Research On Intelligent Question Answering System Based On Domain Knowledge Graph

This thesis uses deep learning technology to build a knowledge graph in the medical field.As an example,splitting the process from user input to answer feedback into three stages.A question and answer system based on domain knowledge graph is formed,aiming to integrate domain knowledge and provide users with simple and fast information retrieval services.Based on this objective,the main research of this thesis includes the following three aspects:(1)Firstly,a Chinese NER model DB-Attention is proposed in this thesis.To avoid the problem of too small word separation granularity,the model uses character-level embeddings incorporating lexical information as input,ensuring that important word information is not lost.The user input questions are extracted by the Biaffine dual affine attention mechanism and the lexical information is fused in the embedding layer.Experiments show that Glo Ve word embedding this method of using co-occurrence matrices instead of global matrix singular value decomposition.It not only allows for smoother training,but also effectively saves around 4% of system overhead while taking into account global corpus information.(2)Secondly,an improved intention parsing model Bi-Text CNN is proposed.firstly,A set of n-gram pre-trained word embeddings are added to the Text CNN model as static semantic features.On top of this multiple convolutional kernels are used for feature extraction of both word embeddings.Finally maximum pooling is performed to obtain the classification probability of the utterance.Experiments show that Glo Ve word embeddings,which use cooccurrence matrix instead of global matrix singular value decomposition,can not only make the training smoother,but also can effectively save about 4% of system overhead while considering global corpus information.(3)Finally,a detailed overall design of the intelligent Q&A system based on the knowledge graph of the medical and health care domain is made.To make a good data support for the Q&A system.40,000 medical-related entities and 240,000 inter-entity relationships were integrated,and the construction of a knowledge graph in the medical domain was completed.Afterwards,the DB-Attention model and Bi-Text CNN model were integrated to build an intelligent Q&A system with B/S architecture.Finally,each module of the system was elaborated and the front-end page of the Q&A system was demonstrated.Through testing,the accuracy rate of the system was 66.3%,which verified the usability of the system.The model proposed in this thesis achieves good results in the entity recognition task,not only improving the efficiency of recognizing entities,but also supporting the recognition of nested entities.In addition,the dual-channel intent resolution model also achieves some performance advantages in terms of both recognition performance and system overhead,when considering global information.The domain knowledge graph-based medical Q&A system exemplifies the superiority of a new type of information retrieval tool.It provides users with a simple and reliable alternative to search engines for information retrieval.