An Improved Relation Extraction Method Based on Information Control Injection and Attention-Guided Densely Connected Graph Convolutional Network
Zirui Zhang, Fanfang Meng, Xiaoxia Liu, Yuanhui Meng, Yiyu Yang, Benhui Chen · 2023
The extraction of entities and multi-type relations in overlapping triples has been a challenging problem. This paper proposes a relation extraction method based on information control injection and attention-guided densely connected graph convolutional networks by dividing the model into two stages inspired by the GraphRel [1] approach. Considering that the shallow graph convolutional network (GCN) used by GraphRel can only capture local structural information on large graphs, this paper uses the multi-head attention mechanism to form different weight matrices. It extracts deeper structural information in the text by combining it with a densely connected graph convolutional network (DCGCN). In addition, the features extracted by DCGCN are not entirely correct. Therefore, this paper uses KL-divergence to control the degree of information injection to obtain a better feature representation. To verify the model's effectiveness, we conducted extensive experiments on two widely used public datasets: NYT and WebNLG. The results show that compared to GraphRel, the F1 value of the model improves by 24.8% on the NYT dataset and 35.9% on the WebNLG dataset. More importantly, the model significantly improved in extracting overlapping relations.