Intention Recognition Of Medical Question And Answer Based On Knowledge Graph

Problem intent recognition(also known as problem analysis)refers to detecting the true intent of the problem through various methods.At present,problem intent recognition methods are divided into two categories: a class of retrieval-based problem recognition methods.This type of method implements the similarity calculation between a given problem and the problem in the knowledge base,and uses the problem with the maximum similarity as the matching result to achieve Problem analysis,such methods fail to consider sentence structure information,resulting in low recognition accuracy.The other type is the problem intent recognition method based on knowledge graph.This type of method has template-based problem analysis and semantic-based problem analysis method.Among them,template-based problem analysis method needs to manually build a large number of problem templates for matching with problems.Semantic-based problem analysis methods need to manually construct a large number of entity mapping tables and relationship mapping tables.Both of the above knowledge graph-based problem intention recognition methods have a large workload.Aiming at the problems of existing problem intention recognition methods,this paper proposes a problem intention recognition method that combines named entity recognition and problem type recognition.This method uses named entities to identify named entities in a sentence,uses question type recognition to identify sentence categories,and finally combines named entity recognition and question type recognition to transform a sentence into a subgraph of the knowledge graph,and realizes problem intention recognition through matching.The specific research contents are as follows:(1)Improvement of named entity recognition: There are two improvements in the named entity recognition part of problem intention recognition.One is in the word embedding part of named entity recognition.It is proposed to use skipgram pre-trained word embedding matrix,which is better than traditional random initialization word.The embedded matrix is easier to train.Another point is that in the part of the recurrent neural network for named entity recognition,a method of optimizing LSTM with peephole connection is proposed to make the neural network learn sentence structure better.Experiments show that these two optimizations are effective.(2)Improvement of problem category recognition: In the problem type recognition of problem intention output,it is proposed to add the sentence pattern representation of the sentence on the basis of the word representation of the traditional sentence.Experiments show that this addition can indeed improve the accuracy of problem classification in the future.(3)Improvement of problem intention recognition: In problem intention recognition,a method combining named entity recognition and problem type recognition is proposed to realize problem intention recognition.Experiments show that this method can greatly reduce the workload of constructing manual rules compared with the traditional problem intention recognition method.