SEPC: Improving Joint Extraction of Entities and Relations by Strengthening Entity Pairs Connection

Joint extraction of entities and relations aims at recognizing relational triples (subject s, relation r, object o) from unstructured text. For any entity pair (s, o) in correct relational triples, they do not appear independently, but depending on each other. While existing approaches usually model entity pairs only by sharing the encoder layer, which is insufficient to exploit entity pair intrinsic connection. To solve this problem, we propose to strengthen entity pairs connection (SEPC) by utilizing the duality property of entity pairs, which can further improve the joint extraction. The entity pairs recognization is transformed to finding subject conditioned on the object and finding object conditioned on the subject, and the dual supervised learning is introduced to model their connection. We finally demonstrate the effectiveness of our proposed method on two widely used datasets NYT and WebNLG (Code and data available: https://github.com/zjp9574/SEPC).