 1 Introduction In the biomedical domain, terminologies are the keystone of many applications. They are used for structuring the knowledge as well as retrieving and formalizing information contained in documents. For instance, the well-known MeSH®(Medical Subject Headings) www.nlm.nih.gov/mesh terminology is developed to index the very popular PubMED database (www.pubmed.gov). In most Indo-European languages, biomedical terms also have interesting inner characteristics in that they tend to be complex morphological constructions. Indeed, they are often resulting from the composition of several Greek or Latin roots, prefixes, and suffixes. This morphological complexity is an important point to take into account for basic operations like handling, understanding, translating or building semantic relationships between these terms, and furthermore for higher level applications like machine translation or, as we demonstrate in this paper, Information Retrieval (IR). In this paper, we investigate the development of morphological resources and show how a biomedical Information Retrieval task can benefit from such resources. More precisely, we present several techniques aiming at breaking up a term into its morphological components, namely morphs1, while labeling these morphs with some semantic information. To the contrary of existing studies (Deléger et al., 2008; Markó et al., 2005a, for example) which are chiefly based on human expertise, the techniques proposed here rely on unsupervised or semi-supervised approaches. The original idea at the heart of our approach is to use the multilingualism of existing terminological databases. We exploit Japanese as a pivot language, or more precisely terms written in Kanjis, to help decompose the terms of other languages into morphs and associate them with the corresponding Kanjis, in a fully automatic way. Thus, Kanjis play the role of a semantic representation for morphs. The main advantage of Kanjis in this respect is that Japanese terms can be seen as a concatenation of elementary independent words that may even be found in general language dictionaries. For example, the term photochemotherapy can be translated in Japanese by 光化学法; splitting and aligning these two terms gives: photo ↔ 光 (’light’), chemo ↔ 化学 (’chemistry’, ’medicine’), therapy ↔ 法 …

Content truncated to 500 words