Semantic Judgement of Medical Concepts: Combining Syntagmatic and Paradigmatic Information with the Tensor Encoding Model

Semantic Judgement of Medical Concepts: Combining Syntagmatic and Paradigmatic Information with the Tensor Encoding Model Mike Symonds1, Guido Zuccon2, Bevan Koopman1,2, Peter Bruza1, Anthony Nguyen2 michael.symonds@qut.edu.au, guido.zuccon@csiro.au, bevan.koopman@csiro.au p.bruza@qut.edu.au, anthony.nguyen@csiro.au 1 School of Information Systems, Queensland University of Technology 2 Australian e-Health Research Centre, CSIRO Brisbane, Australia Abstract This paper outlines a novel approach for modelling semantic relationships within medical documents. Medical terminologies contain a rich source of semantic informa- tion critical to a number of techniques in medical informatics, including medical in- formation retrieval. Recent research sug- gests that corpus-driven approaches are ef- fective at automatically capturing seman- tic similarities between medical concepts, thus making them an attractive option for accessing semantic information. Most previous corpus-driven methods only considered syntagmatic associations. In this paper, we adapt a recent approach that explicitly models both syntagmatic and paradigmatic associations. We show that the implicit similarity between certain med- ical concepts can only be modelled using paradigmatic associations. In addition, the inclusion of both types of associations over- comes the sensitivity to the training corpus experienced by previous approaches, mak- ing our method both more effective and more robust. This finding may have impli- cations for researchers in the area of medi- cal information retrieval. 1 Introduction Semantic similarity measures are central to sev- eral techniques used in medical informatics, in- cluding: medical search (Voorhees and Tong, 2011; Cohen and Widdows, 2009), literature- based discovery (e.g., drug discovery (Agarwal and Searls, 2009)), clustering (e.g., gene cluster- ing (Glenisson et al., 2003)), and ontology con- struction or maintenance (Cederberg and Wid- dows, 2003). Automatically determining the similarity be- tween medical concepts presents a number of spe- cific challenges, including vocabulary mismatch. For example, the phrases heart attack and my- ocardial infarction are synonymous, referring to the same medical concept. Beyond vocabulary mismatch are situations where semantic similar- ity is based on implied relationships, for exam- ple the mention of an organism (e.g. Varicella zoster virus) suggests the presence of a disease (e.g. chickenpox). Existing approaches for measuring medical se- mantic similarities fall into two major categories: (i) those that utilise path-based measures be- tween concepts in medical thesauri/ontologies, and (ii) corpus-based approaches that derive sim- ilarity judgements from the occurrence and co- occurrence of concepts within text, e.g., using La- tent Semantic Analysis (LSA). Research comparing path-based methods with corpus-based methods highlighted the ability for corpus-based methods to provide superior per- formance on medical concept similarity judge- ments (Pedersen et al., 2007). However, re- search evaluating eight different corpus-based ap- proaches found that the performance was sensi- tive to the choice of training corpus (Koopman et al., 2012). This finding means it is difficult to ap- ply a corpus-based approach which is both robust and effective. It is important to note that the corpus based ap- proaches referred to in this paper do not rely on syntactic information found in extra-linguistic re- sources, and use solely the co-occurrence statis- tics of words found in natural language to model word associations. Therefore, research modelling …

Content truncated to 500 words