Abstract: In view of the complex structure of the existing entity relationship extraction model and the poor extraction effect, a joint extraction method of entity relationship based on pre -training coding feature fusion of bidirectional encoder representation from transformers (BERT) and convolutional neural network (CNN) is proposed. First, the position of the beginning and end of the subject is predicted based on the sentence vector encoded by BERTCNN. Secondly, the feature vector of the predicted first and last position index sentence is used as the head and tail vector of the predicted subject, and then the head and tail vectors of the predicted subject are feature-fused to obtain the subject vectors based on the product method. Then, the subject vector and the sentence vector are fused by the product method to obtain a new sentence encoding vector, which guides the prediction of the head and tail positions of the object under different relationships and obtains entity relationship triplet. In order to verify the effect of the proposed model, this model and other similar algorithms are tested with the public data sets of NYT and WebNLG. The accuracy and recall rate of this model are better than those of the compared model, and the F1 values reach to 92.75% and 93.19%, respectively

Key words: font-size:15.04px, ">bidirectional encoder representation from transformers(BERT), convolutional neural network(CNN), feature fusion, binary classification, joint extraction of entity relationship, entity relationship triplet