A clustering based approach for domain relevant relation extraction

Most existing corpus based relation extraction techniques focus on predefined relations. In this paper, a clustering based method is presented for domain relevant relation extraction including both relation type discovery and relation instance extraction. Given two raw corpora, one in the general domain, one in an application domain, domain specific verbs connecting different instances are extracted based on syntactic dependency as well as a small set of domain concept instance seeds. Relation types are then discovered based on verb clustering followed by relation instance extraction. The proposed approach requires no predefined relation types, no prior training of domain knowledge, and no need for manually annotated corpora. This method is applicable to any domain corpus and it is especially useful for knowledge-limited and resource-limited domains. Evaluations conducted on Chinese football domain for relation extraction show that the approach discovers various relations with good performance.