Research On Knowledge Extraction Method For Intelligent Medical Guidance

With the development of knowledge graph and artificial intelligence,the development of intelligent guidance is becoming faster and faster.The intelligent medical guidance system provides clear medical information for patients on the basis of the knowledge map medical Knowledge Extraction method,and can also provide doctors with accurate patient information feedback.Due to the existence of problems such as "difficulty in seeking medical treatment",intelligent guidance systems represented by Baidu Medical and JD Medical have emerged,quickly occupying half of the medical online Q&A field.The data in the field of intelligent medical guidance has the characteristics of large data volume,high professional level of terminology,and obscure and difficult to understand medical language.Therefore,there are two problems with the current intelligent guidance system.For patients,the accuracy of intelligent guidance recognition is low and cannot provide more effective case information about the disease;On the other hand,for doctors,intelligent guidance has low flexibility and cannot accurately understand patient information.In order to solve the above problems,this thesis studied the knowledge extraction method for intelligent medical guidance with the help of knowledge mapping technology,including the following research contents:(1)To solve the problem of low accuracy in intelligent guidance recognition,this thesis proposed a entity recognition method integrating BERT network and position features(A Entity Recognition Method Integrating BERT Network and Position Features,BERT_POS).This method proposed a new recognition strategy to identify keywords based on the extraction of word vector features in the BERT network.On this basis,an improved calculation method based on relative position features was proposed,which considered words with relative position as the benchmark and absolute distance as the starting point,ensuring the extraction of meaningful entity information.Through entity recognition experiments,it had been proven that BERT_POS method improved the P indicator by 0.15% on the medical data set CEMRNER, by about 1% on the data set cMedQANER,and by 0.1% on the F1 indicator.(2)To solve the problem of low flexibility in intelligent guidance,this thesis proposed a relationship extraction method based on CNN model and word sentence level Transformer network(A Relationship Extraction Method Based on CNN Model and Word Sentence Level Transformer Network,WS_CNN_Transformer).After predefined 19 relationship categories,this method distinguished synonyms from non synonyms by improving the word embedding layer of multiple sliding windows,enhanced the semantic features of text using CNN layers with different convolutional kernel heights,and improved the Multi-head Attention mechanism in Transformer,proposing a Word Sentence Level Multi-head Attention mechanism,realized the weight change of all text features at the sentence level to improve the accuracy of relationship extraction.Furthermore,entity recognition and relationship extraction were jointly optimized,and a knowledge extraction method for intelligent medical guidance was proposed.Through relationship extraction experiments,it had been proven that WS_CNN_Transformer method improved the P indicator by 1% on the medical data set CEMRNER,by about 1% on the data set cMedQANER,and by 0.3% on the F1 indicator.Through the knowledge extraction experiment,it was proved that the P index and F1 index of the knowledge extraction method on the medical data set CEMRNER had a weak improvement,and the P index and F1 index on the data set cMedQANER also had a weak improvement.