基于迁移学习的维吾尔语命名实体识别  

Uyghur named entity recognition based on transfer learning

作　　者：孔祥鹏 吾守尔·斯拉木[2] 杨启萌 李哲 KONG Xiang-peng;SILAMU·Wushouer;YANG Qi-meng;LI Zhe(College of Software,Xinjiang University,Urumqi 830046,China;College of Information Science and Engineering,Xinjiang University,Urumqi 830046,China)

机构地区：[1]新疆大学软件学院,新疆乌鲁木齐830046 [2]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《东北师大学报（自然科学版）》2020年第2期58-65,共8页Journal of Northeast Normal University(Natural Science Edition)

基　　金：国家“973”重点基础研究计划项目(2014CB340506);国家自然科学基金资助项目(61363063);国家“十三五”科研规划2017年度科研项目(ZDA135-6).

摘　　要：针对维吾尔语命名实体识别存在无法关注词序列依赖关系等问题,提出一种基于迁移学习的联合深度模型(TBIBC).首先通过BERT(Bidirectional Encoder Representations from Transformers)预训练中文数据集生成具有语义信息的词向量,再将词向量序列通过空洞卷积神经网络(IDCNN)进行膨胀操作以减少神经元层数和参数,输出向量输入到双向门控循环单元(BiGRU)进行上下文语义信息提取,然后通过CRF层得到最优标签序列.模型采用共享深度神经网络隐藏层的方法将训练好的中文实体识别模型迁移到维吾尔语模型上.结果表明,该模型的准确率为91.39%,召回率为90.11%,F 1值达到90.75%,能显著提升维吾尔语命名实体识别性能.Aiming at the problem that Uygur named entity recognition can not pay attention to word sequence dependency,a hybrid depth model based on transfer learning(TBIBC)is proposed.Firstly,the Chinese data set is trained by BERT(Bidirectional Encoder Representations from Transformers)to generate word vectors with semantic information.Then,the word vectors are expanded by IDCNN to reduce the number and parameters of neurons.The output vectors are then input to BiGRU for context semantic information extraction.The optimal tag sequence is obtained through CRF layer.The model uses the method of shared depth neural network hidden layer to transfer the trained Chinese entity recognition model to the Uygur language model.The accuracy of the model is 91.39%,the recall rate is 90.11%,and the F 1 value is 90.75%,which can significantly improve the performance of Uygur named entity recognition.

关 键 词：迁移学习 BERT BiGRU 空洞卷积网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]