Chinese Medical Question Answering Matching Method Based On Knowledge Graph And Keyword Attention Mechanism

With the development of intelligent information technology,more and more people choose to ask for help on online Q&A community when they encounter health problems.As a result,the automatic Q&A systems for medical community arise,which return responses to users based on the existing medical Q&A records in the community.However,there are usually multiple responses to a question,so it is necessary to automatically select the response among them that best matches the question.This task mainly includes two aspects: 1)constructing an appropriate machine learning model to learn the features of the text to achieve accurate question-and-answer matching;2)improving the accuracy of question-and-answer matching when the number of samples is small.For the first task,this paper proposes a Chinese medical question-and-answer matching method based on knowledge graph and keyword attention mechanism,named KK-BERT.First,the pretrained model BERT is used to learn the features of texts and a keyword attention module is introduced on this basis to better learn the information contained in the keywords in the question-answer pair.Then,the external knowledge in the knowledge graph is added and the entity link between the dataset and the medical knowledge graph is built.The experimental results show that the introduction of external knowledge can effectively improve the performance of the model,and the attention mechanism of keywords can further reduce the knowledge noise.For the second task,this paper investigates the data enhancement methods commonly used in NLP and then uses back translation,synonym replacement,question-and-answer exchange and random transformation to further improve the above methods.The experimental results show that the data enhancement method can effectively improve the performance of the model when the sample size is small.In addition,in the random de-information method,the performance of question-andanswer matching is optimal when the text is transformed with a probability of 15%.