A Novel Text Relation Extraction Model Based on PCNNs with Dynamic Learning Rates

作     者：He, Chen Yuan, Ying-Chun Liu, Zhen-Chao Guo, Jiang-Hua 

作者机构：Department of Information Science and Technology Hebei Agricultural University No 289 Lingyusi Street Hebei Baoding China Hebei Key Laboratory of Agricultural Big Data Hebei Agricultural University China 

出 版 物：《Journal of Network Intelligence》 (J. Network Intell.)

年 卷 期：2022年第7卷第4期

页      面：921-934页

核心收录：

基　　金：Acknowledgements. The research in this paper was supported by Research and Practice Project of Higher Education Teaching Reform in Hebei Province (grant number 2020GJJG076  2018GJJG140)  Hebei Province Master’s Student Innovation Ability Training Funding Project in 2022 (grant number CXZZSS2022100)The research in this paper was supported by Research and Prac-tice Project of Higher Education Teaching Reform in Hebei Province (grant number 2020GJJG076  2018GJJG140)  Hebei Province Master’s Student Innovation Ability Training Funding Project in 2022 (grant number CXZZSS2022100) 

主　　题：Extraction 

摘      要：Distant Supervised Relation Extraction (DSRE) is the mainstream relation extraction field, but most extraction models use a fixed learning rate, which leads to the model constantly learning noise labels when performing relation classification. PCNNs is widely used in distantly supervised relation extraction. In this paper, a model named ERNIE-DLRPCNNs is proposed, which combines pre-training model ERNIE and dynamic learning rate without fixed period into PCNNs. Globally, the convergence of the whole model is improved by gradient accumulation, and adds textual semantic representation with the ERNIE model. Locally, the whole training process is stratified by dynamic learning rate, and the learning rate of each layer is adjusted to improve the accuracy of the relational extraction model. When tested on the well-known New York Times (NYT) dataset, our ERNIE-DLRPCNNs model has a prediction accuracy of up to 82.0 percent, outperforming existing mainstream relational models such as GPCNNs in terms of P-R curve and AUC improvement. © 2022.