Learning from Negative Examples in Set-Expansion

Abstract

This paper addresses the task of set-expansion on free text. Set-expansion has been viewed as a problem of generating an extensive list of instances of a concept of interest, given a few examples of the concept as input. Our key contribution is that we show that the concept definition can be significantly improved by specifying some negative examples in the input, along with the positive examples. The state-of-the art centroid-based approach to set-expansion doesn't readily admit the negative examples. We develop an inference-based approach to set-expansion which naturally allows for negative examples and show that it performs significantly better than a strong baseline.