摘要
电子病历中包含着医疗领域的丰富知识,对于医疗健康信息服务有着重要的意义。其中的概念实体之间的关系是医疗知识的重要组成部分,对于获取医疗领域中疾病、治疗、检查之间关系有着重要的意义。针对于电子病历中文本结构稀疏的特点,原有的基于词的特征表示效果有限,所以从特征选择的角度出发,提出了一种基于深度学习的特征学习,将有限的上下文特征进行进一步抽象表示的方法。实验中使用深度稀疏自动编码来对实体上下文的向量表示进行再表示,来得到更抽象和更有识别意义的特征。实验表明,本文使用的深度学习进行特征的再表示方法对于识别的召回率对比于基线实验有比较明显的提高。
Electronic medical records contain huge quantity of medical knowledge, and it has great importance to the clini- cal decision support system. The relations of concepts and entities are very important in the medical knowledge and have significance in getting the relation of diseases, treatment and test. According to the sparsity of the text in the EMR, original method based on the word feature can be limited. This paper starts from the feature selection and makes a research on the feature learning based on deep learning to extract abstract features from the limited context among the entities. Then this pa- per uses the deep sparse auto - encoder to make a representation of the vector of context for getting more abstract and Dis- criminative features. The experiment shows that the method of learning features by deep architecture can reach a better re- suit than the baseline experiment by improving the recall rate of the relation extraction.
出处
《智能计算机与应用》
2014年第3期35-38,41,共5页
Intelligent Computer and Applications
关键词
电子病历
实体关系抽取
特征选择
深度学习
EMRs
Entity Relation Extraction
Feature Selection
Deep Architecture
