The Method of Construction Knowledge Triples Under Joint Extraction of Entity Relations Based on Distant Supervision

Abstract

Aiming at the problems of cascading mistakes and overlap relation problems in the current methods of relation extraction. A joint entity relation extraction method based on Bert is proposed. In this method, Bert and bilstm are adopted, and attention mechanism is introduced to improve the existing HRL model. By replacing the traditional word2vec word vector with Bert, richer semantic information is considered. The attention mechanism is introduced to consider the dependence between words on the basis of bilstm. At last, entity relation is extracted and entity relation triples are output. The experimental result shows that this method can deal with the overlap relation well and improve the result obviously. Experiments on the NYT10 dataset show that the proposed method has higher accuracy and recall rate. Compared with the latest joint extraction method, the P, R, and F1 values are increased by 5.2%, 4.6%, and 4.86%.