Extraction And Application Of Key Information In Doctor-patient Communication Text For Esophageal Cancer During Perioperative Period

With the development of the Internet in recent decades,the way people review diseases has gradually changed from going to the hospital to communicating with doctors online through chat software.However,with the increasing number of patients,the number of doctors appears to be slightly thin.The same doctor may need to recheck and chat with multiple patients at the same time,and many patients need to recheck many times.Repeated checking of chat records will greatly waste doctors’ time.Therefore,this thesis constructs a key information extraction system for doctor-patient question and answer,aiming at saving doctors’ time and reducing the workload required by doctors.The key information extraction system of doctorpatient question and answer analyzes the doctor-patient dialogue text entered by doctors,and processes the question and answer text by using named entity recognition,text intention recognition and text matching,and finally returns the correct results according to the key sentence template.The specific research contents are as follows:(1)Named entity recognition system based on Mac BERT and medical terminology.This thesis analyzes the problems of BERT in the recognition of named entities,introduces the Mac BERT model to improve the recognition effect,further improves the rationality of the prediction results through CRF,and introduces the use of the medical domain proper noun dictionary for parallel extraction to ensure that the extracted medical proper nouns can be correctly classified.The accuracy of this model in the named entity recognition dataset in this thesis is 81.2%,which is a certain improvement over other named entity recognition domain models.(2)Text matching model based on twin BERT-Bilstm.This thesis analyzes the common matching methods in the field of text matching,and improves them in view of the existing problems.In addition to using the twin BERT for text matching,it also uses the temporal characteristics of Bilstm to improve the ability of the model to express the sentence context,and then prompts the directional representation of the statement to make the accuracy of similarity matching higher.According to the data set in this thesis,some adjustments have been made to make the model better in the field of doctor-patient question and answer.The model in this thesis achieves 90.4% accuracy in the doctor-patient question and answer data set,which is greatly improved compared with other models in the same field.(3)Sentence intention analysis model based on ERNIE-CNN model.This thesis discusses the difficulties of sentence intention analysis,analyzes the advantages of the ERNIE model,and makes improvements in the field of doctorpatient question and answer.It uses ERNIE to express the sentence vector and CNN to improve the prediction accuracy of the model,so that the model achieves 79.89%accuracy in this data set.At the same time,a key sentence matching model is proposed.The results extracted from the named entity recognition model,text matching,and text classification model are matched according to the template,thus forming the final key sentence.(4)Implementation of key information extraction system for doctor-patient question and answer.Based on the Flask framework,this thesis designs the frontend display interface of the system using network programming technology,integrates the named entity recognition model,text matching model,and text classification model in the back-end,and sends the inherited results to the frontend using Socket,and displays them in the front-end.At the same time,in order to prevent the system from blocking and other problems,the multithreading technology is applied to the system,and the pressure test is passed,which proves the feasibility of the system in the hospital and has high practical value.