Diabetes Knowledge-base Question Answering Based On BERT-BiGRU And Global Pointer Network

With the rapid development of the Internet,people need more and more knowledge and answers from the Internet.The traditional search engine based on keyword matching has been unable to meet people’s needs.Question answering in knowledge-base alleviates the above problems.It can get answers by parsing natural language questions and then querying knowledge-base.In recent years,the rise of knowledge graph has brought about structured and high-quality knowledgebase,and the ability of knowledge representation and feature extraction in deep learning has become stronger and stronger.Question answering of knowledge base based on deep learning has become a hot research direction.In the knowledge-based question answering based on deep learning,the BERT model does not fully mine the contextual information between the input question vector sequences.For long nested named entity recognition,the Global Pointer network based method takes up too large a proportion of negative samples,which will drown the loss of positive samples.For the above problems,the research content of this paper is mainly divided into the following three parts:(Ⅰ)The BERT model is insufficient to mine the context information between the input question vector sequences,so the BERT-BiGRU model is proposed,which can extract the context association information between sequences from the output vector sequences of BERT.The results of the ablation experiment show that the accuracy of the three main tasks in the question answering based on deep learning: named entity recognition task,relational vector representation task,and entity vector representation task is improved in the three tasks and the final answer retrieval task.In addition,multi-task learning is introduced.Experimental data show the effectiveness and enhancement of BiGRU for a certain task.(2)For long nested named entity recognition based on Global Pointer network,the negative sample proportion is too large.In this paper,we propose a long nested named entity recognition based on Global Pointer network and multi-label text classification.The model takes multi-label text classification as an auxiliary task and selects multi-category tags whose entity type is input text.In this way,the proportion of positive samples is increased in the training,and the problem that the proportion of negative samples is too large in the classification of a single Global Pointer network is alleviated.Finally,the experimental results show that the recognition accuracy of long nested named entities is improved on three medical data sets with long nested named entities.The effectiveness of BERT-BiGRU is also verified.(3)In order to alleviate the doctor-patient conflict in the field of diabetes,this paper designs and implements a diabetes knowledge-base question answering system.The system is based on the knowledge-base question answering and long nested named entity recognition tasks of deep learning in the first two parts.Firstly,Dia KG is imported into the graph database Neo4 j,and then the interface of the former model and the database query are encapsulated by the Flask back-end framework.Finally,the Vue front-end framework is used to design the web page and display the results.Through the above work,this paper improves the BERT model and the Global Pointer network to improve the semantic parsing ability of questions,and then uses the improved algorithm and Dia KG diabetes knowledge graph to realize a diabetes question answering system,which can solve the needs of users.