Research On Entity Relation Extraction Based On Adversarial Learning And Global Pointer Generation Network

The existing knowledge map is usually used as the external knowledge base of search engine and intelligent customer service to enhance the task performance.Entity relation extraction is an important basic method in the construction of knowledge graph,which aims to automatically extract relation triples from existing massive unstructured texts.In addition,entity relation extraction method has important applications in text extraction of key information in vertical fields such as finance,tourism,and medical care.Therefore,the method of entity relation extraction has received extensive attention from industry and academia,and how to construct an efficient and accurate entity relation extraction model has become a research hotspot.The entity relation extraction method adopts a pipeline method to independently perform the two sub-tasks of entity recognition and relation classification,which lacks interaction,resulting in a large amount of error propagation.Therefore,the joint extraction method has become the mainstream method of research because of its good interaction.The development of deep learning has greatly improved the accuracy of the end-to-end entity relation extraction model,but the problems brought about by the improvement of model performance are that there are more model parameters and a wider and deeper network structure.As a result,the model's scale is larger and its inference time is longer,which makes it difficult to train and deploy on servers with limited computing resources.Aiming at the above problems,this paper firstly proposes an entity relation extraction model AGPGN based on adversarial learning and global pointer generation network.The model is based on the shared encoding layer of the BERT model as the interactive dependency of entity recognition and relation classification.At the same time,a global pointer generation network that integrates specific relation features is proposed in entity recognition to better deal with the problem of entity overlapping in relation extraction.In order to improve the generalization ability of the model,the adversarial training based on PGD is added to the model training.The performance on the dataset improves the F1 value of the benchmark model by 0.8%,outperforming other existing models.Then this paper applies knowledge distillation to entity relation extraction model.Knowledge distillation is a common way of model compression.In this model,the model AGPGN based on adversarial learning and global pointer generation network is used as the teacher model,the LSTM-encoded global pointer generation network model is used as the student model,and the BERT self-distillation-based global pointer generation network model is used as the student model.Experiments are carried out through the method of knowledge distillation.The experimental results show that the model can still achieve good results after knowledge distillation.