Joint Entity and Relation Extraction with Triple Discrimination

Joint entity and relation extraction, extracting relational triples (subject, relation, object) from unstructured natural texts, is a significant task in information extraction and automatic knowledge graph constructions. Existing methods usually utilize identified possible entities to predict relations. However, the extraction results of almost existing methods contain two types of errors: invalid triples and unrecognized positive triples. Intuitively, generating the confidence of triples and learning the gold triples information by utilizing knowledge graph embedding techniques are effective to solve issues above. In this research, we propose a novel joint method with triple discrimination for extracting entities and relations. Specifically, two modules are contained in the proposed method: a triple extractor, a triple discriminator. The former is used to extract triple candidates, and the latter is used to generate the confidence of triples. Experiment results indicate that the proposed approach achieves the state-of-the-art performance on public NYT and WebNLG datasets. Moreover, the results prove the method is effective to reduce errors usually contained in the results obtained by previous methods.