基于弱监督学习的细粒度中医临床医学实体识别方法研究

项目名称： 基于弱监督学习的细粒度中医临床医学实体识别方法研究

项目编号： No.61501063

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 王亚强

作者单位： 成都信息工程大学

项目金额： 19万元

中文摘要： 中医临床医学实体识别是实现计算机准确地理解半结构化和非结构化中医临床记录的关键任务。目前中医临床医学实体识别主要基于需要满足“强监督假设”条件的训练数据，构建有监督序列化模型，采用语块划分的方法直接从中医临床记录中识别粗粒度的医学实体。该类方法存在（1）复合型中医临床医学实体识别结果不能准确表示；（2）充分满足“强监督假设”条件的训练数据在实际条件下难以构建的问题。根据前期研究发现，弱监督学习和细粒度命名实体识别方法是解决中医临床医学实体识别现存问题的重要手段。因此，本课题拟以中医临床记录中的“主诉和现病史”为研究载体，以其中包含的中医临床医学实体（如症状实体、疾病实体等）为研究对象，依据弱监督学习和细粒度命名实体识别方法的框架和最新成果，开展对高效、鲁棒且实用的中医临床医学实体识别模型和算法的研究。从而，为中医临床医学实体识别提供新思路和新方法，推进中医信息获取与处理领域的研究进展。

中文关键词： 中医信息获取；中医信息处理；语义建模；语义信息；语义关联

英文摘要： Medical entity recognition from clinical records of traditional Chinese medicine (TCM) is the key to achieve better understanding of the semi-structure and unstructured clinical records based on computer. Most of the currently existing methods used to recognize medical entities from clinical records of TCM are based on supervised sequential models which need satisfy strong supervision assumption, and they recognize the medical entities directly through chunking in coarse-grained forms. While these methods have some disadvantages that (1) they cannot appropriately and accurately recognize the composite entities which are frequently appear in clinical records of TCM; (2) it is impossible to construct a training dataset which satisfy the strong supervision assumption in practice. Recently, we find that weakly-supervised learning and fine-grained entity recognition methods could be used to cover these disadvantages. Therefore, in this project, we would study on the methods of fine grained medical entity recognition from clinical records of TCM based on weakly-supervised learning. It would provide new ideas and methods for other researchers who are working on medical entity recognition from clinical records of TCM in the field of TCM information acquisition and processing. Moreover, we expect that our research would promote the development of the field.

英文关键词： TCM Information Acquisition;TCM Information Processing;Semantic Modeling;Semantic Information;Semantic Association