Research On Vertical Domain Question Answering Based On Deep Learning And Knowledge Graph

The purpose of automatic question answering is to allow the machine to understand the questions asked by people and give accurate answers in the form of natural language.This technology requires machines to store a lot of background knowledge.In recent years,the rapid development of knowledge graph technology has led to more research on question answering systems based on knowledge graphs.The research field of automatic question answering can be divided into open field and vertical field.This thesis mainly focuses on the medical vertical field.However,in research there are the following problems:insufficient labeling data in actual scenes and machines have deviations in understanding user problems in various vertical fields.So the aim of this research is to solve the above problems.Firstly,in order to solve the problem of insufficient annotation data when training the model,based on the latest developments in natural language modeling and text enhancement algorithms,a text enhancement method that combines representation-oriented and original text-oriented text enhancement is proposed,and reverse translation technology is used to improve the enhancement.The effectiveness of the text enhancement algorithm proposed in this thesis is verified by variety of text,different scenarios on multiple Chinese and English data sets.Secondly,aiming at the problem that the machine has a biased understanding of the user's problem,the identifying the user's intention is used.Based on the existing intention recognition methods,an improved algorithm using the capsule network is proposed in this thesis.The improved algorithm has the advantages of the capsule network and can solve the problem of some features lost in the pooling stage in the original algorithm.Moreover,the information is divided into different layers to solve more complex problems.The algorithm uses dynamic routing algorithms in the training process to increase the weight of important features and discover more hidden features,thereby the performance of intent recognition is improved.Finally,a knowledge graph for the medical field is constructed in this thesis.An automatic question and answer system in the medical vertical field is realized based on the research of text enhancement and intention recognition in the first two parts,which verifies the feasibility of the algorithm proposed in this thesis in actual scenarios.In addition,the algorithms proposed in this thesis can also be transferred to other vertical fields.