Literature underpins research, providing the foundation for new ideas. But as the pace of science accelerates, many researchers struggle to stay current. To expedite their searches, some scientists leverage hypothesis generation (HG) systems, which can automatically inspect published papers to uncover novel implicit connections. With no foreseeable end to the driving pace of research, we expect these systems will become crucial for productive scientists, and later form the basis of intelligent automated discovery systems. Yet, many resort to expert analysis to validate such systems. This process is slow, hard to reproduce, and takes time away from other researchers. Therefore, we present a novel method to validate HG systems, which both scales to large validation sets and does not require expert input. We also introduce a number of new metrics to automatically identify plausible generated hypotheses. Through the study of published, highly cited, and noise predicates, we devise a validation challenge, which allows us to evaluate the performance of a HG system. Using an in-progress system, MOLIERE, as a case-study, we show the utility of our validation and ranking methods. So that others may reproduce our results, we provide our code, validation data, and results at bit.ly/2EtVshN.
View on arXiv