Minimising semantic drift with Mutual Exclusion Bootstrapping

Iterative bootstrapping techniques are commonly used to extract lexical semantic resources from raw text. Their major weakness is that, without costly human intervention, the extracted terms (often rapidly) drift from the meaning of the original seed terms. In this paper we proposeMutual Exclusion bootstrapping (MEB) in which multiple semantic classes compete for each extracted term. This significantly reduces the problem of semantic drift by providing boundaries for the semantic classes. We demonstrate the superiority of MEB to standard bootstrapping in extracting named entities from the GoogleWeb 1T 5-grams. Finally, we demonstrate that MEB is a multi-way cut problem over semantic classes, terms and contexts.