中国科学院自动化研究所机构知识库

Joint extraction of entities and relations from unstructured texts is a crucial task in large-scale knowledge graph construction. Recent methods achieve promising performance but still suffer from some inherent limitations, such as ignorance of the importance of relations in linking entities, overreliance on alignment between entity pairs, and decoding inefficiency. To deal with such problems, we propose a novel joint extraction framework, which is based on (entity, relation) pair linking, a new perspective to solve joint extraction. The framework is called H2O2Net since its decoding process is similar to the decomposition of H2O2. Specifically, two identical components are designed to predict (head entity, relation) and (tail entity, relation) pairs respectively, which is exploited by a linking strategy to generate triples. Such relation plays a natural role of connection, which alleviates the dependency of entity pairs alignment. Experimental results on benchmarks demonstrate that H2O2Net achieves state-of-the-art performance with higher efficiency and delivers consistent performance gain on complex scenarios of different overlapping patterns and multiple triples.