Extraction of key-phrases from biomedical full-text with supervised learning techniques

Key-phrase extraction plays useful a role in the research area of Information Systems (IS) such as digital libraries. Short metadata like key phrases could be beneficial for searchers to understand the concepts of documents' concept. This paper evaluates the effectiveness of different supervised learning techniques on biomedical full-text: Naïve Bayes, linear regression, SVMs (reg1/2), all of which could be embedded inside an IS for document search. We use these techniques to extract key phrases from PubMed. We evaluate the performance of these systems using the well-established holdout validation method. The contributions of the paper are comparison among different classifier techniques, and a comparison of performance differences between full-text and abstract. We conducted experiments and found that SVMreg-1 improves the performance of key-phrase extraction from full-text while Naïve Bayes improves from the abstracts. These techniques should be considered for use in information system search functionality. Additional research issues also are identified.