Iterative Set Expansion of Named Entities using the Web

Set expansion refers to expanding a given partial set of “seed” objects into a more complete set. One system that does set expansion is SEAL (Set Expander for Any Language), which expands entities automatically by utilizing resources from the Web in a language independent fashion. In a previous study, SEAL showed good set expansion performance using three seed entities; however, when given a larger set of seeds (e.g., ten), SEAL’s expansion method performs poorly. In this paper, we present an Iterative SEAL (iSEAL), which allows a user to provide many seeds; briefly, iSEAL makes several calls to SEAL, each call using a small number of seeds. We also show that iSEAL can be used in a “bootstrapping” manner, where each call to SEAL uses a mixture of user-provided and self-generated seeds. We show that the bootstrapping version of iSEAL obtains better results than SEAL using fewer user-provided seeds. In addition, we compare the performance of various ranking algorithms used in iSEAL, and show that the choice of ranking method has a small effect on performance when all seeds are user-provided, but a large effect when iSEAL is bootstrapped. In particular, we show that Random Walk with Restart is nearly as good as Bayesian Sets with userprovided seeds, and performs best with bootstrapped seeds.