【作者】 王杰；

【导师】 张会雄；

【作者基本信息】 电子科技大学 ， 生物医学工程， 2020， 硕士

【摘要】 随着移动通信技术的发展,越来越多的人在面对健康问题时会通过手机、电脑等电子设备在互联网上搜索相关问题。针对这一现象,各大医疗平台纷纷开发了多个在线医疗问答社区平台,但是面对着越来越多的健康问题,医生的资源是有限的,因此如何通过分析现有的医疗健康问答数据来快速自动回答患者所提出的问题是一个关键的问题。针对这个问题,医疗问题答案如何正确选择是构建自动医疗问答系统的关键内容,因此我们研究了一系列方法尝试从众多候选答案中找出与问题相对应的准确答案。本文主要研究了中文医疗健康领域的问答匹配问题,主要包括两个关键问题:1.如何将文本信息准确的表示为计算机可以处理的向量形式。2.如何构建神经网络模型可以在文本的向量表示中准确捕捉其中的语义信息。针对第一个问题,我们尝试了“分词”与“分字”两种文本表示方法,发现针对中文医疗领域由于目前的分词工具分词的不准确性,直接表示为字向量的方式既可以减少表示矩阵的维度,减少内存和计算需求,还可以提高最终模型的准确率。针对第二个问题,首先我们验证了传统的CNN、RNN以及其变种LSTM与BiGRU神经网络模型在cMedQA、cMedQA2数据集上的问答匹配性能,分析了不同模型的优缺点。然后将多种神经网络模型相结合构建多尺度卷积神经网络模型与BiGRU-CNN模型,验证不同组合模型在多个数据集上的性能,分析不同组合模型的特点。为了进一步评估与提高医疗问答领域问答匹配模型的准确率与泛化能力,我们利用爬虫技术收集了多个医疗健康网站的医疗问答数据,构建了据我们所知目前最大的中文医疗健康问答数据集cMQA,并在上面验证了上述提到的模型。最后将注意力池化与传统神经网络模型相结合构建新的组合模型BiGRUATT与BiGRU-CNN-ATT,验证不同注意力池化组合模型在多个数据集上的性能,证明了引入注意力池化机制后模型的准确率有了一定程度的提升,为构建自动医疗问答系统提供了一定的理论与算法支持。更多还原

【Abstract】 With the development of mobile communication technology,more and more people will search related problems on the Internet through mobile phones,computers and other electronic devices when facing health problems.In response to this phenomenon,major medical platforms have developed multiple online medical question and answer platforms,but facing more and more health problems,doctors ’resources are limited,so how to use existing medical health data to quickly analyze answering questions posed by patients automatically is a key question.In response to these questions,how to choose the correct answers to the medical questions is the key content of building an automatic medical question answering system,so we study a series of methods to try to find the accurate answer corresponding to the question from many candidate answers.This article mainly studies the question and answer matching problem in the field of Chinese medical health,mainly including two key questions: 1.How to accurately represent the text information into a vector form that can be processed by the computer.2.How to build a neural network model which can accurately capture the semantic information in the vector representation of text.For the first problem,we use two text representation methods: "word segmentation" and " character segmentation",and find that for the Chinese medical field,due to the inaccuracy of the current word segmentation tool,the direct representation as a character vector can reduce the dimension of the matrix,reduce memory and computing requirements,and can also improve the accuracy of the final model.For the second question,the matching performance of the traditional CNN,RNN and its variants LSTM and BiGRU neural network models on the cMedQA and cMedQA2 datasets is verified.Then combine a variety of neural network models to build a multiscale convolutional neural network model and BiGRU-CNN model,verify the performance of different combination models on multiple datasets,and analyze the characteristics of different combination models.In order to evaluate and improve the accuracy and generalization ability of the question and answer matching model in the field of medical question and answer,we collect the medical question and answer data from several medical and health websites by using crawler technology,construct the largest Chinese medical health question and answer dataset as far as we know,and verifiy the model mentioned above.Finally,we combine attention pooling with traditional neural network model to construct new combination models BiGRU-ATT and BiGRU-CNN-ATT to verify the performance of different attention pooling combination models on multiple data sets,and prove that the accuracy of the model has been improved to a certain extent after the introduction of attention pooling mechanism,which provides theoretical and algorithmic support for the construction of automatic medical question answering system.更多还原