Research on entity relation extraction based on active learning
Relation classification is an important NLP task to extract relations between entities.In this paper,we report our method for a largest schema-based Chinese information extraction dataset.We incorporate BERT into a new framework and apply active learning for joint entity relation extraction.This model extends existing approaches from three perspectives.First,our method could solve the problem that multiple entities belongs to multiple triplets.We design this framework based on the idea of probabil-ity graph and develop a new"head-tail"labeling method.Second,we proposed an innovative approach that apply active learning on relation extraction problem.Third to transmit information between subject entities,predicate and object entity,we propose a new normalization method called conditional layer normalization.fourthly,a new loss function is designed to avoid class imbalance.Therefore,we enhance the information extracting ability of the model and achieve F1-score 0.840 on test set with a single model,and achieve better performance with much less data than the original deep models trained by full data.
NETL
NSTL
万方数据
