Research On Healthcare Question Answering Retrieval Algorithms And Evaluation Methods Based On Deep Learning

With the development of "artificial intelligence healthcare",the problems of insufficient supply and uneven distribution of medical resources that are common in medical people’s livelihood are solved.With the popular deep learning technology in the field of artificial intelligence,how to improve the performance in the healthcare question answering retrieval for structured and unstructured medical data and how to build effective evaluation methods remain unsolved.To solve these problems,this thesis conducts in-depth research.The main works are as follows:First,this thesis proposes a similar medical case retrieval algorithm SMCR based on multi-modal patient profile graph to solve the the similar medical pattern missing and data insufficiency problems in healthcare question answering retrieval for structured medical data.SMCR first proposes a novel representation method(i.e.,multi-modal patient profile graph)and then develops the deep graph similarity learning module to compute the similarity between two graphs.This thesis also conducts extensive experiments on the clinical dataset MIMIC-Ⅲ.The experimental results show that the multi-modal patient profile graph in SMCR mitigates the similar medical pattern missing and data insufficiency problem and improves the performance of healthcare question answering retrieval for structured medical data.Second,this thesis proposes a healthcare question answering algorithm CK-HQA based on the context embedding and knowledge embedding to solve the data insufficiency and lacking of context information and knowledge information problems in healthcare question answering retrieval for unstructured medical data.CK-HQA first develops the context encoder and knowledge encoder to obtain the context embedding and knowledge embedding respectively.Secondly,CK-HQA proposes a new joint model to learn the knowledge-enhanced context embedding based on the context embedding and knowledge embedding.Finally,the neural ranking models utilize the knowledge-enhanced context embedding to compute the final ranking score between question and document.This thesis also conducts extensive experiments on two datasets with the IR metrics.The experimental results show that CK-HQA achieves the state-of-the-art performance in healthcare question answering retrieval for unstructured medical data with the knowledge-enhanced context embedding and mitigates the existing problems.Third,this thesis further proposes automatic healthcare question answering evaluation methods HQADeepHelper based on grid search to solve the excessive manual evaluation and the lack of unified evaluation metrics and technology problems.This method first constructs unified evaluation metrics based on the IR metrics,then it proposes a novel automatic evaluation framework based on Grid Search,which combines the data selection and preprocessing module and automatic healthcare question answering retrieval module,to realize several model evaluation methods.This thesis also conducts several automatic experiments to demonstrate the efficiency of the automatic evaluation methods.