End-to-end Relation Extraction Using Graph Convolutional Network with a Novel Entity Attention

Abstract

There are more and more researches on joint relation extraction, however, the current popular joint extraction method has more or less limitations, either the training time is too long or the effect is not very good. In this paper, we propose an end-to-end relation extraction model, without using handcraft features, and propose a novel graph convolutional neural network based on entity attention mechanism which can perform better feature extraction on tree nodes. In addition, for preserving relevant information on the dependency tree to the greatest extent, we use a path-centric pruning strategy to remove irrelevant content, it makes the model more robust. Our model consists of five parts: Bert layer for vector representation, BiGRU layer, CRF layer for sequence labeling, GCN layer and Predict layer. To evaluate our method, we conduct experiments on the public dataset NYT and ACE05. Our model achieve the state of the art on the task of entity and relation extraction.