Jiho Kim, Sangha Nam and Key-Sun Choi.
Abstract: Universal Schemas are a remarkable approach in solving relation extraction. Universal Schemas can be constructed by a union of all usable schemas whose relation types come from both knowledge bases and surface form predicates. New facts are extracted by jointly learning latent feature vectors of entity pairs and relation types through matrix factorization models. However, in free word order languages where surface form predicates do not fully reveal the characteristics of a sentence, Universal Schemas cannot be constructed in the same way. In this study, we introduce a novel expansion of Universal Schemas: using shortest dependency paths, which contain rich linguistic features, and entity types instead of surface form predicates. For performance verification, we constructed and evaluated a Universal Schema in Korean with a combination of Korean DBpedia and Korean Wikipedia distant supervision data.
Keywords: Universal Schemas; Relation Extraction; Shortest Dependency Paths; Knowledge Base Completion