Research On Chinese Electronic Medical Record Entities Recognition And Entity Relation Extraction Based On Semi-Supervised Learning

The electronic medical record is a digital record written by the medical staff describing individual medical activities for a patient.The electronic medical records contain comprehensive patient individual health information.Analyzing and mining electronic medical records entities and extracting relation information can help to build clinical decision support systems and provide personalized health information services.Due to electronic medical records' unique textual features and structural features,it is difficult to apply the traditional methods of named-entity recognition and relation extraction in medical fields.Therefore,this paper proposes a Chinese electronic medical record named-entity recognition and relation extraction methods based on semi-supervised.Extensive experiments based on our datasets have shown that this approach has great potential.There are three main contributions to this research,as follows:(1)This paper analyzes the structural features of the electronic medical records'text language and the electronic medical record itself,and the differences in the data tagging specifications in the medical fields and evaluation methods.And based on the existing entities and relations' tagging specification,a small corpus is.(2)The joint model of Bidirectional RNN and the CRF is one of the most widely used algorithm in the named-entity recognition tasks.Based on it,we add a new recurrent neural network to extract more text language features and structural features in a small amount of label corpus and a large number of unlabeled data.Extensive experiments based on small-scale tagged corpora have shown that our model not only is more suitable for Chinese electronic medical records but also can effectively achieve better recognition results and get a higher F-value.(3)In traditional relation extraction methods,the two tasks of named-entity recognition and relation extraction studied as independent tasks,ignoring the correlation between these two tasks.By changing the tagging method of relationships,named-entity recognition and relation extraction are trained as joint tasks.Experiments show that the improved relation extraction model can complete the extraction task under a small number of training sets.