Deep Learning Based Relation Extraction Of Birth Defects And The Construction Of Knowledge Graph

Birth defects refer to the abnormal body structure,metabolism or function of the fetus that occurs during its formation and development period.It affects the quality of the birth population and the health of children seriously,so that the quality of life of children can't be guaranteed.In addition,it also brings heavy mental and financial burdens to the patient's family.At present,there are relatively few statistical information about birth defects in China,and there is a lack of systematic integration of information related to them,which is very unfavorable for the prevention and treatment of birth defects.To solve this problem,based on the method of co-occurrence,this research selects the sentences that contain both the birth defects and phenotypes/manifestations,genes or teratogens/drugs from PubMed literatures.First of all,we remove the instances of false positives,and use the relationships defined by UMLS to label the relationship between two entities in the same sentence manually based on the idea of distant supervision to establish a training corpus related to birth defects.Then,three different deep learning based relationship extraction models are trained using the manually labeled corpus,namely Bi-LSTM,PCNN + Attention and BERT + Softmax respectively.Based on the prediction results of these three models,we adopt the idea of majority voting and high confidence coefficient to integrate the prediction result of the above three models,and get the final prediction model to obtain the <entity,relationship,entity> triples.Next,the entity relation triples are stored in Neo4 j,which is a famous Graph Database,to construct a knowledge graph related to the birth defects.Finally,by using the pre-trained relation extraction prediction model,the relationship between two entities can be predicted from the subsequent new co-occurrence sentences.Further,the new triples obtained from the prediction model are added to the knowledge graph.In addition,before the training process of prediction model,we also trained the word vectors in the field of biomedicine based on the Word2 vec algorithm,which can be directly used for subsequent natural language processing tasks in various biomedical fields.This research builds an entity relationship prediction system based on deep learning technology.It crawls literature data from PubMed and store them on the local server in a periodic automatic way,and then uses the pre-trained relationship extraction model to obtain relationship triples from these medical text data and stores them in the local Neo4 j graph database for visualization.Knowledge Graph of birth defects provides query function based on Cypher language,which can query entities and their relationships,and presents the results to users graphically.The results of this study are helpful for the research work of scholars and doctors to a certain extent,and also play a positive role in the prevention and treatment of birth defects.