结合近邻分析的小样本命名实体识别方法  

Few shot named entity recognition method combinedwith neighborhood component analysis

作　　者：江汀莹 线岩团[1,2] 王红斌[1,2] JIANG Tingying;XIAN Yantuan;WANG Hongbin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]昆明理工大学云南省人工智能重点实验室,云南昆明650500

出　　处：《现代电子技术》2023年第19期88-94,共7页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(62266028);云南省重大科技专项计划项目课题(202202AD080003⁃02)。

摘　　要：小样本命名实体识别通过少量标注样本构建可应用于未知类别的实体识别模型,是命名实体识别的新兴研究方向。小样本实体识别具有实体边界检测不准确以及训练数据数量少、可见实体类别少的特点,导致模型性能表现不佳,模型极易过拟合。针对这个问题,提出结合近邻分析的小样本命名实体识别方法,将命名实体识别任务分成实体范围检测和实体类别预测两部分完成,其中实体范围检测部分使用片段打分的方法,实体类别预测部分使用近邻分析的方法。此外,为了缓解训练数据少带来的模型过拟合问题,还在实体特征的隐空间上进行Mixup数据增强以扩增伪样本。实验结果表明,所提出的结合近邻分析的小样本命名实体识别方法有效地提高了小样本实体识别的性能。Few shot named entity recognition(NER)is a new research direction of NER,which is to build an entity recognition model that can be applied to unknown categories by a small number of labeled samples.Few shot entity recognition is characterized by inaccurate entity boundary detection,small amount of training data and few visible entity categories,which leads to poor model performance and prone to overfitting.In view of this,a few shot named entity recognition method combined with neighborhood component analysis(FSNCA)is proposed.In the method,the NER task is divided into two parts of entity range detection and entity category prediction.The former employs fragment scoring method,while the later employs neighborhood component analysis(NCA)method.In addition,Mixup data enhancement is carried out on the hidden space of entity features for pseudo⁃sample amplification,so as to alleviate the model overfitting caused by less training data.Experimental results show that the proposed FSNCA can improve the performance of few shot entity recognition effectively.

关 键 词：小样本 实体识别 实体分类 过拟合 近邻分析算法 数据增强 泛化性能 深度神经网络 

分 类 号：TN911.1-34[电子电信—通信与信息系统] TP391.1[电子电信—信息与通信工程]