Abstract
Joint entity and relation extraction is an essential task in natural languageprocessing and knowledge graph construction. Existing approaches usuallydecompose the joint extraction task into several basic modules or processingsteps to make it easy to conduct. However, such a paradigm ignores the factthat the three elements of a triple are interdependent and indivisible.Therefore, previous joint methods suffer from the problems of cascading errorsand redundant information. To address these issues, in this paper, we propose anovel joint entity and relation extraction model, named OneRel, which castsjoint extraction as a fine-grained triple classification problem. Specifically,our model consists of a scoring-based classifier and a relation-specific hornstagging strategy. The former evaluates whether a token pair and a relationbelong to a factual triple. The latter ensures a simple but effective decodingprocess. Extensive experimental results on two widely used datasets demonstratethat the proposed method performs better than the state-of-the-art baselines,and delivers consistent performance gain on complex scenarios of variousoverlapping patterns and multiple triples.