Sequence to sequence learning for joint extraction of entities and relations

The relations of two entities contained in the sentence are often complicated. There exists multiple relation tuples which owns one or both the same entities among them. Extracting those overlapped multiple relation tuples faces great challenges. Most existing works employ the famous Seq2Seq model for joint extraction of entities and relations, which cannot handle the multiple relations properly due to the missing of graph structure information of the relations and entities in a sentence. In this paper, we propose Seq2Seq-RE, an end-to-end relation extraction model, which first utilizes the gate graph neural networks (GGNNs) for joint extraction of entities and relations. Unlike previous works, we take the interaction of entities and relations through a GGNNs-based sequence-to-sequence with attention mechanism for better extracting multiple relation tuples. Besides, our graph structure based on the forward-edge, backward-edge, self-edge, and dependency-edge. To our knowledge, we are the first to employ four types of edge into graph-based neural networks for joint extraction of entities and relations and conduct a comprehensive qualitative analysis to investigate what types of edges benefit our task. Experimental results show that our model surpasses the current strong baseline methods by 1.7% in NYT29 and 0.8% in NYT24 (F1 score). (c) 2022 Published by Elsevier B.V.