A Two-stage Bootstrapping Algorithm for Relation Extraction
Ang Sun · Recent Advances in Natural Language Processing · 2009
Bootstrapping has been empirically proved to be a powerful method in learning lexico-syntactic patterns for extracting specific relations such as book-author and organizationheadquarters. However, it is not clear how to adapt this method to extract more general relations such as the employment-organization (EMP-ORG) relation. Relations like EMP-ORG are actually a set of relations which involves many nominals such as executive, secretary, officer, editor and soldier. To address this challenge, we propose a two-stage bootstrapping algorithm in this paper. The first stage is a commonly used bootstrapping framework, starting with a small set of seeds (entity pairs) and a large corpus to learn relation patterns which are further used to extract more seeds. We combined it with a second stage bootstrapping which takes as input the relation patterns learned in the first stage and aims to learn relation nominals and their contexts. After the two-stage bootstrapping learning, we incorporate features extracted from learned nominals and their contexts into a state-of-the-art SVM based relation extractor and we observe a 2% gain in F-measure.