Appraising UMLS Coverage for Summarizing Medical Evidence

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 513–524, Osaka, Japan, December 11-17 2016. Appraising UMLS Coverage for Summarizing Medical Evidence Elaheh ShafieiBavani, Mohammad Ebrahimi, Raymond Wong, Fang Chen University of New South Wales, Sydney, Australia Data61 CSIRO, Australia {elahehs,mohammade,wong,fang}@cse.unsw.edu.au Abstract When making clinical decisions, practitioners need to rely on the most relevant evidence avail- able. However, accessing a vast body of medical evidence and confronting the issue of informa- tion overload, can be challenging and time consuming. This paper proposes an effective sum- marizer for medical evidence by utilizing both UMLS and WordNet. Given a clinical query and a set of relevant abstracts, we aim to generate a fluent, well-organized, and compact summary that answers the query. Analysis via ROUGE metrics shows that using WordNet as a general- purpose lexicon helps to capture the concepts not covered by the UMLS Metathesaurus, and hence significantly increases the summarization performance. The effectiveness of our proposed approach is demonstrated by conducting a set of experiments over a specialized evidence-based medicine (EBM) corpus - which has been gathered and annotated for the purpose of biomedical text summarization. 1 Introduction Over the past two decades, clinical guidelines urged practitioners to move towards evidence-based medicine, which is formally defined as conscientious and judicious use of current best evidence in mak- ing decisions about the care of individual patients (Sackett et al., 1996). Evidence-based medical practice heavily relies on research evidence, rather than intuition, unsystematic clinical experience, or pathologic rationale (Group and Others, 1992). However, searching through and evaluating primary medical liter- ature is extremely time consuming (Sarker et al., 2015). Even targeted searches tend to return a large set of relevant documents, and not summaries or answers to the queries. Hence, the explosive growth of content of medical evidence requires development of techniques to present information to physicians and researchers in an effective way. Automatic text summarization has been introduced as a natural language processing technique to address this problem (Reeve et al., 2007). Even though the problem of information overload and the advantages of summarization are critical in the biomedical domain, the majority of summarizers are designed to be general-purpose. They usually work with a simple representation of the summary comprising of information that can be directly ex- tracted from the document, such as terms, phrases, or sentences (Mihalcea and Tarau, 2004). However, recent studies (e.g. (Fiszman et al., 2004)) have demonstrated the benefits of summarization based on richer representations that make use of domain-specific knowledge sources. These approaches represent the documents using concepts instead of words, and may also be enriched by using semantic associations among concepts (e.g. synonymy, hypernymy, etc.) (Plaza et al., 2011). While a query is asked in the field of biomedicine, one of the main challenges is to understand the un- derlying semantic relatedness of the query and document sentences, and consequently extract the most non-redundant, query-relevant parts from the documents. Documents in biomedicine are very different from documents in other fields, and …

Content truncated to 500 words