Research and Application of Relation Extraction based on Triple Relation Graph Convolutional Networks

Abstract

In recent years, data in non-Euclidean spaces is becoming more and more. Traditional methods cannot perform feature extraction on these data. Most of existing methods just extract contextual semantic features from relational instances. Their structural features in corpora are ignored. To solve this problem, the paper proposed a relation extraction method based on triple relation graph convolutional networks (TRGCN). Based on the extraction of semantic features of sentences using convolutional neural networks, this method used the concept of triple relation graphs to represent structural features. In other words, triple relation graphs were formed by considering triples formed by the relation between two entities in one sentence as nodes and triples with common entities and same relations as edges. Finally, multiple-layer graph convolutional networks were used for training. As shown by experimental results, the method proposed in this paper achieved an F1 value of 86.8% on the SemEval 2010 Tesk 8 data set, indicating that it is better than mainstream convolutional neural networks and recurrent neural networks.