医疗文本的小样本命名实体识别  

Few-shot Named Entity Recognition for Medical Text

作　　者：秦健 侯建新 谢怡宁[1] 何勇军[1] QIN Jian;HOU Jian-xin;XIE Yi-ning;HE Yong-jun(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,哈尔滨150080

出　　处：《哈尔滨理工大学学报》2021年第4期94-101,共8页Journal of Harbin University of Science and Technology

基　　金：国家自然科学基金面上项目(61673142);黑龙江省自然科学基金杰出青年项目(JJ2019JQ0013);哈尔滨市杰出青年人才基金(2017RAYXJ013);黑龙江省自然科学基金面上项目(F2017013);黑龙江省普通本科高等学校青年创新人才项目(UNPYSCT-2016034);黑龙江省教育厅科学技术研究项目(12511096);哈尔滨理工大学青年拔尖创新人才(20152);中国博士后基金(20132303120003).

摘　　要：针对医疗文本命名实体识别缺乏足够标记数据的问题,提出了一种新的命名实体识别神经网络和数据增强方法。该方法首先利用汉字的拼音和笔画来扩展Bert词向量,使Bert词向量能够包含更多先验知识;接着将命名实体识别模型与分词模型进行联合训练,以增强模型对于实体边界的判别能力;最后采用改进的数据增强方法处理训练数据,能够在避免模型过拟合的同时增加模型对于命名实体的识别效果。在CCKS-2019提供的电子病历文本上的实验结果表明,所提出的方法在小样本情况下能有效提高命名实体识别的准确率,在训练数据减少一半的情况下,识别率仍能保持没有明显下降。Aiming at the problem that medical text named entity recognition lacks sufficient labeled data,a newly named entity recognition deep neural network and data enhancement method is proposed.First of all,the Bert word vector is extended with pinyin and strokes of Chinese characters to make it contain more useful information.Then the named entity recognition model and the word segmentation model are jointly trained to enhance the model′s ability to recognize entity boundaries.Finally,an improved data enhancement method is used to process the training data,which can increase the recognition effect of the model on named entities while avoiding overfitting of the model.The experimental results on the electronic medical record text provided by CCKS-2019 show that the proposed method can effectively improve the accuracy of named entity recognition in the case of small samples and the recognition rate can still be maintained without a significant decrease when the training data is reduced by half.

关 键 词：命名实体识别 小样本 数据增强 联合训练 特征融合 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]