A Span-based Model for Joint Entity and Relation Extraction with Relational Graphs

Extracting related entity pairs from natural language texts is essential to constructing knowledge graphs. However, prior work is limited by the long distance between entities and cannot accurately extract entities and relationships in the case of relatively complex overlap. To address these limitations, we present Span-RG, a span-based relation extraction model which uses graph convolutional networks to jointly learn entities and relations. First, a boundary prediction task is designed to determine whether a word is an entity boundary, which enhances the span representation to obtain a high quality candidate entity spans. Then, we improve GCN and constructe a span-based entity relationship graph, where the edge is the relationship and the node is the span. We evaluate Span-RG on three public datasets: ACE05, NYT and WebNLG. Experimental results demonstrate that the performance of the model is significantly better than the baseline method.