Relation Extraction with Matrix Factorization and Universal Schemas

摘要: © 2013 Association for Computational Linguistics. Traditional relation extraction predicts relations within some fixed and finite target schema. Machine learning approaches to this task require either manual annotation or, in the case of distant supervision, existing structured sources same The need datasets can be avoided by using a universal schema: union all involved schemas (surface form predicates as OpenIE, preexisting databases). This schema has an almost unlimited set (due surface forms), supports integration with data (through types To populate database such we present matrix factorization models that learn latent feature vectors entity tuples relations. We show achieve substantially higher accuracy than traditional classification approach. More importantly, operating simultaneously on observed text pre-existing DBs Freebase, are able reason about unstructured mutually-supporting ways. By doing so our approach outperforms stateof- the-Art supervision.