Research On Key Technologies Of Non-Factoid Medical Question Answering System Based On Deep Learning

Question answering system allows users to ask questions in natural language,which can give accurate answers quickly.It has the characteristics of high real-time,high accuracy and friendly interaction interface.It has attracted wide attention from academia and business circles,and has developed rapidly in various application fields in recent years.At present,the question answering system is mainly based on retrieval.This kind of question answering system can ensure that the returned answers are grammatically correct and fluent.However,the retrieval answering system relies heavily on the existing data sets and can only match the answers in the existing data sets.There are some problems,such as slow matching speed,insufficient real-time performance,and the answers returned are fixed and not diverse.The generative question answering system can automatically sample and generate personalized and diverse answers by understanding the question,which can solve the problem of insufficient real-time and diversity of retrieval question answering system.Recently,text generation tasks usually use sequence-to-sequence learning framework based on attention mechanism as the basic model and the construction of pointer network generation model has achieved good results.The source input or additional knowledge base of this kind of research can provide enough information to copy and guide answer generation.However,in the non-factual medical field,questions and answers are often organized by patients in a more colloquial form according to their own situation.When the source input carries insufficient information,the pointer mechanism can not copy enough effective information,and the model generation effect is not good.In view of the above problems,this paper studies the construction of a generative nonfactual medical question answering system.The goal is to analyze the illness of the non-factual questions put forward by patients,to give recommendations for diagnosis and treatment,and to return them to users in the form of natural language.The main works are as follows:Firstly,in order to solve the problem of over-fitting in the training of deep learning model for non-factual medical question classification task,this paper proposes a convolutional autoecoding network medical question classification model based on char level.The model introduces auto-ecoding structure on the basis of text convolution neural network.Experiments on char-level and word-level semantic units are carried out respectively.Experiments show that the experimental accuracy based on char-level achieves better results.Adding auto-ecoding structure can effectively filter noise and reduce over-fitting.Secondly,in order to solve the problem of insufficient information carried by single source input question,insufficient coding of traditional encoder-decoder model and insufficient copying of effective information by pointer generation network,resulting in poor fluency and correlation index of generating answers,this paper proposes an answer generation model combining graph convolution auto-ecoding inferencing and pointer copy.Considering the exposure bias in the training and inferencing stages,the model uses the graph convolution auto-encoding model to infer the key information of the global answer,and fuses the key information of the question and the answer on the basis of the pointer network.The model chooses the more critical information to guide the generation of the answer.Experiments show that the improved pointer generation model proposed in this paper can effectively improve the correlation and fluency of the generated answers.The medical question answering system constructed in this paper can learn medical knowledge and experience from abundant historical data of user-doctor inquiry.According to the description of user's illness,key features are extracted,and the patient's condition is analyzed.Then,further diagnosis and treatment suggestions are given.Various candidate answers are generated and pushed to the doctor.Finally,the doctor determines the optimal answer and returns it to the user.In addition,the model can automatically generate medical answer template.Compared with the fact-based question answering system,the generated answer carries more information,has longer sentences and is more humanized.