Research On Chinese Medical Entity Relation Extraction Based On Graph Neural Network

In recent years,deep learning methods have made breakthrough achievements in various fields,and the field of medical information extraction has also successfully made great progress through deep learning methods,but most of the medical information extraction research is in the English domain,and there are still many difficulties in the research of Chinese medical entity relation extraction,such as ignoring the semantic relationship of sentences,the problem of overlapping entity relations and the lack of Chinese medical data sets.This thesis focuses on the problems in the entity relation extraction task,and the entity relation extraction algorithm is studied based on graph neural network and combined with the Chinese medical domain.In general,this thesis focuses on three tasks of entity relation extraction in Chinese medicine,and the main research work is as follows:(1)A relation extraction model based on syntactic dependency structure information is proposed for the problem that sentence syntax information is often ignored in Chinese medical text relation extraction tasks.The model captures sentence semantic dependency structure information and sequence information by Graph Convolutional Network(GCN)and Bidirectional Long Short-Term Memory(Bi-LSTM).In addition,a new pruning operation is added to the model considering the effect of noise in the dependency tree.Finally,a multihead attention mechanism is introduced to assign different weights to entities.The effectiveness of the model on the Chinese medical text entity relation extraction task is verified by experimental comparison,with an improvement of 1.94% in F1 value compared with the baseline model.(2)A joint extraction model of Chinese medical entity relations based on graph convolutional neural network and pointer network is proposed for the problem of nested entities and overlapping relations that often occur in medical texts.The model optimizes the structure and introduces a global pointer network.The sentence feature information is learned through Bi-LSTM and attention mechanism,and then the grammatical relations between entities are parsed using GCN layer.Finally,multiple decoders are used for joint extraction of entity relations through the global pointer network.The model learns the graph structure data by GCN,while the problem of overlapping entity relations is solved by using the pointer network.The experimental results show that the model achieves an F1 value of 83.77% on the dataset containing overlapping entity relations,which is better than other comparative models.(3)A semi-supervised learning based Chinese medical entity relation extraction model is proposed to address the problem that the above two models have poor recognition results due to insufficient labeled data in supervised training scenarios.The purpose of introducing semisupervised learning is to improve the performance of the model by using unlabeled Chinese medical text.The training dataset is expanded by Bootstrapping algorithm using limited sample resources for multiple repetitive sampling.The model uses Bi-GRU(Bi-directional Gating Recurrent Unit)to learn sentence sequence information,which reduces the computational effort compared to Bi-LSTM.In addition,the model also introduces Graph Attention Networks(GAT),which learns the weight relations between nodes through a selfattentive mechanism.The experimental results show that the model can achieve entity relation extraction in low-resource scenarios by introducing a semi-supervised algorithm,and the F1 value of the model reaches 84.37%.