An Advanced BERT-Based Decomposition Method for Joint Extraction of Entities and Relations

Abstract

Joint extraction of entities and relations is an important task in the field of natural language processing and the basis of many NLP high-level tasks. However, most existing joint models cannot solve the problem of overlapping triples well. We propose an efficient end-to-end model for joint extraction of entities and overlapping relations. Firstly, the BERT pre-training model is introduced to model the text more finely. Next, We decompose triples extraction into two subtasks: head entity extraction and tail entity extraction, which solves the problem of single entity overlap in the triples. Then, We divide the tail entity extraction into three parallel extraction sub-processes to solve entity pair overlap problem of triples, that is the relation overlap problem. Finally, We transform each extraction sub-process into sequence tag task. We evaluate our model on the New York Times (NYT) dataset and achieve overwhelming results compared with most of the current models, Precise =0.870, Recall = 0.851, and F1 = 0.860. The experimental results show that our model is effective in dealing with triples overlap problem.