Evaluation of Literature-Based Discovery Systems
edited by: J. Mackenzie Owen, Peter Bruza, Marc Weeber
Evaluating discovery systems is a fundamentally challenging task because if they are successful, by definition they are capturing new knowledge that has yet to be proven useful. To overcome this difficulty, many researchers in literature-based discovery (LBD) replicated Swanson’s discoveries to evaluate the performance of their systems. They reported overall success if one of the discoveries generated by their system was the same as Swanson’s discovery. This type of evaluation is powerful yet incomplete because it does not inform us about the quality of the rest of the discoveries identified by the system nor does it test the generalizability of the results. Recently, alternative evaluation methods have been designed to provide more information on the overall performance of the systems. The purpose of this chapter is to review and analyze the current evaluation methods for LBD systems and to discuss potential ways to use these evaluation methods for comparing performance of different systems, rather than reporting the performance of only one system. We will also summarize the current approaches used to evaluate the graphical user interfaces of LBD systems.