论文部分内容阅读
【目的】利用Word2Vec深度学习技术从面向大众的健康信息中寻找疾病关联,解决非医学人士通常不了解多种疾病之间存在的关联,从而影响到健康信息搜寻中的全面性和有效性的问题。【方法】由专家选取30个常见疾病主题,从高质量医学新闻网站上采集对应疾病的文档,运用Word2Vec技术对各疾病的相关文档构造词向量,计算向量距离判断疾病关联。通过与专家评分的相关分析衡量判断结果的准确性。【结果】最优情况下,Word2Vec得到的结果与专家评分相关系数达到0.635。通过对比不同的算法模型、优化方法、数据规模及重要参数对结果的影响,发现Skip-Gram模型结合负样本数为20的Negative Sampling优化方法在大规模数据集上的实验结果最优。【局限】疾病主题选取宽泛时,影响Word2Vec判断准确性,本文的疾病主题选取粒度有待改善。【结论】利用Word2Vec技术在面向大众的健康信息源中也可以探测疾病关联,其有效性表明该技术可用于改善大众的健康信息搜寻的个性化服务。
【Objective】 Using Word2Vec deep learning technology to find disease association from public health information and solving the problem that non-medical people generally do not understand the relationship between multiple diseases, thus affecting the comprehensiveness and validity of health information search . 【Methods】 30 common disease topics were selected by experts. The corresponding disease documents were collected from high-quality medical news websites. Word2Vec was used to construct word vectors of related documents for each disease and vector distance was used to judge the disease association. The accuracy of judgment results is measured by correlation analysis with expert scores. 【Result】 Under the best condition, the correlation coefficient between Word2Vec and expert score reached 0.635. By comparing the effects of different algorithm models, optimization methods, data size and important parameters on the results, it is found that the Skip-Gram model combined with Negative Sampling optimization method with negative sample number 20 has the best experimental results on large-scale data sets. [Limitations] Wide selection of disease topics, affecting Word2Vec to determine the accuracy of the selection of the disease theme needs to be improved. [Conclusion] Using Word2Vec technology to detect disease association in health-oriented health information sources, its effectiveness shows that the technology can be used to improve the public health information search personalized service.