[1]梁兵涛,倪云峰.基于集成学习的中文命名实体识别方法[J].南京师大学报(自然科学版),2022,45(03):123-131.[doi:10.3969/j.issn.1001-4616.2022.03.016]
　Liang Bingtao,Ni Yunfeng.Chinese Named Entity Recognition Method Based on Ensemble Learning[J].Journal of Nanjing Normal University(Natural Science Edition),2022,45(03):123-131.[doi:10.3969/j.issn.1001-4616.2022.03.016]
点击复制

基于集成学习的中文命名实体识别方法()
分享到：

《南京师大学报（自然科学版）》[ISSN:1001-4616/CN:32-1239/N]

卷:
第45卷
期数:
2022年03期
页码:
123-131
栏目:
计算机科学与技术
出版日期:
2022-09-15

文章信息/Info

Title:
Chinese Named Entity Recognition Method Based on Ensemble Learning
文章编号:
1001-4616(2022)03-0123-09
作者:
梁兵涛1; 倪云峰2
(1.杭州优行科技有限公司,浙江 杭州 310000)(2.西安科技大学通信与信息工程学院,陕西 西安 710600)
Author(s):
Liang Bingtao1; Ni Yunfeng2
(1.Hangzhou Youxing Technology CO.,LTD.,Zhejiang 310000,China)(2.College of Communication and Information Engineering,Xi'an University of Science and Technology,Xi'an 710600,China)
关键词:
命名实体识别; BERT模型; 集成学习; 注意力机制; 迭代膨胀卷积网络
Keywords:
named entity recognition; BERT model; ensemble learning; attention mechanism; IDCNN
分类号:
TP391
DOI:
10.3969/j.issn.1001-4616.2022.03.016
文献标志码:
A
摘要:
针对中文命名实体识别经典的BiLSTM-CRF(bi-directional long short-term memory-conditional random field)模型存在的嵌入向量无法表征多义词、编码层建模时注意力分散以及缺少对局部空间特征捕获的问题,本文提出一种融合BERT-BiGRU-MHA-CRF和BERT-IDCNN-CRF模型优势的集成模型完成命名实体识别. 该方法利用裁剪的BERT模型得到包含上下文信息的语义向量; 再将语义向量输入BiGRU-MHA(bi-directional gated recurrent unit-multi head attention)及IDCNN(Iterated Dilated Convolutional Neural Network)网络. 前者捕获输入序列的时序特征并能够根据字符重要性分配权值,后者主要捕获输入的空间特征,利用平均集成方式将捕获到的特征融合; 最后通过CRF层获得全局最优的标注序列. 集成模型在人民日报和微软亚洲研究院(Microsoft research asia,MSRA)数据集上的F1值分别达到了96.09%和95.01%. 相较于单个模型分别提高了0.74%和0.55%以上,验证了本文方法的有效性.
Abstract:
Aiming at the problems existing in the classical BiLSTM-CRF(bi-directional long short-term memory-conditional random field)model of Chinese named entity recognition,such as the inability of the embedding vector cannot represent polysemy,the attention of the coding layer is distracted and lack of capturing local spatial features. This paper proposes an ensemble model that combines the advantages of the BERT-BiGRU-MHA-CRF and BERT-IDCNN-CRF models to complete named entity recognition. This method uses the BERT model to obtain a semantic vector containing contextual information,and then inputs the semantic vector into BiGRU-MHA(bi-directional gated recurrent unit-multi head attention)and IDCNN(Iterated Dilated Convolutional Neural Network)networks. The former captures the timing characteristics of the input sequence and can assign weights according to the importance of the characters,the latter mainly captures the spatial characteristics of the input,and uses the mean ensemble method to fuse the captured features. Finally,the global optimal annotation sequence is obtained through the CRF layer. The F1 values of the ensemble model on the datasets of People's Daily and Microsoft Research Asia(MSRA)reached 96.09% and 95.01%,respectively. Compared with the single model,it has increased by more than 0.74% and 0.55%,respectively,which verifies the effectiveness of the method in this paper.

