基于FCN-CRF的医疗命名实体识别  

FCN-CRF based medical named entity recognition

作　　者：潘胜星 唐雅娟[1] PAN Shengxing;TANG Yajuan(Department of Electronic Engineering,Shantou University,Shantou 515063,China)

机构地区：[1]汕头大学电子工程系,广东汕头515063

出　　处：《电子设计工程》2022年第17期60-63,68,共5页Electronic Design Engineering

基　　金：国家自然科学基金(82071992);广东省基础与应用基础研究基金(2020B15120061);广东省高校重点领域(人工智能)项目(2019KZDZX1013);国家重点科研项目(2020YFC0122103);广东省科技计划重点项目(2015B020233018)。

摘　　要：基于词典增强的方法存在搜索潜在匹配词较慢、难以批量计算等问题。针对该问题,对比计算机视觉领域的语义分割任务与自然语言处理领域的命名实体识别任务之间的相似性,提出了引入图像语义分割中的全卷积网络模型(Fully Convolutional Network),并将其应用于医疗文本的命名实体识别任务。在瑞金糖尿病数据集上进行实验,结果表明,相比基于词典增强的方法,该方法能达到F1值差距在1%内,同时无需引入词级别信息且便于进行批量计算,而在实验数据集上实际运行时间仅为基于词典增强方法的不到2%。The lexicon-based methods is slow in searching latent match words and is hard to be calculated in batch. According to that,compares the similarity of semantic segmentation task in CV and named entity recognition task in NLP,introduces the fully convolutional network model which is used in semantic segmentation,and applied it to named entity recognition task in medical text. The test result on Ruijin diabetes dataset shows that compared to lexicon-based methods,this method can achieve similar F1score that diffirence in less than 1% without lexicon information and is easy to be calculated in batch.Meanwhile,the actual running time on the dataset is only less than 2% of times that lexicon-based methods cost.

关 键 词：自然语言处理 命名实体识别 全卷积网络 条件随机场 语义分割 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]