One recent estimate for the annual growth rate of scientific
publications is 8% per year (Bornmann & Mutz, JAIST 2014), and in some
discplines growth is even higher. This creates problems for
scientific understanding: researchers and decision makers want to be
informed by the best, most recent, and most reliable scientific data.
One approach to address the rising amount of published literature is
reliance on systematic reviews: meta reviews that comprehensively
survey the extant literature and synthesize results from previously
published studies.
This talk discusses the use of supervised text classification to
partially automate the literature search process in systematic
reviews. A systematic review is unlike most information retrieval
applications because it is important to not only find documents that
are useful (e.g., as in Web search), but to find the preponderance of
documents that meet specified selection criteria. Similar "high
recall" problems arise in patent application review and legal
e-discovery. Some experimental results will be presented from a
recent collaboration with the U.S. Centers for Disease Control and
Prevention to identify scientific articles about methods to improve
health care worker effectiveness in the developing world.
Paul McNamee is a member of the Principal Professional Staff at the
Johns Hopkins University Applied Physics Laboratory and a Senior
Research Scientist with the JHU Human Language Technology Center of
Excellence. Dr. McNamee earned a Ph.D. in computer science from the
University of Maryland Baltimore County in 2008, where his
dissertation research focused on innovative methods for effective
multilingual text retrieval. He has published over 75 scholarly
publications related to the automated processing of language data and
he has participated in numerous international competitions for text
retrieval technologies.