参考文献/References:

[1]HAMMERTON J. Named entity recognition with long short-term memory[C]//Proceedings of the Seventh Conference on Natural Language Learning at Annual Conference of the North American Chapter of the Association for Computational Linguistics:Human Language Technologies 2003. 2003:172-175.
[2]LAMPLE G,BALLESTEROS M,SUBRAMANIAN S,et al. Neural architectures for named entity recognition[J]. arXiv preprint arXiv:1603.01360,2016.
[3]ZHANG Y,YANG J. Chinese NER using lattice LSTM[J]. arXiv preprint arXiv:1805.02023,2018.
[4]GUI T,MA R,ZHANG Q,et al. CNN-Based Chinese NER with Lexicon Rethinking[C]//Twenty-Eighth International Joint Conference on Artificial Intelligence. Macao:Springer,2019:4982-4988.
[5]WANG C Q,CHEN W,XU B. Named entity recognition with gated convolutional neural networks[M]//Chinese Computational Linguistics and Natural Language Processing Based on Naturally Annotated Big Data. Springer,Cham,2017:110-121.
[6]YU F,KOLTUN V. Multi-scale context aggregation by dilated convolutions[J]. arXiv preprint arXiv:1511.07122,2015.
[7]YU B H,WEI J X. IDCNN-CRF-based domain named entity recognition method[C]//2020 IEEE 2nd International Conference on Civil Aviation Safety and Information Technology(IEEE 2nd International Conference on Civil Aviation Safety and Information Technology:542-546.
[8]杨晓辉,毕雪华,张琳琳,等. 基于多任务的中文电子病历中命名实体识别研究[J]. 东北师大学报(自然科学版),2020,52(1):81-87. DOI:10.16163/j.cnki.22-1123/n.2020.01.016.
[9]孙弋,梁兵涛. 基于BERT和多头注意力的中文命名实体识别方法[J/OL]. 重庆邮电大学学报(自然科学版):1-10[2022-02-13]. http://kns.cnki.net/kcms/detail/50.1181.N.20211209.2010.004.html.
[10]张柯文,李翔,严云洋,等. 基于多特征双向门控神经网络的领域专家实体抽取方法[J]. 南京师大学报(自然科学版),2021,44(1):128-135.
[11]孔祥鹏,吾守尔·斯拉木,杨启萌,等. 基于迁移学习的维吾尔语命名实体识别[J]. 东北师大学报(自然科学版),2020,52(2):58-65. DOI:10.16163/j.cnki.22-1123/n.2020.02.010.
[12]李妮,关焕梅,杨飘,等. 基于BERT-IDCNN-CRF的中文命名实体识别方法[J]. 山东大学学报(理学版),2020,55(1):102-109.
[13]周志华. 机器学习[M]. 北京:清华大学出版社,2016.
[14]石春丹,秦岭. 基于BGRU-CRF的中文命名实体识别方法[J]. 计算机科学,2019,46(9):237-242.
[15]杨飘,董文永. 基于BERT嵌入的中文命名实体识别方法[J]. 计算机工程,2020,46(4):40-45,52. DOI:10.19678/j.issn.1000-3428.0054272.

相似文献/References:

[1]张柯文,李 翔,严云洋,等.基于多特征双向门控神经网络的领域专家实体抽取方法[J].南京师大学报(自然科学版),2021,44(01):128.[doi:10.3969/j.issn.1001-4616.2021.01.018]
　Zhang Kewen,Li Xiang,Yan Yunyang,et al.Domain Expert Entity Extraction Method Based on Multi-FeatureBidirectional Gated Neural Network[J].Journal of Nanjing Normal University(Natural Science Edition),2021,44(03):128.[doi:10.3969/j.issn.1001-4616.2021.01.018]

备注/Memo

备注/Memo:
收稿日期:2022-01-10.
通讯作者:倪云峰,教授,博士,研究方向:煤矿信息化技术研究等. E-mail:635129613@qq.com
更新日期/Last Update: 2022-09-15